Keynote Speech (Ⅲ) Statistical Machine Translation without Explicit Linguistic Structures

摘要

Different linguistic structures exist at various levels of natural language. A word is composed of a lexeme and a number of morphemes. A phrase consists of multiple words. Multiple phrases form a sentence. A paragraph is a sequence of more than one sentences, and so on. These structures, be them syntactic or semantic, facilitate our analysis of natural languages. It is however unclear whether these structures are necessary for machine translation. In this talk, I will present some of the latest research in neural machine translation, where nearly no such linguistic structure is being exploited, however, with comparable to, or often better than, many existing machine translation systems. The goal of this talk is not to strongly assert that those linguistic structures are not necessary, but to stimulate active discussion on this issue.