Using Past Speaker Behavior to Better Predict Turn Transitions

This paper explores using a summary of past speaker behavior to better predict turn transitions. We computed two types of summary features that represent the current speaker’s past turn-taking behavior: relative turn length and relative floor control. Relative turn length measures the current turn length so far (in time and words) relative to the speaker’s average turn length. Relative floor control measures the speaker’s control of the conversation floor (in time and words) relative to the total conversation length. The features are recomputed for each dialog act based on past turns of the speaker within the current conversation. Using the switchboard corpus, we trained two models to predict turn transitions: one with just local features (e.g., current speech act, previous speech act) and one that added the summary features. Our results shows that using the summary features improve turn transitions prediction.