Chapter 5: Machine Translation Evaluation and Optimization

The evaluation of machine translation (MT) systems is a vital field of research, both for determining the effectiveness of existing MT systems and for optimizing the performance of MT systems. This part describes a range of different evaluation approaches used in the GALE community and introduces evaluation protocols and methodologies used in the program. We discuss the development and use of automatic, human, task-based and semi-automatic (human-in-the-loop) methods of evaluating machine translation, focusing on the use of a human-mediated translation error rate HTER as the evaluation standard used in GALE. We discuss the workflow associated with the use of this measure, including post editing, quality control, and scoring. We document the evaluation tasks, data, protocols, and results of recent GALE MT Evaluations. In addition, we present a range of different approaches for optimizing MT systems on the basis of different measures. We outline the requirements and specific problems when using different optimization approaches and describe how the characteristics of different MT metrics affect the optimization. Finally, we describe novel recent and ongoing work on the development of fully automatic MT evaluation metrics that have the potential to substantially improve the effectiveness of evaluation and optimization of MT systems. Progress in the field of machine translation relies on assessing the quality of a new system through systematic evaluation, such that the new system can be shown to perform better than pre-existing systems. The difficulty arises in the definition of a better system. When assessing the quality of a translation, there is no single correct answer; rather, there may be any number of possible correct translations. In addition, when two translations are only partially correct – but in different ways – it is difficult to distinguish quality. Moreover, quality assessments may be dependent on the intended use for the translation, e.g., the tone of a translation may be crucial in some applications, but irrelevant in other applications.