Anaphora Resolution in Multi-Person Dialogues

Prateek Jain, Manav Ratan Mital, Sumit Kumar, Amitabha Mukerjee, Achla M. RainaSIGDIAL Workshop 2004 · OpenReview

TL;DR

This paper introduces "Jepthah," a novel system for resolving anaphoric references in multi-person dialogues. It tackles the complexity of dialogue anaphora by integrating syntactic, semantic, and world knowledge through a graph-based approach, enhanced by genetic algorithms for heuristic weight tuning and a new "anaphora chain" concept for broader anaphor coverage.

What Problem Does This Paper Address?

Resolving anaphoric references (like pronouns, e.g., "he," "it," "they") is a fundamental challenge in natural language processing, crucial for understanding discourse. While significant work exists for single-speaker text, multi-person dialogues present unique complexities. These include a higher density of anaphoric references, the need to track multiple speakers and their conversational turns, and the presence of various anaphor types (pronouns, reflexives, deictic anaphors, cataphora) that can be ambiguous or context-dependent. Existing heuristic-based approaches often require laborious manual tuning of weights, and statistical methods can struggle with limited dialogue corpora. This paper aims to build a more robust and adaptable system for dialogue anaphora resolution.

Methods

The core of Jepthah is a graph-theoretic approach to model the dialogue. All nouns and pronouns are represented as nodes in a graph. Directed edges connect potential anaphor-antecedent pairs. The system then defines "anaphora chains" as paths within this graph, representing sequences of co-referential elements. This chaining concept is key to resolving a wider range of anaphors, including plural anaphora and cataphora.

The resolution process involves two phases:

  1. Constraint Application: Initial pruning of infeasible co-references using a set of constraints (e.g., gender, number, person agreement; semantic consistency; reflexive pronoun rules).
  2. Heuristic Application: Assigning weights to the remaining edges using a combination of standard heuristics (e.g., definiteness, recency, syntactic/semantic parallelism) and novel pragmatic heuristics specifically designed for dialogue settings (e.g., speaker-questioner relationships).

A significant innovation is the use of genetic algorithms to automatically tune the weights of these heuristics. This relieves the programmer from manual, ad-hoc weight assignment and allows the system to adapt to specific dialogue corpora. The system also incorporates a naive character building component, using a limited knowledge base of noun ontologies and functionalities to inform semantic and world-knowledge-based heuristics.

Datasets

The paper evaluates Jepthah on two types of corpora:

  • Manually written dialogues: Developed by the authors in a student environment.
  • Fragments from plays by G. B. Shaw: Specifically, "Shaw's play - Pygmalion" and "Shaw's play - Man and Superman."

The authors also refer to "hand-crafted dialogues" which are sample dialogues they wrote themselves for testing. No new benchmark datasets are introduced.

Key Insights and Ideas

  • Anaphora Chains as a Unifying Concept: The idea of anaphora chains provides a powerful framework to represent and resolve complex referential relationships, extending beyond simple pronoun-noun links to encompass cataphora and plural anaphora.
  • Genetic Algorithms for Adaptive Heuristic Tuning: Automating the weight tuning process with genetic algorithms is crucial for building systems that can adapt to the nuances of different dialogue corpora without extensive manual intervention.
  • Dialogue-Specific Pragmatic Heuristics: Identifying and incorporating heuristics that leverage the turn-taking and speaker dynamics inherent in dialogues is essential for improving resolution accuracy in this domain.
  • Graph-based Modeling for Complex Relations: Representing the dialogue as a graph allows for a structured and systematic way to explore and rank potential anaphor-antecedent relationships.

Key Results

The system achieved notable accuracy on the tested corpora:

  • Shaw's plays: Approximately 65% accuracy.
  • "Hand-crafted" dialogues: Approximately 80% accuracy.

The paper highlights that the genetic algorithm tuning led to an "appreciable increase in accuracy." The system is described as automatic and knowledge-based, capable of resolving a wide variety of anaphors, including cataphora and plural anaphora.

Limitations and Future Work

  • Lack of Standard Corpora: The authors acknowledge that the lack of standard corpora for anaphora resolution, especially for dialogues, is a major drawback for general system evaluation and comparison.
  • Naive Character Building: The "naive character building" is a simplified approach. More sophisticated semantic and world knowledge integration could further improve performance.
  • Extensive Evaluation: The authors state that an "extensive evaluation of the various resolution algorithms as well as the system as a whole remains to be done." This suggests that while promising results were shown, further validation and comparison with other state-of-the-art systems are needed.
  • Unambiguous Plural Pronouns: The system currently deals only with unambiguous cases of plural pronouns (e.g., "both of us," "two of you"), not more complex cases like "all of us" or "they." Expanding this capability would be a natural next step.

Summary updated

BibTeX
@inproceedings{jain2004anaphora,
  author = {Prateek Jain and Manav Ratan Mital and Sumit Kumar and Amitabha Mukerjee and Achla M. Raina},
  title = {{Anaphora Resolution in Multi-Person Dialogues}},
  booktitle = {SIGDIAL Workshop},
  year = {2004},
  url = {https://openreview.net/forum?id=nKwPyZMjfVn},
}
Figure