Topic Classification of Small Sample News Based on Prompt Engineering

This paper proposes a classification framework that integrates dynamic prompt generation and knowledge distillation to address the issue of performance degradation in news topic classification in small sample scenarios. Research background shows that over 65% of vertical news datasets have a sample size of less than 500, and traditional methods generally have F1 scores below 0.6 in small sample sizes. This framework consists of three core modules: dynamic prompt generation, extracting news features through GRU and attention mechanisms, and generating adaptive templates by combining topic semantic libraries; Constructing a "teacher student" model through knowledge distillation, using 20000 unlabeled data to transfer ERNIE 3.0 knowledge; The classification decision module integrates features through a gating mechanism and sets dynamic thresholds based on sample distribution. The experiment was validated on a 10 topic news dataset (3200 articles), with an accuracy of 0.785 in the 5-shot scenario and 0.763 in Macro-F1, which were 9.2% and 7.4% better than the mixed prompt model, respectively; The ablation experiment proves that the knowledge distillation module contributes the most to performance improvement (accuracy reduction of 9.1%). Research has shown that this method can effectively utilize the contextual learning ability of prompt engineering and unlabeled data, providing a new solution for low resource news classification.