Coreference Resolution

Coreference Resolution Linking Entity Nlp Needs-Review

What it is

Coreference resolution identifies and links all mentions of the same entity in a document. For example, “John”, “he”, and “the CEO” might all refer to the same person. The output is a partition of mentions into coreference clusters, one per unique entity. Coreference is critical for understanding document-level meaning and for downstream tasks like summarization and QA.

[illustrate: Text with mentions colored by coreference cluster; pronouns linked to antecedents via arcs]

How it works

  1. Mention detection: Identify candidate mentions (usually noun phrases, pronouns)

    • “John Smith”, “he”, “the CEO”, “John”

  2. Linking: Determine which mentions refer to same entity

    • Pairwise linking: Score all pairs (mention_i, mention_j); cluster based on scores
    • Global clustering: Jointly cluster all mentions

  3. Neural approaches:

    • BiLSTM or BERT encoder for mention representations
    • Pairwise classifier: P(coreferent | mention_i, mention_j)
    • Beam search or agglomerative clustering to form clusters

Example

Text:
"John met Mary. He gave her a gift. The gift was expensive.
Mary appreciated it. She thanked John."

Coreference clusters:
- Cluster 1 (John): "John" (sent 1), "He" (sent 2), "John" (sent 5)
- Cluster 2 (Mary): "Mary" (sent 1), "her" (sent 2), "She" (sent 5)
- Cluster 3 (gift): "a gift" (sent 2), "The gift" (sent 3), "it" (sent 4)

Variants and history

Coreference resolution dates to the 1990s with rule-based systems. Mention-pair models (2004+) trained pairwise classifiers. Entity-mention models score mentions against entity representations. Span-based models (Lee et al., 2017) directly score mention spans without pre-detected mentions. Contextualized embeddings (BERT, 2018+) improved performance. Modern systems achieve 75–80% F1 on OntoNotes benchmark.

When to use it

Use coreference resolution for:

  • Document understanding and question-answering
  • Summarization (tracking who did what)
  • Information extraction (entity-centric)
  • Machine translation (pronoun handling)
  • Dialogue systems (tracking referents)

Coreference is challenging; performance drops significantly on out-of-domain text. Joint approaches (NER + coreference) sometimes improve results.

See also