Word Embeddings at Post-Editing

Marcio Lima Inácio, Helena de Medeiros Caseli

Abstract: After more than 60 years of research in Machine Translation, it has not been possible yet to develop a perfect fully automatic translation system for unlimited purposes. Thus, it is still necessary post-editing to correct possible mistranslations output by the Machine Translation system. Several approaches have been proposed in order to also automate the post-editing task. This work addresses one of the main steps of an automatic post-editing tool: the automatic proposition of word replacements for a Machine Translation output. To do so, we propose a novel method based on bilingual word embeddings. In the experiments present in this paper we show the effectiveness of this approach in two of the most frequent lexical errors: ‘not translated word’ and ‘incorrectly translated word’.