Multi-perspective prompt fusion for zero-shot classification of agricultural news texts with large language models

In the era of smart agriculture, agricultural information and related event news spread rapidly across social media and online platforms. Efficient classification of such massive agricultural texts holds practical value for optimizing production decisions and enhancing risk early warning. The wide dissemination of these texts also raises the demand for accurate and timely content analysis. However, the domain vocabulary sparsity, semantic ambiguity, and annotated data scarcity of agricultural texts pose notable challenges for classification tasks. Existing zero-shot classification methods based on large language models predominantly design prompt templates from a single perspective, making it difficult to simultaneously cover multiple levels of text analysis, while classification judgments also lack cross-validation mechanisms from multi-dimensional information. To address these issues, this paper proposes a multi-perspective prompt fusion framework that constructs multiple complementary analytical perspectives based on hierarchical differences in text analysis, forming multi-level coverage from local terminology features to overall thematic attribution, and guiding large language models to perform classification judgment from different cognitive dimensions. This paper further constructs a dual-dimensional consistency measurement mechanism integrating decision-level and confidence-level metrics, and employs an adaptive threshold-driven disagreement detection and arbitration mechanism to conduct secondary judgment on low-consistency samples. Experiments on the public datasets PestObserver-France and HumAID demonstrate that the proposed method achieves classification accuracy of 85.3% and 68.5%, corresponding to improvements of 3.2% and 3.7% over the strongest baseline ChatAgri. The expected calibration error decreases from 0.135 and 0.148 to 0.098 and 0.112, representing reductions of 27% and 24% respectively. Perspective heterogeneity analysis shows that the pairwise prediction agreement rate among the five perspectives is 18 to 21 percentage points lower than that of the five-path sampling of Self-Consistency, confirming the contribution of heterogeneous perspective design to classification robustness. The framework currently relies on text input only and provides a basis for extension to multimodal agricultural information.