Those who got fired up about Chomsky’s difficult comments regarding empiricism, including myself, will be gratified to see that Peter Norvig, patron saint of data-driven computational linguistics (inter alia), has released his own comments, along the same lines as mine, only better researched, more broadly researched, more respectful, more thorough, and, well, coming from the keyboard of Peter Norvig.
Norvig says it better
by
Tags:
Comments
7 responses to “Norvig says it better”
-
While I very much enjoyed reading Norvig’s piece, I’m disheartened by the context in which it came to being. Chomsky is reported to have made some short, semi-coherent remark, and the computational linguistosphere feels the need to send out one of their greatest champions to dismantle it piece-by-piece in an article many times as long as the original comment. Meanwhile, Chomsky doesn’t care, and possibly doesn’t know, because he has installed himself in such an impenetrable mental and professional bubble that he can act as a knowledge source with no external inputs, despite being provably wrong on many occasions and arguably a detriment to the advancement of the field. Why isn’t Chomsky irrelevant? Is it the cult of personality? Is it the culture of academia? Is it all just first-mover advantage? I do not understand why so many allow themselves and their professional lives to be in thrall to him. I am exasperated beyond my words to express.
-
That’s an excellent point. I would dispute Chomsky’s position as a “knowledge source” (outside of the sociopolitical stuff, on which I think he is often quite insightful), but I concur that he’s dropping dogma where the rest of us –including Norvig– are doing our best to drop science.
Chris has been complaining, accurately, that Norvig characterizes Chomsky as “inventing modern linguistics”, but I think Norvig’s characterization is accurate with respect to the way so much of the ling field feels obliged to respond to everything Chomsky has to say, no matter how off-the-cuff (or, frankly, ignorant of the state-of-the-art) C. seems to be.
-
I said “act as a knowledge source”, not actually *be* a knowledge source. Like you, I do not personally regard him as one. But he sure acts that way. He seems only to spew out, without ever taking in outside information. He seems to be dogmatic and patriarchal when it comes to linguistics, which is ironic, given his stance on social issues.
-
-
-
I guess even Chomsky doesn’t understand what he said… But I love that everyone wants to tell his own thoughts on every sentece from the “old man” (yes, including me :D).
I love Norvig and his intellectual development is an example of how statistical methods won in AI and computational linguistics. His classic PAIP (published in 1992) is full of old school rule based examples, and his fellow Pereira co-authored the classic Prolog and Natural Language Analysis. This smart guys changed their minds so this must have a cause!
But I think Norvig (and Jeremy in his previous post) made a few mistakes. Commercial success is not the measure of science. The founding fathers of logic were regarded as idiots, later electrical computers were built on their theories. Only a few theoretical physician made billions simply by doing science.
Sure, there are two cultures, but not only in statistics! And only a few understands the connection between the two! Aumann’s Interactive Epistemology I and II is the best example! Game theory (stats and probability fans love it) is good to describe a situation and making predictions (or choosing a strategy), but the old-school, rule-based logic is the tool of analysing epistemic states. Sounds philosophical? Yes, but we still need this sort of thinking!
I interpret Chomsky as a thinker (yes, a thinker or a philosopher). His questions are out of the scope of one type of science! He is interested in the epistemic background of language and he’s using the language of analytic philosophy and logic. This type of thinking yields no direct answers and results – but what would life be without philosophy, literature and other humanities? This is still serious science, but not in the sense of natural science! But (discrete) math and logic are still sciences!
Yorick Wilks’ What would a Wittgensteinian computational linguistics be like? paper tells a similar story about the two cultures and their connections in computational linguistics.
-
I just don’t get why Chomsky is still so highly regarded in linguistic circles. Geoffrey Sampson has clearly demonstrated how flawed his mathematics are, and that it’s all a bit of a cult – Chomsky is at MIT, so he has to be clever, and I as a poor linguist lack the understanding of his genius…
But everybody (myself included) always feels they have to justify themselves when doing something different from him. Instead he should just be ignored (for his linguistic work). But that is hard to do due to his fame.
-
I wish this all weren’t happening when I have job talks and a (highly statistically-oriented) dissertation to finish. I think he started going off the rails when he started talking about the “quake” example, and I think that there are a number of problems and mischaracterisations of Chomsky’s work (or, really, the field of generative grammar for whom Chomsky is still spokesman for better or for worse).
Look: the idea that people say all kinds of variable things is so obvious as to be banal. What we are arguing about are the constraints not only on producing and disambiguating the signal, but converting it into representations—yes, representations—that interact with the rest of cognition.
And I emphasize the point about conversion into representation (and, worse, acquiring the limits on that conversion). In all the claims about probability, statistics, modeling, there is the little problem that all of these approaches, when they are used as a criticism of generative linguistics, are playing a shell game, no matter what their merits are: they’re hiding the framework in which the data is structured, which itself contains its own biases, and almost none of them have any plausibility when it comes to describing what is necessary for a language to be acquired.
Finally, why do people think that the success of particular approaches in machine translation and other rather, um, eclectic language-related activities have any validity on the scientific questions that Chomsky asks?
-
As a computational linguist of a statistical nature, I can say that on one had we’ve one a lot of exciting things.
On the other hand, we’ve proven the following about language: when you have a whole lotta events, they tend to look like a whole lotta distributions of a whole lotta events.
Huzzah! Hallelujah! Event distribution is distributed.
What’s for lunch?
Leave a Reply