Question Answering through Enhanced Question Classification for Domain Adaptation

The field of Question Answering (QA) has achieved substantial advancements, particularly in specialized domains. However, adapting QA systems to new domains when trained in a distinct source domain and deployed in a different target domain presents significant challenges. This study explores the role of enhanced question classification in facilitating domain adaptation in QA systems. We introduce a comprehensive framework termed Enhanced Question Classification for Adaptive QA (EQCA-QA). This framework employs a sophisticated question classifier to categorize questions from both source and target datasets, followed by a joint training mechanism that utilizes self-supervised learning through pseudo-labeling. To optimize performance, we reduce the inter-domain discrepancies using the maximum mean discrepancy (MMD) method while also minimizing intra-class variances among QA pairs of identical question types for more precise adaptation. To our knowledge, this is the pioneering attempt at integrating enhanced question classification with self-supervised learning techniques to address domain adaptation challenges in QA. Our results show that EQCA-QA consistently surpasses the contemporary benchmarks across various datasets.