coreference resolution

Existing coherence modeling approaches treat entity consistency and discourse relations in isolation, failing to capture their interdependence. Method: We propose the first joint modeling framework that unifies entity linking—including coreference resolution—with explicit and implicit discourse relation classification, enabling neural cross-sentence entity tracking and discourse structure representation in a single, end-to-end trainable architecture. Results: Evaluated on three standard coherence assessment benchmarks, our model achieves significant improvements over single-feature baselines (average +3.2% F1), demonstrating strong complementarity between entity-level and discourse-level information. Contribution: This work introduces a novel, unified paradigm for discourse coherence modeling, establishing a principled foundation for jointly learning entity dynamics and rhetorical structure within cohesive text.

This study investigates how large language models (LLMs) internally represent and distinguish named entity mentions, addressing the many-to-many mapping between entities and their surface mentions. Method: We propose the first clustering-inspired evaluation framework for entity recognition, integrating representation clustering, low-dimensional linear subspace detection, and knowledge-structure isomorphism modeling to quantify both intra-entity mention cohesion and inter-entity mention separation. Contribution/Results: We find that entity information is compactly encoded in low-dimensional linear subspaces as early as the initial Transformer layers, and that these subspaces exhibit structural isomorphism with real-world entity knowledge. Evaluated on five mainstream LLMs, our framework achieves entity recognition precision/recall of 0.66–0.90. Results confirm that strong entity discrimination emerges early in the model hierarchy, and that the quality of entity representations significantly influences downstream token prediction performance.

Traditional entity linking adopts a two-stage paradigm (mention detection followed by disambiguation), suffering from error propagation, high computational overhead, and poor cross-domain generalization. This paper proposes an end-to-end joint modeling framework that unifies mention detection and entity disambiguation. It leverages fine-tuned large language models (LLMs) to construct context-aware mention representations, effectively mitigating domain shift. Key contributions include: (i) a lightweight adapter mechanism that fuses deep semantic features from LLMs with local structural cues—avoiding full-parameter fine-tuning; and (ii) a cross-domain robust joint optimization objective. The method achieves state-of-the-art performance on multiple standard benchmarks (AIDA, MSNBC, ACE2005), notably improving F1 scores by an average of +4.2% in zero-shot cross-domain settings, thereby demonstrating superior effectiveness and generalization capability.

Low accuracy and high ambiguity in entity linking for short-text question answering remain critical challenges. This paper proposes a large language model (LLM)-based entity linking agent framework that emulates human cognitive processes by decoupling entity recognition, candidate retrieval, and disambiguation into collaborative, modular agents. The framework supports active reasoning and tool invocation—including knowledge base queries and context expansion—to enable context-aware, end-to-end linking. Its core innovation lies in a dynamic, agent-driven reasoning mechanism that overcomes the rigidity and error propagation inherent in conventional pipeline approaches. Extensive experiments on multiple short-text QA benchmarks demonstrate substantial improvements in linking accuracy. The method consistently outperforms state-of-the-art baselines in both tool-augmented linking and end-to-end QA tasks, validating its robustness and effectiveness.

This work addresses multimodal entity linking with joint text-image inputs—given a mention comprising a textual paragraph and multiple images, the goal is to accurately retrieve the target entity from a multimodal knowledge base integrating textual descriptions, visual content, and structured attributes. We introduce fine-grained structured attributes to this task for the first time, establishing Ameli, the first attribute-aware benchmark, along with its dedicated multimodal knowledge base. We propose a trimodal fusion model that integrates BERT (for text), ViT (for images), an attribute embedding module, and cross-modal attention mechanisms to jointly align textual, visual, and attribute representations. Our method achieves significant improvements over state-of-the-art approaches on Ameli; ablation studies confirm that explicit attribute modeling boosts accuracy by over 8%. To foster reproducibility and further research, we fully open-source our code, pre-trained models, and a large-scale dataset comprising 16K+ annotated mentions and 34K+ entities.

This work addresses the challenge of implicit entity recognition in first-person retrospective narratives, where entities such as persons, locations, or events are often omitted without explicit mention. To tackle this problem, the authors propose a novel task formulation that leverages non-local contextual cues for inference and introduce IRC-Bench, a benchmark dataset pairing explicit and implicit narrative instances to systematically evaluate model capabilities. The study integrates large language model (LLM) generation, dense retrieval, retrieval-augmented generation (RAG), and QLoRA fine-tuning across 19 distinct model configurations. Experimental results demonstrate that a QLoRA-finetuned Llama-3.1-8B achieves a 38.94% exact match accuracy under open-world settings, while a finetuned DPR retriever attains a Hit@1 score of 35.38% in closed-world retrieval, confirming both the efficacy of the proposed approach and the difficulty of the benchmark.

This work addresses the limitations of existing ultra-fine entity typing methods, which rely solely on sentence-level context and struggle to accurately classify long-tail entity types. To overcome this, the authors propose a novel approach that leverages controllable generation to produce coherent multi-sentence narratives that expand entity context. They introduce the first Narrative-UFET dataset and design two variants of synthetic narratives—type-preserving and type-altering—to systematically investigate how discourse structure influences entity type disambiguation. Experimental results demonstrate that multi-sentence contexts derived from synthetic narratives substantially outperform sentence-level baselines, particularly for long-tail types. Notably, type-altering narratives provide stronger discriminative signals, and synthetic narratives consistently surpass natural contexts in classification performance, confirming the effectiveness and potential of controllable narrative generation for enhancing ultra-fine-grained entity typing.

Traditional entity alignment methods rely on graph structures and textual attributes, struggling to accurately identify identical entities and distinguish semantically similar ones under heterogeneous descriptions. This work proposes IRIS, a novel framework that, for the first time, extracts context-independent entity identity representations from frozen large language models. By designing identity-oriented prompts, IRIS constructs a unified comparison space, enabling direct similarity-based alignment without any training and allowing one-time encoding to be reused across knowledge graphs. The approach eliminates dependence on specific candidate sets or graph pairs, achieving state-of-the-art Hits@1 scores of 100.00, 99.38, 98.31, and 97.99 on four standard benchmarks—significantly outperforming existing methods.

This work addresses three core challenges in conversational information seeking—modeling personal context, generating personalized responses, and automatic evaluation—by proposing an integrated solution. It introduces the ConEL dataset and the CREL method for conversational entity linking, constructs LAPS, a large-scale personalized dialogue dataset, and presents FACE, the first reference-free, globally aware automatic evaluation framework. The proposed approaches substantially enhance the system’s ability to understand and leverage users’ personal contexts, yielding responses better aligned with individual preferences. Moreover, FACE demonstrates high agreement with human judgments, establishing a new paradigm for both the development and evaluation of personalized dialogue systems.

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.