Improving Intent Detection Accuracy Through Token Level Labeling

Abstract

Intent detection is traditionally modeled as a sequence classification task where the role of the models is to map the users’ utterances to their class. In this paper, however, we show that the classification accuracy can be improved with the use of token level intent annotations and introducing new annotation guidelines for labeling sentences in the intent detection task. What is more, we introduce a method for training the network to predict joint sentence level and token level annotations. We also test the effects of different annotation schemes (BIO, binary, sentence intent) on the model’s accuracy. 2012 ACM Subject Classification Computing methodologies → Natural language processing