Research And Implementation Of Simple Question Answering Algorithm Based On Knowledge Graph

The knowledge graph is extracted from a large number of facts.It is roughly divided into two categories.One is the entity extracted from a knowledge base such as Wikipedia and the relationship between them.The other is a triple with entity relationships extracted from a large number of pages.The content is more rich,which also makes it more noisy.With the emergence of these knowledge graphs,people began to use effective methods to obtain a lot of valuable content in the knowledge graph.For this reason,many query languages have been proposed.However,it is not easy to understand the grammar in these query languages.Therefore,many question answering algorithms based on knowledge graph are proposed.In the field of artificial intelligence,Q & A based on knowledge graph uses knowledge graph as a key element to answer human questions.This is an effective way to solve the problem,and has promoted the development of artificial intelligence.There are two main research lines for Q & A tasks based on knowledge graphs.The first is a semantic parsing method,which maps the problem to its logical form and then transforms it into a structured query.The second is a neural network-based method,which also has two branches.Roughly pipeline framework and end-to-end framework.This paper focuses on the pipeline framework for single relation question.Unlike end-to-end frameworks,each of its modules is tightly connected,and many deep learning frameworks are applied in the middle.The framework needs to get the best results for each module to ensure that the final conclusion is better than other algorithms.The overall structure of the pipeline framework is as follows:(1)the entity detection module used to identify the entities mentioned in the question;(2)the entity linking module used to generate entity candidates in the knowledge map pointed to by the question;(3)relation detection module for semantic similarity between relation candidates.In the entity detection module,we give a problem,the goal of entity detection is to identify consecutive markers in the mentioned range that relate to the subject entity in the question.This article use the BiLSTM-CRF model to label entities,which is most commonly used.And replace the entities with special symbols to convert the probleminto problem pattern.In the entity linking module,this paper adds the multi-label classification model,which based on the traditional string matching algorithm,and obtains a new algorithm,which is called entity linking algorithm based on problem pattern classification.For the relation detection module,this paper proposes a new model,which is called attention-based question pattern relation matching model.The model involves two levels of relation encoding,namely,word level and relation level.In this paper,a verification experiment is carried out on the simple questions data set.The experimental results(80.80%)show that the pipeline framework implemented in the paper is feasible and effective compared with the previous methods.For the entity detection module,four models are used.And the results(97.41%)show that BiGRU-CRF model can achieve higher accuracy.For the entity linking module,the entity linking algorithm based on problem pattern classification is also effective,much better than the traditional string matching method,and has universality.For the entity detection module,attention-based question pattern relation matching model has the same high accuracy.