A Linear-time Model of Language Production: some psychological implications (extended abstract)

David D. McDonaldACL 1980

TL;DR

This paper proposes a computational model of language production, MUMBLE, that operates in linear time and produces utterances incrementally. It argues that computationally motivated constraints, rather than just observed behavior, are crucial for building psychologically plausible models of language generation.

What Problem Does This Paper Address?

Traditional psycholinguistic models of language production often focus on identifying the different levels of linguistic representation and the constraints on them, primarily based on analyzing speech errors and reaction times. However, these approaches don't fully explain how the process of generating language actually works. The paper highlights that many different computational mechanisms could produce the same observed behavior, making it difficult to determine the true underlying process. The core problem is to develop a computational model that is not only consistent with empirical data but also informed by principled, computationally motivated constraints, leading to a more robust and psychologically realistic theory of language production.

Methods

The paper introduces MUMBLE, a computational model of language production. Its core design is based on two key, intertwined stipulations:

  1. Linear-time execution: The processing time is directly proportional to the number of elements in the input message. Each decision about realization is made only once and cannot be revised.
  2. Surface-level syntactic phrase structure: A working structure, representing pending realization decisions and planned linguistic actions, is a surface-level syntactic phrase structure. This structure is augmented with explicit labels for its constituent positions. This "tree" is used for control (determining the next action), specifying constraints, and representing linguistic context.

These constraints lead to "indelibility" (decisions are final) and a limited-size buffer for linguistic processing, preventing "talking oneself into a corner." The model also adopts a "lexicalist" position on transformations, meaning they must apply at the moment a construction is instantiated and based on available information, not later.

Datasets

Not discussed in the paper.

Key Insights and Ideas

  • Computational constraints are key: The paper strongly advocates that computationally motivated constraints are essential for building psychologically plausible models of language production, going beyond simply matching empirical observations.
  • Linear-time processing is a powerful constraint: Requiring the production process to operate in linear time forces specific decision-making strategies and limits the possibilities for monitoring and self-correction, which aligns with observed human limitations.
  • "Indelibility" as a psychological principle: The concept of decisions being final ("written in indelible ink") is proposed as a potentially key characteristic of psychologically plausible performance theories of natural language.
  • Explaining speech errors through model design: The paper suggests that by incorporating features motivated by speech error analysis (like the distinction between functional and content morphemes) into the model's design, it can provide explanations for these phenomena. This creates a "two-way street" between computational modeling and psycholinguistics.

Key Results

The paper outlines the design principles of MUMBLE and discusses how they lead to properties like incremental utterance production, a well-defined point-of-no-return, limited buffer size, and grammatical robustness. It posits that these properties, derived from computational constraints, can explain certain classes of speech errors (e.g., word-exchange errors where functional morphemes are "stranded") and hesitation/self-correction data.

Limitations and Future Work

  • Focus on text, not speech: MUMBLE currently produces texts, not speech. Therefore, it has no knowledge of syllable structure or intonation and cannot contribute to explaining errors at that level.
  • Further exploration needed: The full paper is intended to further develop the notion of a linear-time production process, explore its implications for explaining speech errors, hesitation/self-correction data, and linguistic constraints.
  • Synthetic experiments: The authors suggest performing "synthetic experiments" by incorporating specific evidence from speech errors into MUMBLE's framework to explore the impact of hypotheses on the model's design.

Summary updated

BibTeX
@inproceedings{mcdonald1980linear,
  author = {David D. McDonald},
  title = {{A Linear-time Model of Language Production: some psychological implications (extended abstract)}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {1980},
}

Loading related content...

Figure