Bridging Anaphora Resolution

81 concepts62 papers1 proposal

When reading a text, humans naturally connect concepts that are related but not explicitly identical. Consider the sentences: "We found a house we liked. The roof was in good condition." As readers, we instantly understand that "the roof" refers to the roof of the house.

In computational linguistics, this phenomenon is known as bridging anaphora (or associative anaphora). Unlike identity coreference—where a pronoun or noun phrase refers to the exact same entity (e.g., mapping "the house" to "it")—bridging requires a system to link an anaphor to a non-identical but conceptually related anchor.

As Vincent Ng 0001 notes, while identity coreference resolution has matured significantly, resolving bridging references remains an open frontier. It requires moving past surface-level string matching toward deep semantic reasoning and world knowledge.


The Core Challenge: Moving Beyond Identity

Traditional natural language processing models often limit their scope to identity coreference. However, as David Klappholz et al. argued early on in The Use of Dynamically Extracted Context for Anaphoric Reference Resolution, this narrow view fails to account for complex references where an entity's presence is merely implied by the broader context.

To resolve a bridging description, a system must solve two distinct problems formulated by Renata Vieira et al. in Towards Resolution of Bridging Descriptions:

  1. Anchor Identification: Finding the correct preceding entity that triggers the reference.
  2. Link Classification: Determining the specific semantic relationship (e.g., part-whole, set-member, or associative) connecting them.
[We found a house we liked.] <-- (Anchor)
       |
       | (Relationship: Part-Whole)
       v
[The roof was in good condition.] <-- (Bridging Anaphor)

Because bridging anaphors frequently lack explicit determiners or direct lexical overlap, standard syntactic rules fall short. Naoya Inoue et al. demonstrated in Resolving Direct and Indirect Anaphora for Japanese Definite Noun Phrases that indirect anaphora resolution shows much slower performance gains with increased training data than direct coreference, highlighting its heavy reliance on extensive world knowledge.


Overcoming the Knowledge Bottleneck

How can a machine acquire the common-sense knowledge needed to connect "house" to "roof" or "space shuttle" to "astronauts"? Researchers have approached this knowledge bottleneck from several angles.

1. Hybrid Lexical and Salience Models

Early machine learning approaches sought to combine discourse structure with external knowledge bases. In Learning to Resolve Bridging References, Massimo Poesio et al. demonstrated that combining discourse salience (how prominent an entity is in the text) with lexical distance (measured via Google search counts or WordNet) could push resolution accuracy past 80% on balanced datasets.

2. Corpus-Driven Case Frames

Instead of relying on static, hand-coded taxonomies, later techniques looked to large-scale text corpora to discover how words naturally co-occur. Ryohei Sasano et al. in A Probabilistic Model for Associative Anaphora Resolution framed associative anaphora as a form of "zero anaphora" (omitted arguments). By automatically extracting nominal case frames from a massive 1.6-billion-sentence web corpus, their system learned the typical semantic arguments that nouns expect, significantly outperforming purely salience-based baselines.

3. Deterministic and Relational Embeddings

To handle unrestricted bridging—where anaphors may not fit neat part-whole hierarchies—Yufang Hou designed a deterministic algorithm in A Deterministic Algorithm for Bridging Anaphora Resolution. By creating specialized word embeddings that blend general lexical knowledge with syntactic relational patterns, the author showed that a targeted "average-and-compare" approach could rival complex machine learning architectures.


The Role of Discourse Structure and Cognitive Focus

Resolving an anaphor is not just a matter of scanning a dictionary; it is constrained by how the human mind tracks a conversation. Cognitive models, such as the one explored by Raymonde Guindon in Anaphora Resolution: Short-Term Memory and Focusing, emphasize that our brains maintain a dynamic "focus" that limits the search space for potential anchors.

When a text introduces an entity, it opens up a mental model of related concepts. If a system fails to track these implicit shifts in focus, its downstream reasoning suffers. For example, Shachar Mirkin et al. showed in Assessing the Role of Discourse References in Entailment Inference that discourse references like bridging are mandatory for nearly half of all textual entailment inferences. Without them, models cannot maintain the logical chains required to verify if one statement proves another.


Bridging in the Era of Transformers

With the advent of deep pre-trained language models, a natural question emerged: do massive transformer models inherently understand these implicit relationships without explicit training?

In Probing for Bridging Inference in Transformer Language Models, Onkar Pandit et al. probed the internal layers of state-of-the-art architectures. They discovered that while transformers capture a surprising amount of relational knowledge in their attention heads, they still struggle to consistently perform full bridging inference out of the box.

This limitation becomes highly visible in complex tasks like semantic parsing and question answering:

To push past these limits, modern architectures often integrate explicit entity tracking. For instance, Tomohide Shibata et al. utilized a dynamic "entity buffer" to jointly model coreference and predicate-argument structures, ensuring that even when an entity is omitted or indirectly referenced, its semantic footprint is preserved across the discourse.

Key Foundational Work

Anaphora for Limited Domain Systems (1981) — contributed 1 concept elements to this direction.
The Use of Dynamically Extracted Context for Anaphoric Reference Resolution (1977) — contributed 1 concept elements to this direction.
Search and Inference Strategies in Pronoun Resolution: an Experimental Study (1981) — contributed 1 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_bridging_anaphora_resolution_fb14bf1a,
  title = {bridging anaphora resolution},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/bridging-anaphora-resolution/rdr_fb14bf1a1e8e46ccb986778211259d79},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_bridging_anaphora_resolution_fb14bf1a_v1,
  title = {bridging anaphora resolution},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/bridging-anaphora-resolution/rdr_fb14bf1a1e8e46ccb986778211259d79/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake