Research On Question Answering Based On Open Domain Knowledge Base

In recent years,a number of public knowledge bases covering a wide range of fields and large data scales have emerged,such as YAGO,Freebase,DBpedia,and Google Knowledge Graph,which has made the question answering based on the open domain knowledge base a very hot research direction.In this research direction,knowledge base question answering technology research combined with the deep learning method is becoming more and more popular,and shows a bright future in performance.Learning from the existing mainstream methods,the knowledge base question answering is regarded as the idea of multiple task modules on the assembly line.It is divided into three modules: subject detection,entity linking and relation detection,and they are solved one by one.Aiming at the performance impact of the upstream model's error propagation on the downstream model,the subject detection task is regarded as a special sequence labeling model.The advanced language model method is used to generate the word embedding by using the context string language model to make it have semantic contextualization ability and enhances model performance.In order to improve the recall rate of the subject entity,two candidate pool strategies are designed and the candidate subject entity set is pruned from the text surface level and the context semantic level to improve the quality of the entity linking result.To make up for the shortcomings of the pruning step,relation detection model is combined to prune candidate entity sets to further improve the overall accuracy of question answering.In order to capture the context and deeper semantic features in relation detection model,a multi-layer residual BiLSTM encoder is designed.Then the question is matched with the multi-granularity relations in four modes to achieve a more comprehensive matching effect from local information and global information.The experimental part is compared with the existing representative methods.Freebase was selected as knowledge base and experiments were conducted on the popular dataset SimpleQuestions.The results demonstrate the strong competition between the proposed method and the most advanced methods in the performance of each task module.