Research On Entity Relation Classification In Few-shot Paradigm

Entity relation classification is an important task in natural language processing,aiming to identify the relations that exist between entity pairs in text,and is the basis for many tasks such as knowledge graph construction and automatic QA systems.Traditional rule-based meth-ods require domain experts to develop rules,which are labor-intensive and poorly generalized?supervised learning methods based on deep learning rely on a large amount of annotated data?distantly supervised methods can align knowledge from knowledge bases to text for automatic annotation,but they cannot solve the problem of long-tail distribution of samples and introduce sample noise.Therefore,it is of great research significance to realize entity relation classifica-tion in few-shot paradigm.The existing few-shot relation classification methods combine few-shot learning with rela-tion classification models,among which the prototypical network-based methods have achieved better results.Most of the current prototypical network-based approaches use convolutional neu-ral networks(CNNs)and static word embeddings to encode text,and the model performance is limited by the semantic modeling ability of word embeddings and insufficient modeling ability of entities on semantic information.In addition,the original prototypical network ignores the differences of the support set samples and cannot construct a prototype suitable for the current classification task for specific query samples.To address the above problems,we proposed a few-shot relation classification model based on entity information enhancement and selective attention.The model uses Ro BERTa to con-struct a text encoder that can adequately extract text contextual semantic features.To enhance the utilization of entity semantic information,we proposed entity information augmentation based on entity pairwise mutual attention,which fuses entity semantic information with overall contextual information to obtain better entity relationship features.In addition,we proposed a prototype construction method based on selective attention.In the process of constructing the prototype,selective attention assigns different weights to the support samples based on the features of the current query samples,so that a more accurate prototype can be constructed.Fi-nally,we added support set similarity loss to the loss function of the model to further improve the accuracy of relation classification.The results of the experiments on the Few Rel 1.0 and Few Rel 2.0 datasets,which are com-monly used for few-shot relation classification,demonstrate that the model we proposed effec-tively improves the accuracy of few-shot relation classification and has stronger domain migra-tion ability compared with the mainstream methods.We also tested the model on the traditional relation classification task datasets Sem Eval2010 Task 8 and NYT-10 to verify the applicabil-ity of the model.In addition,we designed several sets of ablation experiments to analyze the specific impact of each module in the model.Finally,we tested the accuracy and computa-tional efficiency of the model under different task conditions,and verifies the feasibility of the application of the model.