Abstract
Translation Quality Assessment (TQA) conducted by human translators is a process widely used, both in estimating the increasingly used Machine Translation performance and in finding an agreement between customers and translation providers in translation industry. While translation scholars are aware about the importance of having a reliable way to conduct the TQA process, it seems that there is limited literature facing the issue of reliability with a quantitative approach. Here we consider the TQA as a complex process, considering it from the physics of complex systems point of view, and we face the reliability issue with a Bayesian approach. Using a dataset of translation quality evaluations, in an error annotation setting, entirely produced by the Language Service Provider Translated Srl, we build two Bayesian models that parameterise the features involved in the TQA process, namely the translation difficulty, the characteristics of the translators involved in producing the translation and assessing its quality (reviewers). After validating the models in an unsupervised setting, showing that it is possible to get meaningful insights about translators even with just one review per translation job, we extract information about the translators and reviewers and we show that reviewers reliability cannot be taken for granted even if they are expert translators: the translator's expertise could induce also a cognitive bias when reviewing a translation produced by another translator. The most expert translators, though, show the highest level of consistency, both in the task of translating and in the one of assessing translation quality.
Abstract (translated)
URL
https://arxiv.org/abs/2203.07135
