Knowing What and How: A Multi-modal Aspect-Based Framework for Complaint Detection

With technological advancements, the proliferation of e-commerce websites and social media platforms has created an avenue for customers to provide feedback to enterprises based on their overall experience. Customer feedback serves as an independent validation tool that could boost consumer trust in the brand. Whether it is a recommendation or review of a product, it provides insight allowing businesses to understand what they are doing right or wrong. By automatically analyzing customer complaints at the aspect-level enterprises can connect to their customers by customizing products and services according to their needs quickly and deftly. In this paper, we introduce the task of Aspect-Based Complaint Detection (ABCD). ABCD identifies the aspects in the given review about a product and also finds if the aspect mentioned in the review signifies a complaint or non-complaint. Specifically, a task solver must detect duplets (What, How) from the inputs that show WHAT the targeted features are and HOW they are complaints. To address this challenge, we propose a deep-learning-based multi-modal framework, where the first stage predicts what the targeted aspects are, and the second stage categorizes whether the targeted aspect is associated with a complaint or not. We annotate the aspect categories and associated complaint/non-complaint labels in the recently released multi-modal complaint dataset (CESAMARD), which spans five domains (books, electronics, edibles, fashion, and miscellaneous). Based on extensive evaluation our methodology established a benchmark performance in this novel aspect-based complaint detection task and also surpasses a few strong baselines developed from state-of-the-art related methods (Resources available at: https://github.com/appy1608/ECIR2023_Complaint-Detection ).