Dialog-Act Recognition Using Discourse and Sentence Structure Information

Automatic recognition of Dialog-act (DA) is one of the most important processes in understanding spontaneous dialog. Most existing studies have been working on how to use various classifying methods in DA recognition; meanwhile, less attention has been paid to feature selection specifically. This paper introduces several textual features for DA recognizing, and proposes a novel usage for sentence structure features. Especially, this paper investigates the effect of discourse structure features in DA recognition, which are little studied before. The experimental results on both Chinese corpus and English Corpus show the selected features and feature combination rules significantly improve the overall performance. The accuracy of DA recognition rises from 77.05% to 88.21% on Chinese corpus, and from 59.08% to 64.92% as well on English corpus.