A deep learning model Font_MBAFF was proposed for the task of text similarity calculation, which was mainly applied to the matching of question pairs in Chinese agricultural short texts. In order to solve the problems of sparse semantic features and inadequate understanding of specialized vocabulary in agricultural short texts, it was firstly optimized in the feature representation stage. By introducing the unique font features of Chinese characters to expand the features, including side radicals and four corner numbers, thus enriching the semantic representation of features. In the feature extraction layer, the multi-scale convolution attention channel weighted network MSCN and the bidirectional long short-term memory network Multi_SAB based on multi-head self-attention mechanism were combined respectively, so that the model can further optimize the feature extraction from the spatial and temporal relationship sequences of semantic features. Finally, TEXTAFF, an improved attention fusion mechanism for text, was used in the intelligent fusion stage of features. The experimental results indicated that the Font_MBAFF model can effectively compensate for the lack of feature words in short texts, optimizing text feature extraction and feature fusion. The accuracy of semantic matching reached 96.42%. Compared with five other semantic matching models, including MaLSTM, BiLSTM, BiLSTM_Self-attention, TEXTCNN_Attention, and Sentence-BERT, the Font_MBAFF model demonstrated significant advantages, achieving a correctness rate that was at least 2.07 percentage points higher. Furthermore, the model proved resilient in experiments with datasets of different sizes, showing rapid response times during testing. Font_MBAFF deep learning model exceled at determining the similarity of Chinese agricultural short texts.