Abstract
Machine translation produces marginal accuracy rates for low-resource languages, but its deep learning model expects to yield improved accuracy with time. This longitudinal study investigates how Google Translate's Urdu-to-English translated output has evolved between 2018 and 2021. Accuracy and acceptability of the translations have been determined by, a) an interlinear gloss that identifies core semantic units and grammatical functions to be translated and, b) a descriptive comparison of the translated text's syntactic and semantic properties with those of the source text. Overall, despite a 50 % error rate that persists over the three-year interval, the research reports significant improvement in the overall intelligibility of the translations, in contrast to initial results from 2018, which exhibited rampant non-localized errors. Working backwards from instances of errors to morphosyntactic and semantic patterns underlying them, the study concludes that the pro-drop feature of Urdu, Urdu's case-marking system, identification of clause boundaries, polysemous terms, and orthographically similar words pose the greatest difficulty in neural machine translation. These results point to the need for incorporating syntactic information in training data.