May 29, 2004

Natural language and artificial intelligence

I hesitate to say a good word for modern software (it might encourage the further production of the sort of bug-ridden word processors I use so much), but I think Craig Silverstein (Google's technology director), in the interview that Mark Liberman recently cited, might actually be underestimating recent successes at simulating aspects of common sense. Try typing "Simmons" and "Beautyrest" into the Google search box (notice, no use of the word mattress) and watch the mattress ads spring up in the margin of your display of search results. If that isn't effective computer simulation of common sense I don't know what would be: the name Simmons is common enough (one might imagine it thinking), but there is a Simmons firm that has trademarked the name Beautyrest, and it's the name of a line of mattress products, so...

The classic dreams of GOFAI (Good Old-Fashioned Artificial Intelligence) may have been quixotic. But the work on impossible projects about replicating common sense (like inferring in context that someone who goes into a restaurant probably intends to eat a meal there, and will be paying for it, and so forth) used programmers who stayed in the business and ended up working on more sensible things that turned out to be considerably more successful. Google's subject-linked advertisements really do quite often turn out to be relevant, even if you do get silly mistakes based on superficial word similarities sometimes. This is not GOFAI, but it is reasonably called AI, and it is not completely brain-dead. I'm fully aware that what's really going on may be just a matter of mattress-sellers including words like Simmons and Beautyrest on the lists of words that they pay to have their ads associated with, which renders it almost stupidly simple; but my point is not that wondrously clever techniques of reasoning are really being used, but rather that it takes so little of this sort of elementary rigging and accessing of files and lists to make it appear that the system is intelligently responsive to one's interests, and the folks at Google are doing it quite well.

And it's not just Google. Yesterday I checked the details of a book on and I noticed that not only did it have some suggestions for me about the usual kinds of books I buy, it also asked me if I would like the book I had just looked up to be delivered to me the next day, and told me that if I placed my order within one hour and ten minutes that could be done. While I stared at this, the screen did a sudden update and "one hour and ten minutes" changed to "one hour and nine minutes". Now, the software accomplishment involved shouldn't be compared to to proving Fermat's last theorem, but looks intelligent. It's a closer approach to helpful, relevant, and timely information than I've had in most phone calls to retailers.

What is notable about these advances in advertising and retailing software, though, is that the symbolic computational linguistics of the 1980s has contributed nothing to them. None of the small wonders of convenience and user-friendliness found on the very best of the commercial websites involves anything you could reasonably call natural language processing.

(I know, I know, boasts of natural language processing, and says its product is "able to understand the context of what you are asking" and can offer "answers and search suggestions in the same human terms in which we all communicate". Puh-lease! can you say pa-thet-ic? Ask Jeeves "Show me some cars that are not Japanese." The results are all about Japanese cars. The NLP claims about AskJeeves appear to be a load of nonsense.)

Perhaps Silverstein is right to talk in terms of it being centuries before you can talk to a computer at the library reference desk and find it as intelligent as the human being who currently staffs it. But I'm not prepared to concede that yet. I think in due course we have to get back to real natural language front ends: it is possible to pull together (1) literal sentence understanding based on grammatical analysis, and (2) modern computational techniques of spotting likely relevance. But no one is trying to do it at the moment.

The people working on (2) are doing brilliantly. (Notice, the spotting of likely spelling mistakes in Google and advanced word processors is an aspect of (2), and deserves some real respect.) But work on (1) was all but abandoned in industry some ten years or more ago. It didn't have to be. Nothing was discovered that made people decide syntax or literal meaning were impossible to grapple with algorithmically. I am not ready to believe that centuries will have to elapse before I can send email to a shopping robot "Get me price information on Simmons Beautyrest mattresses from at least five stores in Northern California" and get some useful data back in response. Sure, there are some syntactically irresolvable ambiguities in there (is it mattresses from five stores, or information from five stores?). But Google's ad-relevance technology could pretty much figure out what I want even without analyzing the grammar. If we brought even just a little grammar together with some educated guesswork and smart search technology, we could have something really remarkable. Not intelligence and empathy — we're going to be talking about something vastly less intelligent than a cockroach here, and an equivalent level of empathy, i.e., none — but it could still be something very effective, something that you could easily mistake for rapid and effective assistance from an intelligent entity that had understood the gist of what you said.

Posted by Geoffrey K. Pullum at May 29, 2004 04:41 PM