Coreference Chain

The most common type of coreference chain, where a named entity is introduced and subsequently referenced by pronouns.

Example:

• Mention 1: Victoria
• Mention 2: she
• Mention 3: her

Chain: {Victoria₁, she₂, her₃}

This chain demonstrates anaphora, where the pronouns she and her depend on the antecedent Victoria for their interpretation. The head-finding heuristic identifies Victoria as the syntactic head, anchoring the entire chain.

Chains often include appositives and nominal descriptions that provide additional information about the entity.

Example:

• Mention 1: Dr. Elena Torres
• Mention 2: the chief of cardiology
• Mention 3: she
• Mention 4: Dr. Torres
• Mention 5: the renowned surgeon

Chain: {Dr. Elena Torres₁, the chief of cardiology₂, she₃, Dr. Torres₄, the renowned surgeon₅}

This illustrates how mention detection must identify both named entities and descriptive noun phrases as referring to the same individual.

A split antecedent occurs when a plural pronoun refers to multiple distinct entities introduced separately, requiring the merging of multiple antecedents into a single chain.

Example:

• Mention 1: Alice
• Mention 2: Bob
• Mention 3: they
• Mention 4: the two engineers

Chain: {Alice₁, Bob₂, they₃, the two engineers₄}

This is challenging for mention pair models, which classify pairwise coreference independently. Higher-order inference is often required to transitively merge these distinct antecedents.

Cataphora reverses the typical antecedent-pronoun order, with the pronoun appearing before the noun phrase it refers to, creating forward-looking chains.

Example:

• Mention 1: he
• Mention 2: the CEO
• Mention 3: Marcus Chen
• Mention 4: his

Chain: {he₁, the CEO₂, Marcus Chen₃, his₄}

This structure is common in literary openings and requires the coreference system to defer resolution until the explicit antecedent appears later in the discourse.

Bridging anaphora creates chains where mentions are not strictly coreferent but are inferentially linked through world knowledge rather than direct identity.

Example:

• Mention 1: the conference room
• Mention 2: the door
• Mention 3: it

Chain: {the conference room₁} → {the door₂, it₃}

While the door and it form a coreference chain, their relationship to the conference room is one of meronymic association (part-whole), not identity. This distinction is critical for accurate entity resolution.

A singleton entity is mentioned exactly once in a document with no coreferring mentions, yet it must still be identified as a distinct discourse referent.

Example:

• Mention 1: the quarterly earnings report

Chain: {the quarterly earnings report₁}

While singletons lack the complexity of multi-mention chains, proper identification is essential for downstream tasks like knowledge graph population and relationship extraction, where every entity must be accounted for.

The prerequisite step for building any coreference chain. Mention detection identifies all spans of text that refer to an entity, including:

• Named entities ("Barack Obama")
• Nominal phrases ("the president")
• Pronouns ("he")

Without accurate mention detection, downstream coreference models cannot form complete chains.

A neural architecture that scores all candidate antecedents for a given mention and selects the highest-ranked one. Unlike mention-pair models that make independent pairwise decisions, mention-ranking models:

• Consider the entire candidate set simultaneously
• Learn to compare antecedents against each other
• Produce more globally coherent coreference chains

This approach is the foundation of modern end-to-end systems like e2e-coref.

An iterative refinement technique that enables transitive reasoning across coreference chains. After an initial antecedent assignment, span representations are updated based on the representations of their predicted antecedents:

• If mention A refers to B, and B refers to C, the model infers A refers to C
• Multiple iterations propagate information through the entire chain
• Dramatically improves resolution of long-distance dependencies

An entity mentioned exactly once in a document with no coreferring mentions. Singleton entities still form a valid coreference chain of length one. Key considerations:

• They must be identified as distinct discourse referents
• Many benchmark evaluations (e.g., CoNLL-2012) exclude singletons from scoring
• Real-world applications often require singleton detection for complete entity inventories

A fixed-length vector encoding a contiguous sequence of tokens, critical for neural coreference. Computed by:

• Concatenating hidden states of the span's start and end tokens
• Applying attention over all tokens within the span
• Using models like SpanBERT optimized for span-level tasks

High-quality span representations enable the model to compare mentions and build accurate coreference chains.

A computational efficiency technique that restricts the candidate antecedent search space. For each mention, the model applies heuristic filters:

• Distance-based pruning: limit to nearby sentences
• Syntactic constraints: filter by agreement features (number, gender)
• Head-matching: require compatible head words

Pruning dramatically reduces the quadratic complexity of pairwise scoring while preserving high recall for valid coreference chains.