Sentiment classification for telugu using transformed based approaches on a multi-domain dataset.

Area of Science:

• Natural Language Processing (NLP)
• Computational Linguistics
• Machine Learning

Background:

• Sentiment analysis is crucial for resource-rich languages but underdeveloped for low-resource languages like Telugu due to data scarcity.
• Transformer-based models show promise for NLP tasks, with growing interest in multilingual pre-trained models for low-resource languages.

Purpose of the Study:

• To assess the effectiveness of four pre-trained transformer models (IndicBERT, RoBERTa, DeBERTa, XLM-RoBERTa) for sentence-level sentiment analysis in Telugu.
• To establish a reliable benchmark for Telugu sentiment analysis.

Main Methods:

• Evaluated four transformer models: IndicBERT, RoBERTa, DeBERTa, and XLM-RoBERTa.
• Utilized a custom dataset, 'Sentikanna,' comprising diverse domain datasets for Telugu.
• Compared model performance across three distinct datasets for robust evaluation.

Main Results:

• XLM-RoBERTa demonstrated strong performance, achieving 79.42% accuracy in binary sentiment classification.
• All four models showed promising results, indicating the viability of transformer-based approaches for Telugu sentiment analysis.

Conclusions:

• Transformer-based models are effective for Telugu sentiment analysis, even with limited resources.
• XLM-RoBERTa provides a strong baseline for future research and development in Telugu NLP.
• The 'Sentikanna' dataset serves as a valuable resource for evaluating sentiment analysis models in Telugu.