BanglaNewsClassifier: A machine learning approach for news classification in Bangla Newspapers using hybrid stacking...

Area of Science:

• Natural Language Processing
• Machine Learning
• Computational Linguistics

Background:

• The proliferation of Bangla online news necessitates advanced classification methods.
• Traditional models struggle with the complexity and linguistic nuances of Bangla datasets.
• Hybrid and deep learning techniques offer potential for improved accuracy in Bangla text classification.

Purpose of the Study:

• To develop and evaluate machine learning and deep learning models for classifying Bangla news articles into eight categories.
• To compare the performance of traditional algorithms, deep learning architectures, and hybrid models.
• To identify the most effective approach for Bangla news classification, addressing low-resource language challenges.

Main Methods:

• Utilized a large dataset of 118,404 Bangla news articles.
• Applied feature extraction techniques: TF-IDF vectorization and word2Vec embeddings.
• Experimented with traditional machine learning, deep learning (e.g., bidirectional long short-term memory), and hybrid stacking classifiers.

Main Results:

• The best-performing model was a stacking meta-classifier combining bidirectional long short-term memory and support vector machine, achieving 94% accuracy.
• This hybrid model significantly surpassed the performance of individual baseline models.
• Comprehensive performance analysis included confusion matrices, ROC curves, and error analysis.

Conclusions:

• Ensemble methods and deep learning, particularly the proposed stacking classifier, are highly effective for Bangla news classification.
• The study demonstrates the efficacy of advanced techniques for low-resource natural language processing tasks.
• This research provides valuable insights into optimizing Bangla text classification systems.