Sentiment Analysis for Transliterated Hindi and Marathi Language using Machine Learning Approach

Abstract

Sentiment analysis for local transliterated languages such as Hindi and Marathi has gained increasing research interest due to the linguistic diversity and informal nature of user-generated content. However, most existing approaches are limited by insufficient datasets that fail to capture the wide range of transliteration-based spelling variations inherent in such languages. To address this gap, the present study introduces a manually curated sentiment word dictionary for Hindi and Marathi, enriched with diverse transliterated spellings and associated sentiment weights. Using this resource, multiple sentence-level datasets were developed, including Hindi, Marathi, and real-world YouTube comment datasets, where each sentence is annotated with an average sentiment score derived from constituent sentiment words. A comprehensive sentiment classification framework was then designed using three feature extraction strategies: Count Vectorizer (CV), TF-IDF Vectorizer, and a Graph Embedding Technique (GET) combined with Rank-Based Selection (RBS). These features were used to train and evaluate three machine learning classifiers, Support Vector Machine (SVM), Logistic Regression (LR), and Random Forest (RF), which relies mainly on manually engineered linguistic features and graph-based representations. Experimental results demonstrate that SVM consistently outperforms LR and RF across all feature configurations. Among all combinations, SVM with TF-IDF achieved the highest accuracy, while SVM with GET+RBS demonstrated robust performance across datasets. Furthermore, the Hindi, Marathi, and mixed Hindi-Marathi datasets yielded comparable and higher accuracies than the YouTube comments dataset, confirming the advantage of structured transliterated corpora in sentiment analysis.