Automatic Sarcasm Detection in Twitter Messages

Abstract

In the past decade, social media like Twitter have become popular and a part of everyday life for many people. Opinion mining of the thoughts and opinions they share can be of interest to, e.g., companies and organizations. The sentiment of a text can be drastically altered when figurative language such as sarcasm is used. This thesis presents a system for automatic sarcasm detection in Twitter messages. To get a better understanding of the field, state-of-the-art systems for detecting sarcasm in Twitter messages are explored. Many such systems already exist, and a common theme among them is the use of automatically annotated data for both training and testing. In addition to presenting a system for detecting sarcasm, this thesis also looks into the use of manually annotated data for testing. To this end, a dataset of tweets manually annotated with respect to the presence of sarcasm was built. The result was very similar to that of a previously made set, and both of them showed considerable deviation from automatic annotation. This implies that using automatically annotated data for the task of sarcasm detection in tweets is a mediocre approximation. Experiments with both of the manually annotated datasets also gave very similar results, showing that they are well annotated and reasonably representative for sarcasm detection in tweets.