Improving Japanese Sentiment Analysis with Text Normalization

Abstract

We are working on sentiment polarity classification for Japanese text on social networking services (SNS). Various types of noise appear in SNS text, including typographical errors such as misspellings and SNS-specific phrases. Such textual diversity may degrade the performance of sentiment analysis. Therefore, we improve the performance of sentiment analysis for Japanese SNS texts with text normalization. Experimental results on 6,000 manually normalized posts show that our text normalization improves the performance of sentiment analysis. Our detailed analysis revealed that most of the 33 categories of our normalization typology contributed to the sentiment analysis.