A Linear-Time Model of Language Production: Some Psychological Implications

David D. McDonaldACL 1980 · DOI

Introduction

Traditional psycholinguistics has historically focused on identifying the fundamental units of language—such as phonemes, morphemes, and phrases—by analyzing speech errors and hesitations. However, as David D. McDonald argues, identifying these units is only the first step. To truly understand human communication, we must define the computational processes that manipulate these units in real-time.

This paper introduces MUMBLE, a computational model of language production designed to adhere to strict psycholinguistic constraints. By treating production as a linear-time process governed by an "indelible" series of decisions, MUMBLE provides a structural explanation for the specific patterns of errors and fluencies observed in human speech.

The MUMBLE Architecture

The MUMBLE system is built upon two foundational architectural hypotheses that distinguish it from traditional generative models:

  1. Linear-Time Execution: The time required to produce an utterance is directly proportional to the number of elements in the input message. In this framework, the system makes a realization decision for each element exactly once. There is no backtracking, no look-ahead beyond a limited buffer, and no revision of previously committed segments.
  2. The Syntactic Tree as a Program: Rather than treating a phrase structure tree as a static data structure, MUMBLE views the surface-level syntactic tree as a "program of scheduled events." This tree simultaneously manages three roles: it controls the sequence of actions, specifies grammatical constraints, and maintains the linguistic context for subsequent decisions.

The Principle of Indelibility

A central concept in MUMBLE is "Indelibility." Decisions in the model are effectively written in "indelible ink." Once a syntactic construction is instantiated within the tree, the relative positions of its constituents cannot be altered. This forces the system to operate with "blinders," making decisions based solely on locally available information. This constraint mimics the human "point of no return" in speech, where once a word is uttered, the speaker must work within the grammatical framework they have already established.

Modeling Human Speech Errors

MUMBLE’s primary strength lies in its ability to explain human speech errors as an inevitable consequence of its architecture, rather than through the addition of arbitrary rules.

Stranded Morphemes and Word Swaps

A classic phenomenon in psycholinguistics is the "combined-form error," where word roots are swapped but their grammatical suffixes remain in their intended positions. For example, a speaker intending to say "My variables are more local" might instead say "My locals are more variable."

In this error, the plural marker "-s" stays attached to the second noun slot, even though the root words have switched places. MUMBLE explains this by maintaining a strict separation of position and content:

  • The phrase structure tree defines the grammatical "slots" and their properties (such as plurality or tense).
  • The message elements (the concepts or roots) are mapped into those slots.

Because the tree is instantiated first as a program of scheduled events, the plural "tag" is already fixed in the second noun position. If a mapping error occurs and the wrong message element is assigned to that slot, it simply inherits the grammatical properties already designated for that position.

Theoretical Implications

By enforcing a linear, no-backtracking constraint, MUMBLE achieves computational parsimony. It demonstrates that language production can be efficient and robust without the heavy overhead of complex backtracking or global scanning.

Furthermore, the model accounts for "grammatical robustness." Humans rarely make purely ungrammatical errors (e.g., violating basic word order rules) compared to the frequency of lexical or planning errors. In MUMBLE, this is because the syntactic framework is the "program" itself; as long as the program runs, the output will follow the structural rules of the language, even if the content mapped into that structure is incorrect.

Limitations and Future Directions

As an early model of the production process, MUMBLE has defined boundaries:

  • Phonological Scope: The current model produces text rather than acoustic speech, meaning it does not yet account for errors involving syllable structure or intonation.
  • Self-Correction: While the model acknowledges that humans occasionally catch their own errors, the strict indelibility constraint makes modeling "repair" a complex challenge for future iterations.
  • Empirical Expansion: Future work aims to use MUMBLE to simulate specific reaction time and hesitation data, further grounding the computational model in experimental psychological findings.

Conclusion

MUMBLE represents a shift from describing what language is to how it is produced. By imposing human-like limitations—such as linearity and indelibility—onto a computational framework, the model provides a principled explanation for why our speech fails in the specific, patterned ways that it does.

Summary updated

BibTeX
@inproceedings{mcdonald1980linear,
  author = {David D. McDonald},
  title = {{A Linear-Time Model of Language Production: Some Psychological Implications}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {1980},
  doi = {10.3115/981436.981454},
}
Figure