Research For Cross-domain Few-shot Entity Relation Classification

Entity relationship classification is one of the important technologies in information extraction,and plays an important role in scenarios such as intelligent question answering,knowledge graph generation,and information retrieval.Existing entityrelationship classification models are limited by large-scale labeled data,and it is difficult to improve model performance in fields such as medical care and finance,where labeled corpus is scarce,difficult to obtain,and expensive to label.And when the two kinds of data for model training and testing come from different fields,due to the deviation of feature distribution in different fields,the model will perform well in the field of training data but poorly in the field of test data.In response to the above problems,this paper conducts research on the crossdomain few-shot entity relationship classification method,focusing on how to improve the generalization ability of the classification model when the sample size of the target domain is limited,thereby improving the classification performance of the model.The main contributions of this paper are as follows:(1)The main body of the cross-domain small-sample entity relation classification model adopted in this paper is divided into two parts: encoder and classifier.Among them,the encoder adopts the BERT pre-training model.The input is augmented by inserting labels before and after entities in the input sentence.The vector corresponding to the entity start position label in the hidden layer of the BERT output is spliced as the vector representation of the entity pair relationship in the input sentence.The classifier adopts two metric methods,namely dot product method and prototype network.The model based on few-shot learning can quickly acquire the ability to solve problems through a small number of training samples,which relieves the model’s dependence on labeled data.(2)A method to fuse multi-dimensional attention mechanisms is proposed.The multi-dimensional attention mechanism and the measurement method are integrated to improve the classification ability of the model.Multi-dimensional attention mainly includes instance-level attention and feature-level attention.Instance-level attention is used to select instances that are beneficial for classification,avoiding the influence of noisy data.Feature-level attention is used to highlight the easily distinguishable feature dimensions in the feature space and alleviate the feature sparsity problem.(3)A method based on learning feature transformation layers is proposed.A feature transformation layer is integrated on the encoder to learn various distributions of features,thereby improving the generalization ability of the model.Due to different metric functions and different domain conditions,the hyperparameters of the feature conversion layer are difficult to find effective general parameters by manual tuning.Therefore,a self-learning method is used to optimize the hyperparameters of the feature conversion layer.