Siamese meta-learning network for social disputes based on multi-head attention.

Abstract

Few-shot learning has been widely used in scenarios where labeled data is scarce, where meta-learning based few-shot classification is widely used, such as the Siamese network. Although the Siamese network has achieved good results in some applications, there are still some problems: (1) When computing prototype vectors with external knowledge of class labels, it depends on the quality and correctness of class labels. (2) When processing data, the Siamese network is not sufficient to capture dependencies between long distance. (3) When the data is complex or the samples are unbalanced, the Siamese network does not achieve the best performance. Therefore, this article proposes a multi-head attention siamese meta-learning network (MASM). Specifically, this article uses synonym substitution to solve the problem that the computation of prototype vectors will be transitionally dependent on class label. In addition, we use the multi-head attention mechanism to capture long-distance dependence by exploiting its global perception capability, which further improves the model performance. We conducted experiments on four benchmark datasets, all of which achieved good performance, and also applied the model for the first time in the field of social disputes, and experimented on a homemade private dispute dataset, which also achieved good results.