Mention Detection

The primary mechanism involves identifying the start and end tokens of a mention. Modern systems use learned span representations rather than hand-crafted rules. Key aspects include:

• Detecting nested mentions (e.g., 'the CEO of Microsoft' contains 'Microsoft')
• Handling discontinuous mentions in some languages
• Resolving head-finding heuristics to identify the syntactic head of each span
• Pruning invalid spans using mention likelihood scores to reduce computational overhead

Detection systems must distinguish between different mention types to enable accurate coreference. The standard OntoNotes taxonomy includes:

• Named entities: Proper nouns referring to specific people, organizations, locations
• Nominal mentions: Common noun phrases that refer to entities ('the company', 'a large city')
• Pronominal mentions: Pronouns that anaphorically refer to antecedents ('she', 'it', 'they')
• Singleton detection: Identifying entities mentioned exactly once, which still require representation as distinct discourse referents

Modern systems like the e2e-coref architecture compute a mention score for every possible span up to a maximum length. This score represents the model's confidence that a span constitutes a valid mention. The process involves:

• Enumerating all spans within a length constraint (typically up to 10-30 tokens)
• Computing span representations from token embeddings using attention or concatenation of boundary vectors
• Applying a feedforward network to produce a scalar mention score
• Using span pruning to retain only the top-k highest-scoring spans for downstream coreference scoring

Detection accuracy improves by incorporating linguistic constraints that filter implausible mentions. Common filters include:

• Part-of-speech tag filtering: Requiring mentions to contain at least one noun or pronoun
• Syntactic parse constraints: Using dependency parses to identify valid noun phrase boundaries
• Binding theory constraints: Applying syntactic principles governing pronoun-antecedent relationships
• Head-finding rules: Identifying the semantic head word to ensure the span represents a coherent entity reference

Mention detection performance is measured independently from coreference linking using precision, recall, and F1 on identified mention boundaries. Key metrics include:

• Mention Detection F1: Evaluates exact span boundary matching against gold annotations
• B-cubed precision and recall: Measures entity-level detection accuracy across coreference chains
• CoNLL-2012 shared task metrics: The standard benchmark combining MUC, B-cubed, and CEAF scores
• Boundary errors vs. type errors: Distinguishing between incorrect span offsets and misclassified mention types

In pro-drop languages like Japanese, Chinese, and Spanish, entities may be referenced without explicit textual mentions. Detection systems must handle:

• Zero pronouns: Syntactically omitted arguments that are semantically understood
• Implicit arguments: Verb arguments inferred from context without surface realization
• Empty category detection: Using syntactic cues to identify positions where a mention should be inferred
• Cross-linguistic variation: Adapting detection strategies for languages with different pro-drop parameters