Translationese and Translation Universals in Machine Translation

Antonio Toral - Faculty of Arts - University of Groningen

 

 

Translationese and Translation Universals in Machine Translation

Abstract:

Machine Translation (MT) has improved dramatically since the recent adoption of neural networks. Given the high quality attainable by MT nowadays, I deem it timely to go beyond the common practices used in MT evaluation to date, for which I propose to borrow concepts from Translation Studies. This talk presents two such cases:

1. Translationese, which is commonly present in the source side of the test sets that are used to evaluate MT systems. Our research shows evidence that (i) such use of translationese results in inflated human evaluation scores for MT systems; (ii) in some cases system rankings in shared tasks change and (iii) the impact translationese has on a translation direction is inversely correlated to the translation quality attainable by state-of-the-art MT systems for that direction.

2. Translation universals and laws of translation: simplification, normalisation and interference. We compare human translations (HT), raw MT and post-edited MT (PE) and find out that MT and PE are simpler and more normalised and have a higher degree of interference from the source language than HT. These findings suggest that the wide adoption of MT and PE could have a negative impact on the target language.

If you would like to meet with the speaker, please contact Cristina España.