Aiming to address the limited classification performance of non-canonical short texts for agricultural technology consultation, which stemed from semantic sparsity and insufficient feature representation. Based on the Jilin Provincial Agricultural Information Service Platform, it constructed a dataset of Chinese agricultural technology short texts under real scenarios (average length <30 characters) and proposed a fusion optimization model (ErcNet) based on deep semantic reconstruction and dual-driven feature enhancement. Differentiated from traditional methods that only improved single structures for standard agricultural texts, this work achieved dual breakthroughs in the entire agricultural technology field: firstly, it adopted a fixed weighted fusion strategy for the last three encoder layers of pre-trained language models, built a deep semantic representation completion mechanism, and this mechanism effectively resolved semantic fragmentation in spontaneous questions from farmers;secondly, it designed a coordinated optimization module combining multi-scale convolution and efficient channel attention (ECA), established a "semantic density enhancement-discriminative boundary sharpening" dual-driven architecture during feature extraction, and this architecture significantly improved the feature discriminability of non-canonical short texts. Finally, it proposed a global-local feature fusion mechanism, which further complemented semantic extraction. Experimental results showed that on the self-constructed dataset, the model outperformed ERNIE, TextCNN, and other models, with precision at 96.82%, recall at 96.96%, and F1 score at 96.88%;in cross-domain testing on the THUCNews dataset, it achieved 91.70% accuracy, which verified the method's generalization ability.