Enhancing English-Persian Neural Machine Translation with a Large-Scale Parallel Dataset and Relative Position Representations

Transformer-based models have revolutionized neural machine translation (NMT), particularly with the introduction of the encoder-decoder architecture. However, training these models effectively often requires large amounts of parallel data or pretraining on massive unlabeled corpora. In the context of English-Persian translation, the lack of extensive parallel datasets has hindered progress. To address this, we introduce a new dataset of 4 million English-Persian parallel sentences that span various topics. Without any pretraining on unlabeled data, our model achieves a BLEU score of 47 on the PEPC benchmark and 35 on the MIZAN benchmark, demonstrating strong performance. We used Transformers with relative position representations, enabling the model to generalize to sequence lengths not seen during training. To promote further research and reproducibility, we have open-sourced both the dataset1 and the trained model2, supporting advancements in English-Persian NMT.