RoQUEST: A Robust QUESTion Classification Framework in Consumer Electronics Domain

Question-answering (QA) has been a crucial NLP (Natural Language Processing) task for a long time. As an umbrella term, QA has multiple sub-tasks under its hood. Question classification (QC) is one of the most important sub-tasks of any QA pipeline. By providing hints about the classes or types of incoming questions and/or potential answers, a QC module enables the QA system to implement appropriate approaches/strategies to handle different types of questions to produce better answers. Although QA systems have evolved significantly over time in the last few decades and witnessed a paradigm shift from earlier IR (Information Retrieval) based systems to more recent RAG (Retrieval Augmented Generation) based systems, there are very few QC frameworks available in the research community that can match and support such a wide range of QA systems. Moreover, a robust QA pipeline must be able to answer any type of question, which is essential in a QA system, regardless of the paradigm. However, most of the research works and datasets available in the literature have been proposed considering QA systems consisting of only factoid questions, and the QC frameworks for those systems are supposed to help in performance improvement by predicting their subtypes (also referred to as answer types in the literature). In our present work, we propose a QC framework to bridge these gaps. Our proposed QC framework consists of three types of questions, viz., factoid, confirmation, and descriptive questions in the consumer electronics domain. Consumer electronics is an important domain where the users have plenty of questions that could be addressed by a robust automated QA system. However, a suitable QC framework is not available in the literature, which could help to build a robust QA pipeline in this domain. In our current research, we address this issue by annotating our proposed QC dataset in the above mentioned domain. Moreover, the questions in our proposed dataset have two novel features, namely factoid/descriptive questions without “wh-words” and use of “polite phrases” in the questions. The inclusion of these two features makes our proposed dataset more robust and realistic, but it also makes the question classification task more challenging, since none of the existing datasets support these features. To address this challenge, we suggest a novel data augmentation technique to train our models. Finally, we propose a novel question classifier that works as a benchmark setup for the proposed dataset. We compare the performance of our proposed architecture with previously proposed models for the QC task and demonstrate the superiority of our proposed architecture.