Bridging Anaphora is the referential phenomenon where an entity is introduced through its conceptual relationship to a previously mentioned entity rather than by direct repetition or pronominalization — requiring commonsense or world knowledge to infer the connection between the new reference and its antecedent, unlike standard coreference where the relationship is identity.
The Core Distinction from Standard Coreference
Standard coreference (identity anaphora) links expressions that refer to the same entity:
"Apple released a new phone. It was praised by reviewers." → "It" = "a new phone" (identity).
Bridging anaphora links expressions where one entity is associated with but not identical to the antecedent:
"I drove to the conference. The parking lot was full." → "The parking lot" is not the conference — it is PART OF the conference venue. The connection requires world knowledge: conferences are held in buildings with parking lots.
The "bridge" is the implicit relationship connecting the new expression to its antecedent: part-whole, set-member, attribute, event-participant, or functional association.
Taxonomy of Bridging Relations
Part-Whole Relations (Meronymy):
- "I bought a car yesterday. The engine was making a strange noise." → engine is part-of car.
- "She entered the building. The elevator was out of service." → elevator is part-of building.
- "The patient had surgery. The incision was carefully closed." → incision is part-of surgical procedure.
Set-Member Relations:
- "The committee voted. The chair abstained." → chair is a member-of committee.
- "I love Italian food. The pasta here is exceptional." → pasta is instance-of Italian food.
- "Several engineers attended. The most senior gave a presentation." → most senior is member-of engineers.
Event-Participant / Event-Result Relations:
- "There was a car accident on Main Street. The victim was taken to the hospital." → victim is participant-of accident.
- "The company went bankrupt. The creditors received nothing." → creditors are participants-in bankruptcy.
- "The bomb exploded. The debris scattered for blocks." → debris is result-of explosion.
Functional / Attribute Relations:
- "She checked into the hotel. The room had a view of the bay." → room is functionally-associated-with hotel stay.
- "He applied for the job. The salary was competitive." → salary is attribute-of job.
Why Bridging Requires Commonsense Knowledge
Standard coreference can be resolved largely through surface features: number agreement, gender agreement, proximity, and syntactic constraints. Bridging resolution requires:
1. Ontological knowledge: Knowing that cars have engines, buildings have elevators, committees have chairs.
2. Script knowledge: Understanding typical event structures — accidents have victims; surgeries have incisions; job applications have salaries.
3. Context-sensitive inference: The same phrase may bridge differently in different contexts. "The driver" bridges to a car in one context and to a sports event in another.
No surface-level feature reliably indicates a bridging relation. The system must infer that a definite noun phrase ("The parking lot") is bridging rather than introducing a new entity, and then identify the antecedent from all previously mentioned entities.
Corpus Resources
ISNotes: 10,000 bridging instances in news text, annotated for bridging type and antecedent. The most widely used benchmark for English bridging resolution.
BASHI: Bridging anaphora annotation in the Heidelberg Text Corpus. Focuses on German, testing cross-linguistic bridging patterns.
Prague Discourse Treebank: Czech corpus with bridging annotations, enabling cross-linguistic study of bridging phenomena.
Why Standard Coreference Systems Fail
Standard coreference resolvers (trained on OntoNotes) are optimized for identity coreference and fail on bridging for two reasons:
Mention Scope: Standard resolvers learn to link mentions that share lexical roots, pronominal forms, or gender/number agreement. Bridging links "the parking lot" to "the conference" — completely different lexical items with no pronominal connection.
Training Signal: OntoNotes does not annotate bridging relations, so standard models are never trained to recognize them. They either ignore the bridging expression entirely (treating it as a new entity) or incorrectly link it as an identity coreference to a superficially similar antecedent.
Approaches to Bridging Resolution
Relation Classification: Enumerate candidate antecedents and classify the relation type (part-whole, set-member, event-result, none). Requires training on bridging-annotated corpora.
Knowledge Graph Grounding: Use ConceptNet, Wikidata, or FrameNet to enumerate known part-whole and functional relationships between entity types, providing bridging candidates consistent with structured world knowledge.
Large Language Model Prompting: GPT-4 class models, trained on massive text, implicitly encode many bridging relationships and can resolve bridging in few-shot settings by leveraging their broad world knowledge.
Discourse Coherence Models: Bridging references are motivated by discourse coherence — they connect the current sentence to an entity already in the discourse model. Coherence-aware models that track the discourse state are better positioned to identify bridging.
Practical Implications
Bridging anaphora failures cause subtle but systematic errors in downstream NLP systems:
- Summarization: "The door" appearing in a summary without establishing the house creates an unresolved reference.
- Information Extraction: "The victim was a teacher" attached to no specific accident loses its informational value.
- Reading Comprehension: Questions about parts or participants of events cannot be answered if the bridge is not resolved.
Bridging Anaphora is inference linking — connecting entities through conceptual relationships of containment, membership, causation, and function rather than identity, requiring the world knowledge that standard coreference systems do not possess.