Co-Attention Based Few-Shot Relation Classification Model with Dynamic Routing

Abstract

With the development of natural neural networks, supervised methods are usually confronted with the problem of lacking labeled data. Few-shot learning methods are now a mainstream research method that allows models to classify relation base on a small amount of data. Relation classification is a basic task in natural language processing and it is the most critical step in the construction of a knowledge graph. This paper focus on few-shot relation classification and we propose a co-attention based few-shot relation classification model with dynamic routing. This model is divided into three parts: encoder layer, aggregation layer and matching layer. The encoder layer is used to extract the important features from support set and query set and convert it into vectors. Aggregation layer is to aggregate the vectors of instances in the same class. The matching layer is to compute the score between the query instances which is extracted form encoder layer and the class vector which is output by aggregation layer. We apply this model on FewRel dataset and the experiment result shows that our method is better than other methods.