My Photo

« Animated Social Network Visualization | Main | SimuTerra: SketchUp for Terrain »

January 22, 2007


Bob Carpenter

I like the example: "pick up the car at the airport". Does the prepositional phrase "at the airport" attach to the noun "car" or to the verb phrase "pick up the car"? Usually, the car's going to be at the airport when you pick it up. But there's also tense. The car might not be at the airport now, but it'll be there when you pick it up. Furthermore, "the car" is really attributive here, and only gets instantiated when an agent assigns the rental car. In that way, it's like "the winner of this year's Super Bowl", which is not yet determined, but will be after the teams play.

There's a huge literature on the psycholinguistics of parsing. And its interaction with everything from phonetics to world knowledge and context.


I am by no means an expert or even a novice in this area. I'm speaking more on speculation than experience, but I have always been under the impression that this was a perfect example of where you start when you have an ambiguous sentence. If you parse it and there are two potential "structures", THEN you use context to determine which one is truly the case. Is this completely off base?

Matthew Hurst


My belief is that we process such sentences into an ambiguous form (not storing multiple parses, but something else). Only when there is a need to disambiguate, do we consider resolving the ambiguity - and even then there may be no need to create a complete unique structure. For example, in the case I cite, unless someone asks specifically what the shoppers were intending to buy we probably don't need to create a single structure.


About the parsing:

Okay, well, I have a different view. I believe that there is quite a complex cascade of inference and prediction, starting at the level of raw sense data from the cochlear and retina and skin and limb proprioceptors, ending at high-level constructs with which we reason, with information moving simultaneously up and down. So consciousness (and by this I mean being conscious OF something) is a process rather than a state.

This isn't a view unique to me, its something I've absorbed from many different sources and it seems to be what the emerging academic consensus is. But most intriguingly, I came along this quite recently:

I know, it's easy to dismiss a commercial entity that claims to have an abstract mathematical model of neocortical function and intends to commercialize it rather than build a academic revolution around it, but this is an interesting century we are living in.

So in the case you provided, there would be a battle between the various layers: in the parlance of data mining, n-gram world models at the bottom and semantic models at the top, maybe grammar models in the middle, trying to find a global optimum of what reduces confusion.

The key idea is that there is no single, unitary algorithm that makes sense of input -- which certainly makes a bit more sense of the historical failure to produce anything resembling AI using simple-minded application of purely bottom-up methods like genetic algorithms and purely top-down methods like symbolic reasoning systems.

The comments to this entry are closed.

Twitter Updates

    follow me on Twitter

    March 2016

    Sun Mon Tue Wed Thu Fri Sat
        1 2 3 4 5
    6 7 8 9 10 11 12
    13 14 15 16 17 18 19
    20 21 22 23 24 25 26
    27 28 29 30 31    


    Blog powered by Typepad