FII FUNNY at SemEval-2021 Task 7: HaHackathon: Detecting and rating Humor and Offense

Proceedings of the 15th International Workshop on Semantic Evaluation (SemEval-2021), pages 1226–1231 Bangkok, Thailand (online), August 5–6, 2021. ©2021 Association for Computational Linguistics 1226 FII FUNNY at SemEval-2021 Task 7: HaHackathon: Detecting and rating Humor and Offense Mihai Samson Alexandru Ioan Cuza University Iasi mihai.samson@info.uaic.ro Daniela Gifu Alexandru Ioan Cuza University Iasi daniela.gifu@info.uaic.ro Abstract The “HaHackathon: Detecting and Rating Hu- mor and Offense” task at the SemEval 2021 competition focuses on detecting and rating the humor level in sentences, as well as the level of offensiveness contained in these texts with humoristic tones. In this paper we present an approach based on recent Deep Learning techniques by both trying to train the models based on the dataset solely and by trying to fine tune pretrained models on gigantic corpus. 1 Introduction The figurative language of Social Media is one of the most challenging topics facing natural language processing (NLP). In this study, we refer at humor that requires a multidisciplinary approach for its de- tection (Dan Alexandru and Daniela Gˆıfu, 2020). Imagine, a viral topic on social media as elec- tions (Gˆıfu, 2010) or a political crisis (Delmonte, Rodolfo and Tripodi, Rocco and Gˆıfu, Daniela, 2013). Social media users themselves introduce a specific language based on common practices (e.g., humor, irony), making their message analy- sis very challenging (Reyes, Antonio and Rosso, Paolo and Buscaldi, Davide, 2012). The legitimate research questions of this paper intend to answer: Is humor an insurmountable barrier for Artificial In- telligence (AI)? We propose an approach based on recent DL techniques for sentiment analysis (SA). Furthermore, we experimented with multiple types of DL architectures ranging from Convolutional Neural Networks (CNN) to Recurrent Neural Net- works (RNN) which we tried to train using only the dataset provided by the SemEval-2021 Task 7 competition, but we also used pretrained Trans- former architectures which we fine-tuned using the available data. The rest of the paper is organized as follows: section 2 describes the literature related to sentiment analysis and humor detection, section 3 presents the dataset and method of this study, section 4 summarizes the results of the conducted experiments, with discussions after experiments, followed by section 5 with the conclusions. 2 Related Work This topic has attracted significant attention in re- cent years, evidenced by increasing number of workshops of the same competition (e.g., SemEval- 2017 Task 6: HashtagWars: Learning a Sense of Humor or SemEval-2020 Task 7: Assessing Humor in Edited News Headlines). Such a competition is attractive, especially since the problem of labeled data is somewhat solved, considering the fact that the automatic humor recognition depends on these. For the binary task, as in this case, there are many computational models to solve it or to detect the humor intensity or humor dimension (Yang, Diyi and Lavie, Alon and Dyer, Chris and Hovy, Ed- uard, 2015) (Chen, Peng-Yu and Soo, Von-Wun, 2018). Thus, work on this topic was never followed by high results, as this problem is still almost sub- jective and text classification even …

Content truncated to 500 words