TUG-CIC at SemEval-2021 Task 6: Two-stage Fine-tuning for Intended Sarcasm Detection

Proceedings of the 16th International Workshop on Semantic Evaluation (SemEval-2022), pages 951 - 955 July 14-15, 2022 ©2022 Association for Computational Linguistics TUG-CIC at SemEval-2021 Task 6: Two-stage Fine-tuning for Intended Sarcasm Detection Jason Angel CIC, Instituto Politécnico Nacional Mexico City, Mexico ajason08@gmail.com Segun Taofeek Aroyehun TU Graz Graz, Austria aroyehun.segun@gmail.com Alexander Gelbukh CIC, Instituto Politécnico Nacional Mexico City, Mexico www.gelbukh.com Abstract We present our systems and findings for the iSarcasmEval: Intended Sarcasm Detection In English and Arabic at SEMEVAL 2022. Specif- ically we take part in the Subtask A for the English language. The task aims to determine whether a text from social media (a tweet) is sarcastic or not. We model the problem us- ing knowledge sources, a pre-trained language model on sentiment/emotion data and a dataset focused on intended sarcasm. Our submission ranked third place among 43 teams. In addi- tion, we show a brief error analysis of our best model to investigate challenging examples for detecting sarcasm. 1 Introduction According to the Freedictionary1, sarcasm is a cut- ting statement to express contempt or ridicule, of- ten using words to convey a meaning that is the opposite of their literal or actual meaning. Due to its ambiguous nature, sarcasm plays an important role for resolution of several NLP tasks, such as Sentiment Analysis (Liu et al., 2010; Joshi et al., 2017; Maynard and Greenwood, 2014), Hate speech (Frenda et al., 2022), disagreement clas- sification (Ghosh et al., 2021), Opinion Mining (Kannangara, 2018) among others. One example of benefits of modeling sarcasm is presented by Bouazizi and Ohtsuki (2015) about the use of sar- castic tweets to improve Sentiment Analysis. In this work we present our submission to iSar- casmEval at Semeval-2022 for subtask A (Given a text, determine whether it is sarcastic or non- sarcastic). Our solution system is at the top three teams among 43 teams. The proposed approach uses the pre-trained language model well informed of this task, because it uses sentiment/emotion data 1https://www.thefreedictionary.com/ and we enhanced it with a dataset about intended sarcasm texts. The rest of the document goes as follows: Sec- tion 2 overviews some related works about sarcasm detection tasks. In Section 3, we give a detailed de- scription of our conducted experiments, including the dataset used. Section 4 summarizes our results and offers an interpretation to our findings. Finally, Section 5 presents our conclusion, contributions and future work. 2 Related works Beside the importance of detecting sarcasm for industry applications related to understanding com- ments on social networks or commercial products reviews (Yavanoglu et al., 2018), sarcasm detection has received great attention from the NLP commu- nity, influencing the creation of diverse approaches, benchmark datasets and evaluation methods. The methods for modeling sarcasm range from rule-based system and statistical approaches by ex- ploiting handcrafted features using traditional ma- chine learning (Joshi et al., 2015; Wicana et al., 2017) to modern techniques using deep learning architectures together with word embeddings and pretrained language models (Mehndiratta and Soni, 2019; Srivastava et al., …

Content truncated to 500 words