The Influence of Context on Dialogue Act Recognition

This article presents a deep analysis of the influence of context information on dialogue act recognition. We performed experiments on the annotated subsets of three different corpora: the widely explored Switchboard Dialog Act Corpus, as well as the unexplored LEGO and Cambridge Restaurant corpora. In contrast with previous work, especially in what concerns the Switchboard corpus, we used an event-based classification approach, using Support Vector Machines (SVMs), instead of the more common sequential approaches, such as Hidden Markov Models (HMMs). We opted for such an approach so that we could control the amount of provided context information and, thus, explore its range of influence. Our base features consist of n-grams, punctuation, and whwords. Context information is obtained from previous utterances and provided in three ways – n-grams, n-grams tagged with relative position, and dialogue act classifications. A comparative study was conducted to evaluate the performance of the three approaches. From it, we were able to assess the importance of context information on dialogue act recognition, as well as its range of influence for each of the three selected representations. In addition to the conclusions originated by the analysis, this work also produced results that advance the state-of-the-art, especially considering previous work on the Switchboard corpus. Furthermore, since, to our knowledge, the Email addresses: [email protected] (Eugénio Ribeiro), [email protected] (Ricardo Ribeiro), [email protected] (David Martins de Matos) Preprint submitted to Computer Speech and Language June 3, 2015 ar X iv :1 50 6. 00 83 9v 1 [ cs .C L ] 2 J un 2 01 5 remaining datasets had not been previously explored for this task, our experiments can be used as baselines for future work on those corpora.