Research On And Application Of Relation Extraction Algorithms Based On Deep Learning

Relation extraction is a way of extracting semantic relations between entities from text.As an important part of information extraction,relation extraction provides convenience for constructing structured data that can be recognized by machines.Relation extraction algorithms based on deep learning refer to the algorithms that use deep learning method to solve the problem of relation extraction.The supervised relation extraction algorithms make full use of manually labeled information to train the model.It is the mainstream method of relation extraction at present,but the cost of labeling data is too high.The distantly supervised relation extraction algorithm automatically builds a large amount of training data by aligning the knowledge base with unstructured text,reducing the model's dependence on artificially labeled data.In recent years,countless experts and scholars have used deep learning methods to conduct deep research in the fields of supervised relation extraction and distantly supervised relation extraction,and have proposed some effective relation extraction algorithms,but these algorithms basically face the following problems: 1)The supervised relation extraction algorithms mostly choose to the features of entities from a single sentence,the one-sided features learned in this way will affect the accuracy of relation extraction.2)Because of the data from distantly supervised relation extraction is constructed by aligning the knowledge base with the corpus,there is a lot of noise in the data,which will greatly interfere with the results of relation extraction.3)Whether it is supervised relation extraction or distantly supervised relation extraction,the accuracy of extracting entity and sentence features needs to be improved.The topic of the thesis comes from the general project of the National Natural Science Foundation of China.In order to overcome the shortcomings of the existing relation extraction algorithms,the author has conducted deep research on the relation extraction algorithms at home and abroad,and respectively proposed an efficient supervised relation extraction algorithm and an efficient distantly supervised relation extraction algorithm,the main work of this paper is as follows.1)Aiming at the problem that existing supervised relation extraction algorithms ignore external knowledge,an entity semantic relationship graph that can represent the connection relationship between different entities is established,which can store the knowledge of the corpus,and for this purpose,a graph neural network based on semantic similarity is designed to effectively filter knowledge to enrich the understanding of entities.On this basis,A Novel Entity Relation Graph for Supervised Relation Extraction(ERGSRE)algorithm is proposed.2)In view of the defects such as incomplete feature extraction,low accuracy of noise identification and difficult to make full use of noise data,etc.,a graph neural network is used to enrich sentence structure information,and a noise recognition method based on margin metric is designed to identify the noise in the data.In addition,we use the incorrectly labeled samples with correctly labeled samples to train the algorithm to improve the accuracy of relation extraction.Based on this,A Margin Metrics Denoising Method for Distantly Supervised Relation Extraction algorithm(MMDSRE)is proposed.3)The ERGSRE algorithm and the MMDSRE algorithm proposed in this article are simulated on real datasets.The test results show that the performance of the ERGSRE algorithm and the MMDSRE algorithm is higher than the current state-of-the-art relation extraction algorithms.The two algorithms proposed in this paper are integrated into Open NRE,an open relation extraction toolkit,which can conveniently extract the relationship of entity pairs with one click,and meanwhile improve the performance of Open NRE.Although the two algorithms proposed in this article can respectively improve the performance of supervised relation extraction and distantly supervised relation extraction algorithms,this deep learning-based approach relies too much on training data,and it is easy to forget the old task after learning a new task,so it is difficult to meet the actual demand.While the lifelong learning mechanism can effectively solve this phenomenon.Therefore,how to introduce the lifelong learning mechanism to improve the anti-forgetting ability of the model is the future research direction of the author.