Sentiment Analysis in Sundanese Using Pre-trained Multilingual Language Models

This paper proposes a document-based sentiment analysis in Twitter for mixed language of Indonesian and Sundanese, 2nd most active speakers of regional language in Indonesia. The existing pre-train language model is finetuned using our new-built Sundanese sentiment training data. We compared the usage of multilingual pre-train language models (XLM-R and XLM-Tweet) with monolingual pre-train language model for Sundanese (SundaBERT). The data used in this study were taken from the social media Twitter, with as many as 7,771 tweets that have been annotated with neutral, negative and positive. The data is divided into 60% training data, 20% validation data and 20% test data. The experimental results showed that the highest performance of 87% F1-score was achieved by using XLM-Tweet pre-train language model. Furthermore, we conducted experiments on the available Sundanese sentiment data set (translated from Indonesian sentiment data set) and obtained F1-score of 82%.