Modifying SO-PMI for Japanese Weblog Opinion Mining by using a balancing factor and detecting neutral expressions

Abstract

We propose a variation of the SO-PMI algorithm for Japanese, for use in Weblog Opinion Mining. SO-PMI is an unsupervised approach proposed by Turney that has been shown to work well for English. We first used the SO-PMI algorithm on Japanese in a way very similar to Turney's original idea. The result of this trial leaned heavily toward positive opinions. We then expanded the reference words to be sets of words, tried to introduce a balancing factor and to detect neutral expressions. After these modifications, we achieved a well-balanced result: both positive and negative accuracy exceeded 70%. This shows that our proposed approach not only adapted the SO-PMI for Japanese, but also modified it to analyze Japanese opinions more effectively.