Comparative Human and Automatic Evaluation of Glass-Box and Black-Box Approaches to Interactive Translation Prediction

Abstract

AbstractInteractive translation prediction (ITP) is a modality of computer-aided translation that assists professional translators by offering context-based computer-generated continuation suggestions as they type. While most state-of-the-art ITP systems follow a glass-box approach, meaning that they are tightly coupled to an adapted machine translation system, a black-box approach which does not need access to the inner workings of the bilingual resources used to generate the suggestions has been recently proposed in the literature: this new approach allows new sources of bilingual information to be included almost seamlessly. In this paper, we compare for the first time the glass-box and the black-box approaches by means of an automatic evaluation of translation tasks between related languages such as English–Spanish and unrelated ones such as Arabic–English and English–Chinese, showing that, with our setup, 20%–50% of keystrokes could be saved using either method and that the black-box approach outperformed the glass-box one in five out of six scenarios operating under similar conditions. We also performed a preliminary human evaluation of English to Spanish translation for both approaches. On average, the evaluators saved 10% keystrokes and were 4% faster with the black-box approach, and saved 15% keystrokes and were 12% slower with the glass-box one; but they could have saved 51% and 69% keystrokes respectively if they had used all the compatible suggestions. Users felt the suggestions helped them to translate faster and easier. All the tools used to perform the evaluation are available as free/open–source software.