Transcript Seminar Slides

Knowledge Structure
Vijay Meena (99005027)
Gaurav Meena (00005020)
Introduction

Human being involve a lot of common sense
- A child is always youngr than his mother
- You can push something with a straight stick
But computer fails to recognize them
We can write programs that exceeds the capability
of experts, but yet we cant write program that match
the level of a three years old child at recognizing
objects.
Common sense knowledge


Computer lacks common sense
Why has it been so hard to give computer
this common sense ?
- involves a great amount of knowledge
-
Many kind of representation.
AI has become the gold mine of techniques
We don’t want to give computers knowledge about
particular areas, but instead want to give common
sense.
Common sense is problem of
great scale and diversity
Two parts
1. How to give millions of piece of knowledge to
computer?
- need a database.
2. How to give computers the capacity for common
sense reasoning?
- having database is not enough
- don’t have enough idea about how to
represent, organize and use common sense
knowledge
Some projects

CyC
– Started by Douglous lelant in 1994

WordNet
– Started by Fellbaum in 1998

ConceptNet
- Started by MIT Media LabTeam after
WorldNet
CyC - CyCorp



To create the world’s first true AI having both
common sense and the ability to reason
about it
CyC represents it’s common knowledge in a
language call CyCL.
CyCL is a second order logical language with
second order features such as quantification
over predicates.
WordNet




WordNet is an online lexical reference
system
Have semantic network representation
Nodes have lexical notation
English nouns, verbs, adjectives and adverbs
are organized into synonym sets, each
representing one underlying lexical concept.
Concept Net




Very large semantic network of common
sense knowledge.
Structured as a network of semi structured
natural language fragment.
Captures a wide range of common sense
concept and relations.
Ease of use camparable to WordNet.
ConceptNet



Presently consists of 250,000 elements of
common sense knowledge.
Combined the best of both CyC and
WordNet.
Extended Sementic network representation of
WordNet.
WordNet v/s ConceptNet


WordNet
lexical notation of node
Small ontology of
semantic relations


ConceptNet
Conceptual notation of
node
Richer set of relation
appropiate to concept
level node
At present there are 19 semantic relations used in
ConceptNet representing different categories.
Representation of Knowledge




Logic representation : unambigious
Natutal language representation : ambigious
Maintaining some ambiguity lends us greater
flexibility.
Methodology for reasoning over natural
language fragment
Concept Net
Focus on the Knowledge
representation aspects of
ConceptNet
Origin



machine-usable resource mined OMCS
CRIS mined predicate argument structures
from OMCS
Is produced by an automatic process which
applies a set of ‘common sense extraction
rules’ to the semi-structured English
sentences of the OMCS corpus.
Structure
Semi-Structured natural-language fragment
nodes falls into three general classes :
 Noun Phrases (things, places, people)
 Attributes (modifiers)
 Activity Phrases (actions and actions
compounded with a noun phrase or
prepositional phrase)
Grammer for partially structuring
natural language cncepts
Node class
A portion of the
grammar
Examples of valid nodes
Noun Phrases
NOUN,
NOUN NOUN,
ADJ NOUN,
NOUN PREP NOUN
“apple”; “San Francisco”; “fast
car”; “life of party”
Attributes
ADJ,
ADV ADJ
“red”; “very red”
Activity Phrases
VERB,
VERB NOUN,
ADV VERB,
VERB PREP NOUN,
VERB NOUN,
PREP NOUN
“eat”; “eat cookie”; “quickly eat”;
“get into accident”; “eat food
with fork”
Semantic Relation Types
currently in ConceptNet








Category
Things
Events
Actions
Saptial
Goals
Functions
Generic
Methodology for Reasoning over
Natural Language Concepts


Computing conceptual similarity
Flexible Inference:
context finding,
inference chaining,
conceptual analogy.
Conclusion



To support several kinds of practical
inferences over text
To maintain an easy-to-use knowledge
representation
ConceptNet follows the easy-to-use semantic
network structure of WordNet, but
incorporates a greater diversity of relations
and concepts inspired by Cyc.
References




www.openmind.org
www.conceptnet.org
http://www.cogsci.princeton.edu/~wn/
Focusing on ConceptNet's natural
language knowledge representation, Liu,
H. & Singh, P. (2004)