Informative Manual Evaluation of Machine Translation Output

Proceedings of the 28th International Conference on Computational Linguistics, pages 5059–5069 Barcelona, Spain (Online), December 8-13, 2020 5059 Informative Manual Evaluation of Machine Translation Output Maja Popovi´c ADAPT Centre, School of Computing Dublin City University, Ireland maja.popovic@adaptcentre.ie Abstract This work proposes a new method for manual evaluation of Machine Translation (MT) output based on marking actual issues in the translated text. The novelty is that the evaluators are not assigning any scores, nor classifying errors, but marking all problematic parts (words, phrases, sentences) of the translation. The main advantage of this method is that the resulting annotations do not only provide overall scores by counting words with assigned tags, but can be further used for analysis of errors and challenging linguistic phenomena, as well as inter-annotator disagreements. Detailed analysis and understanding of actual problems are not enabled by typical manual evaluations where the annotators are asked to assign overall scores or to rank two or more translations. The proposed method is very general: it can be applied on any genre/domain and language pair, and it can be guided by various types of quality criteria. Also, it is not restricted to MT output, but can be used for other types of generated text. 1 Introduction While automatic evaluation metrics are very important and invaluable tools for rapid development of ma- chine translation (MT) systems, they are only a substitution for human assessment of translation quality. Various methods have been proposed and used for the human evaluation of MT output, such as (ALPAC, 1966; White et al., 1994; Koehn and Monz, 2006; Vilar et al., 2007; Graham et al., 2013; Forcada et al., 2018; Barrault et al., 2019), and all of them are essentially based on some of the following three quality criteria: adequacy (accuracy, fidelity), comprehensibility (intelligibility) and fluency (grammaticality). Adequacy measures how well the meaning of the original text is conveyed to the translated text. Com- prehensibility reflects how well a reader is able to understand the translated text without access to the original text. Fluency describes the grammar of the target language in the translated text. The choice of criteria often depends on the task and on the purpose of the translation: for example providing transla- tions to users requires high comprehensibility and adequacy, comparing different MT systems is based mainly on adequacy, etc. Regardless of the quality criterion, the annotators are usually asked to assign an overall quality score for the given MT output, or to rank two or more competing outputs from best to worst. One advantage of direct scoring over ranking is that the resulting annotations include not only the information that one MT output is better than another, but also the degree to which that output was better than the other. Also, it gives an estimation about the absolute quality of each MT output. Another advantage is that direct estimation of translation quality extends the usefulness of the annotated data to other tasks such as quality estimation. Still, neither of these two annotation methods provides …

Content truncated to 500 words