Reinforcement Learning for Abstractive Question Summarization with Question-aware Semantic Rewards

The growth of online consumer health questions has led to the necessity for\nreliable and accurate question answering systems. A recent study showed that\nmanual summarization of consumer health questions brings significant\nimprovement in retrieving relevant answers. However, the automatic\nsummarization of long questions is a challenging task due to the lack of\ntraining data and the complexity of the related subtasks, such as the question\nfocus and type recognition. In this paper, we introduce a reinforcement\nlearning-based framework for abstractive question summarization. We propose two\nnovel rewards obtained from the downstream tasks of (i) question-type\nidentification and (ii) question-focus recognition to regularize the question\ngeneration model. These rewards ensure the generation of semantically valid\nquestions and encourage the inclusion of key medical entities/foci in the\nquestion summary. We evaluated our proposed method on two benchmark datasets\nand achieved higher performance over state-of-the-art models. The manual\nevaluation of the summaries reveals that the generated questions are more\ndiverse and have fewer factual inconsistencies than the baseline summaries\n