Shining Lights

A question I hear a lot these days is, "How's work?" It always leaves me sort of tongue-tied. I mean, work's work, I guess. But here's another stab at an answer.

I worked for around a year, up until last summer, on "surface realization" (which amounts to putting together sentences in natural language). Since I came to Bosch my focus has been officially switched to "user modeling" (which amounts to predicting what a human user knows/prefers/needs/etc.). There's been a lot of culture shock and stuff moving to Bosch, but I think I finally put my finger on why my user modeling research has been (by and large) so slow and unfulfilling, in contrast with my earlier work.

I have no inspiration. I don't mean no inspiration at all, or no inspiration to do research; I have plenty of that. But in my earlier work in surface realization, solid papers were thick on the ground--the kind that get at central issues, decompose the problem effectively, apply well-established techniques, and present meaningful experimental results. I had heroes who had gone before--minor demigods in the academic pantheon, perhaps, but heroes nonetheless: Ratnaparkhi 2000, or Kay 1996, or (a little further afield, but excellent) Klein and Manning 2003. (Note: I never noticed until just now that one of my computational linguistics muses has the exact same name as my improv coach. Crazy.) These are papers that lift your spirits: you see a sense of continuity and building on one another's work--you see a recognition of the underlying structures that unite the topic with the rest of the discipline. Surface realization is a perilous subject in a lot of ways (evaluation is notoriously slippery, for one thing), but there is good inspiring exploration going on.

Not so for user modeling. It's a jungle. The literature abounds with sketchy and semi-apocryphal guidebooks: Fantastick Travels in the New Worlde (with Brilliante and Statistickal Illustrations). Papers tend to present entire systems in their manifold glory, which is nearly useless for gleaning the good ideas from the hacks. It's an isolated territory cut off from the rest of the discipline--it's even unclear which discipline that would be. It's a savage untamed land without corpora, without evaluation metrics--and thus far, without the brilliant lights of inspiration I have sought.

What's interesting, though, is recognizing that this is important to me, and what this says about my adventuresome spirit. Apparently the uncharted wilderness doesn't by itself move me to chart it: I am Sir Stanley, and not Dr. Livingstone. I need those who have gone before. Or maybe that's not the real reason at all. Maybe I need examples of excellence not just to inspire, but to instruct: without them I really don't know what I'm striving toward. I don't know what good user modeling research is until I've seen it.

(Either way, perhaps this is one reason I'm inspired to study philosophy--a subject with a greater cloud of witnesses than almost any other.)