Kannada-to-English Neural Machine Translation Using an LSTM Encoder-Decoder Architecture

Abstract

Machine Translation (MT) plays an important role in enabling communication and information access across different languages. However, translating Indian regional languages remains challenging due to limited parallel corpora, morphological richness, flexible word order, and vocabulary variations. Kannada, a widely spoken Dravidian language, presents several challenges for developing accurate and context-aware translation systems, particularly when translating into English. This study proposes a deep learning-based Kannada-to-English Neural Machine Translation (NMT) system using a Long Short-Term Memory (LSTM) encoder–decoder architecture. The proposed approach employs text normalization, tokenization, vocabulary construction, word embedding, and sequence padding during preprocessing. An LSTM-based encoder learns the semantic representation of Kannada input sequences, while the decoder generates the corresponding English translation sequentially. An attention mechanism can be incorporated to improve the model's ability to focus on relevant source-language words, particularly for longer sentences. The proposed system is evaluated using standard machine translation metrics such as BLEU, ROUGE, and METEOR. The performance is compared with baseline sequence-to-sequence and alternative neural translation models to assess translation quality. The study aims to demonstrate the applicability of LSTM-based neural architectures for low-resource Kannada-to-English translation and identify the major challenges affecting translation accuracy. The findings can contribute to the development of more effective and accessible machine translation systems for Kannada and other Indian regional languages.