How MT errors correlate with postediting effort: a new ranking of error types
Kevin Hu · Asia Pacific Translation and Intercultural Studies · 2020
As a result of the rapid development of translation technologies such as machine translation (MT), the global translation industry has inevitably entered a stage of intensive human-machine interaction. A clear manifestation of this is the industrial shift from the traditional workflow of translating from scratch to a new paradigm in which translations are produced by postediting, i.e., translating by fixing the errors in machine-translated texts). Recently, this shift has been expedited by the advent of neural machine translation, a new MT approach that delivers higher output quality in various domains and language pairs. However, although the core of postediting is the identification and correction of the errors generated by MT systems, more research is still in need on how MT errors correlate with postediting effort (the human effort demanded to postedit a machine-translated text). This paper reports on a preliminary study that examines the correlation between translation error types and cognitive postediting effort using empirical data on participants’ pauses in the postediting process. It is found that cognitive effort expended on an error is positively correlated with the stretch of text in which the error is embedded.