Because going between any two languages is a separate task. Unless you have a universal language that contains every linguistic feature everywhere and can achieve lossless transition, using X->Y and X->Z to make Y->Z will introduce an unacceptable level of error. English is not a universal language.
My Natural Language Systems professor just went through that topic yesterday -- that in the case of all machine translation applications, the ultimate goal is to get as high up in a 'natural language pyramid' where the tip is a single universal language. So we get the most use out of elements of any two given languages that have as much in common with each other as possible; these common elements could then be applied to other similar languages.
Interestingly, he also proposed that the closest thing we have right now to that 'tip' language is sanskrit.
In the 80's the Wang corporation (of Wang wordprocessing fame) was working on using Aymara as such a bridge language. I don't know what happened to the project.
