Feature-Optimized Machine Learning Framework for Multiclass Kashmiri Text Classification: A Comparative Analysis

Abstract

Classifying text in low resource languages poses significant difficulties because of the sparsity of annotated data sets and morphological complexity of low resource languages. This paper fixes this deficiency by establishing a strong machine learning system to perform multi-classification of Kashmiri text. The main objective of this study is the evaluation of the conventional methods called as machine learning techniques on the large scale, self-curated Kashmiri corpus consisting of about 27,000 sentences in nine semantic categories. The data is systematically preprocessed to normalize and tokenize it, remove noise and then using the feature extraction approach that quantifies term importance using frequency within documents and inverse corpus-level distribution called as (TF-IDF) having the vocabulary size 5000. This approach includes the implementation and optimization of five supervised learning algorithms, such as linear classification models, Bayesian probabilistic classifiers, distance-based instance learning technique, maximum-margin classifiers and ensemble learning approaches based on decision trees with the help of GridSearchCV and RandomisedSearchCV with k-fold cross-validation. F1-score, precision, Accuracy, recall and Receiver Operating Characteristic-Area Under Curve (ROC-AUC) are used to assess the models and are backed-up with analysis of the confusion matrix. The findings show that SVM has the best classification accuracy of 93%, which shows that its performance is outstanding in comparison to other models. Results demonstrate the usefulness of optimization of the features representation and model-tuning in low-resource context and offer a predictable baseline to the future evolution of natural language processing investigations in the Kashmiri language.