Sentiment Analysis through Transfer Learning for Turkish Language

Abstract

Sentiment Analysis (SA) has received much attention in recent years. In this paper, we proposed a model based on the transfer learning technique to address SA problem. First, we utilize word embeddings that are trained on 322K documents from Turkish Wikipedia. The model employs a regular Long Short-Term Memory (LSTM) with dropout. Secondly, we fine-tuned the pre-trained language model on two different target datasets (restaurant and product reviews) independently. Finally, the LSTM is trained to classify reviews according to positive and negative sentiments and its associated performance is assessed. This study is also considered to be the important attempt that uses transfer learning by applying a fine-tuning technique and deep learning architecture to address SA problem for Turkish Language.