Enhanced Question Classification for Marathi: A Comparative Study of Rule-Based and Machine Learning–Based Approaches

Accurate question classification is a crucial component in developing effective Marathi Question Answering Systems, as it enables improved information retrieval and automated responses. However, Marathi, being a morphologically rich and low-resource language, poses significant challenges for automatic classification. This paper presents a comparative study of rule-based and machine learning approaches for classifying Marathi questions into WH and Non-WH categories. The study evaluates a rule-based classifier based on linguistic patterns and WH keywords, alongside machine learning models using Term Frequency-Inverse Document Frequency-based word- and character-level features. By comparing their performance under the same experimental conditions, the study identifies the most effective approach for Marathi question classification. Multiple classifiers, including Naive Bayes, Decision Tree, Support Vector Machine (SVM), and K-Nearest Neighbors, are employed for classification. Experimental results show that the Decision Tree classifier achieves the highest accuracy of 99.66%, followed closely by SVM with 99.41%, demonstrating the effectiveness of pattern-based learning in Marathi. The rule-based approach also performs strongly, achieving an accuracy of 95.64%, highlighting the importance of lexical features such as WH keywords. Further evaluation using precision and recall indicates that Decision Tree and SVM provide balanced and reliable performance with minimal misclassification. While rule-based methods are effective for structured patterns, they lack generalization for complex sentence variations, which are better handled by machine learning models. Future research will focus on expanding datasets, implementing deep learning architectures, and exploring hybrid models that combine linguistic rules with advanced AI techniques to further enhance Marathi question classification.