Aspect Based Sentiment Analysis Model for Hotel Services in Amharic Language Using Machine Learning Techniques

Abstract

In the past years, the World Wide Web (www) has come to be a large source of user-generated content and opinionated data. used by social media, such as YouTube, Facebook, organizational websites, etc. Opinion mining focuses on the sentiment which may be positive, negative, and neutral. Aspect-based sentiment analysis where particular aspects are extracted, sentiment polarity of the aspects is determined. Sentimental analysis carries through in one of three different levels namely, sentence, document, and aspect /feature. Among the three levels, aspect level sentimental analysis is detail and complex but has a better advantage to meet customers and the organization's needs. Comments written by customers have a huge advantage on the success of the hotel. contrarily, it would be difficult for a hotel to manually analyze a numerous amount of commented data to know whether a customer is satisfied or not. In this study, to alleviate this problem, Aspect Based Sentimental Analysis Model through implementing machine learning techniques is proposed. In this research, four Machine learning classification approaches are used. these include Naive Bayesian, Logistic Regression, Support Vector Machine, and Gradient Boosting which were experimented for building and evaluating the sentiment (polarity) model with the extracted features based on the 2124 datasets collected from 10 hotels’ social media pages as well as their websites and annotated by Amharic linguistic expert. the cosine similarity technique for extracting aspects (features) of the hotels is also used. The cross-validation method is known for its skill in the less biased or less optimistic estimation over the previous simple train/split approach. Dataset is partitioned into k equal-sized groups or folds, and each fold is treated as a validation set, while the rest k-1 is used as a training set to fit the model. The 10-fold cross validation of the Gradient boosting model based on the countvetorizer+unigram+Trigram feature with an accuracy of 92.8% and an f1-score of 95% has shown the best performance as compared to SVM, NB, and LR models. Aspect based SA has evolved as an active research area that dominates all the sciences in the world, so this research proposes aspect-based sentimental analysis to predict the aspect (like food, ambiance, drink, price) and polarity (+ve, -ve or neutral) by gathering data from Facebook pages and websites. Due to a shortage of time stemmer is not apply, so we recommend for the others researchers to apply stemmer