Entity Linking & Coreference Resolution

• Entity Linking and Coreference Resolution are core NLP tasks that map text mentions to knowledge base entities and cluster co-referring expressions to resolve ambiguity.
• Recent joint models leverage deep contextual encoders, graph-based inference, and unified loss functions, achieving significant F1 score improvements and enhanced consistency.
• Advanced frameworks extend these techniques to cross-document, conversational, and multimodal settings, addressing challenges in low-resource and complex narrative texts.

Entity linking and coreference resolution are core tasks in natural language processing that enable the construction of structured knowledge from unstructured or semi-structured text. Entity linking (EL) grounds textual mentions to real-world entities in a knowledge base (KB), addressing ambiguity and synonymy. Coreference resolution (CR) identifies expressions—pronouns, nominals, and proper names—that refer to the same underlying entity within or across texts. Historically, these tasks have been approached separately, but recent research demonstrates that their integration yields substantial gains in coherence, consistency, and downstream knowledge extraction. Advances in deep contextual encoders, graph-based inference, and LLMs have resulted in sophisticated joint models that extend to cross-document, conversational, multimodal, and low-resource settings.

1. Formal Definitions and Theoretical Foundations

Entity linking is defined as the mapping $f_{\mathrm{EL}}: M \rightarrow E$ from textual mention set $M$ to knowledge base entity set $E$. This requires two sub-tasks: mention detection and entity disambiguation (given a candidate set per mention, select the correct entity).

Coreference resolution is the mapping $f_{\mathrm{CR}}: M \rightarrow \mathcal{C}$ from mentions to a set of clusters $\mathcal{C}$, such that all mentions in a cluster refer to the same entity. CR encompasses resolving pronouns, noun phrases, and sometimes conditionals or eventive expressions.

Both tasks are typically evaluated via precision, recall, and $F_1$, with additional coreference-specific metrics such as MUC, $B^3$, and CEAF$_{\phi_4}$, and entity linking–specific cluster-hard and mention-level $F_1$ (Sarkar et al., 24 Sep 2025, Zaporojets et al., 2021).

A theoretical link between EL and CR arises from the latent entity hypothesis: accurate entity disambiguation constrains coreference chains, and robust coreference enhances contextualizing EL (Zaporojets et al., 2021, Bai et al., 2021).

2. Algorithmic Paradigms and Model Architectures

Modern EL and CR models can be categorized as follows:

• Pipeline architectures: NER → coreference → entity linking, as seen in early and domain-specific systems (Yan et al., 2017).
• Joint inference models: Simultaneously optimize both EL and CR assignments, enforcing cluster-level entity consistency via structured prediction, mention-level self-attention, or joint loss functions (Bai et al., 2021, Zaporojets et al., 2021, Verlinden et al., 2021).
• Generative and unsupervised approaches: EM-based ranking models with resolution-mode variables enable unsupervised learning from raw text and straightforward EL extensions via a $\mathtt{kb\_link}$ mode (Ma et al., 2016).
• Graph-based algorithms: Knowledge graph (KG) node embeddings, GNN-based propagation, and message passing are employed for both within-document and cross-document coreference propagation, especially for indirect or alias-rich references (Dong et al., 8 Apr 2025).
• Multimodal and LLM-driven frameworks: Weakly supervised mention–region alignment, prompt-cache mechanisms, and end-to-end LLM prompting have advanced the scalability and robustness for long, ambiguous legal and literary texts (Goel et al., 2022, Meher et al., 30 Oct 2025, Sarkar et al., 24 Sep 2025).

Key architectural patterns include: span-based mention encoders (SpanBERT, Bi-LSTM), cluster-aware scoring functions (feed-forward networks over mention/entity pairs), and attention/graph-based candidate integration.

3. Joint Modeling: Consistency, Optimization, and Inference

Recent work demonstrates that enforcing a single entity-link assignment per coreference cluster enhances cluster consistency, error robustness, and coverage of "hard" mentions (those with incomplete candidate sets) (Zaporojets et al., 2021, Verlinden et al., 2021). For example, Zaporojets et al. introduce globally normalized models over mention–mention and mention–entity graphs, optimized via cross-entropy or Matrix-Tree Theorem–driven objectives.

Joint architectures (e.g., C$^2$) share representations between coref and entity linking heads, with mention-level self-attention integrating global and speaker cues (Bai et al., 2021). Joint loss is typically balanced as $$L_{\mathrm{total}} = \lambda_1 L_{c} + \lambda_2 L_{\ell}$$. Integration of KB embeddings further regularizes span representations, increasing CR and EL $F_1$ (up to +5 points) and benefiting rare entity types (Verlinden et al., 2021).

Cross-document and conversational modeling further extend these frameworks. Dynamic linking via contextual encoders and KG propagation enables coreference across documents, outperforming text-only baselines by 3–6 $F_1$ points (Dong et al., 8 Apr 2025). In conversation, incorporating personal entities and deictic coreference with long-history context (e.g., via LongFormer) is necessary for end-to-end dialog understanding (Joko et al., 2022).

4. Datasets, Evaluation Protocols, and Empirical Results

Entity linking and coreference systems are evaluated on varied corpora:

Dataset	Genre / Domain	Mentions	Entities	Languages
OntoNotes	News, Conv., ...	194K	44K	Multi
DWIE	News, Entity-centric	28K	~12K	English
ConEL-2	Wizard-of-Wikipedia	2.4K	-	Conversational
Mahānāma	Literary (epic)	109K	5.5K	Sanskrit, EN

Evaluation metrics include mention-level and cluster-level $F_1$ (EL, coref), coreference-specific measures (MUC, $B^3$, CEAF$_{\phi_4}$), and recall of gold entities in clusters (Sarkar et al., 24 Sep 2025, Zaporojets et al., 2021). In challenging settings, joint models exceed standalone $F_1$ by up to 5 points; on hard mention cases, joint inference achieves up to 50% better recall than mention-local EL (Zaporojets et al., 2021).

In conversational EL, the CREL toolkit achieves up to 72.9 mention detection $F_1$ and 65.1 end-to-end EL $F_1$, outperforming “document-trained” systems by 10–40 $F_1$ points (Joko et al., 2022). In literary Sanskrit, even advanced models drop from 74.8 to 51.6 CoNLL F1 under global context, reflecting the difficulty of robust long-range entity resolution (Sarkar et al., 24 Sep 2025).

5. Specialized and Multimodal Settings

• Conversational agents: Personal entity linking relies on coreference-style matching with explicit mention scoring, requiring new datasets and adapted token-classification architectures (Joko et al., 2022).
• Multiparty dialog: Joint coref+character linking with speaker-aware mention representations is essential for accurate resolution of pronouns and role phrases (Bai et al., 2021).
• Legal and narrative texts: LLM-prompted, three-stage coreference pipelines with persistent caches (LINK-KG) yield marked reductions in knowledge graph node duplication and noise (−45.21% node duplication, −32.22% noise compared to RAG-only baselines) (Meher et al., 30 Oct 2025).
• Visual grounding: Weakly supervised models jointly align text mentions and image regions, regularized by linguistic priors; performance gains are observed vs. both text-only and unsupervised grounding baselines (+6–10% F1 in multimodal MUC/BLANC) (Goel et al., 2022).

6. Open Challenges, Error Modes, and Future Directions

Current limitations include recall loss on creative/indirect deictics, system brittleness in global narrative tracking, and KB-coverage constraints in low-resource and informal domains (Joko et al., 2022, Sarkar et al., 24 Sep 2025). Multilingual and cross-lingual adaptation remains challenging: transfer learning and cross-lingual alignment have demonstrated competitive performance in some settings without in-language supervision (Kundu et al., 2018).

Future directions highlighted include:

• Detection and linking of non-canonical personal/nominal mentions via neural mention detectors leveraging world knowledge (Joko et al., 2022).
• Extension of joint modeling to handle implicit arguments, event coreference, and frame semantics (Aralikatte et al., 2019).
• End-to-end learning of entity and relation embeddings, moving beyond span-based heuristics (Verlinden et al., 2021).
• Advanced graph propagation and candidate-pruning for scalable cross-document models (Dong et al., 8 Apr 2025).
• Integration of multimodal signals, generative PLMs, and prosody/morphological analyzers for narrative and low-resource texts (Goel et al., 2022, Sarkar et al., 24 Sep 2025).

The trajectory of research on entity linking and coreference resolution is toward tightly integrated, modality-flexible, and knowledge-aware systems that support robust, scalable information extraction across genres, languages, and modalities.