In sentiment analysis, sarcasm is rarely taken into account. The performance of sentiment systems can improve by applying sarcasm detection. Indications for the detection of sarcasm in tweets are discussed in this study. Two thousand tweets from 2015 with the hashtag #sarcasme or #not are used to train a SVM classifier. Features aimed at punctuation and content are used. Machine learning is used to measure the performance at various ratios sarcastic and non-sarcastic tweets. Testing is done on a set of 230 manually annotated tweets (supplemented by non-sarcastic tweets). The system achieves an accuracy of 62 percent in a set of 50 percent sarcastic and 50 percent non-sarcastic tweets. Intensifiers, ellipses, exclamation marks and to a lesser extent emoticons are the best predictors of sarcasm. The system outperforms the baseline, but it is impossible to classify all tweets properly.

The complete Bachelor Thesis can be downloaded here:
Thesis