Abstract

After more than 60 years of research in machine translation (MT), the resulting technology has finally begun to be adopted for purposes beyond merely gisting. Many factors contributed for this significant change in the way MT is perceived and used, including noticeable improvements in translation quality. Evaluation campaigns (such as those organized by NIST ) have been showing steady progress over the years for certain language-pairs. Industry-focused experiments have reported significant gains in productivity by using MT systems as part of their translation workflow, as well as significant cost reductions by using MT instead of humans for inbound translation (e.g., internal communication) or non-crucial content such as product reviews.

The decision on whether or not to adopt MT for a given purpose is directly connected to the level of quality that is expected by users. Different groups of users prioritise different aspects of translation quality (e.g. meaning over grammar). However, a metric of quality that accommodates these different needs is yet to be defined. In this talk I will introduce an initiative towards formalising the notion of quality for human and machine translation as part of QTLaunchPad, an EC coordination and support action, and describe some of my recent work on estimating some aspects of translation quality through automatic evaluation metrics that do not rely on human references.