entity linking

This work addresses cross-document coreference resolution in multi-document knowledge graphs. We propose a knowledge-graph-contextualized dynamic linking mechanism that jointly models entity mentions, knowledge graph structure, and cross-document semantic relations via an integrated framework combining contextual embeddings and graph neural networks, enabling end-to-end entity linking and coreference inference. Our key contributions are: (i) the first integration of knowledge graph embeddings with context-aware graph reasoning for modeling cross-document consistency; and (ii) dynamic updating of entity representations to accommodate heterogeneous, multi-source documents. Evaluated on benchmark datasets—including WD100K and CrossDoc—our method achieves substantial improvements in coreference resolution performance, yielding an average +4.2% F1-score gain over prior approaches. The results demonstrate enhanced capability in multi-document information extraction and knowledge fusion.

Over 698 million references in Crossref lack DOIs, severely impeding the structural construction of citation networks. This paper proposes a systematic approach integrating heuristic rules, metadata matching, and fuzzy text parsing to accurately align unstructured references to target文献 entities in OpenCitations Meta. Methodologically, it combines an interpretable rule engine with semantic matching strategies—specifically designed to enhance linkage accuracy under sparse or inconsistent metadata conditions. Contributions include: (1) a novel hybrid alignment framework balancing precision and robustness; (2) a manually curated gold standard and a validated Crossref subset for rigorous evaluation. Experimental results demonstrate high precision, substantially expanding both the coverage breadth and linkage quality of the open citation network. The method provides a reproducible, scalable technical pathway for large-scale reference DOI enrichment.

To address the need for interactive exploration and editing of named entities and their relationships within document collections, this paper proposes a tripartite graph representation model integrating documents, entity mentions, and entities as distinct node types. We design an interactive visualization system supporting coordinated multi-perspective analysis. Key innovations include a fuzzy-view mechanism and an editable graph structure, enabling users to dynamically refine entity relations, define novel associations, and receive real-time visual feedback. The system incorporates linked multi-view interfaces, hierarchical filtering, thumbnail-based navigation, and semantic-aware graph layout—balancing macro-level overviews with micro-level validation. It supports export of structured JSON data and high-resolution images. Evaluation demonstrates significant improvements in named entity clustering efficiency and relational inference accuracy, thereby enhancing human-AI collaboration for deep semantic understanding of textual relationships.

Existing citation extraction tools struggle to process footnotes in humanities and legal scholarship due to their embedded placement within main text, inclusion of commentary and cross-references, and highly variable formatting. To address this challenge, this work introduces FOSSIL, the first multilingual open dataset specifically designed for footnote citations, comprising over 7,600 annotated references from 96 scholarly papers. The authors also develop a dedicated PDF-TEI Editor collaborative annotation platform, standardize a seven-annotator labeling protocol, and implement a Grobid-based customized footnote parsing model. This end-to-end pipeline substantially improves performance, increasing the micro F1-score from 0.36 to 0.72 with notable gains in recall, thereby demonstrating the approach’s effectiveness while highlighting remaining challenges in handling cross-references and mixed-content footnotes.

Existing benchmarks for reference extraction and parsing are largely confined to well-formatted English bibliographies, struggling to handle the multilingualism, footnote-embedded citations, abbreviations, and diverse historical citation styles prevalent in social sciences and humanities. This work introduces the first unified evaluation benchmark tailored to this domain, comprising three real-world datasets. It systematically assesses the performance of large language models—including DeepSeek-V3.1, Mistral-Small, Gemma-3, and Qwen3-VL—on reference extraction, parsing, and end-to-end document processing tasks, with GROBID serving as a strong baseline. The study further proposes a hybrid deployment strategy combining lightweight LoRA fine-tuning with task routing, which achieves near-saturated extraction performance on medium-sized models and significantly enhances parsing accuracy and system robustness in complex citation scenarios.

Existing knowledge graphs modeling academic citation networks typically capture only topological structure, neglecting the stance and evaluative content embedded in citations. This work proposes a novel "claim network" representation framework that reformulates citations as typed claim units, each comprising source, target, claim text, and one of four stance labels, along with a general-purpose construction pipeline. By explicitly modeling citation intent for the first time, the approach enables fine-grained characterization of scholarly evaluative relationships. The authors construct a claim network from 127 papers on 3D point cloud semantic segmentation, yielding 8,260 typed claims, and demonstrate significant performance gains in downstream tasks including retrieval-augmented reasoning, stance-aware summarization, and topological analysis.

Large language models are prone to generating hallucinations or dubious citations in academic writing, undermining research credibility. This study presents the first systematic evaluation and comparison of mainstream citation verification tools—CheckIfExist, HalluCiteChecker, Hallucinator, HalRef, and RefChecker—on real-world academic documents. The analysis reveals significant limitations in current approaches, particularly concerning citation extraction accuracy, breadth of database coverage, and consistency in verification. While these tools can offer preliminary alerts for potentially fabricated references, their overall effectiveness remains constrained. This work provides an empirical foundation and clear directions for improving the verification of citation authenticity in scholarly communication.

Multimodal entity linking faces significant challenges in simultaneously achieving high accuracy, computational efficiency, and storage efficiency. This work proposes FAST-MEL, a lightweight encoder architecture that constructs unified, compact, fixed-length multimodal vector representations for both knowledge base entities and text–vision mentions, enabling efficient indexing and matching. By doing so, FAST-MEL maintains state-of-the-art entity linking accuracy while accelerating inference by three orders of magnitude and reducing storage overhead by an order of magnitude compared to the fastest existing system. This represents the first approach to successfully overcome the longstanding trade-off among these three critical objectives, demonstrating that they can be jointly optimized without compromising performance.

This study addresses the limitations of traditional Retrieval-Augmented Generation (RAG) systems, which often neglect structured metadata and entity relationships within documents, thereby constraining retrieval efficacy and answer quality. To overcome this, the authors propose an agent-optimized enhanced entity page format that leverages linked data—specifically Schema.org annotations and dereferenceable entity pages—and integrates llms.txt instructions, navigational pathways, and neural search capabilities. An Enhanced+ variant is further introduced to strengthen inter-entity linking and navigation support. The system is implemented using Vertex AI Vector Search 2.0 and the Google Agent Development Kit, combining JSON-LD, HTML, and linked data platform technologies. Experimental results demonstrate a 29.6% accuracy improvement over standard RAG and a 29.8% gain in full agent workflows, with the Enhanced+ variant achieving the best performance (accuracy: 4.85/5; completeness: 4.55/5).

This study investigates how to enhance retrieval effectiveness over RDF datasets while preserving the semantic faithfulness of metadata generated by large language models (LLMs) to the original data. It formulates metadata generation for the first time as a system-level information retrieval problem and systematically evaluates six LLM-based strategies—ranging from unconstrained rewriting to knowledge graph–grounded agent approaches—in terms of the trade-off between retrieval performance and content faithfulness. Experimental results show that unconstrained rewriting yields the greatest retrieval gains but suffers from the lowest faithfulness, whereas profile-guided rewriting achieves the best balance between the two objectives. The work reveals that retrieval improvements may stem from unfaithful semantic expansions and proposes a new paradigm that jointly optimizes effectiveness and trustworthiness in metadata generation for semantic data retrieval.