Friday, October 16, 2009

Google Defends Wittgenstein?

I continue to be amused by the fact that one of my best sources for criticizing the positivist dreams of a "semantic knowledge layer" on the World Wide Web has been Google, particularly as personified by Director of Search Peter Norvig. Thus far I have concentrated on the question of whether such a "semantic knowledge layer" can "offer a substantive improvement over the kind of brute-force search that Google does so well." However, yesterday I had an opportunity to watch a video of a lecture that Norvig had given at Berkeley; and, when I saw some of the demonstrations he had cooked up as examples of "language understanding" (even if in the loosest sense of the phrase), it occurred to me that he may have hit on the fundamental principle behind Ludwig Wittgenstein's opposition of Bertrand Russell's positivism.

That principle is the one that states that "the life of the sign" lies in how that sign is used. Now there are basically two ways to take on the question of how a sign (such as a word) is used. The first is to get out all the guns that linguistic theory has provided since the beginning of the twentieth century. We can get a good sense of those guns from The Handbook of Discourse Analysis, edited by Deborah Schiffrin, Deborah Tannen, and Heidi E. Hamilton and published by Blackwell in 2001. There we can find over 800 pages of content guaranteed to leave even the most intrepid reader (here comes Anna Russell again!) "as befogged as before." The second approach is to build up an enormous database of use cases, which may then be applied as examples with the analytic support of little more than statistical inference. From Google's point of view, that database is the World Wide Web; and they have the search tool to seek out everything that is there. This is the approach Norvig has taken in his research; and I suspect that, if he even has his own copy of the discourse analysis Handbook, it has gathered as much dust as my own. His philosophy is one of the common variants of Murphy's Law: When brute force it unsuccessfully applied, it means you are not applying enough of it. (The alternative statement is, "When in doubt, use a larger hammer.")

Does this really work? Norvig certainly gave some impressive examples of how this approach can be applied to language translation. He did not say very much about examples that did not turn out very well, but he had some interesting plots of quality of translation against size of database. When you think about it, this may be exactly what Wittgenstein had in mind. The only way we know how a word-sign is used is by actually looking at the cases in which it is used. Wittgenstein could not have fathomed that we could do this for the number of cases that Norvig could use, so he could never get beyond philosophizing about what the results would be. At the very least, there is a delicious irony to the consequence that Google may eventually beat down the Semantic Web crowd with a big stick that was originally handled by Wittgenstein!

No comments: