Exploring deep learning approaches for Tamil text processing: A comprehensive review and experimental comparison

Abstract

The integration of deep learning techniques into natural language processing (NLP) has led to significant advancements in language-based applications. However, these rapid developments present challenges for researchers, particularly in understanding their applicability to low-resource languages. Tamil, a morphologically rich Dravidian language, continues to face limitations in NLP due to resource scarcity and linguistic complexity. This study aims to investigate the applicability of various deep learning models for Tamil text processing and to identify effective, computationally efficient approaches suitable for low-resource language settings. To achieve this, the study first provides a focused review of recent deep learning approaches applied to NLP, with particular emphasis on Tamil language processing. It then presents a comparative experimental analysis of 20 deep learning models, including word embedding-based architectures (Word2Vec and FastText combined with ANN, CNN, RNN, LSTM, GRU, and Bi-GRU) and transformer-based models (such as mBERT, DistilBERT, and XLM-RoBERTa), evaluated across four Tamil NLP tasks using five datasets. The experimental results show that the FastText-based Bi-GRU model consistently outperforms other approaches across most tasks. Furthermore, word embedding methods demonstrate performance comparable to or better than transformer-based models while requiring significantly fewer computational resources. These findings highlight that efficient and well-designed models can achieve strong performance for Tamil NLP. The study also emphasizes the importance of improving language resources, such as large-scale corpora and annotated datasets, to further advance NLP research in low-resource languages.