Crossing Coreference in Discourse Representation Theory

Michael HessCOLING 1988

TL;DR

This paper tackles the notoriously tricky problem of "crossing coreference" in natural language sentences, where pronouns in one part of a sentence refer to noun phrases in another, and vice-versa. By adapting Discourse Representation Theory (DRT) with insights from Prolog's second-order logic constructs, the authors propose a novel framework that elegantly explains these complex sentences without ambiguity.

What Problem Does This Paper Address?

Natural language is rife with pronouns that refer back to previously mentioned entities (anaphora) or, less commonly, forward to entities yet to be mentioned (cataphora). Sentences exhibiting "crossing coreference," often called Bach-Peters sentences, present a significant challenge to linguistic theories. These sentences feature a reciprocal dependency: a pronoun in one noun phrase refers to the other noun phrase, and vice-versa. For example, "The hunter who shot at it hit the lion that chased him." Traditional linguistic theories struggle to provide a coherent explanation for these sentences, often leading to analyses that suggest ambiguity where none intuitively exists or infinite embeddings of grammatical structures. The core problem is how to represent and resolve these intertwined references in a way that accurately reflects the sentence's meaning and truth conditions.

Methods

The paper proposes a solution by integrating concepts from Discourse Representation Theory (DRT) with the expressive power of second-order Prolog constructs. The authors leverage Prolog's setof operator, which handles uniqueness requirements, and its relational logic for variable binding. This approach allows for a more direct representation of the dependencies in Bach-Peters sentences.

Key methodological steps include:

  • Adapting DRT: The authors suggest modifying DRT to explicitly represent the embedding of noun phrases, moving beyond traditional DRT's focus on flat structures or embedding entire DRSs.
  • Prolog's setof: This operator is used to represent definite noun phrases, ensuring that each refers to a unique entity. The uniqueness requirement is crucial for resolving the specific dependencies in Bach-Peters sentences.
  • Relational Logic and Variable Binding: Prolog's mechanism for variable sharing across clauses is adapted to handle the "spreading" of variable bindings, mirroring how pronouns can refer across different parts of a sentence. This contrasts with first-order logic's limitations where variables within quantifiers are typically inaccessible from the outside.
  • Mapping Variables to Pronouns: A distinction is made between variables used as computational artifacts (to flatten embeddings) and those representing true anaphoric relations. The latter are mapped to pronouns in the surface sentence, with specific rules for handling level-crossing co-occurrences to ensure correct syntactic generation.

Datasets

Not discussed in the paper. The paper focuses on theoretical linguistic examples and the logical representation of sentences rather than empirical data from large corpora.

Key Insights and Ideas

  • Prolog as a Model for Coreference: The paper highlights how Prolog's relational logic and variable binding mechanisms provide a surprisingly effective model for capturing complex coreferential dependencies, particularly those found in Bach-Peters sentences.
  • DRT Enhancement: The authors demonstrate that DRT can be enhanced by allowing explicit embedding of conditions within DRSs and by leveraging Prolog's meta-call facility to handle the dual role of expressions as data and executable procedures.
  • Irreducibility of Cataphora: The analysis reinforces the idea that cataphora in Bach-Peters sentences cannot be reduced to simple anaphora, as traditional theories often attempt. The proposed framework naturally accounts for this distinction.
  • Unified Explanation: The approach offers a single, coherent framework that explains Bach-Peters sentences, simple cataphora, and anaphora without resorting to ambiguity or infinite embeddings.

Key Results

The paper presents a formalization that:

  • Correctly derives the truth conditions for Bach-Peters sentences and related examples (like sentences 3 and 4 in the text) without ambiguity.
  • Avoids the problematic repetition of logical expressions that arises in first-order logic translations.
  • Provides a mechanism for mapping complex semantic representations back to surface-level sentences, handling the generation of pronouns and full noun phrases appropriately.

Limitations and Future Work

  • Syntactic Generation Rules: While the paper proposes rules for mapping semantic representations to surface sentences, it acknowledges that these rules might be "ad-hoc" and suggests the need for further refinement.
  • DRT Modifications: The proposed modifications to DRT, particularly regarding the handling of definite noun phrases and the suspension of evaluation for certain expressions, require further investigation and implementation.
  • Generalization: The paper primarily focuses on definite noun phrases. Extending the framework to handle indefinite noun phrases and other quantifiers more comprehensively is a natural next step.
  • Empirical Validation: While the theoretical framework is compelling, empirical validation with real-world linguistic data and computational implementation would be valuable.

Summary updated

BibTeX
@inproceedings{hess1988crossing,
  author = {Michael Hess},
  title = {{Crossing Coreference in Discourse Representation Theory}},
  booktitle = {International Conference on Computational Linguistics},
  year = {1988},
}
Figure