Skip to content
Back to search results

Coreference Resolution

Introduced by
Hobbs (1978); Lee et al. (2017, neural end-to-end)
Year
1978
Type
NLP information-extraction task
Font size

What does it do?

Coreference resolution is a natural-language-processing task that detects when different expressions in a text refer to the same entity — for example a name, a later pronoun, and a descriptive phrase all pointing at one person.

Rooted in early linguistic work by Hobbs (1978) and advanced by the end-to-end neural model of Lee et al. (2017), it improves the quality of information extraction and text understanding.

The intuition behind the method

When we read, we automatically know that "she", "the author", and "Marie Curie" can all be the same person. A machine does not. Coreference resolution teaches a system to make those links: it spots every expression that could refer to an entity (a mention), decides which mentions point to the same thing, and groups them together so the rest of a pipeline reads the text the way a human does.

History

Coreference resolution grows out of early computational linguistics, notably Hobbs (1978), who formalised how pronoun references can be resolved over syntactic structure. The field later moved from such rule-based and feature-based systems toward learned models, with Lee et al. (2017) introducing an end-to-end neural approach that scores mentions and their links directly without a hand-built pipeline of intermediate steps.

When to use it

Coreference resolution fits when you have text in English or another supported language and want to track entities across a document — typically as a building block for information extraction or deeper text understanding. A NER pass beforehand helps. It is descriptive and explanatory in purpose and needs at least a modest amount of text (around 20 units) to be meaningful; without text data it cannot run.

Limitations and considerations

  • Works only for English or other supported languages; coverage outside those is limited.
  • Accuracy depends on reliable upstream steps — weak mention detection or missing NER degrades the clusters.
  • Ambiguous or long-distance references remain hard, so some links will be wrong.

Source conditions do not replace independent scientific review or validation by an analysis engine.

Strengths

  • Links scattered mentions — names, pronouns, and descriptive phrases — back to a single entity, which sharpens information extraction.
  • Improves the quality of downstream text-understanding tasks by resolving who or what each expression refers to.
  • Supports both rule-based linguistic approaches and modern end-to-end neural models.

Applications

  • Information extraction — consolidating all references to an entity across a document.
  • Improving question answering and reading comprehension by resolving pronouns to their referents.
  • Building knowledge graphs and entity-centric summaries from text.
  • Cleaning text for downstream analysis in health, social, economic, engineering, and business research corpora.

How it works

How it works

1

The text is prepared for analysis: it is segmented into sentences and tokens and run through the supporting language pipeline. A named-entity recognition (NER) step is a useful precursor here, because knowing where entities sit makes the later mention detection more reliable.

2

Candidate mentions are identified — the spans of text that could refer to an entity, such as proper names, noun phrases, and pronouns. This produces the set of expressions whose referents the method must work out.

3

For each pair of candidate mentions the system decides whether they refer to the same entity. Classical approaches apply linguistic rules over syntax and recency, while neural end-to-end models score mention pairs directly from learned representations.

4

The pairwise links are resolved into clusters: every mention judged to point at the same entity is grouped together. Each cluster represents one real-world entity, and downstream tasks can now treat "she", "the author", and the original name as a single referent.

Common pitfalls

  • Skipping a NER or mention-detection step leaves the linker working from poor candidates and produces noisy clusters.
  • Applying a model built for one language to another supported-language gap silently lowers accuracy.
  • Treating every pairwise link as certain — ambiguous pronouns and distant antecedents are common error sources and should be inspected.

Frequently asked questions

What is the difference between coreference resolution and named-entity recognition?

NER finds and labels entity mentions (such as people or organisations) in text. Coreference resolution goes a step further by deciding which of those mentions — plus pronouns and descriptive phrases — refer to the same real-world entity, grouping them into clusters. NER is often a useful precursor to coreference.

Do I need a NER step first?

It is not strictly required, but a NER pass is a helpful precursor: knowing where entities sit makes mention detection and the resulting clusters more reliable.

Does it work in languages other than English?

It works for English or other supported languages. Coverage outside the languages your model supports is limited, so the chosen model must match the language of your text.

Rule-based or neural — which approach should I use?

Rule-based methods, in the tradition of Hobbs (1978), apply linguistic rules over syntax and recency and are transparent. End-to-end neural models, such as Lee et al. (2017), learn to score mentions and their links directly and generally handle harder cases better at higher computational cost.

Related methods

Sources

  1. Lee, K. et al. (2017). End-to-end Neural Coreference Resolution. EMNLP.

    Open publication reference
    Copied.
  2. Hobbs, J.R. (1978). Resolving Pronoun References. Lingua, 44(4), 311-338.

    Open DOI reference
    Copied.