Answer-Aware Multi-Hop Question Generation Based on Variant Convolutional Networks

The task of multi-hop question generation requires aggregating and reasoning about multiple dispersed evidence points from a text to generate complex questions related to the answer. Whereas most sentences in the text are redundant or contain irrelevant information, most current methods are oriented towards full-text encoding, which can lead to the problem of not being able to effectively identify and aggregate evidence points in different text passages. In this paper, we propose a multi-hop problem coding fusion network based on variant convolutional networks, feature enhancement of full-text coded information by gated convolutional networks to improve the effect of identifying evidence points, then using graph convolutional networks to identify and aggregate scattered evidence points, and finally using Maxout pointer decoding to generate multi-hop problems. In the comparative experiments on the dataset HotpotQA, the model achieves significant performance improvement in multiple metrics of multi-hop problem generation.