Samuel Murray wrote:
So, if I understand correctly, the QE designer would figure out which things make a sentence less likely to be translated well, based on the language combination. Right?
So for example, if it is known that sentences in Upper Slobovian that contain many nouns will likely produce a poor translation in Mid Slobovian but likely produce a reasonably accurate translation in Lower Slobovian, then such sentences would be graded with a low QE score if the target language is Mid Slobovian (which means that they'll go to the top of the pile of segments to be checked by a human) but a high QE score if the target language is Lower Slobovian (which means they'll move to the bottom of that pile). Do I understand this correctly?
So QE is not an evaluation of an existing translation by comparing it with a human translation, but a guess about the quality of the translation based on the elements of sentences in a given language combination. Yes?
Thanks for your interest, Samuel. Here's a very simplistic explanation: QE is based on language combination, yes. Now, the QE designer, so to call it, is usually a scientist - not a lot of linguists can deal with this. The main reason is that it's a complex process. There are standard set of features (the features are what indicate the potential presence of an issue in a sentence) and there are more specific ones. Knowing which features you want to use, a QE model is created using supervised machine learning on parallel data (source, target). So these features are "learned" from already translated text. That model can then be create to estimate the quality of other texts for which you don't have a reference translation.
What you say is correct - QE is only a prediction of the quality, not an evaluation of it. However, we always need to keep in mind that with language, 1+1 is not always 2. I did some work on quality estimation based only on linguistic features, with the sole purpose of making QE more accessible to translators and linguists. I think I posted links in this forum.
If you want to learn more, let me recommend some papers:
Machine Translation Quality Estimation Across Domains - http://www.aclweb.org/anthology/C14-1040
Confidence estimation for Machine Translation - http://web.eecs.umich.edu/~kulesza/pubs/confest_coling04.pdf
Estimating the Sentence-Level Quality of Machine Translation Systems - http://www.mt-archive.info/EAMT-2009-Specia.pdf