Enhanced Tamil Morphological Analysis through Deep Learning and Embedding-Based Similarity Techniques

Abstract

Abstract Morphological analysis is vital in NLP, especially for morphologically rich languages like Tamil, which pose challenges due to complex inflectional and derivational forms. Traditional rule-based methods struggle with scalability and unseen words, while deep learning remains underexplored for Tamil. This research proposes a hybrid approach combining deep learning for lemma prediction and an embedding-based similarity method for grammatical feature prediction. Various architectures—including Recurrent Neural Network (RNN), Long Short term memory (LSTM) and Gradient recurrent unit (GRU)—are evaluated for lemma prediction, while FastText embeddings enable effective feature transfer for unseen words, addressing the out-of-vocabulary problem. The model is trained on curated word-lemma pairs and grammatical annotations, demonstrating high accuracy and generalization. This work offers a scalable, low-resourcefriendly solution for Tamil morphological analysis and contributes to advancing Tamil NLP.