Sentiment Analysis of Japanese Tweets Using Auto-Augmented Sentiment Polarity Dictionaries and Advanced Word Embedding

Abstract

Social media posts offer a potent source for extracting public sentiment. Hence, a wide spectrum of methods for automatic sentiment analysis of social media posts – from classical lexicon-based approaches to the more recent and currently trendy machine-learning-based approaches – has emerged. The main drawback of machine-learning-based sentiment analysis is that it requires a huge amount of annotated data, while the main challenge associated with lexicon-based approaches is data sparsity – as social media posts often include out-of-lexicon informal words. In order to minimize these challenges, in this paper, we propose a hybrid approach for sentiment analysis of Japanese tweets by automatically augmenting standard sentiment polarity dictionaries and leveraging state-of-the-art fastText word representation and text classification library. Our experimental results confirmed that such hybrid approaches can improve accuracy of sentiment analysis.