Children Use Speech Disfluencies to Make Social Inferences

• In Event: Children’s Utilization of Shared Information Across Contexts

Abstract

Human communication has the remarkable capacity to transmit information between agents. However, information is not beamed telepathically between agents; instead, it must be packaged into natural language messages and delivered in a timely manner. When burdens on processing time prevent timeliness, speakers use disfluencies (e.g., “um”) to hold the conversational floor while retrieving information. These disfluencies can thus provide insight into a speaker’s mental states. Indeed, adult listeners can use disfluencies to infer speaker knowledge (Brennan & Williams, 1995), and also silent pauses to infer speaker desires (Roberts et al., 2011). Do children make these social inferences from disfluencies?

Children produce disfluencies in their own speech from a young age (Casillas, 2014) and, children age 2 can use disfluencies to predict what object a speaker is talking about (Kidd et al., 2011). In two studies, we show that children (ages 4-9) are able to draw social inferences (knowledge in study 1, and preferences in study 2) based on speech disfluencies, while content is held constant.

In study 1 (ages 4-9, n = 101, pre-registered n = 120, data collection paused due to COVID-19), children saw a story where two speakers were asked to label familiar animals. In the labelling condition, both speakers accurately labelled an animal, but one used a disfluency (e.g., saying “This one is a… uh… tiger”). Children were then asked who knows more about animals and, across the age range, reported that the fluent speaker knows more than the disfluent speaker (M = 1.67 of out 2 trials, 95% CI [1.49, 1.83]). Importantly and in line with adult work (Brennan & Williams, 1995), in a separate condition, children did not infer a lack of knowledge from a speech disfluency when both people gave non-answers (saying “Uh… I don’t know”) (M = 0.90 out of 2 trials, 95% CI [0.67, 1.15]), demonstrating that children recognize that disfluencies do not always signal incompetence. Before COVID-19, data collection was completed for 4-5 year-olds and 6-7 year-olds, and this pattern of results is evident even in our 4-5 year-old sample.

In study 2, we tested how children (ages 4-9, current n = 57, pre-registered n = 60, data collected remotely) use disfluencies to recover preference information. In this story, speakers were asked “Which animal do you like the best?” and both speakers independently stated the same preference, with one using a disfluency (e.g., saying “Um… the tiger is my favorite.”). Data collection is ongoing, but consistent with a separate pilot sample suggesting that children by age 6 infer that an agent who states their preference quickly has a stronger preference.

In sum, this work suggests that young children consider the answer someone gives, but also how they give it. These data also show that children are not heuristically responding to auditory disfluency, but instead understand contexts when such cues are less relevant (i.e. for non-answers). Broadly, our account emphasizes that disfluencies serve as general cues and contextualized interpretations allow a range of social inferences.

Authors

• Ben Morris, University of Chicago
  Presenting Author

• Daniel Yurovsky, Carnegie Mellon University
  Non-Presenting Author

• Alex Shaw, University of Chicago
  Non-Presenting Author