Research On Knowledge Base Question Answer Based On Joint Learning

With the advent of the era of massive data,people's requirements for searching complex and complex data are getting higher and higher on the Internet.Traditional search engine systems based on keyword matching have been increasingly unable to meet people's requirements.Therefore,how to provide users with a the accurate and efficient automatic question answering system has become an important research project in the industry.In recent years,with the vigorous development of the knowledge graph,Whether in the field of English or Chinese,the question-and-answer system based on knowledge graph increasingly reflects its importance in the field of automatic question and answer.Unlike traditional search engines,the question and answer system based on knowledge graph no longer returns a series of matching documents for users,instead,it uses its intelligent and accurate recommendations to push accurate answers to users,which significantly improves the user experience,improves the stickiness of users in using search engine products,and further creates economic value for enterprises.Therefore,more and more researchers are investing in the research of knowledge answering system.Knowledge base based question answering system(KBQA)is mainly divided into based on method of semantic parsing and based on method of distributed semantic representation.Among them,the method of semantic parsing is affected by the semantic gap,which makes the accuracy and recall of question and answer low.With the deepening of research,the method of distributed semantic representation gradually surpassed the method of semantic parsing.But at this stage,the performance of KBQA based on distributed semantic representation still needs to be improved.In previous research,it was restricted by the accuracy of semantic representation and the lack of connection between entity and relation in research of KBQA.This article focuses on simple question answering over knowledge bases task,and conducts research from multiple aspects such as data annotation,multi-dimensional representation of questions,and joint model construction.The main research content includes the construction of entity recognition and relation prediction models for joint learning,and entity linking based on inverted index and answer search based on path search.Aiming at the problems of neglecting the correspondence between entity and relation caused by mutually independent entity detection and relation prediction and training efficiency in simple question answering over knowledge bases,a knowledge base question and answer model of joint entity recognition and relation prediction is proposed,using CNN-BiLSTM-CRF to identify entity in question and combining text features extracted by CNN-BiLSTM with label embedding features of the question for relation prediction,As a result,the F1 value of entity recognition is improved by 1.1% and the accuracy of relationship prediction is increased by 1.6% compared with the independent training method?An n-grams model is combined with TF-IDF to build an entity alias index for entity linking,which speeds up the training efficiency of model and reduce the difficulty of training complex networks,then build a path search index to find the answer.The experimental results show that the proposed method can effectively improve the accuracy of question and answer and the efficiency of training.The recall rate of entity links has also been significantly improved,and the final accuracy rate of the Q & A system on the test set has increased by 2.6%.