A zero-shot prompt learning approach on fine-grained text classification

Fine-grained classification of social media texts in the military domain is crucial for public opinion analysis yet remains challenging. This is primarily due to the scarcity of high-quality annotated data, the need for deep domain knowledge, and the complexity of processing long-context documents. To solve these problems, this paper proposes a novel framework that combines text screening with prompt learning. Specifically, we first introduce a method that fuses a domain-specific lexicon and an improved Term Frequency-Inverse Document Frequency (TF-IDF) algorithm to efficiently screen military texts. Subsequently, we design a structured prompting strategy that systematically injects detailed category definitions, boundary constraints, and decision principles into the Large Language Model (LLM) to guide zero-shot reasoning. Experimental results on a rigorously annotated gold-standard test set demonstrate that the framework, when combined with the Deepseek-Reasoner model, achieves a macro-F1 score of 89.69%, representing a 19.82% improvement over the baseline prompt. Notably, reasoning-optimized models significantly outperform dialogue-optimized models in adhering to complex domain rules, with a 9.33% higher F1-score. These findings confirm the framework’s robust efficacy in low-resource settings, providing a reliable technical solution for analyzing highly specialized military texts and guiding model selection for other domain-specific tasks.