I know SFA about NLP, but I'd imagine that Perl would be quite suitable: it has excellent text segmentation support, which I'd guess would be useful. (Although I'd think semantic structure would be more important than text chunking... argh. Too much image processing in brain.)

As far as Perl's "suitability" goes: Perl's really good for hashing out proof-of-concept code and dealing with text, which would (likely) make it good for some aspects of AI programming. By the same token, Lisp is good for, say, genetic programming (since its code/data separation is even thinner than Perl's), Prolog's good for planning/reasoning (because its logical syntax makes encoding knowledge easier), and C's good for search-based and general number-crunching (neural nets, anyone?), because, well, you can write some bloody fast C code.

In my (limited) experience, a lot of AI involves a lot of number crunching, which Perl isn't so good at. (Not that it can't crunch numbers, but it can't crunch them fast enough. And let's face it: you're going to be a happier camper if your program takes one second to run than if it takes five minutes, especially when you're tweaking parameters.) I'd be most likely to write a prototype in Perl (provided Lisp, Prolog, and Haskell are unsuited for the job), make sure that I'm on the right track, then recode in C for the efficiency win. As always, YMMV.