The translation between languages is necessarily lossy - some of the authors intention (in terms of style and meaning) is lost. By creating a dataset that maps different translations to a source, a more configureable mapping can be provided that potentially captures the original author’s intention.