Troll detection by domain-adapting sentiment analysis

A troll is a user intent on sowing discord on the internet. We propose an approach to detect such users from the sentiment of the textual content in online forums. Since trolls typically express negative sentiments in their posts, we derive features from sentiment analysis, and use SVM^rank to do binary and ordinal classification of trolls. With a small labeled training set of 20 users, we achieved 60% and 58% generalized receiver operating characteristic (ROC) for binary and ordinal troll classification on our forum data respectively. In our experiments, we used features derived from a recursive neural tensor network sentiment analysis model trained on a movie reviews data set written in standard English. However, our forum data set contains messages in a wide spectrum of topics, and are often written in Colloquial Singapore English. We applied domain adaptation techniques to the sentiment analysis model using un-annotated forum data, and achieved a final result of 78% and 69% generalized ROC for binary and ordinal troll classification respectively.