Research On COVID-19 Few Shot Relation Extraction Based On Deep Learning

The novel coronavirus has swept the world,causing severe loss of life and property to people in all countries,and the virus is still evolving.The researchers used relation extraction techniques to extract knowledge triples from COVID-19 texts to assist epidemic prevention and control through text analysis or knowledge mapping.Mainstream deep learning relationship extraction methods rely on large-scale data sets,while large annotated corpora of COVID-19 texts are lacking and difficult to construct.The rise of few-shot learning provides new methods and ideas for relationship extraction under low resources.It is of both theoretical significance and practical value to study the problem of few-shot relation extraction of novel coronavirus texts.This paper studies from the following three aspects:(1)A multi-feature fusion few-shot relation extraction method based on semantic enhancement is studied.In order to solve the problems of insufficient semantic learning,inadequate text encoding expression and poor accuracy of deep learning model in low-resource scenarios,this paper uses piece-based convolutional networks to integrate information such as location,part of speech and syntactic dependency as text representation supplement,and captures more semantics by means of multi-feature learning.The fine-grained semantic information is extracted from Wikipedia,converted into fine-grained semantic embedding and aggregated into word embedding,and feature extraction is carried out by piecewise maximum pooling to further enrich the semantic expression ability of word embedding.Few Rel data sets are used for experimental verification,and the results show that the proposed method can achieve better prediction with limited learning.(2)A small novel coronavirus text relation extraction dataset is constructed.In view of the lack of novel coronavirus corpus,a large number of unstructured texts were collected and sorted out,text entity types and relational categories were defined,and data annotation and corpus analysis were completed in a standard format.(3)A few-shot relation extraction model based on entity fine-tuning of Prompt framework was established.In order to solve the problem that the cost of fine-tuning the downstream task of the deep learning pretraining-finetuning model is too large,Prompt framework was applied after analyzing the data characteristics,and a prompt template was constructed.To solve the problem of multi-entity text in data set,the ambiguity of multi-entity text is reduced by using type entity labeling and multiple rating system.Experiments on public data sets and self-built COVID-19 data sets show that the established model achieves better long-term relation prediction and significantly improves its accuracy.