This paper explores the performance of natural language processing in financial sentiment classification. We collected people's views on U.S. stocks from the Stocktwits website. The messages on this website reflect investors' views on the stock. These messages are classified into positive or negative sentiments using a BERT-based language model. Investor sentiment can be further analyzed to help more investors, businesses or organizations make effective decisions. The experimental results show that the pre-trained BERT model has been fine-tuned on the labeled sentiment dataset, and can recognize the sentiment of investors with an accuracy of more than 87.3%.