Towards an interlingua for speech translation in limited domains Pour une interlangue utile en traduction automatique de la parole dans des domaines limités

We present a methodology for constructing automatically verifiable and human-readable interlinguas, which simultaneously combine expressivity and simplicity. By defining the interlingua using the same formalism as the one used to define the analysis and generation grammars, it is feasible, in a natural way, both (1) to generate a human-readable gloss for an interlingua representation, (2) to determine whether said representation is well-formed. As well as serving as a pivot language for carrying out translation, we show how the interlingua can be used for two other important purposes: improving the quality of recognition, by filtering out hypotheses which are not semantically well-formed, and simplifying the development cycle, which thanks to the gloss form can be transformed into a disjoint set of monolingual tasks.