With the rapid increase in online text data, sentiment analysis has become increasingly important in various applications. However, current research on neutral sentiment analysis is relatively limited, and there is still no effective method for accurately identifying neutral emotions. While Bert has shown superior performance in classifying positive and negative emotions compared to traditional approaches like the Naive Bayes algorithm, its accuracy in identifying neutral sentiment remains a challenge. To address this, we propose a method that combines Bert with the Bayesian algorithm and employs a sentiment dictionary constructed from the text data in the database to calculate correction weights for the two sentiments. This approach aims to improve Bert's accuracy in identifying neutral sentiment texts and thus enhance the overall performance of sentiment analysis.