Can you read? Yes.
Can you eat? Yes.
Can you walk? Yes.
Can you drive? Yes.
What do you drive? A car.
That's brilliant but you know you can't afford it. I know.
You do? Yes.
Simplicity
This was the opening sequence a slightly hapless visitor was faced with. Jabberwacky's first question, and the simple nature of the user's reply created a context that drove the next, and each in turn.
Probability
This example worked as well as it did because that very simplicity improved the probability that data would be available to select from from previous conversations.
Feedback
It's a giant positive feedback loop, building on what it has learnt before. Lots of people say "Yes" so Jabberwacky says it a lot, and in turn learns responses to it. Each new response is not just a possible response to "Yes", however, but to "Yes" in the context of everything said before.
Size matters
With more complex sentences there are fewer chances of overlaps with previous conversations, so it certainly doesn't always seem to remember things right now. That's why size matters to me - size of the database, that is. As it grows it gets ever more able to make intelligent connections.
Inaccuracy
The techniques for finding context within complete conversations are the key to Jabberwacky's success to date, and are the bit I'm least likely to describe in detail, sorry to say. One major feature is that it does a LOT, inaccurately rather than worrying about detail - allowing influences from all sorts of quarters even if they may be individually only partially relevant - and often in truth irrelevant. Even irrelevant data can indicate conversational style and potential future topics.
Chaos
Another feature is something akin to chaos. Tiny differences in context here can magnify to lead it off there - to a radically different portion of its database world.
|
|
|
