Tom Wolfe takes on linguistics
languagelog.ldc.upenn.edu
languagelog.ldc.upenn.edu
NLP/Comp Ling is another story entirely, of course, as are semantics, pragmatics, experimental ling, phono pursuits, etc. But for hard syntactic theory I don't think the generative ling is going anywhere.
Generative linguistics is in fact increasingly relevant to NLP, since people are moving towards more complex language models. That's not to say that cutting edge research in syntax is going to be helpful to people trying to solve practical NLP problems, but by the same token, string theory isn't much use if you're trying to build a bridge. It tends to be the basic and well-established results within a field that have useful applications outside it.
Chomsky is brilliant, no doubt, and critically showed the world how much latent structure there is in language. However, he is IMHO pretty wrong in thinking that this latent structure can't be learned (e.g. inferring a probabilistic context-free grammar).
One thing to be said for generative linguistics is that a whole lot of super-interesting phenomena have only been characterized within specific generative frameworks--language documentation and psycholinguistic research can only progress with a formal system in which language can be rigorously characterized. So one of the huge projects in the future will be translating the 50+ years of research into the formalisms in the modern computer science / statistical NLP / Bayesian cognitive science / learning-centric developmental psychology stack.
The rest of academia has agreed on a set of methodologies and premises? No wonder we feel left out!
>increasingly ornate innate machinery is needed to explain how kids arrive at the 'right' grammar
This is too vague to respond to. But theories do tend to get more complex as we learn more, since there's more data to account for.
>where the structural representations were never great at predicting processing difficulty, nor implementable in a machine.
Which structural representations assumed by (say) GB theory are not implementable in a machine?
>he is IMHO pretty wrong in thinking that this latent structure can't be learned (e.g. inferring a probabilistic context-free grammar).
Chomsky always agreed that it was possible to learn context-free grammars via statistical methods. That's why he placed such a great emphasis in the 60s on showing that context-free grammars are not a suitable model for natural language.
Your last paragraph is fairly astonishing, insofar as it admits that generative linguistics has obtained lots of interesting results which cannot be characterized in "modern" terms. That sounds like a pretty strong indication that generative linguistics has got something right!
Most of the individual constraints proposed in the myriad papers on GB and Minimalism are probably implementable by machine. But no one in Chomskyan generative syntax seems interested in explicitly spelling out the full set of principles that would underly a large-coverage grammar--except maybe Steedman's Minimalist Grammar formalism, which is ignored by most people who call themselves "syntacticians". In contrast, the HPSG and LFG communities have attempted to provide large-scale grammars and a lot of NLP work has used them in a serious way, but those communities are no longer very active.
Is there a work which lays out modern Minimalist generative syntax in full formal detail, and shows how this formal system handles a very large range of different syntactic phenomena? Such that it would be possible to produce a large hand-parsed corpus? It seems like this is what would be needed for generative syntax to have relevance outside linguistics departments. If it exists and I'm just unaware of it, I'd be glad to hear about it!
It's also a bit odd to suggest that any framework for which there isn't a wide coverage parser lacks any relevance outside linguistics departments. I know of lots of examples of cross-disciplinary work involving generative linguistics, but most of it doesn't relate to parsing at all. I'd say that wide coverage parsers are actually a pretty niche interest, which is one reason why people don't tend to work on them much.
Please correct me if I'm wrong but your representation of Chomsky's thesis seems to be the opposite of what it is.
Chomsky says that the ability to learn a language is innate - you don't need to go to school to do it, nor do you need to learn grammar rules.
You say that Chomsky says that the "latent structure in language" cant be learned and if my representation is correct, then yours is way off and seems partisan.
Presumably you have a critique of his poverty of stimulus argument and not merely a blanket dismissal?
From my experience most people who say Chomsky is wrong have in most cases never actually read Chomsky's papers. Trust me it's really hard to finish any of his actual papers, even as someone who majored computational linguistics. I'm pretty sure 99% of people on this thread who are throwing around the term "generative linguistics" casually have no idea what it actually is in detail. Maybe they read some wikipedia article, maybe they watched a 45 minute Youtube video of some famous guy talking about it.
I see it as a same phenomenon as people criticizing that "Blue ocean" theory is wrong, "Black swan" theory is wrong, "Lean startups" theory is wrong, all without actually having read and thought about any of those books and basing their criticism on the shallow knowledge they picked up from blogs, which even I have been guilty of.
My thought: The approach by Chomsky and others in this field may not be in fashion nowadays, but it does provide a significant foundation on top of which many others build their theories. Also for these significant theories there is no "right" or "wrong", the whole point is the theory does exist and it gives us something to think about. Saying Karl Marx was wrong is a foolish thing to say because right or wrong is not what matters. Maybe thousands of years later in the future things may change and communism may end up becoming the perfect ideology for our society then. It's all contextual.
http://norvig.com/chomsky.html
HN discussion of related article:
Chomsky's work is still seen as something of a foundation of modern linguistics and it's still taught in basic classes at university. It's also been very seminal in that it's brought about a vast array of models and theories that build upon it, including modern models such as the various phrase structure grammars (HPSG for instance).
However, the world of linguistics has changed quite a bit since the 1960s and through areas such as NLP and statistical models in particular we have a far better understanding of the building blocks of human language. Chomsky appears to be somewhat desperately trying to keep up with these changes and accommodate them in his models, which in turn have grown vastly more complex over the years compared with his original simple and elegant x' theory. In recent years, he also comes across as quite stubborn regarding his rejection of statistic models and machine learning.
Not sure where you're getting that from. Which aspects of Chomsky's recent work do you think are drive by some kind of desire to keep up with developments in NLP?
>his models, which in turn have grown vastly more complex over the years compared with his original simple and elegant x' theory
X' theory was not really 'original'. It was first introduced in Remarks on Nominalization in 1970. The models have not in fact grown vastly more complex. In a number of respects quite significant simplifications have been achieved (e.g. the reduction of large numbers of construction-specific transformational rules to a small number of generalizations about A and A' movement).
Chomskyian/generative linguistics completely dominated a generation of young linguists in the 1970s and 1980s: it provided a research program for writing dissertations. But the results were often dubious; generative grammar has great things to say about syntax, but little about morphology, phonology, semantics, &c. Moreover, Chomsky has radically changed his theories several times: his earliest theories, that were most influential, have been all but completely abandoned. Then he went to principles-and-parameters (which is where he lost me, with the argument that pro-drop is conditioned by binary brain switches), and then to the Minimalist Program, which I haven't really bothered to read as much about. But this is a minority paradigm these days. Chomsky's ideas start to really fall apart when you look too closely at some of his premises, like the poverty of the stimulus.
Then there are more traditional linguists (who do more fieldwork and comparative linguistics), applied linguists, and cognitive linguists. All of these groups tend to have separate conferences and have all the problems of scientists in different paradigms, always talking past each other.
EDIT: Chomsky made real, lasting contributions in formalizing the field. He completely transformed the discipline of linguistics for good and bad—but I think his good contributions will endure for a long time, once we get over the Chomskyan hangover.
Our goal is to build high-quality, sophisticated NLP solutions for "cleantext" : text that has been professionally written and copy-edited, so that it doesn't have obvious errors of spelling, grammar, punctuation, etc.
What makes our work different is that, unlike almost all other research in NLP, we don't depend on human-annotated data (like the Penn TreeBank) for training and evaluation. Instead, we build our models into lossless data compressors, and evaluate by invoking the compressor on a large text corpus.
In addition to this change in evaluation methodology, we are also deeply interested in the empirical study of text - that is, in the questions of traditional syntax research. For example, an important part of English syntax relates to verb argument structure: verbs differ in the types of arguments they can accept. Some verbs can accept infinitive complements, that-complements, or indirect objects; others cannot. Our system contains a module that handles argument structure and forwards this information to the parser and compressor.
I wrote a book presenting a rationale for why scientists should be deeply interested in large-scale, lossless data compression. Roughly, empirical data compression research is a variation of the scientific method: if theory A achieves better compression than theory B, then A is closer to the truth. Scientists can use this principle to search systematically through theory-space and find high-quality theories.
http://arxiv.org/abs/1104.5466
Interestingly, compression has a direct link to generation. By sending random bits into the decompressor, you obtain sample sentences that illustrate the model's "idea" of what English text looks like. By studying the discrepancy between the sample sentences and real text, you can discover ways to improve the model. This is a very productive technique that we use routinely. Chomsky declared that his goal was to build a machine that generates grammatical English sentences, but none of his followers actually tried to build such a device. So in some sense our research is actually closer to Chomsky's stated goal than what Chomsky's followers are doing.
I would love to hear from people who might want to collaborate on linguistics work, or use the data compression idea to explore other fields (eg. computer vision, bioinformatics, astronomy).
Brilliant. I suspect wolfe specifically extends the usually abbreviated latin word alii because alii in Hawaiian means "royalty".
It's time to read the Old Testament story of the Tower of Babel once more. That's always been a fascinating story for me ... the notion that people were forming a real or virtual tower that threatened divine order and so needed to be split by making their speech mutually unintelligible.