Research On Chinese Semantic Parsing Based On LSTM Neural Network

Semantic parsing aims at mapping natural language to a complete formal meaning representation which can be automatically understood and executed by computers.It is a core research field of natural language processing and key technology to implement various intelligent systems,such as question answering systems over knowledge bases(KBQA)and robot controlling.The study of semantic parsing of English has experienced a long history,and a variety of classical methods have been proposed.Particularly,with the development of neural network,approach based on LSTM neural network has been a new research trend.For Chinese is more complex compared to English and the public Chinese data set about semantic parsing is not sufficient,there is little research on semantic parsing of Chinese.In this thesis,by using question answering over knowledge bases as potential application scenario,the existing semantic parsing approach based on LSTM Encoder-Decoder model is improved and applied to Chinese semantic parsing task.The main work in this thesis include:(1)Syntactic information of natural language is introduced into the existing semantic parsing based on LSTM Encoder-Decoder according to the perplexity and diversity of Chinese.In addition,tree structured LSTM neural network is integrated with the Encoder.It encodes natural language according to the syntax tree which could reduce the impact of the diversity of Chinese.(2)The process of Decoder is improved which makes Decoder be able to utilize knowledge base to improve the accuracy.Furthermore,the result of semantic parsing can be used in the scenario of question answering directly.(3)A data set for semantic parsing of Chinese which makes up for the lack of public data set for Chinese semantic parsing is constructed.The data set consists of Chinese questions and corresponding meaning representations which are descripted with lambda calculus and annotated manually by people.(4)The improved semantic parsing approach is implemented based on LSTM Encoder-Decoder model,and the approach proposed in this thesis is experimented on the above mentioned data set.Experiment results show that the improved semantic parsing approach can achieve better accuracy.