A Unified Abstractive Model for Generating Question-Answer Pairs

Anonymous

Abstract: Large-scale question-answer pairs (QAP) are valuable for many applications, such as knowledge bases construction and machine reading comprehension. Although its importance has been widely recognized, existing approaches are still faced with critical challenges. On the one hand, QAPs are obtained by selecting spans from original texts as their answers, while abstractive answer generation is more suitable and natural for complex QA applications. On the other hand, the interaction between the sub-tasks of answer generation and question generation should be well captured to enhance each other mutually. To this end, we propose a Unified Abstractive model for Question-Answer Pairs generation (UA-QAP). Specifically, we devise the joint model with a query-guided gate to collectively model the two sub-tasks simultaneously and capture the interaction information between them. Therefore, our model can generate semantically comprehensive question-answer pairs. We conduct extensive experiments on three large-scale datasets. The experimental results demonstrate that our model achieves state-of-the-art performance.