AI Research Analysis

Understanding Nagamese Language

Nagamese is an Assamese-lexified creole, spoken widely across Nagaland. It features 28 phonemes (6 vowels, 22 consonants) and exhibits disyllabic, trisyllabic, and tetrasyllabic word structures. Grammatically, it is a Subject-Object-Verb (SOV) language, similar to other Indian languages like Hindi and Assamese.

Key resources for Nagamese include a Nagamese-English-Assamese dictionary by Bhim Kanta Boruah (with POS information), online news articles from nagamesekhobor.com, the Nagamese Bible, and various books/hymns. Social media content also serves as a source for language data.

Context from Assamese NLP Research

Given Nagamese's Assamese-lexified nature, this section highlights related works in Assamese sentiment analysis, which inform the approach for a low-resource language.

• Das et al. (2021) developed a sentiment polarity model for Assamese news using lexical features and ML classifiers.
• Dev et al. (2021) adapted "Vader" for Assamese texts, leveraging "Bengali-Vader."
• Recent works (Das et al. 2022, 2023, 2024; Dev et al. 2023a, 2023b) have explored multimodal sentiment analysis, deep neural networks (LSTM-CNN), and supervised ML on translated and user-generated Assamese content, demonstrating various feature engineering and model approaches.

Building the Nagamese Corpus

A crucial step for any learning model is the creation of a training dataset. For this study, an annotated corpus of 594 Nagamese sentences was manually labeled for sentiment polarity (positive, negative, neutral) and basic emotions (anger, anticipation, disgust, fear, joy, sadness, surprise, trust).

The unlabelled dataset comprised approximately 25k tokens and 1,195 unique words, collected from online Nagamese News articles (nagamesekhobor.com) and bible passages (bible.com). This mixed corpus provides varied content from current affairs to religious texts, ensuring a diverse representation of the language.

AI Methodology & Feature Engineering

The core methodology involves building a sentiment dictionary of 1,195 unique Nagamese words, categorized into positive (162), negative (162), and neutral (871) sentiments. This dictionary is instrumental in generating features for classification.

Features generated include counts of positive, negative, and neutral words, intensity words, sentence length, and occurrences of specific Nagamese words ("bisi", "olop"), emoticons (😊, ☹️), exclamation marks, and question marks. A refined set of 9 best features was ultimately used for prediction.

Experimental Design & Model Selection

The experiments utilized Naïve Bayes (GaussianNB) and Support Vector Machines (SVC) classifiers, chosen for their established performance in sentiment analysis tasks. The class labels for sentence polarity were positive, negative, and neutral, while emotions included anger, anticipation, disgust, fear, joy, sadness, surprise, and trust.

The annotated dataset of 594 sentences was split into 494 sentences for training and 100 sentences for testing. Implementation was carried out using the scikit-learn library in Python. For SVM, experiments were conducted with linear, poly, and rbf kernels, with the best results reported using the rbf kernel.

Performance Outcomes & Insights

SVM significantly outperformed Naïve Bayes, achieving 71% accuracy for sentiment polarity and 67% for emotion classification. Detailed performance metrics (precision, recall, f1-score) revealed strong results for specific sentiment classes and emotions.

Emotion classification showed F1-scores of 1.00 for fear and surprise, but 0.00 for anger and trust, indicating challenges in detecting these specific emotions with the current dataset and methods. Confusion matrices highlighted common misclassifications, such as anticipation to joy and trust to anticipation.

Conclusion & Future Directions

This work successfully demonstrates the first sentiment polarity and emotion classification for the Nagamese Language. By building an annotated corpus of 594 sentences and a sentiment dictionary of 1,195 Nagamese words, the study provides foundational resources for future NLP efforts.

The use of Naïve Bayes and SVM classifiers yielded promising results, with SVM achieving 71% accuracy for polarity and 67% for emotion. Future work includes creating a larger dataset and exploring more advanced machine and deep learning models to further improve accuracy and address classification challenges for specific emotions.