An Adaptive Multilingual and Morphology-Aware Framework for Low-Resource Indian Language Translation
Kavitha Thiyagarajan, Zahrah Sataar, N. Srinivasan, E Vanitha., Ponmurugan Panneer Selvam
Abstract
In recent years, Neural Machine Translation (NMT) has gained importance for enabling cross-lingual communication, yet low-resource Indian languages continue to face difficulties owing to limited corpora and complex morphology. However, a recent study on Kannada-Tulu translation introduced a Transformer-based NMT model with part-of-speech (POS) tagging, but it was constrained by a very small dataset and achieved only modest performance improvements. Therefore, Adaptive Multilingual Morphology-Aware Neural Machine Translation (AMM-NMT) was proposed. Initially, data are collected from the newly released English-Tulu parallel corpus and large-scale Samanantar corpus of Indic language datasets. Then, the data undergoes cleaning, normalization, and back-translation to enrich scarce Tulu resources, followed by Sentence-Piece subword tokenization to handle vocabulary diversity. Furthermore, training was performed using multilingual pre-training on related Dravidian pairs from Samanantar. Furthermore, fine-tuning of English-Tulu was enhanced through curriculum learning and adapter layers for adaptive transfer. Finally, the model integrates morphological feature embeddings (POS, case, tense, number, and gender) to capture linguistic richness, enabling the generation of fluent Tulu translations from Kannada or English inputs. Experimental results show that the proposed AMM-NMT achieves a Bilingual Evaluation Understudy (BLEU) score of 27.3, character n-gram F-score (chrF) of 41.6%, and Word Error Rate (WER) of 42.1%.