Question-Answer Pairs Generation Based on Interaction-Guided Joint Abstractive Model
Automatically generating large-scale question-answer pairs is valuable for many applications such as knowledge base construction and machine reading comprehension.Although its importance has been widely recognized,existing approaches to question-answer pair generation still face serious challenges.First,in traditional question-answer pair generation models,extractive answer acquisition methods are difficult to apply to complex natural interaction scenarios.In contrast,generative models can automatically generate answers with more natural expressions through semantic understanding of text.Second,for the question-answer pair generation task,the interaction between the two subtasks of answer generation and question generation needs to be captured and enhanced more comprehensively in order to prevent semantic mismatches between the generated answers and questions.Finally,due to the difference in task difficulty between answer extraction and question generation,the joint learning of the two tasks of answer extraction and question generation can lead to an optimization imbalance between the two subtasks during training.For this reason,this paper proposes an Interaction-Guided Joint Abstractive QAPs Generation Model(IGJA-QAP).Specifically,this paper designs a joint generation model with an answer-guided multiheaded gate mechanism that simultaneously models two subtasks in a unified manner and efficiently captures and enhances the information interactions between them,so that semantically matching question-answer pairs can be generated.In this paper,a comprehensive experimental analysis is conducted on three large-scale datasets,SQuAD,NewQA and CoQA.The proposed model outperforms the best methods by 3.0%,5.9%and 4.3%on average for the answer generation task and 1.5%,0.5%and 2.1%on average for the question generation task,respectively.The experimental results demonstrate that our model achieves state-of-the-art performance.
万方数据
维普
