HAIN: Multilabel Classification With Hierarchical Attention-Based Interaction Network for Multiturn Dialogue Texts

Abstract

Multi-turn dialogue systems with cognitive capabilities have been popularly used in many applications and generate a large volume of dialogue texts. The dialogue texts of many dialogue systems contain multiple labels, and a typical case is that the customer requirements or complaints in an online multi-turn dialogue service system often need to be classified into multiple categories or forwarded to multiple departments, that is, multi-label classification on multi-turn dialogue texts. Traditional methods treat the dialogue text as the plain text, which makes it difficult for classifiers to learn important features. On the other hand, regarding each utterance as the basic semantic unit may destroy the semantic information integrity of the dialogue. Furthermore, the above methods all ignore the relationship between labels. To address above issues, we propose to use the utterance pair as the basic semantic unit and design a novel multi-label classification model for multi-turn dialogue texts in this paper. Specifically, we propose to use multi-head attention mechanism to learn the key features of multiple labels and design a new method to learn the correlation among labels. Besides, we propose to use penalty term to allow multi-head attention to focus on different semantics. Finally, we conduct sufficient experiments on four real-world multi-turn dialogue datasets and the experimental results show that our model produces superior performance.