Begin the News is a kind of text using natural language and are highly publicized for new agencies are experts in collecting information, organize materials and identifying false facts. News titles are mostly used for state the facts and express emotion indirectly, while it is important for many organizations to analyze and collect the instant data of news sentiment and their spreading. In this research, the dataset is from Kaggle that contains more than 50,000 samples with the news titles and their sentiment score. Word embedding is used for word tokenizing and long short-term memory (LSTM) algorithm is used on the classification task. A binary classifier that can divide the samples into positive news and negative news reach an accuracy of 87.25%, and a triple classifier that can divide the samples into positive news, negative news and neutral news have 81.38% accuracy after tuning the hyperparameters. Both of the outcome is better than other methods in comparison. LSTM can be a reliable method to use for semi-supervised news title sentiment analysis.