Intentions and Information in Discourse

Nicholas Asher, Alex LascaridesACL 1994

TL;DR

This paper proposes a formal framework for understanding how our intentions and the information conveyed in a discourse interact during interpretation. It argues for separating reasoning about mental states from discourse structure, while still allowing them to influence each other, offering a more nuanced model of how we understand language.

What Problem Does This Paper Address?

When we communicate, we don't just process the literal meaning of words. We also infer the speaker's intentions – what they want us to believe or do – and how this intention shapes the structure of their message. Conversely, the way a discourse is structured can reveal information about the speaker's intentions. Existing theories of discourse interpretation, like Rhetorical Structure Theory (RST), often struggle to simultaneously represent and reason about both the "intentional import" (speaker's goals) and the "informational import" (content of the message). This paper aims to bridge this gap by developing a unified framework that can handle the complex interplay between these two aspects of discourse.

Methods

The authors propose extending a framework called SDRT (Segmented Discourse Representation Structures) and its associated logical engine, DICE.

  • Separation of Reasoning: A key methodological innovation is the argument for separating the logical rules and axioms used for reasoning about beliefs and desires (intentional structure) from those used for inferring rhetorical relations (discourse structure).
  • Formalizing Mental States: They introduce a modal, dynamic logic to represent propositional attitudes like beliefs (B), wants (W), and intentions (I).
  • Practical Syllogism: To model reasoning about intentions, they adapt the "Practical Syllogism," a classical form of reasoning about action, into a defeasible (non-monotonic) rule. This allows an interpreter to infer an agent's intentions based on their beliefs and desires.
  • Abduction for Intentions: They introduce an "Abductive Practical Syllogism" (APS1), which allows for explanatory reasoning (abduction) when inferring intentional structure, a capability not typically present in discourse relation inference.
  • Cooperation Axiom: A crucial axiom, "Cooperation," links intentional and discourse structure. It states that if an utterance is intended to support another, the discourse relation between them must be one that licenses the intended inference (e.g., Result or Evidence).
  • Belief Property: Discourse relations like "Result" and "Evidence" are shown to have a "Belief Property," meaning they license an inference from the speaker's belief in the first part of the discourse to a belief in the second.
  • New Rules for Discourse Attachment: They introduce specific rules like the "Result Rule" and "Evidence Rule" that use the inferred intentional support to determine the appropriate discourse relation and, importantly, can infer new domain information that makes the discourse coherent.

Datasets

Not discussed in the paper. The paper uses example discourses for illustration rather than formal datasets.

Key Insights and Ideas

  • Intentions are not just about "what to say," but "why to say it": The paper highlights that understanding why a speaker said something (their intention) is crucial for interpreting what they said and how it relates to other parts of the discourse.
  • Reasoning about intentions is different: The authors argue that the inferential mechanisms for beliefs and desires are distinct from those for discourse structure. Specifically, abduction is useful for intentions but not typically for discourse relations.
  • Discourse structure can reveal information: A surprising insight is that the structure of a discourse can be used to infer new information about the domain, even if that information wasn't explicitly stated. This happens when the structure helps satisfy the speaker's inferred intentions.
  • Coherence is driven by intentional goals: A discourse is coherent not just because its parts logically connect, but because the connections serve the speaker's communicative goals. If the discourse structure doesn't align with the inferred intentions, it can lead to incoherence.
  • The "Belief Property" is key: Discourse relations that have the "Belief Property" are essential for bridging the gap between what a speaker wants the listener to believe and what the listener actually comes to believe.

Key Results

The paper provides a formal framework that:

  • Successfully models the flow of inference between communicative intentions and domain information in discourse.
  • Demonstrates how to integrate reasoning about mental states with discourse structure analysis using SDRT and DICE.
  • Explains how different contexts can lead to different interpretations of the same discourse, including cases where information flows from discourse structure to intentions, and vice-versa.
  • Shows how the framework can handle cases where new domain information is inferred to make a discourse coherent and satisfy the speaker's intentions.

Limitations and Future Work

  • Complexity of Real-World Discourse: The paper acknowledges that the analysis is for "very simple discourses" and that real-world communication involves more complex interactions between attitudes.
  • Integration with Speech Acts: The authors suggest that future work should integrate their model with theories of speech acts and attitudes to better capture how intentions are communicated.
  • Revision of Intentions: While they touch upon how agents update their intentions, a full formalization of intention revision in discourse is left as a future direction.
  • Choice of Discourse Relations: The paper notes that their current framework doesn't fully address how to choose among multiple possible discourse relations when several are plausible, suggesting the need for strategies to handle such choices.
  • Abductive Reasoning for Discourse Structure: While abduction is used for intentional structure, its role in inferring discourse relations is not fully explored.

Summary updated

BibTeX
@inproceedings{asher1994intentions,
  author = {Nicholas Asher and Alex Lascarides},
  title = {{Intentions and Information in Discourse}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {1994},
}
Figure