Semantic Parsing and Text Generation of Complex Questions Answering Based on Deep Learning and Knowledge Graph

The current semantic parsing method can accurately parse simple question, but its lack of ability to analyze complex questions. Especially, there are many complex problems in the medical, legal fields. Therefore, the semantic analysis method of the complex question and the method of generating the complex answer are particularly important. However, the current complex question answering technology has the problems of low efficiency of compound question parsing methods and loss of semantic information in complex answer generation. To solve this problem, this paper proposes an complex question parsing method based on improved Bi-LSTM and ccomplex answer generation method based on BERT-LSTM. Firstly, we define a complex question parsing model, in which different parsing methods and answer organization methods are formulated for different kind of complex questions. Then the improved Bi-LSTM model is used to analyze the complex question and decompose the original question into multiple sub-questions that answers in the knowledge graph, according to the complex question analysis model. Finally a BERT-LSTM model extract complex answer from sub-answers based on machine-reading comprehension method. In order to test the effect of this method, we make a Chinese complex question and answer corpus, and construct a Chinese complex question answering system. Experimental results show that the accuracy of this system is better than the others. The score of ROUGE-L evaluation increased by 9.3%.