Grant: Discovery Grants Program - Individual

The main concern of our proposal research program is the development of new methods and new approaches for known methods in the field of statistical machine translation (SMT). These methods are named adaptive and hybrid methods. Adaptation in SMT refers to the tuning of the log-linear model weights (LMs, TMs, ...) such that a sentence can be translated in a better way by software implementing this model. In previous experiments, we have already shown the existence of those desired sentence-dependent weights but have also shown the huge difficulty to exhibit them. Some improvement can be done in this direction. In SMT, one will call hybrid methods those that are combination of two or more methods. The difficulty here lies in finding criterions allowing to decide in an automatic way which method should be applied to translate a specific sentence. We have already developed such criterions but a lot of improvements and attempts need still to be done in this direction too. Hence the development of these general techniques in combination with already known methods, methods to be refined (like the new method we developed called back-translation) and methods still to be discovered is the core of our proposal and is also certainly a long-term research. Some of these methods are new and promising and the ideal expected goal of our works is certainly to improve translations provided by SMT systems, though this problem is an extremely complex one.