Slashdot Mirror

The problem of deciding good and bad would still remain. Perhaps a Wikipedia-style approach with volunteer evaluators?

They're running a project like this, to teach computers from volunteers. (A.k.a., collect volunteer-supplied data for various machine learning tasks.) It's Open Mind. Looks like they're not doing natural language grammar per se, but something similar could be used for that.

Grammar's hard, though. Linguists have been working for some 50 years to explicitly write down the rules differentiating what we percieve as good versus bad; and they still haven't succeeded. We still can't enumerate or describe the intuitions we all have in our heads about this.

The OpenMind project is also fascinating -- they plan to teach computers "common sense" items, to build up a human-type intelligence based on the inputs from thousands of Internet users. Obviously some people will put bogus data in, but they have verification steps as well to weed that out.

I haven't tried their speech project so I can't say whether or how well it works, but it's a step in the right direction at the very least.

Open Mind Commonsense: Teaching computers the stuff we all know

Until Cyc is allowed to self-generate rules this will limite Cyc's growth to the abilites of humans to feed it information on fact at a time. This will greatly limit the database's access to less popular or more technical topics and will slow down the process of learning.

I think it'd be cool to teach Cyc to program. "A bubble sort is less efficient than a quicksort."

Perhaps it could fix all Microsoft's bugs, without access to the source!

Oh, btw there's another couple projects similar to Cyc:

OpenMind and MindPixel .

The Open Mind Commonsense project being done by the MIT media lab. It's a similar project, but this one takes its input from users on the 'net, which I suspect is not such a good idea since someone else still has to sift through all of the contradictions that result. Not to mention people entering nonsense just to be silly.

I think the NASA project is basically make-work. After all, after all of the users have added their time and energy, the results are thrown away. They're not used to teach computers to recognize craters in the next evolution.

Having a speech recognition toolbox is only one part of the problem. As many people in the domain (I used to work in speech recognition) will tell you is that sometimes, the key to a good speech recognition engine is not in the code, but in the speech data used to train it. Speech databases are very expensive and speech recognition companies usually have a lot of "proprietary" databases.

One project which addresses the problem is the Open Mind Initiative, and more specifically the Open Mind Speech Recognition project, for which I am the coordinator. Our goal is to collect data from people on the internet and make that data available to people working on speech recognition with a GPL-like license. I think this is the key to having OSS speech recognition engines perform as well as the proprietary ones. The project is not very advanced yet, but any help would be really welcomed.