Generating Shifting Sentiment for a Conversational Agent

Proceedings of the NAACL HLT 2010 Workshop on Computational Approaches to Analysis and Generation of Emotion in Text, pages 89–97, Los Angeles, California, June 2010. c⃝2010 Association for Computational Linguistics Generating Shifting Sentiment for a Conversational Agent Simon Whitehead University of Melbourne, Australia srwhitehead@gmail.com Lawrence Cavedon RMIT University, Australia lawrence.cavedon@rmit.edu.au Abstract We investigate techniques for generating al- ternative output sentences with varying sen- timent, using (an approximation to) the Valentino method, based on SentiWordNet, of Guerini et al. We extend this method by filter- ing out unacceptable candidate sentences, us- ing bigrams sourced from different corpora to determine whether lexical substitutions are ap- propriate in the given context. We also com- pare the generated candidates against human judgements of whether the desired sentiment shift has occurred: our results suggest limi- tations with the overall knowledge-based ap- proach, and we propose potential directions for improvement. 1 Introduction The design of more natural or believable conver- sational agents (Bates, 1994; Pelachaud and Bilvi, 2003) requires the need for such agents to communi- cate affectively, by the display of emotion or attitude towards objects, other agents, or states of affairs. More engaging or influential agents may seek to ac- tually affect their conversational partner at a deeper level, for example, by influencing their emotional state (van der Sluis and Mellish, 2008). Previous work in this area has explored the use of gestures and facial expression (Caridakis et al., 2007) and rhythm and prosody of speech (Zovato et al., 2008) for ex- pressing affect; however there has been little work on generation of affective language in dialogue. Our general approach is inspired by (Fleischman and Hovy, 2002)’s work on generating different surface-level versions of utterance content, depend- ing on an agent’s appraisals towards objects, char- acters and events in its environment. While their approach is effective, it relies on manual creation of lexical alternatives, customized to the application domain. We are interested in approaches that will scale, and can be applied domain-independently. While our ultimate aim is generation of language that relects emotional state, in this work we in- vestigate the automatic generation of varying “sen- timent” in output utterances; we focus on senti- ment mainly due to the recent development of use- ful resources for this task. (Guerini et al., 2008)’s Valentino system is an approach to automatically generating candidate output utterances with differ- ent sentiment from an original; the authors suggest ECAs as a possible application scenario for their techniques. We explore this suggestion, implement- ing a lexical substitution (McCarthy and Navigli, 2007) approach to dialogue generation with sen- timent, using the Valentino approach and associ- ated resources. Lexical substitution approaches raise well-known challenges, and we investigate a number of techniques to address these in Section 4; for ex- ample, using bigrams and grammatical relations to determine which substitutions are acceptable based on their context in a sentence.1 Our techniques show improvement over naive lex- ical substitution; however, an evaluation with human subjects suggests that a deeper problem is that even “acceptable” …

Content truncated to 500 words