Extracting Social Meaning: Identifying Interactional Style in Spoken Conversation

Human Language Technologies: The 2009 Annual Conference of the North American Chapter of the ACL, pages 638–646, Boulder, Colorado, June 2009. c⃝2009 Association for Computational Linguistics Extracting Social Meaning: Identifying Interactional Style in Spoken Conversation Dan Jurafsky Linguistics Department Stanford University jurafsky@stanford.edu Rajesh Ranganath Computer Science Department Stanford University rajeshr@cs.stanford.edu Dan McFarland School of Education Stanford University dmcfarla@stanford.edu Abstract Automatically extracting social meaning and intention from spoken dialogue is an impor- tant task for dialogue systems and social com- puting. We describe a system for detecting elements of interactional style: whether a speaker is awkward, friendly, or flirtatious. We create and use a new spoken corpus of 991 4-minute speed-dates. Participants rated their interlocutors for these elements of style. Us- ing rich dialogue, lexical, and prosodic fea- tures, we are able to detect flirtatious, awk- ward, and friendly styles in noisy natural con- versational data with up to 75% accuracy, compared to a 50% baseline. We describe sim- ple ways to extract relatively rich dialogue fea- tures, and analyze which features performed similarly for men and women and which were gender-specific. 1 Introduction How can we extract social meaning from speech, de- ciding if a speaker is particularly engaged in the con- versation, is uncomfortable or awkward, or is partic- ularly friendly and flirtatious? Understanding these meanings and how they are signaled in language is an important sociolinguistic task in itself. Extracting them automatically from dialogue speech and text is crucial for developing socially aware computing systems for tasks such as detection of interactional problems or matching conversational style, and will play an important role in creating more natural dia- logue agents (Pentland, 2005; Nass and Brave, 2005; Brave et al., 2005). Cues for social meaning permeate speech at every level of linguistic structure. Acoustic cues such as low and high F0 or energy and spectral tilt are impor- tant in detecting emotions such as annoyance, anger, sadness, or boredom (Ang et al., 2002; Lee and Narayanan, 2002; Liscombe et al., 2003), speaker characteristics such as charisma (Rosenberg and Hirschberg, 2005), or personality features like extro- version (Mairesse et al., 2007; Mairesse and Walker, 2008). Lexical cues to social meaning abound. Speakers with links to depression or speakers who are under stress use more first person singular pro- nouns (Rude et al., 2004; Pennebaker and Lay, 2002; Cohn et al., 2004), positive emotion words are cues to agreeableness (Mairesse et al., 2007), and neg- ative emotion words are useful cues to deceptive speech (Newman et al., 2003). The number of words in a sentence can be a useful feature for extroverted personality (Mairesse et al., 2007). Finally, dia- log features such as the presence of disfluencies can inform listeners about speakers’ problems in ut- terance planning or about confidence (Brennan and Williams, 1995; Brennan and Schober, 2001). Our goal is to see whether cues of this sort are useful in detecting particular elements of conversa- tional style and social intention; whether a speaker in a speed-dating conversation is judged by …

Content truncated to 500 words