Lexicon-Based Sentiment Analysis of Online Customer Ratings as a Quinary Classification Problem

Abstract

Online customer reviews are not only an important decision-making tool for customers, they are also used by e-commerce providers as a source of information to analyze customer satisfaction. In order to reduce the complexity of evaluation comments, written reviews are additionally represented by evaluation stars in many evaluation systems. Numerous studies address the sentiment recognition of written reviews and view polarity recognition as a binary or ternary problem. This study presents the first results of a holistic approach, which takes up the combination of customer reviews with evaluation points realized in platform-dependent evaluation systems. Sentiment analysis is regarded as a quinary classification problem. In this study, 5,000 customer evaluations are analyzed with lexicon-based sentiment analysis at document level with the target to predict the evaluation points based on the determined polarity. For sentiment analysis the data mining tool RapidMiner is used and the categorization of the sentiment polarity is realized by using different NLP techniques in combination with the sentiment dictionary SentiWordNet. The supervised learning algorithms k-Nearest Neighbor, Naïve Bayes and Random Forest are used for classification and their classification quality is compared. Random Forest achieves the most accurate results in conjunction with NLP techniques, while the other two classifiers provide worse results. The results suggest that a stronger scaling of polarity requires a stronger differentiation between classes and thus a more intensive lexical preprocessing.