Research Insider: «Distributional Semantics and Compositional Translation»
The current strategies in statistical and neural machine translation rely on the use of parallel and aligned corpora, as well as on non-compositional phrase segmentation. However, other strategies are possible, namely the use of non-parallel bilingual corpora and compositional segmentation; a new translation method that requires building the contextual meaning of sentences by making use of distributional models in a compositional way.
In this talk, the key aspects and the main applications of compositional distributional semantics will be explored. For instance: how to build the composite meaning of "red car" from the senses of its constituent words if each word is defined as a vector of linguistic contexts in a distributional model? Special attention will be paid to a new method for machine translation based on distributional models and compositional semantics. From this point of view, translating is the process of paraphrasing in a bilingual space.