Abstract: To address the problems of multi-stage encoding of the encoder of the graph-to-sequence question generation model and the easy loss of rich sequence information and semantic structure information in the paragraphs during decoding, this paper designs a model based on multi-stage timing and semantic information enhancement(MS-SIE). The model first fuses the semantic information of the passages encoded at different stages of the encoder and inputs them to the recurrent neural network for encoding. Then, an iterative graph neural network is introduced in the decoding stage to combine the encoded paragraph information with the rich semantic structure information hidden in the previously generated text questions in the decoding stage. Finally, a recurrent neural network based on an attention mechanism is used to generate the questions. The results show that the model proposed significantly outperforms the existing sequence-to-sequence model and graph-to-sequence model in both automatic evaluation metrics and manual evaluation metrics.

Key words: question generation, multi-stage temporal fusion, semantic information enhancement, recurrent neural network, iterative graph neural network