Interactions between turn-taking gaps, disfluencies and social obligation

Rebecca Lunsford, Peter A. Heeman, Jan P. H. van SantenInterspeech 2012

TL;DR

This paper investigates how the length of pauses between speakers (turn-taking gaps) and the presence of disfluencies (like "um" or "uh") interact, especially when a speaker is socially obligated to respond, such as after a question. They found that children, particularly those with typical development, are more likely to be disfluent after a question and when pauses are longer, suggesting a trade-off between timing and fluency driven by social pressure.

What Problem Does This Paper Address?

In human conversation, speakers generally try to minimize the silence between turns. However, the length of these pauses can be influenced by various factors, including the need to clarify, the length of the upcoming utterance, or even the difficulty of the question asked. This paper focuses on a specific aspect: how these pauses, or "turn-taking gaps," might influence the fluency of the speaker's next utterance, particularly the disfluencies that appear at the beginning of their speech. This is crucial for understanding natural human interaction and for building more robust spoken dialogue systems (SDS) that can better handle user disfluencies. The authors are particularly interested in whether the social pressure to respond quickly after a question leads to more disfluencies, and how this might differ in children with typical development (TD) versus those with language impairments like Autism Spectrum Disorder (ASD) or Developmental Language Disorder (DLD).

Methods

The researchers analyzed transcribed speech data from children interacting with a clinician. They used a speech recognizer for forced alignment to accurately measure the duration of silence between turns (gaps). Disfluencies, defined as "mazes" (false starts, repetitions, revisions, fillers, stutters), were identified and coded. They then employed mixed-effects logistic regression models to predict the likelihood of a child's utterance starting with a maze. Key factors included:

  • Gap Length: The duration of silence before the child's turn.
  • Question Status: Whether the clinician's preceding turn was a non-question, a question, or a series of consecutive questions.
  • Utterance Length: The number of words in the child's utterance.
  • Activity Type: The context of the interaction (Conversation, Description, Play).

The models were built to account for individual subject variability and were analyzed both by group (TD, DLD, ASD) and in a combined model to compare the groups.

Datasets

The study utilized data collected during the administration of the Autism Diagnostic Observation Schedule (ADOS) [16]. This dataset included speech from 91 children: 27 with Typical Development (TD), 21 with Developmental Language Disorder (DLD), and 43 with Autism Spectrum Disorder (ASD). The children ranged in age from 3.9 to 8.9 years. The data consisted of interactions between the child and a clinician across three activities: Conversation, Description, and Play.

Key Insights and Ideas

  • The Fluency-Timing Trade-off: Children are not just trying to fill silence; they seem to actively manage the trade-off between speaking quickly and speaking fluently. When a social obligation to respond is high (after a question), they are more likely to risk disfluencies to maintain a short pause.
  • Social Obligation Drives Disfluency: The pressure to respond after a question significantly increases the likelihood of turn-initial disfluencies, especially when combined with longer preceding gaps. This suggests that social cues play a critical role in speech production.
  • ASD and Social Pressure: Children with ASD appear to handle the social pressure of responding to questions differently. They are less likely to be disfluent after a question compared to TD children. This suggests they might be waiting until they have fully planned their response rather than rushing and becoming disfluent.
  • Activity Matters: The type of interaction influences disfluency rates. The "Description" activity, for instance, led to fewer disfluencies, possibly because it reduces the pressure to respond quickly or is less demanding socially.

Key Results

  • For children with TD and DLD, the likelihood of a turn-initial maze increased with the length of the preceding gap.
  • Being asked a question (or a series of questions) significantly increased the likelihood of a child producing a maze.
  • An interaction between gap length and utterance length was observed: a short gap followed by a long utterance was more likely to result in a maze than a long gap followed by a short utterance.
  • The interaction between questions and gap length showed that gap length had a greater effect on maze likelihood following a question.
  • Group Differences:
    • Children with DLD did not significantly differ from TD children in their maze production patterns, suggesting their language processing might not be as heavily impacted by social pressure in this context.
  • Children with ASD were less likely to produce a maze after a question compared to TD children. They were also more likely to produce mazes as utterance length increased.
  • The models showed good fit for TD and ASD groups (r > 0.90) but a less well-fit model for the DLD group.

Limitations and Future Work

  • DLD Model Fit: The authors note that the model for the DLD group had a less robust fit, particularly due to sparse data in certain bins. This suggests further investigation with larger DLD datasets might be beneficial.
  • Future Work for HCI: The findings offer practical implications for designing better spoken dialogue systems. SDS could be designed to be less question-response centric to reduce social pressure, or to anticipate increased disfluencies following longer gaps.
  • Differential Diagnosis: The observed differences in gap lengths and disfluency patterns between ASD and DLD groups could potentially serve as additional tools for differential diagnosis, given that social impairment is a key differentiator.
  • Nature of Interaction: The study suggests that the nature of the interaction itself (e.g., Description vs. Conversation) influences maze production, hinting at future research exploring how different conversational dynamics impact speech.

Summary updated

BibTeX
@inproceedings{lunsford2012interactions,
  author = {Rebecca Lunsford and Peter A. Heeman and Jan P. H. van Santen},
  title = {{Interactions between turn-taking gaps, disfluencies and social obligation}},
  booktitle = {Interspeech},
  year = {2012},
}
Figure