lecture3

CS 337 -- Intro to Semantic Information Processing-- L. Birnbaum

LECTURE 3: REPRESENTING ACTIONS IN CONCEPTUAL DEPENDENCY

To sum up so far:

The representation of an utterance must be unambiguous -- or at

least, far less ambiguous than the utterance itself -- it must

make explicit the implicit content that was inferred, and it must

reflect semantic relatedness to other utterances (e.g.,

paraphrases).

Our vocabulary must allow us to construct such representations.

It must therefore be relatively unambiguous in order to facilitate

plausible inference. It must be fairly restricted to encourage

the use of the same representational elements to represent similar

meanings, so as to make clear how they are related.

Representations must have STRUCTURE. They must display the

relations among the concepts that make them up. These relations

give rise to expectations. A representation must indicate what

relations are necessary or possible, and which have already been

established and which have not.

We now turn to the structure of actions.

John went to New York.

"Going" is the ACTION.

"John" is the ACTOR.

"New York" is the "to" DIRECTION.

"John" is the OBJECT.

"some unknown place" is the "from" DIRECTION.

"past" is the TIME.

John took Mary to New York.

"John and Mary" is the OBJECT.

John drove to New York from Boston.

"Boston" is the "from" DIRECTION.

Let's look at a paraphrase:

John went to New York from Boston by driving.

"driving" is the INSTRUMENTAL ACTION.

John gave Mary a book

"a book" is the OBJECT.

"Mary" is the "to" RECIPIENT.

"John" is the "from" RECIPIENT.

Let's look at an example where the surface structure of the sentence

does not so directly reflect the underlying conceptual structure of

its plausible interpretations.

John likes chocolate.

Let's look at potential paraphrases again:

John likes eating chocolate.

Eating chocolate makes John happy.

(These might be wrong. John might like rubbing chocolate on

his body.)

Representing the most plausible interpretation of this utterance

requires representing implicit actions and causal relations.

The capitalized words are called the CONCEPTUAL CASES of actions.

They denote SLOTS in conceptual case FRAMES (aka "arguments" to

functions or predicates). Thus actions have the form:

(Act ACTOR (--) OBJECT (--) TO (--) FROM (--)

INSTRUMENT (--) TIME (--))

The relevant conceptual cases depend on the particular action.

We now turn to the acts themselves.

The need to restrict the representation vocabulary (principle 2b),

coupled with the need to develop a vocabulary for actions (5b),

suggests that we should look at closely related actions and see to

what extent they can be represented as instances of the same

UNDERLYING action.

(Up to now, we have been discussing the theory of conceptual

dependency, which is a theory of how representations can

facilitate inference and memory processing. We now turn to the

practice: conceptual dependency as an attempt to devise a

representation system that conforms with the theory.)

John sold Bill a bicycle.

Bill bought a bicycle from John.

What's happening here?

John is giving Bill the bicycle.

Bill is giving John some money.

Mary gave John the check.

John received the check from Mary.

These are almost the same, except that perhaps who actually

acted is not so clear in the second.

John took the check from Mary.

Again, almost the same, except that John is the actor.

These utterances are clearly related, and their representations could

make this clear if they referenced the same underlying action.

We could call it GIVE, but by using an English word we might

unconsciously rely on connotations that a computer program would

not know.

The best symbol would be G0054, but that's not very mnemonic.

In conceptual dependency, we compromise by calling this action

ATRANS: to transfer an abstract relationship with respect to some

object, such as possession or control, from a donor to a

recipient.

(Show representation of above in terms of ATRANS.)

Schank & Carbonell:

With any ... representation scheme, the advantage of the symbols

we create can only be in the new symbols or actions that they

spawn. That is, it is the inferences that come from [them] that

are of key importance.

Examples to be worked out in class:

Fred donated his old sofa to the Salvation Army.

Fred sold Bill a house.

Methodological interlude: We are playing the "representation game."

The rules are as follows:

Our discussions involved three different kinds of entities:

Representations of specific interpretations of utterances (R

level entities).

Specific vocabulary items (V level entities).

General claims about properties of representations and

vocabulary items, i.e., the entities of the preceding two

levels (T level).

First, we motivate the need for representations with certain

properties, and hence for a representation vocabulary with certain

properties. These general (T level) claims constitute the core of

conceptual dependency.

Next, we attempt to see how these general claims could be applied

to the problem of representing example utterances, or more

accurately, SPECIFIC INTERPRETATIONS of those utterances. Our

goal is to motivate the introduction of a specific vocabulary (V

level) for representing actions and causal relations.

Finally, after getting a handle on the vocabulary, we go back to

the examples to see how they would be represented (R level) using

it.

I want to make this as clear as possible, because otherwise too much

class time will be taken up with spurious arguments against this last

step, based on the claim that a proposed representation fails to

capture an alternative interpretation of some utterance.

This argument, even when correct, CANNOT affect claims at the R

level, i.e., a proposed representation for some PARTICULAR

INTERPRETATION of an utterance, because it is proposing the

representation for ANOTHER interpretation. An attack on a

proposed representation must be either (1) the claim that it fails

to include something that is clearly part of ANY reasonable

interpretation of the utterance, or (2) the claim that it includes

something that should NOT be part of any reasonable interpretation

of the utterance.

Of course, this argument CAN be used to attack a PROGRAM for

constructing some representation, if you argue that the

program has represented the WRONG interpretation, given

context and task. We have not yet been playing this game,

however.

This argument, if correct, CAN be used as part of an argument

against V level claims, IF our representation vocabulary is UNABLE

to represent the alternative interpretation. The reason is as

follows: Before some process model can construct one or the other

interpretation, depending on context and task, it must be able to

represent both of them.

Finally, such arguments CANNOT affect claims at the T level, i.e.,

our general theoretical principles, unless they would seem to show

that no vocabulary constructed according to those principles could

capture some interpretation or the whole range of possible

interpretations of an utterance.

Back to conceptual dependency; some other ubiquitous underlying

actions (V level):

PTRANS: to physically transfer an object from one location to

another.

MTRANS: to transfer information, either within or between

individuals.

MBUILD: to create or combine thoughts.

ATTEND: to focus a sense organ on a stimulus.

PROPEL: to apply a force to an object, in a given direction.

Examples to be worked out in class:

John went to New York.

John pushed the desk into his office.

John told Mary that he went to New York.

(Note here that restrictions on case slots depend on

action involved -- the object of MTRANS must be a

proposition.)

John decided to go to New York.

John saw that Bill was coming towards him.

MTRANS Actor (JOHN)

Object (PTRANS Actor (BILL)

Object (BILL)

To (LOCATION-OF Object (JOHN))

From (?)

Time (PAST))

From (M-LOCATION-OF Object (EYES Part-of (JOHN)))

To (M-LOCATION-OF Object (CONSCIOUS-MIND Part-of (JOHN)))

Instrument (ATTEND Actor (JOHN)

Object (EYES Part-of (JOHN))

To (LOCATION-OF Object (BILL))

From (?)

Time (PAST))

Why isn't the ATTEND the Object of an MBUILD? I.e., why not

represent this as something like "John concluded that Bill was

coming towards him by looking at Bill (coming towards him)"?

Because this is an INFERENCE. Maybe John WON'T conclude that

Bill is coming towards him. Maybe Bill is dead, and John will

conclude that he's hallucinating. In other words, the MBUILD

is an inference from the MTRANS -- one which is often true,

but not always. Roughly speaking, we must be able to

represent the distinction between seeing and believing.

This example points up the problem of how to make sure that you know

what your representational vocabulary "means." MBUILD is a somewhat

problematic action -- recall the proposal for a "PBUILD" predicate --

and is perhaps too general. Above, we concluded that MBUILD implied

belief much more strongly than MTRANS. In other words, they could be

distinguished on the basis of what could be inferred from them, and to

what degree.

"The real meaning of ATRANS consists of the inferences that are

likely to be true when ATRANS is present." (Schank, 1975)

"With any ... representation scheme, the advantage of the symbols

we create can only be in the new symbols or actions that they

spawn. That is, it is the inferences that come from [them] that

are of key importance." (Schank & Carbonell, 1978)

-------------------------------------------------------------------------------

Reading assignment:

Rieger, ch. 5 in CIP, pp. 157-288

-------------------------------------------------------------------------------