Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Less is more: Probabilistic reduction is best explained by small-scale predictability measures

Cassandra L. Jacobs, Andrés Buxó-Lugo, Anna K. Taylor, Marie Leopold-Hooke

TL;DR

This study interrogates how much context is needed to relate language-model probabilities to speech production, arguing that small-context n-gram representations suffice for probabilistic reduction in phonetic duration. Through analyses of four spontaneous English corpora and dialect-focused datasets, the authors show that intonational-phrase–level probabilities from n-grams predict duration more reliably than large language models (LLMs), and that LLMs contribute minimal, sometimes confounded, explanatory power after accounting for prosodic covariates. The work demonstrates that short-range context captures the relevant probabilistic effects and that LLM-based approaches may overestimate context needs due to position-length confounds, suggesting a more incremental view of production with phrase-level retrieval. They also explore dialectal variation, showing consistent probabilistic reduction across Buckeye and CORAAL and highlighting the small practical effect sizes, with broader implications for cognitive plausibility and methodological practices in psycholinguistics.

Abstract

The primary research questions of this paper center on defining the amount of context that is necessary and/or appropriate when investigating the relationship between language model probabilities and cognitive phenomena. We investigate whether whole utterances are necessary to observe probabilistic reduction and demonstrate that n-gram representations suffice as cognitive units of planning.

Less is more: Probabilistic reduction is best explained by small-scale predictability measures

TL;DR

This study interrogates how much context is needed to relate language-model probabilities to speech production, arguing that small-context n-gram representations suffice for probabilistic reduction in phonetic duration. Through analyses of four spontaneous English corpora and dialect-focused datasets, the authors show that intonational-phrase–level probabilities from n-grams predict duration more reliably than large language models (LLMs), and that LLMs contribute minimal, sometimes confounded, explanatory power after accounting for prosodic covariates. The work demonstrates that short-range context captures the relevant probabilistic effects and that LLM-based approaches may overestimate context needs due to position-length confounds, suggesting a more incremental view of production with phrase-level retrieval. They also explore dialectal variation, showing consistent probabilistic reduction across Buckeye and CORAAL and highlighting the small practical effect sizes, with broader implications for cognitive plausibility and methodological practices in psycholinguistics.

Abstract

The primary research questions of this paper center on defining the amount of context that is necessary and/or appropriate when investigating the relationship between language model probabilities and cognitive phenomena. We investigate whether whole utterances are necessary to observe probabilistic reduction and demonstrate that n-gram representations suffice as cognitive units of planning.
Paper Structure (16 sections, 1 equation, 7 figures, 6 tables)

This paper contains 16 sections, 1 equation, 7 figures, 6 tables.

Table of Contents

  1. Introduction
  2. Prosodic Preliminaries
  3. Structural and lexical influences on duration
  4. Utterance length and position confounds on duration
  5. Probabilistic factors
  6. Covariate selection
  7. Cognitive plausibility of backward language modeling
  8. Establishing short-range probabilistic reduction in clark_relationship_2025
  9. N-gram versus language model predictability and probabilistic reduction in upadhye2025back
  10. Examining probabilistic reduction in smaller and diverse dialect corpora
  11. Data preprocessing
  12. Language model fine-tuning
  13. Reverse language model fine-tuning
  14. Effect size of probabilistic reduction
  15. Discussion
  16. ...and 1 more sections

Figures (7)

  • Figure 1: Sentence-final lengthening and anticipatory shortening effects (relative position) on duration across four spontaneous speech corpora. Only utterances comprised of ten words or fewer are visualized; all are analyzed.
  • Figure 2: Correlation between log language model probability and words' durations (log ms) for short- and long-range language model probabilities.
  • Figure 3: Relationship utterance length and relative position on language model probabilities and duration in Switchboard corpus.
  • Figure 4: Relationship between utterance length and relative position on trigram conditional probabilities in CORAAL and Buckeye.
  • Figure 5: Regression model coefficient estimates for probabilistic reduction by corpus, language model type, and direction.
  • ...and 2 more figures