Hao Gu, Lujun Li, Hao Wang, Lei Wang, Zheyu Wang, Bei Liu, Jiacheng Liu, Qiyuan Zhu, Sirui Han, Yike Guo. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Binxing Xu, Hao Gu, Lujun Li, Hao Wang, Bei Liu, Jiacheng Liu, Qiyuan Zhu, Xintong Yang, Chao Li, Sirui Han, Yike Guo. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Text-to-image diffusion models have made remarkable progress in generating high-quality images, but achieving fine-grained control over visual text remains a challenge. Existing methods fall short in enabling precise manipulation of font and color in visual text. To address this, we propose FC-Render, a method that introduces adaptive Font and Color control into a pre-trained TextDiffuser-2 model. Our approach consists of two stages: first, we develop independent font and color experts for decoupled control over visual text attributes. Then, we introduce an adaptive router that dynamically adjusts the involvement of each expert in font, color, or combined tasks, enabling precise fusion of font and color features. Through extensive experiments, we demonstrate that FC-Render enables fine-grained control over font and color while maintaining text rendering quality, providing new insights into the interaction between font, color, and spelling accuracy.

Large language models (LLMs) have made significant advancements in natural language processing, yet they still face challenges in tasks requiring deep comprehension of structurally complex and visually rich documents. Human reading patterns, captured through eye-tracking, provide valuable insights into how meaning is extracted from text, particularly in Visually Rich Documents (VRDs). In this work, we propose a novel approach to integrating eye-tracking data into LLMs to enhance reading strategies that more closely reflect human cognitive processes. By mimicking human gaze paths when reading a VRD, we explore how these insights can improve LLMs’ comprehension abilities. Our experiments demonstrate that LLMs enhanced with human-like reading orders outperform baseline models in VRD understanding tasks. These findings suggest that incorporating human-like reading strategies can bridge the gap between machine and human understanding. Overall, our results indicate that integrating eye-tracking data significantly enhances LLM performance, paving the way for more human-like document comprehension in future AI systems.

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.

Key-value relations are prevalent in Visually-Rich Documents (VRDs), often depicted in distinct spatial regions accompanied by specific color and font styles. These non-textual cues serve as important indicators that greatly enhance human comprehension and acquisition of such relation triplets. However, current document AI approaches often fail to consider this valuable prior information related to visual and spatial features, resulting in suboptimal performance, particularly when dealing with limited examples. To address this limitation, our research focuses on few-shot relational learning, specifically targeting the extraction of key-value relation triplets in VRDs. Given the absence of a suitable dataset for this task, we introduce two new few-shot benchmarks built upon existing supervised benchmark datasets. Furthermore, we propose a variational approach that incorporates relational 2D-spatial priors and prototypical rectification techniques. This approach aims to generate relation representations that are more aware of the spatial context and unseen relation in a manner similar to human perception. Experimental results demonstrate the effectiveness of our proposed method by showcasing its ability to outperform existing methods. This study also opens up new possibilities for practical applications.

Summary Event schema refers to the use of a template to depict similar events, and it is a necessary prerequisite for event causality extractions. The induction of event schemas is a difficult task, especially for texts in the open domain, due to the complex and diverse manifestations of events. Previous models considered participants in event mentions are independent or compositional, ignoring the high‐order correlations among participants, which limit their capability of induce event schema. To remedy this, we propose constructing an Event Structure Hypergraph (ESH) to better utilizes the event structural information for event schema induction. In particular, we first extract event mentions from the open‐domain corpus. and then construct an ESH by representing event mentions as a hyperedges. ESH contains high‐order information between participants in event mention. To, learn event mentions representation based on ESH, we propose a weighted attentive hypergraph neural network (WHGNN) to model event high‐order correlations and then integrate node‐category weight matrix into the training of network by improving event representation. By applying jointly cluster algorithm on the event mentions representation, we can induce reliable event schemas. Experimental results on three datasets demonstrate that our approach can induce salient and high‐quality event schemas on open‐domain corpus.

Extracting meaningful entities belonging to predefined categories from Visually-rich Formlike Documents (VFDs) is a challenging task.Visual and layout features such as font, background, color, and bounding box location and size provide important cues for identifying entities of the same type.However, existing models commonly train a visual encoder with weak cross-modal supervision signals, resulting in a limited capacity to capture these nontextual features and suboptimal performance.In this paper, we propose a novel Visually-Asymmetric coNsistenCy Learning (VANCL) approach that addresses the above limitation by enhancing the model's ability to capture finegrained visual and layout features through the incorporation of color priors.Experimental results on benchmark datasets show that our approach substantially outperforms the strong LayoutLM series baseline, demonstrating the effectiveness of our approach.Additionally, we investigate the effects of different color schemes on our approach, providing insights for optimizing model performance.We believe our work will inspire future research on multimodal information extraction.

Few-shot link prediction is an important recent research direction in the field of knowledge graph.To learn better entity and relation representations, previous works utilize the neighborhood information of entities but they ignore the associated relations.More than this, the link invovles few-shot entities cannot be well linked by these methods.Therefore, this paper proposes a novel method that dynamically aggregates local neighbourhood information from entities and relations and introduces the information bottleneck principle to filter irrelevant features to better capture the feature of few-shot entities.Experimental results on three benchmark datasets demonstrate that our model outperforms other models and that our approach can effectively improve the model performance on those entities in the tail of the long-tailed distribution.

Answer selection aims to identify the most relevant answers to a given question from a set of candidates. It is the fundamental component of intelligent question answering system. To improve performance, it gradually becomes an effective strategy to integrate external structured knowledge bases (KBs) into the answer selection model. Due to expensive cost of construction and maintenance of such KBs, these models are suffering from domain barriers and information incompleteness. In this paper, we propose a two-stage extraction–comprehension answer selection model, which can extract ad-hoc knowledge from unstructured web texts to enhance the performance of answer selection. For the extraction, two types of snippets are extracted from unstructured web pages and utilized as the source of ad-hoc knowledge. For the comprehension, a selective attention mechanism is employed to extract and integrate ad-hoc knowledge from multiple text snippets obtained in the first stage, which can enrich the representation of question–answer pairs and more accurately identify the correct answers. By incorporating ad-hoc knowledge extracted from both types of snippets, the proposed model achieves state-of-the-art results on two public available benchmark datasets. In particular, on WikiQA, in terms of the two evaluation metrics (mean average precision and mean reciprocal rank), it achieves 9.9[Formula: see text] and 8.4[Formula: see text] higher than the previous non-pretraining-based models, and 3.4[Formula: see text] and 3.2[Formula: see text] higher than the pretraining-based models.

Abstract In recent years, the extraction of overlapping relations has received great attention in the field of natural language processing (NLP). However, most existing approaches treat relational triples in sentences as isolated, without considering the rich semantic correlations implied in the relational hierarchy. Extracting these overlapping relational triples is challenging, given the overlapping types are various and relatively complex. In addition, these approaches do not highlight the semantic information in the sentence from coarse-grained to fine-grained. In this paper, we propose an end-to-end neural framework based on a decomposition model that incorporates multi-granularity relational features for the extraction of overlapping triples. Our approach employs an attention mechanism that combines relational hierarchy information with multiple granularities and pretrained textual representations, where the relational hierarchies are constructed manually or obtained by unsupervised clustering. We found that the different hierarchy construction strategies have little effect on the final extraction results. Experimental results on two public datasets, NYT and WebNLG, show that our mode substantially outperforms the baseline system in extracting overlapping relational triples, especially for long-tailed relations.

Abstract Metonymy resolution (MR) is a challenging task in the field of natural language processing. The task of MR aims to identify the metonymic usage of a word that employs an entity name to refer to another target entity. Recent BERT-based methods yield state-of-the-art performances. However, they neither make full use of the entity information nor explicitly consider syntactic structure. In contrast, in this paper, we argue that the metonymic process should be completed in a collaborative manner, relying on both lexical semantics and syntactic structure (syntax). This paper proposes a novel approach to enhancing BERT-based MR models with hard and soft syntactic constraints by using different types of convolutional neural networks to model dependency parse trees. Experimental results on benchmark datasets (e.g., ReLocaR, SemEval 2007 and WiMCor) confirm that leveraging syntactic information into fine pre-trained language models benefits MR tasks.

Abstract Causal event extraction (CEE) aims to identify and extract cause‐effect event pairs from texts, which is a fundamental task in natural language processing. Recent research treat CEE as a sequence labeling problem. However, the linguistic complexity and ambiguity of textual description results in the low accuracy of extractors. To address the above issues, considering the prior knowledge like the causal network constructed based on the causal indicators, which can represent information transition between cause and effect, may helpful for CEE. In this article, we propose causality‐associated graph neural network to incorporate in‐domain knowledge by taking important causal words into account. External causal knowledge is modeled as causal associated graph (CAG). Then we use graph neural networks (GNN) to capture the complex relationship of intraevent mentions and interevent causality in a sentence based on the relationship obtained from CAG. Finally, sentence sequence and prior causal knowledge of GNN embedding are fed into multiscaled convolution and bidirectional long short‐term memory networks. Experimental results on two datasets show that our method outperforms the state‐of‐the‐art baseline.

Abstract Link prediction plays an important role in constructing knowledge graph. Recently, graph representation learning models yield state‐of‐the‐art results. However, existing models concentrate merely on triples or graph structures and mostly ignore textual descriptions, resulting in incomplete or partial information. In this paper, we propose a novel graph representation learning model to address this challenge, namely multi‐source heterogeneous information fusion with adaptive importance sampling. Our model leverages multiple sources, such triple, graph structure and textual description, and generate rich‐attribute embeddings for entities, encapsulating relations simultaneously. We also propose an adaptive importance sampling algorithm to boost aggregation of useful features from local neighbours. Additionally, we also boost node aggregation of useful features from local neighbours by adaptive importance sampling algorithm in our model. Experimental results on two benchmark datasets show that our proposed model significantly outperforms state‐of‐the‐art methods.

Causal Event Extraction (CEE) is a joint extraction task of events and causality, which can help text understanding, event prediction and so on. Recent research has achieved state-of-the-art performance in various Natural Language Processing (NLP) tasks by combining pre-trained models with neural networks. However, ambiguity of event description and long-distance dependence of event causality result in the low accuracy of extractors. In this paper, we propose a model to incorporate in-domain knowledge by taking frequent expression of event causality into account, and use iterated dilated convolutions to expand the perception field of event causality. External causal knowledge is modeled as frequent n-grams with different length, which is used as convolution filters during kernel initialization, enhancing the ability of model to capture the boundary of event description. To obtain long-distance dependence of event causality, we use iterated dilated convolutions to aggregate context from the entire sentence. Experimental results show that our method significantly outperform the baselines with faster convergence speed.

Summary Event evolutionary graph (EEG) reflects sequential and causal relations between events, which is of great value for event prediction. However, lacking event context in the EEG raises the problems of direction uncertainty and low accuracy when making predictions. In this article, we propose a conditional event evolutionary graph (CEEG) to deal with these problems. CEEG extends EEG with an additional four types of event context, including state, cause, sub‐type, and object. We first extract event context by matching the input with self‐adaptive semantic templates and generalize the context for each event. To identify the evolution direction, we treat it as a binary classification problem and calculate the event transition probability for each direction given the generalized context. Experimental results show that CEEG has a strong ability to generate better event evolutionary paths compared with NAR, EEM, and other non‐context‐based methods.

Abstract Given a question and its answer candidates (named QA corpus), answer selection is the task of identifying the most relevant answers to the question. Answer selection is widely used in question answering, web search, and so on. Current deep neural network models primarily utilize local features extracted from input question‐answer pairs (QA pairs). However, the global features contained in QA corpora are under‐utilized, and we argue that these global features substantially contribute to the answer selection task. To verify this point of view, we propose a novel model that combines local and global features for answer selection. In our model, two different global feature extractors are employed to extract statistical global features and deep global features from a QA corpus, respectively. Furthermore, we investigate the integration of these global features with local features in various experimental settings: statistical global features, deep global features, and a combination of statistical and deep global features. Our experimental results show that the global features are effective for answer selection. Our model obtains new state‐of‐the‐art results on two public answer selection datasets and performs especially well on YahooCQA, where it achieves 9.2 and 6% higher precision@1 (P@1) and mean reciprocal rank (MRR) scores than previously published models.

Summary Taxonomic relations play an important role in various Natural Language Processing (NLP) tasks (eg, information extraction, question answering and knowledge inference). Existing approaches on embedding‐based taxonomic relation learning mainly rely on the word embeddings trained using co‐occurrence‐based similarity learning. However, the performance of these approaches is not quite satisfactory due to the lack of sufficient taxonomic semantic knowledge within word embeddings. To solve this problem, we propose an improved embedding‐based approach to learn taxonomic relations via incorporating relation descriptions into word embeddings. First, to capture additional taxonomic semantic knowledge, we train special word embeddings using not only co‐occurrence information of words but also relation descriptions (eg, taxonomic seed relations and their contextual triples). Then, using the trained word embeddings as features, we employ two learning models to identify and predict taxonomic relations, namely, offset‐based classification model and offset‐based similarity model. Experimental results on four real‐world domain datasets demonstrate that our proposed approach can capture additional taxonomic semantic knowledge and reduce dependence on the training dataset, outperforming the state‐of‐the‐art compared approaches on the taxonomic relation learning task.

In different contexts, one abstract concept (e.g., fruit) may be mapped into different concrete instance sets, which is called abstract concept instantiation. It has been widely applied in many applications, such as web search, intelligent recommendation, etc. However, in most abstract concept instantiation models have the following problems: (1) the neglect of incorrect label and label incompleteness in the category structure on which instance selection relies; (2) the subjective design of instance profile for calculating the relevance between instance and contextual constraint. The above problems lead to false prediction in terms of abstract concept instantiation. To tackle these problems, we proposed a novel model to instantiate the abstract concept. Firstly, to alleviate the incorrect label and remedy label incompleteness in the category structure, an improved random-walk algorithm is proposed, called InstanceRank, which not only utilize the category information, but it also exploits the association information to infer the right instances of an abstract concept. Secondly, for better measuring the relevance between instances and contextual constraint, we learn the proper instance profile from different granularity ones. They are designed based on the surrounding text of the instance. Finally, noise reduction and instance filtering are introduced to further enhance the model performance. Experiments on Chinese food abstract concept set show that the proposed model can effectively reduce false positive and false negative of instantiation results.

Relation classification is one of the most fundamental upstream tasks in natural language processing and information extraction. State-of-the-art approaches make use of various deep neural networks (DNNs) to extract higher-level features directly. They can easily access to accurate classification results by taking advantage of both local entity features and global sentential features. Recent works on relation classification devote efforts to modify these neural networks, but less attention has been paid to the feature design concerning syntax. However, from a linguistic perspective, syntactic features are essential for relation classification. In this article, we present a novel linguistically motivated approach that enhances relation classification by imposing additional syntactic constraints. We investigate to leverage syntactic skeletons along with the sentential contexts to identify hidden relation types. The syntactic skeletons are extracted under the guidance of prior syntax knowledge. During extraction, the input sentences are recursively decomposed into syntactically shorter and simpler chunks. Experimental results on the SemEval-2010 Task 8 benchmark show that incorporating syntactic skeletons into current DNN models enhances the task of relation classification. Our systems significantly surpass two strong baseline systems. One of the substantial advantages of our proposal is that this framework is extensible for most current DNN models.

State-of-the-art methods for relation extraction consider the sentential context by modeling the entire sentence. However, syntactic indicators, certain phrases or words like prepositions that are more informative than other words and may be beneficial for identifying semantic relations. Other approaches using fixed text triggers capture such information but ignore the lexical diversity. To leverage both syntactic indicators and sentential contexts, we propose an indicator-aware approach for relation extraction. Firstly, we extract syntactic indicators under the guidance of syntactic knowledge. Then we construct a neural network to incorporate both syntactic indicators and the entire sentences into better relation representations. By this way, the proposed model alleviates the impact of noisy information from entire sentences and breaks the limit of text triggers. Experiments on the SemEval-2010 Task 8 benchmark dataset show that our model significantly outperforms the state-of-the-art methods.

Distantly supervised relation extraction leverages knowledge bases to label training data automatically. However, distant supervision may introduce incorrect labels, which harm the performance. Many efforts have been devoted to tackling this problem, but most of them treat relation extraction as a simple classification task. As a result, they ignore useful information that comes from related tasks, i.e., dependency parsing and entity type classification. In this paper, we first propose a novel Multi-Task learning framework for Relation Extraction (MTRE). We employ dependency parsing and entity type classification as auxiliary tasks and relation extraction as the target task. We learn these tasks simultaneously from training instances to take advantage of inductive transfer between auxiliary tasks and the target task. Then we construct a hierarchical neural network, which incorporates dependency and entity representations from auxiliary tasks into a more robust relation representation against the noisy labels. The experimental results demonstrate that our model improves the predictive performance substantially over single-task learning baselines.