This paper explores the use of external sources of bilingual information available on-line for word-level machine translation quality estimation (MTQE). These sources of bilingual information are used as a black box to spot sub-segment correspondences between a source-language (SL) sentence S to be translated and a given translation hypothesis T in the target-language (TL). This is done by segmenting both S and T into overlapping sub-segments of variable length and translating them into the TL and the SL, respectively, using the available bilingual sources of information on the fly. A collection of features is then obtained from the resulting sub-segment translations, which is used by a binary classifier to determine which target words in T need to be post-edited. Experiments are conducted based on the data sets published for the word-level MTQE task in the 2014 edition of the Workshop on Statistical Machine Translation (WMT 2014). The sources of bilingual information used are: machine translation (Apertium and Google Translate) and the bilingual concordancer Reverso Context. The results obtained confirm that, using less information and fewer features, our approach obtains results comparable to those of state-of-the-art approaches, and even outperform them in some data sets.
展开▼
