Establishing semantic relationships across languages using word vectors

摘要

In this paper, we improvise an existing word vector based machine translation system to measure semantic relatedness across languages and apply it to the language pair of English and Hindi. We also evaluate the model using human scored word relatedness datasets. Unlike most systems performing a similar task, the system does not make use of parallel corpora, which are cumbersome and not practical to build between all possible pairs of languages. The system learns a linear transformation between multidimensional word vector spaces of a pair of languages using a set of known translations, called a bilingual dictionary. An approach to reduce the effort in building a usable bilingual dictionary in the system is also proposed.