I read your essay about why NLP UI’s had not succeeded, and I think it’s a much more thoughtful analysis than most I’ve ever seen. But here are a few comments.

When forms work just fine, there is probably no need for using natural language. You’re right that forms have the big advantage of letting you know what things the system does and does not know how to represent and manipulate (an excellent point that I’d actually never heard before). In the forseeable future, nobody should try using NLP UI’s where something tried-and-true works well.

As you predicted, the need for typing is less of an issue now, and speech recognition has progressed quite a bit.

Applications do indeed need ways to refer to things. You might want to take a look at Kim Patch’s company, http://www.redstartsystems.com, which uses speech recognition (but not NLP) to let you refer to things in a computer UI very easily.

I think the biggest problems are still the classical ones: it’s hard to parse natural language statements because human languages are so complex by their very nature, and you have to be able to relate the sentences to an underlying knowledge base in order for them to have meaning, and that’s all very hard.

One reason there hasn’t been much progress in this area is that so many of the needs for NLP have been met by systems that are statistical in nature, rather than actually understanding human language. The success of statistics-based systems has pushed the need for “real” NLP back so far that industry hasn’t been working on it. I believe the pendulum will swing back; we’ll want abilities that statistical approaches inherently cannot do, and the time will come again for real NLP work in industry.

It’s been commonplace to say that “Artificial Intelligence” has not succeeded. But if you look at what I was taught about A.I. when I was an undergrad, successful A.I. is all around us. My cell phone and even my car can recognize speech. Genuine robots applications have succeeded and walking robots work fine. Deep Blue beat Karporov, and on and on. The usual thing happens: people say that what we’ve really learned is that speech recognition and robots and chess don’t require intelligence, after all. In this way, A.I. can never succeed, by definition. We all knew that even back in 1975: the bar gets moved back so you have to run as fast as you can to stay in the same place. But when you look from a historical perspective, the pattern is clear. We’ll have NLP. And it will “turn out” that you didn’t need “intelligence” for NLP after all.

– Dan