Transliterating Urdu to Roman Urdu using neural network

Transliteration is a process of converting a word from the alphabet of one lan-guage to another language. Previously used techniques are related to statistical, phrase level and rule-based approaches but Neural Machine Translation (NMT) has replaced it because of its heterogeneous, scalable and dynamic structure. NMT is used for rich resource languages e.g. German, Vietnam, Chinese and also performing well for poor resource languages like Myanmar, Hindi, and Roman-Urdu. Urdu is a low resource language and there is no sig-nificant work done to transliterate using NMT models. In this paper, we are working on Urdu to Roman-Urdu transliteration using sequence-to-sequence and attention-based models. This model uses the Encoder-Decoder architecture that takes one language as input (source) and decoder transforms it to desire output (target). Its results are phenomenal for rich resource languages, provid-ing context aware and scalable solutions. In the field of language translitera-tion, Long Short-Term Memory (LSTM) and Bi-directional models are com-monly employed to effectively deal with long-term dependencies. To handle unseen data, a combination of Byte Pair Encoding (BPE) and subword tech-niques is utilized, employing a hybrid approach that incorporates both word and character level embeddings. In order to evaluate the performance of the transliteration system, experiments are conducted on a parallel corpus consist-ing of 60k samples, which were generated from scratch. The system undergoes extensive testing to fine-tune the hyperparameters, ultimately achieving state-of-the-art results measured by the BLEU score on both the training and testing datasets. Additionally, the NMT model provides scalable, robust, context aware structure and can handle out-of-vocabulary (OOV) words.