Adversarial Multi-task Model for Emotion, Sentiment, and Sarcasm Aided Complaint Detection

Abstract

Automatic identification of consumer complaints about products or services purchased can be crucial for businesses and online merchants since they can utilize this knowledge to address the needs of their clients, including handling and resolving complaints. Previous studies on complaint detection do not consider sarcasm, which is often used to express a breach of expectation without directly stating the complaint. Furthermore, since every speech act is influenced by emotions, the customer’s emotional state has a considerable impact on the complaint expression. In this paper, we hypothesize that sarcasm, along with two closely related tasks of sentiment and emotion, could aid the process of complaint identification and thereby propose a deep multi-task framework to solve the four problems jointly. We manually annotate the recently released Complaints dataset with the emotion, sentiment, and sarcasm classes. We present an attention-based adversarial multi-task deep neural network model for complaint detection. Experimental results on the extended version of the Complaints dataset show the effectiveness of our proposed approach for complaint detection over the existing state-of-the-art system. The evaluation also demonstrates that the proposed multi-task system improves performance for the primary task, i.e., complaint detection, with the assistance of the three auxiliary tasks, emotion recognition, sentiment analysis, and sarcasm detection.