337 -- Intro
to Semantic Information Processing-- L. Birnbaum
2: CONCEPTUAL DEPENDENCY:
CRITERIA FOR MEANING REPRESENTATIONS
To repeat our conclusions from the last lecture:
language comprises understanding the situations
actions, objects, etc.) it describes, in conjunction
goals of the speaker or writer in speaking or writing as he did.
other words, understanding means being able to explain why and
what happened happened and why it was reported as it was.
understanding requires plausible inferences based on the
on context, and on background world knowledge.
the case of language understanding, these inferences
and reveal implicit content.
need a KNOWLEDGE REPRESENTATION system capable of representing
the MEANING of texts and the BACKGROUND KNOWLEDGE used in
ability to distinguish among the distinct meanings of
words and utterances shows that the underlying
of those distinct meanings are different.
is, the representation itself must be relatively unambiguous.
We now approach the question of criteria for
There are two sorts:
A representation must be able to express the ideas
opinions, etc.) that are needed to perform the tasks at hand.
What are those tasks, and how do they affect the form
the representation should take?
slogan form: STRUCTURE = FUNCTION +
course, the primary functional requirement is the need to
the appropriate content.)
It is hard to say anything concrete about how to go
a representation system that can express the content
you need. You
must look at a lot of examples, and try to represent
a lot of facts.
When you can't find a way to represent something you
need, you change
seems obvious to us that if one wants to describe a domain, one
about it by looking for the concepts in that domain that serve
organize the most information. For
each domain then, there
be some set of [such] primitives. (Schank
symptom of having got [the representation] wrong is that it
hard to say anything very useful about the concepts one has
It is easier, fortunately, to recognize when one [has got
right]: assertions suggest themselves faster than one can write
AI has more to say about functional concerns,
however. In order to
understand the relevance of such concerns to
representation, we must
first distinguish three (related) levels at which
themselves (R): Elements of this level of analysis
the representations of facts and utterances (or rather, specific
of utterances) themselves -- the kinds of structures
computer program would compute during understanding and inference.
system, or vocabulary (V): Elements of this level
analysis are the vocabulary out of which representations at the
level are built, and the inference rules by which those
items are related to each other.
theory (T): The elements of this level of analysis
claims and hypotheses about the other two levels and their
As pointed out abive, these claims and hypotheses
as much as possible, be motivated by the considerations of
and CONTENT. In fact, a large part
of the theory is
delineating these considerations.
The best possible theory, i.e., set of claims in T,
determine every decision about V and R.
No such T exists for any
reasonable-sized domain, so parts of every V, and
representation, are somewhat arbitrary.
Nevertheless, there is a
great deal that we can indeed say.
Functional considerations form the basis of Schank's
good representation must facilitate the plausible inference and
processing needed to understand. (This
is a claim about R;
claim itself is part of T.)
a good representation system must make it possible to build
representations. (This is a claim
about V; again, the claim
is part of T.)
Content considerations: What is the domain of facts and utterances that
we wish to represent? The answers form part of T.
are concerned primarily with the problem of representing
discourse. (Claim about R.)
we need vocabulary that allow us build representations of
about such things. (Claim about V.)
is a tall order; we must be satisfied with partial success.
IS IMPORTANT IS THE ABILITY TO EXTEND THE VOCABULARY IN A
content must be represented to adequately state facts about a
is a T level question. From the
point of view of the R and
levels, the answer is a given.
Turning to how functional considerations apply:
immediately rule out words and sentences of natural
themselves as an acceptable meaning representation,
ambiguity and ellipsis.
"Mary gave John a million dollars."
that John possesses a million dollars.
gave John a kiss."
imply that John possesses anything.
is difficult to prevent such confusions, however, if the
inference rules must be posed in terms of the words
So we can conclude the following (claims about R and
1a R: Unambiguous representation of utterances,
i.e., one that
differences in meaning, facilitates plausible inference.
1b V (following from 1a): Unambiguous
construction of unambiguous representations.
Another problem with natural language as a meaning
that the same meaning can be expressed in many ways
superficially quite different.
good knowledge representation must, to the greatest extent
represent similar (or related) meanings in similar (or
This facilitates memory search, for example in
was the author of "Hamlet"?
It also facilitates inference: to the extent that
meaning are reflected in similarities in
representation, the plausible
inferences they share can be captured by shared
sold Bill a bicycle.
bought a bicycle from John.
Finally, it is also important for sharing conceptual
when learning French for example, new inference rules
have to be learned to associate French sentences with their
rather than relying on knowledge acquired learning
So we can conclude:
2a R: Similarities in meaning between utterances
should be reflected
similar representations, to facilitate memory search and
of inference rules.
The need to restrict the representational vocabulary
meanings must, to the greatest extent possible, be
in the sharing of representational elements.
representational elements lead to unnecessary
of memory and inferential processing.
the other hand, if the vocabulary is too restricted, one may
be able to express the necessary concepts, or only be able to
so in a cumbersome way (a nebulous way of saying that other
considerations may impose a "conciseness" requirement).
These considerations lead us to:
2b V (following from 2a): Restricting the
much as is possible will require similar meanings to be
in terms of shared vocabulary items.
To this we can add:
3a R: The representation of an utterance should make
that were needed in order to understand it.
depending on context and purpose.)
3b V (following from 3a): Not much to say here,
vocabulary must be able to express the inferences
are made in understanding.
To sum up so far: the representation of an utterance
unambiguous -- or at least, far less ambiguous than
itself -- it must make explicit the implicit content
inferred, and it must reflect semantic relatedness
to other utterances
How else can representations facilitate inference?
Most directly, by
pointing out what inferences are necessary or
is already known about some instance of a concept being
and what remains to be known, should (as much as
be obvious by inspection.
should supply EXPECTATIONS.
We can get at this another way by recognizing that
cannot just be unstructured bundles of semantic
killed the bear" does not mean the same thing as "The bear
John." Because representations must be unambiguous, they
reflect such distinctions.
must have STRUCTURE. They must
among the concepts that make them up. In
These structural relations give rise to
representation must indicate what relations are
necessary or possible,
and which have already been established and which
To sum up:
4a R: The representation of a concept should make
clear what is known
it and what remains to be known, i.e., representations should
4b V (following from 4a and 1a): Vocabulary items
must have structure,
to specify the inference
rules that determine how
can be related to each other.
Let's look briefly at physical objects.
What attributes do they have,
and how are they related?
density, mass, function, ownership, etc.
= volume x density
who own an object with some function probably need to make
of that function.
Let's now turn to claims that stem from
considerations of content (the
claims themselves are T level):
5a R: Because people talk about actions, we must be
able to represent
the V level, we know that this entails a vocabulary capable of
such representations. But, do we
that represent actions? Or can we
get by with representing
McDermott's example of a man who runs around a track
times. At the end, the state of the
world is not changed,
that the man is more tired -- but this could be a result of
sit-ups the entire time, or just running in place. To
these, we must represent the action of running itself.
consider another example: How to distinguish movement from
5b V (follows from 5a and 1a): The representation
system must include
vocabulary of actions.
More informally (still at T level), conceptual
dependency is most
concerned with vocabulary that allows us to
physical actions (e.g., eating)
mental actions (e.g., concluding)
temporal and causal relations among actions and states
is somewhat concerned with:
physical states (e.g., health)
mental states (e.g., happiness)
has very little to say about physical or mental objects (e.g.,
Charniak & McDermott, ch. 1, pp. 8-28
Schank, CIP, ch. 3, pp. 22-82
Write an ELIZA (see Winston & Horn, ch. 17); the
purpose of this
assignment is to get up to speed in LISP, and to