Abstract: Aiming at  the problem that when the semantics of the input text were relatively complex, traditional end-to-end models generated questions with incomplete semantics,  we proposed a text encoder architecture based on semantic feature extraction. Firstly, we constructed bidirectional long short-term memory network to obtain  basic contextual information. Secondly,  self-attention mechanism was used to extract global features of semantics and bidirectional convolutional neural network model was used to extract local features of semantics respectively. Finally, we designed a hierarchical structure to merge the features and input their own information  to obtain the final text representation for question generation. Experimental results on the SQuAD dataset show that the question generation based on semantic feature extraction and hierarchical structure are significantly effective, and the results are better than the existing methods. Moreover, semantic feature extraction and hierarchical structure are improved  in each evaluation index of the task.

Key words: question generation, bidirectional long short-term memory network, convolutional neural network, self-attention mechanism, hierarchical structure