Optimizing Tokenization Choice for Machine Translation across Multiple Target Languages

Abstract

Tokenization is very helpful for Statistical Machine Translation (SMT), especially when translating from morphologically rich languages. Typically, a single tokenization scheme is applied to the entire source-language text and regardless of the target language. In this paper, we evaluate the hypothesis that SMT performance may benefit from different tokenization schemes for different words within the same text, and also for different target languages. We apply this approach to Arabic as a source language, with five target languages of varying morphological complexity: English, French, Spanish, Russian and Chinese. Our results show that different target languages indeed require different source-language schemes; and a context-variable tokenization scheme can outperform a context-constant scheme with a statistically significant performance enhancement of about 1.4 BLEU points.