An Indonesian Sentiment Classification Model Based on Multi-task Learning

Abstract

AbstractWith the rapid development of social network, people are keen on giving voice to their feelings on social media containing their attitude towards products or services, which is significant for company to have a better understanding of the customers’ opinions. Nowadays, many researchers are concentrating on sentiment classification research in universal language such as English and Chinese. However, sentiment classification research has not been widely applied for low-resource language. In this paper, we constructed an Indonesian sentiment classification dataset consisting of 108,000 hotel comments, dividing into positive, negative and neutral polarities. Moreover, we transformed three-label sentiment classification into three binary classification models and proposed a multi-task learning (MTL) model for Indonesian sentiment classification. We utilized Bi-LSTM as the general representation layer shared across three binary sentiment classification models, aiming to better capture the bidirectional semantic dependency information. In the sentiment-specific representation layer, the sentiment-specific features will be extracted through the max-pooling layer, and the three polarities probability distributions were predicted respectively through the sigmoid non-liner layer. The experimental results demonstrated that, both on the existing open source dataset and the self-constructed dataset, our model outperforms other models that we used as baseline.KeywordsIndonesian sentiment classificationMulti-task learning modelMulti-Label classification