Using XAI Techniques to Persuade Text Classifier Results: A Case Study of Covid-19 Tweets

Abstract

Background: To offer a transparent decision support system able of classifying tweets' sentiment into positive, neutral, and negative sentiment and explains the prediction result by XAI techniques Methods: We started by data preprocessing phase. For data representation, we used TF-IDF, and we applied four machine-learning algorithms including Naive Bayes, random forest, logistic regression, and support vector machine, as well as four deep learning RNN, LSTM, GRU, and Bi-directional RNN. To raise model trust, we used LIME and SHAP to improve model explainability. Findings: The empirical findings show that the Logistic Regression model and SVM model using the TF-IDF feature extraction approach have the best performance when compared to the other models, with an average accuracy of 84% and 86% respectively. The data balancing step pushed the accuracy of the Random Forest model from 47% to 73%, other models slightly changed. The performance of deep learning models was better than traditional machine learning models, LSTM and GRU achieve approximately 78%, and Bi-directional RNN achieve 79% for dataset 2. Novelty and applications: we propose a highly accurate approach for SA which has been tested on two datasets. Also, to increase trust in model prediction, we explain the predicted sentiment.