Multi-class sentiment analysis of urdu text using multilingual BERT.

Area of Science:

• Natural Language Processing
• Computational Linguistics
• Artificial Intelligence

Background:

• Sentiment analysis (SA) is crucial for opinion mining but lacks resources for low-resource languages like Urdu.
• Existing SA research primarily focuses on English, leaving a gap in Urdu opinion analysis.

Purpose of the Study:

• To develop a novel, manually annotated, multi-class Urdu dataset for sentiment analysis.
• To establish baseline performance metrics for Urdu SA using diverse computational methods.
• To evaluate the efficacy of advanced models, including Multilingual BERT (mBERT), for Urdu SA.

Main Methods:

• Creation of a 9312-review Urdu dataset across domains like food, movies, and politics, annotated into positive, negative, and neutral classes.
• Implementation and comparison of rule-based, machine learning (SVM, NB, Adaboost, MLP, LR, RF), and deep learning (CNN-1D, LSTM, Bi-LSTM, GRU, Bi-GRU) models.
• Fine-tuning of Multilingual BERT (mBERT) using BERT word embeddings for Urdu sentiment classification.

Main Results:

• The proposed mBERT model, utilizing BERT pre-trained word embeddings, significantly outperformed traditional machine learning, deep learning, and rule-based approaches.
• The mBERT model achieved a high F1 score of 81.49%, establishing a new benchmark for Urdu sentiment analysis.
• Comparative analysis highlighted the superior performance of deep learning models over traditional machine learning and rule-based methods.

Conclusions:

• The developed Urdu sentiment analysis dataset is a valuable resource for advancing research in low-resource languages.
• Multilingual BERT (mBERT) demonstrates exceptional capability in Urdu sentiment analysis, outperforming other evaluated techniques.
• This research provides a strong foundation and benchmark for future Urdu Natural Language Processing (NLP) tasks.