The scaling problem in the pattern recognition approach to machine translation

Statistical machine translation (SMT) has proven to be an interesting pattern recognition framework for automatically building machine translations systems from available parallel corpora. In the last few years, research in SMT has been characterized by two significant advances. First, the popularization of the so called phrase-based statistical translation models, which allows to incorporate local contextual information to the translation models. Second, the availability of larger and larger parallel corpora, which are composed of millions of sentence pairs, and tens of millions of running words. Since phrase-based models basically consists in statistical dictionaries of phrase pairs, their estimation from very large corpora is a very costly task that yields a huge number of parameters which are to be stored in memory. The handling of millions of model parameters and a similar number of training samples have become a bottleneck in the field of SMT, as well as in other well-known pattern recognition tasks such as speech recognition or handwritten recognition, just to name a few. In this paper, we propose a general framework that deals with the scaling problem in SMT without introducing significant time overhead by means of the combination of different scaling techniques. This new framework is based on the use of counts instead of probabilities, and on the concept of cache memory.