Guoym at SemEval-2020 Task 8: Ensemble-based Classification of Visuo-Lingual Metaphor in Memes

Proceedings of the 14th International Workshop on Semantic Evaluation, pages 1120–1125 Barcelona, Spain (Online), December 12, 2020. 1120 guoym at SemEval-2020 Task 8: Ensemble-based Classification of Visuo-Lingual Metaphor in Memes Yingmei Guo1, Jinfa Huang2, Yanlong Dong1, Mingxing Xu1 1Department of Computer Science&Technology, Tsinghua University 2School of Electronics and Computer Engineering (SECE), Peking University (guoym18, dongyl17)@mails.tsinghua.edu.cn xumx@tsinghua.edu.cn, jinfahuang@stu.pku.edu.cn Abstract In this paper, we describe our ensemble-based system designed by guoym Team for the SemEval- 2020 Task 8, Memotion Analysis. In our system, we utilize five types of representation of data as input of base classifiers to extract information from different aspects. We train five base classifiers for each type of representation using five-fold cross-validation. Then the outputs of these base classifiers are combined through data-based ensemble method and feature-based ensemble method to make full use of all data and representations from different aspects. Our method achieves the performance within the top 2 ranks in the final leaderboard of Memotion Analysis among 36 Teams. 1 Introduction Memotion analysis is a task to understand the emotions of memes (Sharma et al., 2020). A meme is an idea, behavior, or skill that can be transferred from one person to another by imitation: stories, fashions, inventions, recipes, songs, ways of plowing a field, or throwing a baseball or making a sculpture (Black- more and Blackmore, 2000). With the popularization of the Internet and instant message applications, Internet memes on social media platforms such as Facebook, Instagram, and Twitter have become an effective way of communication. (Shifman, 2014) However, the latest Internet memes have facilitated the prevalence of hate speech in online social media which makes this study problem increasingly significant (Williams et al., 2016). Detecting an offensive meme is more complex than detecting an offensive text because it involves visual and language understanding. So it is necessary to find a hybrid approach for the computational processing of Internet memes. Similar to the text content on social media, memes need to be analyzed and processed to extract the information and the emotion conveyed. (Peirson et al., 2018; Oliveira et al., 2016) have tried to automate the meme generation process. (French, 2017) have tried to extract its inherent sentiment in the recent past. On this basis, more studies need to be done to distinguish fine-grained emotions in memes. In this paper, we describe our approach to SemEval-2020 task 8, Memotion Analysis, which aims to identify the emotions of memes. There are three subtasks: task A aims for sentiment classification, task B aims for humor classification and task C is to quantify the scales of semantic classes. The system performance is evaluated by the macro F1 score for task A and the macro-averaged F1 score for task B and task C. According to our observation of the dataset and single classifier’s performance, we design an ensemble-based classification system (Rokach, 2010) which is composed of multiple based classifiers. These base classifiers are combined using data-based ensemble method and feature-based ensemble method which are two steps of voting to get …

Content truncated to 500 words