Phrase table pruning for Statistical Machine Translation

Phrase-Based Statistical Machine Translation systems model the translation process using pairs of corresponding sequences of words extracted from parallel corpora. These biphrases are stored in phrase tables that typically contain several millions such entries, making it di cult to assess their quality without going to the end of the translation process. Our work is based on the examplifying study of phrase tables generated from the Europarl data, from French to English. We give some statistical information about the biphrases contained in the phrase table, evaluate the coverage of previously unseen sentences and analyse the e ects of pruning on the translation