CLZT: A Contrastive Learning Based Framework for Zero-Shot Text Classification

Zero-shot text classification aims to predict classes which never been seen in training stage. The lack of annotated data and huge semantic gap between seen and unseen classes make this task extremely hard. Most of existing methods employ binary classifier-based framework, and regard it as a relatedness (yes/no) prediction problem between instances and every candidate class. However, these methods only consider the similarities between one instance and one class at a time, and ignore semantic relations between candidate classes. To alleviate this problem, we propose a novel Contrastive Learning based Zero-shot Text classification framework (CLZT). With the contrastive optimized objects, we can capture the semantic relations between classes that need to be predicted and build more discriminative embeddings. Main experiment shows that our method achieves the best overall f1 score compared with baselines in three different datasets.