PERFORMANCE ANALYSIS OF LARGE LANGUAGE MODELS IN DIALOGUE PROCESSING SYSTEMS FOR LOW-RESOURCE LANGUAGES COMPARED TO ENGLISH LANGUAGE
Sauvik Bal, Lopa Mandal · JOURNAL OF MECHANICS OF CONTINUA AND MATHEMATICAL SCIENCES · 2025
This study investigates the performance of dialogue processing systems in low-resource languages, specifically Bengali and Hindi, using advanced transformer-based models. English, a high-resource language, is used as a benchmark for comparison. Transformer models such as BERT, RoBERTa, FLAN-T5, DistilBERT, and GPT-2 are fine-tuned for question answering tasks across these languages. The evaluation includes metrics like F1 Score, Precision, Recall, and Exact Match to assess language-specific performance. The experiment reveals that GPT-2 delivers the highest exact match scores in Bengali and Hindi, while RoBERTa achieves superior F1 scores, indicating balanced performance. The study emphasizes the importance of monitoring training and validation losses to ensure effective model convergence and to identify overfitting. These findings highlight the potential of transformer models in improving dialogue systems for low-resource linguistic contexts.