UWB at SemEval-2018 Task 3: Irony detection in English tweets

Proceedings of the 12th International Workshop on Semantic Evaluation (SemEval-2018), pages 520–524 New Orleans, Louisiana, June 5–6, 2018. ©2018 Association for Computational Linguistics UWB at SemEval-2018 Task 3: Irony detection in English tweets Tom´aˇs Hercig1,2 1NTIS – New Technologies for the Information Society, Faculty of Applied Sciences, University of West Bohemia, Czech Republic 2Department of Computer Science and Engineering, Faculty of Applied Sciences, University of West Bohemia, Czech Republic tigi@kiv.zcu.cz http://nlp.kiv.zcu.cz Abstract This paper describes our system created for the SemEval-2018 Task 3: Irony detection in En- glish tweets. Our strongly constrained system uses only the provided training data without any additional external resources. Our system is based on Maximum Entropy classifier and various fea- tures using parse tree, POS tags, and morpho- logical features. Even without additional lexi- cons and word embeddings we achieved fourth place in Subtask A and seventh in Subtask B in terms of accuracy. 1 Introduction Frequent use of creative and figurative language on social media has important implications for natural language processing tasks such as senti- ment analysis. The semantics of a sentence with creative or figurative language can be quite dif- ferent from the same sentence with literal mean- ing and misinterpreting figurative language such as irony represents a significant challenge in sen- timent analysis. Hercig and Lenc (2017) explored the effect of figurative language on sentiment anal- ysis and confirmed that figurative language affects sentiment analysis. The issue of automatic irony and/or sarcasm1 detection has been addressed mostly in English, however there has been some research in other languages as well (e.g. Dutch (Liebrecht et al., 2013), Italian (Bosco et al., 2013), Brazilian Por- tuguese (Vanin et al., 2013), and Czech (Pt´aˇcek et al., 2014)). 2 Task The goal of SemEval-2018 Task 3 (Van Hee et al., 2018) is to detect irony in English tweets. Subtask 1There is only a weak boundary in meaning between irony, sarcasm and satire (Reyes et al., 2012) A detects just binary score for irony and Subtask B also detects more detailed types of irony (non- ironic, ironic by clash, situational irony, and other forms of verbal irony). These subtasks correspond to their respective phases (A and B). Data for sub- task B were available only after phase A was fin- ished. At the evaluation time the following descrip- tions of the submitted system labels were given: • Constrained: only the provided training data were used to develop the system • Unconstrained: additional training data were used Only after the end of the phase A we learned that constrained systems can make use of addi- tional resources like lexicons, dictionaries, embed- dings, etc. Thus we introduce another system label to describe our system – strongly constrained. • Strongly Constrained: using ONLY the the provided training/development data without any additional external resources (such as lexicons, embeddings, etc.) Data statistics for Subtask A and Subtask B are shown in Table 1 and 2 respectively. Label Test Train Non-ironic 473 (60,3%) 1923 (50,2%) Ironic 311 (39,7%) 1911 (49,8%) Table 1: …

Content truncated to 500 words