Collaborative Data Cleaning for Sentiment Classification with Noisy Training Corpus

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