Japanese sarcasm detection in Twitter using PU learning and NU learning

Abstract

Recently, a large number of methods have been proposed for detecting sarcasm in Twitter tweets. In English-speaking countries, it is common to use explicit hashtags when tweeting sarcasm, which can be leveraged to collect training data. However, sarcasm-indicative hashtags are not widely used in Japanese tweets, and thus it is difficult to collect a sufficient amount of Japanese sarcastic tweets in this approach.Therefore, in this study, we aim to enhance the classifier performance by increasing the amount of training data through semi-supervised learning. Specifically, based on the assumption that tweets containing the word ’皮肉’ tend to express sarcasm, we collect tweets that meet this criterion. We then extract highly reliable sarcastic sentences by NU learning with the collected data as unlabeled, and train a classifier by adding them to training data. In an evaluation experiment of the classifier, we compared two methods: one utilizing only PU learning and the other combining PU learning with NU learning. The results demonstrated a 3.4% improvement in accuracy rate for the proposed method that integrated PU learning with NU learning, as compared to the method solely using PU learning.