The (oversimplified) problem with machine translation is that a simple Dictionary 2 Ditctionary solution is not sufficient.
Words put into phrases, changes in 'tense', and thus varrying forms of conjugation across languages all come into play. (and as mentioned dialects, slang, and general 'misspellings')
An XML 'language' object database for each known lauguage will eventually be the key I feel.
Even then it would take an immense amount of logic to breakdown one language and build it up into another in realtime.. that is if such xml databases existed. It could be done on YOUR PC, but would be expensive to run a website that offered such a service to any large number of people.
IE: You can't just translate 'words', you need to translate meaning which single words are only part. (This could be done through an xml type defintion of each language)
Functionally, might take the form of recursive logic with feedback. Also known as a Neural Network.
From inside out... translate word. But thats not enough, take output, use as input and need to translate words 'meaning' in sentance. Then keep working outward into paragraph if necessary, and finally article.
In that case the 'dictionary' would just be a VERY WELL 'trained' Neural Network (one for each language)