Article Preview
Top1. Introduction
The prevalence of computing and mobile technologies and the proliferation of social media such as Twitter, Facebook, and Google+ have extended the means of human communication (Ellison, 2007; Xiaomei et al., 2018). According to Thelwall et al. (2011), Twitter has more than five hundred million registered users, and more than 80% of them record their daily experiences and events on the Internet. Consequently, certain businesses have paid considerable attention to customer satisfaction with their products and services. In particular, opinions or comments that are based on text messages can be further analyzed.
Opinion mining or sentiment analysis poses a critical research problem (Li and Wu, 2010; Bravo-Marquez et al., 2013; Hu et al., 2017; Kaue & Moreira, 2016; Khan et al., 2014). It is usually based on text mining techniques including natural language processing (NLP), text analysis, and computational linguistics for identifying and extracting meaningful information from online opinions and reviews, news, or other sentences to predict the emotional state of users (Chatterjee et al., 2018; Deshmukh & Tripathy, 2018; Hu et al., 2018; Lee et al., 2018; Nguyen & Nguyen, 2018; Piryani et al., 2017; Tian et al., 2014). In general, useful keywords or terms related to emotions are extracted on the basis of NLP algorithms (Manning and Schutze, 1999) and publicly available resources such as SentiWordNet1; conversely, constructed sentiment classifiers can be used to classify the valence of certain selected opinions.
Opinion mining or sentiment analysis is similar to text classification or categorization. However, effective opinion mining poses two challenges. The first one is that the restrictions of writing an opinion differ substantially between different social network websites, which renders extracting the content of an opinion difficult. Consider, for example, Twitter; a single message is restricted to 140 characters. Under this constraint, fully describing a user’s opinion can be extremely difficult.
Researchers in related works have attempted to extract representative keywords to describe sentimental statuses (Bravo-Marquez et al., 2013; Hu et al., 2017; Khan et al., 2014; Li & Wu, 2010; Nguyen & Nguyen, 2018; Tian et al., 2014). This feature extraction step can produce representative features for each opinion and thus enable constructing a classifier. However, the number of features extracted to describe an opinion (or text message) is usually very high, resulting in very high-dimensional feature vectors. Certain features in the high-dimensional feature vectors are not beneficial for the classification task, and such features can be considered noisy information. Therefore, performing feature selection to filter unrepresentative features has been widely studied in research on text classification (Aghdam et al., 2009; Lee & Lee, 2006; Rehman et al., 2017; Zheng et al., 2004). Recently, this topic has been considered in studies on opinion mining (Abbasi et al., 2008; Agarwal & Mittal, 2013; Nicholls & Song, 2010), but it has not been fully explored.