An External Knowledge-Based Text Classification Method for Few-Shot Text

The proliferation of Internet-generated text, particularly in domains such as news and sentiment classification, has led to increased reliance on short texts characterized by brevity and high generalizability. While fine-tuning pre-trained models has shown success, these approaches typically require substantial labeled data and are unsuitable for few-shot settings. To address this, we propose a novel few-shot classification method enhanced with external knowledge. Our approach constructs a tailored prompt template integrated with input texts, reformulating classification as a cloze-style task. Additionally, external knowledge is leveraged to expand label words, with predictions mapped to original labels via a scoring mechanism. Experiments on sampled subsets of three datasets—THUC-News, Toutiao, and Chinese News Titles—demonstrate that our method significantly outperforms baseline models under 1-shot, 5-shot, and 10-shot conditions. Notably, in 1-shot settings, accuracy improvements average 10.1%, 8.4%, 4.8%, and 5.6% across the respective datasets.