Better N-best Translations through Generative n-gram Language Models