Comparative Analysis of Machine Translation for Hindi-Dogri Text Using Rule-Based, Statistical, and Neural Approaches

Abstract

Machine translation has made significant progress in several Indian languages, but some, known as computationally low-resourced languages, have seen very little work in this field. Dogri language which is listed in the 8th Schedule of the Indian Constitution is one such language. The authors has developed a Machine Translation System for the Hindi-Dogri pair in the fixed news domain using three approaches: Rule-Based Machine Translation (developed using linguistic rules), Statistical Machine Translation (built using the Moses toolkit), and Neural Machine Translation (developed using neural networks). A comparison of all three approaches is presented in this paper. The paper also discusses various research challenges identified in each approach used for machine translation. A corpus of around 0.1 million sentences in the news domain was used for the development of corpus-based techniques, i.e., SMT and NMT models. The authors also addressed the question of whether NMT produces equivalent or better results compared to the SMT and RBMT approaches. Evaluation of the test results was performed by language experts along with the Bilingual Evaluation Understudy (BLEU) metric. In expert evaluation, it was observed that the NMT and SMT models' results are less ambiguous compared to RBMT. The BLEU score of RBMTS is (79.65), SMT is (52.39) and Bidirectional Embedding LSTM model of NMT is (52.46). The performances of the SMT and NMT models can improve further with the increase in dataset (bilingual parallel corpus).