graph-to-text explanation

Large-scale Transformer models suffer from poor interpretability and token-level representations lacking semantic completeness. Method: This paper proposes an end-to-end text-to-semantic-graph paradigm: leveraging dependency parsing and semantic role labeling to automatically construct semantics-preserving graph structures, where nodes and edges explicitly encode linguistic units and their logical relations; it further integrates graph neural networks with interpretable attention mechanisms to yield structured, traceable, and intervenable decision explanations. Contribution/Results: By bypassing traditional fragmented token representations, the approach significantly improves explanation consistency (+37%) and inference efficiency—outperforming attribution methods for larger models by 5.2×—across multiple NLP classification tasks. It is the first work to realize end-to-end interpretable modeling from raw text to semantic graphs.

Large language models (LLMs) exhibit limited performance on graph-structured tasks due to weak structural awareness, permutation invariance, and insufficient capacity for complex relational reasoning. This work proposes a human-interpretable graph-to-text encoding method that, for the first time, maps refined Weisfeiler–Lehman similarity classes to semantic color tokens instead of conventional numerical symbols, thereby effectively integrating graph structural information into LLM inputs. By leveraging structure-aware prompt engineering, the approach substantially enhances model performance on both algorithmic reasoning and predictive graph tasks, with particularly notable gains on tasks requiring global structural understanding. The method demonstrates consistent effectiveness across both synthetic and real-world datasets.

Current NLP models lack interpretability in text similarity assessment and logical relation classification (entailment, contradiction, neutral), hindering reliable semantic structure modeling and logical reasoning. To address this, we propose the first hybrid interpretable framework integrating Montague semantics, graph embedding, and first-order logic. Our method synergistically combines generative language models with logical prompting to explicitly encode syntactic structure, logical connectives, and spatiotemporal constraints. It is the first approach to precisely distinguish all three logical relations in text classification tasks. Evaluated on three independently annotated datasets, it significantly outperforms state-of-the-art baselines—particularly excelling in sentence structural equivalence detection, sensitivity to logical connectives, and spatiotemporal reasoning. The framework substantially enhances transparency and trustworthiness in information retrieval systems.

This work addresses three critical challenges in generating graph models from natural language using large language models (LLMs): syntactic violations (deviations from the metamodel), constraint inconsistencies (violations of domain-specific rules), and content hallucinations (introduction of spurious elements). To tackle these, we propose an “abstraction–concretization” framework: first, aggregating outputs from multiple LLMs to construct a probabilistic partial model; then, refining it into a complete, constraint-compliant graph model via a constraint-driven optimization process. The method integrates probabilistic modeling, symbolic constraint reasoning, and self-consistency-based ensemble mechanisms. Extensive experiments across diverse open- and closed-source LLMs demonstrate substantial improvements in syntactic correctness, constraint adherence, and semantic fidelity—effectively suppressing structural errors and hallucinations. Our approach establishes a verifiable, constraint-governed generation paradigm for model-driven engineering powered by LLMs.

Existing LLM-based graph analysis benchmarks rely on direct structural reasoning, limiting scalability to large graphs; in contrast, human experts routinely solve such tasks programmatically using graph libraries (e.g., NetworkX, PyTorch Geometric). Method: We propose ProGraph—the first programming-centric benchmark for graph analysis—comprising three expert-level task categories, multi-scale real-world graphs, and six mainstream graph libraries. We further introduce LLMS4Graph, a high-quality dataset featuring authoritative documentation and automatically generated code. To enhance API comprehension and code generation, we employ documentation-augmented retrieval and fine-tune open-source LLMs on this data. Contribution/Results: Our approach yields 11–32% absolute accuracy gains on ProGraph, with the best model achieving 36% accuracy. All components—including the ProGraph benchmark, LLMS4Graph dataset, and enhanced models—are publicly released to advance LLMs’ programmatic understanding of structured graph data.

本文探索使用基于原子命题的图表示来进行可解释的自然语言推理，通过将句子转换为图并输入到语言模型中，尽管在某些数据集上略低于文本模型，但实现了较高准确性。

本文通过BioGlyph方法将网络拓扑结构编译成可解释的角色语言，显著提升了大型语言模型在复杂网络结构推理中的性能。

This work addresses the challenge of comprehending knowledge-intensive texts, which often contain numerous entities and intricate nested relationships that hinder efficient semantic understanding. To tackle this issue, we propose GraphTide, a dynamic visualization approach that supports progressive exploration through on-demand decomposition of entity relationships, structure-aware force-directed layout optimization, incremental information disclosure, and smooth animated transitions. These mechanisms collectively produce clear, contextually coherent visualizations of nested relational structures. User studies demonstrate that GraphTide significantly improves both the efficiency and accuracy of users’ comprehension of complex textual content compared to conventional graph visualizations and static nested diagrams.

This work addresses the limitations of large language models (LLMs) in structured data processing and multi-hop reasoning by proposing a graph-native collaborative intelligence framework. The approach integrates three key innovations: graph-augmented retrieval-based reasoning, bidirectional synergy between LLMs and knowledge graphs, and graph algorithm–driven agent decision-making. By deeply unifying LLMs, graph neural networks (GNNs), and graph computing techniques, the framework cohesively combines natural language interfaces, hybrid LLM-GNN pipelines, and graph data management systems. This integration substantially enhances fact consistency and context-aware reasoning and decision-making capabilities in complex scenarios, laying the foundation for next-generation graph-native intelligent systems.

This study systematically investigates the applicability and capability boundaries of large language models (LLMs) in graph computation tasks. Addressing structured relational reasoning and algorithmic operations, it proposes the first role-based classification framework that categorizes LLM usage paradigms into “executor” and “planner.” Through a comprehensive literature review and task-specific analysis, the work evaluates LLM performance across diverse graph-related scenarios. The findings indicate that while LLMs demonstrate feasibility on small-scale graph tasks, they remain unreliable for large-scale, precise computations. The paper clarifies the appropriate role of LLMs in graph computing, summarizes existing benchmark datasets, and outlines four key directions for future research to advance this emerging intersection of language models and graph algorithms.