When processing multi-view data, conflicts may occur since different views have unique insights. Existing studies always consider conflict as a bad factor and thus adopt a negative operation, i.e., eliminating or minimizing conflicts. However, in a multi-view decision-making scenario, conflict can highlight view differences and reveal the reliability of individual and collective decisions. To this end, we propose a novel trusted collective learning method (TrustCL) that can actively handle conflicts between individuals and obtain a collective opinion and its reliability by considering conflicts. Specifically, TrustCL first learns view-specific evidence supporting individual opinions. To deal with conflicting opinions across views, TrustCL assembles a reliability-oriented collective learning phase to determine which view owns a higher priority. It further combines priority and view-specific evidence to conclude the final collective opinion and its reliability. Experiments on a real-world multidisciplinary consultation dataset demonstrate the superiority of our method and exhibit some interesting findings regarding conflictive multi-view decision-making. Our code is available at https://github.com/ifbettrer/TrustCL

Document-level relation extraction (DocRE) is a more complex and practically significant task than sentence-level extraction, as it involves identifying relations between entities that span multiple sentences. In this paper, we propose TLAIF_DLRE, a novel architecture aimed at improving entity representation and addressing key training challenges. Our model enhances the expressiveness of head and tail entity representations by integrating potential entity mentions (anaphors). Additionally, we introduce a Two-Stage Loss(TSL) function for knowledge distillation, which mitigates challenges such as long-tail class distributions and positive-negative class imbalance by shifting the training focus during the teacher-student model training process. Experimental results on the widely used DocRED and Re-DocRED datasets demonstrate that TLAIF_DLRE achieves state-of-the-art performance1.

Image-text retrieval (ITR) is a fundamental task in multimodal learning. It aims to bridge image and text by constructing a shared embedding space that achieves accurate semantic alignment across two modalities. Most existing work has been devoted to designing tailored cross-attention modules to pursue retrieval accuracy but ignores the learning potential of the model network architecture. This work introduces a dual encoder that uses global information to enhance the feature interaction between visual regions. Designing an Adaptive Pooling (AP) module enables the model to automatically learn the optimal aggregation strategy from multiple perspectives based on the local features. We further propose a Dynamic Triplet Loss (DTL) to adjust the learning objective of the model network dynamically for efficient training. Experimental results on two benchmark datasets, Flickr30K and MS-COCO, demonstrate our APDTL model achieves state-of-the-art performance. Our code is released at github.com/jinshuaihu/APDTL.

Visual Question Answering with Natural Language Explanations (VQA-NLE) requires models to provide logically grounded explanations while generating accurate answer. Existing methods struggle with complex questions that demand comprehensive logical reasoning and world knowledge integration. We propose that decomposing the target image and question into question-answer-explanation triples containing visual cues, world knowledge, and object relationships can enhance image understanding, question comprehension, and knowledge retrieval for final answer reasoning. We call the proposed method Sub-QAE Prompting that includes three key steps. We train a visual question generation (VQG) model using instruction-aligned question contexts which subsequently utilized to generate the sub-questions from image. Then generate image-based and text-based sub-questions via the VQG model and frozen large language model respectively. Last we construct question-answer-explanation triples by deriving corresponding answers and explanation for generated subquestions, and encode question-answer-explanation triples into visual-aware prompting module for MLLM to generate the final answer and explanation. Experimental results on the two challenge benchmark VQA-X and A-OKVQA demonstrate that our method achieves state-of-the-art performance compared to existing VQA-NLE approaches.

Metonymy resolution (MR) is a crucial challenge in natural language understanding and information retrieval. Recent large-scale pre-trained language models have shown promising results in various natural language processing (NLP) tasks, including MR. Despite these achievements, current models still struggle in many real-world scenarios. Since these models rely heavily on contextual information and ignore entity information, they are prone to extract irrelevant features and overfit when fine-tuned with less training data. In this article, we propose a mixed information bottleneck framework to address the above issues, which learns optimal data representations based on the principle of minimal sufficiency. Our model can effectively mitigate irrelevant features in context and entity by using different types of information bottlenecks for entity and context information separately while reducing the dimensionality of latent representations. We show that our approach achieves state-of-the-art performance on three benchmark datasets for location MR, outperforming previous Bert-based methods by a large margin. Ablation studies and qualitative analysis show the effectiveness of our models in reducing dimensionality while extracting more relevant features.

One challenge of learning from streaming data is that only a limited number of labeled examples are available, making semi-supervised learning (SSL) algorithms becoming an efficient tool for streaming data mining. Recently, the graph-based SSL algorithms have been proposed to improve SSL performance because the graph structure can utilize the interactivity between surrounding nodes. However, graph-based SSL algorithms have two main limitations when applied to streaming data. First, not all the labels of the data in the streaming data may be reliable, and direct classification using a graph can lead to suboptimal performance. Second, graph-based SSL algorithms assume the structure of the graph is static, but the learning environment of streaming data is dynamic. Hence, we propose a competence-aware graph neural network (CA-GNN) to deal with these two limitations. Unlike other models, CA-GNN does not directly rely on graph information that could include mislabeled nodes. Instead, a competence model is used to explore latent semantic correlations in the streaming data and capture the reliability for each data. A streaming learning strategy then evolves CA-GNN's parameters to capture the dynamism of the graph sequences. We conducted experiments using seven real datasets and four synthetic datasets, respectively, and compared the outcomes across various methods. The results demonstrate that CA-GNN classifies streaming data more effectively than the state-of-the-art (SOTA) methods.

The current charge prediction datasets mostly focus on single-defendant criminal cases.However, real-world criminal cases usually involve multiple defendants whose criminal facts are intertwined.In an early attempt to fill this gap, we introduce a new benchmark that encompasses legal cases involving multiple defendants, where each defendant is labeled with a charge and four types of crime elements, i.e., Object Element, Objective Element, Subject Element, and Subjective Element.Based on the dataset, we further develop an interpretable model called EJudge that incorporates crime elements and legal rules to infer charges.We observe that predicting crime charges while providing corresponding rationales benefits the interpretable AI system.Extensive experiments show that EJudge significantly surpasses state-of-the-art methods, which verify the importance of crime elements and legal rules in multi-defendant charge prediction.

Dialogue generation, as a crucial subtask of dialogue systems, is garnering increasing attention in the field of Natural Language Processing (NLP). The success of dialogue generation relies on effectively utilizing context information to ensure coherent and diverse responses. However, current approaches heavily rely on external sources rather than leveraging the inherent dialogue content. We propose a new approach to address this challenge by introducing semantic segmentation from the field of image processing into NLP. Our contribution lies in the development of a Dialogue Generation model with Hierarchical Encoding and Semantic segmentation of dialogue Context, which is called DGHESC. This model is topic and speaker-aware, capturing the flow of topic and speaker information within the dialogue context using a hierarchical transformer-based framework. Specifically, we extract semantic information at the word-level for each utterance, segment the dialogue context based on topic and speaker semantics, and employ attention mechanisms to model the context at the utterance-level. Experimental results on two open-domain datasets demonstrate the effectiveness of DGHESC. It enhances response quality and achieves state-of-the-art performances on the datasets.

In the multi-label text classification task, a text usually corresponds to multiple label categories, and the labels have correlation and hierarchical structure. However, when the label hierarchy is unknown, the number of various labels is not balanced, which makes it difficult for the model to classify low-frequency labels. In addition, labels have semantic similarities that make it difficult for the model to distinguish between them. In this article, we propose a multi-label text classification model based on multi-level constraint augmentation and label association attention. Compared with traditional methods, our method has two contributions: (1) In order to alleviate the problem of unbalanced number of different label categories and ensure the rationality of sample generation, we propose a data augmentation method based on multi-level constraints. In the process of sample generation, this method uses historical generation information, sample original text information, and sample topic to constrain the generated text. (2) In order to make the model recognize the associated labels accurately, we propose an interaction mechanism based on label association attention and filter gate. This method combines text information and label weight information. At the same time, our classification model considers the important weights of text sentences and effectively utilizes the co-occurrence relationship between labels. Experimental results on three benchmark datasets show that our model outperforms state-of-the-art methods on all main evaluation metrics, especially on low-frequency label prediction with sparse samples.

Legal Case Recommendation (LCR) is to find out the documents that are most similar to the input case from the judicial point of view. Since the legal documents are long texts and have strong legal attributes, the traditional recommendation method based on text similarity is difficult to accurately understand the legal documents, resulting in poor effect of LCR. To address this problem, we propose Recursive Graph Representation Learning (ReGRL) to hierarchically learn the information in the graph, and obtain a more informative graph representation to accurately understand the case. ReGRL captures nodes, edge, and community information at different levels to integrate information of different granularity. To achieve this, ReGRL performs top-down graph decomposition and bottom-up graph encoding in a recursive form, which allows ReGRL to flexibly control the depth of the learning layers and provide accurate case representations for LCR. Experimental results show that ReGRL can not only generate a good representation, but has a better performance compared to text representation methods for LCR. In addition, we also performed different experiments to analyze the principle of ReGRL and verify the effectiveness of recursive procedures.

As one of the most common social behavior in human society, communication in multi-turn conversation or dialogue system has always been a research focuses of natural language processing (NLP).The quality of downstream tasks in multi-turn dialogue is often determined by the result of dialogue context modeling.For dialogue generation, the context information will determine the consistency and diversity of the generated responses.However, the current research on dialogue generation increasingly relies on external information rather than mining from the dialogue content itself.In this paper, we propose a topic and speaker-aware hierarchical encoder-decoder (TSHED) model to capture the topic and speaker information flow in the context for response generation with the hierarchical transformer-based framework.Specifically, we obtain semantic information of each utterance at word-level and then apply topic and speaker-aware attention to model context at utterancelevel.Experimental results on two open-domain datasets show that TSHED significantly improves the quality of responses and outperforms strong baselines.

Crowdsourcing is a process of acquisition, integration, and analysis of big and heterogeneous data generated by a diversity of sources in urban spaces, such as sensors, devices, vehicles, buildings, and human. Especially, nowadays, no countries, no communities, and no person are immune to urban emergency events. Detection about urban emergency events, e.g., fires, storms, traffic jams is of great importance to protect the security of humans. Recently, social media feeds are rapidly emerging as a novel platform for providing and dissemination of information that is often geographic. The content from social media usually includes references to urban emergency events occurring at, or affecting specific locations. In this paper, in order to detect and describe the real time urban emergency event, the 5W (What, Where, When, Who, and Why) model is proposed. Firstly, users of social media are set as the target of crowd sourcing. Secondly, the spatial and temporal information from the social media are extracted to detect the real time event. Thirdly, a GIS based annotation of the detected urban emergency event is shown. The proposed method is evaluated with extensive case studies based on real urban emergency events. The results show the accuracy and efficiency of the proposed method.

Association link network (ALN) can organize massive Web information to provide many intelligent services in our big data society. Effective semantic layered technologies not only can provide theoretical support for knowledge discovery in Web resources, but also can improve the searching efficiency of related information systems such as Web information system and industrial information system. How to realize the layer division of association semantic by the hierarchy analysis of ALN is an important research topic. To solve this problem, this paper proposes a hierarchy-cutting model of association semantic. First, experiments of four types of keywords with different linking roles are conducted to discover the possible distribution law. Experimental results show that these keywords with association role reveal previous power-law distribution. Then, based on the discovered power-law distribution, up-cutting and down-cutting points are presented to divide the association semantic into three layers. At the same time, theories of the hierarchy-cutting model are presented. Finally, examples of current core topic and permanent topics belonging to a domain are given. The experiments show that hierarchy-cutting points have high accuracy. The multilayer theory of association semantic can provide a theoretical support for knowledge recommendation with different particle sizes on ALNs.

Summary Recent years have witnessed the emergence of new types of semantic search engines which attempt to overcome the defects of the traditional search engines by providing different search patterns. A big question here is that in order to achieve the semantic search engines (SSEs), what type(s) of search patterns should SSEs support? To help seek one of the many possible answers, in this paper we start with classifying and comparing current search engines, particularly from the perspective of search patterns which consist of index structure, user profiles, and interaction mechanism. We then present a novel search pattern named ExNa by defining its model and basic operations in detail. To validate the ExNa search pattern, we develop a prototype search engine named KNOWLE, and the experimental results show that KNOWLE equipped with ExNa can improve both the efficiency of the entire system when compared with search engines of other search patterns. Copyright © 2016 John Wiley & Sons, Ltd.

Online comment has become a popular and efficient way for sellers to acquire feedback from customers and improve their service quality. However, some key issues need to be solved about evaluating and improving the hotel service quality based on online comments automatically, such as how to use the less trustworthy online comments, how to discover the quality defects from online comments, and how to recommend more feasible or economical evaluation indexes to improve the service quality based on online comments. To solve the above problems, this paper first improves fuzzy comprehensive evaluation (FCE) by importing trustworthy degree to it and proposes an automatic hotel service quality assessment method using the improved FCE, which can automatically get more trustworthy evaluation from a large amount of less trustworthy online comments. Then, the causal relations among evaluation indexes are mined from online comments to build the fuzzy cognitive map for the hotel service quality, which is useful to unfold the problematic areas of hotel service quality, and recommend more economical solutions to improving the service quality. Finally, both case studies and experiments are conducted to demonstrate that the proposed methods are effective in evaluating and improving the hotel service quality using online comments.

This paper advocates for a novel approach to recommend texts at various levels of difficulties based on a proposed method, the algebraic complexity of texts (ACT). Different from traditional complexity measures that mainly focus on surface features like the numbers of syllables per word, characters per word, or words per sentence, ACT draws from the perspective of human concept learning, which can reflect the complex semantic relations inside texts. To cope with the high cost of measuring ACT, the Degree-2 Hypothesis of ACT is proposed to reduce the measurement from unrestricted dimensions to three dimensions. Based on the principle of “mental anchor,” an extension of ACT and its general edition [denoted as extension of text algebraic complexity (EACT) and general extension of text algebraic complexity (GEACT)] are developed, which take keywords' and association rules' complexities into account. Finally, using the scores given by humans as a benchmark, we compare our proposed methods with linguistic models. The experimental results show the order GEACT>EACT>ACT> Linguistic models, which means GEACT performs the best, while linguistic models perform the worst. Additionally, GEACT with lower convex functions has the best ability in measuring the algebraic complexities of text understanding. It may also indicate that the human complexity curve tends to be a curve like lower convex function rather than linear functions.