Practical and efficient out-of-domain detection with adversarial learning
Bo Wang, Tsunenori Mine · Proceedings of the 37th ACM/SIGAPP Symposium on Applied Computing · 2022
Detecting Out-of-Domain (OOD) questions is important in real-world applications and is the subject of active research. Traditional In-Domain (IND) classifier-based methods can easily distinguish OOD questions if they are very different from IND ones, but not otherwise. Although large-scale network-based approaches have been used to solve this problem and obtained better results, the computational costs associated with the training and prediction processes are usually very high. In addition, these methods did not seek to leverage any unlabeled data. To address these issues, adversarial learning methods have been widely used in the field of image processing. At the same time, there are currently only a few practices in the domain of text classification, and their performance remains suboptimal. In this paper, we propose a novel practical and efficient OOD question detection framework using adversarial learning. We conducted a range of experiments on three open datasets. The experimental results demonstrate the advantages of our framework compared to baseline methods.