Answer-driven knowledge graph question generation: a policy optimization approach via Deep Q-Network

Within the domain of natural language processing, the automatic generation of questions constitutes a critical research area, with the objective of formulating questions that exhibit both logical consistency and contextual appropriateness. Currently, mainstream methods rely on sequence-to-sequence neural network models; these models often neglect the effective encoding and utilization of answer information, leading to the generated questions frequently lacking semantic relevance to the preset answers. To enhance the consistency of question-answer pairs, this paper proposes an innovative answer-driven knowledge graph question generation framework. The core of this framework lies in two aspects: first, an efficient answer-context fusion mechanism is designed to explicitly inject answer semantics into context representations; second, The decoding procedure is innovatively guided through the application of Deep Q-Network (DQN) reinforcement learning. By employing a reward mechanism, the proposed method directly maximizes the sustained advantages of query quality and response relevance, instead of solely depending on conventional maximum likelihood estimation. Experimental results on benchmark WebQuestions and PathQuestions datasets reveal that the novel approach achieves improvements across automated evaluation metrics. Through comprehensive ablation analyses, we not only substantiate the individual efficacy of the designed answer-aware component and the DQN-based reinforcement learning scheme, but also uncover the critical role of their synergistic interaction in enhancing both the precision and the variety of the questions generated.