DR.QG: Enhancing Closed-Domain Question Answering via Retrieving Documents for Question Generation

Zhi-Wei Tong, Yao-Chung Fan, Fang-Yie Leu

Abstract: In this paper, we propose a new framework called Document-Retrieval Question Generation (DR.QG). The goal of DR.QG is to generate a question corresponding to a given answer. Differing from the common question generation setting, DR.QG takes only answers for question generation, while existing question generation takes a context passage and an answer as input for generating. To achieve this goal, we explored the possibility of importing document retrieval. Through the performance evaluation on the Question Answering (QA) task, we demonstrate the feasibility of DR.QG. The result shows that our method improves QA performance by up to 13%. Furthermore, we simulate the closed-domain situation on the open-domain dataset and show that we improved performance by 3%.