Petar Kovač*, Jelena Bratulić*, Ivan Stresec*
Sentiment analysis is the task of processing data with the goal of gauging general public opinion by classifying a text as either positive, negative or neutral. One particularly difficult aspect of sentiment analysis is irony detection since it can, and often does, change text sentiment. Text-based irony detection is a difficult problem owing to the additional textual cues that are often used and which can be ambiguous. In this paper, we discuss the effect of punctuation and punctuation-based features on the ability of a model to detect irony. The data we used is taken from the SemEval 2018 competition, task 3. We show that including punctuation substantially improves model performance on several neural network models commonly used in NLP (CNNs, RNNs and LSTMs).
Made as part of the Text Analysis and Retrieval (TAR) course at UniZg-FER, academic year 2020/2021.