What Is Your Cause for Concern? Towards Interpretable Complaint Cause Analysis

Apoorva Singh, Prince Jha, Rohan Bhatia, Sriparna Saha

Abstract: The abundance of information available on social media and the regularity with which complaints are posted online emphasizes the need for automated complaint analysis tools. Prior study has focused chiefly on complaint identification and complaint severity prediction: the former attempts to classify a piece of content as either complaint or non-complaint. The latter seeks to group complaints into various severity classes depending on the threat level that the complainant is prepared to accept. The complainant’s goal could be to express disapproval, seek compensation, or both. As a result, the complaint detection model should be interpretable or explainable. Recognizing the cause of a complaint in the text is a crucial yet untapped area of natural language processing research. We propose an interpretable complaint cause analysis model that is grounded on a dyadic attention mechanism. The model jointly learns complaint classification, emotion recognition, and polarity classification as the first sub-problem. Subsequently, the complaint cause extraction and the associated severity level prediction as the second sub-problem. We add the causal span annotation for the existing complaint classes in a publicly available complaint dataset to accomplish this. The results indicate that existing computational tools can be repurposed to tackle highly relevant novel tasks, thereby finding new research opportunities (Resources available at: https://bit.ly/Complaintcauseanalysis ).