ICML · LLMs and Cognition Workshop

2024

Bilingual disfluencies

Modeling Bilingual Disfluencies with Large Language Models

Negin Raoof†, Yating Wu†, Carlos Bonilla†, Junyi Jessy Li, Stephanie M Grasso, Alexandros G. Dimakis, Zoi Gkalitsiou

ICML Workshop on LLMs and Cognition · 2024

† Equal contribution.

paperBibTeX

Overview

We study which properties of words predict disfluencies in monolingual and bilingual speech. Separate models combine linguistic features with GPT-2 surprisal, a measure of how unexpected a word is in its preceding context.

Higher surprisal is associated with disfluencies in both groups. Other predictors differ, including word frequency and phonological neighborhood density. We examine words on both sides of each disfluency.

Paper

Which word comes before the pause?

These examples from the appendix show why the words before and after a pause are treated separately.

Pause after the noun

give me the stethoscope uhm to hear your heart

Preceding word: stethoscope. Following word: to.

Pause before the noun

give me the uhm stethoscope to hear your heart

Preceding word: the. Following word: stethoscope.

In the bilingual model, lower frequency of the preceding word is associated with between-word disfluencies. In the monolingual model, the corresponding association involves fewer phonological neighbors of the following word.

Source

Appendix A.0.1, page 7. Illustrative sentences from Appendix A.0.1, not participant transcripts. Feature interpretations paraphrased.

What the models predict

We fit one logistic regression classifier per disfluency type for each speaker group. GPT-2 provides word surprisal; the classifiers also use lexical, syntactic, and phonetic features.

Between words
Pauses and fillers between words.
Within a word
Broken words and repetitions of sounds or syllables.
Repetitions
Repeated words or short phrases.

Training and test sets are separated by participant. Features of the preceding and following words are analyzed separately.

Source

Tasks, features, and evaluation: Sections 2.3–2.5

PDF
Features, models, and results.

Citation

Download BibTeX
View BibTeX
@inproceedings{bilingualdisfluencies2024,
  title={{Modeling Bilingual Disfluencies with Large Language Models}},
  author={Raoof, Negin and Wu, Yating and Bonilla, Carlos and Li, Junyi Jessy and Grasso, Stephanie M and Dimakis, Alexandros G. and Gkalitsiou, Zoi},
  year={2024},
  booktitle={ICML 2024 Workshop on LLMs and Cognition},
  url={https://openreview.net/forum?id=rrNAqNYRLA}
}