BERT at SemEval-2020 Task 8: Using BERT to Analyse Meme Emotions

Proceedings of the 14th International Workshop on Semantic Evaluation, pages 1094–1099 Barcelona, Spain (Online), December 12, 2020. 1094 BERT at SemEval-2020 Task 8: Using BERT to analyse meme emotions Adithya Avvaru1,2 and Sanath Vobilisetty2 1 International Institute of Information Technology, Hyderabad, India 2Teradata India Pvt. Ltd, India 1adithya.avvaru@students.iiit.ac.in 2sanath.vobilisetty@teradata.com Abstract Sentiment analysis, being one of the most sought after research problems within Natural Language Processing (NLP) researchers. The range of problems being addressed by sentiment analysis is ever increasing. Till now, most of the research focuses on predicting sentiment, or sentiment categories like sarcasm, humor, offense and motivation on text data. But, there is very limited research that is focusing on predicting or analyzing the sentiment of internet memes. We try to address this problem as part of “Task 8 of SemEval 2020: Memotion Analysis” (Sharma et al., 2020). We have participated in all the three tasks of Memotion Analysis. Our system built using state-of-the-art pre-trained Bidirectional Encoder Representations from Transformers (BERT) (Devlin et al., 2018) performed better compared to baseline models for the two tasks A and C and performed close to the baseline model for task B. In this paper, we present the data used for training, data cleaning and preparation steps, the fine-tuning process of BERT based model and finally predict the sentiment or sentiment categories. We found that the sequence models like Long Short Term Memory(LSTM) (Hochreiter and Schmidhuber, 1997) and its variants performed below par in predicting the sentiments. We also performed a comparative analysis with other Transformer based models like DistilBERT (Sanh et al., 2019) and XLNet (Yang et al., 2019). 1 Introduction Social Media (Mandiberg, 2012) has gained a lot of traction since it’s inception. Almost 50% of the world’s population is on social media. Be it Facebook, Instagram, Twitter or any other social media platform, content generation is growing exponentially day by day. Recently, there has been a surge of using memes1 as a communication medium in social media platforms. Memes can take many forms; the most prominent usage of memes consisted of images combined with text. Memes, being images, can have text in one or more languages. This makes it even more difficult for an algorithm to understand and decode its sentiment or any other characteristic. Of late, smaller videos as memes are also gaining popularity. Irrespective of the type of meme, the meme may get changed, remixed or recreated while communicating through social media networks (French, 2017). Memes are used in contexts involving political discussion (Nave et al., 2018), to add a sarcastic perspective to the discussion, for social purposes, etc. Meticulously analysing the memes can help us understand the involvement of societal factors like race and gender (Milner, 2013), implications on culture and the values promoted by the memes. Moreover, analysing the meme’s underlying emotion can help us understand and possibly eradicate propagation of fake news, offensive content through memes (earlier identification of offensive content is only based on text (Zampieri et al., 2019)) and might also help in …

Content truncated to 500 words