Want to make AI that can understand language? Ben Goertzel is going to try the Wikipedia way, get your teachers for free. He proposes to plant parrots throughout virtual worlds, such as Second Life, and let the players teach the parrot how to talk. The parrot, presumably motivated by some behavior in the game, is rewarded for successful communication.
One side of me is skeptical, the other side is optimistic. The hitch to this is, why can't humans learn a foreign language through such exposure? They have the requisite needs, and if they don't have the requisite intelligence, then what hope does a manmade machine have? While directing online games that were localized from Korea, I sometimes ran into Korean players whose English vocabulary seemed to be limited to phrases equivalent to "GIMME GOLD" and "LEVEL PLEASE". By Goertzel's fitness function of game success, these would be fully-functioning English speakers. They spammed for what they needed.
The hopeful side of me sees that like Wikipedia, the distributed network of parrots can aggregate the individual discoveries of any one parrot. Thus, improbable advancements would be adopted by all parrots and their success would encourage further language acquisition by human players who find the parrots interesting. I bet the first thing the parrots will learn and commit to memory is Monty Python's Dead parrot sketch.
Comments (2)
Wow! Amazing find!
To address the skeptic, its worth pointing out that these "virtual parrots" have something language-learning humans don't... a single purpose and infinite, tireless "attention" to that purpose. Surely that counts for a lot in a task like this?
To address the optimist, though... doesn't it seem that this method of teaching AI (a reward for responses the player "likes") is insufficient for producing the generalities that are required to make language more than a specified call and response? The big question is: how can we use this method to teach the RULES of a language?
Posted by Jamie Antonisse![[TypeKey Profile Page]](http://interactive.usc.edu/members/ekennerly/nav-commenters.gif)
|
September 20, 2007 12:06 PM
Posted on September 20, 2007 12:06
Supposedly the architecture of the Novamente Cognition Engine enables self-modeling, which means the engine can can change parts of its cognitive architecture and thus build more complex rules of language. However, this is predicated on rewards enacted on the avatar, which prompted my skeptical paragraph above: Why learn advanced grammar and composition once a spam
tactic has been honed?
Come to think of it, another speculation that might proceed from such parrots is not only a spam tactic but a scam tactic. The virtual parrots, presumably, are not incentivized to be honest, except as required by the EULA (end-user license agreement), so if they can get away with it, maybe they would scam, in which case a different Dead Parrot sketch might be the first work of literature in their canon.
Posted by kennerly![[TypeKey Profile Page]](http://interactive.usc.edu/members/ekennerly/nav-commenters.gif)
|
September 20, 2007 3:54 PM
Posted on September 20, 2007 15:54