FII UAIC at SemEval-2022 Task 6: iSarcasmEval - Intended Sarcasm Detection in English and Arabic

Proceedings of the 16th International Workshop on Semantic Evaluation (SemEval-2022), pages 970 - 977 July 14-15, 2022 ©2022 Association for Computational Linguistics FII UAIC at SemEval-2022 Task 6: iSarcasmEval - Intended Sarcasm Detection in English and Arabic Tudor Manoleasa Faculty of Computer Science, Alexandru Ioan Cuza University of Iasi, Romania tudor.manoleasa@info.uaic.ro Iustin Sandu Faculty of Computer Science, Alexandru Ioan Cuza University of Iasi, Romania iustin.sandu@info.uaic.ro Daniela Gifu Faculty of Computer Science, Alexandru Ioan Cuza University of Iasi, Romania Institute of Computer Science, Romanian Academy - Iasi Branch daniela.gifu@info.uaic.ro Diana Trandabat Faculty of Computer Science, Alexandru Ioan Cuza University of Iasi, Romania diana.trandabat@info.uaic.ro Abstract The “iSarcasmEval - Intended Sarcasm Detec- tion in English and Arabic” task at the Se- mEval 2022 competition focuses on detecting and rating the distinction between intended and perceived sarcasm in the context of tex- tual sarcasm detection, as well as the level of irony contained in these texts. In the context of SemEval, we present a binary classification method which classifies the text as sarcastic or non-sarcastic (task A, for English) based on five classical machine learning approaches by trying to train the models based on this dataset solely (i.e., no other datasets have been used). This process indicates low performance com- pared to previously studied datasets, which in- dicates that the previous ones might be biased. 1 Introduction One of the most challenging tasks facing natural language processing (NLP) is the automatic figura- tive language detection (Gifu Daniela, Samson Mi- hai, 2021), (C. Van Hee, E. Lefever, and V. Hoste, 2018), such as humor, sarcasm or irony. In gen- eral, this way of expressing takes advantage of linguistic elements in order to project complex and explainable meanings. In this paper, we investigate the binary classification models for figurative lan- guage, as is sarcasm (Dan Alexandru and Daniela Gˆıfu, 2020). In general, it is omnipresent on the public space, being disruptive of computational sys- tems that harness this data to perform tasks such as opinion mining and sentiment analysis in elec- tions (Gˆıfu, Daniela, 2010). The political actors themselves introduce a specific language based on sarcasm or irony, making their message analysis very challenging (Reyes, Antonio and Rosso, Paolo and Buscaldi, Davide, 2012). In order to identify the figurative meaning of a specific message, it is required to encode each sentence separately. The research question guiding this paper is what are the most efficient sarcasm detection algorithms? We propose an approach based on three classi- cal machine learning (ML) approaches by trying to train the models based on this dataset solely (i.e., no other datasets have been used). Further- more, we experimented with architectures rang- ing from Na¨ıve Bayes (multinomial, complement and bernoulli variants), Support Vector Machines (SVMs with linear, polynomial and RBF kernels) to Logistic Regression which we tried to train using only the dataset provided by the SemEval-2022 Task 6 competition (Ibrahim Abu Farha, Silviu Oprea, Steven Wilson, and Walid Magdy, 2022). The rest of the paper is organized as follows: sec- tion 2 …

Content truncated to 500 words