Evaluation of the accuracy of transfer learning on sentiment analysis for Indonesian tweets

Sentiment analysis is an automatic process of understanding, extracting and processing textual data to obtain the sentiment information. From machine learning point of view, the sentiment analysis is a supervised learning problem whose training and predicting data come from a similar domain. When domain changes, the machine learning model must be rebuilt from scratch using new training data. New training data requires manual labeling process which is very costly and time-consuming. Therefore, it would be more effective and efficient using transfer learning which uses the training data from an already available domain to deal with the estimating data on different domains. In this paper, we evaluate the accuracy of the transfer learning on sentiment analysis for Indonesian tweets. Our simulations show that the accuracy of the transfer learning is still lower than that of the supervised learning. Moreover, the bi-gram features can improve the accuracy of the transfer learning.