Comparative analysis of machine translation for Hindi-Dogri text using rule-based, statistical, and neural approaches

Machine translation has made significant progress in several Indian languages; however, some, known as computationally low-resourced languages, have seen very little work in this field. The Dogri language, which is listed in the 8th Schedule of the Indian Constitution, is one such language. The authors have developed a machine translation system for the Hindi-Dogri language 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 article. The article also discusses various research challenges identified in each approach used for machine translation. A corpus of approximately 0.1 million sentences in the news domain was used to train the corpus -based statistical machine translation (SMT) and neural machine translation (NMT) models. The authors also addressed whether NMT produces results equivalent to or better than those of SMT and rule-based machine translation (RBMT). To ensure a comprehensive evaluation, the outputs of all systems were evaluated using two approaches: manual evaluation by language experts and automatic evaluation using standard metrics—Bilingual Evaluation Understudy (BLEU), TER (Translation Edit Rate), METEOR (Metric for Evaluation of Translation with Explicit Ordering), and WER (Word Error Rate). Although RBMT achieved the highest overall scores in both automatic and manual evaluations, expert analysis revealed that translations produced by NMT and SMT exhibited less ambiguity. The study concludes that the performance of SMT and NMT systems are likely to improve further with the availability of larger bilingual parallel corpora.