Learning to model domain-specific utterance sequences for extractive summarization of contact center dialogues

This paper proposes a novel extractive summarization method for contact center dialogues. We use a particular type of hidden Markov model (HMM) called Class Speaker HMM (CSHMM), which processes operator/caller utterance sequences of multiple domains simultaneously to model domain-specific utterance sequences and common (domain-wide) sequences at the same time. We applied the CSHMM to call summarization of transcripts in six different contact center domains and found that our method significantly outperforms competitive baselines based on the maximum coverage of important words using integer linear programming.