improving english-persian neural machine translation system through filtered back-translation method

This study utilizes the neural machine translation (nmt) approach to improve the vru english-persian nmt system. in an nmt system, the encoder takes a sequence of source words as inputs and the decoder takes the source vectors through an attention mechanism as input and returns the target words. as english-persian is a low resource language pair and few researches have been carried out on this language pair, it is important to augment the nmt system with various data. the study explores two methods to enhance the vru system: back-translation and data filtering. at first, we created nmt models using two corpora, amirkabir and persica. to see whether higher ratios of synthetic data leads to decreases or increases in translation performance, we modeled different ratios using back-translation. we found that back-translation significantly improved the vru nmt system. second, the filtering method is applied to eliminate noisy data by applying sentence-bleu, average alignment similarity (aas), maximum alignment similarity (mas), combination of aas and mas, combination of aas, mas, and sent-bleu. results show that the combination of aas, mas, and sent-bleu produced the highest growth, with a bleu score of 30.65. the study concludes that the proposed methods effectively enhance the vru english-persian nmt system.