Xinyu Zhang, Yuchen Wan, Boxuan Zhang, Zesheng Yang, Lingling Zhang, Bifan Wei, Jun Liu. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Geometry problem solving (GPS) requires high-level symbolic and logical reasoning based on geometry theorem knowledge to arrive at the answer. Despite the remarkable advances achieved by Large Language Models (LLMs) in various problem-solving tasks, they still struggle to perform rigorous multi-step geometry reasoning, which is essential for GPS. In this paper, we propose a dynamic tree-based geometry problem solver named GeoTree, which combines a knowledgeable LLM with a rigorous symbolic solver to perform geometry reasoning cooperatively. Specifically, an iterative multi-step geometry reasoning process is performed dynamically based on a tree-like structure, thereby emulating divergent and deliberate human problem-solving thinking. Each geometry reasoning step is completed collaboratively through four components, consisting ofTheorem Seeker,Symbolic Solver,Evaluator, andController. First,Theorem Seekerprompts LLMs to seek out candidate theorems with their inherent geometry theorem knowledge. Subsequently,Symbolic Solverapplies the theorems on the known conditions to obtain new additional conditions. Then,Evaluatorassesses the availability of the theorems and prompts LLMs to judge the usefulness of these new conditions for the problem target, which serves as the heuristic guidance for subsequent reasoning. Finally,Controllerdetermines the termination state, which decides whether to continue invoking the other three components for further attempts. Extensive experiments on Geometry3K demonstrate the superiority of GeoTree in accuracy, efficiency, and explainability.

In this work, we address the challenging task of Generalized Referring Expression Comprehension (GREC). Compared to the classic Referring Expression Comprehension (REC) that focuses on single-target expressions, GREC extends the scope to a more practical setting by further encompassing no-target and multi-target expressions. Existing REC methods face challenges in handling the complex cases encountered in GREC, primarily due to their fixed output and limitations in multi-modal representations. To address these issues, we propose a Hierarchical Alignment-enhanced Adaptive Grounding Network (HieA2G) for GREC, which can flexibly deal with various types of referring expressions. First, a Hierarchical Multi-modal Semantic Alignment (HMSA) module is proposed to incorporate three levels of alignments, including word-object, phrase-object, and text-image alignment. It enables hierarchical cross-modal interactions across multiple levels to achieve comprehensive and robust multi-modal understanding, greatly enhancing grounding ability for complex cases. Then, to address the varying number of target objects in GREC, we introduce an Adaptive Grounding Counter (AGC) to dynamically determine the number of output targets. Additionally, an auxiliary contrastive loss is employed in AGC to enhance object-counting ability by pulling in multi-modal features with the same counting and pushing away those with different counting. Extensive experimental results show that HieA2G achieves new state-of-the-art performance on the challenging GREC task and also the other 4 tasks, including REC, Phrase Grounding, Referring Expression Segmentation (RES), and Generalized Referring Expression Segmentation (GRES), demonstrating the remarkable superiority and generalizability of the proposed HieA2G.

Diagram question answering (DQA) is a challenging task that requires models to combine with domain-specific knowledge and reason over the diagrams to answer questions. Multimodal Large Language Models (MLLMs) have recently made notable strides in combining textual and visual information, emerging as a promising solution for addressing the DQA task. However, they still encounter challenges in deliberate multimodal reasoning over the fine-grained visual details of content-rich and knowledge-grounded diagrams. The tight interweaving of visual and textual reasoning for MLLMs is also susceptible to hallucinations. To overcome these limitations, we propose a Glance-then-Focus Multimodal Reasoning framework named GlFoMR for DQA, which features a flexible architecture for comprehensive visual and text interaction. Firstly, the diagram is parsed into a hierarchical structure spanning different granularities including isolated single-object, object-group, and whole-diagram. Subsequently, the Glance-Plan and Focus-Reason stages collaborate to decouple the complex reasoning process. Glance-Plan first generates a preliminary plan by glancing at the multimodal context, specifying sub-goals related to knowledge extraction, visual perception, and visual reasoning. Based on these sub-goals, Focus-Reason further integrates domain-specific knowledge and visual details to enable more deliberate reasoning. The parsed multi-granularity diagram information is seamlessly incorporated into the corresponding sub-goal achievement process, enhancing the perception and reasoning capabilities of MLLMs for better DQA performance. Extensive experimental results on four DQA datasets demonstrate that GlFoMR achieves substantial improvements, showcasing its potential to advance the development of multimodal reasoning.

Automatically generated questions often suffer from problems such as unclear expression or factual inaccuracies, requiring a reliable and comprehensive evaluation of their quality.Human evaluation is widely used in the field of question generation (QG) and serves as the gold standard for automatic metrics.However, there is a lack of unified human evaluation criteria, which hampers consistent and reliable evaluations of both QG models and automatic metrics.To address this, we propose QGEval, a multi-dimensional Evaluation benchmark for Question Generation, which evaluates both generated questions and existing automatic metrics across 7 dimensions: fluency, clarity, conciseness, relevance, consistency, answerability, and answer consistency.We demonstrate the appropriateness of these dimensions by examining their correlations and distinctions.Through consistent evaluations of QG models and automatic metrics with QGEval, we find that 1) most QG models perform unsatisfactorily in terms of answerability and answer consistency, and 2) existing metrics fail to align well with human judgments when evaluating generated questions across the 7 dimensions.We expect this work to foster the development of both QG technologies and their evaluation.

Conversational question generation aims to generate questions that depend on both context and conversation history.Conventional works utilizing deep learning have shown promising results, but heavily rely on the availability of large-scale annotated conversations.In this paper, we introduce a more realistic and less explored setting, Zero-shot Conversational Question Generation (ZeroCQG), which requires no human-labeled conversations for training.To solve ZeroCQG, we propose a multi-stage knowledge transfer framework, Synthesize, Prompt and trAnsfer with pRe-Trained lAnguage model (SPARTA) to effectively leverage knowledge from single-turn question generation instances.To validate the zero-shot performance of SPARTA, we conduct extensive experiments on three conversational datasets: CoQA, QuAC, and DoQA by transferring knowledge from three single-turn datasets: MS MARCO, NewsQA, and SQuAD.The experimental results demonstrate the superior performance of our method.Specifically, SPARTA has achieved 14.81 BLEU-4 (88.2% absolute improvement compared to T5) in CoQA with knowledge transferred from SQuAD.

Textbook Question Answering (TQA) task requires answering questions by reasoning based on both the given diagrams and text context. There are mainly two challenges for the task. First, the diagrams are different from the natural images. Similar shapes or color blocks may express different semantics and there is also a large intra-topic variation for diagrams. Hence, the characteristics of visual semantic ambiguity and variable visual appearance make the diagram understanding more challenging. Second, for the text, the specific education domain with terminologies exists a great gap with the general domain. Therefore, it is difficult to represent the text semantics effectively using a text encoder pretrained in the general domain. In this paper, we propose a Spatial-Semantic Collaborative Graph Network (SSCGN) for TQA task, which can help enhance the diagram and text understanding and facilitate multimodal reasoning. Specifically, the Spatial-guided Semantic Enhancing (SSE) module fully exploits the spatial and semantic relationships between visual objects and OCR tokens to collaboratively enhance the diagram semantic understanding. Moreover, based on the semantically enhanced region representations of the SSE module, the Fine-grained Spatial-Aware Graph Network (FSA-GN) can help obtain richer relation-aware region representations for joint reasoning by capturing more fine-grained spatial relationships. We further propose multiple self-supervised auxiliary tasks to enhance the initial diagram and text semantic representations by pretraining the diagram encoder and text encoder. Extensive experiments and ablation studies are conducted to validate the effectiveness of SSCGN.

Diagram Question Answering (DQA) aims to correctly answer questions about given diagrams, which demands an interplay of good diagram understanding and effective reasoning. However, the same appearance of objects in diagrams can express different semantics. This kind of visual semantic ambiguity problem makes it challenging to represent diagrams sufficiently for better understanding. Moreover, since there are questions about diagrams from different perspectives, it is also crucial to perform flexible and adaptive reasoning on content-rich diagrams. In this paper, we propose a Disentangled Adaptive Visual Reasoning Network for DQA, named DisAVR, to jointly optimize the dual-process of representation and reasoning. DisAVR mainly comprises three modules: improved region feature learning, question parsing, and disentangled adaptive reasoning. Specifically, the improved region feature learning module is designed to first learn robust diagram representation by integrating detail-aware patch features and semantically-explicit text features with region features. Subsequently, the question parsing module decomposes the question into three types of question guidance including region, spatial relation and semantic relation guidance to dynamically guide subsequent reasoning. Next, the disentangled adaptive reasoning module decomposes the whole reasoning process by employing three visual reasoning cells to construct a soft fully-connected multi-layer stacked routing space. These three cells in each layer reason over object regions, semantic and spatial relations in the diagram under the corresponding question guidance. Moreover, an adaptive routing mechanism is designed to flexibly explore more optimal reasoning paths for specific diagram-question pairs. Extensive experiments on three DQA datasets demonstrate the superiority of our DisAVR.

The typical RDF querying method is a structured query language such as SPARQL. However, a structured query language requires issuers to understand the data model and query language syntax, limiting the accessibility of RDF data. The keyword query is regarded as an intuitive way for searching the RDF data. This paper proposes a keyword query approach based on the community structure of the RDF entity graph, which utilizes the characteristics of the community structure on the RDF graph. We propose a novel RDF entity graph model by designing an efficient entity index mechanism by treating the RDF entity as the virtual document, which reduces the graph scale while improving the efficiency of retrieving entities. We design an effective hierarchical and overlapping community-based index mechanism on the RDF entity graph, improving the efficiency and accuracy of the subgraph (sub-community) matching. Our approach takes full advantage of the schema of the RDF data and the community structural characteristic of the RDF entity graph. We evaluate our approach using two real-world datasets, and the experiment demonstrates the effectiveness and efficiency of our approach.

Text segmentation is a fundamental step in natural language processing (NLP) and information retrieval (IR) tasks. Most existing approaches do not explicitly take into account the facet information of documents for segmentation. Text segmentation and facet annotation are often addressed as separate problems, but they operate in a common input space. This article proposes FTS, which is a novel model for faceted text segmentation via multitask learning (MTL). FTS models faceted text segmentation as an MTL problem with text segmentation and facet annotation. This model employs the bidirectional long short-term memory (Bi-LSTM) network to learn the feature representation of sentences within a document. The feature representation is shared and adjusted with common parameters by MTL, which can help an optimization model to learn a better-shared and robust feature representation from text segmentation to facet annotation. Moreover, the text segmentation is modeled as a sequence tagging task using LSTM with a conditional random fields (CRFs) classification layer. Extensive experiments are conducted on five data sets from five domains: data structure, data mining, computer network, solid mechanics, and crystallography. The results indicate that the FTS model outperforms several highly cited and state-of-the-art approaches related to text segmentation and facet annotation.

Most community question answering (CQA) websites manage plenty of question-answer pairs (QAPs) through topic-based organizations, which may not satisfy users' fine-grained search demands. Facets of topics serve as a powerful tool to navigate, refine, and group the QAPs. In this work, we propose FACM, a model to annotate QAPs with facets by extending convolution neural networks (CNNs) with a matching strategy. First, phrase information is incorporated into text representation by CNNs with different kernel sizes. Then, through a matching strategy among QAPs and facet label texts (FaLTs) acquired from Wikipedia, we generate similarity matrices to deal with the facet heterogeneity. Finally, a three-channel CNN is trained for facet label assignment of QAPs. Experiments on three real-world data sets show that FACM outperforms the state-of-the-art methods.

Discovering hyponym relations among domain-specific terms is a fundamental task in taxonomy learning and knowledge acquisition. However, the great diversity of various domain corpora and the lack of labeled training sets make this task very challenging for conventional methods that are based on text content. The hyperlink structure of Wikipedia article pages was found to contain recurring network motifs in this study, indicating the probability of a hyperlink being a hyponym hyperlink. Hence, a novel hyponym relation extraction approach based on the network motifs of Wikipedia hyperlinks was proposed. This approach automatically constructs motif-based features from the hyperlink structure of a domain; every hyperlink is mapped to a 13-dimensional feature vector based on the 13 types of three-node motifs. The approach extracts structural information from Wikipedia and heuristically creates a labeled training set. Classification models were determined from the training sets for hyponym relation extraction. Two experiments were conducted to validate our approach based on seven domain-specific datasets obtained from Wikipedia. The first experiment, which utilized manually labeled data, verified the effectiveness of the motif-based features. The second experiment, which utilized an automatically labeled training set of different domains, showed that the proposed approach performs better than the approach based on lexico-syntactic patterns and achieves comparable result to the approach based on textual features. Experimental results show the practicability and fairly good domain scalability of the proposed approach.