Coreference Resolution is the NLP task of identifying all expressions in a document that refer to the same real-world entity and grouping them into coreference chains — determining that "Obama," "The President," "he," and "the former senator from Illinois" within a document all refer to the same individual, enabling models to track entity identity across discourse boundaries.

The Core Challenge

Natural language routinely refers to entities using different expressions across a document:

"Barack Obama delivered his inaugural address in 2009. The 44th President of the United States outlined an ambitious agenda. He called for unity across party lines, and his speech was praised for its rhetorical power."

A system must recognize: {Barack Obama, his, The 44th President of the United States, He, his} → one coreference cluster referring to Barack Obama. {his inaugural address, his speech} → one cluster referring to the same speech. Without coreference resolution, each pronoun and definite description is an ungrounded dangling reference.

Types of Referring Expressions

The Two-Stage Pipeline

Stage 1 — Mention Detection: Identify all candidate mention spans in the document. A mention is any noun phrase that could potentially refer to an entity. Challenges: mentions can be nested ("the president of [Apple]"), and not all noun phrases are mentions (generic references like "a dog can bite" do not corefer with specific dogs).

Modern systems use span-based approaches: enumerate all spans up to a maximum length, score each as a valid mention using a trained classifier, then prune to a manageable candidate set (typically top K spans by mention score).

Stage 2 — Pairwise Coreference Scoring: For each pair of candidate mention spans, score the probability that they corefer. The score combines:

• Mention representations: BERT-based span representations encoding the mention text and surrounding context.
• Antecedent features: Distance between mentions (nearby mentions are more likely to corefer), genre (document type), speaker (in dialogue).
• Compatibility: Gender agreement ("he" cannot corefer with "Angela Merkel" unless the model is wrong about the gender), number agreement, animacy.

Stage 3 — Cluster Formation: Group pairwise coreferent mention pairs into entity clusters using transitivity: if A coreferss B and B coreferss C, then A, B, and C form a cluster. The standard approach uses a "mention-ranking" algorithm where each mention selects its most likely antecedent, and clusters are formed by following antecedent links.

Neural Coreference Resolution

The field was transformed by end-to-end neural approaches that jointly learn mention detection and coreference scoring:

e2e-coref (Lee et al., 2017): First end-to-end neural coreference system. Uses LSTMs to encode documents, spans as pooled token representations, and learns mention and coreference scoring jointly. Eliminated hand-crafted feature engineering.

SpanBERT Coreference (Joshi et al., 2019): Uses SpanBERT's specially pre-trained span representations (optimized for span boundary prediction) within the e2e-coref framework. SpanBERT's span-level self-supervised objective aligns directly with coreference's span-centric structure. Achieved large gains over LSTM-based systems.

LingMess and Cluster-Level Models: More recent approaches maintain cluster representations that update as mentions are resolved, enabling the model to use the accumulated cluster context (all previously resolved mentions of an entity) when resolving new mentions.

Why Coreference Resolution Matters

• Document Summarization: Summaries must avoid dangling pronouns ("He said the deal was complete" — who is "he"?). Coreference chains enable substituting the canonical entity name for pronouns.
• Information Extraction: Extracting "the company's CEO said profits increased" as a fact about a specific company requires resolving "the company" to its antecedent.
• Question Answering: "When did she found the company?" requires resolving "she" to the named individual from earlier context.
• Machine Translation: Grammatical gender agreement across sentences requires tracking which entities are being referred to.
• Knowledge Base Population: Aggregating facts about an entity across a document requires knowing which mentions all refer to that entity.

Evaluation and Benchmarks

OntoNotes: The primary benchmark. Multi-genre corpus (newswire, web, broadcast, telephone conversations) with coreference annotations. Evaluated using CoNLL F1 — the average of MUC, B³, and CEAF F1 metrics, each capturing different aspects of cluster quality.

GAP (Gender-Ambiguous Pronouns): Tests coreference for gender-ambiguous English pronouns, revealing biases in systems that associate gender with occupation names.

Winograd Schema Challenge: Binary pronoun resolution requiring commonsense inference — represents the hardest end of the coreference spectrum.

Remaining Challenges

• Long Documents: Standard neural approaches process 512-token windows. Full-book or long-report coreference remains challenging.
• Singular "they": Non-binary pronoun usage requires resolving "they" to singular entities, violating traditional number agreement heuristics.
• Cross-Document Coreference: Recognizing that "Apple" in two different news articles refers to the same company, even without shared document context.

Coreference Resolution is connecting the dots across a document — linking every varied reference (pronoun, definite description, proper name, nominal) back to the single real-world entity it represents, enabling models to track who and what a discourse is actually about.