Sentiment Analysis of Product Reviews in the Absence of Labelled Data Using Supervised Learning Approaches

Abstract

With the growing pace of internet usage, there is a vast variety of diverse individual opinions and thoughts avail-able online. Consumer reviews can act as a feedback and as well as a pool of ideas for which they can be of immense importance to any business. With the growth and popularity of opinion-rich resources such as online review sites and personal blogs, people now can and do, actively use information technology to seek out and un-derstand the opinions of others to decide whether to buy a product or not. Social media websites such as Facebook, Twitter, and e-commerce websites such as eBay, Amazon, etc. are being widely used to communicate viewpoints effectively. Assigning a positive or a negative sentiment to these reviews can help companies understand their users and also help users to make better decisions. Sentiment analysis being a challenging task can be tackled using supervised machine learning techniques or through unsupervised lexicon based approaches if labelled data is unavailable. In this study, we show that in absence of labelled product reviews of a particular website, labelled product reviews from a different website can be effectively used to train the supervised techniques to achieve a comparable performance to the unsupervised lexicon based approaches. This approach also benefits by covering all of the product reviews which the lexicon based approaches fail to do so. We deduce this by comparing five su-pervised approached and three lexicon based approaches on iPhone 5s reviews gathered from Amazon, Facebook, and Reevoo blog. Furthermore, we also found that unigram features combined with bigram features give the best results, and the effect of varying the training data size on the performance of ML classifiers in some cases was significant whereas in other cases it did not have any effect. Our results also suggest that reviews from Amazon are easiest to classify, followed by reviews from Reevo, and Facebook reviews are the hardest to classify.

Mobile Legend Game Prediction Using Machine Learning Regression Method

“…Precision shows the results of the level of accuracy between the requested data and the results predicted by the algorithm model, while recall serves to display the model's success in retrieving information, and the f1-score is the weighted average value of the results of precision and recall values. Classification report on the recommended algorithm model [29] [30]. 4 describes the precision, recall, and F1-score from the artificial neural network and random forest algorithm models.…”

Cognitive Relationship-Based Approach for Urdu Sarcasm and Sentiment Classification

“…Naïve Bayes (NB): In text classification, NB is considered amongst the classical methods. It is a probabilistic classifier that uses Bayes' theorem with the assumption that the features are mutually independent [78]- [80]. For the document X , the prediction for class L l can be calculated through the following equation: where the Bayes' rule can (in equation 11) can be expanded w.r.t individual features of the document X (such that X ≡ {x 0 , x 1 , • • • , x n }) as:…”

“…There has been much research on sentiment classification in which the F1-score has been used as the main evaluation metric, as demonstrated in works like [18], [38], [78]. However, since the dataset used in this paper is slightly imbalanced, the authors are placing more emphasis on achieving a higher balanced accuracy [83], which is calculated as the arithmetic mean of the true positive rate and true negative rate [91]- [93].…”

Microblog Emotion Analysis Using Improved DBN Under Spark Platform

“…The COAE dataset contained 30,000 articles from 2011, and 40,000 articles from 2014, while the NLP&CC dataset contained 4,000 articles collected in 2013. Some training data were annotated manually using four different emotional categories, namely joy, anger, disgust, and depression (Muhammad et al, 2020).…”