Nasty.Semantic.Coreference.Scorer (Nasty v0.3.0)

View Source

Generic scoring module for coreference resolution.

Scores pairs of mentions for coreference likelihood using multiple features:

  • Sentence distance (recency)
  • Gender and number agreement
  • String matching (exact and partial)
  • Entity type compatibility
  • Mention type patterns (pronoun-name, etc.)

All weights are configurable to allow tuning for different languages and domains.

Summary

Functions

distance_score(m1, m2, max_distance, weight)

Scores based on sentence distance (recency).

entity_type_score(m1, m2, weight)

Scores based on entity type match.

gender_agreement_score(m1, m2, weight)

Scores based on gender agreement.

number_agreement_score(m1, m2, weight)

Scores based on number agreement.

partial_match_score(m1, m2, weight)

Scores based on partial string match.

pronoun_name_boost(m1, m2, weight)

Boost score for pronoun-name pairs.

score_cluster_pair(cluster1, cluster2, opts \\ [])

Scores a pair of mention clusters for merging.

score_mention_pair(m1, m2, opts \\ [])

Scores a pair of mentions for coreference likelihood.

string_match_score(m1, m2, weight)

Scores based on exact string match (case-insensitive).

Functions

distance_score(m1, m2, max_distance, weight)

@spec distance_score(
  Nasty.AST.Semantic.Mention.t(),
  Nasty.AST.Semantic.Mention.t(),
  pos_integer(),
  float()
) :: float()

Scores based on sentence distance (recency).

Closer mentions get higher scores. Mentions beyond max_distance get 0 score.

entity_type_score(m1, m2, weight)

@spec entity_type_score(
  Nasty.AST.Semantic.Mention.t(),
  Nasty.AST.Semantic.Mention.t(),
  float()
) ::
  float()

Scores based on entity type match.

Returns weight if both have same entity type, 0 otherwise.

gender_agreement_score(m1, m2, weight)

@spec gender_agreement_score(
  Nasty.AST.Semantic.Mention.t(),
  Nasty.AST.Semantic.Mention.t(),
  float()
) ::
  float()

Scores based on gender agreement.

Returns weight if genders agree, 0 otherwise.

number_agreement_score(m1, m2, weight)

@spec number_agreement_score(
  Nasty.AST.Semantic.Mention.t(),
  Nasty.AST.Semantic.Mention.t(),
  float()
) ::
  float()

Scores based on number agreement.

Returns weight if numbers agree, 0 otherwise.

partial_match_score(m1, m2, weight)

@spec partial_match_score(
  Nasty.AST.Semantic.Mention.t(),
  Nasty.AST.Semantic.Mention.t(),
  float()
) ::
  float()

Scores based on partial string match.

Returns weight if one text contains the other, 0 otherwise.

pronoun_name_boost(m1, m2, weight)

@spec pronoun_name_boost(
  Nasty.AST.Semantic.Mention.t(),
  Nasty.AST.Semantic.Mention.t(),
  float()
) ::
  float()

Boost score for pronoun-name pairs.

These are common coreference patterns (e.g., "John... he"). Returns weight if one is pronoun and other is proper name, 0 otherwise.

score_cluster_pair(cluster1, cluster2, opts \\ [])

@spec score_cluster_pair(
  [Nasty.AST.Semantic.Mention.t()],
  [Nasty.AST.Semantic.Mention.t()],
  keyword()
) ::
  float()

Scores a pair of mention clusters for merging.

Uses average linkage: averages scores of all mention pairs between clusters.

Options

  • :merge_strategy - Linkage type (default: :average)
    • :average - Average of all pairwise scores
    • :best - Maximum pairwise score
    • :worst - Minimum pairwise score

score_mention_pair(m1, m2, opts \\ [])

@spec score_mention_pair(
  Nasty.AST.Semantic.Mention.t(),
  Nasty.AST.Semantic.Mention.t(),
  keyword()
) ::
  float()

Scores a pair of mentions for coreference likelihood.

Returns a float score between 0.0 and ~2.0, where higher scores indicate stronger coreference evidence.

Parameters

  • mention1 - First mention
  • mention2 - Second mention
  • opts - Scoring options
    • :max_distance - Maximum sentence distance (default: 3)
    • :weights - Custom weight configuration (default: @default_weights)

Examples

iex> score = Scorer.score_mention_pair(m1, m2, max_distance: 3)
0.85

string_match_score(m1, m2, weight)

@spec string_match_score(
  Nasty.AST.Semantic.Mention.t(),
  Nasty.AST.Semantic.Mention.t(),
  float()
) ::
  float()

Scores based on exact string match (case-insensitive).

Returns weight if texts match exactly, 0 otherwise.