Thread Links Date Links
Thread Prev Thread Next Thread Index Date Prev Date Next Date Index

Re: Directions for future research

Title:

Hello,

   How can you design systems with the flexibility
   of natural language that can have the perceptual
   and learning ability of a two-year-old child?

That's the real problem.  I don't claim to have the
answer, but I am concerned that the focus of much of
the current research is not addressing that problem.
Instead, some of it may make it harder to solve.
I also believe that trying to make systems to learn and acquire the world
knowledge like children is a basic and important attempt to build large
knowledge bases, ontologies, lexicons and even inference engines.
We have done a project which I think is toward a part of the challenge you
mentioned. In this work (called Hasti) we assumed that the system is going to
learn about concepts and their relations (ontological knowledge) and also
words which refer to these concepts (lexical knowledge) through receiving
simple texts. It starts from scratch, there is no ontology and no lexicon (except
a very small kernel) at the beginnings. Just like a child, while Hasti learns more
 it can process more complex texts. The main idea is that babies are born with a
kernel in their mind knowing how to learn (and not what to learn).
The tests show good results in processing both general and specific texts for first
grade of primary school children. Hasti processes Farsi (Persian) texts but its
main learning algorithms is language independent.
The main conceptual problem in Hasti is that although it can discover concepts
and relations from simple texts but it has no visualization about them. It just learns
like blind, deaf, unfeeling children who can read simple texts!!
a paper on Hasti is published as;

Shamsfard " Learning Ontologies from Natural Language Texts" ,
              Int J.of Human-Computer Studies, No. 60 , pp. 17-63, Jan.2004.

Now we are continuing the project to enhance Hasti's learning process and
map the results to existing ontologies like SUMO. Your comments and critiques
are welcome.

Thanks
Mehrnoush Shamsfard