Deep neural models tremendously improved Machine Translation systems, and these systems are translating an immense amount of text every day. The automatically translated text finds a way to Translation Memories and storing these translation units in a TM is not ideal. In this context, we investigate whether distinguishing neural from human translations is feasible. To the best of our knowledge, this is the first time extensive experiments are reported on distinguishing human from machine translation produced by a state-of-the-art neural MT systems (in-domain and out-of-domain) under both monolingual and bilingual settings. Also, this research is one of its kind to deploy transformers-based language models for this task. We achieve an accuracy of 80% on several test sets (in-domain and out-of-domain), which is way above our random baseline.

See recording of this presentation here: https://drive.google.com/file/d/1miE7yvHJ4WQeuTOmvSI6BmS4O_Nj7wEM/view?usp=sharing