Introduction

Due to the sheer volume of digital contents such as customer reviews, blogs and news corpora, sentiment classification has received enormous attention from large number of scholars as well as practitioners. Sentiment classification, also known as sentiment analysis, in electronic commerce is a task of judging the opinions (positive or negative) of customers about products and services (document, sentence, paragraph, etc.) based on computational intelligence such as machine learning. Sentiment classification provides organizations with a tool to transform data into ‘actionable knowledge’ that decision maker can use in pursuit of improved organizational performance. Customer review data can be used for development of market strategy and decision making for product/ service requirements for customer satisfaction, strategic analysis, and commercial planning (Gamon et al. 2005; Ye et al. 2009; Li and Wu 2010; Yu et al. 2013; Kang and Park 2014; Meisel and Mattfeld 2010; Yan et al. 2015; García-Moya et al. 2013). Government and public sector can also take advantages of analysing public sentiment from their blog and social media to obtain citizen feedback on new policy implementation (Ceron et al. 2014; Cheong and Lee 2011). Due to the strategic importance of sentiment classification, the literature is abundant of many studies that propose various algorithms for sentiment classification to improve its accuracy particularly in business and management research domains (Bai 2011; Duric and Song 2012; Fersini et al. 2014; Kontopoulos et al. 2013; Sobkowicz et al. 2012), computer science (Denecke 2008; Melville et al. 2009; Prabowo and Thelwall 2009; Hung and Lin 2013), and computational linguistics (Mullen and Collier 2004; Pang and Lee 2004; Aue and Gamon 2005; Okanohara and Tsujii 2005; Davidov et al. 2010; Liu and Yu 2014) among others. With more sophisticated machine learning algorithms or auxiliary resources for word polarity, researchers tried to make an improvement in accuracy. However, in spite of such strategic values of sentiment classification techniques, the literature still lacks studies that provide practitioners and scholars with clear guidance on how and when to apply different sentiment classification algorithms to data obtained from different problem domains. While previous studies are focusing on increasing the accuracy of the algorithms, less effort was made to understand the impact of the linguistic properties of the dataset they use on the performance of the algorithms. The lack of clear guideline on the use of the algorithms against different datasets makes decision makers underutilize their data that may lead to underoptimal and sometimes wrong decisions by neglecting fits between data and algorithms. Some studies (Pang et al. 2002; Moraes et al. 2013) showed a performance comparison among existing sentiment classification algorithms but they fail to suggestfactors that can affect the performance of each algorithms as they just compared the performance with regards to the different test data or the experiment results from previous works. The lack of literatures regarding systematic comparison among various sentiment classification can be a critical barrier to apply the sentiment classification for researchers and practitioners.