With the development of information technology, agricultural information consultant service based on mobile Internet has become an important part of agro-technical extension system. More than ten million questions in all have been collected by agro-technical extension Q&A community. With the continuous popularization of Q&A community, answering questions manually only by agricultural experts and technicians can neither follow the rapid growth of the questions nor meet the needs of farmers who want to be answered quickly and accurately. Agricultural intelligent Q&A is one of the effective ways to solve the problem. High quality text matching for new questions is the key technology. The accuracy of text matching is limited by the characteristics of agricultural text, such as large amount of data, poor standardization, wide range, much noise, and sparse features. In order to improve the accuracy, the deep semantics, word co-occurrence and maximum matching degree of agricultural short text were extracted and Co_BiLSTM_CNN model composed of bi-long short-term memory, convolutional neural networks, dense networks and Siamese network of shared parameters, was proposed to extract multi-semantic features. The precision, recall, F1, accuracy and time complexity were selected as evaluation indexes to comprehensively measure the performance of the model. The experimental results showed that the model could extract text features more comprehensively, with an accuracy of 94.15%. Compared with the other six text matching models, the experimental results showed obvious advantages.