Research And Implementation On Knowledge Base Question Answering Oriented Entity Linking Technology

The development of the Internet has brought about an explosive growth of information.With the emergence of large knowledge bases,Knowledge Base Question Answering System has emerged at the historic moment.The Knowledge Base Question Answering System accepts natu-ral language questions,understands the intention of the questions,and que-ries the knowledge base to output answers that match the questions,con-veniently and effectively meeting people's needs.Entity linking is the first step in a knowledge base question answering system.It identifies the entity mention in the question and links it to a specific entity in the knowledge base.At present,the entity linking mostly adopts the pipeline mode,which is divided into two stages:mention detection and entity disambiguation.Because these two stages are separately modeled and trained,there are problems of error diffusion and localization of supervision information.The main work of this paper includes:First,the paper analyzes the data set of the Knowledge Base Question Answering System released by NPLCC in 2016,revises the three catego-ries of questions in the data:1)the mismatch of questions and their answers;2)the triples corresponding to the question and the answer not included in the knowledge base;3)the wrong questions.In addition,the mention in the question is labeled backward.On the base of these,an entity linking data set NLPCC-H is formed.Secondly,a new entity linking combination model is proposed.In the stage of mention detection,the sequential labeling model of the combina-tion of bidirectional short-term memory and conditional random field(BiLSTM-CRF)is adopted,and beam search is introduced to obtain mul-tiple candidate mentions,so as to reserve more possibility for the subse-quent stage.At the same time,the re-ranking mechanism based on syntactic information is introduced.In the entity disambiguation stage,the semantic similarity model is adopted,and the model samples according to the clus-tering results to alleviate the problem of imbalance between positive and negative samples.An attention mechanism is introduced to obtain better candidate pair vector representation.The top1 accuracy of the model is 84.12%,which achieves the best performance of the existing models.Finally,the entity linking model proposed in this paper is applied to a KBQA system,which is designed and implemented in combination with other existing technologies.The system shows good performance in prac-tical applications.