The Brooklyn Multi-Interaction Corpus for Analyzing Variation in Entrainment Behavior

Proceedings of the 13th Conference on Language Resources and Evaluation (LREC 2022), pages 1721–1731 Marseille, 20-25 June 2022 © European Language Resources Association (ELRA), licensed under CC-BY-NC-4.0 1721 The Brooklyn Multi-Interaction Corpus for Analyzing Variation in Entrainment Behavior Andreas Weise1, Matthew McNeill1, Rivka Levitan1,2 1Department of Computer Science, The Graduate Center, CUNY, New York, NY, USA 2Department of Computer and Information Science, Brooklyn College, CUNY, Brooklyn, NY, USA {aweise, mmcneill}@gradcenter.cuny.edu, rlevitan@brooklyn.cuny.edu Abstract We present the Brooklyn Multi-Interaction Corpus (B-MIC), a collection of dyadic conversations designed to identify speaker traits and conversation contexts that cause variations in entrainment behavior. B-MIC pairs each participant with multiple partners for an object placement game and open-ended discussions, as well as with a Wizard of Oz for a baseline of their speech. In addition to fully transcribed recordings, it includes demographic information and four completed psychological questionnaires for each subject and turn annotations for perceived emotion and acoustic outliers. This enables the study of speakers’ entrainment behavior in different contexts and the sources of variation in this behavior. In this paper, we introduce B-MIC and describe our collection, annotation, and preprocessing methodologies. We report a preliminary study demonstrating varied entrainment behavior across different conversation types and discuss the rich potential for future work on the corpus. Keywords: Entrainment, Variation, Dialogue Corpus 1. Introduction Entrainment is the phenomenon of conversational part- ners adapting to one another to become more similar. This affects a wide variety of linguistic dimensions, including lexical choice (Brennan and Clark, 1996), syntax (Reitter et al., 2006), phonetics (Pardo, 2006), and prosody (Levitan and Hirschberg, 2011; Gravano et al., 2014). It has also been found in many con- versational contexts, from collaborative, personal in- teractions, in pairs (Brennan and Clark, 1996; Levitan and Hirschberg, 2011; Lubold and Pon-Barry, 2014) and in groups (Rahimi et al., 2017), to conversations over the phone (Cohen Priva and Sanker, 2020) or on Twitter (Danescu-Niculescu-Mizil et al., 2011), to even human-computer interaction (Coulston et al., 2002; Suzuki and Katagiri, 2007; Thomason et al., 2013). Notably, it correlates with, among others, success in tasks (Reitter and Moore, 2007) and learning (Thoma- son et al., 2013), rapport (Lubold and Pon-Barry, 2014) and social behavior (Levitan et al., 2012), and the per- ceived trustworthiness and likability of artificial inter- locutors (Levitan et al., 2016; Metcalf et al., 2019). The degree and valence with which entrainment occurs vary substantially by the speakers and context (Lubold and Pon-Barry, 2014; Pardo et al., 2018; Weise et al., 2019). Despite theories purporting to broadly explain the behavior (Giles et al., 1991; Chartrand and Bargh, 1999; Pickering and Garrod, 2004), the exact mecha- nisms governing its emergence in individual conversa- tions are still poorly understood. Attempts to attribute differences to gender, for instance, have yielded dis- parate results or no effect (Pardo et al., 2018; Weise et al., 2019). This lack of theoretical understanding also hinders broader practical application, as appro- priate settings for entraining conversational avatars are difficult to predict. This can even result in …

Content truncated to 500 words