Tens of thousands of question and answer data have been increased per second in the internet agricultural technology extension community, these massive data have features of recessive part of speech, emotion and unwanted vectors, and how to implement data aggregation and data block reduction is the difficult problem in this field. An analytical model for the extraction of emotional polarity in agricultural question and answer based on convolutional neural network was proposed, the training set was transformed into a 256-dimensional word vector by using the Skip-gram model after segmenting the dataset with agricultural word segmentation dictionary. The convolution neural network after batch-normalization specification was used to train the dataset, and the neural network model parameters used to identify the part of speech emotional similarities in the agricultural technology promotion community question and answer were obtained. The experimental results showed that the method could accurately identify redundant queues in the test sample set, and by comparing with the other four text classification methods, there were also obvious advantages in each index, the accuracy of the semantic feature extraction for the test set was up to 82.7%.