Incremental Dialogue Act Recognition: Token- vs Chunk-Based Classification

This paper presents a machine learning based approach to incremental dialogue act classification with a focus on the recognition of communicative functions associated with dialogue segments in a multidimensional space, as defined in the ISO 24617-2 dialogue act annotation standard. The main goal is to establish the nature of an increment whose processing will result in a reliable overall system performance. We explore scenarios where increments are tokens or syntactically, semantically or prosodically motivated chunks. Combing local classification with meta-classifiers at a late fusion decision level we obtained state-of-the-art classification performance. Experiments were carried out on manually corrected transcriptions and on potentially erroneous ASR output. Chunk-based classification yields better results on the manual transcriptions, whereas token-based classification shows a more robust performance on the ASR output. It is also demonstrated that layered hierarchical and cascade training procedures result in better classification performance than the single-layered approach based on a joint classification predicting complex class labels.