Machine Translation Post-editing and Effort : Empirical Studies on the Post-editing Process
Maarit Koponen · Työväentutkimus Vuosikirja · 2016
This dissertation investigates the practice of machine translation post-editing and the various aspects of effort involved in post-editing work. Through analyses of edits made by post-editors, the work described here examines three main questions: 1) what types of machine translation errors or source text features cause particular effort in post-editing, 2) what types of errors can or cannot be corrected without the help of the source text, and 3) how different indicators of effort vary between different post-editors. The dissertation consists of six previously published articles, and an introductory summary. Five of the articles report original research, and involve analyses of post-editing data to examine questions related to post-editing effort as well as differences between post-editors. The sixth article is a survey presenting an overview of the research literature. The research reported is based on multiple datasets consisting of machine translations and their post-edited versions, as well as process and evaluation data related to post-editing effort. The dissertation presents a mixed methods study combining qualitative and quantitative approaches, as well as theoretical and analytical tools from the fields of language technology and translation studies. Data on edits performed by post-editors, post-editing time, keylogging data, and subjective evaluations of effort are combined with error analyses of the machine translations in question, and compared for various post-editors. The results of this dissertation provide evidence that, in addition to the number of edits performed, post-editing effort is affected by the type of edits as well as source text features. Secondly, the results show that while certain language errors can be corrected even without access to the source text, certain other types that more severely affect the meaning cannot. Thirdly, the results show that post-editors' speed and the amount of editing they perform differ, and that various profiles can be identified in terms of how the edits are planned and carried out by the post-editors. The results of this work may have both theoretical and practical implications for the measurement and estimation of post-editing effort.