The architecture I used is completely different from what is described here, but the goals are very similar. I had to handle any curse word in any language, including curses from one language translated into another language, as well as offensive phrases, and their translated equals, as well as offensive slang, and mispelt offensive slang translated from other languages.
I ended up with an offense dictionary of about 700K words and phrases. This was back in '99, so my memory may not be 100% here, but I remember using Perfect Hash to generate a compiled hash table for the dictionary, and then a trie to organize the dictionary lookups. The entire system was about 150K of a downloaded exe to access the NHL simulcast chat, as all the offensive language filtering occurred on the client side. Chatting anything that could be offensive turned into a series of words with their interiors all asterisk '*', and it ran in something like 500 ms. Fun times. That company and project died with the dot com bust.