Sentiment analysis on big sparse data streams with limited labels

Sentiment analysis is an important task in order to gain insights over the huge amounts of opinionated texts generated on a daily basis in social media like Twitter. Despite its huge amount, standard supervised learning methods won't work upon such sort of data due to lack of labels and the impracticality of (human) labeling at this scale. In this work, we leverage distant supervision and semi-supervised learning to annotate a big stream of tweets from 2015 which consists of 228 million tweets without retweets (and 275 million with retweets). We present the insights from our annotation process regarding the effect of different semi-supervised learning approaches, namely Self-Learning, Co-Training and Expectation-Maximization. Moreover, we propose two annotation modes, the batch mode where all labeled and unlabeled data are available to the algorithms from the beginning and a lightweight streaming mode that processes the data in batches based on their arrival time in the stream. Our experiments show that stream processing with a sliding window of three months achieves comparable results to batch processing while being more efficient. Finally, to tackle the class imbalance problem, as our dataset is imbalanced toward the positive sentiment class, and its aggravation by the semi-supervised learning methods, we employ data augmentation in the semi-supervised learning process in order to equalize the class distribution. Our results show that semi-supervised learning coupled with data augmentation outperforms significantly the default semi-supervised annotation process. We make the so-called TSentiment15 sentiment-annotated dataset available to the community to be used for evaluation purposes and for developing new methods.