Irony, which is a way of expression through the use of the opposite, commonly occurs in daily social media posts. Hence, automatic detection of irony is essential to understand the semantics of informal texts more accurately. The literature has several sentiment analysis studies on Turkish texts, but those focusing on irony detection are very few. This paper investigates the effectiveness of a rich set of supervised learning methods varying from traditional to deep neural solutions on Turkish texts. Traditional irony detection methods such as Support Vector Machine (SVM) and tree-based binary classifiers are analyzed on Turkish informal texts. Furthermore, such methods are extended by polarity-based information and graph-based similarity scores as features. Additionally, neural architecture based solutions including BERT and various LSTM network models are adapted for the problem. Irony detection performance of all the methods are comparatively analyzed on a data set collected within this study, which is larger than the previously used irony detection data sets in Turkish.

No abstract