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

LECTURE 5:  INFERENCE: AN INTRODUCTION

We now turn to a discussion of Rieger's theory of inference.  Like all

AI theories, it consists of two parts:

    CONTENT (Rieger and Schank): 16 classes of inference that can be

    drawn given an act, state, or causal relation.

      We have already seen three of these classes: enablement

      inference, result inference, function inference.

    PROCESS (Rieger): Spontaneous, automatic, "bottom-up" inference

    from inputs and inferred concepts, searching for matches with

    previously known concepts in memory, or others inferred from the

    input itself.

Let's consider the process first.  Rieger's claim is that just about

all inferences that it is possible to draw, should in fact be drawn.

    The purpose of inference in his system is to discover unforeseen

    relations among concepts.

    "The 'goal' of inferencing is rather amorphous: make an inference,

    then test to see whether it looks similar to, is identical to, or

    contradicts some other piece of information in the system.  When

    one of these situations occurs, the memory can take special action

    in the form of discontinuing a line of inferencing, asking a

    question, revising old information, creating new causal

    relationships, or perhaps invoking some sort of higher-level,

    goal-directed belief pattern..."

    This is perhaps not so clear as it should be.  The biggest problem

    with Rieger's process theory is that the system never really does

    much with the inferences it draws.

In more detail:

    Indexed under each predicate and object in each input concept, by

    the name of the inference type, are "inference molecules" that

    specify (1) what the appropriate inference of that type is for

    that predicate or object, and (2) how to apply it to the current

    input concept.  Inference molecules are really just little

    fragments of LISP code -- or they may be viewed as productions or

    inference rules to be executed by a special rule interpreter.

    Each input concept is placed on a queue, and all of the inferences

    that can be drawn from it are drawn, and the resulting concepts

    are placed on the end of the queue for later inference.  In other

    words, breadth-first expansion through inference space.

    "When a new inference is generated, one of three conditions can

    apply:

      (1) The new inference can match something else in memory.

        When this happens, the new information is said to CONFIRM the

        old.

      (2) The new inference CONTRADICTS (is incompatible with) some

      old information.  This means either that something is

      Conceptually peculiar about the utterance, or that memory has

      made an incorrect decision about some referent, or has

        generated a probabilistic inference which turns out to be

        unlikely.

      (3) The new inference can neither be determined to contradict

      nor confirm old knowledge.  In this case, the new information

        is simply remembered, and is said to AUGMENT existing

        knowledge."

Let's look at an example:

    John told Mary that Bill wants a book.

    1   John believes that Bill wants a book.

    2   Mary now knows that Bill wants a book.

    3   Bill wants a book.

    4   Bill wants to come to possess a book.

    5   Bill probably wants to read a book.

    6   Bill might want to know the concepts contained in a book.

    7   A book about what?

    8   Bill might get himself a book.

    9   John might give Bill a book.

    10  Mary might give Bill a book.

    11  John may want Mary to give Bill a book.

    12  John and Mary may have been together recently.

    Some of these seem quite reasonable (e.g., 1-4, 7-8, 12) or even

    insightful (11).  On the other hand, others seem a bit weird (9).

A clearer example of the problem here:

    John hit Mary.

    1   John probably used his hand.

    2   Mary was probably hurt.

    3   John probably wanted to hurt Mary.

    4   John might have been angry at Mary.

    5   Mary now wants to feel better.

    6   John probably wants Mary to feel better.

    What's wrong with 6?  The inference being used here is

    motivational -- people usually want the outcomes of their actions.

    (This is how 3 was derived).  But obviously it has been taken too

    far.

In other words, the problem with the lack of goal-direction in

Rieger's theory is that it's hard to know when to STOP making

inferences.

    There are several ways to handle this problem: (ask for class

    input)

    Unprincipled: Stop expanding an inference chain after it reaches a

    specified length.

    Principled: Stop expanding an inference chain after it leads to

    confirmation of or contradiction with other information

    Principled: (special case of above) Keep checking for inherent

    reasonableness of inferences.  Stop expanding an inference chain

    when it generates an unreasonable inference.

Turning now to the 16 inference classes themselves:

    SPECIFICATION: The representation of an action has certain

    conceptual cases which must be filled in (specified).  If the

    input language leaves them unspecified, they must be inferred as

    specifically as available information allows.

      John picked up a rock.  He hit Bill.

      JOHN HIT BILL WITH THE ROCK.

    CAUSATIVE: What were the likely causes of an action or state?

      John hit Mary.

      JOHN WAS PROBABLY MAD AT MARY.

    RESULTATIVE: What are the likely results (effects on the world) of

    an action?

      Mary gave John a car.

      JOHN HAS THE CAR.

    MOTIVATIONAL: Why did (or would) an actor want to perform an

    action?  What were his intentions?  In the absence of information

    to the contrary, people can be assumed to perform actions for the

    probable consequences of those actions.

      Mary hit John.

      MARY PROBABLY WANTED JOHN TO BE HURT.

    ENABLEMENT: What states of the world must be (or have been) true

    in order for some action to occur?

      Pete went to Europe.

      WHERE DID HE GET THE MONEY?

    FUNCTION: Why do people desire to possess objects?  Usually, in

    order to perform their normal functions.

      John wants the book.

      JOHN PROBABLY WANTS TO READ IT.

    ENABLEMENT-PREDICTION: If a person wants a particular state of the

    world to exist, is it because of some predictable action that

    state would enable?

      Dick looked in the cookbook to find out how to make a roux.

      DICK WILL NOW BEGIN TO MAKE A ROUX.

    MISSING-ENABLEMENT: If a person cannot perform some action he

    Desires, can it be explained by some missing prerequisite state of

    the world?

      Mary couldn't see the horses finish.  She cursed the man in

      front of her.

      THE MAN BLOCKED HER VISION.

    INTERVENTION: If an action in the world is causing (or will cause)

    undesired results, what might an actor do to prevent or curtail

    the action?

      The baby ran into the street.  Mary ran after him.

      MARY WANTS TO PREVENT THE BABY FROM GETTING HURT.

    ACTION-PREDICTION: Knowing a person's needs and desires, what

    actions is he likely to perform to attain those desires?

      John wanted some nails.

      HE WENT TO THE HARDWARE STORE.

    KNOWLEDGE-PROPAGATION: Knowing that a person knows certain things,

    what other things can he also be predicted to know?

      Pete told Bill that Mary hit John with a bat.

      BILL KNEW THAT JOHN HAD BEEN HURT.

    NORMATIVE: Relative to knowledge of what is normal about the

    world, determine how strongly a piece of information should be

    believed in the absence of specific knowledge.

      Does Pete have a gall bladder?

      IT'S HIGHLY LIKELY.

    STATE-DURATION: Approximately how long can some state or

    protracted action be predicted to last?

      John handed a book to Mary yesterday.  Is Mary still holding

        it?

      PROBABLY NOT.

    FEATURE: Knowing some features of an entity, and the situations in

    which that entity occurs, what additional things can be predicted

    about that entity?

      Andy's diaper is wet.

      ANDY IS PROBABLY A BABY.

    SITUATION: What other information surrounding some familiar

    situation can be imagined or inferred?

      Mary is going to a masquerade.

      SHE WILL PROBABLY WEAR A COSTUME.

    UTTERANCE-INTENT: What can be inferred from the WAY in which

    something was said?  Why did the speaker say it?  (Once his

    reasons have been inferred, the proper response, e.g., to offer

    help, might become clearer.)

      Mary couldn't jump the fence.

      WHY DID SHE WANT TO?

Let's look at how specification inferences get made:

    John hit Bill.

    (RESULTS

     Action (PROPEL Actor (JOHN) Object (X) To (BILL) From (JOHN))

     State (IN-PHYSICAL-CONTACT Object1 (BILL) Object2 (X)))

    What is the X that John PROPELled?

    A general rule: If all else fails, find an X in memory which has 

    the appropriate conceptual features, and which was recently

    created or referenced.

    A specific rule for determining the OBJECT of PROPEL: If the ACTOR

    had has something in his hand at the time the PROPEL takes place,

    assume that it is the OBJECT of the PROPEL.  Else, assume that the

    OBJECT of the PROPEL is the ACTOR's hand.

Causal chains:  Input causals are often incomplete.

    Mary didn't like Bill.

    Mary kissed John because he hit Bill.

    Working forwards from "hit":

    1   John propelled his hand towards Bill.

          results

    2   John's hand came in physical contact with Bill.

          enables

    3   John propelled Bill.

          results

    4   Bill suffered a negative change in physical state.

          initiates

    5   Mary's happiness increases.

          initiates, in conjunction with 3

    6   Mary likes John.

    Working backwards from "kiss":

    7   Mary kissed John.

          Reason-is

    8   Mary likes John.

    Recall that memory is always looking for matches between concepts.

    Because 6 and 8 match, they are merged, and a connected causal

    chain is the result.

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

Assignment

    Write a Rieger inference program.  Implement the following classes

    of inference, where applicable,

      Causative

      Resultative

      Motivational

      Enablement

      Function

    for, at least, the following actions, states, and objects

      ATRANS

      PTRANS

      MTRANS

      PROPEL

      ATTEND

      EATING

      LOCATION

      POSSESSION

      KNOWLEDGE

      WANTING

      HEALTH

      HAPPINESS

      BOOKS

      TELEPHONES

      FOOD

And more if you like.  The program will take instances of concepts as

input (not English), and will be used in two modes:

    (1) Simply to draw inferences from a single input concept,

    stopping an inference chain at some fixed length.

    (2) To build causal chains.

      Including the following examples:

          John went over to the book.

          John looked up something in it.

          Mary gave John some money.

          John bought a car.

          Mary went over to the telephone.

          She told John that she would be home at 8.

          John bought a book.

          He read the book.

      And any others you want.

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