Extracting Pseudo-Labeled Samples for Sentiment Classification Using Emotion Keywords

Sophia Yat Mei Lee,Daming Dai,Shoushan Li,Kathleen Ahrens

DOI: https://doi.org/10.1109/ialp.2011.61

2011-01-01

Abstract:Sentiment and emotion analysis have been traditionally established as independent research topics in NLP. Although they are two important aspects of subjective information and are closely related, there have been few attempts to combine the two analyses. As a preliminary attempt, we integrate emotion information into sentiment analysis by employing emotion keywords to help automatically extract pseudo-labeled samples. The extracted pseudo-labeled samples are then used as the initial training data to perform semi-supervised learning for sentiment classification. Experimental results across four domains show that our approach using emotion keywords is capable of extracting pseudo-labeled samples with high precision (about 90%). Moreover, the pseudo-labeled samples along with the semi-supervised learning approach further improve the classification performance.