Turkish Sentiment Analysis Using BERT

While sentiment analysis is a popular research area, most of the research has been conducted for English and the number of studies for Turkish are rather limited. Limited resources for Turkish natural language processing (NLP) is one of the major challenges for Turkish NLP research. In order to overcome these limitations, we propose two approaches for Turkish sentiment analysis: 1) fine tuning multilingual model of BERT 2) using main model of BERT after machine translation of Turkish texts into English. We conducted experiments on Turkish movie and hotel review datasets where each review is labeled either positive or negative. Our methods achieve high accuracy scores such that in the movie dataset, our BERT models outperform existing methods.