Emotion and behaviour in automatic dialogue summarisation

This paper presents an overview of a six-year research project on automatic summarisation of emotional and behavioural features in dialogues. It starts by describing some evidence for the hypothesis that whenever a dialogue features very impolite behaviour, this behaviour will tend to be described in the dialogue's summary, with a bias influenced by the summariser's viewpoint. It also describes the role some experiments played in providing useful information on when and how assessments of emotion and behaviour should be added to a dialogue summary, along with the necessary steps (such as the development of a multi-dimensional annotation scheme) to use these experimental results as a starting point for the automatic production of summaries. Finally, it introduces an automatic dialogue summariser capable of combining technical and emotional or behavioural information in its output summaries.