Roman Urdu sentiment analysis using Machine Learning with best parameters and comparative study of Machine Learning algorithms

Abstract

People talks on the social media as they feel good and easy way to express their feelings about topic, post or product on the ecommerce websites. In the Asia mostly the people use the Roman Urdu language script for expressing their opinion about the topic. The Sentiment analysis of the Roman Urdu (Bilal et al. 2016)language processes is a big challenging task for the researchers because of lack of resources and its non-structured and non-standard syntax / script. We have collected the Dataset from Kaggle containing 21000 values with manually annotated and prepare the data for machine learning and then we apply different machine learning algorithms(SVM , Logistic regression , Random Forest, Naïve Bayes ,AdaBoost, KNN )(Bowers et al. 2018) with different parameters and kernels and with TFIDF(Unigram , Bigram , Uni-Bigram)(Pereira et al. 2018) from the algorithms we find the best fit algorithm , then from the best algorithm we choose 4 algorithms and combined them to deploy on the data set but after the deployment of the hyperparameters we get the best model build by the Support Vector Machine with linear kernel which are 80% accuracy and F1 score 0.79 precision 0.79 and recall is 0.78 with (Ezpeleta et al. 2018)Grid Search CV and CV is 5 fold. Then we perform experiments on the Robust linear Regression model estimation using (Huang, Gao, and Zhou 2018)(Chum and Matas 2008)RANSAC(random sample Consensus) that gives us the best estimators with 82.19%.

Keywords

• machine learning
• tfidf
• kaggle
• svm
• rf
• logistic regression
• naïve bayes
• adaboost
• ransac
• hyper parameter