Machine Reading Comprehension for Multiclass Questions on Thai Corpus

Abstract

Previous Thai question answering and machine reading comprehension researches focus on small scale dataset and do not utilize the deep learning approach to build the models. In this research, we develop a Thai machine reading comprehension (MRC) model on Thai MRC dataset provided by NECTEC. This dataset consists of 17,000 question-answer pairs and has two classes of questions, which are factoid and yes-no questions. We use BIDAF as the based MRC architecture. We have performed experiments with 3 different multiclass model designs, which includes special tokens, joint, and cascade model. We also utilize contextual embeddings for Thai language to enhance the model’s performance. As the results suggest that cascade architecture has the best F1 performance. We then incorporate transfer learning and modify the attention mechanisms to increase the model’s accuracy on yes-no questions.