Possibility to first train on a big corpus and then on smaller corpora to update the phrase table Our usage in Wipo: launch a SMT on all patents compiled so far (takes between one and 4 weeks of CPU), then update every week with newly published patents.