面向小样本关系抽取的自适应胶囊网络.

The few-shot relationship extraction task is a hot issue in natural language processing. It aims to train the relationship extraction model using low-cost label data. The widely used prototype network has some problems, such as inaccurate and incomplete expression of class prototypes. This paper proposed an Adaptive Capsule Network (ACNet) to solve this problem. ACNet generates a class prototype with the inductive capability of the capsule network. On this basis, the dynamic routing process is evaluated so that it can adjust network parameters adaptively to different samples. At the same time, a memory iteration mechanism is introduced in ACNet to help the model determine the class representation quickly. Experiments on a few-shot relational dataset FewRel show that ACNet can handle few-shot relational extraction tasks. [ABSTRACT FROM AUTHOR]