Research And Application Of Entity Relationship Extraction Method Based On Deep Learning

At present,with people's desire for quick knowledge acquisition and the rapid development of natural language processing,entity relationship extraction technology has become an enduring and important research direction in this field.It has important contributions to the construction of domain knowledge graphs or various natural language processing applications.Although the relationship extraction technology has developed various theories,the accuracy of automatic extraction cannot achieve better practicability.Methods based on deep learning,such as traditional pipeline methods,have error propagation and information redundancy,and joint extraction methods have the problems of entity overlap and relationship category label imbalance,which will eventually lead to low overall accuracy of the relationship extraction model.Therefore,this article is based on the deep learning united entity relationship extraction method,in view of the above-mentioned problems,conducted theoretical research and experimental analysis,and proposed some improvements and innovations to the entity relationship extraction algorithm model:(1)Aiming at the difficulty of the entity relationship extraction algorithm to accurately extract the overlapping entity relationships in a single sentence and the multiple triples in the office,in-depth study of the Bi-LSTM entity relationship extraction framework,and proposes an ordered structure coded pointer network decoding method Entity relationship extraction model.The improved joint entity relationship extraction model in this paper adds generation of anti-disturbance items in the input layer,uses an improved ordered long and short-term memory network in the encoding layer,and introduces a pointer network in the decoding layer,the decoding layer adopts an extraction strategy of first identifying the head entity,and then jointly identifying the tail entity and relationship.Experiments show that the improved network has a better overall accuracy in extracting sentence overlapping entity relationships,and also has a good improvement in the extraction of sentences containing different numbers of triples.This shows that the improved network in this paper has better performance than other networks.Which shows that the improved network compare other networks have better extraction effects.(2)Aiming at the problem of the pointer network exacerbating the imbalance of relationship category labels,based on the encoding and decoding network,an entity relationship extraction method based on the enhancement of fine-grained semantic information is proposed.In the process of training relationship samples,traditional models tend to train relationships with a large number of samples,which will cause the model accuracy to be falsely high,and it is not very friendly to train and extract relationships with a small number of samples.However,this paper proposes a method of constructing a fine-grained relation category dictionary,using the attention mechanism to give higher weight to relations with fewer examples,thereby reducing the problem of sample imbalance in the process of entity relation extraction.Secondly,the context information and entity type information are introduced to construct a fine-grained semantic information enhancement model to help improve the model's ability to extract entities and relationships.Finally,after experiments show that this model has significantly improved on the problem of label samples in relationship categories,compared to other models increased by 4.5% in the NYT standard data set,the overall accuracy of the WEBNLG standard data set increased by 6.8%.(3)For the current question and answer system,the answer can only be obtained based on the exact match of the question and answer pair,and the user's intention cannot be well understood.This paper uses the improved and proposed relationship extraction algorithm model combined with the related technology of the question answering system to construct an improved question answering system to effectively identify the user's intention and obtain the answer that the user needs.