Research And Application Of Few-shot Relation Extraction Based On Neural Networks

With the rapid development of big data technology,information extraction is of more and more great significance.One important reason is that it can mine accurate structured information from unstructured data with sparse information through neural network model.Entity relation extraction,which has attracted more and more attention from natural language processing researchers in recent years,is an essential part of information extraction.Relationship extraction is mainly to study semantic knowledge information in unstructured data.And then it uses the knowledge learned to extract entity relation in the vast amounts of unstructured data such as text information.In other words,it translates the unstructured data into structured relational data to support knowledge base,question answering system,information retrieval,etc.However,in many practical application scenarios,there is not enough data for models to train.Even for some fields with enough samples,there is also the problem that the cost of relationship annotation is too high.Therefore,the study of relationship extraction based on few-shot learning is of great significance.The research content of this thesis is as follows:(1)Quantitative analysis is made on the model complexity of word embeddings encoder using BERT and Glove to obtain the floating points operations of context-dependent pre-training model theoretically and experimentally.BERT is a context-dependent pretraining model while Glove is a static pre-training model.On this basis,the model of word embedding encoder using Glove is improved,and the trainable data-enhanced network layer and context-dependent sampling method are proposed to make the simple neural attentional meta-learner SNAIL as a sentence feature classifier lose a little accuracy when Glove is used as word feature encoder compared with BERT.The accuracy is 75.71 percents on Few Rel.But the forward propagation speed of the model is hignly improved.It shows that the application of small sample relation extraction to the actual system needs to solve the problem when there is no query sample type in the support set.The existing simple neural attentional element learner is improved to make it have a bidirectional structure,and the experiment shows that the structure proposed in this thesis can improve the accuracy of auxiliary annotation.(2)This thesis applies the model to the accurate knowledge graph construction system,analyzes and explains the main application scenarios of the system,and explains the function modules and design implementation of the system.The experimental results show that the system can assist manual construction of an accurate knowledge graph,the accuracy of which increases from 92.2% to 99.5%.In practical application,the knowledge graph obtained in this thesis is actually a relatively good annotation sample,but it has not been fully applied in this thesis.Future research can explore how to input the knowledge graph into the model to further improve the model capability.