Abstract:

This article, one the few publications on Machine Translation (MT) ever written in Italian, appeared in the Italian journal of cognitive sciences and artificial intelligence Sistemi Intelligenti, which promotes a multidisciplinary approach to natural or artificial intelligent systems, calling for contributions by biologists, computer scientists, engineers, linguists, logicians, philosophers, psychologists, social scientists and others interested in the field. Consistently with such an approach, the article does not report on software technicalities of specific translation systems available on the market, but investigates the foundations of MT as a branch of computational linguistics and artificial intelligence. From the author's theoretical viewpoint (which nevertheless does not exclude practical applications), MT is a discipline dealing with the formal, computational modelling of (i) the linguistic competence of bilingual or multilingual subjects and (ii) the kind of performance required for translating between the languages they know.

In particular, this work discusses linguistic models adopted within the historical evolution of MT since the end of Second World War and points out parellel developments or divergences with respect to the contemporary trends in theoretical linguistics. Applications to MT of phrase structure grammars, transformational grammars, dependency grammars, categorial grammars and unification-based grammars are described and evaluated, accounting for the influence of scientists like R. Carnap, Y. Bar-Hillel, N. Chomsky, L. Tesnière, R. Montague, M. Kay, J. Bresnan and R. Kaplan, among others. Some discussion is also devoted to classical, vexed questions concerning the (non-)feasibility of fully automatic high-quality translation (in absence of world knowledge), or the architectural alternative between MT systems based on a language-independent interlingua and those exploiting bilingual transfer. The final section of the article introduces the readers to more recent trends, combining unification-based grammars with hybrid stratificational architectures that include both interlingual and transfer components, like MT research applying Bresnan and Kaplan's Lexical-Functional Grammar, or the Eurotra project in which the present author partecipated (see Curriculum Vitae).