Collective Entity Linking

Unlike local models that disambiguate each mention in isolation, collective approaches enforce a global interdependence constraint. The system evaluates the compatibility of all linked entities simultaneously, seeking a configuration where every pair of entities exhibits high semantic relatedness. This prevents a single high-confidence but contextually incoherent link from derailing the entire document's interpretation.

The core objective function is the maximization of semantic coherence across the candidate entity set. The system computes pairwise relatedness scores using measures like Wikipedia Link-based Measure (WLM) or Normalized Google Distance (NGD). The final entity assignment is the subgraph of the knowledge base that maximizes the aggregate relatedness score, ensuring all linked entities belong to a consistent topical domain.

Collective linking is typically implemented using graph algorithms on a mention-entity graph:

• Dense subgraph approximation: Finds a high-coherence subgraph connecting all mentions.
• Random walks: Uses PageRank-like algorithms to propagate topical evidence across the graph.
• Loopy Belief Propagation: Iteratively passes messages between mention nodes to converge on a globally consistent assignment. These methods explicitly model the conditional dependencies between linking decisions.

To reduce the combinatorial search space, collective linkers often inject a topical coherence prior. The system first infers the document's dominant topics using a model like Latent Dirichlet Allocation (LDA), then biases entity selection toward candidates strongly associated with those topics. This prior acts as a regularizer, penalizing entity configurations that are semantically scattered and favoring those that cluster around a central theme.

The joint optimization of N mentions over K candidates per mention yields a search space of K^N configurations, making exhaustive search intractable. Production systems employ approximation strategies:

• Greedy hill-climbing: Iteratively flips the assignment of the most incoherent mention.
• Integer Linear Programming (ILP): Formulates the problem with binary variables and solves for a globally optimal configuration under relaxed constraints.
• Beam search: Maintains a fixed-width beam of the most promising partial configurations.

A local model disambiguates each mention independently using only its immediate context window and prior probability P(e|m). A collective model adds a pairwise compatibility term ψ(e_i, e_j) that scores the semantic relatedness between entity assignments. The collective approach resolves cases where a mention is locally ambiguous but globally constrained—for example, 'Michael Jordan' in a document about machine learning is correctly linked to the Berkeley professor, not the athlete, because the professor co-occurs coherently with other CS entities.

The core subtask that collective entity linking solves at scale. While standard NED resolves a single mention in isolation, collective approaches use the disambiguation of one mention to constrain the possible identities of others. This resolves cases like distinguishing 'Apple' the company from the fruit by analyzing surrounding entities like 'iPhone' or 'orchard'.

The initial retrieval phase that produces a shortlist of possible knowledge base entries for each mention. Collective systems rely on high-recall candidate sets because the global optimization step can later filter out false positives. Techniques include:

• Surface form dictionaries built from anchor text and redirects
• Approximate nearest neighbor search over entity embeddings
• Alias tables from Wikipedia disambiguation pages

Dense vector representations of knowledge graph entities that encode semantic properties and relational structure. Models like TransE and DistMult learn embeddings that position related entities close together in vector space. Collective entity linking uses these embeddings to compute pairwise coherence scores between candidate entities, measuring how well they fit together in a unified topic.

The task of identifying all expressions that refer to the same real-world entity, including pronouns and definite noun phrases. Collective entity linking depends on accurate coreference chains because:

• A pronoun like 'it' inherits its entity link from its antecedent
• Resolving 'the company' to 'Apple Inc.' prevents contradictory links
• Building coherent discourse models enables topical agreement scoring

The capability to correctly identify when a mention refers to an entity not present in the target knowledge base. In collective linking, a false positive link to an existing entity can cascade errors across the document. Nil prediction prevents this by:

• Setting a coherence threshold below which a mention is left unlinked
• Using out-of-distribution detection on entity embedding similarity scores
• Flagging mentions for downstream knowledge base population

The task of clustering mentions of the same entity across multiple documents. While collective entity linking operates within a single document, cross-document coreference extends this to corpus scale. The two tasks share the same global optimization principle: decisions about one mention constrain others, but across documents instead of within them.