Reinforcement learning-driven deep question generation with rich semantics

Abstract

Deep question generation (DQG) refers to generating a complex question from different sentences in context. Existing methods mainly focus on enhancing information extraction based on the encoder–decoder neural networks though they cannot perform well in DQG tasks. To address this issue, we consider combining reinforcement learning with semantic-rich information to generate deep questions in this paper. In particular, we propose a Semantic-Rich Reinforcement Learning Deep Question Generation (SRL-DQG) model, which better utilizes the semantic graphs of document representations based on the Gated Graph Neural Network (GGNN). In order to generate high-quality questions, we also optimize specific objectives via reinforcement learning with consideration of four evaluation factors including naturality, relevance, answerability, and difficulty. Empirical evaluations demonstrate that our SRL-DQG effectively improves the quality of generated questions and achieves superior performance than existing methods in terms of multiple performance metrics. Specifically, we show that several BLEU-n scores were improved by 3.5% to 10% after running SRL-DQG on 6072 samples of HotPotQA.