Cross-Model Sentiment Analysis of Tweets on the Russia-Ukraine War: A Comparative Study of Lexicon-Based and Transformer Models

Aditya Arora, Anand Zutshi

Abstract: This research explores sentiment analysis of social media discourse sur-rounding the Russia-Ukraine war by leveraging two distinct sentiment prediction models - Vader (lexicon-based) and Transformer (deep learning-based). A comprehensive pipeline was developed to extract, preprocess, and classify tweets into positive, negative, and neutral sentiments. Exploratory Data Analysis (EDA), visualization, and clustering techniques were employed to identify key patterns and features across sentiment categories. To enhance the robustness of sentiment classification, various machine learning models, including XGBoost, Random Forest, Support Vector Machine (SVM), Naive Bayes, and Logistic Regression, were trained on the Vader-labeled dataset and subsequently tested on the Transformer-labeled dataset. This cross-model evaluation approach provided insights into the generalizability and consistency of machine learning classifiers across different sentiment annotation techniques. The findings highlight disparities and alignments between lexicon-based and neural network-driven sentiment labeling, shedding light on the reliability and effectiveness of hybrid methodologies for social media sentiment analysis in dynamic geopolitical contexts.

Keywords: Sentiment Analysis, Social Media, Russia-Ukraine Conflict, VADER, Transformer Models, Machine Learning, Cross-Model Evaluation