Aspect-Based Complaint and Cause Detection: A Multimodal Generative Framework with External Knowledge Infusion

Abstract

Customer reviews often contain valuable feedback about a product or service, but it can be challenging to extract specific complaints and their underlying causes from the text. Despite the use of various methods to detect and analyze complaints, no studies have concentrated on thoroughly examining complaints at the aspect-level and the underlying reasons for such aspect-level complaints. We add the rationale annotation for the aspect-based complaint classes in a publicly available benchmark multimodal complaint dataset (CESAMARD), which spans five domains (books, electronics, edibles, fashion, and miscellaneous). Current multimodal complaint detection methods treat these tasks as classification problems and do not utilize external knowledge. The present study aims to tackle these concerns. We propose a knowledge-infused unified Multimodal Generative framework for Aspect-based complaint and Cause detection (MuGACD) by reframing the multitasking problem as a multimodal text-to-text generation task. Our proposed methodology established a benchmark performance in the novel aspect-based complaint and cause detection task based on extensive evaluation. We also demonstrated that our model consistently outperformed all other baselines and state-of-the-art models in both full and few-shot settings (The dataset and code are available at https://github.com/Raghav10j/ECML23 ).