Topic-Adaptive Sentiment Analysis on Tweets via Learning from Multi-sources Data

Twitter sentiment analysis model trained from data of one topic may per-forms worse on another topic. While tweets have diverse topics, and the topic of target data of sentiment analysis task is change with the application requirement, which makes it hard to get a good performance. We also notice that one model cannot performs well on every topic, and tweets have no sentiment label to train the model. On the other hand, there are plenty of available text data with sentiment label and topic information such as online movie and product reviews. So we propose a method based on transfer learning, which use topic information and term distribution as a bridge between target tweets and texts from other sources. It could quickly find appropriate in-stances from other sources, and then use them to train a model adapting to target tweets of specific topic for getting better performance. The experiment result shows the effectiveness of the proposed method.