Is bridging the same as coreference?

In the non-identity sense used here, no. It links related referents instead of identifying two mentions as the same referent.

Does the antecedent always name an object?

No. Some treatments also allow events or larger stretches of discourse.

Can world knowledge settle every case?

No. It can suggest relationships while leaving several candidates plausible.

AI Glossary / Natural language and speech / Advanced

Bridging inference

Definition

A bridging inference connects new information in a text to information already available by supplying an unstated relationship. In the referential sense used here, it helps a reader identify a newly mentioned entity through its association with an earlier entity or event. The two referents need not be the same thing.

What it means

Consider “We bought a bicycle. The saddle was uncomfortable.” The saddle has not previously been mentioned, yet a reader can connect it with the bicycle using knowledge about bicycle parts. That connection helps the second sentence fit into the developing account.
This differs from “We bought a bicycle. It was expensive”, where it refers to the bicycle itself. Bridging adds a related referent; identity coreference returns to an existing one. Terminology varies across discourse research, and broader uses of bridging can include other coherence-building inferences. This entry focuses on the referential phenomenon important for language-understanding systems.

How it works

An interpreter notices a mention whose referent needs contextual support, considers earlier mentions or events, and looks for a plausible relationship. Possibilities include part - whole relations, people associated with organizations, and participants associated with events. Linguistic cues, topic continuity, and background knowledge help distinguish candidates.
For an AI system, recognizing that a mention requires bridging and choosing its antecedent are separate decisions. A system may correctly identify the saddle as a bridging mention but attach it to the wrong bicycle when several bicycles are being discussed.
A useful representation preserves separate entities and records a relationship between them. It should also preserve uncertainty when the context does not identify a unique antecedent. Familiar associations provide evidence for interpretation; they do not license arbitrary additions to a factual summary.

Three practical examples

A product-support assistant reads “I unpacked a desk. The drawer was jammed.” It considers the drawer part of the desk rather than treating drawer and desk as interchangeable names.
A news-reading tool connects “The exhibition opened yesterday” with a subsequent mention of “the curator”. The inferred role relationship helps organize the account without assuming that every exhibition has identical staffing.
A document search system interprets a later reference to “the battery” in a discussion of a portable speaker. If the same passage also discusses a phone, it must use context rather than simply choose the closest device word.

Think of it this way

Imagine adding a new stop to a map and drawing a road to a place already marked. The road explains how the locations connect; it does not merge them into one location. Bridging inference supplies that connecting road in a reader's mental account of a text.

Worked example

Use this passage: “Nadia inspected a bicycle. The saddle was torn. The bicycle needed repairs.” Assume the intended reading is that the saddle belongs to the inspected bicycle.
Identify four mentions in order: M1 is Nadia, M2 is a bicycle, M3 is the saddle, and M4 is the bicycle. Create three entities: a person N, a bicycle B, and a saddle S. Link M1 to N, M2 and M4 to B, and M3 to S.
The identity clusters are therefore {M1}, {M2,M4}, and {M3}. Add a bridging relationship part-of(S,B). M3 is connected to M2 through that relationship, while M4 is an ordinary identity-coreference continuation of M2.
Now compare two faulty representations. A system that merges M3 with M2 produces only two entities and makes the saddle identical to the bicycle. A system that keeps three entities but omits the part-of relation preserves their distinctness while missing the contextual connection. These errors are different and should not be counted as the same kind of success or failure.
The passage explicitly attributes torn to S and needing repairs to B. It does not explicitly identify a broken wheel, a repair price, or an accident. A bridge between saddle and bicycle cannot justify adding those details.
Finally, insert “She also inspected another bicycle” before the saddle sentence. The candidate set changes: background knowledge now supports saddles for both bicycles. The earlier single-bicycle interpretation is no longer enough to determine the intended link without further context.

Illustrative case study

Fictional example: RepairBrief builds a summarizer for equipment inspection notes. Its initial entity table merges device names with later component mentions whenever they occur near each other. A report about a damaged drawer becomes a statement that an entire desk is damaged, losing the original level of detail.
The team separates identity links from component relationships and displays inferred relationships for review. It also creates examples with two nearby devices so reviewers can test antecedent selection. Its intended improvement is a faithful account of what each note says and what it merely suggests. No accuracy gain or customer outcome is claimed for this invented case.

Common misconceptions and mistakes

Bridging is not just matching repeated words. The expressions can be different because they denote different entities. Nor is every definite phrase a bridge: it may repeat an earlier referent, identify something situationally obvious, or depend on another source of familiarity.
Part - whole examples are convenient, but bridging is not limited to physical components. Annotation schemes differ in the relations and mention types they include. Another mistake is treating a plausible link as logically unavoidable: multiple candidates, unusual situations, or missing context may defeat the default interpretation.

Business use

Bridging analysis can help document review, support-note summarization, and information extraction retain relationships that ordinary identity clustering misses. Evaluate both recognition of bridging mentions and selection of their antecedents. Keep inferred links distinguishable from explicitly stated facts, especially when generated summaries will guide operational decisions. Include ambiguous passages and domain-specific relationships in evaluation data rather than relying entirely on examples.

Check your understanding

How many entities occur in the worked example's intended representation?

Answer: Three: Nadia, the bicycle, and its saddle.

Which mentions are coreferent?

Answer: M2 and M4.

What makes M3 a bridge?

Answer: The saddle is introduced through an inferred part-of relationship to the bicycle, rather than through identity with it.

Related terms

Learn next

Compare identity anaphora, bridging anaphora, and brand-new mentions in short passages. Then study annotation guidelines that state which relationships count as bridging, and examine how recognition errors differ from choosing the wrong antecedent after a bridge has been recognized.

Sources

Jurafsky and Martin, Speech and Language Processing, chapter 24, information status and inferrables: https://web.stanford.edu/~jurafsky/slp3/24.pdf Hou, Markert, and Strube, Unrestricted Bridging Resolution, definition and task distinctions: https://aclanthology.org/J18-2002.pdf

Previous: Breusch - Pagan test Next: Brier score