Sentiment Analysis Techniques of Online Product Reviews

Abstract

This research proposes a model for Sentiment analysis of emotional levels of users’ opinions towards online products and services. The techniques used in this research are Machine Learning, Text Mining integrated with 3 types of Word Segmentation and Bag of Words. Four sentiment analysis methods: K-nearest neighbors Classifier, Random Forest Classifier, Logistic Regression and Support Vector Machines (SVM) were used for analysis. The model was developed through 5 steps including 1) Data Collection and Preparation Phase, 2) Text Processing Phase, 3) Training & Streaming Phase, 4) Classification Phase, and 5) Model Evaluation Phase. Consumers’ opinions were gathered from social blogs related to online products and services. 252 articles with 1,412 comments were collected making up 83,670 words in the dataset. The model can help consumers make a decision for purchasing of goods and services, and help entrepreneurs gain the information to improve products and services in the future. This proposed method classified opinions into 3 scales: positive, neutral and negative opinions.  Each method offered different levels of forecast ability: the K-Nearest Neighbors revealed the accuracy index at 56.2%, Random Forest Classifier at 71.6%, Logistic Regression at 77.2% and Support Vector Machines (SVM) at 79%.  However, the combination of SVM and Longest Word Segmentation performed better compared to other techniques and is more appropriate as part of sentimental analysis model for Thai reviews and comments.