Research And Implementation Of Entity Relation Extraction Algorithm Based On Deep Learning

The arrival of the era of big data not only brings opportunities and changes to the development of all walks of life,but also contains a large amount of valuable knowledge wealth.In this context,knowledge map emerges.Knowledge map is a semantic web that reveals the relationship between entities,which can provide application basis for AI scenarios such as intelligent search,machine question answering,personalized recommendation,etc.In order to meet the demand of constructing large-scale knowledge graph,it is an important and urgent task to extract effective structured information from unstructured text accurately and effectively when facing massive data.In recent years,the information extraction task has received the attention of the field researchers,and the entity and relation extraction task is the core task.There is a common complex situation of overlapping triples in entity relationship extraction.Solving the problem of triples extraction of various types of relationship overlap and improving the extraction performance are the focus of current research.In addition,it is costly to manually label data by entity relationship extraction.How to enlarge the value of data is an important research point.The specific research contents of this thesis are as follows:First,the existing model ignores the impact of the complete semantic information of the relational label on the triplet extraction model.This thesis proposes a three-dimensional three-tuple extraction model which focuses on the semantic information of relational tags,and obtains the semantic information of relational tags from both static and dynamic aspects for the first time.The static aspect is the name of the relationship label and the description of the relationship label.The dynamic aspect is the prior semantic knowledge and contextual semantic information that the model automatically learns from the sentences related to the relationship.In addition,in order to enable the model to think about problems at the triple level and make better use of the semantic information of relational labels,this thesis proposes a decoder at the triple level.Finally,the loss function is modified to alleviate the category imbalance problem.The experimental results show that the model has achieved the best results in this field on two common data sets in this field,WebNLG and NYT.The F1 values in the two public data sets are as high as 93.5%and 94.4%,and the effect is outstanding in various complex situations such as overlapping triples.Second,the existing entity and relation extraction field has the problem of insufficient labeled samples and poor quality.We proposes five data augmentation methods for this task:add,replace,swap,delete and Relation-Type.These methods consider the unique characteristics of the field that both the entity type information and the entity shallow reference information will have a great impact on the extraction results,and achieve great results in two public data sets and real application data sets.Among them,the Relation-Type method has a baseline model effect gain of 0.5%,0.9%and 2.1%on the three data sets,which is more prominent when the amount of data is small.Third,faced with massive data,the original movie information query system uses keyword matching technology to query is not only a waste of user time but also not intelligent.In this thesis,a visual movie information query prototype system based on entity relation extraction is constructed.The threedimensional triplet extraction model of concern semantic information is used to construct knowledge graph,and the Relation-Type data enhancement method is used to enrich the training data.By using this method,we can mine the entities and relations in unstructured text,obtain more complete knowledge structure,and provide more comprehensive and accurate search results for users.