Comparative Evaluation of Rule-Based, SVM, CRF and BiLSTM Models for Gujarati Part-of-Speech Tagging Using an Art and Culture Corpus

Part-of-Speech (POS) tagging is a fundamental Natural Language Processing (NLP) task that assigns grammatical categories to words in a sentence. Gujarati, being a low-resource language, has limited benchmark datasets and standardized evaluation frameworks. This study presents a comparative analysis of Rule-Based, Support Vector Machine (SVM), Conditional Random Fields (CRF), and Bidirectional Long Short-Term Memory (BiLSTM) models for Gujarati POS tagging using an Art and Culture corpus containing approximately 20,000 manually annotated tokens based on BIS tag standards. A unified experimental framework is employed to ensure fair comparison across all models. The results show that BiLSTM achieved the highest accuracy of 96.7%, followed by CRF (94.2%), SVM (91.5%), and the Rule-Based approach (79.1%). The study demonstrates that deep learning models provide superior contextual understanding, while CRF remains an effective choice for low-resource Gujarati POS tagging.