Research On Key Techniques Of Knowledge Base Question Answering Based On Multi-granularity Semantic Matching

In the information age,with the rapid development of the Internet and the explosive growth of knowledge,people's demand for fast and accurate information has prompted the birth of automatic question answering technology.The knowledge base question answering(KBQA)is one of the most important branches,it can automatically answer the natural language questions containing facts in knowledge base,which is more accurate,fast and concise.Recent years,the method of deep learning has also been widely used in KBQA.However,due to the variety of questions asked by users in the real world,and the complexity and difficulty of natural language,the existing KBQA technology still has some shortcomings,such as ignoring the original information,the difficulty of distinguishing between entities with the same name and so on.The key to deal with KBQA problem lies in two aspects: entity detection/linking and relation detection.Relation detection is one of the most important steps in knowledge base question answering.To enhance the effect of relation detection and retain more comprehensive primitive information,an improved model based on multi-granularity semantic matching is proposed in this paper,which uses long-short term memory network and convolutional neural network.This model can take advantage of the advantages of LSTM and CNN to obtain semantic level and word level matching information respectively.The left side of the model is a deep two-way long-short term memory network(Bi-LSTMs)for semantic modeling based on different relationship granularity and hierarchy,which first extracts the relationship information from three levels :relation-level and word-level as well as the topic entity type related to the relation.Secondly,the deep bi-directional long-term memory network(BiLSTMs)is used to learn different levels of problem representation.Thirdly,an attention mechanism is used to track both entities and relationships at the same time.Finally,a residual learning method is also used to complete the hierarchical matching of problems and relationships.On the right side of the model,a convolutional neural network based on three-dimensional convolution is used to model from a literal point of view.To test the effect of this model,a simple KBQA system is built.Experimental results show that this method improves the accuracy of relationship detection,and it helps the KBQA system in this paper to perform better in both single relation(Simple Questions)and multiple relation(Web QSP)QA benchmark tests.