Semi-supervised Learning for Sentiment Classification using Small Number of Labeled Data

Abstract

Sentiment analysis is an essential task to gain insights over a huge amount of opinions and thoughts. Timeliness of data is important in making major decision. However, the manual data labeling method is slow and expensive, it also cannot cope with this enormous amount of data. We investigated the literature of sentiment analysis and discovered most of the works using manual data labeling. We propose semi-supervised learning as a method that helps to reducing the effort and time needed in data labeling as it uses a combination of small amount of labeled data and a large pool of unlabeled data for model training. In our work, we trained semi-supervised deep neural network with different settings and compared the model performances to a baseline, the supervised deep neural network trained with same number of labeled data. From the results, we can see that the unlabeled data is useful in improving the data performances. But it is not a guarantee, the unlabeled data must be handled with care otherwise degraded the model performances.