An ascription-based approach to speech acts

Mark Lee, Yorick Wilks

TL;DR

This paper proposes a new approach to understanding speech acts by focusing on "belief ascription" rather than pre-computing complex nested beliefs. This allows for a more flexible and computationally efficient way to model how agents understand each other's intentions and beliefs in dialogue, particularly in planning and plan recognition tasks.

What Problem Does This Paper Address?

Natural Language Processing (NLP) research often grapples with understanding the nuances of human conversation, particularly "speech acts" – the actions performed by utterances (e.g., asking a question, making a request, informing). A key challenge is modeling the mental states of dialogue participants, specifically their beliefs, goals, and intentions. Traditional approaches often rely on "belief modeling," trying to represent these mental attitudes in complex, nested structures. However, these methods have a significant drawback: they assume agents can compute potentially infinite levels of nested beliefs, which is computationally expensive and arguably unrealistic for real-time dialogue. This paper addresses the limitations of these simplistic and computationally demanding belief modeling techniques in speech act processing.

Methods

The core of the paper's contribution lies in its "ascription-based approach" to speech acts, implemented in a system called ViewGen. Instead of pre-computing and storing extensive nested belief structures, ViewGen uses "belief ascription" – inferring or attributing beliefs, goals, and intentions to other agents as needed.

Two key ascription methods are introduced:

  • Default Ascription: This method assumes that rational agents share common beliefs about the world unless there's evidence to the contrary. For example, if the system believes the world is round, it can default-attribute this belief to another agent.
  • Stereotypical Ascription: This method leverages stereotypes associated with certain classes of agents. For instance, if an agent is identified as a "doctor," ViewGen can ascribe typical medical knowledge and goals to them.

These ascription mechanisms allow ViewGen to dynamically build nested belief structures only when required for understanding or planning, rather than maintaining them constantly. The system also uses a non-linear POCL planner for planning and plan simulation, which can operate within these dynamically constructed belief environments.

Datasets

Not discussed in the paper.

Key Insights and Ideas

  • "On-demand" belief computation is more realistic: The paper argues against the necessity of pre-computing deeply nested belief structures for every dialogue interaction. Instead, beliefs are "ascribed" or inferred when relevant, making the process more efficient and plausible.
  • Ascription as a bridge: Belief ascription acts as a crucial mechanism for bridging the gap between an agent's own mental states and their understanding of another agent's mental states, especially in the context of speech acts.
  • Separating minimal conditions from effects: The authors emphasize that speech act theories should focus on the minimal conditions required for an act to be successful, rather than trying to predict all possible effects. This leads to a more parsimonious and extendable theory.
  • Solipsism in speech act theory: The paper advocates for a "solipsist" approach, meaning the theory should refer solely to finite belief representations of the speaker or hearer, avoiding the problematic notion of mutual belief.

Key Results

The paper doesn't present quantitative results in terms of accuracy metrics on specific benchmarks. Instead, it focuses on the conceptual and architectural contributions of its ascription-based approach. The key "results" are the demonstration of how ViewGen can:

  • Represent mental attitudes (beliefs, goals, intentions) using nested environments.
  • Dynamically construct nested belief structures through default and stereotypical ascription.
  • Use these ascription techniques to model speech acts, enabling planning and plan recognition within nested belief environments.
  • Simulate other agents' planning and recognize their plans by reasoning about beliefs at different nesting levels.

Limitations and Future Work

  • Limitations: The authors acknowledge that their set of 20 speech act types is not exhaustive and is based on a partial taxonomy. They also note that while their approach allows for on-demand computation of highly nested beliefs, it doesn't explicitly detail how to handle non-conventional language use that might require such deep nesting.
  • Future Work: The authors plan to:
  • Develop a more robust dialogue parser.
  • Link ViewGen to the LaSie information extraction platform to empirically derive belief sets from a medical domain corpus.
  • Further explore the implications of their approach for understanding non-conventional language use.

Summary updated

Abstract

The two principal areas of natural language processing research in pragmantics are belief modelling and speech act processing. Belief modelling is the development of techniques to represent the mental attitudes of a dialogue participant. The latter approach, speech act processing, based on speech act theory, involves viewing dialogue in planning terms. Utterances in a dialogue are modelled as steps in a plan where understanding an utterance involves deriving the complete plan a speaker is attempting to achieve. However, previous speech act based approaches have been limited by a reliance upon relatively simplistic belief modelling techniques and their relationship to planning and plan recognition. In particular, such techniques assume precomputed nested belief structures. In this paper, we will present an approach to speech act processing based on novel belief modelling techniques where nested beliefs are propagated on demand.
BibTeX
@misc{lee1999ascription,
  author = {Mark Lee and Yorick Wilks},
  title = {{An ascription-based approach to speech acts}},
  year = {1999},
  eprint = {cs/9904009},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/cs/9904009},
}

Loading related content...

Figure