Large language models (LLMs) have become indispensable, but the most celebrated efficiency methods---mixture-of-experts (MoE), speculative decoding, and complex retrieval-augmented generation (RAG)---were built for hyperscale providers with vast infrastructure and elite teams. Outside that context, their benefits collapse into overhead, fragility, and wasted carbon. The result is that a handful of Big Tech companies benefit, while thousands of hospitals, schools, governments, and enterprises are left without viable options. We argue that the next frontier is not greater sophistication at scale, but robust simplicity: efficiency that thrives under modest resources and minimal expertise. We propose a new research agenda: retrofitting pretrained models with more efficient architectures without retraining, inventing lightweight fine-tuning that preserves alignment, making reasoning economical despite long chains of thought, enabling dynamic knowledge management without heavy RAG pipelines, and adopting Overhead-Aware Efficiency (OAE) as a standard benchmark. By redefining efficiency to include adoption cost, sustainability, and fairness, we can democratize LLM deployment---ensuring that optimization reduces inequality and carbon waste rather than amplifying them.

While annotated corpora are crucial in the field of natural language processing (NLP), those containing copyrighted material are difficult to exchange among researchers.Yet, such corpora are necessary to fully represent the diversity of data found in the wild in the context of NLP tasks.We tackle this issue by proposing a method to lawfully and publicly share the annotations of copyrighted literary texts.The corpus creator shares the annotations in clear, along with a non-reversible hashed version of the source material.The corpus user must own the source material, and apply the same hash function to their own tokens, in order to match them to the shared annotations.Crucially, our method is robust to reasonable divergences in the version of the copyrighted data owned by the user.As an illustration, we present alignment experiments on different editions of novels.Our results show that our method is able to correctly align 98.7 to 99.79% of tokens depending on the novel, provided the user version is sufficiently close to the corpus creator's version.We publicly release novelshare, a Python implementation of our method.

This paper presents a practical problem in dialogue systems: the capability to adapt to changing user intentions and resolve inconsistencies in conversation histories. It is crucial in scenarios like train ticket booking, where travel plans often change dynamically. Notwithstanding the advancements in NLP and large language models (LLMs), these systems struggle with real-time information updates during conversations. We introduce a specialized dataset to evaluate LLM-based chatbots on such conversational adaptability by asking a broad range of open-domain questions, focusing on scenarios where users modify their requests mid-conversation. Additionally, as LLMs are susceptible to generating superfluous sentences, we propose a novel, Chain-of-Thought-free evaluation framework to distill the user intention from their responses. Through extensive investigations on four LLMs, we observe that these contemporary LLMs are not well-aligned with the latest user intent in long-term conversations; they often fail to capture the nuances of natural conversations in a zero-shot setting. Interestingly, the results demonstrate that GPT-4, widely recognized as having the most advanced reasoning capabilities to date, is bested by GPT-3.5 in this task. This work aims to improve the practicality of LLM-based chatbots, bridging the gap between the current capabilities of dialogue systems and the fluidity of human interactions.

Abstraction--the ability to recognize and distill essential computational patterns from complex problem statements--is a foundational skill in computer science, critical both for human problem-solvers and coding-oriented large language models (LLMs). Despite recent advances in training LLMs for code generation using reinforcement learning (RL), most existing approaches focus primarily on superficial pattern recognition, overlooking explicit training for abstraction. In this study, we propose AR2 (Adversarial Reinforcement Learning for Abstract Reasoning), a novel framework explicitly designed to enhance the abstraction abilities of LLMs. AR2 employs a teacher model to transform kernel problems into narrative-rich, challenging descriptions without changing their fundamental logic. Simultaneously, a student coding model is trained to solve these complex narrative problems by extracting their underlying computational kernels. Experimental results demonstrate that AR2 substantially improves the student model's accuracy on previously unseen, challenging programming tasks, underscoring abstraction as a key skill for enhancing LLM generalization.

Transformer-based Optical Character Recognition (OCR) models perform well on common tasks but struggle to generalize to low-resource settings, handwritten text, or diverse scripts such as Jawi. Traditional adaptation methods often require labeled target data or fall short under severe domain shifts. We tackle this challenge with Unsupervised Multi-Source Heterogeneous Domain Adaptation (UMSHDA) for Handwritten Text Recognition (HTR). We adapt Domain-Adversarial Neural Networks (DANN) and Prototype Learning for Adversarial Domain Adaptation (PLADA) within a Transformer-based OCR architecture. We introduce two novel multi-adversarial adaptation strategies: (1) an Additive strategy that jointly optimizes a domain classification loss and a prototype-based alignment loss, and (2) an Integrative strategy that uses a prototype-driven adversarial signal to augment the domain classifier with semantic constraints. Experimental results show that our methods significantly outperform non-adaptive and DANN baselines on challenging handwriting tasks. For Jawi handwriting, our models achieve a relative error reduction of up to 22%. We also demonstrate superior cross-script generalization to German handwriting, achieving near-perfect performance, with a CER reduced by more than 50% relative to the DANN baseline. This work demonstrates that tailored multi-adversarial domain adaptation can effectively bridge significant domain gaps, enabling robust recognition accuracy for complex, low-resource HTR.

We introduce MESAQA, a novel dataset focusing on multi-span contextual understanding question answering (QA).Unlike traditional single-span QA systems, questions in our dataset consider information from multiple spans within the context document. MESAQA supports evidence-grounded QA, demanding the model’s capability of answer generation and multi-evidence identification. Our automated dataset creation method leverages the MASH-QA dataset and large language models (LLMs) to ensure that each Q/A pair requires considering all selected spans. Experimental results show that current models struggle with multi-span contextual QA, underscoring the need for new approaches. Our dataset sets a benchmark for this emerging QA paradigm, promoting research in complex information retrieval and synthesis.

This work presents a novel trie (prefix-tree)based parallel decoding method that addresses the memory inefficiency of batch-based beam search.By sharing a single KV cache across beams with common prefixes, our approach dramatically reduces memory usage and enables efficient decoding.We evaluated our method across three attention architectures, Multi-Head Attention (Phi-3.5-miniinstruct),Grouped Query Attention (Llama-3.1-8B-Instruct),and Sliding Window Attention (Mistral-Small-24B-Instruct-2501), using CN-N/DailyMail for abstractive summarization and HumanEval for code generation.Our experiments demonstrate substantial memory savings (4-8×) and up to 2.4× faster decoding, without compromising generation quality.These results highlight our method's suitability for memory-constrained environments and largescale deployments.

Sycophancy causes models to produce answers that cater to user expectations rather than providing truthful responses.Sycophantic behavior in models can erode user trust by creating a perception of dishonesty or bias.This lack of authenticity may lead users to question the reliability and objectivity of the system's responses.Although Reinforcement Learning from Human Feedback (RLHF) is effective in aligning models with human preferences, previous studies have observed that it can simultaneously amplify sycophantic behavior.However, these studies primarily focused on proprietary models and employed indirect analysis to demonstrate the influence of human feedback.Our study focuses on sycophancy in open-source models, which are more reproducible and transparent for research.We investigated the impact of human feedback on sycophancy by directly comparing models aligned with human feedback to those not aligned.To address sycophancy, we proposed assessing the user's expected answer rather than ignoring it.Consequently, we developed the Sycophancy Answer Assessment (SAA) dataset 1 and introduced Self-Augmented Preference Alignment, demonstrating that these methods effectively enhance the model's assessment ability and significantly reduce sycophancy across tasks.

Stable Diffusion models have made remarkable strides in generating photorealistic images from text prompts but often falter when tasked with accurately representing complex spatial arrangements, particularly involving intricate 3D relationships. To address this limitation, we introduce SmartSpatial, an innovative approach that not only enhances the spatial arrangement capabilities of Stable Diffusion but also fosters AI-assisted creative workflows through 3D-aware conditioning and attention-guided mechanisms. SmartSpatial incorporates depth information injection and cross-attention control to ensure precise object placement, delivering notable improvements in spatial accuracy metrics. In conjunction with SmartSpatial, we present SmartSpatialEval, a comprehensive evaluation framework that bridges computational spatial accuracy with qualitative artistic assessments. Experimental results show that SmartSpatial significantly outperforms existing methods, setting new benchmarks for spatial fidelity in AI-driven art and creativity.

Vision-language models (VLMs) often struggle with culturally specific content — a challenge largely overlooked by existing benchmarks that focus on dominant languages and globalized datasets. We introduce TᴀɪᴡᴀɴVQA, a VQA benchmark designed for Taiwanese culture to evaluate recognition and reasoning in regional contexts. TᴀɪᴡᴀɴVQA contains 2,736 images and 5,472 manually curated questions covering topics such as traditional foods, public signs, festivals, and landmarks. The official benchmark set includes 1,000 images and 2,000 questions for systematic assessment, with the remainder of the data used as training material. Evaluations on state-of-the-art VLMs reveal strong visual recognition but notable weaknesses in cultural reasoning. To address this, we propose a data augmentation strategy that combines human-annotated and synthesized dialogues to enhance cultural understanding. Fine-tuning yields significant gains on TᴀɪᴡᴀɴVQA while maintaining stable performance on other multimodal tasks. To further explore the models’ cultural understanding, we conducted an open-ended question answering experiment. The results indicate a notable decline in cultural knowledge generation ($\approx$10–20\%), suggesting challenges remain. TᴀɪᴡᴀɴVQA offers a scalable framework for building culturally grounded AI models in low-resource cultures, promoting diversity and fairness in multimodal AI. Our dataset and code are publicly available on [Hugging Face](https://huggingface.co/datasets/hhhuang/TaiwanVQA) and [GitHub](https://github.com/hhhuang/TaiwanVQA).

Texts, printed or written, are the most important form of cultural preservation. This is true for Chinese studies and many cultures. For this reason, optical character recognition (OCR) is one of the most important tools for heritage preservation. With OCR and reading order detection (ROD), we convert the information in their hard-copy forms to digitized documents, allowing both lasting storage and further analysis. Given a page image, the step of ROD helps us determine the order of the texts on the page. An effective ROD reduces the cost of post-OCR human labor for correction. Rule-based and n-gram models are conceivable and popular techniques for implementing ROD, considering the coordinates and the alternative orders of the individually detected characters. In this paper, we report a block-based approach that does not rely on the n-gram techniques for ROD. Instead, we identified the text blocks, and ordered the blocks based on the ideas of learning-to-rank, and concatenated the texts within the blocks by their orders. We evaluated our approach on the MTHv2 dataset. Experimental results show that our approach achieved impressive results, reducing the position error rate (PER) to 5.4%, compared to 61.7% for heuristic methods and 56.5% for MLP, representing a significant improvement of 91.3% and 90.4%, respectively. Additional experiments show that we could have used less data than one might expect to achieve a satisfactory quality of ROD. We implemented and demonstrated a fully functioning prototype at a recent AAAI conference. $$^{*}$$ The actual performance of the prototype provides solid evidence of the effectiveness and extensibility of our design to more general contexts, e.g., Japanese texts.

Chinese historical documents, with their unique layouts and reading patterns, pose significant challenges for traditional Optical Character Recognition (OCR) systems. This paper introduces a tailored OCR system designed to address these complexities, particularly emphasizing the crucial aspect of Reading Order Detection(ROD). Our system operates through a threefold process: text detection using the Differential Binarization++ model, text recognition with the SVTR Net, and a novel ROD approach harnessing raw image features. This innovative method for ROD, inspired by human perception, utilizes visual cues present in raw images to deduce the inherent sequence of ancient texts. Preliminary results show promising reductions in page error rates. By preserving both content and context, our system contributes meaningfully to the accurate and contextual digitization of Chinese historical manuscripts.

People often struggle to remember specific details of past experiences, which can lead to the need to revisit these memories. Consequently, lifelog retrieval has emerged as a crucial application. Various studies have explored methods to facilitate rapid access to personal lifelogs for memory recall assistance. In this paper, we propose a Captioning-Integrated Visual Lifelog (CIVIL) Retrieval System for extracting specific images from a user’s visual lifelog based on textual queries. Unlike traditional embedding-based methods, our system first generates captions for visual lifelogs and then utilizes a text embedding model to project both the captions and user queries into a shared vector space. Visual lifelogs, captured through wearable cameras, provide a first-person viewpoint, necessitating the interpretation of the activities of the individual behind the camera rather than merely describing the scene. To address this, we introduce three distinct approaches: the single caption method, the collective caption method, and the merged caption method, each designed to interpret the life experiences of lifeloggers. Experimental results show that our method effectively describes first-person visual images, enhancing the outcomes of lifelog retrieval. Furthermore, we construct a textual dataset that converts visual lifelogs into captions, thereby reconstructing personal life experiences.

Headline generation, a key task in abstractive summarization, strives to condense a full-length article into a succinct, single line of text. Notably, while contemporary encoder-decoder models excel based on the ROUGE metric, they often falter when it comes to the precise generation of numerals in headlines. We identify the lack of datasets providing fine-grained annotations for accurate numeral generation as a major roadblock. To address this, we introduce a new dataset, the NumHG, and provide over 27,000 annotated numeral-rich news articles for detailed investigation. Further, we evaluate five well-performing models from previous headline-generation tasks using human evaluation in terms of numerical accuracy, reasonableness, and readability. Our study reveals a need for improvement in numerical accuracy, demonstrating the potential of the NumHG dataset to drive progress in number-focused headline generation and stimulate further discussions in numeral-focused text generation.

Large Language Models (LLMs) have demonstrated capability in “instruction induction,” generating instructions from demonstrations (input-output pairs). However, existing methods often rely on large datasets or numerous examples, which is impractical and costly in real-world scenarios. In this work, we propose a low-cost, task-level framework called Induct-Learn. It induces pseudo instructions from a few demonstrations and a short phrase, adding a CoT process into existing demonstrations. When encountering new problems, the learned pseudo instructions and demonstrations with the pseudo CoT process can be combined into a prompt to guide the LLM’s problem-solving process. We validate our approach on the BBH-Induct and Evals-Induct datasets, and the results show that the Induct-Learn framework outperforms state-of-the-art methods. We also exhibit cross-model adaptability and achieve superior performance at a lower cost compared to existing methods.

The popularization of networks fosters the convenience of communication. People can easily share their life experiences and thoughts with relatives and friends via instant messaging software. As time passes, individuals may forget certain details of life events, leading to difficulties in effectively communicating with others. The propensity of individuals to forget or mix up life events highlights the importance of services aimed at retrieving information about past experiences. This paper presents a conversational information recall system, ConvLogRecaller, which proactively supports real-time memory recall assistance during online conversations. Given a conversation of the user with others, ConvLogRecaller suggests a message if the user forgets the details of the life experiences. The services provided by our system can avoid hesitations or memory lapses that might hinder the efficiency of a conversation.

Contributions are essentially the core of every scientific research, highlighting their key values to the academic community. Systems that are capable of identifying the contributions from scientific papers precisely and organizing them into well-structured summaries can facilitate both text processing and human comprehension. In this paper, we present ContributionSum, a dataset consisting of 24K computer science papers with contributions explicitly listed by the authors, which are further classified into different contribution types based on a newly-proposed annotation scheme. In addition, we study the task of generating disentangled contributions that summarize the values of scientific papers into key points. We propose a fine-grained post-training strategy tailored to our task and leverage salient information of different contribution types in the papers. To assess the coherency and coverage of each contribution aspect, we perform summary-level and contribution-level evaluations for our task. Experimental results show that our method improves upon mainstream baselines.

This work proposes a new natural language processing (NLP) task to tackle the issue of multi-round, sequential text-based knowledge update. The study introduces a hybrid learning architecture and a novel self-supervised training strategy to enable generative language models to consolidate knowledge in the same way as humans. A dataset was also created for evaluation and results showed the effectiveness of our methodology. Experimental results confirm the superiority of the proposed approach over existing models and large language models (LLMs). The proposed task and model framework have the potential to significantly improve the automation of knowledge organization, making text-based knowledge an increasingly crucial resource for powerful LLMs to perform various tasks for humans.

As the significant growth of scientific publications in recent years, an efficient way to extract scholarly knowledge and organize the relationship among literature is necessitated. Previous works constructed scientific knowledge graph with authors, papers, citations, and scientific entities. To assist researchers to grasp the research context comprehensively, this paper constructs a fine-grained citation graph in which citation intents and their supporting evidence are labeled between citing and cited papers instead. We propose a model with a Transformer encoder to encode the long-lengthy paper. To capture the coreference relations of words and sentences in a paper, a coreference graph is created by utilizing Gated Graph Convolution Network (GGCN). We further propose a graph modification mechanism to dynamically update the coreference links. Experimental results show that our model achieves promising results on identifying multiple citation intents in sentences.

This paper introduces the DynamicESG dataset, a unique resource for dynamically extracting ESG ratings from news articles. The ESG rating, a novel metric employed annually to gauge a company's sustainability, relies heavily on corporate disclosure and other external information, especially news narratives. Our dataset, comprising a wide spectrum of news over a twelve-year span, annotates articles in accordance with MSCI ESG ratings methodology and SASB standards, with relevance to ESG issues. DynamicESG provides a comprehensive means of investigating the relationship between public discourse, ESG-related events, and subsequent ESG rating adjustments. We detail our data collection, curation, annotation procedure, and inter-rater agreement, ensuring high data quality and usability. Importantly, our dataset includes a temporal dimension, enabling the analysis of longitudinal trends in ESG ratings and their correlation with news coverage. Moreover, the dataset incorporates an opportunity/risk tendency, thus permitting analysis from diverse perspectives to discern if the news is beneficial or detrimental to the company. We believe this dataset will serve as a valuable resource for researchers in fields such as corporate social responsibility, sustainable investing, machine learning, and natural language processing. Initial analysis using the dataset underscores its potential to facilitate new insights into the dynamics of ESG ratings and the influence of news media on these ratings.

This paper investigates a crucial aspect of mental health by exploring the detection of suicidal ideation in spoken phone conversations between callers and counselors at a suicide prevention hotline.These conversations can be lengthy, noisy, and cover a broad range of topics, making it challenging for NLP models to accurately identify the caller's suicidal ideation.To address these difficulties, we introduce a novel, self-adaptive approach that identifies the most critical utterances that the NLP model can more easily distinguish.The experiments use real-world Lifeline transcriptions, expertly labeled, and show that our approach outperforms the baseline models in overall performance with an F-score of 66.01%.In detecting the most dangerous cases, our approach achieves a significantly higher F-score of 65.94% compared to the baseline models, an improvement of 8.9%.The selected utterances can also provide valuable insights for suicide prevention research.Furthermore, our approach demonstrates its versatility by showing its effectiveness in sentiment analysis, making it a valuable tool for NLP applications beyond the healthcare domain.

In various scenarios, such as conference oral presentations, company managers' talks, and politicians' speeches, individuals often contemplate the potential questions that may arise from their presentations.This common practice prompts the research question addressed in this study: to what extent can models generate multiple questions based on a given presentation transcript?To investigate this, we conduct pilot explorations using earnings conference call transcripts, which serve as regular meetings between professional investors and company managers.We experiment with different task settings and methods and evaluate the results from various perspectives.Our findings highlight that incorporating key points retrieval techniques enhances the accuracy and diversity of the generated questions.

Abstract With the aid of recently proposed word embedding algorithms, the study of semantic relatedness has progressed rapidly. However, word-level representations are still lacking for many natural language processing tasks. Various sense-level embedding learning algorithms have been proposed to address this issue. In this paper, we present a generalized model derived from existing sense retrofitting models. In this generalization, we take into account semantic relations between the senses, relation strength, and semantic strength. Experimental results show that the generalized model outperforms previous approaches on four tasks: semantic relatedness, contextual word similarity, semantic difference, and synonym selection. Based on the generalized sense retrofitting model, we also propose a standardization process on the dimensions with four settings, a neighbor expansion process from the nearest neighbors, and combinations of these two approaches. Finally, we propose a Procrustes analysis approach that inspired from bilingual mapping models for learning representations that outside of the ontology. The experimental results show the advantages of these approaches on semantic relatedness tasks.

In this article, we provide an overview of ACM TWEB’s special issue, Financial Technology on the Web . This special issue covers diverse topics: (1) a new architecture for leveraging online news to investment and risk management, (2) a cross-platform analysis of the post quality and users’ behaviors, and (3) an empirical study on disentangling decentralized finance compositions. In addition to a guide for the special issue, we also share a brief opinion on the future of financial technology on the Web.

Keeping knowledge facts up-to-date is labored and costly as the world rapidly changes and new information emerges every second. In this work, we introduce a novel task, news event triggered knowledge update. Given an existing article about a topic with a news event about the topic, the aim of our task is to generate an updated article according to the information from the news event. We create a multi-grained dataset for the investigation of our task. The articles from Wikipedia are collected and aligned with news events at multiple language units, including the citation text, the first paragraph, and the full content of the news article. Baseline models are also explored at three levels of knowledge update, including the first paragraph, the summary, and the full content of the knowledge facts.

Peer review is an essential part of the scientific process in which the research papers are assessed by several reviewers. The author rebuttal phase, which is held at most top conferences, provides an opportunity for the authors to defend their work against the arguments made by the reviewers. The strengths and the weaknesses pointed out by the reviewers, as well as the authors' responses, will be evaluated by the area chair. The final decisions generally accompany meta-reviews regarding the reason for acceptance/rejection. Previous research has studied the generation of meta-review using transformer-based summarization models. However, few of them consider the rebuttals' content and the interaction between reviews and rebuttals' arguments, where the argumentation persuasiveness plays an important role in affecting the final decision. To generate a comprehensive meta-review that well organizes reviewers' opinions and authors' responses, we present a novel generation model that is capable of explicitly modeling the complicated argumentation structure from not only arguments between the reviewers and the authors but also the inter-reviewer discussions. Experimental results show that our model outperforms baselines in terms of both automatic evaluation and human evaluation, demonstrating the effectiveness of our approach.

Handling unanswerable questions in knowledge base question answering (KBQA) has been a focus in recent years. However, how to explain why a given question is unanswerable is rarely discussed. In this work, we seek not only to correct unanswerable questions based on a personal knowledge base, but also to explain the reason of the correction. We argue that different types of questions need heterogeneous subgraphs with different types of connections. We thus propose a heterogeneous subgraph aggregation network with a two-level attention mechanism to detect important entities and relations in subgraphs and attend to informative subgraphs for different questions. We conduct comprehensive experiments on five subgraphs and their combinations, with results that attest the effectiveness of incorporating heterogeneous subgraphs.

Explaining the reasoning of neural models has attracted attention in recent years. Providing highly-accessible and comprehensible explanations in natural language is useful for humans to understand model’s prediction results. In this work, we present a pilot study to investigate explanation generation with a narrative and causal structure for the scenario of health consulting. Our model generates a medical suggestion regarding the patient’s concern and provides an explanation as the outline of the reasoning. To align the generated explanation with the suggestion, we propose a novel discourse-aware mechanism with multi-task learning. Experimental results show that our model achieves promising performances in both quantitative and human evaluation.

When recalling life experiences, people often forget or confuse life events, which necessitates information recall services.Previous work on information recall focuses on providing such assistance reactively, i.e., by retrieving the life event of a given query.Proactively detecting the need for information recall services is rarely discussed.In this paper, we use a humanannotated life experience retelling dataset to detect the right time to trigger the information recall service.We propose a pilot modelstructured event enhancement network (SEEN) that detects life event inconsistency, additional information in life events, and forgotten events.A fusing mechanism is also proposed to incorporate event graphs of stories and enhance the textual representations.To explain the need detection results, SEEN simultaneously provides support evidence by selecting the related nodes from the event graph.Experimental results show that SEEN achieves promising performance in detecting information needs.In addition, the extracted evidence can be served as complementary information to remind users what events they may want to recall.➊ My brother decided to propose to his fiancée, Ellie.Afterward, ➋ She hosted the party in their backyard.It was so fun!They had a beautifully stocked bar with lots of alcohol.➌ The shots and cocktails were fantastic!PRE-RETOLD STORY POST-RETOLD STORY Match Match Conflict Forgotten My brother's engagement party was hosted in my hometown.➍ It was hosted by his fiancée, Ellen.➎ She had it outside in her backyard.Even though it was 90 degrees, we all had so much fun !➏ In addition, it was also the first time that

In a debate, two debaters with opposite stances put forward arguments to fight for their viewpoints. Debaters organize their arguments to support their proposition and attack opponents’ points. The common purpose of debating is to persuade the opponents and the audiences to agree with the mentioned propositions. Previous works have investigated the issue of identifying which debater is more persuasive. However, modeling the interaction of arguments between rounds is rarely discussed. In this paper, we focus on assessing the overall performance of debaters in a multi-round debate on online forums. To predict the winner in a multi-round debate, we propose a novel neural model that is aimed at capturing the interaction of arguments by exploiting raw text, structure information, argumentative discourse units (ADUs), and the relations among ADUs. Experimental results show that our model achieves competitive performance compared with the existing models, and is capable of extracting essential argument relations during a multi-round debate by leveraging argumentative structure and attention mechanism.

Argument structure elaborates the relation among claims and premises. Previous works in persuasiveness prediction do not consider this relation in their architectures. To take argument structure information into account, this paper proposes an approach to persuasiveness prediction with a novel graph-based neural network model, called heterogeneous argument attention network (HARGAN). By jointly training on the persuasiveness and stance of the replies, our model achieves the state-of-the-art performance on the ChangeMyView (CMV) dataset for the persuasiveness prediction task. Experimental results show that the graph setting enables our model to aggregate information across multiple paragraphs effectively. In the meanwhile, our stance prediction auxiliary task enables our model to identify the viewpoint of each party, and helps our model perform better on the persuasiveness prediction.

People often encounter situations where they need to recall past experiences from their daily life. In this paper, we aim to construct a question answering system that enables human to query their past experiences over personal knowledge base. Previous works on knowledge base question answering focus on finding answers for answerable questions. In the real world applications, however, people often muddle up facts and ask those questions that cannot be answered with knowledge base. This work presents a novel system consisting of question answering model and question generation model. It not only answers answerable questions, but also corrects unanswerable questions if necessary. Our question answering model recognizes the question that is inconsistent with the state of the personal knowledge base and suggests facts that can form a feasible question. Then, the facts are converted to an answerable question by the question generation model. For refining question, we propose a question generation model based on the reinforcement learning (RL) with question editing mechanism. Experimental results show that our proposed system is effective for correcting unanswerable questions in personal knowledge base question answering.

Freedom of speech is one of the principles in the constitution of most countries. However, in the 2020 United States presidential election, Donald Trump's Twitter account is suspended due to the risk of further incitement of violence. That leads to the question: Which kind of rumors may undermine society? In this paper, we discuss this question based on the case studies of real-world court orders, which are the judges' official proclamations. We point out the possible research directions that NLP researchers may need to consider before applying our systems to society.

Numeral information plays an important role in narratives of several domains such as medicine, engineering, and finance. Previous works focus on the foundation exploration toward numeracy and show that fine-grained numeracy is a challenging task. In machine reading comprehension, our statistics show that only a few numeral-related questions appear in previous datasets. It indicates that few benchmark datasets are designed for numeracy learning. In this paper, we present a Numeral-related Question Answering Dataset, NQuAD, for fine-grained numeracy, and propose several baselines for future works. We compare NQuAD with three machine reading comprehension datasets and show that NQuAD is more challenging than the numeral-related questions in other datasets. NQuAD is published under the CC BY-NC-SA 4.0 license for academic purposes.

The economic policy uncertainty (EPU) index is one of the important text-based indexes in finance and economics fields. The EPU indexes of more than 26 countries have been constructed to reflect the policy uncertainty on country-level economic environments and serve as an important economic leading indicator. The EPU indexes are calculated based on the number of news articles with some manually-selected keywords related to economic, uncertainty, and policy. We find that the keyword-based EPU indexes contain noise, which will influence their explainability and predictability. In our experimental dataset, over 40% of news articles with the selected keywords are not related to the EPU. Instead of using keywords only, our proposed models take contextual information into account and get good performance on identifying the articles unrelated to EPU. The noise free EPU index performs better than the keyword-based EPU index in both explainability and predictability.

The volatility of stock price reflects the risk of stock and influences the risk of investor's portfolio. It is also a crucial part of pricing derivative securities. Researchers have paid their attention to predict the stock volatility with different kinds of textual data. However, most of them focus on using word information only. Few touch on capturing the numeral information in textual data, providing fine-grained clues for financial document understanding. In this paper, we present a novel dataset, ECNum, for understanding the numerals in the transcript of earnings conference calls. We propose a simple but efficient method, Numeral-Aware Model (NAM), for enhancing the capacity of numeral understanding of neural network models. We employ the distilled information in the stock volatility forecasting task and achieve the best performance compared to the previous works in short-term scenarios.

Filming video blogs, which is shortened to vlog, becomes a popular way for people to record their life experiences in recent years. In this work, we present a novel task that is aimed at extracting life events from videos and constructing personal knowledge bases of individuals. In contrast to most existing researches in the field of computer vision that focus on identifying low-level script-like activities such as moving boxes, our goal is to extract life events where high-level activities like moving into a new house are recorded. The challenges to be tackled include: (1) identifying which objects in a given scene related to the life events of the protagonist we concern, and (2) determining the association between an extracted visual concept and a more high-level description of a video clip. To address the research issues, we construct a video life event extraction dataset VidLife by exploiting videos from the TV series The Big Bang Theory, in which the plot is around the daily lives of several characters. A pilot multitask learning model is proposed to extract life events given video clips and subtitles for storing in the personal knowledge base.

Textual information extraction is a typical research topic in the NLP community.Several NLP tasks such as named entity recognition and relation extraction between entities have been well-studied in previous work.However, few works pay their attention to the implicit information.For example, a financial news article mentioned "Apple Inc." may be also related to Samsung, even though Samsung is not explicitly mentioned in this article.This work presents a novel dynamic graph transformer that distills the textual information and the entity relations on the fly.Experimental results confirm the effectiveness of our approach to implicit tag recognition.

Both the issues of data deficiencies and semantic consistency are important for data augmentation.Most of previous methods address the first issue, but ignore the second one.In the cases of aspect-based sentiment analysis, violation of the above issues may change the aspect and sentiment polarity.In this paper, we propose a semantics-preservation data augmentation approach by considering the importance of each word in a textual sequence according to the related aspects and sentiments.We then substitute the unimportant tokens with two replacement strategies without altering the aspect-level polarity.Our approach is evaluated on several publicly available sentiment analysis datasets and the real-world stock price/risk movement prediction scenarios.Experimental results show that our methodology achieves better performances in all datasets.

This paper presents a conversational lifelog mining system, ConvLogMiner, which detects personal life events from the human online conversation in real-time. Given a daily conversation of two speakers, ConvLogMiner identifies the new life events specific to each speaker that occur in the latest utterances. The lifelogs mined by our system are useful to provide complementary information to support lifestyle analysis and memory assistance service.

With the passage of time, people often have misty memories of their past experiences. Information recall support for people by collecting personal lifelogs is emerging. Recently, people tend to record their daily life via filming Video Weblog (VLog), which contains visual and audio data. These large scale multimodal data can be used to support information recall service that enables users to query their past experiences. The challenging issue is the semantic gap between the visual concept and the textual query. In this paper, we aim to extract personal life events from vlogs shared on YouTube and construct a personal knowledge base (PKB) for individuals. A multitask learning model is proposed to extract the components of personal life events, such as subjects, predicates and objects. The evaluation is performed on a video collection from three YouTubers who are English native speakers. Experimental results show our model achieves promising performance.

With the advance of science and technology, people are used to recording their daily life events via writing blogs, uploading social media posts, taking photos, or filming videos. Such rich repository personal information is useful for supporting human living assistance, such as information recall service. The main challenges are how to store and manage personal knowledge from various sources, and how to provide support for people who may have difficulty recalling past experiences. In this position paper, we propose a research agenda on personal knowledge mining from various sources of lifelogs, personal knowledge base construction, and information recall for assisting people to recall their experiences. Ten research questions are formulated.

Previous studies on the financial news focus mainly on the news articles explicitly mentioning the target financial instruments, and may suffer from data sparsity. As taking into consideration other related news, e.g., sector-related news, is a crucial part of real-world decision-making, we explore the use of news without explicit target mentions to enrich the information for the prediction model. We develop a neural network framework that jointly learns with a news selection mechanism to extract implicit information from the chaotic daily news pool. Our proposed model, called the news distilling network (NDN), takes advantage of neural representation learning and collaborative filtering to capture the relationship between stocks and news. With NDN, we learn latent stock and news representations to facilitate similarity measurements, and apply a gating mechanism to prevent noisy news representations from flowing to a higher level encoding stage, which encodes the selected news representation of each day. Extensive experiments on real-world stock market data demonstrate the effectiveness of our framework and show improvements over previous techniques.

This paper demonstrates FinSense, a system that improves the working efficiency of financial information processing. Given the draft of a financial news story, FinSense extracts the explicit-mentioned stocks and further infers the implicit stocks, providing insightful information for decision making. We propose a novel graph convolutional network model that performs implicit financial instrument inference toward the in-domain data. In addition, FinSense generates candidate headlines for the draft, reducing a significant amount of time in journalism production. The proposed system also provides assistance to investors to sort out the information in the financial news articles.

Social media’s rise in popularity has demonstrated the usefulness of the wisdom of the crowd. Most previous works take into account the law of large numbers and simply average the results extracted from tasks such as opinion mining and sentiment analysis. Few attempt to identify high-quality opinions from the mined results. In this paper, we propose an approach for capturing expert-like rationales from social media platforms without the requirement of the annotated data. By leveraging stylistic and semantic features, our approach achieves an F1-score of 90.81%. The comparison between the rationales of experts and those of the crowd is done from stylistic and semantic perspectives, revealing that stylistic and semantic information provides complementary cues for professional rationales. We further show the advantage of using these superlative analysis results in the financial market, and find that top-ranked opinions identified by our approach increase potential returns by up to 90.31% and reduce downside risk by up to 71.69%, compared with opinions ranked by feedback from social media users. Moreover, the performance of our method on downside risk control is comparable with that of professional analysts.

This work proposes a standalone, complete Chinese discourse parser for practical applications.We approach Chinese discourse parsing from a variety of aspects and improve the shift-reduce parser not only by integrating the pre-trained text encoder, but also by employing novel training strategies.We revise the dynamic-oracle procedure for training the shift-reduce parser, and apply unsupervised data augmentation to enhance rhetorical relation recognition.Experimental results show that our Chinese discourse parser achieves the state-of-the-art performance.

The goal of claim detection in argument mining is to sort out the key points from a long narrative. In this paper, we design a novel task for argument mining in the financial domain, and provide an expert-annotated dataset, NumClaim, for the proposed task. Based on the statistics, we discuss the differences between the claims in other datasets and the claims of the investors in NumClaim. With the ablation analysis, we show that encoding numeral and co-training with the auxiliary task of the numeral understanding, i.e., the category classification task, can improve the performance of the proposed task under different neural network architectures. The annotations in the NumClaim is published for academic usage under the CC BY-NC-SA 4.0 license.

Most recent works in the field of grammatical error correction (GEC) rely on neural machine translation-based models. Although these models boast impressive performance, they require a massive amount of data to properly train. Furthermore, NMT-based systems treat GEC purely as a translation task and overlook the editing aspect of it. In this work we propose a heterogeneous approach to Chinese GEC, composed of a NMT-based model, a sequence editing model, and a spell checker. Our methodology not only achieves a new state-of-the-art performance for Chinese GEC, but also does so without relying on data augmentation or GEC-specific architecture changes. We further experiment with all possible configurations of our system with respect to model composition order and number of rounds of correction. A detailed analysis of each model and their contributions to the correction process is performed by adapting the ERRANT scorer to be able to score Chinese sentences.

This work presents AutoSurvey, an intelligent system that performs literature survey and generates a summary specific to a research draft. A neural model for information structure analysis is employed for extracting fine-grained information from the abstracts of previous work, and a novel evolutionary multi-source summarization model is proposed for generating the summary of related work. This system is extremely used for both academic and educational purposes.

A dialogue dataset is an indispensable resource for building a dialogue system. Additional information like emotions and interpersonal relationships labeled on conversations enables the system to capture the emotion flow of the participants in the dialogue. However, there is no publicly available Chinese dialogue dataset with emotion and relation labels. In this paper, we collect the conversions from TV series scripts, and annotate emotion and interpersonal relationship labels on each utterance. This dataset contains 25,548 utterances from 4,142 dialogues. We also set up some experiments to observe the effects of the responded utterance on the current utterance, and the correlation between emotion and relation types in emotion and relation classification tasks.

In this paper, we investigate the annotation of financial social media data from several angles. We present Fin-SoMe, a dataset with 10,000 labeled financial tweets annotated by experts from both the front desk and the middle desk in a bank’s treasury. These annotated results reveal that (1) writer-labeled market sentiment may be a misleading label; (2) writer’s sentiment and market sentiment of an investor may be different; (3) most financial tweets provide unfounded analysis results; and (4) almost no investors write down the gain/loss results for their positions, which would otherwise greatly facilitate detailed evaluation of their performance. Based on these results, we address various open problems and suggest possible directions for future work on financial social media data. We also provide an experiment on the key snippet extraction task to compare the performance of using a general sentiment dictionary and using the domain-specific dictionary. The results echo our findings from the experts’ annotations.

Chinese discourse parsing, which aims to identify the hierarchical relationships of Chinese elementary discourse units, has not yet a consistent evaluation metric. Although Parseval is commonly used, variations of evaluation differ from three aspects: micro vs. macro F1 scores, binary vs. multiway ground truth, and left-heavy vs. right-heavy binarization. In this paper, we first propose a neural network model that unifies a pre-trained transformer and CKY-like algorithm, and then compare it with the previous models with different evaluation scenarios. The experimental results show that our model outperforms the previous systems. We conclude that (1) the pre-trained context embedding provides effective solutions to deal with implicit semantics in Chinese texts, and (2) using multiway ground truth is helpful since different binarization approaches lead to significant differences in performance.

Sense embedding models handle polysemy by giving each distinct meaning of a word form a separate representation. They are considered improvements over word models, and their effectiveness is usually judged with benchmarks such as semantic similarity datasets. However, most of these datasets are not designed for evaluating sense embeddings. In this research, we show that there are at least six concerns about evaluating sense embeddings with existing benchmark datasets, including the large proportions of single-sense words and the unexpected inferior performance of several multi-sense models to their single-sense counterparts. These observations call into serious question whether evaluations based on these datasets can reflect the sense model’s ability to capture different meanings. To address the issues, we propose the Multi-Sense Dataset (MSD-1030), which contains a high ratio of multi-sense word pairs. A series of analyses and experiments show that MSD-1030 serves as a more reliable benchmark for sense embeddings. The dataset is available at http://nlg.csie.ntu.edu.tw/nlpresource/MSD-1030/.

The advance in wearable technology has made lifelogging more feasible and more popular. Visual lifelogs collected by wearable cameras capture every single detail of individual's life experience, offering a promising data source for deeper lifestyle analysis and better memory recall assistance. However, building a system for organizing and accessing visual lifelogs is a challenging task due to the semantic gap between visual data and semantic descriptions of life events. In this paper, we introduce semantic knowledge to reduce such a semantic gap for daily activity recognition and lifestyle understanding. We incorporate the semantic knowledge derived from external resources to enrich the training data for the proposed supervised learning model. Experimental results show that incorporating external semantic knowledge is beneficial for improving the performance of recognizing life events.

This paper introduces visual story ordering, a challenging task in which images and text are ordered in a visual story jointly. We propose a neural network model based on the reader-processor-writer architecture with a self-attention mechanism. A novel bidirectional decoder is further proposed with bidirectional beam search. Experimental results show the effectiveness of the approach. The information gained from multimodal learning is presented and discussed. We also find that the proposed embedding narrows the distance between images and their corresponding story sentences, even though we do not align the two modalities explicitly. As it addresses a general issue in generative models, the proposed bidirectional inference mechanism applies to a variety of applications.

In this research, we propose 3 different approaches to measure the semantic relatedness between 2 words: (i) boost the performance of GloVe word embedding model via removing or transforming abnormal dimensions; (ii) linearly combine the information extracted from WordNet and word embeddings; and (iii) utilize word embedding and 12 linguistic information extracted from WordNet as features for Support Vector Regression. We conducted our experiments on 8 benchmark data sets, and computed Spearman correlations between the outputs of our methods and the ground truth. We report our results together with 3 state‐of‐the‐art approaches. The experimental results show that our method can outperform state‐of‐the‐art approaches in all the selected English benchmark data sets.

In this paper, we attempt to answer the question of whether neural network models can learn numeracy, which is the ability to predict the magnitude of a numeral at some specific position in a text description.A large benchmark dataset, called Numeracy-600K, is provided for the novel task.We explore several neural network models including CNN, GRU, BiGRU, CRNN, CNN-capsule, GRU-capsule, and BiGRU-capsule in the experiments.The results show that the BiGRU model gets the best micro-averaged F1 score of 80.16%, and the GRU-capsule model gets the best macroaveraged F1 score of 64.71%.Besides discussing the challenges through comprehensive experiments, we also present an important application scenario, i.e., detecting exaggerated information, for the task.

Social trading platforms provide a forum for investors to share their analysis and opinions. Posts on these platforms are characterized by narrative styles which are much different from posts on general social platforms, for instance tweets. As a result, recommendation systems for social trading platforms should leverage tailor-made latent features. This paper presents a representation for these latent features in both textual data and market information. A real-world dataset is adopted to conduct experiments involving a novel task called next cashtag prediction. We propose a joint learning model with an attentive capsule network. Experimental results show positive results with the proposed methods and the corresponding auxiliary tasks.

This paper demonstrates an opinion indicator (OI) generation system, named Crowd View, with which traders can refer to the fine-grained opinions, beyond the market sentiment (bullish/bearish), from crowd investors when trading financial instruments. We collect the real-time textual information from Twitter, and convert it into five kinds of OIs, including the support price level, resistance price level, price target, buy-side cost, and sell-side cost. The OIs for all component stocks in Dow Jones Industrial Average Index (DJI) are provided, and shown with the real-time stock price for comparison and analysis. The information embedding in the OIs and the application scenarios are introduced.

In this paper we propose the task of numeral attachment to detect the attached target of a numeral. Compared with other kinds of named entities, numerals provide richer and more crucial information in some domains. Fine-grained understanding of the information embedded in numerals is a fundamental challenge. We develop NumAttach, a pilot dataset for the proposed task based on tweets. Two main challenges of this task include the informal writing style in tweets and the representation of numerals. To address these challenges, we present an embedding technique that considers word and numeral information simultaneously. Furthermore, we design a joint learning model with the capsule network to accomplish the proposed task. We also release NumAttach to the research community as a resource.

Previous work on lifelogging focuses on life event extraction from image, audio, and video data via wearable sensors. In contrast to wearing an extra camera to record daily life, people are used to log their life on social media platforms. In this paper, we aim to extract life events from textual data shared on Twitter and construct personal knowledge bases of individuals. The issues to be tackled include (1) not all text descriptions are related to life events, (2) life events in a text description can be expressed explicitly or implicitly, (3) the predicates in the implicit events are often absent, and (4) the mapping from natural language predicates to knowledge base relations may be ambiguous. A joint learning approach is proposed to detect life events in tweets and extract event components including subjects, predicates, objects, and time expressions. Finally, the extracted information is transformed to knowledge base facts. The evaluation is performed on a collection of lifelogs from 18 Twitter users. Experimental results show our proposed system is effective in life event extraction, and the constructed personal knowledge bases are expected to be useful to memory recall applications.

People engage in lifelogging by taking photos with cameras and cellphones anytime anywhere and share the photos, intertwined with captions or descriptions, on social media platforms. The image-text intertwined data provides richer information for image recall. When images cannot keep the complete information, the textual information is a complement to describe the life experiences under the photos. This work proposes a multimodal retrieval model for image recall in image-text intertwined lifelogs. Our Attentive Image-Story model combines an Image model, which transfers visual information and textual information to a single representation space, and a Story model, which captures text-based contextual information, with an attention mechanism to reduce the semantic gap between visual and textual information. Experimental results show our model outperforms a state-of-the-art image-based retrieval model and the image/text hybrid system.

Understanding how information propagates among social media users can allow researchers to provide interesting insights into online social networks and lead to applications such as precise advertising and misinformation management. In this work, we focus on information diffusion through post sharing. Given an information propagation tree, our goal is to predict a list of potential users of the tree. A framework based on graph convolutional network (GCN) is proposed to learn the latent representation of a propagation tree and match it with the latent representation of a user. A novel strategy for tree pruning is further investigated to improve the GCN. Experimental results show that our framework outperforms the existing methods for modeling information diffusion.

This paper demonstrates a novel analytics service, CrowdPT, for capturing the key information, price target (PT), of individual investors on social media. PT, which is mentioned as a conclusion in most of analysts' reports, indicates not only the market sentiment (bullish/bearish) of investors, but also the analysis results. In order to provide the latest opinions of individual investors, we monitor Twitter in real time and update the information in price chart daily. For all component stocks in Dow Jones Industrial Average, textual information from numerous tweets is summarized into a single number, PT, in CrowdPT. Case studies confirm the effectiveness of our analytics service in the financial domain, and show that capturing the PT of individual investors is promising for stock price prediction. The Web API of CrowdPT is also provided for academic purpose.

With the aid of recently proposed word embedding algorithms, the study of semantic similarity has progressed and advanced rapidly. However, many natural language processing tasks need sense level representation. To address this issue, some researches propose sense embedding learning algorithms. In this paper, we present a generalized model from existing sense retrofitting model. The generalization takes three major components: semantic relations between the senses, the relation strength and the semantic strength. In the experiment, we show that the generalized model can outperform previous approaches in three types of experiment: semantic relatedness, contextual word similarity and semantic difference.

With more and more people around the world learning Chinese as a second language, the need of Chinese error correction tools is increasing. In the HSK dynamic composition corpus, word usage error (WUE) is the most common error type. In this paper, we build a neural network model that considers both target erroneous token and context to generate a correction vector and compare it against a candidate vocabulary to propose suitable corrections. To deal with potential alternative corrections, the top five proposed candidates are judged by native Chinese speakers. For more than 91% of the cases, our system can propose at least one acceptable correction within a list of five candidates. To the best of our knowledge, this is the first research addressing general-type Chinese WUE correction. Our system can help non-native Chinese learners revise their sentences by themselves.

This work demonstrates a writing assistant system that provides high level advice for Chinese scientific writing. Cross-lingual approaches are investigated to analyze the information structure of a given Chinese abstract and retrieve useful knowledge in the related work written in both English and Chinese. To the best of our knowledge, this is the first study on Chinese information structure identification. Without the need of labeled Chinese data, our novel model is capable of dealing with Chinese instances by acquiring language-invariant knowledge from the labeled English data. Adversarial learning is employed to enhance the cross-lingual sentence representation.

Numerals that contain much information in financial documents are crucial for financial decision making. They play different roles in financial analysis processes. This paper is aimed at understanding the meanings of numerals in financial tweets for fine-grained crowd-based forecasting. We propose a taxonomy that classifies the numerals in financial tweets into 7 categories, and further extend some of these categories into several subcategories. Neural network-based models with word and character-level encoders are proposed for 7-way classification and 17-way classification. We perform backtest to confirm the effectiveness of the numeric opinions made by the crowd. This work is the first attempt to understand numerals in financial social media data, and we provide the first comparison of fine-grained opinion of individual investors and analysts based on their forecast price. The numeral corpus used in our experiments, called FinNum 1.0, is available for research purposes.

Verbalization of knowledge base (KB) facts about an entity allows users to absorb information from KB more easily. The drawback of most previous work is that they cannot generalize to unseen frames. This work introduces the task of precise KB verbalization that is aimed at generating an exact description for the given factual triples. We propose a novel sequence-to-sequence (seq2seq) model with the local pointer network to deal with this task. The approach to training data construction is also explored. Experimental results show our method improves the performances in terms of Meteor and slot error rates. Human evaluation is also performed to confirm the effectiveness of our model.

With the aid of recently proposed word embedding algorithms, the study of semantic relatedness has progressed and advanced rapidly. In this research, we propose a novel structural-fitting method that utilizes the linguistic ontology into vector space representations. The ontological information is applied in two ways. The fine2coarse approach refines the word vectors from fine-grained to coarse-grained terms (word types), while the coarse2fine approach refines the word vectors from coarse-grained to fine-grained terms. In the experiments, we show that our proposed methods outperform previous approaches in seven publicly available benchmark datasets.

This paper demonstrates DISA, a higher-level writing assistant system, which analyzes the information structure of abstracts, and retrieves the knowledge according to the research goals from the related work. By incorporating the latest neural-network technologies including linguistically-informed neural-network and autoencoder, we construct an intelligent system which extends the scope of computer-aided writing.

This paper presents an approach to identify subject, type and property from knowledge base (KB) for answering simple questions. We propose new features to rank entity candidates in KB. Besides, we split a relation in KB into type and property. Each of them is modeled by a bi-directional LSTM. Experimental results show that our model achieves the state-of-the-art performance on the SimpleQuestions dataset. The hard questions in the experiments are also analyzed in detail.

There has been a surge in opinion-sharing in the public domain. Some opinions greatly influence our decisions, e.g., the choice of purchase. Malicious parties or individuals exploit social media by generating fake reviews for opinion manipulation. This paper aims to investigate the phenomenon of online paid restaurant reviews by bloggers. Our research provides an insight into some characteristics of paid reviews and their authors. We then explore a set of features based on our observations and detect paid reviews and paid bloggers using supervised machine learning techniques. Experimental results show the effectiveness of our approach.

This paper analyzes the lexical semantics of domain-specific terms based on various pre-trained specific domain and general domain word vectors, and addresses the semantic drift between domains. To capture lexical semantics in the specific domain, we propose a bridge mechanism to introduce domain-specific data into general data, and re-train word vectors. We find that even a small-scale fusion can result in the similar lexical semantics learned by using the large-scale domain-specific dataset. Experiments on sentiment analysis and outlier detection show that application of word embedding by the fusion dataset has the better performance than applications of word embeddings by pure large domain-specific and pure large general datasets. The simple, but effective methodology facilitates the domain adaptation of distributed word representations.

Opinion words are crucial information for sentiment analysis. In some text, however, opinion words are absent or highly ambiguous. The resulting implicit opinions are more difficult to extract and label than explicit ones. In this paper, cutting‐edge machine‐learning approaches – deep neural network and word‐embedding – are adopted for implicit opinion mining at the snippet and clause levels. Hotel reviews written in Chinese are collected and annotated as the experimental data set. Results show the convolutional neural network models not only outperform traditional support vector machine models, but also capture hidden knowledge within the raw text. The strength of word‐embedding is also analyzed.

Misuse of Chinese prepositions is one of common word usage errors in grammatical error diagnosis. In this paper, we adopt the Chinese Gigaword corpus and HSK corpus as L1 and L2 corpora, respectively. We explore gated recurrent neural network model (GRU), and an ensemble of GRU model and maximum entropy language model (GRU-ME) to select the best preposition from 43 candidates for each test sentence. The experimental results show the advantage of the GRU models over simple RNN and n-gram models. We further analyze the effectiveness of linguistic information such as word boundary and part-of-speech tag in this task.

This paper explores the role of tense information in Chinese causal analysis. Both tasks of causal type classification and causal directionality identification are experimented to show the significant improvement gained from tense features. To automatically extract the tense features, a Chinese tense predictor is proposed. Based on large amount of parallel data, our semi-supervised approach improves the dependency-based convolutional neural network (DCNN) models for Chinese tense labelling and thus the causal analysis.