Geoffrey Pullum is a reputed professor of Linguistics. Yet, the article comes off as written by an awfully ignorant person.
Forget pragmatics. There won't be a generic (cross-domain) solution for that any time soon.
What happened to Powerset? Doesn't really matter. The brains behind it haven't been killed off. For example, Ronald Kaplan, the top-brain behind Powerset, now works for Nuance Communications. I suppose he will focus on making sure your voice interactions with computers will be smoother.
"How could we have drifted into the second decade of the 21st century with absolutely no commercial NLP products?"
What the heck does that mean? Our smartphones understand voice commands. We have decent TTS and speech recognition. The people behind IBM's Watson seem to be targeting medical knowledge discovery. With a quick scan on Wikipedia, I came across Distinguo, automated essay evaluation, Google Translate makes my life so much easier already, etc.
If Pullum meant that he is still not able to buy an oven that understand his voice commands, sure. But saying that we have "absolutely no commercial NLP products" is braindamaged.
Sure, after all, who could possibly compete with IBM. Or Microsoft. Or AltaVista. I mean, AltaVista - backed by DEC - spent a relative massive amount of money at the time on what seemed like an index of insurmountable scale.
Who could ever compete with Facebook? You'd have to spend billions to replicate what they do, just to get off the ground.
The reason you're wrong, is because paradigm shifts is what matters, not chasing the past. That's the opening that makes it possible to replace Google (not compete head-on with them). I don't care how well organized Google is, they're so large now, they will run into the same clumsy elephant problem that every other mega company in world history has.
The problem isn't that Google has no competition. You should not be trying to solve the problem Google solved in 1998. That is over, they're riding the benefits of a solution from 15 years ago, extrapolated forward and polished.
You never compete head on with the past, you're guaranteed to die that way. You chase the future and replace the past. Look for the inflection, and bury Google in the past accordingly.
Need a better example? Windows was feared as a practical god of wrath in technology circa 1997/98. Who could ever hire enough engineers to write the tens of millions of lines of code to compete with that? It was the omnipotent titan that could never be challenged. Want the Windows obituary? It reads as follows: it never was challenged, it was replaced by the future that looked nothing like Windows 95/XP/Vista/8/etc.
The future begins with a new - obvious in hindsight - solution to a massive problem. Typically this solution will start out simple in concept, but can scale massively. It is almost always extremely cheap to instigate (relative to the value long term). See: PageRank.
The future does not begin by going after a leviathan like Google and playing the game the way they do, and spending money like they do. It never works that way. Chasing the past is usually the most expensive thing you can ever do. Just ask Bing.
You are not going to see a 20 year old Woz or a Torvalds today build an NLP system of any serious utility. They just do not have access to the kind of infrastructure required to do it. So comparisons to OS evolution, MS etc are moot.
You need data, clean data and massive amounts of it to do accurate natural language processing. Very few org's have the data required to pull it off.
And that can be seen in the limitations of Siri, WolframAlpha and IBM's Watson. The human brain takes a decade or more of processing all kinds of data, to build up a language model good enough to have a productive conversation. So even if these companies come up with the better learning algorithms, required to build up a decent language model, the bottle neck will always be access to quality data.
Google has the data, but since they are off busy defending their advertising empire, innovation in search and therefore NLP has taken a backseat.
For the paradigm shift in NLP to happen, things like Siri, Watson and WolframAlpha would need access to Google scale/quality data.
Google can spur things along by building a marketplace around search so people can access the data.
Actually NLP has made great progress in the last few years and there are commercial products now based on it: Watson, Siri, Google Now are good examples.
This is amusing because just today, while I was driving decided to attempt to use Siri (something I do about once a month to remind me why I don't use it). I said, "Nirvana Nevermind". "Sorry, I don't know how to do that." "Play Nirvana Nevermind". "Sorry, I don't know how to do that." "Search music Nirvana Nevermind". "Sorry, I can't search for music."
Well, your last example is evidence that there was a successful interpretation of the sentence.
Things like machine translation, text to speech, and natural language understanding will only gradually get better.
If you compare how well something like Siri does now (not very well) to similar “tools” 10 years ago (awful), it’s hard to deny that some progress has been made.
Processing natural language is a complex problem. Most syntactic parsing algorithms have a complexity of O(n^3). Add semantic reasoning and pragmatic understanding on top of that and you're way beyond what any computer system available today can process in reasonable time.
There are approximations using statistical algorithms and heuristics such as chunk parsing and those work quite well. In fact, this might very well be the way how human language understanding works as well, i.e. human brains most likely don't need to completely parse a sentence for understanding it but use statistical reasoning and heuristics, too.
Nevertheless, while studying grammar is a discipline almost as old as philosophy, linguistics that is the search for a true understanding of the building blocks of human language is a rather young discipline. When we still don't even have a comprehensive grammar of the English language including all its idiosyncrasies and varieties how can we expect to accurately model languages in computer programs?
Human language understanding certainly doesn't require a listener to completely parse a sentence before beginning the process of interpreting it. There is a lot of evidence from psycholinguistics that humans begin the interpretation process of a given word before the end of the word is even pronounced.
The only reason humans can do this is because our language interpreter is separate and parallel to our language parser. Unless we are making a language processing chip (transistors can be fully parallel) we want to avoid wasting work.
Neurons are a bit like transistors, it's a lot harder for them to rewire than for them to perform their already-assigned task. As a result, it is most efficient to have blobs of neurons configure into task regions and do the same thing over and over.
If you've already got a hardwired parser and a hardwired interpreter, there's really no reason to not have them run in parallel. (We have plenty of calories to run all the neurons that we need to, so any small chance of a benefit is worth a try.)
Didn't PowerSet become Bing, for the most part? I remember when Bing was originally launched, they flaunted it because it could show useful information instead of just "search results." (The feature seems to be scrapped now, but just in case you don't believe me - here is an ad that 'claimed' this http://www.youtube.com/watch?v=NHmzzLt8WFA).
It seems NLP is a problem that the author seems to underestimate. IIRC, real speech recognition is a Strong AI/AI Complete problem. For a machine to detect the nuances in regular speech requires that the machine be as smart as at least a small child.
I don't have a lot of knowledge in AI so take my statements with a grain of salt. (Artificial Intelligence: A Modern Approach was a great book, but it made me realize I am not really interested in the field).
If we ignore the pragmatic step, you aren't really doing NLP anymore, but just really fancy keyword matching. The query "Can your site be accessed without using Internet Explorer?" is just broken down with its syntax ("Your site", "accessed", "without", "IE", I'd imagine) and then each term is either translated into a keyword or an operator (semantics). The you end up with Google:Hackernews accessed NOT Internet Explorer (maybe) when what you really wanted was something closer too "list of supported web browsers" union "internet explorer".
Without the pragmatic step there isn't any way to translate "accessed .. IE" to "list of supported web browsers"
Maybe I'm missing something, but the whole pragmatic step seems to be the bread and butter of NLP, anything else just seems to be advanced querying. The whole pragmatic step could answer :
"Can your site be accessed without using Internet Explorer?"
Forget pragmatics. There won't be a generic (cross-domain) solution for that any time soon.
What happened to Powerset? Doesn't really matter. The brains behind it haven't been killed off. For example, Ronald Kaplan, the top-brain behind Powerset, now works for Nuance Communications. I suppose he will focus on making sure your voice interactions with computers will be smoother.
"How could we have drifted into the second decade of the 21st century with absolutely no commercial NLP products?"
What the heck does that mean? Our smartphones understand voice commands. We have decent TTS and speech recognition. The people behind IBM's Watson seem to be targeting medical knowledge discovery. With a quick scan on Wikipedia, I came across Distinguo, automated essay evaluation, Google Translate makes my life so much easier already, etc.
If Pullum meant that he is still not able to buy an oven that understand his voice commands, sure. But saying that we have "absolutely no commercial NLP products" is braindamaged.