Thursday, February 9, 2017

Antoine Bordes Artificial Tasks for Artificial Intelligence

Cribbing from a forum post by Merlin on He is discussing:
Antoine Bordes Artificial Tasks for Artificial Intelligence, ICLR keynote, 2015.
Example Task:
John is in the playground.
Bob is in the office.
John picked up the football.
Bob went to the kitchen.
Where is the football?  A:playground
Where was Bob before he went to the kitchen? 
This is very thought provoking. To answer correctly requires knowing the src/target roles for the event words "in", "picked up", and "went to" but not much else. It is like Narwhal without the application keywords. It sure would be worthwhile to be able to extend the internal GENERAL_OP tree of Narwhal, to include an EVENT_OP tree.

Another idea arising from this "Example Task" is that  - surely - this is not all of AI because answering the such questions correctly is content-independent. It would be very worthwhile layering an ability, on top of Narwhal, to match the original word in the text (as in "snippet" of the NarRecord) to the same word appearing in the question or being inserted in an artificial answer. So with something like an EVENT_OP tree what we have is not a narrow world but a narrow world template.

Another idea for Narhwal is that, as a NWSReader's vaults are filled while reading a text segment, a separate structure of "context" can evolve along side. This context includes changes and time differences to handle questions like "Where was Bob before he went to the kitchen?".

No comments:

Post a Comment