Sentiment recognition analysis is an important method for studying social textual information. It has an important position in social text analysis and research. However, at present, the efficiency of official document text sentiment recognition is low, and manual judgment methods are often used, so the subjective consciousness is strong. This article aims to study the sentiment official document text recognition and analysis method based on the neural network BERT model. It extracts the sentiment information contained in the internet text information through the BERT–SVM model algorithm under deep learning and then mines the user's sentiment. Sentiment analysis is carried out on the sentences in the article, considering the influencing factors of individuals, society, and even the country, putting the method in the position of analyzing the different sentiments represented by a sentence or each word. This article first describes the related technologies of text sentiment recognition, such as feedforward neural networks, convolutional neural networks, and recurrent neural networks. By sentiment training using LSTM-RNN and LSTM-RNN-word2vec models, experiments show that the average accuracy of sentiment classification is 95.12%, the K nearest neighbors result is 90.87% and the Bayesian classifier is 86.84%. By comparison, the BERT–SVM model improves the accuracy of text sentiment classification.