WESSA at SemEval-2020 Task 9: Code-Mixed Sentiment Analysis using Transformers

In this paper, we describe our system submitted for SemEval 2020 Task 9,\nSentiment Analysis for Code-Mixed Social Media Text alongside other\nexperiments. Our best performing system is a Transfer Learning-based model that\nfine-tunes "XLM-RoBERTa", a transformer-based multilingual masked language\nmodel, on monolingual English and Spanish data and Spanish-English code-mixed\ndata. Our system outperforms the official task baseline by achieving a 70.1%\naverage F1-Score on the official leaderboard using the test set. For later\nsubmissions, our system manages to achieve a 75.9% average F1-Score on the test\nset using CodaLab username "ahmed0sultan".\n