Research On The Conversion Method Between Chinese Text And SPARQL Based On Deep Learning Algorithm

Semantic search technology is a new trend of the Internet,which is also one of the hottest trends in the development of Internet technology.Semantic search is inseparable from the development of semantic web because it's based on the standards and technologies of semantic web and that enable computers to collect,understand the information on the Internet,and then provide users with semantic search.For such problems,many traditional question answering systems based on domain knowledge base are proposed,with the purpose of meeting users' needs to access to specific knowledge.These question answering system is usually based on the manual and statistical models obtained from the data observation.Recently,the Seq2 Seq deep learning architecture based on neural networks has proved a satisfactory result in converting source sequences into target sequences.Therefore,in this study,based on the deep learning architecture,proposes a new model to replace the manual and statistical models in the traditional question answering system,called SPARQL translation model based on Seq2 Seq.This model can directly transform Chinese natural language questions from users to a sequence of SPARQL tokens,which can reconstruct a SPARQL query.In other words,the model applies end-to-end method to transform natural language expression to the final query.Our study,based on the data of DBpedia,first created a Chinese domain knowledge base for this experiment by using knowledge extracting,including 36714 three tuples and 678 entities.And then we generated 29400 experimental data composed of Chinese questions and corresponding SPARQL statements by using 49 query templates which was constructed manually.In our experiment we used the Google NMT to build the deep learning architecture for this study,and tested above experimental data on Tensor Flow.Taking the Accuracy score as the evaluation criterion of the model,the experimental results of many different model parameters showed that the convergence speed of the two level bidirectional LSTM and the Luong attention mechanism was the fastest on the test data,and got the highest score of 74.8.After adjusting the query templates and the distribution of experimental data,at last the score was raised to 87.6.The above experiments results show that the SPARQL translation model based on Seq2 Seq is a new solution to realize the question answering system based on the domain knowledge base.