Language-Agnostic Method for Sentiment Analysis of Twitter

With the different events and crises that we are witnessing these days, Twitter plays an essential role in sharing thoughts, opinions, and news worldwide in various languages. Understanding the sentiment of user-generated content has garnered much interest in both industrial and academic communities in recent studies. Due to the limited availability of data from low-resource languages, the focus on multilingual resources is a limiting and challenging issue of sentiment analysis task. Considering the importance of pre-processing in the implementation of a sentiment analysis system, we propose a method consisting of two steps for the pre-processing of tweets in different languages i) a language-agnostic step to replace or remove some elements in the Twitter data structure and ii) a text-normalization step based on the main high-resource language. In addition, we used machine translation techniques to translate low-resource language texts into the main language. We evaluated sentiment classification approaches based on four deep models: an RNN model and three BERT-based architectures, namely vanilla-version, a language-specific, and a large-scale pre-trained model for Twitter. The results show that our method had better accuracy when using a large-scale BERT-based pre-trained model.