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

 

 

LECTURE 6:  REPRESENTING CONNECTED TEXT: INFERENCE AND CAUSAL CHAINS

 

 

Returning to Rieger's theory of inference: let's write some inference

molecules.

 

    ENABLEMENT molecule for ATRANS:

 

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

 

      Build (HAS-POSSESSION Actor ( ) Object ( ))

      Fill Actor  with  From  of input ATRANS

           Object with Object

      Establish ENABLES link from this HAS-POSSESSION to the ATRANS

 

    CAUSATIVE molecule for HAS-POSSESSION:

 

      (HAS-POSSESSION Actor (W) Object (Y))

 

      Build (ATRANS Actor ( ) Object ( ) To ( ) From ( ))

      Fill To     with  Actor of input HAS-POSSESSION

           Object with Object

      Fill rest of cases with UNSPECIFIED

      Establish RESULTS link from this ATRANS to the HAS-POSSESSION

 

Another example:

 

    John went over to the telephone.

 

    RESULTATIVE: John is at location of telephone.

 

    MOTIVATIONAL: John wants to be at location of telephone.

 

    FUNCTION: John wants to communicate something to somebody.

 

    Part of FUNCTION molecule for WANT:

 

      (WANT Actor (X) (HAS-LOCATION Object (Y) Loc (Z)))

 

      If Y is human (often X=Y)

         and Z is (LOCATION-OF Object (W))

         and W has FUNCTION FOO

      then

          Build (WANT Actor (  ) (FOO))

          Fill Actor of WANT with ACTOR of input WANT

          Fill Actor of FOO with Object of HAS-LOCATION

          Establish INITIATE link between this WANT and input WANT

 

    telephone FUNCTION = (MTRANS Actor (X) Object ( ) to ( ) From (X))

                 and (MTRANS Actor ( ) Object ( ) to (X) From ( ))

 

(You might need ACTION-PREDICTION inferences as well to handle the

examples I gave.)

 

The important thing about inference is that it enables us to EXPLAIN

what we read by constructing causal connections between events and

situations described in text.  This leads us to the following theory

of text representation (Schank, 1975):

 

    The proper conceptual representation of a text describing an event

    consists not only of representations of the actions, states, and

    objects described therein, but also of causal connections between

    these entities, linking the text into a causally coherent chunk --

    in the simplest case, a CAUSAL CHAIN.

 

Compare the following example texts:

 

    Mary gave John a million dollars.  He bought a new car.

 

    Mary gave John a million dollars.  She visited her aunt in

    Milwaukee.

 

What other theories of text representation might there be?

 

    Undifferentiated links, i.e., causality is not special.

 

    But now consider:

 

      John is married to Sally.  The boss called John into his

        office.  John's boss is a Rosicrucian.  The Rosicrucians are a

      quasi-religious mystical organization.  Sally does not belong

        to the Rosicrucians.  Sally's mother lives in Des Moines.

 

Let's look at a tough example:

 

    Menakretes' wife left for the country.  Within a week, one of the

    slave girls was wearing a new necklace.

 

Rieger's theory can't handle this kind of thing: if inference chains

get too long, combinatorial problems arise in a big way.  But it CAN

handle the simpler examples.