A Two-stage User Intent Detection Model on Complicated Utterances with Multi-task Learning

Shuangyong Song, Xiangyan Chen, Chao Wang, Xiaoguang Yu, Jia Wang, Xiaodong He

Abstract: As one of the most natural manner of human–machine interaction, the dialogue systems have attracted much attention in recent years, such as chatbots and intelligent customer service bots, etc. Intents of concise user utterances can be easily detected with classic text classification models or text matching models, while complicated utterances are harder to understand directly. In this paper, for improving the user intent detection from complicated utterances in an intelligent customer service bot JIMI (JD Instant Messaging intelligence), which is designed for creating an innovative online shopping experience in E-commerce, we propose a two-stage model which combines sentence Compression and intent Classification together with Multi-task learning, called MCC. Besides, a dialogue-oriented language model is trained to further improve the performance of MCC. Experimental results show that our model can achieve good performance on both a public dataset and the JIMI dataset.