coreF

Coreference resolution, often abbreviated as coref, is the task of determining which expressions in a text refer to the same real-world entity. A coreference cluster collects all mentions—such as pronouns, proper names, and noun phrases—that refer to the same entity. For example, in the sentence “Alice arrived late because she was stuck in traffic,” the pronoun “she” corefers with “Alice.”

Approaches to coreference resolution have evolved from rule-based and feature-engineered systems to statistical and neural models. Traditional methods rely on hand-crafted features and pairwise or cluster-level decisions to link mentions, while modern methods use neural architectures to learn representations for mentions and antecedents and often employ end-to-end or graph-based frameworks. End-to-end neural coreference models typically propose candidate spans and score links to form coherent clusters without requiring extensive feature engineering.

Prominent datasets for coreference research include OntoNotes 5.0 and the CoNLL-2012 shared task. Evaluation commonly uses a combination of metrics such as MUC, B3, and CEAF, often reported together as a CoNLL score. These metrics assess how well a system reconstructs the correct mention clusters across documents.

Coreference resolution is a foundational component in many NLP applications, including information extraction, question answering, document summarization, and dialogue systems. Challenges remain, including resolving pronouns across sentence boundaries, handling long-range dependencies, and adapting models to languages with different pronoun and argument structures. Zero pronouns and the integration of world knowledge continue to be active areas of research.