Speech act profiling: a probabilistic method for analyzing persistent conversations and their participants

Abstract

The increase in persistent conversations in the form of chat and instant messaging (IM) has presented new opportunities for researchers. This paper describes a method for evaluating and visualizing persistent conversations by creating a speech act profile for conversation participants using speech act theory and concepts from fuzzy logic. This method can be used either to score a participant based on possible intentions or to create a visual map of those intentions. Transcripts from the Switchboard corpus, which have been marked up with speech act labels according to a SWBD-DAMSL tag set of 42 tags, are used to train language models and a modified hidden Markov model (HMM) to obtain probabilities for each speech act type for a given sentence. Rather than choosing the speech act with the maximum probability and assigning it to the sentence, the probabilities are aggregated for each conversation participant creating a set of speech act profiles, which can be visualized as a radar graphs. Several example profiles are shown along with possible interpretations. The profiles can be used as an overall picture of a conversation, and may be useful in various analyses of persistent conversations including information retrieval, deception detection, and online technical support monitoring.