Multitask-Based Cluster Transmission for Few-Shot Text Classification

Kaifang Dong, Fuyong Xu, Baoxing Jiang, Hongye Li, Peiyu Liu

Abstract: Few-shot text classification aims to perform class prediction by learning from a few examples on labels. Prototypical Network (ProtoNet) is often used to solve the few-shot problem, devoted to constructing a metric space between classes and samples. However, the ProtoNet-based works of building meta-learners have inadequately developed the potential of metric space for the discriminative representation of text, which lead to the deficiency of classification tasks. To improve the above problem, we propose a smoothing strategy combining averaging and prototyping based on ProtoNet. Specifically, we generate a new cluster by cluster transmission algorithm and combine it with a label vector of the pooling function to enrich the distinguishability representation. The proposed algorithm makes the features of the samples more compact and improves the learning efficiency. Then, we use the representation of basic classification features as an auxiliary task to further enhance the diversity of spatial vectors and alleviate the over-fitting problem. Experiments show that our approach further improves the performance of the few-shot text classification task.

External IDs:dblp:conf/ksem/DongXJLL23