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

LECTURE 4:  CAUSAL RELATIONS AND INFERENCE

Turning to the representation of causal relations among actions and

states.  Let's look at an example:

    John bought a car.

    What state results from this action?

      John owns the car, the seller has his money.

    What state enabled this action?

      John owned the money, the seller owned the car.

    What was the reason for the action?

      John wanted the car.

    Both John's decision and his possession of money were necessary

    for his purchase of the car.  But the way in which they were

    necessary seems different.

Because people can distinguish different types of causal relations,

and because representations should be unambiguous (1a), we must be

able to represent different types of causal relations (T level claim).

Conceptual dependency distinguishes four types of causation (V level):

    Acts can RESULT in states (more properly, state changes).

    States can ENABLE actions.

    States or acts can INITIATE mental states.

    Mental states can be REASONS for acts.

    Examples to be worked out in class:

      John killed Mary.

      John grew the plants with fertilizer.

      When John saw the new car, he wanted it.

      John scared the mugger off.

      Why did you go to New York?

      Fred nauseates me.

Let's consider result and enable inferences from the acts: (Examples

to be presented in class)

    ATRANS Actor (X) Object (Y) To (Z) From (W)

      Enabled-by: W has Y

      Result: Z has Y

          Enables: Z will perform normal function with Y

      Result: W no longer has Y

    PTRANS Actor (X) Object (Y) To (Z) From (W)

      Enabled-by: Y is at W

      Result: Y is at Z

          Enables:  If Z is location of  a  physical  object,  Y  will

          perform normal function with that object

      Result: Y is no longer at W

    MTRANS Actor (X) Object (Y) To (Z) From (W)

      Enabled-by: Y is at "mental location" W

      Result: If X=W but not Z, Z now knows Y

    MBUILD Actor (X) Input-object (Y)  Output-object (Z)

      Enabled-by: X knows Y

      Results: X knows Z

Let's consider inference from instrument to main act:

    ATTEND is Instrumental to MTRANS

    PROPEL is Instrumental to PTRANS

    PTRANS may be Instrumental to ATRANS

And, we can infer acts given states:

    POSSESSES Actor (X) Object (Y)

      Results-from: ATRANS Actor (?) Object (Y) To (X) From (?)

    LOCATION Object (X) Loc (Y)

      Results-from: PTRANS Actor (?) Object (X) To (Y) From (?)

    KNOWS Actor (X) Object (Y)

      Results-from: MTRANS Actor (?) Object (Y) To (X) From (?)

Inference and the "but" test.  Consider the representation of

    John went to New York.

    John arrived in New York.

      Now add "...but he never became located in New York."

    The "but" test: If "X but not Y" is clearly anomalous, then Y is

    an intrinsic part of X; else Y is an inference.

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

Assignment

Using conceptual dependency actions and causal relations described in

class, plus whatever states from Schank (1975) seem necessary,

represent the following utterances:

    John told Mary that he saw Bill.

    Mary told John that she wanted to go to New York.

    Mary took John to New York.

    John decided to give Mary a book.

    Mary stabbed John to death.

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