TOWARDS EXPLAINABLE SENTIMENT ANALYSIS: A HYBRID BILSTM-SVMAPPROACH FOR THE 2025 NEPAL PROTESTS

Abstract

The Nepal Protest of 2025 generated a massive discourse among users, creating a huge corpus of comments with varying perspectives, sentiments, and concerns regarding the political climate at that time. It is imperative to analyze this data in order to gain insight into how society perceives this political scenario and its impact on their lives. This paper explores a hybrid explainable sentiment analysis framework utilizing YouTube commentaries made in the context of 2025 Nepal Protests. Atotal of 10,000 comments from YouTube related to the Nepalese protests of 2025 were collected using the YouTube Data API, followed by preprocessing in the form of language filtering, removing noise, converting to lowercase, lemmatizing, negation handling, and tokenization. Baseline models such as Multinomial Naive Bayes, Logistic Regression, SVM, CNN, and BiLSTM were used in this study with TF-IDF and padded sequences. Based on these baselines, a novel BiLSTM-SVM framework is proposed wherein the BiLSTM framework acts as a deep feature extractor and its output features are fed to an SVM classifier for sentiment classification. The performance of this hybrid system was recorded at an average accuracy of 88% along with a precision, recall, and F1-score of 0.87 in a macro-average. To resolve the issue of interpretability in this model, SHAP (SHapley Additive exPlanations) framework was applied, highlighting the most significant words that led to particular sentiment classes. Words like ”corrupt,” ”violent,” and ”destroy” had a strong impact on negative sentiment classes, whereas ”good,” ”love,” and ”peaceful” contributed to positive predictions. Sentiment distribution analysis across the data corpus showed that there was a positive trend in sentiment owing to reformist and civic sentiment classes, with a large number of negative sentiments due to dissatisfaction with political institutions amid the protests.