Few-Shot Relation Classification Research Based On Prototypical Network And Causal Intervention

In recent years,deep learning represented by convolutional neural network has made great progress.Meanwhile,the disadvantage of deep learning relying on massive annotated data has also become prominent.When annotated data acquisition is difficult and cost is high,deep learning is even more difficult to give full play to its due effectiveness.How to use limited or small amount of annotation data to obtain a prediction model with good generalization performance has gradually become one of the research hotspots in the field of artificial intelligence.Few-Shot learning brings hope for solving the above problems,and is widely used in machine learning tasks in Few-Shot scenarios.In this paper,in order to solve the relationship between the Few-Shot situation classification task as a starting point,to build a high performance relationship classification prediction model as the goal,by means of causal intervention,on how to realize the relationship between the Few-Shot scenarios classification prediction,how to weaken the confounding factors effect the performance of the forecasting model,a theoretical and experimental research.First,in view of the negative impact of confounding factors on the performance of the prediction model,a method of confounding weakening based on causal intervention was proposed.Causal intervention using the method based on the back door adjustment method to realize a priori knowledge of the training model of layered,by using the prior knowledge points number as network parameter in model training way to determine the points and the number of layers of optimal value,and the introduction of BN layer to eliminate dispersion problem caused by the back door adjustment gradient,so as to achieve to confounding factor model to predict the purpose of the performance impact.Second,in order to enhance the semantic representation and feature extraction capability of the model,RoBERTa was used to replace the CNN part of the prototype network as a feature extractor,making the model more suitable for relational classification task scenarios,thus effectively improving the prediction performance of the model.Finally,on the basis of the above studies,a Few-Shot relational classification model RBERTI-Proto based on prototype network and causal intervention was constructed,and the model performance comparison and ablation experiment were carried out on the data set Few Rel.In the 5-way 5-shot scenario,the ACC value of RBERTI-Proto model reached 93.38%,which verified the performance superiority of RBERTI-Proto model compared with other models.And the effectiveness of the weakening method of confounder based on causal intervention and RoBERTa as model feature extractor for improving performance of RBERTI-Proto model.