Reserved Self-training: A Semi-supervised Sentiment Classification Method for Chinese Microblogs

International Joint Conference on Natural Language Processing, pages 455–462, Nagoya, Japan, 14-18 October 2013. Reserved Self-training: A Semi-supervised Sentiment Classification Method for Chinese Microblogs Zhiguang Liu Xishuang Dong Yi Guan∗ Jinfeng Yang School of Computer Science and Technology Harbin Institute of Technology, Harbin, China {triggerliu, dongxishuang, yangjinfeng2010}@gmail.com guanyi@hit.edu.cn Abstract The imbalanced sentiment distribution of microblogs induces bad performance of binary classifiers on the minority class. To address this problem, we present a semi- supervised method for sentiment classi- fication of Chinese microblogs. This method is similar to self-training, except that, a set of labeled samples is reserved for a confidence scores computing pro- cess through which samples that are less than a predefined confidence score thresh- old are incorporated into training set for retraining. By doing this, the classifier is able to boost the performance on the mi- nority class samples. Experiments on the NLP&CC2012 Chinese microblog evalu- ation data set demonstrated that reserved self-training outperforms the best run by 2.06% macro-averaged and 2.30% micro- averaged F-measure, respectively. 1 Introduction Sentiment classification aims to label peoples opinions as different categories such as positive and negative from a given piece of text (Pang et al., 2002). Currently, related research on tradi- tional online media, such as blogs, forums, and online reviews, has made great progress (Baner- jee and Agarwal, 2012; Liu et al., 2005). Howev- er, sentiment classification of microblogs is hard to process due to some unique characteristics of microblogs, for example, short length of update messages and language variations. Moreover, top- ic based microblogs are related with peoples daily lives and people are more likely to post some neg- ative messages to show their unsatisfactoriness, which may partially result in imbalanced senti- ment class distributions. For example, the num- ∗Corresponding author: guanyi@hit.edu.cn ber of negative tweets is far more than that of pos- itive in some topics, which is different from the previous work on sentiment classification that as- sumes the balance between positive and negative samples (Chawla et al., 2002; Yen and Lee, 2009). While supervised techniques have been wide- ly used in sentiment classification (Pang et al., 2002), the main problem that supervised method- s suffered is that they rely on labeled data sole- ly. Semi-supervised methods, which make use of both labeled and unlabeled data, are ideal for sen- timent classification, since the cost of labeling da- ta is high whereas unlabeled data are often readily available or easily obtained (Ortigosa-Hern´andez et al., 2012). However, there are some drawback- s of semi-supervised approaches such as most of the work assume that the positive and negative samples in both labeled and unlabeled data set are balanced, otherwise models often bias towards the majority class (Chawla et al., 2002; Yen and Lee, 2009). In addition, most existing studies on imbal- anced classification focus on supervised learning methods, with few on semi-supervised approach- es (Li et al., 2011). In this study, we propose a reserved self-training method for binary sentiment classification inspired by active learning strategies (Ryan, 2011). Active …

Content truncated to 500 words