Interlinguas: Deep and Shallow

In 1629, Descartes proposed a “language of true philosophy” to serve as an interlingua for translating between languages. Over 300 years later, “semantic interlingua” appears on the top of the Vauquois triangle, as the deepest possible analysis guaranteeing the best possible translation. But the main stream of machine translation has considered the interlingua unrealistic and worked on lower levels of the Vauquois triangle, such as syntactic and lexical transfer. However, the interlingua idea has advantages that do not depend on it being a deep semantic representation. An interlingua makes it possible to build highly multilingual systems without a quadratic blow-up of size. It also enables transfer of information, for instance, from high to low resourced languages; a related idea has recently been exploited in the Universal Dependencies project, which uses a shared set of labels and tags as a cross-lingual representation. Grammatical Framework (GF) is a formalism that was designed for building interlingua-based multilingual grammars. Its original purpose was to enable special-purpose interlinguas precisely capturing the semantics of different domains, such as mathematics or touristic phrases. However, GF also enables interlinguas that are not so deep. They can be based on surface syntactic structures or just chunks of words. Recent developments of this idea have led to a translation system that currently works for all 182 pairs of 14 languages, ranging from English and German to Finnish and Chinese. This system has a stack of interlinguas, where a semantic layer produces high-quality translations whenever the input can be analysed by it, whereas the syntactic and chunk-based layers guarantee the robustness of the system. The interlingual grammar makes the system very compact in size, so that it can be run off-line on mobile devices.