Classifying Speech Acts using Verbal Response Modes

Proceedings of the 2006 Australasian Language Technology Workshop (ALTW2006), pages 34–41. Classifying Speech Acts using Verbal Response Modes Andrew Lampert †‡ †CSIRO ICT Centre Locked Bag 17, North Ryde NSW 1670 Australia firstname.lastname@csiro.au Robert Dale ‡ C´ecile Paris † ‡Centre for Language Technology Macquarie University NSW 2109 Australia [alampert,rdale]@ics.mq.edu.au Abstract The driving vision for our work is to provide intelligent, automated assistance to users in understanding the status of their email conversations. Our approach is to create tools that enable the detec- tion and connection of speech acts across email messages. We thus require a mech- anism for tagging email utterances with some indication of their dialogic function. However, existing dialog act taxonomies as used in computational linguistics tend to be too task- or application-specific for the wide range of acts we find repre- sented in email conversation. The Ver- bal Response Modes (VRM) taxonomy of speech acts, widely applied for discourse analysis in linguistics and psychology, is distinguished from other speech act tax- onomies by its construction from cross- cutting principles of classification, which ensure universal applicability across any domain of discourse. The taxonomy cat- egorises on two dimensions, characterised as literal meaning and pragmatic mean- ing. In this paper, we describe a statisti- cal classifier that automatically identifies the literal meaning category of utterances using the VRM classification. We achieve an accuracy of 60.8% using linguistic fea- tures derived from VRM’s human annota- tion guidelines. Accuracy is improved to 79.8% using additional features. 1 Introduction It is well documented in the literature that users are increasingly using email for managing requests and commitments in the workplace (Bellotti et al., 2003). It has also been widely reported that users commonly feel overloaded when managing mul- tiple ongoing tasks through email communication e.g. (Whittaker and Sidner, 1996). Given significant task-centred email usage, one approach to alleviating email overload in the workplace is to draw on Speech Act Theory (Searle, 1969) to analyse the intention behind email messages and use this information to help users process and prioritise their email. The basic tenet of Speech Act Theory is that when we utter something, we also act. Examples of such acts can include stating, questioning or advising. The idea of identifying and exploiting patterns of communicative acts in conversations is not new. Two decades ago, Flores and Winograd (1986) proposed that workplace workflow could be seen as a process of creating and maintaining networks of conversations in which requests and commit- ments lead to successful completion of work. Recently, these ideas have begun to be ap- plied to email messages. Existing work analysing speech acts in email messages differs as to whether speech acts should be annotated at the message level, e.g., (Cohen et al., 2004; Leuski, 2004), or at the utterance or sentence level, e.g., (Corston- Oliver et al., 2004). Our thesis is that a single email message may contain multiple commitments on a range of tasks, and so our work focuses on utterance-level classification, with the aim of be- ing …

Content truncated to 500 words