Influence relation estimation based on lexical entrainment in conversation Tomoharu Iwata ⇑ , Shinji Watanabe 1 NTT Communication Science Laboratories, Sorakugun, Kyoto 619-0237, Japan Received 20 December 2011; received in revised form 24 August 2012; accepted 27 August 2012 Available online 12 September 2012 Abstract In conversations, people tend to mimic their companions’ behavior depending on their level of trust. This phenomenon is known as entrainment. We propose a probabilistic model for estimating influences among speakers from conversation data involving multiple peo- ple by modeling lexical entrainment. The proposed model estimates word use as a function of the weighted sum of the earlier word use of other speakers. The weights represent influences between speakers. The influences can be efficiently estimated by using the expectation maximization (EM) algorithm. We also develop its online inference procedures for sequentially modeling the dynamics of influence rela- tions. Experiments performed on two meeting data sets one in Japanese and one in English demonstrate the effectiveness of the proposed method. Ó 2012 Elsevier B.V. All rights reserved. Keywords: Conversation analysis; Influence; Latent variable model; Entrainment 1. Introduction In conversations, people tend to mimic such aspects of their companions’ behavior as posture (Chartrand and Bargh, 1999), facial expression (Dimberg, 1982), lexicon (Brennan, 1996; Nenkova et al., 2008), syntax (Reitter et al., 2006), acoustic and prosodic features (Levitan and Hirschberg, 2011) and amplitude (Coulston et al., 2002). This phenomenon is known in the literature as entrainment, accommodation, adaptation, or alignment (Giles et al., 1991; Pickering and Garrod, 2004). Entrainment is said to indicate that people are trusting, accommodating and empathic (Chartrand and Bargh, 1999; Scissors et al., 2008). This paper focuses on the entrainment of lexicon in polylogues, or how people are influenced by their compan- ions in terms of word use in conversation with multiple speakers. The degree to which a person exerts an influence and is influenced by others varies from speaker to speaker. A powerful person is likely to be mimicked by others, and a passive person might often be accommodating to others. The influences also differ between pairs of speakers depend- ing on their level of trust. For example, Alice might use words spoken by Bob, but not words spoken by Charlie. The influences therefore have an asymmetric nature. We propose a simple and effective probabilistic model for estimating influence relations among speakers from conver- sation data with multiple people (Iwata and Watanabe, 2011). With the proposed model, we assume that a speaker’s word use depends on the preceding word use of other speak- ers as well as his/her own preceding word use and the gen- eral word use. We estimate the strength of influence for each pair of speakers by fitting the proposed model to the given conversation data using the expectation maximization (EM) algorithm (Dempster et al., 1977). We also develop online inference procedures for sequentially modeling how the influence relations change in the conversation over time. Note that the proposed model estimates influences on the 0167-6393/$ - see front matter Ó 2012 Elsevier B.V. All rights reserved. http://dx.doi.org/10.1016/j.specom.2012.08.012 ⇑ Corresponding author. Tel.: +81 774 93 5161; fax: +81 774 93 5155. E-mail address: iwata.tomoharu@lab.ntt.co.jp (T. Iwata). URL: http://www.kecl.ntt.co.jp/as/members/iwata/ (T. Iwata). 1 Shinji Watanabe is now with Mitsubishi Electric Research Laboratories. www.elsevier.com/locate/specom Available online at www.sciencedirect.com Speech Communication 55 (2013) 329–339