Meme is an expressive medium that often conveys rich emotions and intentions. Recent studies have confirmed the critical role of metaphors in meme understanding. However, existing metaphor research heavily relies on manual annotations, and mainstream vision-language models (VLMs) still struggle with the recognition and comprehension of metaphors. To address these challenges, we introduce MetaGPT, the first vision-language model specifically designed for meme metaphor understanding. MetaGPT is capable of identifying and extracting metaphors in memes, and generating accurate meme interpretations. Furthermore, we construct a dedicated dataset for meme understanding, MUnd, which comprises approximately 32,000 high-quality question-answer (QA) pairs across three core tasks: metaphor detection, metaphor domain extraction, and meme interpretation. Based on MUnd, we further propose an evaluation benchmark for meme understanding and conduct a comprehensive assessment of existing VLMs. Experimental results reveal that current models still face challenges in metaphor comprehension, while MetaGPT consistently outperforms them across all tasks, highlighting its potential in advancing meme understanding.

Multimodal Large Language Models (MLLMs) face critical privacy challenges arising from the indiscriminate memorization of sensitive data.Existing unlearning methods often fail to precisely disentangle specific instances from general concepts, leading to either catastrophic forgetting of useful knowledge or unsafe content substitution.We attribute these failures to a fundamental geometric mismatch: these approaches primarily operate in Euclidean space, which lacks the capacity to model the hierarchical entailment inherent in visual-linguistic concepts.To address this, we introduce LOTUS (LOrentz Transport for Unlearning Strategies), a framework that performs surgical semantic pruning within the Lorentz manifold.LOTUS employs an Inverted Entailment Cone Loss to sever the semantic inheritance of sensitive concepts and a Lorentz Transport mechanism to align pruned features with a safety refusal prior in the tangent space.Extensive experiments on MLLMU-Bench demonstrate that LOTUS significantly outperforms baselines, improving unlearning efficacy by over 9% on LLaVA compared to state-of-the-art constraint-based methods.Crucially, LOTUS achieves this precision while maintaining general utility, effectively resolving the dilemma between thorough erasure and model stability.

Human cognition excels at extending knowledge through analogy, where word meanings evolve along structured pathways from concrete prototypes to abstract senses via metaphor and metonymy.Do Large Language Models (LLMs) internalize this generative logic, or merely mimic statistical patterns?To investigate this, we introduce CogEvolve, a cognitive linguistic benchmark designed to test these evolutionary pathways across textual and visual modalities.Our evaluation reveals a distinct cognitive profile: models function as "Super-Associators" expert at static recognition yet fail at causal reasoning.In text, they exhibit a Frequency-Primacy Conflation, confusing statistical prevalence with cognitive basicness.Crucially, this reasoning collapses further in the visual domain.We term this deficit the Ungrounded Arrow: models possess high-fidelity concept representations (the "dots") but lack the transformational operators (the "arrows") essential for true relational understanding 1 .

Through preliminary studies, we observed that some figurative language detection tasks show consistent biases: a conservative preference for precision over recall and higher agreement within a model’s own judgments than between different models. To address these biases, we first introduce the Inter-Intra Agreement Ratio (IIAR), a novel metric that systematically quantifies LLM generalizability by comparing intra-and inter-model agreement patterns. Then, we propose Consensus-Based LLM Prediction Aggregation (CLPA), which leverages the collective strengths of multiple LLMs to mitigate individual model biases and achieve balanced precision and recall—without fine-tuning or compromising generality. Finally, we present a human–AI co-annotation framework that uses minimal expert input to efficiently determine optimal consensus thresholds, significantly reducing annotation effort. Experiments on the SemEval-2017 pun detection datasets demonstrate that our approach achieves a new state-of-the-art F1 score of 0.963 on heterographic pun detection through multi-LLM consensus aggregation; furthermore, our co-annotation framework requires only 10% of golden labels to accurately determine optimal consensus thresholds for capturing nuanced human annotation boundaries.

Hateful videos have become prevalent on online platforms, highlighting an urgent need for effective detection. However, existing studies primarily focus on binary classification and fail to provide contextual rationales that reveal the implicit meanings behind these judgments, significantly undermining model explainability. To fill this gap, we aim to achieve explainable hateful video detection, enabling models to provide contextual rationales that integrate relevant evidence and logical reasoning alongside decisions. This approach can comprehensively enhance the understanding of video content and the explainability of the decision-making process. We first introduce two datasets, Ex-HateMM and Ex-ImpliHateVid, for explainable hateful video detection. Each dataset provides fine-grained annotations of multimodal harmful elements, along with contextual rationales. We then propose an Information Augmentation and Reasoning Enhancement (IARE) framework designed for explainable detection. The framework employs an information augmentation phase that leverages the multimodal chain-of-thought to integrate harmful elements, thereby enriching rationale evidence. Additionally, IARE incorporates a reasoning enhancement phase, in which Direct Preference Optimization guides the model toward correct reasoning paths and away from incorrect ones, thereby improving the logical coherence of its justifications. We conduct extensive experiments on the two datasets, comparing multiple baselines with our proposed IARE framework. The results demonstrate that IARE achieves state-of-the-art performance while also generating accurate rationales.

Metaphors are a fundamental cognitive tool for articulating abstract and subjective experiences and implicit semantics, making them potent indicators of psychological state, particularly in individuals with depression. The proliferation of social media has created a vast repository of such metaphorical expressions, offering an unprecedented opportunity to understand mental health struggles. These metaphors can provide crucial insights for clinical assessment and therapeutic intervention. However, their potential remains largely untapped in automated depression detection, primarily due to the lack of large-scale, annotated datasets. To bridge this gap, we introduce the Depression-Related Metaphor Dataset (DRMD), a novel resource of social media posts related to depression, incorporating depression levels (severe, moderate, minimum, and null), the presence or absence of metaphors, and their conceptual source domain mappings. We leverage this dataset to fine-tune Large Language Models (LLMs), integrating metaphorical features to enhance detection capabilities. Our results demonstrate that models incorporating metaphorical information achieve superior accuracy in depression detection and, importantly, generate high-quality explanations for their decisions by referencing specific metaphorical expressions. This work underscores the critical role of metaphorical analysis in computational mental health and provides a foundation for future research in explainable AI for depression detection. The dataset is publicly available.

Senqi Yang, Dongyu Zhang, Jing Ren, Ziqi Xu, Xiuzhen Zhang, Yiliao Song, Hongfei Lin, Feng Xia. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2025.

Bragging, as a pervasive social-linguistic phenomenon, reflects complex human interaction patterns.However, the understanding and generation of appropriate bragging behavior in large language models (LLMs) remains underexplored.In this paper, we propose a comprehensive study that combines analytical and controllable approaches to examine bragging in LLMs.We design three tasks, bragging recognition, bragging explanation, and bragging generation, along with novel evaluation metrics to assess the models' ability to identify bragging intent, social appropriateness, and account for context sensitivity.Our analysis reveals the challenges of bragging in the social context, such as recognizing bragging and responding appropriately with bragging in conversation.This work provides new insights into how LLMs process bragging and highlights the need for more research on generating contextually appropriate behavior in LLMs 1 .

Implicit hate speech has become a significant challenge for online platforms, as it often avoids detection by large language models (LLMs) due to its indirectly expressed hateful intent.This study identifies the limitations of LLMs in detecting implicit hate speech, particularly when disguised as seemingly harmless expressions in a rhetorical device.To address this challenge, we employ a Jailbreaking strategy and Energy-based Constrained Decoding techniques, and design a small model for measuring the energy of metaphorical rhetoric.This approach can lead to LLMs generating metaphorical implicit hate speech.Our research reveals that advanced LLMs, like GPT-4o, frequently misinterpret metaphorical implicit hate speech, and fail to prevent its propagation effectively.Even specialized models, like ShieldGemma and LlamaGuard, demonstrate inadequacies in blocking such content, often misclassifying it as harmless speech.This work points out the vulnerability of current LLMs to implicit hate speech, and emphasizes the improvements to address hate speech threats better.

Major depressive disorder (MDD) is a significant global health burden; timely and precise diagnosis is essential to reduce relapse and mortality. However, existing approaches typically frame depression diagnosis as a monolithic classification problem, neglecting the multi-dimensional and hierarchical reasoning that underpins clinical interviews. We introduce DALE, a modular framework that mirrors psychiatrists'reasoning by explicitly disentangling four dimensions of patient narratives-psychological symptoms, somatic symptoms, protective factors and stressors. Leveraging GPT-4, we augment the dataset with dialogue-level annotations that map patient interview onto a list of attributes. On this auxiliary data, we train four domain-specialised LoRA adapters atop a frozen LLM; each adapter conducts a brief diagnostic dialogue and produces a concise report of its domain. A lightweight classifier then integrates these reports to generate a final summary and predict depression and suicide risk. Experiments on the D4 psychiatric dialogue benchmark show that DALE showing strong performance, while requiring far fewer trainable parameters, and yields interpretable, attribute-level evidence for its predictions.

Automatic detection of Adverse Drug Events (ADEs) from social media has become increasingly important for post-market drug safety surveillance and pharmacovigilance. Although existing social media-based ADE detection methods have effectively addressed some challenges such as data sparsity and class imbalance, they still suffer from spurious correlations, where models tend to learn co-occurrence patterns between drugs and symptoms, leading them to incorrectly identify drug inefficacy or therapeutic intent as ADEs. To address this issue, we propose a structured medical knowledge-guided counterfactual generation method that leverages authoritative medical databases and large language models to construct clinically plausible counterfactual samples, thereby mitigating spurious correlations at the data level. Furthermore, we propose a knowledge-enhanced instruction-tuning strategy that injects drug indications as causal cues into model inputs to enhance its causal reasoning capabilities. Experimental results on two benchmark datasets demonstrate that our method consistently outperforms state-of-the-art models, effectively alleviating spurious correlations and exhibiting superior capabilities in drug-symptom relationship identification.

With the rapid development of online medical platforms, consumer health questions (CHQs) are inefficient in diagnosis due to redundant information and frequent non-professional terms. The medical question summary (MQS) task aims to transform CHQs into streamlined doctors' frequently asked questions (FAQs), but existing methods still face challenges such as poor identification of question focus and low summary faithfulness. This paper explores the potential of large language models (LLMs) in the MQS task and finds that direct fine-tuning is prone to focus identification bias and generates summaries with low faithfulness. To this end, we propose an optimization framework based on core focus guidance. First, a prompt template is designed to drive the LLMs to extract the core focus from the CHQs that is faithful to the original text. Then, a fine-tuning dataset is constructed in combination with the original CHQ-FAQ pairs to improve the ability to identify the focus of the question. Finally, a multi-dimensional quality evaluation and selection mechanism is proposed to comprehensively improve the quality of the summary from multiple dimensions. We conduct comprehensive experiments on two widely-adopted MQS datasets using three established evaluation metrics. The proposed framework achieves state-of-the-art performance across all measures, demonstrating a significant boost in the model's ability to identify the critical focus of questions and a notable improvement in the faithfulness of the summary. The source codes are freely available at https://github.com/DUT-LiuChao/FocusMed.

Accurate recognition of biomedical named entities is critical for medical information extraction and knowledge discovery. However, existing methods often struggle with nested entities, entity boundary ambiguity, and cross-lingual generalization. In this paper, we propose a unified Biomedical Named Entity Recognition (BioNER) framework based on Large Language Models (LLMs). We first reformulate BioNER as a text generation task and design a symbolic tagging strategy to jointly handle both flat and nested entities with explicit boundary annotation. To enhance multilingual and multi-task generalization, we perform bilingual joint fine-tuning across multiple Chinese and English datasets. Additionally, we introduce a contrastive learning-based entity selector that filters incorrect or spurious predictions by leveraging boundary-sensitive positive and negative samples. Experimental results on four benchmark datasets and two unseen corpora show that our method achieves state-of-the-art performance and robust zero-shot generalization across languages. The source codes are freely available at https://github.com/dreamer-tx/LLMNER.

Biomedical question answering (QA) requires precise interpretation of complex medical knowledge. Large language models (LLMs) and retrieval-augmented generation (RAG) leverage external medical literature but often produce hallucinations due to noisy retrieval and insufficient verification. We propose MedTrust-Guided Iterative RAG, a framework that improves factual consistency and reduces hallucinations in medical QA. It introduces three innovations. First, citation-aware reasoning grounds generation in retrieved documents and uses Negative Knowledge Assertions when evidence is missing. Second, an iterative retrieval-verification process refines queries through Medical Gap Analysis. Third, the MedTrust-Align Module (MTAM) applies Direct Preference Optimization to align generation with verified evidence and suppress hallucination-prone patterns.

Current research reveals that the herpes zoster vaccine(HZV) can effectively prevent and mitigate Alzheimer's disease(AD). However, the underlying mechanisms linking the HZV and AD remain incompletely understood. We employ a literature-based discovery (LBD) paradigm to investigate these mechanisms. Traditional knowledge graph-based approaches demonstrate strong performance in implicit knowledge discovery tasks. However, existing methods still face critical challenges: (1) The diverse representations of biomedical entities often degrade knowledge graph quality, such as introducing path redundancy; (2) Current path-ranking mechanisms lack rigorous biological plausibility assessment, resulting in a high number of false-positive paths being retained. To address the aforementioned challenges, this paper proposes a Synergistic Integration framework of Knowledge graphs and Large language models (SIKL). First, the PubTator3.0 tool and manual rule-based methods are employed to extract entities and relational triples from abstract content. A knowledge graph is constructed with medical subject headings as nodes, integrating multiple attributes and relations, effectively mitigating path redundancy caused by diverse entity expressions. Next, a twostage LLM-driven path filtering module is designed. In the first stage, a retrieval-augmented large language model assesses whether candidate paths meet causality conditions, performing preliminary filtering to eliminate false positives. The second stage leverages a chain-of-thought large language model to conduct step-by-step reasoning on remaining paths, further evaluating their biological plausibility. Finally, a comprehensive path ranking strategy combines literature support counts and LLM-generated scores to output Top-p high-confidence hypothetical paths. Experimental results reveals that HZV may delay AD progression through pathways such as neuroinflammatory regulation, with partial mechanisms supported by literature.

Biomedical document-level relation extraction plays a crucial role in mining structured knowledge from biomedical texts. However, existing large language model-based methods struggle with information redundancy, noise interference, and insufficient domain-specific knowledge when processing complex documents. To address these challenges, we propose Bio-$\mathbf{R}^{\mathbf{3}}$, a novel framework that enhances document-level relation extraction through entity aware rewriting and domain-adaptive reasoning. Our Bio-R${ }^{3}$follows a “Rewrite-Retrieval-Reason” pipeline: (1) Entity aware rewriting refocuses the document content on target entity pairs, reducing irrelevant context; (2) Retrieval-augmented generation supplements the document with external biomedical evidence to support relation inference; and (3) Chain-of-Thought prompting and few-shot learning improve reasoning accuracy, while Low-Rank Adaptation fine-tuning enhances domain adaptation. Experiments on the CDR and GDA benchmarks demonstrate that Bio-$\mathbf{R}^{\mathbf{3}}$significantly outperforms existing methods, demonstrating its effectiveness in the extraction of biomedical document relations.

Large Language Models (LLMs) have shown strong abilities in language understanding and reasoning, drawing increasing attention in medical question answering (QA). While retrieval-augmented generation (RAG) methods improve factual accuracy, existing approaches still struggle to retrieve and organize relevant knowledge effectively. To overcome this, we propose NaviPath, a knowledge graph-based RAG framework for medical QA. It enhances LLM responses through a structured prompt built in three steps: (1) extended entity retrieval, (2) multiperspective reasoning path construction, and (3) natural language transformation for better comprehension. Experiments on two medical QA benchmarks show that NaviPath achieves state-of-the-art performance in diagnostic accuracy and factual consistency. The implementation is available at https://github.com/zyr319/Navipath.git.

Generative retrieval enhances retrieval effectiveness by generating natural language represented document identifiers. However, current methods often struggle with two major challenges: limited identifier quality and insufficient query-document interaction, leading to limited retrieval performance. To tackle these challenges, we propose a novel generative retrieval framework integrated with semantic-aligned multi-layer identifiers and reinforcement learning. To improve identifier quality, we design a prompt-driven multi-task learning strategy to generate three types of hierarchical identifiers: summary, keyword, and pseudo-query, to capture multi-granularity document semantics. Furthermore, we adopt supervised fine-tuning to integrate these identifiers. To improve query-document interaction, we devise a multi-view ranking fusion mechanism that combines retrieval results across multi-layer identifiers. We further employ a GRPO-based reinforcement learning based on dense similarity rewards and a difficulty-aware negative sampling strategy to optimize the generated identifiers. Experiments on multiple benchmark datasets show that our framework significantly outperforms existing generative retrieval methods, offering a promising solution for building more effective and semantically aligned retrieval systems. The code for our model is publicly available at https://github.com/yicentian02/GRAM-RL.

Multimodal hate detection aims to identify hate content across multiple modalities for promoting a harmonious online environment. Despite promising progress, three critical challenges, the absence of implicit hateful cues, the cross-modal-induced hate, and the diversity of hate target groups, inherent in the multimodal hate detection task, have been overlooked. To address these challenges, we propose a hypergraph-based prompting fusion model. Our model first uses tailored prompts to infer implicit hateful cues. It then introduces hyperedges to capture cross-modal-induced hate and applies a diversity-oriented hyperedge expansion strategy to account for different hate target groups. Finally, hypergraph convolution fuses diverse hateful cues, enhancing the exploration of cross-modal hate and targeting specific groups. Experimental results on two benchmark datasets show that our model achieves state-of-the-art performance in multimodal hate detection.

Stance detection aims to identify the attitude expressed in text towards a specific target.Recent studies on zero-shot and few-shot stance detection focus primarily on learning generalized representations from explicit targets.However, these methods often neglect implicit yet semantically important targets and fail to adaptively adjust the relative contributions of text and target in light of contextual dependencies.To overcome these limitations, we propose a novel two-stage framework: First, a data augmentation framework named Hierarchical Collaborative Target Augmentation (HCTA) employs Large Language Models (LLMs) to identify and annotate implicit targets via Chain-of-Thought (CoT) prompting and multi-LLM voting, significantly enriching training data with latent semantic relations.Second, we introduce DyMCA, a Dynamic Multi-level Contextaware Attention Network, integrating a joint text-target encoding and a content-aware mechanism to dynamically adjust text-target contributions based on context.Experiments on the benchmark dataset demonstrate that our approach achieves state-of-the-art results, confirming the effectiveness of implicit target augmentation and fine-grained contextual modeling.Our code is publicly available at https: //github.com/EliaukoaYoW/DyMCA.

Patronizing and Condescending Language (PCL) is a form of discriminatory toxic speech targeting vulnerable groups, threatening both online and offline safety. While toxic speech research has mainly focused on overt toxicity, such as hate speech, microaggressions in the form of PCL remain underexplored. Additionally, dominant groups’ discriminatory facial expressions and attitudes toward vulnerable communities can be more impactful than verbal cues, yet these frame features are often overlooked. In this paper, we introduce the PCLMM dataset, the first Chinese multimodal dataset for PCL, consisting of 715 annotated videos from Bilibili, with high-quality PCL facial frame spans. We also propose the MultiPCL detector, featuring a facial expression detection module for PCL recognition, demonstrating the effectiveness of modality complementarity in this challenging task. Our work makes an important contribution to advancing microaggression detection within the domain of toxic speech.

Early detection of Alzheimer's disease (AD) through spontaneous speech analysis represents a promising, non-invasive diagnostic approach. Existing methods predominantly rely on fusion-based multimodal deep learning, effectively integrating linguistic and acoustic features. However, these methods inadequately model higher-order interactions between modalities, reducing diagnostic accuracy. To address this, we introduce SpeechHGT, a multimodal hypergraph transformer designed to capture and learn higher-order interactions in spontaneous speech features. SpeechHGT encodes multimodal features as hypergraphs, where nodes represent individual features and hyperedges represent grouped interactions. A novel hypergraph attention mechanism enables robust modeling of both pairwise and higher-order interactions. Experimental evaluations on the DementiaBank datasets reveal that SpeechHGT achieves state-of-the-art performance, surpassing baseline models in accuracy and F1 score. These results highlight the potential of hypergraph-based models to improve AI-driven diagnostic tools for early AD detection.

The rapid development of large language models has greatly advanced human-computer dialogue research. However, applying these models to specialized fields like maternity and infant care often leads to subpar performance due to a lack of domain-specific datasets. To address this problem, we have created MicDialogue, a Chinese dialogue dataset for maternity and infant care. MicDialogue involves a wide range of specialized topics, including gynecological health, pediatric care, pregnancy preparation, emotional counseling and other related topics. This dataset is curated from two types of Chinese social media: short videos and blog posts. Short videos capture real-time interactions and pragmatic dialogue patterns, while blog posts offer comprehensive coverage of various topics within the domain. We have also included detailed annotations for topics, diseases, symptoms, and causes, enabling in-depth research. Additionally, we developed a knowledge-driven benchmark model using LLM-based prompt learning and multiple knowledge graphs to address diverse dialogue topics. Experiments validate MicDialogue's usability, providing benchmarks for future research and essential data for fine-tuning language models in maternity and infant care.

With the development of wearable cardiac ultrasound devices, it is no longer sufficient to solely rely on doctors for diagnosing long-term echocardiogram videos. Automated diagnosis of echocardiogram videos has now become a research hotspot. Existing studies only analyze echocardiogram video through discriminative models, which have limited question-answering capabilities. Therefore, this study innovatively proposes a large language model with cardiac ultrasound diagnostic capabilities—EchoGPT. EchoGPT integrates the robust communication and comprehension capabilities of large language models (LLMs) with the diagnostic prowess of traditional medical models, empowering patients to obtain accurate medical indicator data and comprehend their health conditions through interactive questioning with the model. The model is capable of local deployment on personal computers, effectively safe guarding user privacy. EchoGPT operates through three main components: left ventricle segmentation, left ventricular ejection fraction LVEF prediction, and finetuning of video-text LLMs. Experimental results demonstrate EchoGPT’s superior accuracy in predicting LVEF compared to other models, and positive feedback from professional physicians through questionnaire surveys, validating its potential in practical applications. The demo is available at https://github.com/zhuqh19/EchoGPT.

Haohao Zhu, Junyu Lu, Zeyuan Zeng, Zewen Bai, Xiaokun Zhang, Liang Yang, Hongfei Lin. Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers). 2025.

News recommender systems aim to provide personalized news reading experiences for users based on their reading history. Behavioral science studies suggest that screen-based news reading contains three successive steps: scanning, title reading, and then clicking. Adhering to these steps, we find that intra-news entity interest dominates the scanning stage, while the inter-news entity interest guides title reading and influences click decisions. Unfortunately, current methods overlook the unique utility of entities in news recommendation. To this end, we propose a novel method called IP2 to probe entity-guided reading interest at both intra- and inter-news levels. At the intra-news level, a Transformer-based entity encoder is devised to aggregate mentioned entities in the news title into one signature entity. Then, a signature entity-title contrastive pre-training is adopted to initialize entities with proper meanings using the news story context, which in the meantime facilitates us to probe for intra-news entity interest. As for the inter-news level, a dual tower user encoder is presented to capture inter-news reading interest from both the title meaning and entity sides. In addition to highlighting the contribution of inter-news entity guidance, a cross-tower attention link is adopted to calibrate title reading interest using inter-news entity interest, thus further aligning with real-world behavior. Extensive experiments on two real-world datasets demonstrate that our IP2 achieves state-of-the-art performance in news recommendation.

Hateful memes are prevalent on the Internet, raising the urgent need for effective detection.Given their implicit nature, incorporating rationales with background knowledge is crucial for enhancing model understanding.However, existing methods often suffer from limited quality of external rationales and misalignment with original meme information.These challenges hinder model comprehension, leading to reduced accuracy and explainability.To address these challenges, we propose a Multimodal Multi-agent Knowledge Enhanced (M2KE) framework for hateful meme detection.M2KE introduces a multi-agent rationale discovery mechanism to extract high-quality rationales relevant to meme content and an adaptive knowledge interaction mechanism to ensure alignment between original meme information and external rationales.Specifically, multi-agent rationale discovery mechanism improves the reliability of rationales by collaboratively verifying and refining them with multiple agents, supported by large language models (LLMs) due to their extensive knowledge.And adaptive knowledge interaction mechanism uses information entropy to dynamically balance the model's attention between original meme information and external rationales, preventing over-reliance on rationales and enabling a more comprehensive understanding.Experimental results on three datasets demonstrate that M2KE significantly outperforms existing models.Further analysis underscores the importance of effectively integrating accurate rationales to enhance model performance.Disclaimer: Samples in this paper may be considered offensive.

MOTIVATION: With the advancement of large language models (LLMs), the field of biomedical document-level relation extraction (BioDocRE) has encountered new opportunities. However, LLMs often face challenges such as hallucinated generation, insufficient reasoning capabilities, and a lack of interpretability when performing relation extraction tasks. RESULTS: To address these issues, we propose the SyRACT (Synergistic Retrieval Augmented Generation and Chain of Thought) framework for high precision relation extraction in biomedical documents. This framework is built around three core strategies: (i) reframing the relation extraction task as a question answering problem to better align with the processing logic of LLMs; (ii) leveraging an external database constructed from PubMed to provide LLMs with rich and reliable contextual information, thus mitigating hallucination generation; and (iii) construct a specific Chain of Thought for BioDocRE tasks, thereby enhancing the model's reasoning ability and the interpretability of its output. We validated this approach on three biomedical relation extraction datasets: CDR, GDA, and ADE. Experimental results show that the SyRACT model improves F1 scores by 11.04%, 9.10%, and 41.00% on three datasets, respectively, compared to the DocRE method, which uses standard prompts for LLMs. AVAILABILITY AND IMPLEMENTATION: Our source code and data are available at https://github.com/donggggxin/SyRACT.

Abstract Large language models (LLMs) demonstrate remarkable text comprehension and generation capabilities but often lack the ability to utilize up-to-date or domain-specific knowledge not included in their training data. To address this gap, we introduce KEDiT, an efficient method for fine-tuning LLMs for knowledge-grounded dialogue generation. KEDiT operates in two main phases. First, it employs an information bottleneck to compress retrieved knowledge into learnable parameters, retaining essential information while minimizing computational overhead. Second, a lightweight knowledge-aware adapter integrates these compressed knowledge vectors into the LLM during fine-tuning, updating less than 2% of the model parameters. The experimental results on the Wizard of Wikipedia and a newly constructed PubMed-Dialog dataset demonstrate that KEDiT excels in generating contextually relevant and informative responses, outperforming competitive baselines in automatic, LLM-based, and human evaluations. This approach effectively combines the strengths of pretrained LLMs with the adaptability needed for incorporating dynamic knowledge, presenting a scalable solution for fields such as medicine.1

Empathetic response generation, aiming to understand the user’s situation and feelings and respond empathically, is crucial in building human-like dialogue systems. Traditional approaches typically employ maximum likelihood estimation as the optimization objective during training, yet fail to align the empathy levels between generated and target responses. To this end, we propose an empathetic response generation framework using reinforcement learning (EmpRL). The framework develops an effective empathy reward function and generates empathetic responses by maximizing the expected reward through reinforcement learning. EmpRL utilizes the pre-trained T5 model as the generator and further fine-tunes it to initialize the policy. To align the empathy levels between generated and target responses within a given context, an empathy reward function containing three empathy communication mechanisms—emotional reaction, interpretation, and exploration—is constructed using pre-designed and pre-trained empathy identifiers. During reinforcement learning training, the proximal policy optimization algorithm is used to fine-tune the policy, enabling the generation of empathetic responses. Both automatic and human evaluations demonstrate that the proposed EmpRL framework significantly improves the quality of generated responses, enhances the similarity in empathy levels between generated and target responses, and produces empathetic responses covering both affective and cognitive aspects.

Named Entity Recognition (NER) in specialized domains for low-resource languages remains a significant challenge due to data scarcity and the complexity of domain-specific terminology. Existing cross-lingual approaches—spanning model-transfer and data-transfer paradigms—often suffer from semantic drift and inadequate domain adaptation. To address these limitations, we introduce BiLegalNERD, the first bilingual Chinese–Uyghur legal NER dataset, constructed via a semantics-aware label transfer strategy. We further propose CUTLM, a cross-lingual annotation method that combines dual translation with Levenshtein-based alignment to ensure high-fidelity preservation of entity boundaries across languages. In addition, we present BiLegalNER, a domain-adapted multilingual NER model incorporating vocabulary expansion and bilingual fine-tuning, significantly enhancing performance on Uyghur legal texts. Experiments demonstrate that BiLegalNER achieves state-of-the-art results, with F1-scores of 86.65% on automatically generated training data and 89.11% on fully human-annotated data—outperforming the strongest multilingual baseline by 4.18% and 4.64%, respectively. Moreover, CUTLM surpasses prior cross-lingual transfer methods by up to 9.89%, confirming its effectiveness in preserving entity integrity during label projection. These findings establish a new benchmark for Uyghur legal NER and provide a scalable framework for cross-lingual NER in low-resource settings.

Session-based recommendation is gaining increasing attention due to its practical value in predicting the intents of anonymous users based on limited behaviors. Emerging efforts incorporate various side information to alleviate inherent data scarcity issues in this task, leading to impressive performance improvements. The core of side information-driven session-based recommendation is the discovery and utilization of diverse data. In this survey, we provide a comprehensive review of this task from a data-centric perspective. Specifically, this survey commences with a clear formulation of the task. This is followed by a detailed exploration of various benchmarks rich in side information that are pivotal for advancing research in this field. Afterwards, we delve into how different types of side information enhance the task, underscoring data characteristics and utility. Moreover, we discuss the usage of various side information, including data encoding, data injection, and involved techniques. A systematic review of research progress is then presented, with the taxonomy by the types of side information. Finally, we summarize the current limitations and present the future prospects of this vibrant topic.

Temporal network embedding (TNE) has promoted the research of knowledge discovery and reasoning on networks. It aims to embed vertices of temporal networks into a low-dimensional vector space while preserving network structures and temporal properties. However, most existing methods have limitations in capturing dynamics over long distances, which makes it difficult to explore multihop topological associations among vertices. To tackle this challenge, we propose LongTNE, which learns the long-range dynamics of vertices to endow TNE with the ability to capture high-order proximity (HP) of networks. In LongTNE, we employ graph self-supervised learning (Graph SSL) to optimize the establishment probability of deep links in each network snapshot. We also present an accumulated forward update (AFU) module to fathom global temporal evolution among multiple network snapshots. The empirical results on six temporal networks demonstrate that, in addition to achieving state-of-the-art performance on network mining tasks, LongTNE can be handily extended to existing TNE methods.

Retrieving appropriate records from the external knowledge base to generate informative responses is the core capability of end-to-end task-oriented dialogue systems (EToDs). Most of the existing methods additionally train the retrieval model or use the memory network to retrieve the knowledge base, which decouples the knowledge retrieval task from the response generation task, making it difficult to jointly optimize and failing to capture the internal relationship between the two tasks. In this paper, we propose a simple and unified generative model for task-oriented dialogue systems, which recasts the EToDs task as a single sequence generation task and uses maximum likelihood training to train the two tasks in a unified manner. To prevent the generation of non-existent records, we design the prefix trie to constrain the model generation, which ensures consistency between the generated records and the existing records in the knowledge base. Experimental results on three public benchmark datasets demonstrate that our method achieves robust performance on generating system responses and outperforms the baseline systems. To facilitate future research in this area, the code is available at https://github.com/dzy1011/Uni-ToD.

In the biomedical field, information is widely distributed across numerous pieces of literature. Extracting events between entities from biomedical texts has garnered significant attention in recent years. However, previous research primarily focus on extracting flat biomedical events, with less attention given to nested biomedical events. Moreover, existing methods for extracting nested events often overlook the long-distance dependencies and global information between trigger words and arguments within events, and they lack sufficient interaction with event type information. To address these issues, we propose a semantic segmentation-based method for extracting nested biomedical events. We introduce U-Net to capture global information and interdependencies between event entities. Additionally, we map event types to natural language text and combine them with sentences for encoding to enhance interaction. We also employ two auxiliary tasks to improve the identification of trigger words and arguments. Finally, events are extracted by identifying the four vertices of the segmented region. Experimental results on two benchmark datasets show that our method excels in recognizing nested biomedical events and outperforms current state-of-the-art methods.

The task of document-level biomedical relation extraction involves identifying relational facts between entities across sentences, given specific entities. However, most current methods overlook the associations between entity pairs and generate fixed entity representations merely through mentions, leading to irrelevant mentions interfering with the determination of relational facts. Additionally, these methods fail to consider the global information and dependencies between relational entities. To address these issues, we propose a document-level relation extraction model based on relation-guided entity-level graphs. Our model aggregates all mentions of the same entity through a relation-guided attention mechanism to obtain flexible entity representations. Furthermore, by using U-Net to generate entity-level feature graphs, it facilitates global interactions and dependency capture between entity pairs. Experimental results on two benchmark datasets demonstrate the advantages of our approach in document-level biomedical relation extraction.

Large language models (LLMs) perform well in many NLP tasks but frequently generate inaccurate information in the biomedical domain, due to hallucination issues. Retrieval-Augmented Generation (RAG) has been introduced to address this issue by integrating external knowledge, enhancing the factual accuracy of outputs. However, naive RAG encounters challenges in effectively utilizing retrieved content, particularly in specialized domains like biomedicine. LLMs often struggle to integrate retrieved content as irrelevant information can interfere with the model’s judgment. Even if relevant documents are retrieved, the model may be unable to accurately comprehend and utilize the domain-specific features due to its inherent knowledge limitations. To overcome these limitations, we propose Document Embeddings Enhanced Biomedical RAG (DEEB-RAG), a framework that incorporates document embeddings along with the original retrieved text. DEEB-RAG uses MedCPT to generate document embeddings and these embeddings are then aligned with the LLM’s semantic space using a two-stage training process on a simple projector. Experimental results on biomedical QA datasets show that DEEB-RAG improves accuracy, with an average performance increase of 2.3% over naive RAG. This demonstrates DEEB-RAG’s ability to mitigate the challenges of utilizing complex biomedical information, thereby enhancing the reliability and effectiveness of LLMs in biomedical domain.

Biomedical document-level relation extraction is a crucial technology for mining the biomedical relationships necessary for clinical diagnosis, treatment, and medical discovery. Although existing intrasentential relation extraction methods have achieved significant results, the complexity and scattered nature of information in biomedical literature require relation extraction techniques to effectively handle cross-sentence information. For example, existing methods have not been able to explicitly model the phenomena of coreference and anaphor in documents, thus affecting the model’s understanding of complex semantics within the document. To address this issue, we propose a new document-level relation extraction model with coreference and anaphor graphs. By abstracting the document into an undirected graph that includes coreference and anaphor information, the framework effectively models the interactions between entities and leverages graph convolutional network in conjunction with pretrained language model to dynamically understand graph structures. Additionally, the shift from fine-grained entity-pair level to coarse-grained document-level training and inference significantly enhances the model’s efficiency while maintaining high extraction performance. Extensive experiments demonstrate that our model achieves a 5.3% increase in F1-score over baseline models on the BioRED dataset with higher efficiency, confirming its effectiveness in handling relation extraction tasks in complex biomedical literature.

Biomedical Named Entity Recognition (NER) serves as a crucial task in biomedical information extraction, aiming to identify and classify entities from unstructured biomedical texts. However, obtaining high-quality annotated biomedical texts is scarce due to their high privacy and the specialized expertise required. Therefore, more researchers are focusing on few-shot NER in the biomedical field. Recent methods mainly fall into three categories: 1) transferring knowledge from high-resource data and fine-tuning with low-resource data, 2) modifying data using various methods to generate new samples, and 3) decompose the NER task into two subtasks with multi-task learning. However, these methods either suffer from domain shift, generate low-quality synthetic data, or encounter error propagation issues. To address these limitations, we have investigated an few-shot NER method, which firstly employs specific prompt templates to guide Large Language Models in generating high-quality new samples. Then, we propose a novel more fine-grained grid-tagging strategy and single-end sensitive multi-task learning framework. This method sets multiple losses to enable the model to learn from multiple task objectives: what constitutes a fully correct, partially correct, and incorrect entity. Additionally, through the new grid-tagging strategy, the model can decode entities based on more tagging clues. To better align with real-world scenarios, we trained in a few-shot scenario and evaluated on the full test set. Extensive experimental results on the NCBI, BC5CDR, BioNLP11EPI, and BioNLP13GE datasets confirm that our method outperforms previous state-of-the-art methods in most scenarios.1

Automatically identifying human phenotype ontology (HPO) concepts from text is important for disease analysis. Existing ontology-driven methods for phenotype concept recognition mainly rely on concept names and synonym information from the ontology, without fully exploiting the rich ontology information. In this paper, we present an improved phenotype concept recognition method by incorporating rich HPO information. We first design prompts with HPO information and use a cutting-edge large language model GPT-4 to generate synonym augmentation for expanding distant supervised training data. We then propose an ontology vector-enhanced phenotype concept classification model to efficiently integrate the taxonomic hierarchical structure of HPO. Additionally, we employ noisy data augmentation to improve the model’s recognition ability in noisy texts and implement a negation detection function. Experimental results on three standard corpora and two typo corpora show our method compares favorably to previous methods and achieves a significant improvement in noisy texts. The source code and data are freely available at https://github.com/DUTIR-BioNLP/PhenoTagger-Updates.

Adverse drug reactions (ADRs) are harmful reactions that occur with qualified drugs under normal dosage. Due to its significant impact on patients’ health, timely detection for ADRs is of great significance. Current approaches typically introduce sentiment to enhance the detection of ADRs because sentiment information can reflect the subjective feelings of the patient about the drug. However, none of the existing works has considered both explicit and implicit sentiment knowledge. To this end, we propose a novel approach named DSN to adaptively fuse explicit and implicit sentiment knowledge for adverse drug reactions detection. Specifically, for explicit sentiment knowledge, we first propose a dimension extension approach to avoid the information loss problem, and then utilize an aware network based on attention mechanism to obtain a textual representation that incorporates explicit sentiment knowledge; As to implicit sentiment knowledge, we design a fusion network based on a gating mechanism to adaptively learn text and implicit sentiment knowledge in order to help and guide the deeper fusion of text and two kinds of sentiment knowledge. Extensive experiments conducted on TwiMed_Twitter dataset demonstrate the superiority of our proposed DSN over existing methods.

Style transfer has been proven to be effective in mitigating domain shift in clinical settings, enhancing the adaptability of pathology image models. However, existing methods assume that the texture of the entire image is domain-specific and irrelevant to class-specific representations. These methods enhance all regions with a single style, which can result in information loss. In this work, we propose CLAP (Cross-modality Local Augmentation for histoPathology), a data augmentation approach that enables cross-modality local manipulation for pathology images. Specifically, the combination of a text extractor network and a feature mapping network enables the integration of CLIP embeddings for style descriptions into editable latent spaces at a fine-grained level. This approach prevents the loss of regional information in whole-slide images, eliminates the need to painstakingly select directions in latent space, and enhances creativity style selection. Experimental results demonstrate that CLAP improves the regional accuracy of cross-modality editing and achieves state-of-the-art performance by enhancing generalization in histopathology classification tasks.

Polypharmacy is the combined use of multiple drugs for the treatment of diseases, which also often comes with a higher risk of side effects. In the medical industry, acquiring rich and comprehensive information about the side effects of multiple drug therapy becomes a crucial task. However, data collection for many side effects is often sparse, so the features of these data cannot be adequately learned, resulting in poor performance in side effects prediction. In this paper, we propose a framework based on knowledge graph embedding (KGE) models which improves KGE by using LTE operations and subsampling methods (called LTESampleKGE). LTESampleKGE consists of two main modules i.e., Entity embedding enhancement module and KGE subsampling module. The former applies linear transformation to entity representation instead of GCN structure to enhance entity embedding, while the latter utilizes subsampling methods for KGE negative sampling (NS) loss to pay more attention to sparse data. Thus, LTESampleKGE can effectively alleviate the problem of data sparsity in the polypharmacy side effects prediction task. Experimental evaluations indicate that our method demonstrates superior performance compared with baseline models. For example, LTESampleKGE outperforms MSTE by 1.20% in PR-AUC score on TWOSIDES dataset and by 0.46% in AP@n score on Drugbank dataset.

Generative conversational agents, driven by large language models, have gained widespread popularity. However, a significant drawback lies in their tendency to produce uncontrollable and unpredictable contents, thereby increasing the risk of generating malevolent responses that potentially exacerbate users’ mental health issues. Although existing research on malevolence detection in dialogues addressed the modeling of interaction patterns in dialogues, the implicitly expressed emotion and user-specific context are often neglected. Addressing this gap, we propose a hypergraph-enhanced context modeling approach for detecting malevolence in mental health counseling dialogues. Our approach harnesses the emotion reasoning capabilities of large language models to generate implicit emotional prompts. Employing hypergraph neural networks, our approach effectively integrates emotional context, user-specific context, and interactive context, fusing them into high-order semantic representations using hypergraph convolution. Experimental results on two benchmark datasets, MDRDC and Dialogue Safety, demonstrate the superiority of our model over state-of-the-art baseline models, particularly in complex contextual scenarios.

Large language models (LLMs) have been widely used in society due to their amazing emergent capabilities, but they also bring security issues. A large amount of content on the Internet may be generated by AI. Whether it is social forums or more professional academic fields, the abuse of AI has become a problem. Especially in some professional fields, Blindly trusting what the Internet says is dangerous. For this reason, for some fields that are risky and need to limit the use of AI, such as medicine, a more comprehensive benchmark is needed to test the ability of AI-generated text detection tasks. Considering the popularity of LLMs, the data distributions used in the training process of different LLMs may lead to the differences in generated data distributions, especially for some LLMs for non-native English speakers. To address this issue, this article introduces an AI-generated text detection dataset in the field of medical question answering. This dataset is generated by various models and prompting methods, and cross validation is performed on multiple types of data between texts generated by different methods to verify the effectiveness of AI-generated text detection and model classification tasks, and to study the generalization of the dataset in different tasks. We have published the dataset for future research on https://github.com/Hellpoop/MAT.

Document-level biomedical relation extraction (Bio-DocRE) involves automatically identifying relation facts between multiple entity pairs from a document. Logical reasoning has become a recent research focus since many relations of long-distance entity pairs in biomedical documents need to be predicted through reasoning from other existing relations. However, this task faces two crucial challenges: long dependency and evidence selection. Existing graph-based or transformer-based models independently model the internal structure of entities and ignore additional global interactions between relation triples. In addition, evidence graph reasoning plays an important role in boosting the logical reasoning between entities, but it is ignored by these methods. To tackle these issues, we propose a novel Bio-DocRE framework called EGALR based on evidence graph assisted logical reasoning. This allows EGALR to focus on important sentences while predicting implicit relations between entities with relation reasoning. Extensive experiments show that our EGALR model achieves state-of-the-art performance on two widely-used biomedical datasets, CDR and GDA.

Document-level biomedical relation extraction (Bio-DocRE) involves identifying relations between entities distributed across multiple sentences in biomedical literature. Most existing approaches construct a document graph structure to model entity-level relation maps and achieve local implicit reasoning at the entity-pair granularity. However, these methods have two main drawbacks. On the one hand, mention-pair granularity is crucial for modeling entity-level relation map and global reasoning but is often ignored in most approaches. On the other hand, these approaches primarily achieve local reasoning of cross-sentence relation instances with the interaction of sentence or document nodes in the graph. Such an approach has difficulty modeling global reasoning and cannot completely address the sub-optimal performance caused by the long dependency issue. To tackle these issues, we propose a novel framework based on hierarchical multi-granularity learning and reasoning (HMGLR) for Bio-DocRE. Experimental results on two widely used biomedical datasets demonstrate that our model achieves new state-of-the-art performance.

Malevolence detection in dialogues aims to identify harmful or inappropriate utterances, significantly impacting dialogue quality and user satisfaction. Although existing studies have shown promising performance by modeling interaction patterns from dialogue history, various malevolence-invoking factors, such as fine-grained emotions, evolving topics and user profiles, are often overlooked. To comprehensively consider these factors, we propose a hypergraph fusion model by employing multi-view LLM-driven prompts for malevolence detection in dialogues. Our model integrates emotion context, topic context, user profile context and interaction context, utilizing hypergraphs to establish high-order contextual relationships from multi views for deducing malevolence-invoking semantics. Experimental results on two benchmark datasets demonstrate that our model achieves the state-of-the-art performance.

Researchers have attempted to mitigate lexical bias in toxic language detection (TLD). However, existing methods fail to disentangle the “useful” and “misleading” impact of lexical bias on model decisions. Therefore, they do not effectively exploit the positive effects of the bias and lead to a degradation in the detection performance of the debiased model. In this paper, we propose a Counterfactual Causal Debiasing Framework (CCDF) to mitigate lexical bias in TLD. It preserves the “useful impact” of lexical bias and eliminates the “misleading impact”. Specifically, we first represent the total effect of the original sentence and biased tokens on decisions from a causal view. We then conduct counterfactual inference to exclude the direct causal effect of lexical bias from the total effect. Empirical evaluations demonstrate that the debiased TLD model incorporating CCDF achieves state-of-the-art performance in both accuracy and fairness compared to competitive baselines applied on several vanilla models. The generalization capability of our model outperforms current debiased models for out-of-distribution data.

Emotion recognition in conversation (ERC) is essential for dialogue systems to identify the emotions expressed by speakers. Although previous studies have made significant progress, accurate recognition and interpretation of similar fine-grained emotion properly accounting for individual variability remains a challenge. One particular under-explored area is the role of individual beliefs and desires in modelling emotion. Inspired by the Belief-Desire Theory of Emotion, we propose a novel method for conversational emotion recognition that incorporates both belief and desire to accurately identify emotions. We extract emotion-eliciting events from utterances and construct graphs that represent beliefs and desires in conversations. By applying message passing between nodes, our graph effectively models the utterance context, speaker’s global state, and the interaction between emotional beliefs, desires, and utterances. We evaluate our model’s performance by conducting extensive experiments on four popular ERC datasets and comparing it with multiple state-of-the-art models. The experimental results demonstrate the superiority of our proposed model and validate the effectiveness of each module in the model.

With the development of the Internet, social media has produced a large amount of user-generated data, which brings new challenges for humor computing. Traditional humor computing research mainly focuses on the content, while neglecting the information of interaction relationships in social media. In addition, both content and users are important in social media, while existing humor computing research mainly focuses on content rather than people. To address these problems, we model the information transfer and entity interactions in social media as a heterogeneous graph, and create the first dataset which introduces the social context information - HumorWB, which is collected from Chinese social media - Weibo. Two humor-related tasks are designed in the dataset. One is a content-oriented humor recognition task, and the other is a novel humor evaluation task. For the above tasks, we purpose a graph-based model called SCOG, which uses heterogeneous graph neural networks to optimize node representation for downstream tasks. Experimental results demonstrate the effectiveness of feature extraction and graph representation learning methods in the model, as well as the necessity of introducing social context information.

In the realm of artificial intelligence and linguistics, the automatic generation of humor, particularly puns, remains a complex task. This paper introduces an innovative approach that employs a Generative Adversarial Network (GAN) and semantic pruning techniques to generate humorous puns. We initiate our process by identifying potential pun candidates via semantic pruning. This is followed by the use of contrastive learning to decode the unique characteristics of puns, emphasizing both correct and incorrect interpretations. The learned features from contrastive learning are utilized within our GAN model to better capture the semantic nuances of puns. Specifically, the generator exploits the pruned semantic tree to generate pun texts, while the discriminator evaluates the generated puns, ensuring both linguistic correctness and humor. Evaluation results highlight our model’s capacity to produce semantically coherent and humorous puns, demonstrating an enhancement over prior methods and approach human-level performance. This work contributes significantly to the field of computational humor, advancing the capabilities of automatic pun generation.

Understanding a meme is a challenging task, due to the metaphorical information contained in the meme that requires intricate interpretation to grasp its intended meaning fully. In previous works, attempts have been made to facilitate computational understanding of memes through introducing human-annotated metaphors as extra input features into machine learning models. However, these approaches mainly focus on formulating linguistic representation of a metaphor (extracted from the texts appearing in memes), while ignoring the connection between the metaphor and corresponding visual features (e.g., objects in meme images). In this paper, we argue that a more comprehensive understanding of memes can only be achieved through a joint modelling of both visual and linguistic features of memes. To this end, we propose an approach to generate Multimodal Metaphorical feature for Meme Classification, named MMMC. MMMC derives visual characteristics from linguistic attributes of metaphorical concepts, which more effectively convey the underlying metaphorical concept, leveraging a text-conditioned generative adversarial network. The linguistic and visual features are then integrated into a set of multimodal metaphorical features for classification purpose. We perform extensive experiments on a benchmark metaphorical meme dataset, MET-Meme. Experimental results show that MMMC significantly outperforms existing baselines on the task of emotion classification and intention detection. Our code and dataset are available at https://github.com/liaolianfoka/MMMC.

By integrating various modules with the Visual Transformer (ViT), we facilitate a interpretation of image processing across each layer and attention head. This method allows us to explore the connections both within and across the layers, enabling a analysis of how images are processed at different layers. Conducting a analysis of the contributions from each layer and attention head, shedding light on the intricate interactions and functionalities within the model's layers. This in-depth exploration not only highlights the visual cues between layers but also examines their capacity to navigate the transition from abstract concepts to tangible objects. It unveils the model's mechanism to building an understanding of images, providing a strategy for adjusting attention heads between layers, thus enabling targeted pruning and enhancement of performance for specific tasks. Our research indicates that achieving a scalable understanding of transformer models is within reach, offering ways for the refinement and enhancement of such models.

Material property prediction, essential in AI-driven science, faces complexity due to the intricate influence of elemental combinations on superconducting characteristics. Acknowledging that the elemental composition and interactions within material molecules naturally align with graphical representation, we herein apply a graph neural network (GNN)-based methodology. By integrating molecular features alongside chemical descriptors, our approach comprehensively captures elemental synergy. Upon evaluation with the SuperCon dataset, our model showcases a significant leap in prediction accuracy, demonstrating its proficiency in estimating critical temperatures of superconductors. This breakthrough notably propels guided material synthesis and deepens the comprehension of superconductivity, thereby accelerating progress in the development of high-temperature superconductors.

Harmful memes have proliferated on the Chinese Internet, while research on detecting Chinese harmful memes significantly lags behind due to the absence of reliable datasets and effective detectors.To this end, we present the comprehensive detection of Chinese harmful memes.We introduce ToxiCN MM, the first Chinese harmful meme dataset, which consists of 12,000 samples with fine-grained annotations for meme types. Additionally, we propose a baseline detector, Multimodal Knowledge Enhancement (MKE), designed to incorporate contextual information from meme content, thereby enhancing the model's understanding of Chinese memes.In the evaluation phase, we conduct extensive quantitative experiments and qualitative analyses on multiple baselines, including LLMs and our MKE. Experimental results indicate that detecting Chinese harmful memes is challenging for existing models, while demonstrating the effectiveness of MKE.

Session-based recommendation aims to predict intents of anonymous users based on their limited behaviors. Modeling user behaviors involves two distinct rationales: co-occurrence patterns reflected by item IDs, and fine-grained preferences represented by item modalities (e.g., text and images). However, existing methods typically entangle these causes, leading to their failure in achieving accurate and explainable recommendations. To this end, we propose a novel framework DIMO to disentangle the effects of ID and modality in the task. DIMO aims to disentangle these causes at both item and session levels. At the item level, we introduce a co-occurrence representation schema to explicitly incorporate co-occurrence patterns into ID representations. Simultaneously, DIMO aligns different modalities into a unified semantic space to represent them uniformly. At the session level, we present a multi-view self-supervised disentanglement, including proxy mechanism and counterfactual inference, to disentangle ID and modality effects without supervised signals. Leveraging these disentangled causes, DIMO provides recommendations via causal inference and further creates two templates for generating explanations. Extensive experiments on multiple real-world datasets demonstrate the consistent superiority of DIMO over existing methods. Further analysis also confirms DIMO's effectiveness in generating explanations.

Sequential recommendation is dedicated to offering items of interest for users based on their history behaviors. The attribute-opinion pairs, expressed by users in their reviews for items, provide the potentials to capture user preferences and item characteristics at a fine-grained level. To this end, we propose a novel framework FineRec that explores the attribute-opinion pairs of reviews to finely handle sequential recommendation. Specifically, we utilize a large language model to extract attribute-opinion pairs from reviews. For each attribute, a unique attribute-specific user-opinion-item graph is created, where corresponding opinions serve as the edges linking heterogeneous user and item nodes. Afterwards, we devise a diversity-aware convolution operation to aggregate information within the graphs, enabling attribute-specific user and item representation learning. Ultimately, we present an interaction-driven fusion mechanism to integrate attribute-specific user/item representations across all attributes for generating recommendations. Extensive experiments conducted on several real-world datasets demonstrate the superiority of our FineRec over existing state-ofthe-art methods. Further analysis also verifies the effectiveness of our fine-grained manner in handling the task.

MOTIVATION: Significant progress has been achieved in biomedical text mining using deep learning methods, which rely heavily on large amounts of high-quality data annotated by human experts. However, the reality is that obtaining high-quality annotated data is extremely challenging due to data scarcity (e.g. rare or new diseases), data privacy and security concerns, and the high cost of data annotation. Additionally, nearly all researches focus on predicting labels without providing corresponding explanations. Therefore, in this paper, we investigate a more realistic scenario, biomedical few-shot learning, and explore the impact of interpretability on biomedical few-shot learning. RESULTS: We present LetEx-Learning to explain-a novel multi-task generative approach that leverages reasoning explanations from large language models (LLMs) to enhance the inductive reasoning ability of few-shot learning. Our approach includes (1) collecting high-quality explanations by devising a suite of complete workflow based on LLMs through CoT prompting and self-training strategies, (2) converting various biomedical NLP tasks into a text-to-text generation task in a unified manner, where collected explanations serve as additional supervision between text-label pairs by multi-task training. Experiments are conducted on three few-shot settings across six biomedical benchmark datasets. The results show that learning to explain improves the performances of diverse biomedical NLP tasks in low-resource scenario, outperforming strong baseline models significantly by up to 6.41%. Notably, the proposed method makes the 220M LetEx perform superior reasoning explanation ability against LLMs. AVAILABILITY AND IMPLEMENTATION: Our source code and data are available at https://github.com/cpmss521/LetEx.

MOTIVATION: Biomedical relation extraction at the document level (Bio-DocRE) involves extracting relation instances from biomedical texts that span multiple sentences, often containing various entity concepts such as genes, diseases, chemicals, variants, etc. Currently, this task is usually implemented based on graphs or transformers. However, most work directly models entity features to relation prediction, ignoring the effectiveness of entity pair information as an intermediate state for relation prediction. In this article, we decouple this task into a three-stage process to capture sufficient information for improving relation prediction. RESULTS: We propose an innovative framework HTGRS for Bio-DocRE, which constructs a hierarchical tree graph (HTG) to integrate key information sources in the document, achieving relation reasoning based on entity. In addition, inspired by the idea of semantic segmentation, we conceptualize the task as a table-filling problem and develop a relation segmentation (RS) module to enhance relation reasoning based on the entity pair. Extensive experiments on three datasets show that the proposed framework outperforms the state-of-the-art methods and achieves superior performance. AVAILABILITY AND IMPLEMENTATION: Our source code is available at https://github.com/passengeryjy/HTGRS.

OBJECTIVE: Most existing fine-tuned biomedical large language models (LLMs) focus on enhancing performance in monolingual biomedical question answering and conversation tasks. To investigate the effectiveness of the fine-tuned LLMs on diverse biomedical natural language processing (NLP) tasks in different languages, we present Taiyi, a bilingual fine-tuned LLM for diverse biomedical NLP tasks. MATERIALS AND METHODS: We first curated a comprehensive collection of 140 existing biomedical text mining datasets (102 English and 38 Chinese datasets) across over 10 task types. Subsequently, these corpora were converted to the instruction data used to fine-tune the general LLM. During the supervised fine-tuning phase, a 2-stage strategy is proposed to optimize the model performance across various tasks. RESULTS: Experimental results on 13 test sets, which include named entity recognition, relation extraction, text classification, and question answering tasks, demonstrate that Taiyi achieves superior performance compared to general LLMs. The case study involving additional biomedical NLP tasks further shows Taiyi's considerable potential for bilingual biomedical multitasking. CONCLUSION: Leveraging rich high-quality biomedical corpora and developing effective fine-tuning strategies can significantly improve the performance of LLMs within the biomedical domain. Taiyi shows the bilingual multitasking capability through supervised fine-tuning. However, those tasks such as information extraction that are not generation tasks in nature remain challenging for LLM-based generative approaches, and they still underperform the conventional discriminative approaches using smaller language models.

Dialogue summarization is a task that aims to condense dialogues while retain the salient information. However, due to different domains involved in the dialogue, the corresponding format of reference summary varies from each other, e.g., QA pairs for customer service and SOAP notes in medical field. To address the common challenges encountered in various fields and alleviate the differences due to the format in the generation process, we introduce a novel unified topic-guided dialogue summarization framework, by which we can first capture the topic structure of the conversation and leverage it to guide the generation of the summary. This framework is the first to model fine-grained topic structure of the dialogue and pose its identification as a Seq2Seq task, as well as introduce the topic-guided segment-wise attention to produce the final summary in segments following the specific format in each domain. Such a concise but effective method avoids the trouble of customizing decoding schemes while retains the topic structure of a dialogue in its summary as much as possible. Comprehensive experiments were conducted on four benchmark datasets in different domains and the results show the better performance and generalization of our method compared with the baselines.

In bioinformatics, protein function prediction stands as a fundamental area of research and plays a crucial role in addressing various biological challenges, such as the identification of potential targets for drug discovery and the elucidation of disease mechanisms. However, known functional annotation databases usually provide positive experimental annotations that proteins carry out a given function, and rarely record negative experimental annotations that proteins do not carry out a given function. Therefore, existing computational methods based on deep learning models focus on these positive annotations for prediction and ignore these scarce but informative negative annotations, leading to an underestimation of precision. To address this issue, we introduce a deep learning method that utilizes a heterogeneous graph attention technique. The method first constructs a heterogeneous graph that covers the protein-protein interaction network, ontology structure, and positive and negative annotation information. Then, it learns embedding representations of proteins and ontology terms by using the heterogeneous graph attention technique. Finally, it leverages these learned representations to reconstruct the positive protein-term associations and score unobserved functional annotations. It can enhance the predictive performance by incorporating these known limited negative annotations into the constructed heterogeneous graph. Experimental results on three species (i.e., Human, Mouse, and Arabidopsis) demonstrate that our method can achieve better performance in predicting new protein annotations than state-of-the-art methods.

Session-based recommendation is devoted to characterizing preferences of anonymous users based on short sessions. Existing methods mostly focus on mining limited item co-occurrence patterns exposed by item ID within sessions, while ignoring what attracts users to engage with certain items is rich multi-modal information displayed on pages. Generally, the multi-modal information can be classified into two categories: descriptive information (e.g., item images and description text) and numerical information (e.g., price). In this paper, we aim to improve session-based recommendation by modeling the above multi-modal information holistically. There are mainly three issues to reveal user intent from multi-modal information: (1) How to extract relevant semantics from heterogeneous descriptive information with different noise? (2) How to fuse these heterogeneous descriptive information to comprehensively infer user interests? (3) How to handle probabilistic influence of numerical information on user behaviors? To solve above issues, we propose a novel multi-modal session-based recommendation (MMSBR) that models both descriptive and numerical information under a unified framework. Specifically, a pseudo-modality contrastive learning is devised to enhance the representation learning of descriptive information. Afterwards, a hierarchical pivot transformer is presented to fuse heterogeneous descriptive information. Moreover, we represent numerical information with Gaussian distribution and design a Wasserstein self-attention to handle the probabilistic influence mode. Extensive experiments on three real-world datasets demonstrate the effectiveness of the proposed MMSBR. Further analysis also proves that our MMSBR can alleviate the cold-start problem in SBR effectively.

Emotion recognition in conversations (ERC), the task of recognizing the emotion of each utterance in a conversation, is crucial for building empathetic machines. Existing studies focus mainly on capturing context- and speaker-sensitive dependencies on the textual modality but ignore the significance of multimodal information. Different from emotion recognition in textual conversations, capturing intra- and inter-modal interactions between utterances, learning weights between different modalities, and enhancing modal representations play important roles in multimodal ERC. In this paper, we propose a transformer-based model with self-distillation (SDT)11The code is available athttps://github.com/butterfliesss/SDT.for the task. The transformer-based model captures intra- and inter-modal interactions by utilizing intra- and inter-modal transformers, and learns weights between modalities dynamically by designing a hierarchical gated fusion strategy. Furthermore, to learn more expressive modal representations, we treat soft labels of the proposed model as extra training supervision. Specifically, we introduce self-distillation to transfer knowledge of hard and soft labels from the proposed model to each modality. Experiments on IEMOCAP and MELD datasets demonstrate that SDT outperforms previous state-of-the-art baselines.

Session-based recommendation intends to predict next purchased items based on anonymous behavior sequences. Numerous economic studies have revealed that item price is a key factor influencing user purchase decisions. Unfortunately, existing methods for session-based recommendation only aim at capturing user interest preference, while ignoring user price preference. Actually, there are primarily two challenges preventing us from accessing price preference. First, the price preference is highly associated to various item features (i.e., category and brand), which asks us to mine price preference from heterogeneous information. Second, price preference and interest preference are interdependent and collectively determine user choice, necessitating that we jointly consider both price and interest preference for intent modeling. To handle above challenges, we propose a novel approach Bi-Preference Learning Heterogeneous Hypergraph Networks (BiPNet) for session-based recommendation. Specifically, the customized heterogeneous hypergraph networks with a triple-level convolution are devised to capture user price and interest preference from heterogeneous features of items. Besides, we develop a Bi-Preference Learning schema to explore mutual relations between price and interest preference and collectively learn these two preferences under the multi-task learning architecture. Extensive experiments on multiple public datasets confirm the superiority of BiPNet over competitive baselines. Additional research also supports the notion that the price is crucial for the task.

The samples presented by this paper may be considered offensive or vulgar.The widespread dissemination of toxic online posts is increasingly damaging to society.However, research on detecting toxic language in Chinese has lagged significantly.Existing datasets lack fine-grained annotation of toxic types and expressions, and ignore the samples with indirect toxicity.In addition, it is crucial to introduce lexical knowledge to detect the toxicity of posts, which has been a challenge for researchers.In this paper, we facilitate the finegrained detection of Chinese toxic language.First, we build MONITOR TOXIC FRAME, a hierarchical taxonomy to analyze toxic types and expressions.Then, a fine-grained dataset TOXICN is presented, including both direct and indirect toxic samples.We also build an insult lexicon containing implicit profanity and propose Toxic Knowledge Enhancement (TKE) as a benchmark, incorporating the lexical feature to detect toxic language.In the experimental stage, we demonstrate the effectiveness of TKE.After that, a systematic quantitative and qualitative analysis of the findings is given. 1

Sarcasm, as a form of irony conveying mockery and contempt, has been widespread in social media such as Twitter and Weibo, where the sarcastic text is commonly characterized as an incongruity between the surface positive and negative situation.Naturally, it has an urgent demand to automatically identify sarcasm from social media, so as to illustrate people's real views toward specific targets.In this paper, we develop a novel sarcasm detection method, namely Sarcasm Detector with Augmentation of Potential Result and Reaction (SD-APRR).Inspired by the direct access view, we treat each sarcastic text as an incomplete version without latent content associated with implied negative situations, including the result and human reaction caused by its observable content.To fill the latent content, we estimate the potential result and human reaction for each given training sample by [xEffect] and [xReact] relations inferred by the pre-trained commonsense reasoning tool COMET, and integrate the sample with them as an augmented one.We can then employ those augmented samples to train the sarcasm detector, whose encoder is a graph neural network with a denoising module.We conduct extensive empirical experiments to evaluate the effectiveness of SD-APRR.The results demonstrate that SD-APRR can outperform strong baselines on benchmark datasets.

The Relational Triple Extraction (RTE) task is a fundamental and essential information extraction task.Recently, the table-filling RTE methods have received lots of attention.Despite their success, they suffer from some inherent problems such as underutilizing regional information of triple.In this work, we treat the RTE task based on table-filling method as an Object Detection task and propose a one-stage Object Detection framework for Relational Triple Extraction (OD-RTE).In this framework, the vertices-based bounding box detection, coupled with auxiliary global relational triple region detection, ensuring that regional information of triple could be fully utilized.Besides, our proposed decoding scheme could extract all types of triples.In addition, the negative sampling strategy of relations in the training stage improves the training efficiency while alleviating the imbalance of positive and negative relations.The experimental results show that 1) OD-RTE achieves the state-of-the-art performance on two widely used datasets (i.e., NYT and WebNLG).2) Compared with the best performing table-filling method, OD-RTE achieves faster training and inference speed with lower GPU memory usage.To facilitate future research in this area, the codes are publicly available at https://github.com/NingJinzhong/ODRTE.

Electronic health records (EHRs) comprehensively describe the health status of many patients during their visits. Combining the records in EHRs with the patient’s current medical treatment status can generate personalized medication combinations. However, the increasing number of drugs poses significant challenges for clinical experts to recommend combination drugs. In recent years, deep learning models have been widely studied and applied in drug recommendation task. Currently, the existing models either lack sufficient mining of patients’ health data or ignore the modelling of patients’ longitudinal medical information. Therefore, we propose a multi-visit interactive recalibration network (MIRNet) for drug recommendation with a triple graph encoder. Specifically, we design a medical recalibration module to capture the feature representations in patient diagnosis and procedure information through cascaded convolutions. To achieve friendly interaction of medical codes between relatively necessary historical visits and the current visit in the drug recommendation process, we propose a multi-visit filter module. Furthermore, we design a triple graph encoder to fuse molecular, EHR, and Drug-Drug Interation (DDI) graphs, which aims to extract implicit drug feature representations from different medical knowledge. We perform experiments on the real-world MIMIC-III dataset, and the experimental results reveal that the model MIRNet outperforms other competitive baselines regarding major indicators.

Automatic extraction of biomedical entities and their relations plays a significant role in biomedical curation tasks. Currently, the table-filling methods have received lots of attention in the general domain. However, the presence of complex lengthy sentences and overlapping relations in biomedical texts makes automatic extraction a challenging task. To address this challenge, we propose a joint extraction table-filling method based on the vertices of the triple region. We extract triples by using multi-label classification to mark the boundaries of the triples, fully utilizing the boundary information of the entities. To incorporate the information of the distance between entity pairs, distance embedding is introduced and dilated convolutions are utilized to capture multi-scale contextual information. We evaluated our model on the CHEMPROT and DDIExtraction2013 datasets. The experimental results demonstrate that our model achieves the state-of-the-art performance on both datasets.

In the realm of medicine and healthcare, adverse drug reactions (ADRs) are a significant contributor to mortality and morbidity. Consequently, it is of paramount importance to closely observe the adverse effects of marketed drugs to minimize associated risks. While current methods for Adverse Drug Reaction (ADR) detection have demonstrated notable efficacy, a significant number of researchers have failed to acknowledge the integral role that sentiment information plays in this process. In this paper, we propose Sentiment Enhanced Dual Graph Convolutional Networks (SEDGCN), a novel method for ADRs detection by incorporating sentiment information. In particular, we first introduce the concept of prompt learning and reformulate the ADR detection task as an aspect-level sentiment analysis task. Subsequently, we construct a sentimentenhanced dependency matrix for each sentence to capture the sentiment knowledge and syntactic information of the sentence. The matrix is then input into the graph convolutional networks to obtain a graph representation of the sentence. Finally, to capture global information, we construct a heterogeneous graph based on all words and sentences and fuse this heterogeneous graph with the sentence-level graph representation for ADR detection. Extensive experimentation on two publicly available datasets, namely TwiMed and Twitter, yielded F1 scores of 78.24% and 75.43%, respectively. These results underscore the efficacy of our proposed model.

Automatic extraction of entities and their relations from unstructured literature to form structured triples is essential for biomedical knowledge construction. Although most existing joint methods have effectively addressed some challenging problems in the biomedical corpora, i.e., the prevalent overlapping issue, they still suffer from a lack of consideration for the intrinsic correlations between entities and relations, as well as low computational efficiency. In this paper, we present a joint entity and relation extraction model with unified interaction maps. Specifically, we concatenate all relations in the natural language form with the input text to integrate the semantic information of relations through a deep Transformer-based encoder. In addition, we apply unified interaction maps to capture the correlations, which can naturally handle the overlapping issue. Extensive experiments on the CHEMPROT and DDIExtraction2013 datasets demonstrate the effectiveness of our model, achieving the state-of-the-art performance with higher efficiency.

Automated International Classification of Diseases (ICD) coding tasks are designed to assign diagnosis and procedure codes to patients’ electronic medical records (EMRs). Recent works have applied deep neural network models and related techniques for code assignment to clinical notes. However, most existing methods have overlooked the advantageous complementary information presented in the tabular data from EMRs and the Wikipedia knowledge. Therefore, we propose a Multimodal Knowledge Fusion Network (MKFN) to effectively integrate clinical notes, tabular data, and Wikipedia knowledge, enhancing the model’s predictive capabilities. We incorporate structured tabular data and clinical notes into an initial multimodal representation using label attention and self-attention mechanisms. We propose a knowledge fusion network to leverage tabular data and Wikipedia knowledge for accurate predictions when code descriptions are absent in clinical notes. Experiments on the MIMIC dataset show that our proposed model achieves competitive results among existing ICD coding methods.

The liver is an internal organ that is most commonly affected by cancer metastasis. Hepatocellular carcinoma (HCC) is a major cause of cancer-related deaths worldwide. Identifying the histological category of this disease is a crucial step in determining the most appropriate treatment plan. Currently, the classification of HCC primarily relies on the observation of pathologists through a visual microscope, which is undoubtedly subjective. To address this issue, the present study conducted pathological examination of liver tumors based on few-shot learning. In order to achieve automatic identification of the three categories in HCC tissue, we propose a few-shot learning based HCC histological image classification method that combines contrastive learning (CL) using different backbone architectures with several well-known machine learning methods. We evaluate 12 backbone-classifier combinations, and the best performance is achieved by combining CL using DeiT-TinyR as the backbone architecture with support vector machine (SVM), with an accuracy of 74.42% and an F1 score of 74.58% in the 3-way-10-shot setting. Experimental results demonstrate the effectiveness of this method in NB histological image classification.

Biomedical Named Entity Recognition (BioNER) is a crucial task in extracting entities from biomedical literature. It plays a key role as the initial step in various biomedical information extraction tasks. Pre-training models have gained popularity in BioNER. However, these models require a large number of parameters, which poses a barrier for many researchers due to the hardware resource requirements. To address this challenge, we propose REIN-NER, a novel reinforcement learning-based approach for BioNER. REIN-NER employs a two-stage training mechanism: Firstly, the basic named recognition model (BaseNER) is pre-trained to capture the mapping knowledge from input words to the corresponding labels. Secondly, after initializing the agent model (AgentNER) with the BaseNER, the reinforcement training is carried out to further obtain the dependency knowledge between the output labels. Both BaseNER and AgentNER are built with lightweight Bi-LSTM networks, which significantly reduce parameter sizes compared to pre-training models. Remarkably, REIN-NER achieves superior performance with F-scores of 84.04% and 93.34% on the CDR-disease and CDR-chemical corpora, respectively, outperforming BERT-based models. Our work presents a pioneering exploration of reinforcement learning in biomedical named entity recognition and demonstrates its effectiveness through state-of-the-art results.

Humor recognition is one of the hot topics in the field of natural language processing in recent years. Most existing methods focus on whether something is humorous, while there is less research on humor level recognition. This article proposes a method based on prompt learning and contrastive learning for humor level recognition. Firstly, the context with prompt learning templates and the context without prompt learning templates are input into a pre-trained language model separately. Then, during the fine-tuning stage, a prompt learning strategy based on P-tuning is used to learn semantic information related to humor levels in the text. Subsequently, a contrastive learning loss function is introduced to increase the distance between the vector representations of different classes and reduce the classification difficulty of strong humor and weak humor samples. Finally, a multi-task learning strategy is used to simultaneously perform prompt learning, contrastive learning, and sentence binary classification tasks. Experiments on the Reddit public humor dataset show that the model’s accuracy on the corpus improves by 1.1% compared to the previous best results. Experimental results indicate that the model proposed in this article can effectively recognize humor levels.

The task of extracting overlapping and nested events has received significant attention in recent times, as prior research has primarily focused on extracting flat events, overlooking the intricacies of overlapping and nested occurrences. In this work, we present a new approach to Event Extraction (EE) by reformulating it as an object detection task on a table of token pairs. Our proposed one-stage event extractor, called ODEE, can handle overlapping and nested events. The model is designed with a vertex-based tagging scheme and two auxiliary tasks of predicting the spans and types of event trigger words and argument entities, leveraging the full span information of event elements. Furthermore, in the training stage, we introduce a negative sampling method for table cells to address the imbalance problem of positive and negative table cell tags, meanwhile improving computational efficiency. Empirical evaluations demonstrate that ODEE achieves the state-of-the-art performance on three benchmarks for overlapping and nested EE (i.e., FewFC, Genia11, and Genia13). Furthermore, ODEE outperforms current state-of-the-art methods in terms of both number of parameters and inference speed, indicating its high computational efficiency. To facilitate future research in this area, the codes are publicly available at https://github.com/NingJinzhong/ODEE.

Pre-trained language models have become the prevailing approach for handling natural language processing tasks in recent years. Given the similarities in sequential features between symbolic music and natural language text, it is fairly logical to adopt pre-training methods to symbolic music data. However, the disparity between music and natural language text makes it difficult to comprehensively model the unique features of music through traditional text-based pre-training strategies alone. To address this challenge, in this paper, we design the quad-attribute masking (QM) strategy and propose the key prediction (KP) task to improve the extraction of generic knowledge from symbolic music. We evaluate the impact of various pre-training strategies on several public symbolic music datasets, and the results of our experiments reveal that the proposed multi-task pre-training model can effectively capture music domain knowledge from symbolic music data and significantly improve performance on downstream tasks.

Image-grounded dialogue systems benefit greatly from integrating visual information, resulting in high-quality response generation. However, current models struggle to effectively utilize such information in zero-resource scenarios, mainly due to the disparity between image and text modalities. To overcome this challenge, we propose an innovative multimodal framework, called ZRIGF, which assimilates image-grounded information for dialogue generation in zero-resource situations. ZRIGF implements a two-stage learning strategy, comprising contrastive pre-training and generative pre-training. Contrastive pre-training includes a text-image matching module that maps images and texts into a unified encoded vector space, along with a text-assisted masked image modeling module that preserves pre-training visual features and fosters further multimodal feature alignment. Generative pre-training employs a multimodal fusion module and an information transfer module to produce insightful responses based on harmonized multimodal representations. Comprehensive experiments conducted on both text-based and image-grounded dialogue datasets demonstrate ZRIGF's efficacy in generating contextually pertinent and informative responses. Furthermore, we adopt a fully zero-resource scenario in the image-grounded dialogue dataset to demonstrate our framework's robust generalization capabilities in novel domains.

MOTIVATION: Few-shot learning that can effectively perform named entity recognition in low-resource scenarios has raised growing attention, but it has not been widely studied yet in the biomedical field. In contrast to high-resource domains, biomedical named entity recognition (BioNER) often encounters limited human-labeled data in real-world scenarios, leading to poor generalization performance when training only a few labeled instances. Recent approaches either leverage cross-domain high-resource data or fine-tune the pre-trained masked language model using limited labeled samples to generate new synthetic data, which is easily stuck in domain shift problems or yields low-quality synthetic data. Therefore, in this article, we study a more realistic scenario, i.e. few-shot learning for BioNER. RESULTS: Leveraging the domain knowledge graph, we propose knowledge-guided instance generation for few-shot BioNER, which generates diverse and novel entities based on similar semantic relations of neighbor nodes. In addition, by introducing question prompt, we cast BioNER as question-answering task and propose prompt contrastive learning to improve the robustness of the model by measuring the mutual information between query-answer pairs. Extensive experiments conducted on various few-shot settings show that the proposed framework achieves superior performance. Particularly, in a low-resource scenario with only 20 samples, our approach substantially outperforms recent state-of-the-art models on four benchmark datasets, achieving an average improvement of up to 7.1% F1. AVAILABILITY AND IMPLEMENTATION: Our source code and data are available at https://github.com/cpmss521/KGPC.

Chinese poetry generation has been a challenging part of natural language processing due to the unique literariness and aesthetics of poetry. In most cases, the content of poetry is topic related. In other words, specific thoughts or emotions are usually expressed regarding given topics. However, topic information is rarely taken into consideration in current studies about poetry generation models. In this article, we propose a topic-enhanced Chinese poetry generation model called TPoet in which the topic model is integrated into the Transformer-based auto-regressive text generation model. By feeding topic information to the input layer and heterogeneous attention mechanism, TPoet can implicitly learn the latent information of topic distribution. In addition, by setting multiple identifiers such as segment, rhyme, and tone, the model can explicitly learn the constraints of generated poems. Extensive experimental results show that the quality of TPoet-generated poems outperforms the current advanced models or systems, and the topic consistency and diversity in generated poems have been significantly improved as well.

The fast spread of hate speech on social media impacts the Internet environment and our society by increasing prejudice and hurting people. Detecting hate speech has aroused broad attention in the field of natural language processing. Although hate speech detection has been addressed in recent work, this task still faces two inherent unsolved challenges. The first challenge lies in the complex semantic information conveyed in hate speech, particularly the interference of insulting words in hate speech detection. The second challenge is the imbalanced distribution of hate speech and non-hate speech, which may significantly deteriorate the performance of models. To tackle these challenges, we propose a novel dual contrastive learning (DCL) framework for hate speech detection. Our framework jointly optimizes the self-supervised and the supervised contrastive learning loss for capturing span-level information beyond the token-level emotional semantics used in existing models, particularly detecting speech containing abusive and insulting words. Moreover, we integrate the focal loss into the dual contrastive learning framework to alleviate the problem of data imbalance. We conduct experiments on two publicly available English datasets, and experimental results show that the proposed model outperforms the state-of-the-art models and precisely detects hate speeches.

Rapid and effective utilization of biomedical literature is paramount to combat diseases like COVID19. Biomedical named entity recognition (BioNER) is a fundamental task in text mining that can help physicians accelerate knowledge discovery to curb the spread of the COVID-19 epidemic. Recent approaches have shown that casting entity extraction as the machine reading comprehension task can significantly improve model performance. However, two major drawbacks impede higher success in identifying entities (1) ignoring the use of domain knowledge to capture the context beyond sentences and (2) lacking the ability to deeper understand the intent of questions. In this paper, to remedy this, we introduce and explore external domain knowledge which cannot be implicitly learned in text sequence. Previous works have focused more on text sequence and explored little of the domain knowledge. To better incorporate domain knowledge, a multi-way matching reader mechanism is devised to model representations of interaction between sequence, question and knowledge retrieved from Unified Medical Language System (UMLS). Benefiting from these, our model can better understand the intent of questions in complex contexts. Experimental results indicate that incorporating domain knowledge can help to obtain competitive results across 10 BioNER datasets, achieving absolute improvement of up to 2.02% in the f1 score.

The automatic extraction of the chemical-disease relation (CDR) from the text becomes critical because it takes a lot of time and effort to extract valuable CDR manually. Studies have shown that prior knowledge from the biomedical knowledge base is important for relation extraction. The method of combining deep learning models with prior knowledge is worthy of our study. In this paper, we propose a new model called Knowledge Guided Attention and Graph Convolutional Networks (KGAGN) for CDR extraction. First, to make full advantage of domain knowledge, we train entity embedding as a feature representation of input sequence, and relation embedding to capture weighted contextual information further through the attention mechanism. Then, to make full advantage of syntactic dependency information in cross-sentence CDR extraction, we construct document-level syntactic dependency graphs and encode them using a graph convolution network (GCN). Finally, the chemical-induced disease (CID) relation is extracted by using weighted context features and long-range dependency features both of which contain additional knowledge information We evaluated our model on the CDR dataset published by the BioCreative-V community and achieves an F1-score of 73.3%, surpassing other state-of-the-art methods. the code implemented by PyTorch 1.7.0 deep learning library can be downloaded from Github: https://github.com/sunyi123/cdr.

Ontologies are widely utilized in the biological domain for data annotation, integration, and analysis. Some representation learning methods have been proposed to learn the representation of entities to assist intelligent applications, such as knowledge discovery. However, most of them neglect the class information of entities in the ontology. In this article, we propose a unified framework, named ERCI, which jointly optimizes the knowledge graph embedding model and self-supervised learning. In this way, we can generate embeddings of bio-entities by fusing the class information. Moreover, ERCI is a pluggable framework that can be easily incorporated with any knowledge graph embedding model. We validate ERCI in two different ways. In the first way, we utilize the protein embeddings learned by the ERCI to predict protein-protein interactions on two different datasets. In the second way, we leverage the gene and disease embeddings generated by the ERCI to predict gene-disease associations. In addition, we create three datasets to simulate the long-tail scenario and evaluate ERCI on these. Experimental results show that ERCI has superior performance on all metrics compared with the state-of-the-art methods.

Proteins are the main undertakers of life activities, and accurately predicting their biological functions can help human better understand life mechanism and promote the development of themselves. With the rapid development of high-throughput technologies, an abundance of proteins are discovered. However, the gap between proteins and function annotations is still huge. To accelerate the process of protein function prediction, some computational methods taking advantage of multiple data have been proposed. Among these methods, the deep-learning-based methods are currently the most popular for their capability of learning information automatically from raw data. However, due to the diversity and scale difference between data, it is challenging for existing deep learning methods to capture related information from different data effectively. In this paper, we introduce a deep learning method that can adaptively learn information from protein sequences and biomedical literature, namely DeepAF. DeepAF first extracts the two kinds of information by using different extractors, which are built based on pre-trained language models and can capture rudimentary biological knowledge. Then, to integrate those information, it performs an adaptive fusion layer based on a Cross-attention mechanism that considers the knowledge of mutual interactions between two information. Finally, based on the mixed information, DeepAF utilizes logistic regression to obtain prediction scores. The experimental results on the datasets of two species (i.e., Human and Yeast) show that DeepAF outperforms other state-of-the-art approaches.

Automatic extraction of drug-drug interactions and their adverse effects can promote the research of pharmacovigilance and thus attracts attention from both academia and industry. Recent efforts focus on span-based approaches and show more promising results. However, span-based methods enumerate all possible candidate entity spans while ignoring boundary information of spans. Meanwhile, lacking sufficient interactions in intra-span and inter-span further hinders the performance of the nested entity and overlapping relation extraction. To this end, we propose a syntactic type-aware graph attention network for drug-drug interactions and their adverse effects extraction. Specifically, a boundary heuristic module is designed firstly to generate the boundary of linguistically legitimate entity spans. And then, different from the general syntactic graph (i.e., only considering dependency edges), we construct a syntactic type-aware graph attention network (STG) to capture interactions in intra-span and inter-span by considering syntactic edges and types simultaneously. Results1achieved on two biomedical benchmark datasets, including drug-drug interaction (DDI) and adverse drug effect (ADE), indicate that our model obtains significantly more performance than the state-of-the-art methods, achieving improvements in the relation F1 score of 1.63% on ADE and 2.07% on DDI dataset, respectively.

As an important task of natural language processing, medication recommendation aims to recommend medication combinations according to the electronic health record, which can also be regarded as a multi-label classification task. But patients often have multiple diseases simultaneously, and the model must consider drug-drug interactions (DDI) of medication combinations when recommending medications, making medication recommendation more difficult. There is little existing work to explore the changes in patient conditions. However, these changes may point to future trends in patient conditions that are critical for reducing DDI rates in recommended drug combinations. In this paper, we proposed the Patient Condition Change Network (PCCNet), which models the current core medications of patient by mining the temporal and spatial changes of patient medication order and patient condition vector, and allocates some auxiliary medications as the currently recommended medication combination. The experimental results show that the proposed model greatly reduces the recommended DDI of medications while achieving results no lower than the state-of-the-art results.11The code is available at https://github.com/master032/PCCNet

Medicine recommendation assists physicians in automatically providing medicine combinations, which is critical in health care. Existing efforts focus on making medicine recommendations based on the patient’s electronic health record(EHR). However, they ignore external medicine knowledge and are vulnerable to the missing EHR. In this paper, a knowledge-enhanced dual graph neural network (KDGN) is proposed to recommend medicine sets. KDGN combines diagnosis-level and procedure-level attention mechanisms to encode multiple types of medical codes. In order to mine medicine from medical knowledge, KDGN further designs a dual-graph neural network, which constructs a medicine co-occurrence graph and molecular connection graph, and retrieves potential therapeutic drugs. Furthermore, during the training phase, we introduce the automatic correction loss based on maximum likelihood estimation to mitigate the impact of missing EHR and enhance the robustness of KDGN. We evaluate the proposed model on the public MIMIC-III dataset, and experimental results show that KDGN outperforms the state-of-the-art model in 4 out of 5 evaluation metrics. Our dataset and code are available at: https://github.com/Benjamin-cell/KDGN.

Biomedical named entity recognition (Bio-NER) is an essential task for biomedical information extraction. In this paper, we regard word-level features and character-level features as two different modalities from a novel perspective and propose a biomedical named entity recognition model based on character feature enhancement with multimodal method (called BioNER-CFEM). BioNER-CFEM can not only capture interactions between modalities, but also learn interactions within modalities. In addition, our proposed cross-attention based sparse selection mechanism can effectively alleviate the noise in the interaction process of the two ‘modalities’. Experimental results show the effectiveness of BioNER-CFEM for the Bio-NER task: it achieves performance boost over SOTA models with competitive efficiency on all six Bio-NER datasets, i.e., $+0.89, +0.64, +0.40$, $+1.40, +5.57, +2.81$ on NCBI-Disease, BC5CDR-Disease, BC5CDR-Chem, BC2GM, JNLPBA, BC4CHEMD, respectively.

Automatic extraction of biomedical relations is important for many tasks, such as drug discovery, protein prediction. However, since the biomedical corpus contains a large number of complex long sentences and overlapping triples, most general domain joint modeling methods are unapplicable to the biomedical domain. Aiming at the problem of overlapping triples in the biomedical field, we design a token pair tagger to achieve single-stage joint biomedical entity and relation extraction. In addition, in order to make the tagger better understand the semantic information and structural information, we design both global-aware module and distance-aware module to introduce location information. Compared with previous methods, our model makes better use of location information, enhances the understanding of complex long sentences, and improves the ability to extract overlapping triples. We evaluated our model on the CHEMPROT and DDIExtraction2013 datasets. Experimental results show that our model significantly outperforms a range of baseline models, achieving the state-of-the-art performance.

Gene Ontology (GO) containing a set of standard concepts (or terms) is launched to unify the functional descriptions of proteins. Developing computational models based on GO to automatically annotate protein functions has been a longstanding active research area. In this paper, we propose a novel method to adaptively fuse functional and topological information between GO Terms. Our method is composed of a pre-trained language model for encoding protein sequences and an adaptive multi-view graph convolutional network (Multi-view GCN) for representing GO terms. Particularly, the Multi-view GCN considers multiple views from functional information, topological structures, and their combinations, and extracts multiple corresponding representations of GO terms. Then, an attention mechanism is applied to adaptively learn the importance weights of these representations. Finally, the predicted scores are calculated by using a dot product between protein sequence features and GO term representations. Experimental results on the datasets of two species (i.e., Human and Yeast) show that our method outperforms other state-of-the-art methods. The code of our proposed method is available at: https://github.com/Candyperfect/Master.

The identification of protein complexes can help understand cellular organization principles and the mechanism of biological evolution. In recent years, researchers have proposed numerous computational methods to identify protein complexes through their interaction networks. Most of these methods identify protein complexes based on the topological structure of the PPI network. However, the topological structure contained in the PPI network is very complicated, and the applicability of advanced representation learning methods has not been researched in depth. This paper proposes a contrastive self-supervised representation learning method to identify protein complexes. Our method uses a mix-hop aggregator based on graph neural network (GNN) to capture high-order interaction in the PPI network and leverage a contrastive self-supervised method to train our model without introducing protein labels. Then, we get the vector representation for each protein and construct a weighted PPI network based on the vector representation similarity. Finally, we apply clustering aggregation to identify protein complexes based on a weighted PPI network. In order to access our method, different PPI networks, DIP, Kroganl4k and Biogrid, are used as datasets. By comparing the competing methods including COACH, CMC, MCODE, ClusterONE, GANE and COAN, experimental results show that our method outperforms classic and state-of-the-art methods.

Intelligent medical services have attracted great research interests for providing automated medical consultation. However, the lack of corpora becomes a main obstacle to related research, particularly data from real scenarios. In this paper, we construct RealMedDial, a Chinese medical dialogue dataset based on real medical consultation. RealMedDial contains 2,637 medical dialogues and 24,255 utterances obtained from Chinese short-video clips of real medical consultations. We collected and annotated a wide range of meta-data with respect to medical dialogue including doctor profiles, hospital departments, diseases and symptoms for fine-grained analysis on language usage pattern and clinical diagnosis. We evaluate the performance of medical response generation, department routing and doctor recommendation on RealMedDial. Results show that RealMedDial are applicable to a wide range of NLP tasks with respect to medical dialogue.

Chinese Named Entity Recognition (NER) has continued to attract research attention. However, most existing studies only explore the internal features of the Chinese language but neglect other lingual modal features. Actually, as another modal knowledge of the Chinese language, English contains rich prompts about entities that can potentially be applied to improve the performance of Chinese NER. Therefore, in this study, we explore the bilingual enhancement for Chinese NER and propose a unified bilingual interaction module called the Adapted Cross-Transformers with Global Sparse Attention (ACT-S) to capture the interaction of bilingual information. We utilize a model built upon several different ACT-Ss to integrate the rich English information into the Chinese representation. Moreover, our model can learn the interaction of information between bilinguals (inter-features) and the dependency information within Chinese (intra-features). Compared with existing Chinese NER methods, our proposed model can better handle entities with complex structures. The English text that enhances the model is automatically generated by machine translation, avoiding high labour costs. Experimental results on four well-known benchmark datasets demonstrate the effectiveness and robustness of our proposed model.

Memes have become the popular means of communication for Internet users worldwide. Understanding the Internet meme is one of the most tricky challenges in natural language processing (NLP) tasks due to its convenient non-standard writing and network vocabulary. Recently, many linguists suggested that memes contain rich metaphorical information. However, the existing researches ignore this key feature. Therefore, to incorporate informative metaphors into the meme analysis, we introduce a novel multimodal meme dataset called MET-Meme, which is rich in metaphorical features. It contains 10045 text-image pairs, with manual annotations of the metaphor occurrence, sentiment categories, intentions, and offensiveness degree. Moreover, we propose a range of strong baselines to demonstrate the importance of combining metaphorical features for meme sentiment analysis and semantic understanding tasks, respectively. MET-Meme, and its code are released publicly for research in \urlhttps://github.com/liaolianfoka/MET-Meme-A-Multi-modal-Meme-Dataset-Rich-in-Metaphors.

Session-based recommendation aims to predict items that an anonymous user would like to purchase based on her short behavior sequence. The current approaches towards session-based recommendation only focus on modeling users' interest preferences, while they all ignore a key attribute of an item, i.e., the price. Many marketing studies have shown that the price factor significantly influences users' behaviors and the purchase decisions of users are determined by both price and interest preferences simultaneously. However, it is nontrivial to incorporate price preferences for session-based recommendation. Firstly, it is hard to handle heterogeneous information from various features of items to capture users' price preferences. Secondly, it is difficult to model the complex relations between price and interest preferences in determining user choices.

External document knowledge is helpful for dialogue systems to generate high-quality responses. Although several knowledge-grounded dialogue models have been designed, external knowledge cannot be comprehensively exploited due to the complex relationships among dialogue context, knowledge, and responses. To this end, we propose a novel transformer-based model, named TransIKG, which incorporates external document knowledge for dialogue generation. TransIKG comprises a two-step integration mechanism, including correlation integration and overall integration. Correlation integration is designed to fully exploit the pairwise mutual information among dialogue context, knowledge, and responses, while overall integration adopts an integration gate to capture global information. Furthermore, we utilize the positional information of dialogue turns to better represent the dialogue context and enhance the generalization ability of our model on out-of-domain documents. Finally, we propose a novel knowledge-aware pointer network to generate knowledge-enhanced response tokens. Experimental results on two benchmark datasets demonstrate that our model outperforms state-of-the-art models on both open-domain and domain-specific dialogues.

Chemical-induced disease (CID) relation extraction from biomedical articles plays an important role in disease treatment and drug development. Existing methods are insufficient for capturing complete document level semantic information due to ignoring semantic information of entities in different sentences. In this work, we proposed an effective document-level relation extraction model to automatically extract intra-/inter-sentential CID relations from articles. Firstly, our model employed BERT to generate contextual semantic representations of the title, abstract and shortest dependency paths (SDPs). Secondly, to enhance the semantic representation of the whole document, cross attention with self-attention (named cross2self-attention) between abstract, title and SDPs was proposed to learn the mutual semantic information. Thirdly, to distinguish the importance of the target entity in different sentences, the Gaussian probability distribution was utilized to compute the weights of the co-occurrence sentence and its adjacent entity sentences. More complete semantic information of the target entity is collected from all entities occurring in the document via our presented document-level R-BERT (DocR-BERT). Finally, the related representations were concatenated and fed into the softmax function to extract CIDs. We evaluated the model on the CDR corpus provided by BioCreative V. The proposed model without external resources is superior in performance as compared with other state-of-the-art models (our model achieves 53.5%, 70%, and 63.7% of the F1-score on inter-/intra-sentential and overall CDR dataset). The experimental results indicate that cross2self-attention, the Gaussian probability distribution and DocR-BERT can effectively improve the CID extraction performance. Furthermore, the mutual semantic information learned by the cross self-attention from abstract towards title can significantly influence the extraction performance of document-level biomedical relation extraction tasks.

Semantic network is a network of concepts connected by semantic relations. It contains two forms ofbinary semantic networkandmultiplex semantic network. The associated fact prediction is a link prediction task that aims to infer the implicitly connected facts by mining the high-level representation of the network. Previous methods for associated fact prediction put much emphasis on the topological feature of network but not utilize the information of semantic expression. This paper proposes aSemanticNetworkEncoder (SemNE), which learns a feature mapping function from the binary semantic networks and can be applied to the multiplex semantic networks in a pre-training manner. SemNE is a two-stage framework that contains an embedding encoder and a prediction decoder. It jointly models the semantic information and network topology to enrich the network representation. A word self-organization method based on the factual boundary is proposed to unify the topological feature and the semantic feature representations. Experimental results on binary semantic networks show that SemNE achieves the state-of-the-art results in associated fact prediction and experimental results on multiplex semantic networks show that SemNE is scalable and can effectively improve the performance of existing models.

Dongyu Zhang, Minghao Zhang, Heting Zhang, Liang Yang, Hongfei Lin. Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers). 2021.

Xianbing Zhou, Yang Yong, Xiaochao Fan, Ge Ren, Yunfeng Song, Yufeng Diao, Liang Yang, Hongfei Lin. Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers). 2021.

Biomedical Named Entity Recognition (BioNER) is often modeled as a sequence labeling task, which assigns the predefined label to each token in given input sequence. Although these sequential labeling models achieve significant achievements, they often fail to give precise boundaries of the named entity. In addition, a vast amount of work focuses more on textual sequence representation but ignores label information. To tackle these problems, in this paper, we directly model span-level named entity recognition, specifically, we treat the BioNER as a joint task of boundary detection and span classification under a multitask framework. In order to enhance boundary supervision, we introduce an entity type label as an additional guide and propose a co-attentive interactive mechanism to improve the span representation. Extensive experiments1on four benchmark datasets demonstrate that our proposed method obtains competitive results, achieving 90.26%, 78.04%, 90.21%, and 86.58% on BC5CDR, JNLPBA, NCBI, and BC2GM datasets, respectively, in terms of F1 score.

The emergence and spread of severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2) have created an enormous socioeconomic impact. Although there are several promising drug candidates in clinical trials, none of them are approved yet. Thus, the drug repositioning approach may help to overcome the current pandemic. However, the sparse dataset of COVID-19 limits the accuracy of existing drug repositioning. To overcome this problem, we propose a novel drug repositioning framework (named Drug2Cov). Drug2Cov can learn an effective representation via integrating self-supervised learning with sparse data. Meanwhile, Drug2Cov uses a heterogeneous graph neural network to capture the complex interaction between viruses, targets, and drugs that enhance the accuracy of drug repositioning. The experimental results demonstrate the effectiveness and feasibility of our proposed Drug2Cov framework. Source code and dataset are freely available at https://github.com/lhf3291109/Drug2Cov.

The goal of relation extraction task is to classify texts containing entity pairs into predefined relation types. Biomedical relation extraction can extract high-quality information from massive medical texts, which plays an important role in biomedical research. In this paper, we propose a self-supervised graph attention network to extract biomedical relations from the complex and noisy biomedical texts. The model incorporates self-supervision within the standard graph attention mechanism. Specifically, the model applies the graph attention mechanism to reduce the influence of noisy words and introduces dependency-based parse trees to construct a self-supervised task. With the supervision of dependency-based parse trees, the graph attention network can not only improve its capacity of learning syntactic information but also alleviate its lack of interpretability. Additionally, we use Gumbel Tree-GRU to obtain sentence information for relation classification. Our model achieves state-of-the-art performance on the DDIExtraction 2013 and ChemProt datasets, respectively, which suggests that our proposed model can effectively improve the performance of biomedical relation extraction.

Automatically extracting entity-pair interactions from biomedical literature plays an important role in promoting the development of the biomedical field. For instance, the interactions between drugs can guide patients to take drugs correctly and avoid clinical adverse drug reactions; The interactions between proteins can help researchers design therapeutic drugs and discover disease mechanisms. However, it is found that relation instances of different classes generated from the same sentence are easily classified into the same class since their context information is almost the same. To address this issue, Triplet Loss based BERT (TL-BERT) approach is proposed in this paper, where the Triplet Loss training strategy is first introduced into the biomedical relation extraction field. Triplet Loss training strategy will increase the distances between these instances generated from the same sentence but belonging to different classes, and decrease the distances between these instances generated from different sentences but belonging to the same class. As a result, our approach can classify these instances generated from the same sentence more correctly. TL-BERT was evaluated on AIMed, BioInfer, and DDI Extraction-2013 corpus, the experimental results demonstrate that the Triplet Loss training strategy can improve the performance on both Protein-Protein interactions extraction tasks and Drug-Drug interactions detection tasks.

Recent metaphor identification approaches mainly consider the contextual text features within a sentence or introduce external linguistic features to the model.But they usually ignore the extra information that the data can provide, such as the contextual metaphor information and broader discourse information.In this paper, we propose a model augmented with hierarchical contextualized representation to extract more information from both sentence-level and discourse-level.At the sentence level, we leverage the metaphor information of words that except the target word in the sentence to strengthen the reasoning ability of our model via a novel label-enhanced contextualized representation.At the discourse level, the position-aware global memory network is adopted to learn long-range dependency among the same words within a discourse.Finally, our model combines the representations obtained from these two parts.The experiment results on two tasks of the VUA dataset show that our model outperforms every other state-of-the-art method that also does not use any external knowledge except what the pre-trained language model contains.

Intent detection and slot filling are two main tasks for building a spoken language understanding (SLU) system. Since the two tasks are closely related, the joint models for the two tasks always outperform the pipeline models in SLU. However, most joint models directly incorporate multiple intent information for each token, which introduces intent noise into the sentence semantics, causing a decrease in the performance of the joint model. In this paper, we propose a Dynamic Graph Model (DGM) for joint multiple intent detection and slot filling, in which we adopt a sentence-level intent-slot interactive graph to model the correlation between the intents and slot. Besides, we design a novel method of constructing the graph, which can dynamically update the interactive graph and further alleviate the error propagation. Experimental results on several multi-intent and single-intent datasets show that our model not only achieves the state-of-the-art (SOTA) performance but also boosts the speed by three to six times over the SOTA model.

Recommendation systems can help users process large amounts of information, and generative adversarial networks (GANs) show great potential in recommendation systems. In this paper, we propose a new GAN model to enhance the information flow within the generator based on the information flow between the original generator and discriminator. Our experimental results indicate that our model reduces the discrepancy between the generator and the discriminator. Both the generator and discriminator yield considerable performance improvements compared to other strong baselines. The improvements by [email protected] and MRR are significant, which can reach 30.98% and 30.17%, respectively.

BACKGROUND: The recognition of pharmacological substances, compounds and proteins is essential for biomedical relation extraction, knowledge graph construction, drug discovery, as well as medical question answering. Although considerable efforts have been made to recognize biomedical entities in English texts, to date, only few limited attempts were made to recognize them from biomedical texts in other languages. PharmaCoNER is a named entity recognition challenge to recognize pharmacological entities from Spanish texts. Because there are currently abundant resources in the field of natural language processing, how to leverage these resources to the PharmaCoNER challenge is a meaningful study. METHODS: Inspired by the success of deep learning with language models, we compare and explore various representative BERT models to promote the development of the PharmaCoNER task. RESULTS: The experimental results show that deep learning with language models can effectively improve model performance on the PharmaCoNER dataset. Our method achieves state-of-the-art performance on the PharmaCoNER dataset, with a max F1-score of 92.01%. CONCLUSION: For the BERT models on the PharmaCoNER dataset, biomedical domain knowledge has a greater impact on model performance than the native language (i.e., Spanish). The BERT models can obtain competitive performance by using WordPiece to alleviate the out of vocabulary limitation. The performance on the BERT model can be further improved by constructing a specific vocabulary based on domain knowledge. Moreover, the character case also has a certain impact on model performance.

The World Health Organization underlines the significance of adverse drug reaction (ADR) reports for patients' safety. Actually, many potential ADRs tend to be under-reported in post-market ADR surveillance. Recognizing ADRs from social media is indispensably important and could complement post-market ADR surveillance for more effective pharmacovigilance studies. However, previous approaches pose two challenges: 1) ADRs show high expression variability in social media, and thus, many potential ADRs are out-of-lexicon ones, which are difficult to be recognized, and 2) most phrasal ADRs are non-standard mentions and their boundaries are difficult to identify accurately. To tackle these challenges, we design three interaction graphs and propose a neural network approach, i.e., Interaction Graph Network (IGN). Specifically, to recognize more out-of-lexicon ADRs, besides the mentions in ADR lexicon, noun phrases in the input sentence are regarded as candidate phrases and their features are taken into considerations. Moreover, in an attempt to accurately identify ADR boundaries, three word-phrase interaction graphs are designed to represent lexicon knowledge and are encoded using graph attention networks (GATs) to directly integrate various boundary and contextual information of candidate phrases into ADR recognition. Experimental results on two benchmark datasets show that IGN can recognize ADR accurately and consistently outperforms other state-of-the-art approaches.

Estimating the similarity of biomedical sentence pair is an important component in such natural language processing (NLP) tasks as text retrieval and text summarization with great amount of biomedical information growing. Deep learning-based approaches have been successfully applied to the task, but they often rely on traditional pre-trained context-independent word embedding. Bidirectional Encoder Representations from Transformers (BERT) is recently employed to pre-train contextualized word/sentence representation models via bidirectional Transformers, outperforming the state-of-the-art for many NLP tasks. The mutual semantic influence between sentences is important for estimating semantic textual similarity, which is neglected in existing methods including BERT. On the other hand, biomedical corpora mainly consist of syntactic complex and long sentences. Owing to the above-mentioned issues, we proposed a hybrid architecture, integrating the pre-trained BERT and downstream bidirectional recurrent neural network (bi-RNN). The proposed model enhanced the sentence semantic representation via employing the self-attention instead of global attention to perform cross attention between sentences. Meanwhile, bi-RNN reduced redundant information in the output of BERT. Experimental results show that the best fine-tuned models consistently outperform previous methods and advance the state-of-the-art for clinical semantic textual similarity in OHNLP 2018 task 2, with up to 0.6% increase in Pearson correlation coefficient.

Severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2) is the strain of coronavirus that causes coronavirus disease 2019 (COVID-19), which leads to over 800,000 deaths and is still no specific medicines. Drug repositioning aiming to infer potential drugs for diseases and achieve much attention during the SARS-CoV-2 epidemic. However, find a specific drug of SARS-CoV-2 is still a large challenge that cannot be addressed well with current methods. To overcome this problem, we present a novel drug repositioning framework of heterogeneous graph convolutional networks for SARS-CoV2. The deep2CoV model can effectively search the potential drugs for SARS-CoV-2, which reduce the number of clinical trials and drug development cycles. The experimental results demonstrate the effectiveness and feasibility of our proposed deep2CoV framework.

Detecting adverse drug events from Twitter is expected to reveal unreported side effects, thereby complementing current spontaneous reporting systems. However, existing studies usually only use word embeddings as the input for deep learning models, which ignores the structural information of sentences. In addition, deep learning models usually require a large number of cases for training, but the scale of annotated corpora that can be used for this task is limited. In order to solve the above problems, we propose a graph-boosted framework, that constructs the text into a graph structure. By using pre-trained graph embeddings and word embeddings for model training, our proposed framework provides richer semantic and structural information for prediction. The experimental results show that the proposed method can be used in different deep learning models and bring improvements when using the TwiMed corpus of different scales.

Extracting protein-protein interaction affected by genetic mutation from biomedical literature automatically is an essential step toward the ultimate goal of precision medicine. However, the existing methods fail to be accurate enough to meet the needs in practice. In this paper, considering the significant progress made by the pre-training model in a wide variety of NLP tasks, we use BioBERT to obtain the representation of the text and adopt a multi-task learning strategy to improve the performance. Evaluated on the BioCreative VI PPIm data set, our proposed model achieves a new state-of-the-art performance that surpassed the previous one by 4.86% in F1-score. The source code is available at https://github.com/dlutwy/ppim.

Automatic extraction of biomedical relations is important for many tasks, such as drug discovery, protein prediction and knowledge graph construction. However, due to the complex and noisy expressions in biomedical texts, existing traditional neural networks, such as recurrent neural networks and convolutional neural networks, fail to capture syntactic information effectively. In this paper, we introduce a multi-head attention mechanism into graph convolutional networks to extract biomedical relations. In our method, the graph convolutional network is exploited to encode the dependency structure of an input sentence and the multi-head attention mechanism is utilized to alleviate the influence of noisy words. We evaluated our method on the ChemProt corpus and the protein-protein interaction corpus which includes five separate sub-datasets and it achieves F-scores of 67.37% and 84.8% on ChemProt corpus and PPI corpora, respectively. The experimental results suggest that Our model can not only alleviate the influence of noisy words, but also obtain more semantic and syntactic information from dependency graph than previous proposed models.

Relations between mutations and diseases hiding in biomedical literature are valuable for the analysis and interpretation of many complex diseases, which can help explore more effective treatment options for corresponding diseases. Most current document-level mutation-disease relation extraction methods are based on classification approaches and suffer the lack of ability to extract inter-sentential relations. To solve this problem, we regard extracting document-level mutation-disease relations as a sequence tagging task and propose a neural network-based method called Star-BiLSTM-LAN. By combining the star transformer and the Bi-directional Long Short-Term Memory network, this method achieves a strong ability to capture semantic and syntactic information at the document level from different aspects, and it can discover internal representations that prove useful for the task of interest. Star-BiLSTM-LAN is evaluated on EMU BCa and PCa datasets, and achieves the state-of-the-art F-scores of 89.20% and 90.43%, which are 4.70% and 2.43% higher than the baseline, respectively. Also, the proposed method achieves an F-score of 94.40% on BRONCO dataset.

In this paper, we propose a gated iterative capsule network model for the ADR detection task, named GICN. To alleviate the impact caused by abbreviations and misspelled words, we add character embedding as part of the input. Most ADRs consist of multiple words, e.g., short-term memory dysfunction. Hence, we apply a convolutional neural network (CNN) to obtain the complete phrase information. To effectively extract deep semantic information, we introduce a capsule network with a gated iteration unit that clusters features from underlying to high capsules. The gated iteration mechanism can remember contextual information, which will be introduced when clustering features. Experimental results show that our approach can achieve significant performance improvement for ADR detection from social media text compared with other state-of-the-art works.

In recent years, the plentiful information contained in Chinese legal documents has attracted a great deal of attention because of the large-scale release of the judgment documents on China Judgments Online.It is in great need of enabling machines to understand the semantic information stored in the documents which are transcribed in the form of natural language.The technique of information extraction provides a way of mining the valuable information implied in the unstructured judgment documents.We propose a Legal Triplet Extraction System for drug-related criminal judgment documents.The system extracts the entities and the semantic relations jointly and benefits from the proposed entity feature and multi-task learning framework.Furthermore, we manually annotate a dataset for Named Entity Recognition and Relation Extraction in Chinese legal domain, which contributes to training supervised triplet extraction models and evaluating the model performance.Our experimental results show that the entity feature introduction and multi-task learning framework are feasible and effective for the Legal Triplet Extraction System.The F1 score of triplet extraction finally reaches 0.836 on the legal dataset.

The Legal Judgement Prediction (LJP) is now under the spotlight. And it usually consists of multiple sub-tasks, such as penalty prediction (fine and imprisonment) and the prediction of articles of law. For penalty prediction, they are often closely related to the trial process, especially the attitude analysis of criminal suspects, which will influence the judgment of the presiding judge to some extent. In this paper, we firstly construct a multi-modal dataset with 517 cases of intentional assault, which contains trial information as well as the attitude of the suspect. Then, we explore the relationship between suspect`s attitude and term of imprisonment. Finally, we use the proposed multi-modal model to predict the suspect's attitude, and compare it with several strong baselines. Our experimental results show that the attitude of the criminal suspect is closely related to the penalty prediction, which provides a new perspective for LJP.

MOTIVATION: The biomedical literature contains a wealth of chemical-protein interactions (CPIs). Automatically extracting CPIs described in biomedical literature is essential for drug discovery, precision medicine, as well as basic biomedical research. Most existing methods focus only on the sentence sequence to identify these CPIs. However, the local structure of sentences and external biomedical knowledge also contain valuable information. Effective use of such information may improve the performance of CPI extraction. RESULTS: In this article, we propose a novel neural network-based approach to improve CPI extraction. Specifically, the approach first employs BERT to generate high-quality contextual representations of the title sequence, instance sequence and knowledge sequence. Then, the Gaussian probability distribution is introduced to capture the local structure of the instance. Meanwhile, the attention mechanism is applied to fuse the title information and biomedical knowledge, respectively. Finally, the related representations are concatenated and fed into the softmax function to extract CPIs. We evaluate our proposed model on the CHEMPROT corpus. Our proposed model is superior in performance as compared with other state-of-the-art models. The experimental results show that the Gaussian probability distribution and external knowledge are complementary to each other. Integrating them can effectively improve the CPI extraction performance. Furthermore, the Gaussian probability distribution can effectively improve the extraction performance of sentences with overlapping relations in biomedical relation extraction tasks. AVAILABILITY AND IMPLEMENTATION: Data and code are available at https://github.com/CongSun-dlut/CPI_extraction. SUPPLEMENTARY INFORMATION: Supplementary data are available at Bioinformatics online.

BACKGROUND: Both intra- and inter-sentential semantic relations in biomedical texts provide valuable information for biomedical research. However, most existing methods either focus on extracting intra-sentential relations and ignore inter-sentential ones or fail to extract inter-sentential relations accurately and regard the instances containing entity relations as being independent, which neglects the interactions between relations. We propose a novel sequence labeling-based biomedical relation extraction method named Bio-Seq. In the method, sequence labeling framework is extended by multiple specified feature extractors so as to facilitate the feature extractions at different levels, especially at the inter-sentential level. Besides, the sequence labeling framework enables Bio-Seq to take advantage of the interactions between relations, and thus, further improves the precision of document-level relation extraction. RESULTS: Our proposed method obtained an F1-score of 63.5% on BioCreative V chemical disease relation corpus, and an F1-score of 54.4% on inter-sentential relations, which was 10.5% better than the document-level classification baseline. Also, our method achieved an F1-score of 85.1% on n2c2-ADE sub-dataset. CONCLUSION: Sequence labeling method can be successfully used to extract document-level relations, especially for boosting the performance on inter-sentential relation extraction. Our work can facilitate the research on document-level biomedical text mining.

BACKGROUND: Biomedical document triage is the foundation of biomedical information extraction, which is important to precision medicine. Recently, some neural networks-based methods have been proposed to classify biomedical documents automatically. In the biomedical domain, documents are often very long and often contain very complicated sentences. However, the current methods still find it difficult to capture important features across sentences. RESULTS: In this paper, we propose a hierarchical attention-based capsule model for biomedical document triage. The proposed model effectively employs hierarchical attention mechanism and capsule networks to capture valuable features across sentences and construct a final latent feature representation for a document. We evaluated our model on three public corpora. CONCLUSIONS: Experimental results showed that both hierarchical attention mechanism and capsule networks are helpful in biomedical document triage task. Our method proved itself highly competitive or superior compared with other state-of-the-art methods.

BACKGROUND: Most biomedical information extraction focuses on binary relations within single sentences. However, extracting n-ary relations that span multiple sentences is in huge demand. At present, in the cross-sentence n-ary relation extraction task, the mainstream method not only relies heavily on syntactic parsing but also ignores prior knowledge. RESULTS: In this paper, we propose a novel cross-sentence n-ary relation extraction method that utilizes the multihead attention and knowledge representation that is learned from the knowledge graph. Our model is built on self-attention, which can directly capture the relations between two words regardless of their syntactic relation. In addition, our method makes use of entity and relation information from the knowledge base to impose assistance while predicting the relation. Experiments on n-ary relation extraction show that combining context and knowledge representations can significantly improve the n-ary relation extraction performance. Meanwhile, we achieve comparable results with state-of-the-art methods. CONCLUSIONS: We explored a novel method for cross-sentence n-ary relation extraction. Unlike previous approaches, our methods operate directly on the sequence and learn how to model the internal structures of sentences. In addition, we introduce the knowledge representations learned from the knowledge graph into the cross-sentence n-ary relation extraction. Experiments based on knowledge representation learning show that entities and relations can be extracted in the knowledge graph, and coding this knowledge can provide consistent benefits.

In order to save time and cost, many disease gene prediction methods have been proposed in recent years. However, the traditional network model uses a binary relationship to represent the relationship between different proteins or gene molecules and phenotypes, which leads to the loss of information. Recently, hypergraph shows that it can overcome this loss of information to some extent and preserve the multivariate relationship, so we transformed the disease gene prediction problem into the problem of ranking the multivariate-relationship object. In this paper, we propose a method of Heterogeneous Probabilistic Hypergraph Ranking (HPHR) to predict disease genes. Firstly, fix a graph centroid for each hyperedge and according to different associations, and add other nodes related to the graph centroid to hyperedges with a certain probability. Then transform the problem of predicting disease genes into the problem of ranking heterogeneous objects, and the candidate genes are sorted by hypergraph ranking. The method is then applied to the integrated disease gene network. Compared with other prediction methods achieved better results, which was verified by this experiment.

Dongyu Zhang, Heting Zhang, Xikai Liu, Hongfei Lin, Feng Xia. Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP). 2019.

Pre-trained distributed word representations have been proven useful in various natural language processing (NLP) tasks. However, the effect of words’ geometric structure on word representations has not been carefully studied yet. The existing word representations methods underestimate the words whose distances are close in the Euclidean space, while overestimating words with a much greater distance. In this paper, we propose a word vector refinement model to correct the pre-trained word embedding, which brings the similarity of words in Euclidean space closer to word semantics by using manifold learning. This approach is theoretically founded in the metric recovery paradigm. Our word representations have been evaluated on a variety of lexical-level intrinsic tasks (semantic relatedness, semantic similarity) and the experimental results show that the proposed model outperforms several popular word representations approaches.

Heterographic pun plays a critical role in human writing and literature, which usually has a similar sounding or spelling structure. It is important and difficult research to recognize the heterographic pun because of the ambiguity. However, most existing methods for this task only focus on designing features with rule-based or machine learning methods. In this paper, we propose an end-to-end computational approach - Pronunciation Spelling Understanding Gated Attention (PSUGA) network. For pronunciation, we exploit the hierarchical attention model with phoneme embedding. While for spelling, we consider the character-level, word-level, tag-level, position-level and contextual-level embedding with attention model. To deal with the two parts, we present a gated attention mechanism to control the information integration. We have conducted extensive experiments on SemEval2017 task7 and Pun of the Day datasets. Experimental results show that our approach significantly outperforms state-of-the-art methods.

Facial appearance matters in social networks. Individuals frequently make trait judgments from facial clues. Although these face-based impressions lack the evidence to determine validity, they are of vital importance, because they may relate to human network-based social behavior, such as seeking certain individuals for help, advice, dating, and cooperation, and thus they may relate to centrality in social networks. However, little to no work has investigated the apparent facial traits that influence network centrality, despite the large amount of research on attributions of the central position including personality and behavior. In this paper, we examine whether perceived traits based on facial appearance affect network centrality by exploring the initial stage of social network formation in a first-year college residential area. We took face photos of participants who are freshmen living in the same residential area, and we asked them to nominate community members linking to different networks. We then collected facial perception data by requiring other participants to rate facial images for three main attributions: dominance, trustworthiness, and attractiveness. Meanwhile, we proposed a framework to discover how facial appearance affects social networks. Our results revealed that perceived facial traits were correlated with the network centrality and that they were indicative to predict the centrality of people in different networks. Our findings provide psychological evidence regarding the interaction between faces and network centrality. Our findings also offer insights in to a combination of psychological and social network techniques, and they highlight the function of facial bias in cuing and signaling social traits. To the best of our knowledge, we are the first to explore the influence of facial perception on centrality in social networks.

BACKGROUND: The number of biomedical research articles have increased exponentially with the advancement of biomedicine in recent years. These articles have thus brought a great difficulty in obtaining the needed information of researchers. Information retrieval technologies seek to tackle the problem. However, information needs cannot be completely satisfied by directly introducing the existing information retrieval techniques. Therefore, biomedical information retrieval not only focuses on the relevance of search results, but also aims to promote the completeness of the results, which is referred as the diversity-oriented retrieval. RESULTS: We address the diversity-oriented biomedical retrieval task using a supervised term ranking model. The model is learned through a supervised query expansion process for term refinement. Based on the model, the most relevant and diversified terms are selected to enrich the original query. The expanded query is then fed into a second retrieval to improve the relevance and diversity of search results. To this end, we propose three diversity-oriented optimization strategies in our model, including the diversified term labeling strategy, the biomedical resource-based term features and a diversity-oriented group sampling learning method. Experimental results on TREC Genomics collections demonstrate the effectiveness of the proposed model in improving the relevance and the diversity of search results. CONCLUSIONS: The proposed three strategies jointly contribute to the improvement of biomedical retrieval performance. Our model yields more relevant and diversified results than the state-of-the-art baseline models. Moreover, our method provides a general framework for improving biomedical retrieval performance, and can be used as the basis for future work.

BACKGROUND: The adverse reactions that are caused by drugs are potentially life-threatening problems. Comprehensive knowledge of adverse drug reactions (ADRs) can reduce their detrimental impacts on patients. Detecting ADRs through clinical trials takes a large number of experiments and a long period of time. With the growing amount of unstructured textual data, such as biomedical literature and electronic records, detecting ADRs in the available unstructured data has important implications for ADR research. Most of the neural network-based methods typically focus on the simple semantic information of sentence sequences; however, the relationship of the two entities depends on more complex semantic information. METHODS: In this paper, we propose multihop self-attention mechanism (MSAM) model that aims to learn the multi-aspect semantic information for the ADR detection task. first, the contextual information of the sentence is captured by using the bidirectional long short-term memory (Bi-LSTM) model. Then, via applying the multiple steps of an attention mechanism, multiple semantic representations of a sentence are generated. Each attention step obtains a different attention distribution focusing on the different segments of the sentence. Meanwhile, our model locates and enhances various keywords from the multiple representations of a sentence. RESULTS: Our model was evaluated by using two ADR corpora. It is shown that the method has a stable generalization ability. Via extensive experiments, our model achieved F-measure of 0.853, 0.799 and 0.851 for ADR detection for TwiMed-PubMed, TwiMed-Twitter, and ADE, respectively. The experimental results showed that our model significantly outperforms other compared models for ADR detection. CONCLUSIONS: In this paper, we propose a modification of multihop self-attention mechanism (MSAM) model for an ADR detection task. The proposed method significantly improved the learning of the complex semantic information of sentences.

The explosive growth of biomedical literature has created a rich source of knowledge, such as that on protein-protein interactions (PPIs) and drug-drug interactions (DDIs), locked in unstructured free text. Biomedical relation classification aims to automatically detect and classify biomedical relations, which has great benefits for various biomedical research and applications. In the past decade, significant progress has been made in biomedical relation classification. With the advance of neural network methodology, neural network-based approaches have been applied in biomedical relation classification and achieved state-of-the-art performance for some public datasets and shared tasks. In this review, we describe the recent advancement of neural network-based approaches for classifying biomedical relations. We summarize the available corpora and introduce evaluation metrics. We present the general framework for neural network-based approaches in biomedical relation extraction and pretrained word embedding resources. We discuss neural network-based approaches, including convolutional neural networks (CNNs) and recurrent neural networks (RNNs). We conclude by describing the remaining challenges and outlining future directions.

With the rapid development of biomedicine, the number of biomedical articles has increased accordingly, which presents a great challenge for biologists trying to keep up with the latest research. Information retrieval seeks to meet this challenge by searching among a large number of articles based on given queries and providing the most relevant ones to fulfill information needs. As an effective information retrieval technique, query expansion has some room for improvement to achieve the desired performance when directly applied for biomedical information retrieval because there exist many domain-related terms both in users' queries and in related articles. To solve this problem, we propose a biomedical query expansion framework based on learning-to-rank methods, in which we refine candidate expansion terms by training term-ranking models to select the most relevant terms. To train the term-ranking models, we first propose a pseudo-relevance feedback method based on MeSH to select candidate expansion terms and then represent the candidate terms as feature vectors by defining both the corpus-based term features and the resource-based term features. Experimental results obtained for TREC genomics datasets show that our method can capture more relevant terms to expand the original query and effectively improve biomedical information retrieval performance.

Multivariate chaotic time series prediction is a hot research topic, the goal of which is to predict the future of the time series based on past observations. Echo state networks (ESNs) have recently been widely used in time series prediction, but there may be an ill-posed problem for a large number of unknown output weights. To solve this problem, we propose a hybrid regularized ESN, which employs a sparse regression with the L1/2regularization and the L2regularization to compute the output weights. The L1/2penalty shows many attractive properties, such as unbiasedness and sparsity. The L2penalty presents appealing ability on shrinking the amplitude of the output weights. After the output weights are calculated, the input weights, internal weights, and output weights are fine-tuning by a Hessian-free optimization method-conjugate gradient backpropagation algorithm. The fine-tuning helps to bubble up the input information toward the output layer. Besides, the largest Lyapunov exponent is used to calculate the predictable horizon of a chaotic time series. Experimental results on benchmark and real-world datasets show that our proposed method is superior to other ESN-based models, as sparser, smaller-absolute-value, and more informative output weights are obtained. All of the predictions within the predictable horizon of the proposed model are accurate.

Spatio-temporal series prediction has attracted increasing attention in the field of meteorology in recent years. The spatial and temporal joint effect makes predictions challenging. Most of the existing spatio-temporal prediction models are computationally complicated. To develop an accurate but easy-to-implement spatio-temporal prediction model, this paper designs a novel spatio-temporal prediction model based on echo state networks. For real-world observed meteorological data with randomness and large changes, we use a cubic spline method to bridge the gaps between the neighboring points, which results in a pleasingly smooth series. The interpolated series is later input into the spatio-temporal echo state networks, in which the spatial coefficients are computed by the elastic-net algorithm. This approach offers automatic selection and continuous shrinkage of the spatial variables. The proposed model provides an intuitive but effective approach to address the interaction of spatial and temporal effects. To demonstrate the practicality of the proposed model, we apply it to predict two real-world datasets: monthly precipitation series and daily air quality index series. Experimental results demonstrate that the proposed model achieves a normalized root-mean-square error of approximately 0.250 on both datasets. Similar results are achieved on the long short-term memory model, but the computation time of our proposed model is considerably shorter. It can be inferred that our proposed neural network model has advantages on predicting meteorological series over other models.

Yufeng Diao, Hongfei Lin, Di Wu, Liang Yang, Kan Xu, Zhihao Yang, Jian Wang, Shaowu Zhang, Bo Xu, Dongyu Zhang. Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing. 2018.

Motivation: In biomedical research, chemical is an important class of entities, and chemical named entity recognition (NER) is an important task in the field of biomedical information extraction. However, most popular chemical NER methods are based on traditional machine learning and their performances are heavily dependent on the feature engineering. Moreover, these methods are sentence-level ones which have the tagging inconsistency problem. Results: In this paper, we propose a neural network approach, i.e. attention-based bidirectional Long Short-Term Memory with a conditional random field layer (Att-BiLSTM-CRF), to document-level chemical NER. The approach leverages document-level global information obtained by attention mechanism to enforce tagging consistency across multiple instances of the same token in a document. It achieves better performances with little feature engineering than other state-of-the-art methods on the BioCreative IV chemical compound and drug name recognition (CHEMDNER) corpus and the BioCreative V chemical-disease relation (CDR) task corpus (the F-scores of 91.14 and 92.57%, respectively). Availability and implementation: Data and code are available at https://github.com/lingluodlut/Att-ChemdNER. Contact: [email protected] or [email protected]. Supplementary information: Supplementary data are available at Bioinformatics online.

Motivation: Adverse events resulting from drug-drug interactions (DDI) pose a serious health issue. The ability to automatically extract DDIs described in the biomedical literature could further efforts for ongoing pharmacovigilance. Most of neural networks-based methods typically focus on sentence sequence to identify these DDIs, however the shortest dependency path (SDP) between the two entities contains valuable syntactic and semantic information. Effectively exploiting such information may improve DDI extraction. Results: In this article, we present a hierarchical recurrent neural networks (RNNs)-based method to integrate the SDP and sentence sequence for DDI extraction task. Firstly, the sentence sequence is divided into three subsequences. Then, the bottom RNNs model is employed to learn the feature representation of the subsequences and SDP, and the top RNNs model is employed to learn the feature representation of both sentence sequence and SDP. Furthermore, we introduce the embedding attention mechanism to identify and enhance keywords for the DDI extraction task. We evaluate our approach using the DDI extraction 2013 corpus. Our method is competitive or superior in performance as compared with other state-of-the-art methods. Experimental results show that the sentence sequence and SDP are complementary to each other. Integrating the sentence sequence with SDP can effectively improve the DDI extraction performance. Availability and implementation: The experimental data is available at https://github.com/zhangyijia1979/hierarchical-RNNs-model-for-DDI-extraction. Contact: [email protected] or [email protected]. Supplementary information: Supplementary data are available at Bioinformatics online.

BACKGROUND: Protein complexes are one of the keys to deciphering the behavior of a cell system. During the past decade, most computational approaches used to identify protein complexes have been based on discovering densely connected subgraphs in protein-protein interaction (PPI) networks. However, many true complexes are not dense subgraphs and these approaches show limited performances for detecting protein complexes from PPI networks. RESULTS: To solve these problems, in this paper we propose a supervised learning method based on network node embeddings which utilizes the informative properties of known complexes to guide the search process for new protein complexes. First, node embeddings are obtained from human protein interaction network. Then the protein interactions are weighted through the similarities between node embeddings. After that, the supervised learning method is used to detect protein complexes. Then the random forest model is used to filter the candidate complexes in order to obtain the final predicted complexes. Experimental results on real human and yeast protein interaction networks show that our method effectively improves the performance for protein complex detection. CONCLUSIONS: We provided a new method for identifying protein complexes from human and yeast protein interaction networks, which has great potential to benefit the field of protein complex detection.

BACKGROUND: Drug discovery is the process through which potential new medicines are identified. High-throughput screening and computer-aided drug discovery/design are the two main drug discovery methods for now, which have successfully discovered a series of drugs. However, development of new drugs is still an extremely time-consuming and expensive process. Biomedical literature contains important clues for the identification of potential treatments. It could support experts in biomedicine on their way towards new discoveries. METHODS: Here, we propose a biomedical knowledge graph-based drug discovery method called SemaTyP, which discovers candidate drugs for diseases by mining published biomedical literature. We first construct a biomedical knowledge graph with the relations extracted from biomedical abstracts, then a logistic regression model is trained by learning the semantic types of paths of known drug therapies' existing in the biomedical knowledge graph, finally the learned model is used to discover drug therapies for new diseases. RESULTS: The experimental results show that our method could not only effectively discover new drug therapies for new diseases, but also could provide the potential mechanism of action of the candidate drugs. CONCLUSIONS: In this paper we propose a novel knowledge graph based literature mining method for drug discovery. It could be a supplementary method for current drug discovery methods.

BACKGROUND: In biomedical information extraction, event extraction plays a crucial role. Biological events are used to describe the dynamic effects or relationships between biological entities such as proteins and genes. Event extraction is generally divided into trigger detection and argument recognition. The performance of trigger detection directly affects the results of the event extraction. In general, the traditional method is used to address the trigger detection as a classification task, as well as the use of machine learning or rules method, which construct many features to improve the classification results. Moreover, the classification model only recognizes triggers composed of single words, whereas for multiple words, the result is unsatisfactory. RESULTS: The corpus of our model is MLEE. If we were to only use the biomedical LSTM and CRF model without other features, the F-score would reach about 78.08%. Comparing entity to part of speech (POS), we find the entity features more conducive to the improvement of performance of detection, with the F-score potentially reaching about 80%. Furthermore, we also experiment on the other three corpora (BioNLP 2009, BioNLP 2011, and BioNLP 2013) to verify the generalization of our model. Hence, F-scores can reach more than 60%, which are better than the comparative experiments. CONCLUSIONS: The trigger recognition method based on the sequence annotation model does not require initial complex feature engineering, and only requires a simple labeling mechanism to complete the training. Therefore, generalization of our model is better compared to other traditional models. Secondly, this method can identify multi-word triggers, thereby improving the F-scores of trigger recognition. Thirdly, details on the entity have a crucial impact on trigger detection. Finally, the combination of character-level word embedding and word-level word embedding provides increasingly effective information for the model; therefore, it is a key to the success of the experiment.

BACKGROUND: The effective combination of texts and knowledge may improve performances of natural language processing tasks. For the recognition of chemical-induced disease (CID) relations which may span sentence boundaries in an article, although existing CID systems explored the utilization for knowledge bases, the effects of different knowledge on the identification of a special CID haven't been distinguished by these systems. Moreover, systems based on neural network only constructed sentence or mention level models. RESULTS: In this work, we proposed an effective document level neural model integrated domain knowledge to extract CID relations from biomedical articles. Basic semantic information of an article with respect to a special CID candidate pair was learned from the document level sub-network module. Furthermore, knowledge attention depending on the representation of the article was proposed to distinguish the influences of different knowledge on the special CID pair and then the final representation of knowledge was formed by aggregating weighed knowledge. Finally, the integrated representations of texts and knowledge were passed to a softmax classifier to perform the CID recognition. Experimental results on the chemical-disease relation corpus proposed by BioCreative V show that our proposed system integrated knowledge achieves a good overall performance compared with other state-of-the-art systems. CONCLUSIONS: Experimental analyses demonstrate that the introduced attention mechanism on domain knowledge plays a significant role in distinguishing influences of different knowledge on the judgment for a special CID relation.

Biomedical relation extraction can automatically extract high-quality biomedical relations from biomedical texts, which is a vital step for the mining of biomedical knowledge hidden in the literature. Recurrent neural networks (RNNs) and convolutional neural networks (CNNs) are two major neural network models for biomedical relation extraction. Neural network-based methods for biomedical relation extraction typically focus on the sentence sequence and employ RNNs or CNNs to learn the latent features from sentence sequences separately. However, RNNs and CNNs have their own advantages for biomedical relation extraction. Combining RNNs and CNNs may improve biomedical relation extraction. In this paper, we present a hybrid model for the extraction of biomedical relations that combines RNNs and CNNs. First, the shortest dependency path (SDP) is generated based on the dependency graph of the candidate sentence. To make full use of the SDP, we divide the SDP into a dependency word sequence and a relation sequence. Then, RNNs and CNNs are employed to automatically learn the features from the sentence sequence and the dependency sequences, respectively. Finally, the output features of the RNNs and CNNs are combined to detect and extract biomedical relations. We evaluate our hybrid model using five public (protein-protein interaction) PPI corpora and a (drug-drug interaction) DDI corpus. The experimental results suggest that the advantages of RNNs and CNNs in biomedical relation extraction are complementary. Combining RNNs and CNNs can effectively boost biomedical relation extraction performance.

In these years, the number of biomedical articles has increased exponentially, which becomes a problem for biologists to capture all the needed information manually. Information retrieval technologies, as the core of search engines, can deal with the problem automatically, providing users with the needed information. However, it is a great challenge to apply these technologies directly for biomedical retrieval, because of the abundance of domain specific terminologies. To enhance biomedical retrieval, we propose a novel framework based on learning to rank. Learning to rank is a series of state-of-the-art information retrieval techniques, and has been proved effective in many information retrieval tasks. In the proposed framework, we attempt to tackle the problem of the abundance of terminologies by constructing ranking models, which focus on not only retrieving the most relevant documents, but also diversifying the searching results to increase the completeness of the resulting list for a given query. In the model training, we propose two novel document labeling strategies, and combine several traditional retrieval models as learning features. Besides, we also investigate the usefulness of different learning to rank approaches in our framework. Experimental results on TREC Genomics datasets demonstrate the effectiveness of our framework for biomedical information retrieval.

Trigger detection plays a key role in the extraction of biomedical events, so it will influence the results of biomedical events extraction directly. The traditional biomedical event trigger recognition method is based on artificial design features and construct feature vectors; Not only does it consume great amounts of manpower, it also lacks system generalization ability. Most of methods of trigger detection are based on the convolutional neural network that identify each word in the text, and regard it as a multi-classification task. However for the multi-word composed of the trigger, there is no useful recognition effect. In this paper, we will use the IBO format and consider the trigger detection as a task of sequence annotation, a solution that improves the recognition accuracy of multi-word triggers by bidirectional LSTM and CRF.

Protein-protein interaction extraction research can be widely applied to the field of life science research. However, most of the machine learning based methods focus on binary PPI relation extraction, which loses rich relationship type information that is critical to the PPIs study. The rule based open information extraction methods can extract the PPI triple (i.e. “protein1, interaction word, protein2”), but suffers from low recall rate problem. In this paper, we propose a hybrid protein-protein interaction triple extraction method. In this method, firstly, machine learning techniques are used to recognize protein entities and extract relational protein pairs. Then, the syntactic patterns and a dictionary are employed to find out corresponding interaction words that represent the relationships between two proteins. This method obtains an F-score of 40.18% on the AImed corpus, which is much higher than the result achieved by the rule based Stanford open information extraction method.

Learning to rank utilizes machine learning methods to solve ranking problems by constructing ranking models in a supervised way, which needs fixed-length feature vectors of documents as inputs, and outputs the ranking models learned by iteratively reducing the pre-defined ranking loss. The document features are always extracted based on classic textual statistics, and different features contribute differently to ranking performance. Given that well-defined features would contribute more to the retrieval performance, we investigate the usage of autoencoders to enrich the feature representations of documents. Autoencoders, as basic building blocks of deep neural networks, have been successfully used in many text mining tasks for generating effective features. To enrich the feature space for learning to rank, we introduce supervision into the loss functions of autoencoders. Specifically, we first train a linear ranking model on the training data, and then incorporate the learned weights into the reconstruction costs of an autoencoder. Meanwhile, we accumulate the costs of documents for a given query with query-level constraints for producing more useful features. We evaluate the effectiveness of our model on three LETOR datasets, and show that our model can generate effective document features to improve the retrieval performance.

BACKGROUND: Drug-drug interactions (DDIs) often bring unexpected side effects. The clinical recognition of DDIs is a crucial issue for both patient safety and healthcare cost control. However, although text-mining-based systems explore various methods to classify DDIs, the classification performance with regard to DDIs in long and complex sentences is still unsatisfactory. METHODS: In this study, we propose an effective model that classifies DDIs from the literature by combining an attention mechanism and a recurrent neural network with long short-term memory (LSTM) units. In our approach, first, a candidate-drug-oriented input attention acting on word-embedding vectors automatically learns which words are more influential for a given drug pair. Next, the inputs merging the position- and POS-embedding vectors are passed to a bidirectional LSTM layer whose outputs at the last time step represent the high-level semantic information of the whole sentence. Finally, a softmax layer performs DDI classification. RESULTS: Experimental results from the DDIExtraction 2013 corpus show that our system performs the best with respect to detection and classification (84.0% and 77.3%, respectively) compared with other state-of-the-art methods. In particular, for the Medline-2013 dataset with long and complex sentences, our F-score far exceeds those of top-ranking systems by 12.6%. CONCLUSIONS: Our approach effectively improves the performance of DDI classification tasks. Experimental analysis demonstrates that our model performs better with respect to recognizing not only close-range but also long-range patterns among words, especially for long, complex and compound sentences.

Adverse drug reactions between chemicals and diseases make chemical-disease relations (CDR) become a research focus. In this paper, we present a chemical-induced disease (CID) relation extraction system, CIDExtractor, to extract CID relations from biomedical literature. CIDExtractor first employs a sentence-level classifier to extract the CID relations located in the same sentence. To construct the classifier, a sentence-level training set is manually annotated and then Co-Training algorithm is used to exploit the unlabeled data with the feature kernel and graph kernel as two independent views. Then CIDExtractor uses a document-level classifier to extract the CID relations spanning multiple sentences. The classifier utilizes the document level information (features) of the chemical and disease pair. Finally, some post-processing rules are applied to the union set of two classifiers and generate the final outputs. Experimental results on the test set of BioCreative V CDR CID subtask show that CIDExtractor can achieve better performance (an F-score of 67.72%) than the state-of-the-art methods. The online CIDExtractor demonstration system is available at http://202.118.75.18:8888/cdr-dut-ir/cid.html.

The two main problems of biomedical event extraction are trigger identification and argument detection which can both be considered as classification problems. In this paper, we propose a distributed representation method, which combines context, consisted by dependency-based word embedding, and task-based features represented in a distributed way on deep learning models to realize biomedical event extraction. The experimental results on Multi-Level Event Extraction (MLEE) corpus show higher F-scores compared to the state-of-the-art SVM method. This demonstrates that our proposed method is effective for biomedical event extraction.

In this paper, we present a deep learning based disease named entity recognition architecture. First, the word-level embedding, character-level embedding and lexicon feature embedding are concatenated as input. Then multiple convolutional layers are stacked over the input to extract useful features automatically. Finally, multiple label strategy, which is firstly introduced, is applied to the output layer to capture the correlation information between neighboring labels. Experimental results on both NCBI and CDR corpora show that ML-CNN can achieve the state-of-the-art performance.

High-throughput experimental techniques have produced a large amount of human protein-protein interactions, making it possible to construct a large-scale human PPI network and detect human protein complexes from the network with computational approaches. However, most of current complex detection methods are based on graph theory which can't utilize the information of the known complexes. In this paper, we present a supervised learning method to detect protein complexes in a human PPI network. In this method, biological characteristics and properties of the network are taken into consideration to construct a rich feature set to train a regression model for protein complex detection. In addition, the specific disease related PPIs are extracted from biomedical literatures and then integrated into the original PPI network for detecting the disease-specific protein complexes more effectively. Experimental results show that the performance of our method is superior to other existing state-of-the-art methods. Furthermore, through the analysis of the breast cancer specific complexes detected with our method, more biological insights for breast cancer (e.g., some candidate susceptible genes of breast cancer) are provided.

MOTIVATION: Detecting drug-drug interaction (DDI) has become a vital part of public health safety. Therefore, using text mining techniques to extract DDIs from biomedical literature has received great attentions. However, this research is still at an early stage and its performance has much room to improve. RESULTS: In this article, we present a syntax convolutional neural network (SCNN) based DDI extraction method. In this method, a novel word embedding, syntax word embedding, is proposed to employ the syntactic information of a sentence. Then the position and part of speech features are introduced to extend the embedding of each word. Later, auto-encoder is introduced to encode the traditional bag-of-words feature (sparse 0-1 vector) as the dense real value vector. Finally, a combination of embedding-based convolutional features and traditional features are fed to the softmax classifier to extract DDIs from biomedical literature. Experimental results on the DDIExtraction 2013 corpus show that SCNN obtains a better performance (an F-score of 0.686) than other state-of-the-art methods. AVAILABILITY AND IMPLEMENTATION: The source code is available for academic use at http://202.118.75.18:8080/DDI/SCNN-DDI.zip CONTACT: [email protected] information: Supplementary data are available at Bioinformatics online.

BACKGROUND: Biomedical literature retrieval is becoming increasingly complex, and there is a fundamental need for advanced information retrieval systems. Information Retrieval (IR) programs scour unstructured materials such as text documents in large reserves of data that are usually stored on computers. IR is related to the representation, storage, and organization of information items, as well as to access. In IR one of the main problems is to determine which documents are relevant and which are not to the user's needs. Under the current regime, users cannot precisely construct queries in an accurate way to retrieve particular pieces of data from large reserves of data. Basic information retrieval systems are producing low-quality search results. In our proposed system for this paper we present a new technique to refine Information Retrieval searches to better represent the user's information need in order to enhance the performance of information retrieval by using different query expansion techniques and apply a linear combinations between them, where the combinations was linearly between two expansion results at one time. Query expansions expand the search query, for example, by finding synonyms and reweighting original terms. They provide significantly more focused, particularized search results than do basic search queries. RESULTS: The retrieval performance is measured by some variants of MAP (Mean Average Precision) and according to our experimental results, the combination of best results of query expansion is enhanced the retrieved documents and outperforms our baseline by 21.06 %, even it outperforms a previous study by 7.12 %. CONCLUSIONS: We propose several query expansion techniques and their combinations (linearly) to make user queries more cognizable to search engines and to produce higher-quality search results.

BACKGROUND: Recently, high-throughput experimental techniques have generated a large amount of protein-protein interaction (PPI) data which can construct large complex PPI networks for numerous organisms. System biology attempts to understand cellular organization and function by analyzing these PPI networks. However, most studies still focus on static PPI networks which neglect the dynamic information of PPI. RESULTS: The gene expression data under different time points and conditions can reveal the dynamic information of proteins. In this study, we used an active probability-based method to distinguish the active level of proteins at different active time points. We constructed dynamic probabilistic protein networks (DPPN) to integrate dynamic information of protein into static PPI networks. Based on DPPN, we subsequently proposed a novel method to identify protein complexes, which could effectively exploit topological structure as well as dynamic information of DPPN. We used three different yeast PPI datasets and gene expression data to construct three DPPNs. When applied to three DPPNs, many well-characterized protein complexes were accurately identified by this method. CONCLUSION: The shift from static PPI networks to dynamic PPI networks is essential to accurately identify protein complex. This method not only can be applied to identify protein complex, but also establish a framework to integrate dynamic information into static networks for other applications, such as pathway analysis.

BACKGROUND: Accurate determination of protein complexes has become a key task of system biology for revealing cellular organization and function. Up to now, the protein complex prediction methods are mostly focused on static protein protein interaction (PPI) networks. However, cellular systems are highly dynamic and responsive to cues from the environment. The shift from static PPI networks to dynamic PPI networks is essential to accurately predict protein complex. RESULTS: The gene expression data contains crucial dynamic information of proteins and PPIs, along with high-throughput experimental PPI data, are valuable for protein complex prediction. Firstly, we exploit gene expression data to calculate the active time point and the active probability of each protein and PPI. The dynamic active information is integrated into high-throughput PPI data to construct dynamic PPI networks. Secondly, a novel method for predicting protein complexes from the dynamic PPI networks is proposed based on core-attachment structural feature. Our method can effectively exploit not only the dynamic active information but also the topology structure information based on the dynamic PPI networks. CONCLUSIONS: We construct four dynamic PPI networks, and accurately predict many well-characterized protein complexes. The experimental results show that (i) the dynamic active information significantly improves the performance of protein complex prediction; (ii) our method can effectively make good use of both the dynamic active information and the topology structure information of dynamic PPI networks to achieve state-of-the-art protein complex prediction capabilities.

Pseudo relevance feedback, as an effective query expansion method, can significantly improve information retrieval performance. However, the method may negatively impact the retrieval performance when some irrelevant terms are used in the expanded query. Therefore, it is necessary to refine the expansion terms. Learning to rank methods have proven effective in information retrieval to solve ranking problems by ranking the most relevant documents at the top of the returned list, but few attempts have been made to employ learning to rank methods for term refinement in pseudo relevance feedback. This article proposes a novel framework to explore the feasibility of using learning to rank to optimize pseudo relevance feedback by means of reranking the candidate expansion terms. We investigate some learning approaches to choose the candidate terms and introduce some state‐of‐the‐art learning to rank methods to refine the expansion terms. In addition, we propose two term labeling strategies and examine the usefulness of various term features to optimize the framework. Experimental results with three TREC collections show that our framework can effectively improve retrieval performance.

Biomedical events can reveal crucial processes in biomedical research. As an important step in biomedical event extraction, biomedical event trigger detection has become a research hotspot. Traditional machine learning methods, which aim to manually design powerful features fed to the classifiers, greatly depend on the understanding of the specific task. In this paper, we propose an approach to automatically learn good features from raw input without manual intervention. The approach is based on dependency-based word embedding and first learns dependency-based word embedding from all available PubMed abstracts. The word embedding contains rich functional and semantic information. Then neural network architecture is used to learn better feature representation based on raw dependency-based word embedding. Meanwhile, we dynamically adjust the embedding while training for adapting to the trigger classification task. Finally, softmax classifier labels the examples by specific trigger class using the features learned by the model. The experimental results show that our approach achieves a micro F1 score of 78.27% and a macro F1 score of 76.94% in significant trigger classes, and performs better than baseline methods. In addition, we can achieve the semantic distributed representation of every trigger word.

Research articles in biomedicine domain have increased exponentially, which makes it more and more difficult for biologists to manually capture all the information they need. Information retrieval technologies can help to obtain the users' needed information automatically. However, it is a great challenge to apply these technologies to biomedicine domain directly because of some domain specific characteristics, such as the abundance of terminologies. To enhance the effectiveness of the biomedical information retrieval, we propose a novel framework based on the state-of-the-art information retrieval methods, called learning to rank, which has been proved effective to rank documents based on their relevance degree. In the framework, we attempt to tackle the problem of the abundance of terminologies by constructing ranking models, which focus on not only retrieving the most relevant documents but also diversifying the searching results to increase the completeness of the resulting list for a given query. In the model training, we propose two novel document labeling strategies, and combine several traditional retrieval models as learning features. Besides, we also investigate the usefulness of different learning to rank approaches in our framework. Experimental results on TREC Genomics datasets demonstrate our proposed framework is effective in improving the performance of biomedical information retrieval.

In these years, the adverse drug reactions (ADRs) have seriously impacted the people's health, and adverse drug event reporting systems become a key means to monitor the drug safety, in which healthcare professionals or drug consumers can submit the adverse drug event reports based on their experience or professional knowledge. However, with the increase of drugs, the number of the submitted reports increases rapidly, making it more and more difficult to capture all the ADRs manually. To tackle the problem, we develop a novel system to compute the similarities among the drugs and adverse reactions automatically from the reports. In the method, we represent the mentions of drugs and adverse reactions as distributed vectors using the skip-gram model, and discover the most potential adverse drug reactions based on the similarities.

This paper presents a deep neural network-based protein-protein interactions (PPIs) information extraction approach which can learn complex and abstract features automatically from unlabeled data by unsupervised representation learning methods. This approach first employs the training algorithm of auto-encoders to initialize the parameters of a deep multilayer neural network. Then the gradient descent method using back-propagation is applied to train this deep multilayer neural network model. Experimental results on five public PPI corpora show that our method can achieve better performance than can a multilayer neural network. In addition, the performance comparison with APG also verifies the effectiveness of our method.

Cross-domain sentiment analysis focuses on these problems where the source domain and the target domain are from different domains. However, traditional sentiment classification approaches usually perform poorly to address cross-domain problems. So, this paper proposed a cross-domain sentiment classification method based on extraction of key sentiment sentence. Firstly, based on the observation that not every part of the document is equally informative for inferring the sentiment orientation of the whole document, the concept of key sentiment sentence was defined. Secondly, taking advantage of three properties: sentiment purity, keyword property and position property, we construct heuristic rules, and combine with machine learning to extract key sentiment sentence. Then, data is divided into key and detail views. Integrating two views effectively can improve performance. Finally, experimental results show the superiority of our proposed method.

BACKGROUND: Revealing protein complexes are important for understanding principles of cellular organization and function. High-throughput experimental techniques have produced a large amount of protein interactions, which makes it possible to predict protein complexes from protein-protein interaction (PPI) networks. However, the small amount of known physical interactions may limit protein complex detection. METHODS: The new PPI networks are constructed by integrating PPI datasets with the large and readily available PPI data from biomedical literature, and then the less reliable PPI between two proteins are filtered out based on semantic similarity and topological similarity of the two proteins. Finally, the supervised learning protein complex detection (SLPC), which can make full use of the information of available known complexes, is applied to detect protein complex on the new PPI networks. RESULTS: The experimental results of SLPC on two different categories yeast PPI networks demonstrate effectiveness of the approach: compared with the original PPI networks, the best average improvements of 4.76, 6.81 and 15.75 percentage units in the F-score, accuracy and maximum matching ratio (MMR) are achieved respectively; compared with the denoising PPI networks, the best average improvements of 3.91, 4.61 and 12.10 percentage units in the F-score, accuracy and MMR are achieved respectively; compared with ClusterONE, the start-of the-art complex detection method, on the denoising extended PPI networks, the average improvements of 26.02 and 22.40 percentage units in the F-score and MMR are achieved respectively. CONCLUSIONS: The experimental results show that the performances of SLPC have a large improvement through integration of new receivable PPI data from biomedical literature into original PPI networks and denoising PPI networks. In addition, our protein complexes detection method can achieve better performance than ClusterONE.

In this paper, we analyze six protein interaction networks widely used for protein complex detection, and compare the performance of six classic computational methods on them in order to find the relations between network characteristics and the performances of these complex detection methods. Furthermore, we explore the difference among the two complexes detected by different methods and the real complex by using a visualization approach which can easily find the difference of two methods and locate the undetected proteins.

In this paper, we present an approach to automatically construct disease related knowledge summarization from biomedical literature. In this approach, first Kullback-Leibler divergence combined with mutual information metric is used to extract disease salient information. Then deep search based on depth first search (DFS) is applied to find hidden relations between biomedical entities. Finally random walk algorithm is exploited to filter out the weak relations. The experimental results show that our approach achieves a precision of 60% and a recall of 61% on salient information extraction, and outperforms the method of Combo. In addition, the method of deep search obtains more hidden relations than the original correlation extraction methods.

The rapidly growing biomedical literature provides a significantly large and readily available source of PPI data. In this paper, we present supervised learning and data integration based complex detection approach. In this approach, a sophisticated natural language processing system, PPIExtractor, is employed to extract new PPI interactions from biomedical literature which are then integrated into original PPI networks. Then a supervised learning model, built by via of the information of available known complexes, is used in the multiple complex detection stages, e.g. the cliques filtering, growth, and candidate complex filtering. The experimental results on three yeast PPI networks demonstrate the effectiveness of our approach.

Ranking plays an important role in information retrieval system. In recent years, a kind of research named 'learning to rank' becomes more and more popular, which applies machine learning technology to solve ranking problems. Lots of ranking models belonged to learning to rank have been proposed, such as Regression, RankNet, and ListNet. Inspired by this, we proposed a novel learning to rank algorithm named GPQ in this paper, in which genetic programming was employed to directly optimize Q-measure evaluation metric. Experimental results on OHSUMED benchmark dataset indicated that our method GPQ could be competitive with Ranking SVM, SVMMAP and ListNet, and improve the ranking accuracies.

BACKGROUND: The amount of biomedical literature available is growing at an explosive speed, but a large amount of useful information remains undiscovered in it. Researchers can make informed biomedical hypotheses through mining this literature. Unfortunately, popular mining methods based on co-occurrence produce too many target concepts, leading to the declining relevance ranking of the potential target concepts. METHODS: This paper presents a new method for selecting linking concepts which exploits statistical and textual features to represent each linking concept, and then classifies them as relevant or irrelevant to the starting concepts. Relevant linking concepts are then used to discover target concepts. RESULTS: Through an evaluation it is observed textual features improve the results obtained with only statistical features. We successfully replicate Swanson's two classic discoveries and find the rankings of potentially relevant target concepts are relatively high. CONCLUSIONS: The number of target concepts is greatly reduced and potentially relevant target concepts gain higher ranking by adopting only relevant linking concepts. Thus, the proposed method has the potential to help biomedical experts find the most useful and valuable target concepts effectively.

BACKGROUND: In recent years, the biomedical literature has been growing rapidly. These articles provide a large amount of information about proteins, genes and their interactions. Reading such a huge amount of literature is a tedious task for researchers to gain knowledge about a gene. As a result, it is significant for biomedical researchers to have a quick understanding of the query concept by integrating its relevant resources. METHODS: In the task of gene summary generation, we regard automatic summary as a ranking problem and apply the method of learning to rank to automatically solve this problem. This paper uses three features as a basis for sentence selection: gene ontology relevance, topic relevance and TextRank. From there, we obtain the feature weight vector using the learning to rank algorithm and predict the scores of candidate summary sentences and obtain top sentences to generate the summary. RESULTS: ROUGE (a toolkit for summarization of automatic evaluation) was used to evaluate the summarization result and the experimental results showed that our method outperforms the baseline techniques. CONCLUSIONS: According to the experimental result, the combination of three features can improve the performance of summary. The application of learning to rank can facilitate the further expansion of features for measuring the significance of sentences.

Given that the amount of biomedical literature available is growing at an explosive speed and people's ability to read literature is limited, the development of a computer-assisted approach for mining central but hidden knowledge from extensive literature is of utmost importance.

This paper presents a learning to rank based gene summarization method using features such as gene ontology relevance, topic relevance and TextRank.

In this paper, we present a supervised learning-based method for predicting protein complexes in protein interaction network. The method extracts rich features from protein interaction network to train a Regression model, which is then used for the cliques filtering, growth, and candidate complex filtering. The experimental results on several protein interaction networks show that our method outperforms other state-of-the-art protein complex detection methods.

Prediction of protein complexes from protein-protein interaction (PPI) networks is crucial to unraveling the principles of cellular organization. Most existing approaches only exploit high-throughput experimental PPI data to predict protein complexes. In this paper, we integrate the multiple biomedical resources for protein complex prediction by constructing attributed PPI networks, which include high-throughput data, co-expression data, genomic data, text mining data and gene ontology data. Multiple biomedical resources are complementary in attributed PPI networks. We propose a novel approach called IMBP based on attributed PPI networks. IMBP can effectively learn the degree of contributions of different biomedical resource for complex prediction. The experimental results show that IMBP can make good use of multiple biomedical data and achieve state-of-the-art performance.

The central issue in language model estimation is smoothing, which is a technique for avoiding zero probability estimation problem and overcoming data sparsity. There are three representative smoothing methods: Jelinek‐Mercer (JM) method; Bayesian smoothing using Dirichlet priors (Dir) method; and absolute discounting (Dis) method, whose parameters are usually estimated empirically. Previous research in information retrieval (IR) on smoothing parameter estimation tends to select a single value from optional values for the collection, but it may not be appropriate for all the queries. The effectiveness of all the optional values should be considered to improve the ranking performance. Recently, learning to rank has become an effective approach to optimize the ranking accuracy by merging the existing retrieval methods. In this article, the smoothing methods for language modeling in information retrieval (LMIR) with different parameters are treated as different retrieval methods, then a learning to rank approach to learn a ranking model based on the features extracted by smoothing methods is presented. In the process of learning, the effectiveness of all the optional smoothing parameters is taken into account for all queries. The experimental results on the Learning to Rank for Information Retrieval (LETOR) LETOR3.0 and LETOR4.0 data sets show that our approach is effective in improving the performance of LMIR.

With the development of Web2.0, social tagging systems in which users can freely choose tags to annotate resources according to their interests have attracted much attention. In particular, literature on the emergence of collective intelligence in social tagging systems has increased. In this article, we propose a probabilistic generative model to detect latent topical communities among users. Social tags and resource contents are leveraged to model user interest in two similar and correlated ways. Our primary goal is to capture user tagging behavior and interest and discover the emergent topical community structure. The communities should be groups of users with frequent social interactions as well as similar topical interests, which would have important research implications for personalized information services. Experimental results on two real social tagging data sets with different genres have shown that the proposed generative model more accurately models user interest and detects high‐quality and meaningful topical communities.

Protein complexes are important for unraveling the secrets of cellular organization and function. Many computational approaches have been developed to predict protein complexes in protein-protein interaction (PPI) networks. However, most existing approaches focus mainly on the topological structure of PPI networks, and largely ignore the gene ontology (GO) annotation information. In this paper, we constructed ontology attributed PPI networks with PPI data and GO resource. After constructing ontology attributed networks, we proposed a novel approach called CSO (clustering based on network structure and ontology attribute similarity). Structural information and GO attribute information are complementary in ontology attributed networks. CSO can effectively take advantage of the correlation between frequent GO annotation sets and the dense subgraph for protein complex prediction. Our proposed CSO approach was applied to four different yeast PPI data sets and predicted many well-known protein complexes. The experimental results showed that CSO was valuable in predicting protein complexes and achieved state-of-the-art performance.

Knowledge about protein-protein interactions (PPIs) unveils the molecular mechanisms of biological processes. In this paper, we present a PPI extraction system, termed PPIExtractor, which automatically extracts PPIs from biomedical text and visualizes them. Given a Medline record dataset, PPIExtractor first applies Feature Coupling Generalization (FCG) to tag protein names, next uses the extended semantic similarity-based method to normalize them, then combines feature-based, convolution tree and graph kernels to extract PPIs, and finally visualizes the PPI network. Experimental evaluations show that PPIExtractor can achieve state-of-the-art performance on a DIP subset with respect to comparable evaluations. PPIExtractor is freely available for academic purposes at: http://202.118.75.18:8080/PPIExtractor/.

Hidden Markov models (HMM) have been widely used in natural language processing (NLP), especially in syntactic level applications, which appears naturally as short-range-dependent sequence recognition problems. But the structure of HMM limits the usage of global knowledge including the sentiment analysis of the text, which has become an increasingly popular research topic in NLP now. In this paper, we propose a novel treatment of HMM model to use the result of sentimental subjectivity analysis in syntactic level task, i.e. part-of-speech (POS) tagging. The subjectivity information is introduced as a pre-classification procedure into the interval-type HMM. The subjectivity degree of the testing sentence is used as a combination factor to choose an appropriate value from the interval. Experiments results on public tagging data sets shows that the proposed approach enhanced the performance of POS tagging.

Wikipedia is characterized by its dense link structure and a large number of articles in different languages, which make it a notableWeb corpus for knowledge extraction and mining, in particular for mining the multilingual associations. In this paper, motivated by a psychological theory of word meaning, we propose a graph‐based approach to constructing a cross‐language association dictionary (CLAD) fromWikipedia, which can be used in a variety of cross‐language accessing and processing applications. In order to evaluate the quality of the minedCLAD, and to demonstrate how the minedCLADcan be used in practice, we explore two different applications of the minedCLADto cross‐language information retrieval (CLIR). First, we use the minedCLADto conduct cross‐language query expansion; and, second, we use it to filter out translation candidates with low translation probabilities. Experimental results on a variety of standardCLIRtest collections show that theCLIRretrieval performance can be substantially improved with the above two applications ofCLAD, which indicates that the minedCLADis of sound quality.

Extracting protein-protein interaction (PPI) from biomedical literature is an important task in biomedical text mining (BioTM). In this paper, we propose a hash subgraph pairwise (HSP) kernel-based approach for this task. The key to the novel kernel is to use the hierarchical hash labels to express the structural information of subgraphs in a linear time. We apply the graph kernel to compute dependency graphs representing the sentence structure for protein-protein interaction extraction task, which can efficiently make use of full graph structural information, and particularly capture the contiguous topological and label information ignored before. We evaluate the proposed approach on five publicly available PPI corpora. The experimental results show that our approach significantly outperforms all-path kernel approach on all five corpora and achieves state-of-the-art performance.

Protein complexes play a key role in many biological processes. Various computational approaches have been developed to identify complexes from protein-protein interaction (PPI) networks. However, high false-positive rate of PPIs makes the identification challenging. In this paper, we propose a protein semantic similarity measure based on the ontology structure of Gene Ontology (GO) terms and GO annotations to estimate the reliability of interactions in PPI networks. Interaction pairs with low GO semantic similarity are removed from the network as unreliable interactions. Then, a cluster-expanding algorithm is applied to identify complexes with core-attachment structure on the filtered network. We have applied our method on three different yeast PPI networks. The effectiveness of our method is examined on two benchmark complex datasets. Experimental results show that our method outperforms other state-of-the-art approaches in most evaluation metrics. Removing interactions with low similarity significantly improves the performance of complex identification.

Learning to rank algorithms are usually grouped into three types: the point wise approach, the pairwise approach, and the listwise approach, according to the input spaces. Much of the prior work is based on the three approaches to learn the ranking model to predict the relevance of a document to a query. In this paper, we focus on the problem of constructing new input space based on groups of documents with the same relevance judgment. A novel approach is proposed based on cross entropy to improve the existing ranking method. The experimental results show that our approach leads to significant improvements in retrieval effectiveness.

It is commonly believed that query logs from Web search are a gold mine for search business, because they reflect users' preference over Web pages presented by search engines, so a lot of studies based on query logs have been carried out in the last few years. In this study, we assume that two queries are relevant to each other when they have same clicked page in their result lists, and we also consider the queries' topics of user's need. Thus, we propose a Two-Stage SimRank (called TSS in this paper) algorithm based on SimRank and some clustering algorithms to compute the similarity among queries, and then use it to discover relevant terms for query expansion, considering the information of topics and the global relationships of queries concurrently, with a query log collected by a practical search engine. Experimental results on two TREC test collections show that our approach can discover qualified terms effectively and improve retrieval performance.

Automatic query expansion technologies have been proven to be effective in many information retrieval tasks. Most existing approaches are based on the assumption that the most informative terms in top-retrieved documents can be viewed as context of the query and thus can be used for query expansion. One problem with these approaches is that some of the expansion terms extracted from feedback documents are irrelevant to the query, and thus may hurt the retrieval performance. In social annotations, users provide different keywords describing the respective Web pages from various aspects. These features may be used to boost IR performance. However, to date, the potential of social annotation for this task has been largely unexplored. In this paper, we explore the possibility and potential of social annotation as a new resource for extracting useful expansion terms. In particular, we propose a term ranking approach based on social annotation resource. The proposed approach consists of two phases: (1) in the first phase, we propose a term-dependency method to choose the most likely expansion terms; (2) in the second phase, we develop a machine learning method for term ranking, which is learnt from the statistics of the candidate expansion terms, using ListNet. Experimental results on three TREC test collections show that the retrieval performance can be improved when the term ranking method is used. In addition, we also demonstrate that terms selected by the term-dependency method from social annotation resources are beneficial to improve the retrieval performance.

In this paper, we use query-level regression as the loss function. The regression loss function has been used in pointwise methods, however pointwise methods ignore the query boundaries and treat the data equally across queries, and thus the effectiveness is limited. We show that regression is an effective loss function for learning to rank when used in query-level. We use neural network to model the ranking function and gradient descent for optimization and refer our method as ListReg. Experimental results show that ListReg significantly outperforms pointwise Regression and the state-of-the-art listwise method in most cases.

Abstract Pseudo‐relevance feedback (PRF) via query expansion (QE) assumes that the top‐ranked documents from the first‐pass retrieval are relevant. The most informative terms in the pseudo‐relevant feedback documents are then used to update the original query representation in order to boost the retrieval performance. Most current PRF approaches estimate the importance of the candidate expansion terms based on their statistics on document level. However, a document for PRF may consist of different topics, which may not be all related to the query even if the document is judged relevant. The main argument of this article is the proposal to conduct PRF on a granularity smaller than on the document level. In this article, we propose a topic‐based feedback model with three different strategies for finding a good query‐related topic based on the Latent Dirichlet Allocation model. The experimental results on four representative TREC collections show that QE based on the derived topic achieves statistically significant improvements over a strong feedback model in the language modeling framework, which updates the query representation based on the top‐ranked documents.

Feature representation is essential to machine learning and text mining. In this paper, we present a feature coupling generalization (FCG) framework for generating new features from unlabeled data. It selects two special types of features, i.e., example-distinguishing features (EDFs) and class-distinguishing features (CDFs) from original feature set, and then generalizes EDFs into higher-level features based on their coupling degrees with CDFs in unlabeled data. The advantage is: EDFs with extreme sparsity in labeled data can be enriched by their co-occurrences with CDFs in unlabeled data so that the performance of these low-frequency features can be greatly boosted and new information from unlabeled can be incorporated. We apply this approach to three tasks in biomedical literature mining: gene named entity recognition (NER), protein-protein interaction extraction (PPIE), and text classification (TC) for gene ontology (GO) annotation. New features are generated from over 20 GB unlabeled PubMed abstracts. The experimental results on BioCreative 2, AIMED corpus, and TREC 2005 Genomics Track show that 1) FCG can utilize well the sparse features ignored by supervised learning. 2) It improves the performance of supervised baselines by 7.8 percent, 5.0 percent, and 5.8 percent, respectively, in the tree tasks. 3) Our methods achieve 89.1, 64.5 F-score, and 60.1 normalized utility on the three benchmark data sets.

Protein complexes can be identified from the protein interaction networks derived from experimental data sets. However, these analyses are challenging because of the presence of unreliable interactions and the complex connectivity of the network. The integration of protein-protein interactions with the data from other sources can be leveraged for improving the effectiveness of protein complexes detection algorithms. We have developed novel semantic similarity method, which use Gene Ontology (GO) annotations to measure the reliability of protein-protein interactions. The protein interaction networks can be converted into a weighted graph representation by assigning the reliability values to each interaction as a weight. Following the approach of previously proposed clustering algorithm IPCA which expands clusters starting from seeded vertices, we present a clustering algorithm OIIP based on the new weighted Protein-Protein interaction networks for identifying protein complexes. The algorithm OIIP is applied to the protein interaction network of Sacchromyces cerevisiae and identifies many well known complexes. Experimental results show that the algorithm OIIP has higher F-measure and accuracy compared to other competing approaches.

Knowledge about protein-protein interactions unveils the molecular mechanisms of biological processes. This paper presents a multiple kernels learning-based approach to automatically extracting protein-protein interactions from biomedical literature. Experimental evaluations show that our approach can achieve state-of-the-art performance with respect to comparable evaluations, with 64.88% F-score and 88.02% area under the receiver operating characteristics curve (AUC) on the AImed corpus.

An essential issue in document retrieval is ranking, and the documents are ranked by their expected relevance to a given query. Multiple labels are used to represent different level of relevance for documents to a given query, and the corresponding label values are used to quantify the relevance of the documents. According to the training set for a given query, the documents can be divided into several groups. Specifically, the documents with the same label are assigned to the same group. If the documents in the group with higher relevance label can always be ranked higher over the ones in groups with lower relevance label by a ranking model, it is reasonable to expect perfect ranking performance. Inspired by this idea, we propose a novel framework for learning to rank, which depends on two new samples. The first one is one-group constituted by one document with higher level label and a group of documents with lower level label; the second one is group-group constituted by a group of documents with higher level label and a group of documents with lower level label. A novel loss function is proposed based on the likelihood loss similar to ListMLE. We demonstrate the advantages of our approaches on the Letor 3.0 data set. Experimental results show that our approaches are effective in improving the ranking performance.

BACKGROUND: Extracting protein-protein interactions from biomedical literature is an important task in biomedical text mining. Supervised machine learning methods have been used with great success in this task but they tend to suffer from data sparseness because of their restriction to obtain knowledge from limited amount of labelled data. In this work, we study the use of unlabeled biomedical texts to enhance the performance of supervised learning for this task. We use feature coupling generalization (FCG) - a recently proposed semi-supervised learning strategy - to learn an enriched representation of local contexts in sentences from 47 million unlabeled examples and investigate the performance of the new features on AIMED corpus. RESULTS: The new features generated by FCG achieve a 60.1 F-score and produce significant improvement over supervised baselines. The experimental analysis shows that FCG can utilize well the sparse features which have little effect in supervised learning. The new features perform better in non-linear classifiers than linear ones. We combine the new features with local lexical features, obtaining an F-score of 63.5 on AIMED corpus, which is comparable with the current state-of-the-art results. We also find that simple Boolean lexical features derived only from local contexts are able to achieve competitive results against most syntactic feature/kernel based methods. CONCLUSIONS: FCG creates a lot of opportunities for designing new features, since a lot of sparse features ignored by supervised learning can be utilized well. Interestingly, our results also demonstrate that the state-of-the art performance can be achieved without using any syntactic information in this task.

We present the application of a recently proposed semi-supervised learning strategy - feature coupling generalization (FCG) - in the task of protein-protein interaction extraction from biomedical literatures. FCG is a framework that generates new features from relatedness of two special types of old features: example-distinguishing features (EDFs) and class-distinguishing features (CDFs). Their relatedness estimated from unlabeled data tends to capture indicative information not available in labeled data. For this task, we designed several EDFs and CDFs derived from the text patterns surrounding the co-occurrence proteins, and combined the new features generated by FCG with Boolean lexical features. The experimental results on AIMED corpus show that the new features yield significant improvement over a strong baseline, and the combined method achieves state-of-the-art performance without using any syntactic information.

In this paper, we propose a graph-based approach to constructing a multilingual association dictionary from Wikipedia, in which we exploit two kinds of links in Wikipedia articles to associate multilingual words and concepts together in a graph. The mined association dictionary is applied in cross language information retrieval (CLIR) to verify its quality. We evaluate our approach on four CLIR data sets and the experimental results show that it is possible to mine a good multilingual association dictionary from Wikipedia articles.

BACKGROUND: Gene named entity classification and recognition are crucial preliminary steps of text mining in biomedical literature. Machine learning based methods have been used in this area with great success. In most state-of-the-art systems, elaborately designed lexical features, such as words, n-grams, and morphology patterns, have played a central part. However, this type of feature tends to cause extreme sparseness in feature space. As a result, out-of-vocabulary (OOV) terms in the training data are not modeled well due to lack of information. RESULTS: We propose a general framework for gene named entity representation, called feature coupling generalization (FCG). The basic idea is to generate higher level features using term frequency and co-occurrence information of highly indicative features in huge amount of unlabeled data. We examine its performance in a named entity classification task, which is designed to remove non-gene entries in a large dictionary derived from online resources. The results show that new features generated by FCG outperform lexical features by 5.97 F-score and 10.85 for OOV terms. Also in this framework each extension yields significant improvements and the sparse lexical features can be transformed into both a lower dimensional and more informative representation. A forward maximum match method based on the refined dictionary produces an F-score of 86.2 on BioCreative 2 GM test set. Then we combined the dictionary with a conditional random field (CRF) based gene mention tagger, achieving an F-score of 89.05, which improves the performance of the CRF-based tagger by 4.46 with little impact on the efficiency of the recognition system. A demo of the NER system is available at http://202.118.75.18:8080/bioner.