Comprehensive Study on Zero-Shot Text Classification Using Category Mapping

Kai Zhang, Qiuxia Zhang, Chung-Che Wang, Jyh-Shing Roger Jang

Abstract: Existing zero-shot text classification methods based on large pre-trained models with added prompts exhibit strong representational capacity and scalability but have relatively poor commercial applicability. Approaches that fine-tune smaller models using label mappings and existing datasets for zero-shot classification are simpler but suffer from weaker generalization capabilities. This paper employs three strategies to improve the accuracy and generalization of pre-trained models in zero-shot text classification tasks: 1) Utilizing a pre-trained model that transforms inputs into a standardized multiple-choice format. 2) Constructing a text classification training set using Wikipedia text data to fine-tune the pre-trained model; 3) Proposing a zero-shot category mapping method based on GloVe text similarity, using Wikipedia categories as substitutes for text labels. Without fine-tuning on the target labels, this method achieves performance comparable to the best models fine-tuned with target labels.