Sentiment Analysis of Turkish Drug Reviews with Bidirectional Encoder Representations from Transformers

Abstract

Sentiment analysis of user generated product or service reviews is significant to enhance quality. Healthcare related computational linguistics studies particularly analysis of drug based user criticisms have principal importance above all. Sentiment analysis of healthcare reviews reveal the relations between patients, doctors and healthcare services. More specifically, sentiment analysis of drug reviews may be used to acquire relations such as adverse drug reactions (ADRs), diagnosis-treatment assist, and personalized therapy recommendations. Most of the drug review sentiment studies are in English. Though Turkish is a widely spoken language, there is limited research conducted on medical domain and there is particularly no study related to drug review sentiment analysis. In this study, we generated a Turkish drug review dataset and we evaluated the generated dataset in detail against (i)traditional machine learning algorithms with language pre-processing steps, stemming and feature selection, (ii)deep learning algorithms with word2vec embedding language model and (iii)various bidirectional encoder representations from transformers (BERT) models in terms of sentiment analysis. The experiments show that neural transformers are promising in Turkish drug review sentiment identification. In particular, Turkish dedicated BERT (BERTurk) resulted in 95.1% weighted-F1 score as the best drug review sentiment prediction performance.