Levantine Arabic Machine Translation - Subjective Ranking

Larger numbers represent outputs preferred more often in a randomized A/B/C comparison.

Results are based on the average human preference scores of 20 randomly selected short strings of Levantine Arabic social media content. Content was translated into English by a Human translator, TWB's dialect adapted MT system, and Google's general Arabic MT system. Each string was randomized and ranked by five translators and the results were averaged. 0 = least preferred; 1 = second preferred; 2 = most preferred.
Source: Translators without Borders