Short-Spoken Language Intent Classification with Conditional Sequence Generative Adversarial Network

Abstract

As one of the most thrilling tasks of natural language understanding (NLU), intent classification in a dialogue system has received a great deal of attention in both industry and academia. The major limiting factor on intent classification is the lack of tagged data. To solve it, in this paper, we propose a conditional sequence generative adversarial network (cSeq-GAN) for intent classification of short-spoken language, in which we simultaneously train a generative model and a discriminative model for two tasks, one to distinguish the generated text from the real spoken one, while the other to predict its intent category. More reliable tagged data obtained by the generator greatly improves the performance of the intent classification task. Extensive experiments on both Air Travel Information System (ATIS) and our selling robot dialogue system for insurance industries demonstrate that our cSeq-GAN achieves competitive classification accuracy with other state-of-art methods of text classification.