Research On Entity Linking And Relations Prediction For Knowledge Base Question Answering

With the development and maturity of knowledge graph technology,it has been applied in various fields as a structured knowledge base,especially in the field of question and answer based on knowledge base.Knowledge base question answering(KBQA)is to accept a natural language question and after semantic understanding combined with the knowledge base to return the answer,different from the traditional keyword-based retrieval question answering,knowledge base question answering directly gives accurate answers,does not need the user to search or reasoning twice,so that the answer acquisition way is more simple and efficient.At present,the accuracy of whole KBQA is not high,and the main reasons include two points: the spelling of entity name is not standard,the question context information is insufficient and the entity description information is missing in the knowledge base,which leads to a high error rate of entity linking;The complex semantics and structure of questions as well as the large number of relations lead to low accuracy of relation prediction.In view of the above problems,this paper has carried out in-depth research,and the specific work and results are as follows:(1)To solve the problems of non-standard spelling of Entity name,insufficient information of question context and missing information of entity description in knowledge base in entity linking,this paper proposes an Entity linking of multi-dimensional matching(MDM-EL).Firstly,the MDM-EL model carries out string dimension matching,which adjusts the entity candidate set to a reasonable size.Then the statistical dimension matching was carried out.In this dimension,the primary entity candidate set was sorted by entity salience.Finally,the entity attribute dimension matching was carried out,and the entity attribute information was used in this dimension to complete the correction of the ranking result,and the entity with the highest ranking was the result of the entity linking.The experimental results show that the accuracy of this model on the data set of Simple Questions reaches 83.43%,which is higher than the stateof-the-art entity linking method in recent years.(2)For the difficult semantic extraction of complex questions in relation prediction,In this paper,a self-attention based Hierarchical semantic extraction network(SA-Hse Net)is proposed.The model is implemented with the framework of encoder and decoder,which is a combination of information extraction and deep learning.In the model,the encoder is responsible for encoding the semantic features of the question and the decoder is responsible for decoding the relational semantics.In order to enhance the semantic extraction ability of the model,the encoder divides the semantic features of the questions into two levels: local shallow semantic and global abstract semantic.Meanwhile,in order to highlight the key semantic information of the two levels,the encoder uses the self-attention mechanism and the cross-attention mechanism to complete the encoding of the questions respectively.The experimental results show that the accuracy of the model on Simple Questions data set reaches 93.36%,and the model has excellent performance in time-consuming aspect.(3)To solve the problem of the large number of original relation sets in relation prediction,this paper proposes a relation constraint subgraph method.In this method,the effective relations is extracted from the original relations set by the relation constraint subgraph,forming the effective relationship set,and then the relation is selected from the effectives relations set.Experimental results show that this method can improve the accuracy of SA-Hse Net model by about 3%.The MDM-EL model proposed in this paper reduces the error rate of entity linking and ensures the correctness of the answer topic.Secondly,the SA-Hse Net model and the relation constraint subgraph method improve the accuracy of relation prediction and ensure the correctness of answer reasoning.Meanwhile,the SA-Hse Net model has excellent performance in time-consuming aspect and ensures the real-time performance of KBQA.Applying the above models and methods to specific KBQA system can provide users with more accurate answers.