Laughter Dynamics in Dyadic Conversations

Bogdan Ludusan, Petra WagnerInterspeech 2019

TL;DR

This paper investigates whether conversational partners synchronize their laughter, a key paralinguistic element, across different languages. They found evidence for synchrony and convergence in laughter dynamics, suggesting that this form of alignment extends beyond verbal communication and may be influenced by language or culture.

What Problem Does This Paper Address?

Human communication is a complex dance of alignment, where partners unconsciously mirror each other's behaviors, from posture to speech patterns. While much research has focused on verbal alignment, the role of paralinguistic cues, like laughter, in this dynamic has been less explored. This paper addresses the question of whether laughter, a fundamental aspect of social interaction, is also subject to conversational alignment. Specifically, it investigates three facets of alignment: synchrony (how similarly speakers use laughter over time), convergence (whether speakers' laughter patterns become more similar towards the end of a conversation), and agreement (whether speakers match the voicing of consecutive laughs). Understanding this could shed light on the subtle ways we connect and influence each other during conversations.

Methods

The researchers employed three novel measures to quantify laughter alignment:

  • Synchrony: This measure quantifies the dynamic likeness in laughter usage between two speakers throughout a conversation. It's calculated using the Spearman correlation coefficient between the number of laughs produced by each participant in ten equal time bins of the conversation. This approach is more robust to outliers than Pearson correlation.
  • Convergence: This metric assesses whether speakers' laughter patterns become more similar over the course of a conversation. It's operationalized by comparing the difference in laughter counts between partners in the first half of the conversation versus the second half. A decrease in this difference suggests convergence.
  • Agreement: This measure focuses on phonetic imitation, specifically looking at the similarity in the voicing of consecutive laughs produced by different speakers. A laugh is considered consecutive if it overlaps with or begins within one second of the end of the other speaker's laugh. Agreement is calculated as the proportion of consecutive laughs that share the same voicing (voiced or unvoiced), compared to a baseline of randomly sampled non-consecutive laughs.

Datasets

The study utilized the DUEL corpus, a multilingual corpus of spontaneous dyadic conversations. This corpus includes interactions in three typologically diverse languages: French, German, and Mandarin Chinese. For each language, ten speaker dyads were recorded, with scenarios designed to elicit conversation, including "Dream Apartment," "Film Script," and "Border Control." For this study, the "Film Script" scenario was chosen due to its high laughter occurrences. After filtering dyads with infrequent laughter, the final dataset comprised seven French dyads, eight German dyads, and six Chinese dyads, totaling over four hours of recordings and more than 1900 laughs and speech-laughs.

Key Insights and Ideas

  • Laughter is a site of alignment: The study provides evidence that laughter, like verbal communication, is subject to conversational alignment, suggesting a deeper, perhaps unconscious, coordination between speakers.
  • Language and culture may play a role: While synchrony and convergence showed similar trends across languages, the degree of alignment varied, hinting at potential language- or culture-specific discourse strategies influencing laughter dynamics.
  • Speech-laughs are important: The paper highlights the prevalence of "speech-laughs" (simultaneous speech and laughter) and advocates for their inclusion in future analyses, as they are often discarded but are a significant part of conversational laughter.
  • Different measures capture different phenomena: Synchrony and convergence, while related, are not perfectly correlated. High overall laughter quantity doesn't guarantee synchronized laughter dynamics, and vice-versa.

Key Results

  • Synchrony: Weak to moderate correlations were observed for synchrony across languages: p = 0.255 for French, p = 0.508 for German, and p = 0.456 for Mandarin.
  • Convergence: A trend towards convergence was observed in all three languages, with the difference in laughter amounts between partners decreasing in the second half of the conversation. However, these differences were not statistically significant, likely due to the small sample size.
  • Agreement (German data only): A significantly higher voicing agreement was found for consecutive laughs (83.1%) compared to randomly sampled non-consecutive laughs (77.4%) in the German data (p < 0.05).

Limitations and Future Work

The authors acknowledge that the small sample size of speaker dyads per language is a significant limitation, potentially explaining the lack of statistical significance for convergence. They also note that laughter annotations in conversational corpora can be unreliable due to definition inconsistencies and technical errors, emphasizing the need for more reliable annotation methods.

Future work suggested includes:

  • Expanding the dataset: Investigating alignment with larger, more reliably annotated corpora.
  • Exploring segmental alignment: Examining whether laughter alignment extends to the phonetic quality of vowels within laughter sounds.
  • Investigating other languages and interaction types: To further understand the influence of language and culture.
  • Controlling for interpersonal dynamics: Future studies could explore how factors like relationship status and perceived closeness influence laughter alignment.

Summary updated

BibTeX
@inproceedings{ludusan2019laughter,
  author = {Bogdan Ludusan and Petra Wagner},
  title = {{Laughter Dynamics in Dyadic Conversations}},
  booktitle = {Interspeech},
  year = {2019},
}
Figure