“…Conventional MT systems, be they phrase-based, n-gram based, syntax-based or hierarchical, are typically trained in two steps: the first step (training) estimates individual features functions; the second one (tuning) learns to combine these features so as to optimize translation quality, for instance using Minimum Error Rate Training (MERT) (Och, 2003). The limitations of MERT, notably its inability to train feature sets containing more than a dozen of features, have long been reported, and more effective discriminative training procedures have been sought (see (Neubig and Watanabe, 2016) for a recent review).…”