A Cooperative Yes-No Query System Featuring Discourse Particles

Kjell Johan SaebøCOLING 1988

TL;DR

This paper introduces the PASSAT system, a cooperative yes-no query system that leverages discourse particles to generate more informative and coherent answers. It proposes a formal model for discourse particles that links their usage to the underlying discourse relations and the context of the conversation, aiming to capture the nuanced way humans provide extended answers.

What Problem Does This Paper Address?

Traditional question-answering systems often provide direct, minimal answers. However, human conversation is far richer; when asked a yes-no question, we often volunteer additional, relevant information beyond a simple "yes" or "no." This phenomenon, known as "cooperative response," is crucial for effective communication. The challenge lies in computationally modeling this behavior, particularly how to determine what extra information to provide and how to signal its relationship to the original question. Discourse particles (like "but," "also," "so") are key linguistic tools that humans use to signal these relationships and ensure coherence in extended answers. This paper addresses the gap in understanding and modeling the role of these particles in generating cooperative, complex responses.

Methods

The paper proposes a formal semantic framework for discourse particles and integrates it into a system called PASSAT. The core idea is that discourse particles are "meta-expressions" that signal discourse relations between sentences, rather than referring to the world directly. These relations are not about truth conditions but about coherence.

The system models discourse relations using a triple: a set of discourse relations (DR), an assignment mapping sentence pairs to these relations (f), and an interpretation mapping discourse particles to subsets of DR (b). A discourse particle is considered coherent if the discourse relation of the sentence pair is not in the denotation of any other particle. This implies that particles are necessary to signal specific discourse relations.

PASSAT itself is a PROLOG II system that uses a register of rules to regulate choices of alternatives and discourse particles. It operates on a sortal hierarchy of natural kinds and scalar structures to identify relevant alternatives. When a direct answer is given, PASSAT seeks to provide an "exhaustive response" by identifying relevant alternatives and selecting the appropriate discourse particle to connect them to the original query. The system uses meaning postulates and lexical relations to determine these alternatives and their relationships.

Datasets

Not discussed in the paper. The paper focuses on the theoretical framework and the PASSAT system's internal logic rather than specific training or evaluation datasets.

Key Insights and Ideas

  • Discourse Particles as Coherence Signals: The paper strongly emphasizes that discourse particles are not about truth but about ensuring coherence between sentences, especially in complex answers. They signal the relationship between pieces of information.
  • Cooperative Response as the Norm: The system is built on the premise that humans naturally provide more information than strictly asked for, and this "cooperative" behavior is the default, not an exception.
  • Context is King for Particles: The meaning and necessity of a discourse particle are determined by the discourse relation it signals and the conversational context, not by the absolute truth of the statements.
  • Formalizing Discourse Relations: The paper offers a formal model (DR, f, b) that captures the denotation of discourse particles as sets of discourse relations, providing a computational basis for their use.
  • Hierarchies and Scales Drive Alternatives: PASSAT uses sortal hierarchies (e.g., types of ships) and scalar structures (e.g., degrees of certainty) to generate relevant alternative information for extended answers.

Key Results

The paper describes the PASSAT system's capabilities rather than presenting quantitative results from empirical evaluations. The key "results" are the system's ability to:

  • Generate extended answers to yes-no queries.
  • Select appropriate discourse particles to signal the relationship between the original query and the additional information provided.
  • Utilize semantic hierarchies and scales to identify relevant alternatives.
  • Mimic a natural, cooperative conversational style by offering more than a minimal answer.

The paper demonstrates this through several example dialogues, showing how PASSAT selects particles like "und zwar" (and specifically), "sondern" (but rather), and "aber" (but) based on the semantic relationships and polarities involved.

Limitations and Future Work

The authors acknowledge several limitations:

  • No Morphology: The system deliberately omits morphological processing (articles, gender variants) to focus on core principles.
  • Minimal Syntax: Only a basic syntax is implemented, primarily relying on total ellipsis. More sophisticated ellipsis handling is a potential area for expansion.
  • Context Sensitivity of Lexical Relations: The system's assessment of lexical relations in meaning postulates and alternative relations is limited by the fact that predicates always appear in predicative position. Introducing contexts where nouns restrict quantification or act as arguments in binary relations would require a more sophisticated semantic representation.
  • Focus Identification: A significant pragmatic challenge is identifying the "focus" of a question, which determines which alternative information is relevant. The current system assumes the focus is always on the one-place predicate. Future work should explore how to handle the ambiguity of yes-no questions and their topic-focus structure to better determine relevant substitutes.

Future work could involve developing methods to determine the extent and direction of relevance in hierarchical structures and integrating more sophisticated syntactic and pragmatic modules.

Summary updated

BibTeX
@inproceedings{saeb1988cooperative,
  author = {Kjell Johan Saebø},
  title = {{A Cooperative Yes-No Query System Featuring Discourse Particles}},
  booktitle = {International Conference on Computational Linguistics},
  year = {1988},
}

Loading related content...

Figure