Translation Table Using von Mises-Fisher Distributions: a directional statistics approach to machine translation
Taylor Turpen · Brandeis Institutional Repository · 2013
This paper describes the building and evaluation of a translation table using the von Mises-Fisher directed statistical distribution. We propose three models to create a directional source to target language translation table. We evaluate two of these translation tables against a baseline multinomial statistical translation table built using the Giza++ open source software package. The best source to target word performance using our model is 6% in the best match category, which is the same as baseline accuracy. The benefits of a directed statistical approach over a traditional translation table are leveraged translation models on minimal data and an abundance of close-match word translations.