Split Antecedent

The most common pattern where a plural pronoun refers to two entities joined explicitly or implicitly.

Example:

• Alice went to the library. Bob stayed home. They planned to meet later.

Resolution Mechanism:

• The system must identify Alice and Bob as separate singleton mentions
• Merge them into a single plural entity for the pronoun they
• This requires tracking discourse referents across sentence boundaries

A pronoun refers to an entity and one of its attributes or possessions mentioned separately.

Example:

• The company announced record profits. Its stock price surged. They are now market leaders.

Complexity:

• Requires understanding that the company and its stock price are distinct but related
• The plural pronoun they merges an organization with its financial instrument
• Demands ontological knowledge about entity–attribute relationships

Multiple participants in a complex event are introduced across separate clauses and later referenced collectively.

Example:

• The plaintiff filed the motion. The defendant responded immediately. They reached a settlement before trial.

Resolution Requirements:

• Identify semantic roles (agent, patient) across sentences
• Recognize that plaintiff and defendant form a natural pair in legal discourse
• Merge them despite no syntactic coordination in the source text

A plural pronoun refers back to multiple items introduced in a list structure rather than a single coordinated phrase.

Example:

• The team needed three things: a new server, updated software, and trained staff. They were all acquired within budget.

Resolution Challenge:

• The antecedents are distributed across a list structure
• The pronoun they must resolve to the set of all three items
• Requires recognizing list semantics and collective reference

The plural pronoun refers to a set that was never explicitly enumerated but is inferable from individual mentions.

Example:

• The CEO approved the budget. The CFO signed the contracts. The COO hired the team. They transformed the company.

Inference Required:

• No explicit coordination phrase like "the three executives"
• The system must construct the set {CEO, CFO, COO} from discourse context
• Relies on recognizing parallel roles and shared organizational membership

Two entities introduced in contrastive or comparative structures are later referenced jointly.

Example:

• Traditional methods failed to scale. Neural approaches required too much data. Both proved inadequate for the task.

Resolution Mechanism:

• The quantifier both explicitly signals a dual antecedent
• System must identify the two contrasted entities
• Requires understanding discourse relations like contrast and comparison

The complete ordered set of all mentions within a discourse that refer to a single entity. A split antecedent creates a chain where a plural pronoun merges two previously distinct chains. For example, 'Alice' and 'Bob' each have their own chain until 'they' links both antecedents into a unified plural reference. Tracking these chains requires higher-order inference to propagate the merged entity representation across subsequent mentions.

A neural architecture that scores all candidate antecedents for a given mention and selects the highest-ranked one. For split antecedents, the model must rank a set of antecedents rather than a single span. Standard pairwise ranking fails here—systems must either enumerate all possible antecedent combinations or use clustering-based approaches that can merge multiple mentions into a composite entity representation.

A non-identity anaphoric relationship where a definite noun phrase refers to an entity inferentially linked to a previously introduced discourse referent. Often confused with split antecedents, bridging anaphora does not require strict coreference. For example, 'I bought a car. The engine is loud.'—'the engine' bridges to 'a car' but is not coreferent. Split antecedents require identity: 'Alice arrived. Bob arrived. They sat down.'—'they' is Alice and Bob.

An iterative refinement technique where span representations are updated based on their predicted antecedents, enabling transitive reasoning. Critical for split antecedent resolution because the model must propagate information across multiple iterations: first linking 'they' to 'Alice', then linking 'they' to 'Bob', and finally merging both into a composite representation. Without higher-order loops, the model cannot accumulate the full set of antecedents.

The standard evaluation dataset for coreference resolution derived from OntoNotes 5.0. Split antecedents are explicitly annotated in this corpus—plural pronouns can be linked to multiple antecedent spans. However, they represent a small fraction of total coreference links, making them a challenging long-tail phenomenon. Models often underperform on split antecedent cases because training data is dominated by singular pronoun-antecedent pairs.

A discourse coherence theory that models local focus by tracking a ranked set of forward-looking centers and a single backward-looking center. Split antecedents violate the standard centering assumption that each utterance has exactly one backward-looking center. When 'they' refers to both Alice and Bob, the backward-looking center is a composite entity, requiring extensions to the theory to handle plural anaphora with multiple antecedents.