Neural Machine Translation and Post-editing: Improving translation of domain specific English terms for a low resource language

Abstract Neural Machine Translation (NMT) has demonstrated salient enhancement and quality output in various aspects, but issues remain when translating specific terms across domains when data are heterogeneous or dealing with rare phrases. To explore this, the translation system for the language pair: English and Persian, was considered, since low resource languages like Persian require more effort and attention to provide accurate translation when dealing with texts in specific domains. An aligned parallel terminology database was created in the Information Technology domain, where all terms were manually annotated. Data were selected from a collection of parallel sentences from the OPUS repository. Terms in the source and the target language were identified and annotated manually. After training an NMT model, the system was manually evaluated on the effectiveness of the application of instance selection to retrieve sentences. Its use led to an overall enhancement in post-editing, thereby achieving a better translation. Besides gaining insights into errors and system performance, this method provides guidance for setting priorities for future extensions and improvements for machine translation. It also sheds light on the use of instance selection across different scenarios in relation to low resource languages like Persian to achieve an overall high quality term translation accuracy.