Head-Finding Heuristic

The heuristic identifies the single word that carries the core semantic and syntactic properties of a noun phrase. For example, in the span 'the large red car', the head is 'car'. Rules typically traverse a parse tree, prioritizing nouns over modifiers. Common rules include selecting the rightmost noun in a flat NP or the child of an NP node that is itself a noun. This reduction from a multi-word span to a single lexical item simplifies downstream feature extraction.

Once the head word is extracted, it becomes the basis for critical coreference features:

• Lexical Match: Checking if the head of a pronoun's candidate antecedent matches the head of another mention.
• Gender Agreement: Using the head word to query a gender lexicon (e.g., 'he' for 'man', 'she' for 'woman').
• Number Agreement: Determining plurality from the head (e.g., 'cars' vs. 'car').
• Semantic Class: Mapping the head to a WordNet synset or named entity type.

Head matching enables computationally cheap, high-precision antecedent pruning. A deterministic sieve can immediately filter out candidate antecedents whose head words do not match the pronoun's expected head. For instance, when resolving 'it', the system can prune all antecedents whose head is not a non-human noun. This drastically reduces the candidate pool before applying more expensive neural scoring functions, improving overall system efficiency.

Heuristics must handle complex syntactic structures robustly:

• Conjunctions: In 'John and Mary', the head is the conjunction itself, but the heuristic may extract both 'John' and 'Mary' as a compound head.
• Prepositional Phrases: In 'the roof of the building', the head is 'roof', not 'building'. Rules must follow the NP → NP PP structure.
• Appositives: In 'Steve Jobs, the CEO of Apple,' the head is 'Steve Jobs', with the appositive providing additional information.

In the canonical deterministic coreference system (Lee et al., 2013), head-finding is a fundamental preprocessing step. The 'Head Match' sieve is one of the earliest and highest-precision passes. It links mentions only if their extracted heads are identical. This sieve resolves straightforward coreference chains like 'Apple Inc.' and 'Apple' before later sieves attempt more complex resolutions using pronominal agreement or discourse salience models.

Rule-based head-finding is brittle against parse errors and fails on non-compositional phrases. In 'the New York Times', the head is the entire multi-word expression, not just 'Times'. Modern neural coreference models (e.g., e2e-coref) bypass explicit head rules by learning span representations directly from token embeddings. However, head-finding heuristics remain crucial for fast, interpretable baselines and for generating silver-standard training data for mention detection.

The prerequisite subtask of identifying all spans of text that refer to an entity. Head-finding is applied immediately after mention detection to extract the syntactic head of each candidate span.

• The head word is used as a lexical feature for mention-pair and mention-ranking models.
• Accurate head-finding prevents the model from learning spurious correlations with determiners or modifiers.
• Example: In the span "the complex neural network," the head is network, not "complex" or "neural."

A computational efficiency technique that restricts the candidate antecedent search space using heuristic filters. Head-finding enables agreement-based pruning by comparing the syntactic heads of mentions.

• Number agreement: A singular head cannot be coreferent with a plural head.
• Gender agreement: Pronominal heads must match the gender of candidate antecedent heads.
• This drastically reduces the number of pairwise scores that must be computed in mention-ranking models.

A fixed-length vector encoding a contiguous token sequence. In neural coreference, the head-finding heuristic determines which token's hidden state receives the highest weight when constructing the span embedding.

• Architectures often concatenate the start token, end token, and head token representations.
• An attention mechanism over the span's tokens, weighted by head likelihood, produces a soft head vector.
• This ensures the representation focuses on the syntactic core of the mention rather than peripheral words.

A neural coreference architecture that scores all candidate antecedents for a given mention and selects the highest-ranked one. The head word is a critical input feature for the pairwise scoring function.

• The model computes a biaffine attention score between the mention's head representation and each candidate antecedent's head representation.
• Head features like part-of-speech tags and dependency labels are concatenated into the scoring vector.
• This architecture replaced independent pairwise classification with a globally informed ranking objective.

A syntactic theory governing the distribution of anaphors, pronominals, and referring expressions. Head-finding identifies the governing category of a pronoun by locating its syntactic head and traversing the parse tree.

• Principle A: Reflexives like "itself" must be bound within their local domain.
• Principle B: Pronouns like "it" must be free within their local domain.
• These structural constraints are encoded as binary features in rule-based sieve architectures and as embeddings in neural models.

The task of analyzing the grammatical structure of a sentence to establish directed relationships between words. Head-finding heuristics are often implemented as deterministic tree traversal rules on dependency parse outputs.

• A common rule: traverse the dependency tree upward from the mention's root token until hitting a noun or proper noun.
• Dependency labels like nsubj and dobj provide strong signals for identifying the semantic head.
• Modern end-to-end systems may learn head-finding implicitly, but explicit parsing remains a powerful inductive bias.