ARCHIVES

This work presents a sentiment analysis system for the Tai Ahom language, addressing the scarcity of Natural Language Processing resources for ancient, low-resource languages of Northeast India. Tai Ahom was the court and literary language of the Ahom kingdom that ruled Assam for nearly six hundred years, and its script is now part of the Unicode Standard, yet no sentiment lexicon, labelled dataset, or analysis pipeline previously existed for it. The proposed system classifies Tai Ahom text directly, in both native script and romanized form, into positive, negative, and neutral sentiment categories, without using any translation step. The framework combines a manually constructed, multi-rater sentiment lexicon of 498 words with a Complement Naive Bayes classifier from scikit-learn, applied directly to native Tai Ahom text. A hybrid sequential pipeline combines both approaches using a confidence-based fallback strategy. The application is implemented using Streamlit for an interactive interface and Ngrok for secure web-based deployment. Experimental evaluation on 392 labelled test sentences shows that the Naive Bayes model achieves 92.1% accuracy, considerably outperforming the standalone lexicon model (52.8%) and the hybrid pipeline (64.5%). These results are compared against AABEG, a related study on Assamese sentiment analysis that used Google Translate followed by VADER and Naive Bayes, where the translation-based VADER approach scored 0% accuracy. This comparison confirms that working directly with native language text gives far better results than translating it first.