Algorithms Of Few Shot Relation Extraction Based BERT In Chinese Domain

As a sub task in the field of IE(Information Extraction),RE(relation extraction)aims to obtain the semantic relations between entities in unstructured text information.Relational extraction method which based Deep Learning can automatically learn semantic features and relational patterns from massive data.However,the model relies heavily on large-scale training data,and it requires a lot of manpower and time to produce high-quality Chinese domain relational extraction dataset.The few shot learning method improves the problem of few training data.By combining the relational extraction model with the few shot learning method,the performance of relational extraction model in the situation of lack of data can be significantly improved.However,in few shot learning,when using few training samples to fine tune the model,it will inevitably cause the problems of abnormal convergence and low accuracy.Therefore,how to improve the accuracy of few shot relationship extraction model has become an urgent problem in the task of few shot relationship extraction.Based on the theory of few shot relation extraction in Chinese domain,this thesis deeply explores the remote supervision method,semantic fusion method and comparative learning method,and uses these methods to improve the few shot relation extraction method.The main research contents and results are as follows:(1)To remedy the flaws that lack of relational extraction data sets in Chinese domain,this thesis proposed an improved remote supervision method named BERT-D which expands the number of existing datasets and types of relationships,the method also weakens the dependence of relational extraction model on annotated data.It also Effectively improves the ability of the model solving the problem of few shot relational extraction.The experimental results show that the BERT-D method performs better than the Distant Supervision method.It not noly improves the extraction accuracy of few-shot datasets but also expands a large number of high-quality data samples.(2)To remedy the flaws that undetected error and false retrieval in relation extraction caused by new words which not in vocabular.This thesis propose a few-shot relation extraction algorithm based on local attention mechanism.The few-shot relation extraction algorithm changes the input mode of entities and special words in the mainstream model,and constructs new word semantic fusion algorithm based on local attention and multi-scale relation extraction algorithm of special words.This method improves the samples' information utilization degree in few shot relation extraction model.Experimental results show that the proposed word vector fusion method based on local attention mechanism significantly improves the ability of the model extracting new words and pronoun relations,it's also improves the accuracy of the whole relation extraction task.(3)To remedy the flaws that model abnormal convergence and model performance affected by outliers in Prototypical Networks,a few-shot relation extraction method ConBERT based on comparative learning is proposed.ConBERT introduces the learning method of contrast learning on the basis of Prototypical Networks,which greatly improves the under fitting problem of Prototypical Networks relation extraction algorithm and increases the dimension of type expression.The experimental results show that the proposed ConBERT method based on contrastive learning improves the accuracy of few shot relation extraction in few shot learning tasks,and improves the ability of relation extraction model to draw inferences from one instance.