FWIW, this strikes me as a pretty good match for Parse::RecDescent. I'd suggest ordering top level named grammar parts from ideal-as-expected to worst case. Probably wouldn't take much work to catch 90% of the input (depending on how loose it really is).

If you provide sample input and show what you're trying and trying to do, you will likely get a lot of help here with it. Also, don't do anything illegal. Make sure you have the right to parse/save the chat data.