17-03-2014, 08:40 PM
Abstract
We explain how complex chains of inferences can be accomplished
by representing existentially quantified sentences, and concepts
denoted by restrictive relative clauses as classification hierarchies.
We describe first the representation structures which make
possible the inferences, and then we explain the algorithms which
draw the inferences from the knowledge structures. All the ideas
explained in this paper have been implemented and are part of the
information retrieval component of SNOWY, a program which understands
scientific paragraphs. An appendix containing a brief session
with the program can be obtained from the authors.
1. Introduction
For the last three years, we have been working on a model of
comprehension of scientific texts 141. We have also built a computer
program, called SNOWY, which embodies this model [7]. SNOWY
is a knowledge-lenient model of comprehension which reads texts
with minimal knowledge or no knowledge at all about their contents.
This aspect differentiates SNOWY, which builds conceptual
representations from scratch by using formation, or learning, rules,
from knowledge-intensive models of comprehension based on the
notions of plans [12], MOPS [ l l ] or TAUS [2]. The long-term goal of
our research is to build a program which will read an encyclopedia
and absorb its knowledge. An essential aspect of such a program is
a question-answering system which retrieves the knowledge that the
program has assimilated. This paper focuses on the knowledge
retrieval aspect of SNOWY'S question-answering module and, in
particular, on the inference techniques to answer certain types of
questions.
The key concept on which SNOWY relies to draw inferences is the
notion of classification. Consider the following paragraph below:
All whales are animals. All animals which live in the Antarctic eat
fish. Some whales live in the Antarctic.
These sentences can be represented in Predicate Calculus as (the
existential quantifier is represented as Vx, and the universal
quantifier as (x)):
1. (x) (Wale(x) --> Animal(x))
2. (x)Vy(Animal(x) and Fish(y) and Live(x,Antarctic)--> Eat(x,y))
3. Vz(Whale(z) and Live(z,Antarctic))
Suppose that given the above paragraph, the following question is
asked: Do whales eat fish?" The Predicate Calculus representation
of the question is VxVy(Whale(x) and Fish(y) and Eat(x,y)). Using
resolution to answer the question is beyond the point, because the
database of facts does not consist of just the facts listed above, and
resolution does not offer any indexing mechanism to access the
right chunk of knowledge.
Yet, the inference can be drawn relatively easily if restrictive relative
clauses and existentially quantified sentences are represented as
classification hierarchies, and a classification algorithm is invoked to
integrate the newly created concepts in the right place in long-term
memory (LTM). In our example, the clause "all animals which live in
the Antarctic" will be represented as a subclass of animals
distinguished by the fact that they live in the Antarctic; the sentence
"whales are animals" will be integrated as "whale -- is-a --> animal;"
and, when the sentence "some whales live in the Antarctic" is read,
a subconcept of whale is created, namely, the concept denoted by
"those whales which live in the Antarctic." Then, when the concept
whales which live in the Antarctic is integrated in LTM, the
classification algorithm determines that it must be integrated as a
subconcept of the concept animals which live in the Antarctic, as
illustrated in the diagram below. (The meaning of the terms used in
the diagram are explained in Sections 2 and 3.)
We explain how complex chains of inferences can be accomplished
by representing existentially quantified sentences, and concepts
denoted by restrictive relative clauses as classification hierarchies.
We describe first the representation structures which make
possible the inferences, and then we explain the algorithms which
draw the inferences from the knowledge structures. All the ideas
explained in this paper have been implemented and are part of the
information retrieval component of SNOWY, a program which understands
scientific paragraphs. An appendix containing a brief session
with the program can be obtained from the authors.
1. Introduction
For the last three years, we have been working on a model of
comprehension of scientific texts 141. We have also built a computer
program, called SNOWY, which embodies this model [7]. SNOWY
is a knowledge-lenient model of comprehension which reads texts
with minimal knowledge or no knowledge at all about their contents.
This aspect differentiates SNOWY, which builds conceptual
representations from scratch by using formation, or learning, rules,
from knowledge-intensive models of comprehension based on the
notions of plans [12], MOPS [ l l ] or TAUS [2]. The long-term goal of
our research is to build a program which will read an encyclopedia
and absorb its knowledge. An essential aspect of such a program is
a question-answering system which retrieves the knowledge that the
program has assimilated. This paper focuses on the knowledge
retrieval aspect of SNOWY'S question-answering module and, in
particular, on the inference techniques to answer certain types of
questions.
The key concept on which SNOWY relies to draw inferences is the
notion of classification. Consider the following paragraph below:
All whales are animals. All animals which live in the Antarctic eat
fish. Some whales live in the Antarctic.
These sentences can be represented in Predicate Calculus as (the
existential quantifier is represented as Vx, and the universal
quantifier as (x)):
1. (x) (Wale(x) --> Animal(x))
2. (x)Vy(Animal(x) and Fish(y) and Live(x,Antarctic)--> Eat(x,y))
3. Vz(Whale(z) and Live(z,Antarctic))
Suppose that given the above paragraph, the following question is
asked: Do whales eat fish?" The Predicate Calculus representation
of the question is VxVy(Whale(x) and Fish(y) and Eat(x,y)). Using
resolution to answer the question is beyond the point, because the
database of facts does not consist of just the facts listed above, and
resolution does not offer any indexing mechanism to access the
right chunk of knowledge.
Yet, the inference can be drawn relatively easily if restrictive relative
clauses and existentially quantified sentences are represented as
classification hierarchies, and a classification algorithm is invoked to
integrate the newly created concepts in the right place in long-term
memory (LTM). In our example, the clause "all animals which live in
the Antarctic" will be represented as a subclass of animals
distinguished by the fact that they live in the Antarctic; the sentence
"whales are animals" will be integrated as "whale -- is-a --> animal;"
and, when the sentence "some whales live in the Antarctic" is read,
a subconcept of whale is created, namely, the concept denoted by
"those whales which live in the Antarctic." Then, when the concept
whales which live in the Antarctic is integrated in LTM, the
classification algorithm determines that it must be integrated as a
subconcept of the concept animals which live in the Antarctic, as
illustrated in the diagram below. (The meaning of the terms used in
the diagram are explained in Sections 2 and 3.)