What is coreference resolution?

The task of linking together mentions in text which corefer, i.e. refer to the same discourse entity in the discourse model, resulting a set of coreference chains (also called clusters or entities).

What can coreference mentions be?

Mentions can be definite NPs or indefinite NPs, pronouns (including zero pronouns) or names.

What is the surface form of an entity mention linked to?

Its information status (new, old, or inferrable), and how accessible or salient the entity is.

Are all NPs referring expressions?

Some NPs are not referring expressions, such as pleonastic it in It is raining.

Provide examples of corpora for human-labeled coreference annotations

OntoNotes for English, Chinese, and Arabic, ARRAU for English, and AnCora for Spanish and Catalan

What can coreference resolution start with?

Mention detection can start with all nouns and named entities and then use anaphoricity classifiers or referentiality classifiers to filter out non-mentions.

What are the three common architectures for coreference using feature-based or neural classifiers?

  1. mention-pair
  2. mention-rank
  3. entity-based

How do modern coreference systems work?

Modern coreference systems tend to be end-to-end, performing mention detection and coreference in a single end-to-end architecture. Algorithms learn representations for text spans and heads, and learn to compare anaphor spans with candidate antecedent spans.

How are coreference systems evaluated?

Coreference systems are evaluated by comparing with gold entity labels using precision/recall metrics like MUC, B3, CEAF, BLANC, or LEA.

What are Winograd Schema Challenge problems?

The Winograd Schema Challenge problems are difficult coreference problems that seem to require world knowledge or sophisticated reasoning to solve.

How can we exhibit gender bias?

Coreference systems exhibit gender bias which can be evaluated using datasets like Winobias and GAP.