1949-2019: 70 anni di TA visti attraverso i dati utilizzati

Randy Scansani e Alberto Barrón-Cedeño, Università di Bologna

Abstract: Similarly to other subfields of computing science and artificial intelligence, machine translation (MT) technology has been in constant development since the 1940s, shifting from human- to data-driven approaches. We provide an overview of the evolution of MT in terms of data requirements after each technology shift – from the rule-based and statistical up to the neural paradigm. Whereas in the early days no parallel data was necessary per se, statistical translation requires large amounts of parallel data. Today, an acceptable output can be obtained even without leveraging data for both languages in a combination when using neural models.