Japanese How-to Tip Machine Reading Comprehension by Multi-task Learning Based on Generative Model

Xiaotian Wang, Tingxuan Li, Takuya Tamura, Shunsuke Nishida, Fuzhu Zhu, Takehito Utsuro

Abstract: In the research of machine reading comprehension of Japanese how-to tip QA tasks, conventional extractive machine reading comprehension methods have difficulty in dealing with cases in which the answer string spans multiple locations in the context. In this paper, we trained a generative machine reading comprehension model of Japanese how-to tip by constructing a generative dataset based on the website “wikihow” as a source of information. We proposed two methods for multi-task learning to fine-tune the generative model, i.e., i) multi-task learning with generative and extractive hybrid training dataset, where both generative and extractive datasets are simultaneously trained on a single model, and ii) multi-task learning with inter-sentence semantic similarity and answer generation, where, drawing upon the answer generation task, the model additionally learns the distance between the sentences of question/context and the answer in the training examples. Evaluation experimental results showed that both of the multi-task learning models significantly outperformed that of the single-task learning model on the generative QA dataset. Especially, that with generative and extractive hybrid training dataset performed the best in terms of the manual evaluation result.