to post-edit or to translate ... That is the question: a case study of a recommender system for Quality Estimation of Machine Translation based on linguistic features
Ona De Gibert Bonet · Communities in ADDI (Universidad del Pais Vasco) · 2018
[EN]The implementation of a machine translation system into production is not enough to warrant its efficient use. There exists the need to know when it is profitable to use machine translation as opposed to translating from scratch. That is why being able to estimate the quality of a machine translation is crucial. This thesis investigates the task of quality estimation of machine translation for a specific machine translation system and a specific domain by developing a recommender system for Spanish to English. The work further investigates how quality estimation can benefit from the use of linguistic characteristics in contrast to the more common shallower features. The data was collected from real translators who performed a post-editing task, and the linguistic features were manually annotated. First, we build a classification model that selects sentences for post-editing or translating. Secondly, we perform a regression task based on three quality indicators: Quality, Time and HTER. Although experimentation shows some promising results, overall the selected features are not discriminative enough for the recommender system to be implemented into production. Results are discussed at different levels, suggesting a replication at a larger scale, with automatic annotation of informative linguistic features.