Collaborative data cleaning for sentiment classification with noisy training corpus

Abstract

Labeled review corpus is considered as a very valuable resource for the task of sentiment classification of product reviews. Fortunately, there are a large amount of product reviews on the Web, and each review is associated with a tag assigned by users to indicate its polarity orientation. We can download such reviews with tags and use them as training corpus for sentiment classification. However, users may assign the polarity tag arbitrarily and inaccurately, and some tags are not appropriate, which results in that the automatically constructed corpus contains many noises and the noisy instances will deteriorate the classification performance. In this paper, we propose the co-cleaning and tri-cleaning algorithms to collaboratively clean the corpus and thus improve the sentiment classification performance. The proposed algorithms use multiple classifiers to iteratively select and remove the most confidently noisy instances from the corpus. Experimental results verify the effectiveness of our proposed algorithms, and the tri-cleaning algorithm is most effective and promising. © 2011 Springer-Verlag.