Unsupervised Approach to Detect Extreme Sentiments on Social Networks

Online Social Network (OSN) platforms enable people freedom of expression to share their ideas, views, and emotions that could be negative or positive. Previous studies have investigated the user's sentiments on such platforms to study people's behavior for different scenarios and purposes. The mechanism to collect information on public views attracted researchers by analyzing data from social networks and automatically classifying the polarity of public opinion(s) due to the use of concise language in posts as tweets. In this paper, we propose an unsupervised approach for the automatic detection of people's extreme sentiments on social networks. The approach is based on two steps: 1) We automatically build a standard lexicon consisting of extreme sentiments terms having high extreme positive and negative polarity, and extend that same lexicon with word embedding method [1]; 2) To validate the lexicon, using an unsupervised approach for automatic detection of extreme sentiments. We further evaluated our system's performance on five different social networks and media datasets. This final task shows that, in these datasets, posts that were previously classified as negatives or positives are indeed extremely negatives or positives in numerous cases.