The dataset is curated for the purpose of evaluating the robustness of Neural Machine Translation (NMT) towards natural occuring noise
(typo, slang, code switching, etc.). The dataset is crawled from Twitter, then pre-processed to obtain sentences with noise.
The dataset consists of a thousand noisy sentences. The dataset is translated into Chinese manually as the benchmark for evaluating the robustness of NMT.
