Some Pragmatic Issues in the Planning of Definite and Indefinite Noun Phrases

Douglas E. AppeltACL 1985

TL;DR

This paper proposes a framework for understanding how language generation systems can pragmatically choose between definite and indefinite noun phrases. By categorizing noun phrase actions into four types based on shared knowledge and identification intentions, the authors provide a more nuanced account that is crucial for building more sophisticated AI systems capable of natural language interaction.

What Problem Does This Paper Address?

Generating natural-sounding and effective language requires more than just knowing grammar and vocabulary. It involves understanding the speaker's intentions and the context of the conversation. Specifically, when a speaker uses a noun phrase (like "the runner" or "a doctor"), they have a specific communicative goal. This paper tackles the challenge of how an AI system, acting as a speaker, can decide which type of noun phrase to use to best achieve that goal. The authors focus on noun phrases that presuppose the existence of an individual described by a given description, excluding those bound by quantifiers or within sentential negation. This focus is motivated by the need for AI systems to effectively communicate in task-oriented dialogues, where identifying and manipulating objects in the environment is often key.

Methods

The paper introduces a plan-based approach to utterance planning, drawing on the KAMP system. Individuals are represented by terms in an intensional logic, and a metalanguage describes the relationship between these terms and the individuals they denote. The core of the paper is a taxonomy of four "concept activation actions" that underlie the use of noun phrases. These actions are distinguished by two key criteria:

  1. Identification Intention: Does the speaker intend for the hearer to identify the referent of the description?
  2. Shared Knowledge: How much mutual knowledge do the speaker and hearer share about the description's denotation?

These criteria lead to four action types:

  • SI (Shared Concept Activation with Identification Intention): Speaker and hearer mutually believe the description applies to a referent, and the speaker wants the hearer to identify it (e.g., "Use the same wrench you used to unfasten the pump").
  • NSI (Nonshared Concept Activation with Identification Intention): Speaker and hearer don't mutually know the referent, but they believe there's a plan the hearer can execute to identify it (e.g., "Get me the largest tomato from the garden").
  • SNI (Shared Concept Activation with No Identification Intention): Speaker and hearer mutually believe the description applies to a referent, but the speaker doesn't want the hearer to identify it (e.g., "The man who murdered Smith is insane").
  • NSNI (Nonshared Concept Activation with No Identification Intention): No mutual belief preconditions; speaker introduces a new individual without expecting the hearer to identify it (e.g., "I met an old friend from high school yesterday").

Datasets

Not discussed in the paper.

Key Insights and Ideas

  • Pragmatics is Key to Generation: The paper strongly emphasizes that language generation is an action, and therefore, pragmatics (the study of how context affects meaning) plays a critical role.
  • Beyond Referential Use: The authors move beyond traditional notions of referring expressions to encompass a broader range of noun phrase uses, arguing that many non-referring phrases still involve pragmatic considerations relevant to planning.
  • A Taxonomy for Planning: The four-way classification of concept activation actions provides a useful level of abstraction for utterance planning systems, allowing for more specific planning strategies based on the communicative intent.
  • Identification is Context-Dependent: What constitutes an "identifying" description depends heavily on the purpose and the hearer's knowledge, not just on the description itself.
  • Mutual Knowledge is Crucial: The degree of shared knowledge between speaker and hearer significantly impacts the preconditions and strategies for successful noun phrase use.
  • Referential vs. Attributive is Orthogonal: The distinction between referential and attributive noun phrases is shown to be independent of the pragmatic distinctions the paper proposes, suggesting a more unified approach to planning.

Key Results

The paper doesn't present quantitative results from experiments. Instead, its key contribution is the theoretical framework and the proposed taxonomy of concept activation actions. The "results" are the definitions and characterizations of these four actions, outlining their distinct preconditions, effects, and associated planning strategies. This framework aims to provide a more comprehensive understanding of noun phrase generation for AI systems.

Limitations and Future Work

The authors acknowledge that the proposed taxonomy accounts for an "important class of noun phrases" but does not "exhaust all possibilities." They identify several areas for future research:

  • Quantificational Noun Phrases: Such as those involving universal quantifiers.
  • Predicate Nominals: Like "John wants to be a doctor."
  • Generics and Bare Plurals: These require further analysis.
  • Coreferring Actions: The activation of terms already introduced into the discourse is a significant class not discussed in detail.

Developing a complete plan-based analysis that subsumes both referring and non-referring noun phrases remains an ongoing goal. The paper also notes that developing adequate criteria for "prima facie identifiable terms" is an important research problem.

Summary updated

BibTeX
@inproceedings{appelt1985some,
  author = {Douglas E. Appelt},
  title = {{Some Pragmatic Issues in the Planning of Definite and Indefinite Noun Phrases}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {1985},
}
Figure