Aligning contextual vector spaces between independent neural translation systems

Abstract

Numerous pre-trained machine translation models are available for translating between different languages. However, these models are limited to a fixed set of languages they were trained for. When there is no translation model available for a specific language pair, we need to translate to one or more intermediate languages, which can result in reduced translation quality. We investigate the possibility of combining two translation models by aligning the vector spaces between them using a simple regressor. We explore the effectiveness of various regression methods for achieving this alignment and evaluate their performance. We show that combining two different translation models is possible, although doing so leads to a decrease in translation quality.