Document Visual Question Answering (DocVQA) aims to generate answers by jointly understanding the textual, layout, and visual elements within document images.Although end-to-end vision-based generative methods have reduced dependency on OCR, they still struggle to achieve precise evidence localization when page semantics are complex and highly similar.However, existing research lacks an in-depth theoretical analysis of the question-driven semantic representation space, failing to fundamentally address the distinguishability problem among semantically similar pages.To fill this theoretical gap, we propose and prove that, given a specific question, each page possesses a unique semantic representation, and there exists a bijective mapping between the page and its unique semantics.Based on this theoretical foundation, we introduce the Flow-Based Page Unique Semantic Mapping Architecture (FUMA), which reconstructs evidence localization from similarity-based retrieval into precise selection on unique semantics.FUMA employs fine-grained cross-modal attention to extract discriminative cues and utilizes flow-based reversible transformations with likelihood regularization to learn bijective mappings, ensuring that each page obtains a unique semantic representation.Moreover, a multi-expert collaboration mechanism complementarily models fine-grained multimodal information within each page, achieving robust answer generation.Experimental results demonstrate that FUMA significantly outperforms existing methods in both evidence localization and answer generation.

The evaluation of Kleene Closure Property Path Queries (KPPQs) over RDF graphs presents a fundamental trade-off. Under the intuitive simple path semantics, which forbids node repetition, the problem is NP-complete. This forces practical systems to adopt existential path semantics, which admits tractable evaluation but produces result sets bloated with redundant and often semantically meaningless paths due to unnecessary detours through cycles. This paper introduces a novel indexing framework that effectively navigates this trade-off. Our key insight is that while enumerating all simple paths is intractable, a carefully selected, compact set of representative simple paths is sufficient to correctly and completely answer connectivity queries in practice. We propose the Node Clustered Index (NCI), which pre-materializes such paths—including the longest acyclic chains and fundamental cycles —transforming online query processing from a costly graph traversal into an efficient index lookup and expansion operation. Crucially, our method guarantees that every result conforms to simple path semantics. Extensive experiments demonstrate that our approach achieves near-perfect recall, outperforming state-of-the-art systems by up to an order of magnitude. This work establishes that for NP-complete path problems, seeking polynomial-time, empirically perfect approximations is a viable and powerful paradigm, delivering both performance and quality without compromise.

Background . Software vulnerabilities are increasing in complexity and scale, posing great security risks to many software systems. Extracting information about software vulnerabilities is a critical area of research that aims to identify and create a structured representation of vulnerability-related information. These structured data help software systems better understand vulnerabilities and provide security professionals with timely information to mitigate the impact of rapidly growing vulnerabilities while guiding future research to develop more secure systems. However, this process relies on the effectiveness of information extraction to transform manual vulnerability analysis from security experts to digital solutions. Despite its importance, the unique nature of vulnerability information and the fast pace at which machine learning-based extraction methods and techniques have evolved make it challenging to assess the current successes, failures, challenges, and opportunities within this research area. This study presents a systematic literature review aimed at clarifying this complex landscape. Methods . In this study, we conduct a systematic literature review (SLR) to explore existing research focusing on extracting information about software security vulnerabilities. We search for 829 primary studies on security vulnerability information extraction from seven widely used online digital libraries, focusing on top peer-reviewed journals and conferences published between 2001 and 2024. After applying our inclusion and exclusion criteria and the snowballing technique, we narrowed our selection to 87 studies for in-depth analysis and addressed four main research questions. We collect qualitative and quantitative data from each study, identifying 34 components such as research problems, methods, contributions, evaluation metrics, results, types of extracted vulnerability information, challenges, and limitations. We use meta-analysis, statistical machine learning, and text-mining techniques to identify themes, patterns, and trends across the primary studies and visualize findings. Result : The study provides an overview of the security vulnerability data landscape, identifies key resources, and guides efforts to improve vulnerability information extraction and analysis. The study finds a diverse landscape of learning algorithms used in security vulnerability information extraction, with Bidirectional Encoder Representations from Transformers (BERT), Long Short-term Memory (LSTM), and Support Vector Machine (SVM) being the most dominant. The study identifies key challenges, including feature engineering complexity, lack of a gold-standard corpus, preprocessing errors, generating accurate training data, addressing imbalanced data, multimodality fusion, and graph sparsity in security knowledge graphs. Insights for Future Research Directions . The study underscores the need for advanced extraction approaches, robust datasets, automated annotation methods, and advanced machine learning algorithms to improve the extraction of security vulnerability information. This study also suggests using large language models (LLMs) and transformer models to facilitate the automatic extraction of security-related words, terms, concepts, and phrases and introduce new filtering parameters for user requirements. We provide all our implementations; it can be found at https://bitbucket.org/slr-svie/vulnerability-information-extraction/src/master/ .

KG-based fact verification verifies the truthfulness of claims by retrieving evidence graphs from the knowledge graph. The faithful inference chains, which are precise relation paths between the mentioned entities and evidence entities, retrieve precise evidence graphs addressing poor performance and weak logic for fact verification. Due to the diversity of relation paths, existing methods rarely extract faithful inference chains. To alleviate these issues, we propose Multi-view Heterogeneous Graph with Causal Intervention (MHGCI): (i) We construct a Multi-view Heterogeneous Graph enhancing relation path extraction from the view of different mentioned entities. (ii) We propose a self-optimizing causal intervention model to generate assistant entities mitigating the out-of-distribution problem caused by counterfactual relations. (iii) We propose a grounding method to extract evidence graphs from the KG by faithful inference chains. Experiments on the public KG-based fact verification dataset FactKG demonstrate that our model provides precise evidence graphs and achieves state-of-the-art performance.

Heterophily emerges as a critical challenge in Graph Anomaly Detection (GAD). Recent studies reveal that neighborhood distributions, rather than heterophily itself, are the fundamental factor for the expressive power of Graph Neural Networks (GNNs). However, two key challenges remain unresolved. First, the overlap in neighborhood distributions between anomalous and normal nodes poses significant difficulties in distinguishing them effectively. Second, the dispersion in neighborhood distributions within the same class prevents the application of a fixed aggregation strategy to accommodate the diverse patterns within the class. To tackle the aforementioned challenges, we propose a novel Graph Neural Network model called Neighborhood Adaptive Aggregation and Spectral Tuning (NAAST-GNN). Specifically, we first design a neighborhood adaptive aggregation module that adjusts the message passing mechanism based on the predicted probabilities for different node classes, ensuring that nodes from distinct classes but with similar neighborhood distributions derive unique aggregated neighborhood information. We then present a spectral tuning module that dynamically selects and combines spectral filters based on the predicted neighborhood distribution, ensuring adaptability to the diverse neighborhood distributions of nodes within the same class. Comprehensive experimental results demonstrate that our method outperforms state-of-the-art baselines.

The recent progress in data and modeling techniques has significantly improved Named Entity Recognition (NER) for well-structured texts and high-resource languages. However, Roman Urdu, written in Latin script, poses unique challenges due to its informal structure, code-switching tendencies, and lack of standardized resources. The existing NER dataset for Roman Urdu suffers from inconsistent annotations and undefined entity boundaries, while traditional deep-learning models struggle with informal and multilingual text. To address these challenges, this study makes several key contributions. First, we re-annotate the Roman Urdu dataset to ensure tagging consistency, adopting the BIOES tagging scheme for entity boundary representation. Second, we fine-tuned multilingual pre-trained language models (PLMs), including mBERT, MuRIL, and XLM-R. Finally, we introduce XLM-R-IDCNN-BiLSTM-H, a novel architecture that combines transformer-based embeddings with advanced neural components, including Iterated Dilated Convolutional Neural Networks (IDCNN) for local dependencies, Bidirectional Long Short-Term Memory (BiLSTM) networks for bidirectional context, and highway networks for smooth information flow. We employ a conditional random field (CRF) to obtain the optimal tag sequence for classification. According to experimental findings, the proposed approach delivers state-of-the-art performance, which yielded an F1 score of 84.26% on the re-annotated Roman Urdu dataset.

Logical rules have been widely used for expressing schema knowledge in various practical applications. It is infeasible to handcraft rules from large knowledge graphs (KGs) and thus many methods have been proposed for learning rules automatically from KGs. However, it is largely ignored how to extract rules in a (target) KG from rules that already exist in some other (source) KGs. In this paper, we propose a framework for KG rule learning based on transfer learning. A major challenge for establishing such a framework is that a suitable alignment mechanism is required for mapping certain subgraph structures between predicates in the source KG and the target KG. Hence, our framework provides a new method for predicate mapping based on graph-structural similarity. The proposed framework can be used as a standalone rule learner but more importantly, it paves a new way for enhancing the state-of-the-art rule learners for large KGs. Extensive experiments are conducted to evaluate the new approach to rule learning, which shows that rules in smaller KGs can be effectively transferred to a large KG.

Abstract As the core operator of regular path query (RPQ), the Kleene closure is essentially a recursive operation based on predicates. It has more expressive power than first-order logic, but its algorithmic complexity is an NP-hard problem and expensive to execute. Most approaches implement Kleene closure via online recursive queries (inefficient for large-scale data), yet significant reusable intermediate results emerge during execution. This paper proposes an efficient query optimization method based on offline preprocessing and tree-structured strategies to address the challenges of NP-hard complexity and low execution efficiency caused by the recursive nature of Kleene closure RPQs (KRPQs). The core contributions of this work include: (i) recursive index tree: by precomputing and materializing intermediate results of predicate paths offline, Kleene closure queries are transformed into non-recursive SPARQL queries. This allows direct retrieval of answer branches from the preconstructed index tree, eliminating redundant computations and reducing online complexity. (ii) Selectivity-driven dynamic optimization: a heuristic cost model estimates intermediate result sizes to guide query plan generation, minimizing redundant operations. (3) Query decomposition tree for nested KRPQs: a hierarchical decomposition algorithm processes nested Kleene closures layer-by-layer from inner to outer expressions, overcoming limitations of existing methods in handling complex nested structures. Experiments on large-scale RDF datasets demonstrate that the method reduces average response times by 50%–70% for single-predicate and expression-based Kleene closure queries compared to systems like Virtuoso, Jena, and KRPQ, with over 60% improvement for complex nested queries. Additionally, multithreaded preprocessing and path pool optimizations achieve up to a 76$\times $ reduction in index construction time. Experiments show that the method can significantly reduce the time of KRPQs based on large RDF graphs.

ABSTRACT Sarcasm detection is a critical and challenging task in sentiment analysis, particularly for low‐resource languages like Urdu, where limited annotated data, linguistic complexity, and subtle contextual cues hinder accurate classification. Traditional machine learning methods often fail to capture the nuanced and often contradictory nature of sarcastic expression. To address these challenges, this paper presents a comprehensive and computationally efficient framework for Urdu sarcasm detection. We first mitigate severe class imbalance through strategic down‐sampling and back‐translation‐based data augmentation. We then conduct extensive benchmarking of traditional deep learning architectures against fine‐tuned pre‐trained language models, including multilingual, monolingual, and Twitter‐specific variants. Building on these insights, we propose DeepSarc, a novel hybrid model that integrates the contextual embeddings from XLM‐T, the multi‐scale feature extraction capabilities of dilated convolutional neural networks (DCNNs), and the sequential dependency modeling of bidirectional long short‐term memory (BiLSTM) networks. While slightly more computationally intensive than simpler alternatives, DeepSarc achieves a state‐of‐the‐art F1‐score, significantly outperforming existing approaches. Our results establish a new benchmark for sarcasm detection in low‐resource languages and provide a scalable, high‐performance framework adaptable to diverse linguistic contexts.

In software development, the raw requirements proposed by users are frequently incomplete, which impedes the complete implementation of software functionalities. With the emergence of large language models, the exploration of generating software through user requirements has attracted attention. Recent methods with the top-down waterfall model employ a questioning approach for requirement completion, attempting to explore further user requirements. However, users, constrained by their domain knowledge, result in a lack of effective acceptance criteria during the requirement completion, failing to fully capture the implicit needs of the user. Moreover, the cumulative errors of the waterfall model can lead to discrepancies between the generated code and user requirements. The Agile methodologies reduce cumulative errors of the waterfall model through lightweight iteration and collaboration with users, but the challenge lies in ensuring semantic consistency between user requirements and the code generated by the agent. To address these challenges, we propose AgileGen, an agile-based generative software development through human-AI teamwork. Unlike existing questioning agents, AgileGen adopts a novel collaborative approach that breaks free from the constraints of domain knowledge by initiating the end-user perspective to complete the acceptance criteria. By introducing the Gherkin language, AgileGen attempts for the first time to use testable requirement descriptions as a bridge for semantic consistency between requirements and code, aiming to ensure that software products meet actual user requirements by defining user scenarios that include acceptance criteria. Additionally, we innovate in the human-AI teamwork model, allowing users to participate in decision-making processes they do well and significantly enhancing the completeness of software functionality. To ensure semantic consistency between requirements and generated code, we derive consistency factors from Gherkin to drive the subsequent software code generation. Finally, to improve the reliability of user scenarios, we also introduce a memory pool mechanism, collecting user decision-making scenarios and recommending them to new users with similar requirements. AgileGen, as a user-friendly interactive system, significantly outperformed existing best methods by 16.4% and garnered higher user satisfaction.

Textual vulnerability descriptions (TVDs) document phrase-based concepts of security vulnerabilities, such asVulnerability TypeandAttack Vector. For example, a TVD might state, “Abuffer overflow vulnerability(Vulnerability Type) in the image processing library allowsremote code execution(Impact).” The manual creation of TVDs often leads to missing phrase-based concepts, hindering severity and exploitability assessments. Large language models (LLMs) are used to complete these missing phrase-based concepts, but they suffer from hallucinations, generating plausible yet inaccurate completion that burden developers. Existing methods mitigate hallucinations by aligning results with knowledge bases, which fails when hallucinated predictions appear believable due to knowledge base similarity. To address this, we propose leveraging 1-hop relationships between phrase-based concepts in a TVD to filter hallucinated predictions. Our approach consists of three modules: 1-hop relationship generation, which enhances prediction diversity by introducing TVD perturbations to counter the LLM’s tendency to generate outputs similar to the context; semantic augmentation using CWE descriptions to enrich the 1-hop relationships and phrase-based concepts, addressing the challenge of capturing concise technical semantics; and chain-of-thought (CoT) reasoning for filtering, which guides the LLM to articulate and verify logical connections, ensuring contextually accurate and consistent 1-hop relationships. Experiments on a TVD dataset show that our method reduces hallucinations, improves completion accuracy from 49% to 62%, and enhances the reliability of LLM-generated phrase-based concepts.

Augmenting missing key aspects in Textual Vulnerability Descriptions (TVDs) is crucial for effective vulnerability analysis. For instance, in TVDs, key aspects includeAttack Vector,Vulnerability Type, among others. These key aspects help security engineers understand and address the vulnerability in a timely manner. For software with a large user base (non-long-tail software), augmenting these missing key aspects has significantly advanced vulnerability analysis and software security research. However, software instances with a limited user base (long-tail software) often get overlooked due to inconsistency software names, TVD limited avaliability, and domain-specific jargon, which complicates vulnerability analysis and software repairs. In this paper, we introduce a novel software feature inference framework designed to augment the missing key aspects of TVDs for long-tail software. Firstly, we tackle the issue of non-standard software names found in community-maintained vulnerability databases by cross-referencing government databases with Common Vulnerabilities and Exposures (CVEs). Next, we employ Large Language Models (LLMs) to generate the missing key aspects. However, the limited availability of historical TVDs restricts the variety of examples. To overcome this limitation, we utilize the Common Weakness Enumeration (CWE) to classify all TVDs and select cluster centers as representative examples. To ensure accuracy, we present Natural Language Inference (NLI) models specifically designed for long-tail software. These models identify and eliminate incorrect responses. Additionally, we use a wiki repository to provide explanations for proprietary terms. Our evaluations demonstrate that our approach significantly improves the accuracy of augmenting missing key aspects of TVDs for log-tail software from 0.27 to 0.56 (+107%). Interestingly, the accuracy of non-long-tail software also increases from 64% to 71%. As a result, our approach can be useful in various downstream tasks that require complete TVD information.

In graph anomaly detection (GAD), the fact that anomalous nodes usually exhibit high heterophily, while most Graph Neural Networks (GNNs) have homophily assumptions, leads to poor performance. Many studies have attempted to solve this problem by employing a set of graph filters covering various frequencies. Their ultimate goal is to design the most appropriate spectral filter to capture the complex signals generated by normals and anomalies. The critical aspect lies in the fusion of information from filters with different frequency response functions. However, existing methods lack a clear indicator to guide the fusion of information at different frequencies. In this paper, we find that local homophily is a valuable metric for assessing the weights of high- and low-frequency information at the node level, and explicitly point out that the accuracy of local homophily is positively correlated with the accuracy of anomaly detection. Moreover, we unveil the phenomenon of camouflage in anomalous, wherein these nodes disguise themselves by making their features resemble those of surrounding normals.

Time-Sensitive Question Answering (TSQA) is to answer questions qualified for a certain timestamp based on the given document. It is split into easy and hard modes depending on whether the document contain time qualifiers mentioned in the question. While existing models have performed well on easy mode, their performance is significant reduced for answering hard time-sensitive questions, whose time qualifiers are implicit in the document. An intuitive idea is to match temporal events in the given document by treating time-sensitive question as a temporal event of missing objects. However, not all temporal events extracted from the document have explicit time qualifiers. In this paper, we propose an Event-AL framework, in which a graph pruning model is designed to locate the timespan of implicit temporal events by capturing temporal relation between events. Moreover, we present an abductive reasoning module to determine proper objects while providing explanations. Besides, as the same relation may be scattered throughout the document in diverse expressions, a relation-based prompt is introduced to instructs LLMs in extracting candidate temporal events. We conduct extensive experiment and results show that Event-AL outperforms strong baselines for hard time-sensitive questions, with a 12.7% improvement in EM scores. In addition, it also exhibits great superiority for multi-answer and beyond hard time-sensitive questions.

Multiple-choice question answering (MCQA) becomes particularly challenging when all choices are relevant to the question and are semantically similar. Yet this setting of MCQA can potentially provide valuable clues for choosing the right answer. Existing models often rank each choice separately, overlooking the context provided by other choices. Specifically, they fail to leverage the semantic commonalities and nuances among the choices for reasoning. In this paper, we propose a novel MCQA model by differentiating choices through identifying and eliminating their commonality, called DCQA. Our model captures token-level attention of each choice to the question, and separates tokens of the question attended to by all the choices (i.e., commonalities) from those by individual choices (i.e., nuances). Using the nuances as refined contexts for the choices, our model can effectively differentiate choices with subtle differences and provide justifications for choosing the correct answer. We conduct comprehensive experiments across five commonly used MCQA benchmarks, demonstrating that DCQA consistently outperforms baseline models. Furthermore, our case study illustrates the effectiveness of the approach in directing the attention of the model to more differentiating features.

Syntactically controlled paraphrase generation is to generate diverse sentences that have the same semantics as the given original sentence but conform to the target syntactic structure. An optimal opportunity to enhance diversity is to make word substitutions during rephrasing based on syntactic control. Existing unsupervised methods have made great progress in syntactic control, but the generated paraphrases rarely have substitutions due to the limitation of training data. In this paper, we propose a Diversity syntactically controlled Paraphrase generation framework (DiPara), in which a novel training strategy is designed to obtain semantic sentences while using the given sentence as training objects. As diverse words vary the syntactic structure around them, we propose a phrase-aware attention mechanism to capture the syntactic structure associated with the current word. To achieve it, the linearized triple sequence is introduced to represent structure singly. Experiment results on two datasets show that DiPara outperforms strong baselines, especially diversity (Self-BLEU4) is improved by 10.18% in ParaNMT-Small.

Social intelligence is essential for understanding complex human expressions and social interactions.While large multimodal models (LMMs) have demonstrated remarkable performance in social intelligence question answering (SIQA), they are still inclined to generate responses relying on language priors and ignoring the relevant context due to the dominant prevalence of text-based data in the pretraining stage.To interpret the aforementioned language bias of LMMs, we employ a structure causal model and posit that counterfactual reasoning can mitigate the bias by avoiding spurious correlations between LMMs' internal commonsense knowledge and the given context.However, it is costly and challenging to construct multimodal counterfactual samples.To tackle the above challenges, we propose an output Distribution Calibration network with Virtual Counterfactual (DCVC) data augmentation framework.DCVC devises a novel output distribution calibration network to mitigate the impact of negative language biases while preserving beneficial priors.Perturbations are introduced to the output distributions of LMMs to simulate the distribution shifts from counterfactual manipulations of the context, which is employed to construct counterfactual augmented data virtually.Experiments on multiple datasets demonstrate the effectiveness and generalizability of our proposed method.

Our research reveals that multi-interest recommendation models in the matching stage tend to exhibit an under-clustered item embedding space, which leads to a low discernibility between items and hampers item retrieval. This highlights the necessity for item embedding enhancement. However, item attributes, which serve as effective side information for enhancement, are either unavailable or incomplete in many public datasets due to the labor-intensive nature of manual annotation tasks. This dilemma raises two meaningful questions: 1. Can we bypass manual annotation and directly simulate complete attribute information from the interaction data? And 2. If feasible, how can we simulate attributes with high accuracy and low complexity in the matching stage?

Textbook question answering (TQA) task aims to infer answers for given questions from a multimodal context, including text and diagrams. The existing studies have aggregated intramodal semantics extracted from a single modality but have yet to capture the intermodal semantics between different modalities. A major challenge in learning intermodal semantics is maintaining lossless intramodal semantics while bridging the gap of semantics caused by heterogeneity. In this article, we propose an intermodal relation-aware heterogeneous graph network (IMR-HGN) to extract the intermodal semantics for TQA, which aggregates different modalities while learning features rather than representing them independently. First, we design a multidomain consistent representation (MDCR) to eliminate semantic gaps by capturing intermodal features while maintaining lossless intramodal semantics in multidomains. Furthermore, we present neighbor-based relation inpainting (NRI) to reduce semantic ambiguity via repairing fuzzy relations with correlations of relations. Finally, we propose hierarchical multisemantics aggregation (HMSA) to guarantee the completeness of semantics by aggregating features of nodes and relations with a reconstruction network (RN). Experimental results show that IMR-HGN could extract the intermodal semantics of answers, achieving a 2.16% improvement on the validation set of the TQA dataset and a 3.04% increase on the test set of the AI2D dataset.

Pre-trained language models (PLMs) have recently enabled rapid progress on sentiment classification under the pre-train and fine-tune paradigm, where the fine-tuning phase aims to transfer the factual knowledge learned by PLMs to sentiment classification. However, current fine-tuning methods ignore the risk that PLMs cause the problem of sentiment bias, that is, PLMs tend to inject positive or negative sentiment from the contextual information of certain entities (or aspects) into their word embeddings, leading them to establish spurious correlations with labels. In this paper, we propose an adaptive Gumbel-attacked classifier that immunes sentiment bias from an adversarial-attack perspective. Due to the complexity and diversity of sentiment bias, we construct multiple Gumbel-attack expert networks to generate various noises from mixed Gumbel distribution constrained by mutual information minimization, and design an adaptive training framework to synthesize complex noise by confidence-guided controlling the number of expert networks. Finally, we capture these noises that effectively simulate sentiment bias based on the feedback of the classifier, and then propose a multi-channel parameter updating algorithm to strengthen the classifier to recognize these noises by fusing the parameters between the classifier and each expert network. Experimental results illustrate that our method significantly reduced sentiment bias and improved the performance of sentiment classification.

Document-level Relation Extraction (DocRE)intends to extract relationships from documents.Some works introduce logic constraints into DocRE, addressing the issues of opacity and weak logic in original DocRE models.However, they only focus on forward logic constraints and the rules mined in these works often suffer from pseudo rules with high standardconfidence but low support.In this paper, we proposes Bidirectional Constraints of Beta Rules(BCBR), a novel logic constraint framework.BCBR first introduces a new rule miner which model rules by beta contribtion.Then forward and reverse logic constraints are constructed based on beta rules.Finally, BCBR reconstruct rule consistency loss by bidirectional constraints to regulate the output of the DocRE model.Experiments show that BCBR outperforms original DocRE models on relation extraction performance (∼2.7 F1) and logic consistency(∼3.1 Logic).Furthermore, BCBR consistently outperforms two other logic constraint frameworks.Our code is available at https://github.com/Louisliu1999/BCBR.

Named entity recognition (NER) is the task of identifying and classifying named entities from texts. NER can benefit from linguistic dependency information, yet existing NER models can only utilize such information on datasets where dependency annotations are readily available. Dependency parsing (DP) models can be used to generate annotations, which are trained independent of the NER task and can cause error propagation to NER. In this paper, we propose a joint NER and DP model through multi-task learning, which allows the NER and DP modules to benefit from the joint training and provides an end-to-end solution to dependency-guided NER. Our model JOINDER uses a shared contextualized embedder, a word encoder, a biaffine dependency classifier, and a multi-hop dependency-guided NER. Experiments on several standard datasets in four languages show the effectiveness of joint learning and the outstanding performance of JOINDER compared to existing models. Moreover, our model can transfer dependency knowledge to other datasets with no dependency annotat.

Multi-interest recommendation methods extract multiple interest vectors to represent the user comprehensively. Despite their success in the matching stage, previous works overlook the long-tail problem. This results in the model excelling at suggesting head items, while the performance for tail items, which make up more than 70% of all items, remains suboptimal. Hence, enhancing the tail item recommendation capability holds great potential for improving the performance of the multi-interest model.

Abstract Named entity recognition (NER) is a fundamental part of other natural language processing tasks such as information retrieval, question answering systems and machine translation. Progress and success have already been achieved in research on the English NER systems. However, the Urdu NER system is still in its infancy due to the complexity and morphological richness of the Urdu language. Existing Urdu NER systems are highly dependent on manual feature engineering and word embedding to capture similarity. Their performance lags if the words are previously unknown or infrequent. The feature-based models suffer from complicated feature engineering and are often highly reliant on external resources. To overcome these limitations in this study, we present several deep neural approaches that automatically learn features from the data and eliminate manual feature engineering. Our extension involved convolutional neural network to extract character-level features and combine them with word embedding to handle out-of-vocabulary words. The study also presents a tweets dataset in Urdu, annotated manually for five named entity classes. The effectiveness of the deep learning approaches is demonstrated on four benchmarks datasets. The proposed method demonstrates notable progress upon current state-of-the-art NER approaches in Urdu. The results show an improvement of 6.26% in the F1 score.

Software vulnerabilities, once disclosed, can be documented in vulnerability databases, which have great potential to advance vulnerability analysis and security research. People describe the key characteristics of software vulnerabilities in natural language mixed with domain-specific names and concepts. This textual nature poses a significant challenge for the automatic analysis of vulnerability knowledge embedded in text. Automatic extraction of key vulnerability aspects is highly desirable but demands significant effort to manually label data for model training. In this article, we propose unsupervised methods to label and extract important vulnerability concepts in textual vulnerability descriptions (TVDs). We focus on six types of phrase-based vulnerability concepts (vulnerability type, vulnerable component, root cause, attacker type, impact, and attack vector) as they are much more difficult to label and extract than name- or number-based entities (i.e., vendor, product, and version). Our approach is based on a key observation that the same-type of phrases, no matter how they differ in sentence structures and phrase expressions, usually share syntactically similar paths in the sentence parsing trees. Specifically, we present a source-target neural architecture that learns the Part-of-Speech (POS) tagging to identify a token’s functional role within TVDs, where the source neural model is trained to capture common features found in the TVD corpus, and the target model is trained to identify linguistically malformed words specific to the security domain. Our evaluation confirms that the proposed tagger outperforms (4.45%–5.98%) the taggers designed on natural language notions and identifies a broad set of TVDs and natural language contents. Then, based on the key observations, we propose two path representations (absolute paths and relative paths) and use an auto-encoder to encode such syntactic similarities. To address the discrete nature of our paths, we enhance the traditional Variational Auto-encoder (VAE) with Gumble-Max trick for categorical data distribution and thus create a Categorical VAE (CaVAE). In the latent space of absolute and relative paths, we further apply unsupervised clustering techniques to generate clusters of the same-type of concepts. Our evaluation confirms the effectiveness of our CaVAE, which achieves a small (85.85) log-likelihood for encoding path representations and the accuracy (83%–89%) of vulnerability concepts in the resulting clusters. The resulting clusters accurately label six types of vulnerability concepts from a TVD corpus in an unsupervised way. Furthermore, these labeled vulnerability concepts can be mapped back to the corresponding phrases in the original TVDs, which produce labels of six types of vulnerability concepts. The resulting labeled TVDs can be used to train concept extraction models for other TVD corpora. In this work, we present two concept extraction methods (concept classification and sequence labeling model) to demonstrate the utility of the unsupervisedly labeled concepts. Our study shows that models trained with our unsupervisedly labeled vulnerability concepts outperform (3.9%–5.14%) those trained with the two manually labeled TVD datasets from previous work due to the consistent boundary and typing by our unsupervised labeling method.

Linjuan Wu, Shaojuan Wu, Xiaowang Zhang, Deyi Xiong, Shizhan Chen, Zhiqiang Zhuang, Zhiyong Feng. Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2022.

The relation classification is to identify semantic relations between two entities in a given text. While existing models perform well for classifying inverse relations with large datasets, their performance is significantly reduced for few-shot learning. In this paper, we propose a function words adaptively enhanced attention framework (FAEA) for few-shot inverse relation classification, in which a hybrid attention model is designed to attend class-related function words based on meta-learning. As the involvement of function words brings in significant intra-class redundancy, an adaptive message passing mechanism is introduced to capture and transfer inter-class differences.We mathematically analyze the negative impact of function words from dot-product measurement, which explains why the message passing mechanism effectively reduces the impact. Our experimental results show that FAEA outperforms strong baselines, especially the inverse relation accuracy is improved by 14.33% under 1-shot setting in FewRel1.0.

Knowledge Graph (KG) representation learning aims to encode both entities and relations into a continuous low-dimensional vector space. Most existing methods only concentrate on learning representations from structural triples in Euclidean space, which cannot well exploit the rich semantic information with hierarchical structure in KGs. In this paper, we propose a novel DataType-aware hyperbolic knowledge representation learning model called DT-GCN, which has the advantage of fully embedding attribute values of data types information. We refine data types into five primitive modalities, including integer, double, Boolean, temporal, and textual. For each modality, an encoder is specifically designed to learn its embedding. In addition, we define a unified space based on Euclidean, spherical, and hyperbolic space, which is a continuous curvature space that combines advantages of three different spaces. Extensive experiments on both synthetic and real-world datasets show that our model is consistently better than the state-of-the-art models. The average performance is improved by 2.19% and 3.46% than the optimal baseline model on node classification and link prediction tasks, respectively. The results of ablation experiments demonstrate the advantages of embedding data types information and leveraging the unified space.

Difficult samples of the minority class in imbalanced text classification are usually hard to be classified as they are embedded into an overlapping semantic region with the majority class.In this paper, we propose a Mutual Information constrained Semantically Oversampling framework (MISO) that can generate anchor instances to help the backbone network determine the re-embedding position of a non-overlapping representation for each difficult sample.MISO consists of (1) a semantic fusion module that learns entangled semantics among difficult and majority samples with an adaptive multi-head attention mechanism, (2) a mutual information loss that forces our model to learn new representations of entangled semantics in the non-overlapping region of the minority class, and (3) a coupled adversarial encoder-decoder that fine-tunes disentangled semantic representations to remain their correlations with the minority class, and then using these disentangled semantic representations to generate anchor instances for each difficult sample.Experiments on a variety of imbalanced text classification tasks demonstrate that anchor instances help classifiers achieve significant improvements over strong baselines.

Query graph as a junction of semantic parsing in question answering over knowledge bases (KBQA) connects questions and logical queries. Though query graph consists of rich information such as structure, relation, etc., the current KBQA's models mainly utilize limited relation information in a naive way. It is not easy to learn the representation of a query graph with that information due to the heterogeneity of the query graph and intricate correlation of relations. In this paper, we propose a Global Semantic-based Message Passing (GSMP) model to model the global semantics of a query graph from its structure and relation information. In GSMP, we present a recurrent-based relational graph convolutional network (RGCN) to capture heterogeneous query graphs where the recurrent unit improves the capability of RGCN in processing small-scale query graphs. Moreover, we present a contextual-based method to remove ambiguity caused by intricate correlations where the contextual adjacency of relations optimizes relation representation. Finally, we present a nonlinear gate-based encoder to learning the representation of questions' syntactic tree, as the structure information of questions, for better matching the global semantics of query graphs. Experiments evaluated on benchmarks show that our model outperforms off-the-shelf models.

People usually describe the key characteristics of software vulnerabilities in natural language mixed with domain-specific names and concepts. This textual nature poses a significant challenge for the automatic analysis of vulnerabilities. Automatic extraction of key vulnerability aspects is highly desirable but demands significant effort to manually label data for model training. In this paper, we propose an unsupervised approach to label and extract important vulnerability concepts in textural vulnerability descriptions (TVDs). We focus on three types of phrase-based vulnerability concepts (root cause, attack vector, and impact) as they are much more difficult to label and extract than name- or number-based entities (i.e., vendor, product, and version). Our approach is based on a key observation that the same-type of phrases, no matter how they differ in sentence structures and phrase expressions, usually share syntactically similar paths in the sentence parsing trees. Therefore, we propose two path representations (absolute paths and relative paths) and use an auto-encoder to encode such syntactic similarities. To address the discrete nature of our paths, we enhance traditional Variational Auto-encoder (VAE) with Gumble-Max trick for categorical data distribution, and thus creates a Categorical VAE (CaVAE). In the latent space of absolute and relative paths, we further use FIt-TSNE and clustering techniques to generate clusters of the same-type of concepts. Our evaluation confirms the effectiveness of our CaVAE for encoding path representations and the accuracy of vulnerability concepts in the resulting clusters. In a concept classification task, our unsupervisedly labeled vulnerability concepts outperform the two manually labeled datasets from previous work.

In this paper, we study the problem of part-of-speech (POS) tagging for security vulnerability descriptions (SVD). In contrast to newswire articles, SVD often contains a high-level natural language description of the text composed of mixed language studded with codes, domain-specific jargon, vague language, and abbreviations. Moreover, training data dedicated to security vulnerability research is not widely available. Existing neural network-based POS tagging has often relied on manually annotated training data or applying natural language processing (NLP) techniques, suffering from two significant drawbacks. The former is extremely time-consuming and requires labor-intensive feature engineering and expertise. The latter is inadequate to identify linguistically-informed words specific to the SVD domain. In this paper, we propose an automatic approach to assign POS tags to tokens in SVD. Our approach uses the character-level representation to automatically extract orthographic features and unsupervised word embeddings to capture meaningful syntactic and semantic regularities from SVD. The character level representations are then concatenated with the word embedding as a combined feature, which is then learned and used to predict the POS tagging. To deal with the issue of the poor availability of annotated security vulnerability data, we implement a finetuning approach. Our approach provides public access to a POS annotated corpus of ~8M tokens, which serves as a training dataset in this domain. Our evaluation results show a significant improvement in accuracy (17.72%-28.22%) of POS tagging in SVD over the current approaches.

Abstract Ontology-mediated querying (OMQ) provides a paradigm for query answering according to which users not only query records at the database but also query implicit information inferred from ontology. A key challenge in OMQ is that the implicit information may be infinite, which cannot be stored at the database and queried by off -the -shelf query engine. The commonly adopted technique to deal with infinite entailments is query rewriting, which, however, comes at the cost of query rewriting at runtime. In this work, the partial materialization method is proposed to ensure that the extension is always finite. The partial materialization technology does not rewrite query but instead computes partial consequences entailed by ontology before the online query. Besides, a query analysis algorithm is designed to ensure the completeness of querying rooted and Boolean conjunctive queries over partial materialization. We also soundly and incompletely expand our method to support highly expressive ontology language, OWL 2 DL. Finally, we further optimize the materialization efficiency by role rewriting algorithm and implement our approach as a prototype system SUMA by integrating off-the-shelf efficient SPARQL query engine. The experiments show that SUMA is complete on each test ontology and each test query, which is the same as Pellet and outperforms PAGOdA. Besides, SUMA is highly scalable on large datasets.

Imbalanced text classification, as practical and essential text classification, is the task to learn labels or categories for imbalanced text data. Existing imbalanced text classification approaches are mostly based on the Imbalance Ratio (i.e. ratio of sizes between categories). Recently, some researchers verified that the imbalance ratio severely affects the performance of classifiers when intrinsic characteristics of data such as class overlapping and small disjuncts occur. However, since the distribution of real-world data is unknown, it is difficult to describe above intrinsic characteristics directly. In this paper, we transform the unknown distribution of data into a graph model and present a graph-based imbalance index named GIR to predict the impact of imbalanced text data on classification performance. Firstly, we introduce an environmental factor that makes the imbalance index sensitive to the intrinsic characteristics of data. Secondly, we propose a graph-based method to calculate this environmental factor. Finally, we use the imbalance index to analyze the performances of imbalanced learning methods and the impact of imbalanced data on text classifiers. The experimental results evaluated on both synthetic data sets and real-world data sets demonstrate the effectiveness of our approach.

Emotional intelligence is a crucial part for human-machine dialogue system. However, the existing research on dialogue mainly faces three problems: (1) focus on the content level of each response while ignoring the impact of emotional factors in the multi-turn dialogue; (2) lacking scalability and adaptability is that only the emotion categories specified by users are generated in a single-turn dialogue; (3) it is difficult to capture and perceive fine-grained emotions and the speaker's emotional state according to the emotional context. To address these problems, we propose an emotional expression model in multi-turn dialogue (EmoEM), which combines emotion-semantic graph with multitask learning mechanism, applying the dialogue generator based on seq2seq network and graph convolution network (GCN) to generate more natural and personalized emotional responses in a structured manner. Generally, EmoEM considers constructing emotion-semantic graph to describe explicit and implicit emotions dynamically. Then, the emotion-semantic graph is applied to the dialogue generator based on the seq2seq neural network, mainly to improve the semantic consistency and text quality in multi-turn dialogue. Moreover, multi-task learning mechanism is introduced to enhance the emotional expression of the text and obtain expected emotional responses. The experimental results show that EmoEM outperforms several baselines in BLEU, diversity and emotional expression.

In this paper, we present a neural network (InteractionNN) for sparse predictive analysis where hidden features of sparse data can be learned by multilevel feature interaction. To characterize multilevel interaction of features, InteractionNN consists of three modules, namely, nonlinear interaction pooling, layer-lossing, and embedding. Nonlinear interaction pooling (NI pooling) is a hierarchical structure and, by shortcut connection, constructs low-level feature interactions from basic dense features to elementary features. Layer-lossing is a feed-forward neural network where high-level feature interactions can be learned from low-level feature interactions via correlation of all layers with target. Moreover, embedding is to extract basic dense features from sparse features of data which can help in reducing our proposed model computational complex. Finally, our experiment evaluates on the two benchmark datasets and the experimental results show that InteractionNN performs better than most of state-of-the-art models in sparse regression.

Existential rules, a family of expressive ontology languages, inherit desired expressive and reasoning properties from both description logics and logic programming. On the other hand, forgetting is a well studied operation for ontology reuse, obfuscation and analysis. Yet it is challenging to establish a theory of forgetting for existential rules. In this paper, we lay the foundation for a theory of forgetting for existential rules by developing a novel notion of unfolding. In particular, we introduce a definition of forgetting for existential rules in terms of query answering and provide a characterisation of forgetting by the unfolding. A result of forgetting may not be expressible in existential rules, and we then capture the expressibility of forgetting by a variant of boundedness. While the expressibility is undecidable in general, we identify a decidable fragment. Finally, we provide an algorithm for forgetting in this fragment.

The pattern satisfiability is a fundamental problem for SPARQL. This paper provides a complete analysis of decidability/undecidability of satisfiability problems for SPARQL 1.1 patterns. A surprising result is the undecidability of satisfiability for SPARQL 1.1 patterns when only AND and MINUS are expressible. Also, it is shown that any fragment of SPARQL 1.1 without expressing both AND and MINUS is decidable. These results provide a guideline for future SPARQL query language design and implementation.

In this paper, we present an embedding-based framework (TrQuery) for recommending solutions of a SPARQL query, including approximate solutions when exact querying solutions are not available due to incompleteness or inconsistencies of real-world RDF data. Within this framework, embedding is applied to score solutions together with edit distance so that we could obtain more fine-grained recommendations than those recommendations via edit distance. For instance, graphs of two querying solutions with a similar structure can be distinguished in our proposed framework while the edit distance depending on structural difference becomes unable. To this end, we propose a novel score model built on vector space generated in embedding system to compute the similarity between an approximate subgraph matching and a whole graph matching. Finally, we evaluate our approach on large RDF datasets DBpedia and YAGO, and experimental results show that TrQuery exhibits an excellent behavior in terms of both effectiveness and efficiency.

The study of paraconsistent reasoning with ontologies is especially important for the Semantic Web since knowledge is not always perfect within it. However, classical OWL reasoners cannot support reasoning with inconsistent ontologies. In this article, the authors present a plugin-based framework called prose to provide rich paraconsistent reasoning services for OWL ontologies, whose architecture contains the three following parts: a classical OWL reasoner, a multi-valued transformer, and an OWL API connecting with them. Within the proposed framework prose, they implement different multi-valued paraconsistent reasoning in the OWL. Moreover, they select three popular classical OWL reasoners and two typical kinds of reasoning services for users. As the authors excepted, prose does exactly enable current classical OWL reasoners to tolerate inconsistency in a simple and convenient way. Finally, they evaluate the three reasoners in a united framework (prose) and, as a result, those results can amend the analysis of the three reasoners on inconsistent ontologies.

In this paper, we present a federated path querying language (FPQ) as itself an extension of the nested regular path querying language with adding an axis operator to support the federation of RDF dataset and relational database. We have proven that FPQ has more expressive power than the nested regular path query language (not to mention regular path querying language). It enjoys the same computational complexity as the regular path query language and its additional expressivity can be exactly used to characterize the conjunction and federation of nested regular path queries. Moreover, we discuss the expressivity of various fragments of FPQ and implement FPQ. Finally, we present an application scenario related to the car-pooling services in real life.

We propose an efficient distributed method for answering regular path queries (RPQs) on large-scale RDF graphs using partial evaluation. In local computation, we devise a dynamic programming approach to evaluate local and partial answers of an RPQ on each computing site in parallel. In the assembly phase, an automata-based algorithm is proposed to assemble the partial answers of the RPQ into the final results. The experiments on benchmark RDF graphs show that our method outperforms the state-of-the-art message passing methods by up to an order of magnitude.

In this paper, we analyze some basic features of SPARQL queries from practical world in a statistical way. In particular, we focus on three statistic features including the occurrence frequency of triple patterns, fragments, and well-designed patterns and four semantic features including monotonicity, non-monotonicity, weak monotonicity and satisfiability. All the features contribute to characterize SPARQL queries in different dimensions. We hope that this statistical analysis would provide some useful observations for researchers and engineers who are interested in what real-word SPARQL queries look like, so that they could develop some practical heuristics for processing SPARQL queries, as well as build SPARQL query processing engines and benchmarks. In addition, our research facilitates to reduce scope of the problems by avoiding some cases that may not occur in practice.

The satisfiability problem for SPARQL 1.0 patterns is undecidable in general, since the relational algebra can be emulated using such patterns. The goal of this paper is to delineate the boundary of decidability of satisfiability in terms of the constraints allowed in filter conditions. The classes of constraints considered are bound-constraints, negated bound- constraints, equalities, nonequalities, constant-equalities, and constant-nonequalities. The main result of the paper can be summarized by saying that, as soon as inconsistent filter conditions can be formed, satisfiability is undecidable. The key insight in each case is to find a way to emulate the set difference operation. Undecidability can then be obtained from a known undecidability result for the algebra of binary relations with union, composition, and set difference. When no inconsistent filter conditions can be formed, satisfiability is decidable by syntactic checks on bound variables and on the use of literals. Although the problem is shown to be NP-complete, it is experimentally shown that the checks can be implemented efficiently in practice. The paper also points out that satisfiability for the so-called ‘well-designed’ patterns can be decided by a check on bound variables and a check for inconsistent filter conditions.

DL-Lite is an important family of description logics. Recently, there is an increasing interest in handling inconsistency in DL-Lite as the constraint imposed by a TBox can be easily violated by assertions in ABox in DL-Lite. In this paper, we present a distance-based paraconsistent semantics based on the notion of feature in DL-Lite, which provides a novel way to rationally draw meaningful conclusions even from an inconsistent knowledge base. Finally, we investigate several important logical properties of this entailment relation based on the new semantics and show its promising advantages in non-monotonic reasoning for DL-Lite.

Navigational queries on graph databases return binary relations over the nodes of the graph. The calculus of relations, popularized by Tarski, serves as a natural benchmark for first-order navigational querying. Recently, nested regular expressions (nre's) have been proposed to extend navigational querying to resource description framework graphs, i.e. ternary relations. This paper investigates the expressiveness of nre's and their relationship to basic SPARQL queries. An elegant proof is given to the effect that nre queries are already expressible as basic SPARQL queries. This result takes exception of nre's involving Kleene star (transitive closure), but on the other hand it holds even when extending nre's with negation (complementation). The resulting language of ‘star-free nre's with negation (sfnre¬)’ can in fact be captured by a precisely delineated fragment of SPARQL, called Tarski-SPARQL. The resulting language is also compared with an alternative extension that adds negation in the form of the difference operator. While sfnre¬ queries are subsumed by first-order logic with three variables (FO3), it is shown that some natural FO3 queries are not expressible in nre¬, even when allowing transitive closure.

Motivated by the current interest in languages for expressing path queries to graph databases, this paper proposes to investigate Walk Logic (WL): the extension of first-order logic on finite graphs with the possibility to explicitly quantify over walks. WL can serve as a unifying framework for path query languages. To support this claim, WL is compared in expressive power with various established query languages for graphs, such as first-order logic extended with reachability; the monadic second-order logic of graphs; hybrid computation tree logic; and regular path queries. WL also serves as a framework to investigate the following natural questions: Is quantifying over walks more powerful than quantifying over paths (walks without repeating nodes) only? Is quantifying over infinite walks more powerful than quantifying over finite walks only? WL model checking is decidable, but determining the precise complexity remains an open problem.