Research On Few-Shot Relation Extraction Methods

There is a large amount of text data on the Internet.Information extraction can extract effective information from text data for us,relation extraction is one of the important steps in information extraction.The existing relation extraction method relies on a large amount of labeled corpus,which requires a lot of manpower and material resources,and the remote supervision method has the problems of noise and long tail distribution.Thus,few-shot relation extraction problem is raised.In few-shot relation extraction task,the model needs to use very little data for relationship extraction.This paper focuses on the problem of fewshot relation extraction and conducts research from the following aspects:1.Research on few-shot relation extraction method based on class feature representation.The few-shot relation extraction method based on category feature representation has faster prediction speed and lower hardware resource occupation,and has a broad application scenario.In this paper,we use pretrained model BERT as the encoder of the relationship example,and study two kinds of extraction feature representation methods,which are based on graph attention network and dynamic routing mechanism respectively.2.Research on the method of few-shot relation extraction based on matching.For the instance feature representation process and the feature loss problem existing in the class feature representation method and class feature representation method,the query instance is matched with each instance in the support set to obtain the similarity between the query instance and the support set.The model BERT+ESIM is proposed,and it is studied from the aspects of encoding layer,sentence matching layer,loss function and adversarial training.3.Further research on matching-based few-shot relation extraction method,proposed a Piecewise Attention Matching Network,optimized the model structure for relationship extraction examples,used BERT-Sequence in the encoding layer and segmented according to the location of the entity,and verified the segment length for different domains for the distribution difference problem,the dynamic piece length is used for domain adaptation.At the sentence matching layer,the segmental attention mechanism is used to calculate the sentence similarity,which can more accurately calculate the similarity between relation instances.The Piecewise Attention Matching Network achieved good result in the Few Rel1.0 evaluation,and achieved the best result in the current leaderboard of Few Rel2.0 evaluation.