The Role of Disfluencies in Topic Classification of Human-Human Conversations

We investigate the impact of disfluencies on the task of classifying natural human-human conversations into topics. Disfluencies are distinctive to spoken language, and their effect on a number of spoken language understanding tasks, including spoken language classification, remains largely unknown. We use a subset of Switchboard-I annotated for disfluencies and topics, and investigate the effect of different disfluency categories with both true and automatically generated transcripts. We show that under the popular bag-of-words representation, even perfect disfluency filtering has a minimal impact on topic classification performance on hand-transcribed data. However, difference are larger with more complex representations (e.g. bigrams) and for some classifiers operating on recognizer transcripts.