Entity Resolution

The foundational step of identifying all spans of text that refer to an entity. Modern systems use span enumeration over all possible n-grams up to a maximum length, then prune low-scoring candidates using a mention scorer. Key signals include syntactic heads, named entity tags, and part-of-speech patterns. Without high-recall mention detection, downstream coreference cannot recover missing entities.

The dominant neural paradigm that scores all candidate antecedents for a given mention and selects the highest-ranked one. Unlike pairwise models that make independent decisions, mention-ranking uses a learned scoring function over span representations. The model computes a similarity score between the mention and each candidate antecedent, typically using biaffine attention or a feedforward network over concatenated features.

The process of encoding a contiguous token sequence into a fixed-length vector. Architectures like SpanBERT pre-train on span-level objectives, learning boundary representations that capture the internal structure of phrases. A span vector is typically computed by concatenating the hidden states of the span's start and end tokens with an attention-weighted sum over all tokens in the span, producing a rich representation that encodes both content and context.

An iterative refinement technique where span representations are updated based on the representations of their predicted antecedents. This enables transitive reasoning across chains: if mention A corefers with B, and B corefers with C, the model can propagate information to infer that A corefers with C. Multiple iterations allow the model to converge on globally consistent coreference chains rather than making greedy local decisions.

A computational efficiency technique that restricts the candidate search space. Without pruning, a mention would be scored against every preceding mention, yielding O(n²) complexity. Heuristic filters based on:

• Distance: only consider antecedents within a fixed window
• Syntactic constraints: filter by agreement in number, gender, and animacy
• Head matching: require compatible syntactic heads This dramatically reduces computation while maintaining high recall.

A deterministic, multi-pass approach that applies a series of high-precision rules in cascading order. Each sieve resolves progressively more ambiguous mentions:

• Exact string matching for repeated names
• Head matching for nominal coreference
• Pronominal resolution using gender and number agreement
• Discourse salience for the most ambiguous cases The sieve architecture provides interpretability and predictable failure modes, making it suitable for high-stakes applications.