The disclosure notably relates to a computer-implemented method for translating a Natural Language Processing (NLP) system given in a source language into at least one target language. The NLP system is based on a lexicalized taxonomy and allows text annotation and classification. The method comprises providing a corpus in the source language. The taxonomy includes annotations allowing determination of the most frequent terms describing a given concept in the corpus. The method further comprises filtering the most frequent terms for each annotation. The method further comprises querying the corpus with the most frequent terms and extracting portions of sentences comprising these terms. The method further comprises tagging the terms in each extracted portion. The method further comprises translating the extracted portions in the at least one target language using a quality machine-translator, thereby obtaining a tagged translation for each portion. The method further comprises normalizing the translations.