Urdu to English-Based Unsupervised Machine Translation

Machine Translation (MT) is a subfield of computational linguistics. The two primary state-of-the-art approaches in MT are Neural Machine Translation (NMT) and Statistical Machine Translation (SMT), both of which require large corpora for training. Urdu counts in low-resource languages due to the fewer resources available for computational work. To build a good translation system, the resources available are not enough. Many languages present in the world have a different structure. Like in Urdu and English, the Urdu structure is based on Subject-Object-Verb (SOV), and the English structure is based on Subject-Verb-Object (SVO). This paper proposes a new method of Urdu-to-English unsupervised translation to overcome one of the biggest obstacles in low-resource languages. The importance of this work is found in the lack of remedies in case of no parallel corpora, especially in structurally differing languages. Possible uses of this model can be extended across integrating two languages, particularly Urdu and English for producing NLP applications such as hate speech detection, opinion mining, and information search in multilingual contexts. We used Tanzil Corpus for Unsupervised Neural Machine Translation model training. From the on-the-fly backtranslation model we obtain the BLEU score of 5.21 and with the supervised technique obtain the BLEU score of 13.85 for Urdu to English language pair. We faced many problems, especially in organizing the text alignments due to the difference in language structures of Urdu (SOV) and English (SVO). However, the preprocessing required in this case was the tokenization, the true casing, and the punctuation differences between the languages.