COMMENT ⊗   VALID 00003 PAGES
C REC  PAGE   DESCRIPTION
C00001 00001
C00002 00002
C00011 00003	Specific Emotions
C00024 ENDMK
C⊗;

%emotio[w90,jmc]		Notes on emotions (for CYC)
% for transcription of 1990 Feb 2 notes, see mcc[w90,jmc]
% see, OUTGO.MSG[1,JMC]/205P/1L,w90.out[let,jmc]/205p/1l for William James
% theory of emotions as sensations.

	The object of these notes is to discuss putting facts about emotions
in CYC.

	The present version of these notes is unsystematic.  Remarks are in
the order I thought of them or squeezed in with topics previously discussed.

  Some more general considerations about putting facts about
particular domains in CYC will be discussed.  Here are some considerations.

	1. We are not concerned with giving CYC emotions.  It is a question
of telling CYC about human emotions.

	2. We would like to give CYC the capability of doing the following.

		a. Inferring from facts about a specific situation that
a person has certain emotions.

		b. Inferring that certain events will give a person certain
emotions.

		c. Inferring that having certain emotions will lead to
certain actions.

		d. Accepting advice about creating or not creating certain
emotions, e.g. don't disappoint people (certain people).

	3. It should be possible for CYC to reach conclusions,
not only about specific situations, but also to infer general laws.

	4. In so far as possible we propose to express
information about emotions in a static way and use CYC's general
facilities for inferring the consequences of events to infer
the consequences of emotional events.  For example, we might write
as an axiom

(1)	hates(P1,P2) ∧ vengeful P1 ⊃ wishes(P1,harm-to P2)).

Axioms about motivation structures would conclude sometimes from

	wishes(P1,X)

that P1 will take action to achieve X, and still other axioms
will tell us about the effects of P1's efforts.

Part of the motivation for proceeding in this modular way is that the general
rules we know how to write about the effects of actions are of limited
scope, and we don't want improvements in action formalisms to require
rewriting everything about emotions.

While the above uses some specific formulas about  hates(P1,P2),  we are
not committed to such a predicate.  It's only illustrative, and
we shall discuss a variety of possible formalisms.

	5. A straightforward situation calculus treatment would
require writing  hates(P1,P2,s)  or better  holds(hates(P1,P2),s)
in (1).  For some purposes, we might do that, but we hope to avoid
it by using contexts (in the sense of unwritten JMC lectures).
We would enter a context referring to a given (variable) situation
and write axioms not mentioning a context.  When we left that
context,  hates(P1,P2)  would be lifted to  holds(hates(P1,P2),s).

Since we advocate a language permitting such lifting, we will ordinarily
avoid writing any parameter that would be replaced by the same
symbol throughout a sentence.  Thus we avoid writing  s, since it
is constant throughout (1).  We don't intend to be commited one
way or the other on whether, when the lifting occurs,  vengeful P1
remains as is or is converted to  holds(vengeful P1,s).

	6. The next question we shall discuss is what to reify.
Here are some possibilities.

		a. The CYC axioms about emotions of 1990 January
regard emotions as unary predicates of a person.  Thus we can
say that someone hates, but we can't say in that formalism whom
he hates.  This is inadequate.  However, it might be worthwhile
to see whether a unary formalism has salvage value.  The idea
is to imagine that the CYC axioms were written within a context
that makes implicit the target of the emotion.  Maybe context
can do this for us, but actually the 1990 January axioms have
too little information to make salvaging them much more than
an exercise in using a context mechanism.

		b. Much of what we want to say about emotions
is conveniently expressed in terms of binary predicates or
binary fluents.  The binary predicates are {\it liftable} to binary
fluents using the context mechanism.

		c. Next consider introducing emotional-state(P)
as a rich object fluent.  (It lifts to  value(emotional-state(P),s),
when the situation is made explicit).  We could say
let

(2)	ee1 = emotional-state P1

and then say

(3)	is-hating ee1

and

(4)	object ee1 = P2.

If we wanted to add that P1 also loved P2, we could add

(5)	is-loving ee1.

To use such a combination would require some rules about combined
hating and loving that would cancel some of the defaults that
operate when only hating is mentioned.

		d. Many of the same considerations apply to different
reifications.  These could include

	hatred P,

i.e. the hate  P  has.  We could write

	P2 = object hatred P1

It might often be worthwhile to enter a context in which the person
with the emotion is assumed, e.g. to be  P1.  Then we could write

	object hatred = P2.


		e. We might introduce completely abstract objects,
and this might be the most elegant solution.  Thus we would have

	is-emotion g0005
	 ∧ actor g0005 = P1
	 ∧ object g0005 = P2
	 ∧ quality g0005 = hatred.

The same actor could then have a variety of emotions at the same
time.

Remarks:

	1. There is no need to make the language  CYC uses to express
facts about emotion correspond to a fixed list of English words.
However, we will start with such a list.

	2. One task for CYC or a program using CYC is answering
questions about a narrative.  The narrative might be expressed
in English or in some logical narrative language, perhaps a
narrative language built into CYC.  Possible domains include
children's stories, Aesop fables or Russian novels.

Specific Emotions

The plan is to discuss them in English first, then write logical
sentences and finally write CYC units.  This way any inadequacies
of the formalisms will not affect our ideas about what needs to
be done, but only their realization.  Actually we might also need
to include some scenarios of what we would hope for CYC to do.
This might turn up some inadequacies of English.

Here first is Guzman's list:

(sadness fear happiness admiration  affection amusement anger anxiety
apathy awe boredom cheerfulness confidence contempt contentment delight
depression resolution disgust dislike elation embarassment enjoyment
excitement friendliness frustration gaiety gratitude grief guilt guile
hate hope impatience inspiration irritation jealousy love nervousness
panic passion pity pride relief remorse resentment reverence serenity
shame solemnity surprise despair courage respect enthusiasm coldness
pleasure security gloominess diffidence buoyancy indecision charm
fearlessness belligerence satisfaction ingratitude innocence patience
humility modesty vanity misery flippancy concern doubt mistrust
certainty abhorrence)

1. Fear.
	Fear usually has an object.  It seems that much of the ground
wanted can be covered if we take this object to be a proposition about
the future.  By default at least, if a person fears a proposition, he
fears any proposition that obviously (to him) implies it.  A person
who fears $p$ wishes that $p$ not happen.  By general reasoning about
wishes, we can infer that preventing  $p$ is likely to be one
of his goals.  We infer that a person fears something by

	a. He or someone else says so.  I suppose this is part of
a general package for inferring the truth of what someone says.

	b. Observing certain physical signs or certain behavior.

	c. Inferring fear from the fact that a certain future
possibility is harmful to him.  Of course, he ought to be aware
that $p$ is possible and harmful.  We need some axioms about harmful
and possible.  We will also need some rules about when a person
can be presumed to be aware of facts.

Digression:

	     There are nonmonotonic shortcut rules that
     bypass reasoning about awareness to go directly from
     facts to conjectures about the actions a person will
     take without even mentioning his awareness.  Such rules
     need to be inferrable from assertions about rules valid
     without special assumptions.  We note that the
     literature on nonmonotonic reasoning doesn't discuss
     inferring the rules themselves.  I incline to the
     opinion that the inference of such rules can be done
     monotonically.  Thus we can monotonically infer a
     nonmonotic rule that a person will leave a burning
     building from a rule that he will leave if he believes
     the building to be burning.  In fact, nonmonotonic
     rules can be inferred rather lightly, so such rules
     need to be selected on the basis of expected
     helpfulness.  If the computer is to infer the rules,
     then it will have to reason about whether they will be
     helful.

	There also need to be some nonmonotonic rules about when a person
doesn't fear something.

	There are some purely psychological facts about fear,
such as that a person experiencing strong fear is likely to
behave irrationally, i.e. without taking into account some facts
that he normally takes into account.

	As with many intentional states, we need to distinguish between
fear as a general attitude and experiencing a specific fear at the moment.
A person may generally fear thunder and lightning but not be experiencing
fear, because the skies are clear.  We could use two terms in a
common sense database, but my inclination is to try to follow English
usage and get by with one, making the distinction with modifiers when
necessary.  In the first place, natural language usages should be given
a presumption of semantic appropriateness.  In the second place there
are intermediate cases.  For example, a person running to get under
shelter to get away from a thunder storm may be worrying about falling
down at the moment, so even though his actions are motivated by fear,
at the moment he is afraid of something else.
This means that the formalism needs to be able to provide for several
simultaneous fears.

	There is a general need to distinguish background mental
fluents from those in consciousness at the moment.  It applies to
beliefs, desires and intentions also.

Digression:

	     The use of functions in a formal language is
     rather like the use of ``the'' in English.  Its use
     assumes that there is exactly one object having a given
     relation to the operand.  Suppose we have both a
     relation $son(x,y)$ meaning that $y$ is one of $x$'s
     sons and a function $son(x)$.  We would like to allow
     the use of $son(x)$ without warning.  What it means
     depends on where it occurs.  For example, a formula
     [1990 Feb 14: details on this postponed till they are
     worked out, but we even want to allow (∃son(x)( ... )].

x fears y ---→ x anticipates harm from  y.
**
x fears y ---→ tendency(x  flees  y)

2. Amusement

	This one will be difficult and probably ought not to be attempted
yet.  Axiomatizing what's funny has always seemed to me very hard.
Perhaps amusement could be treated in a superficial way that would be
somewhat circular, i.e. finds something funny.  At least that would
allow inference of one concept from other related concepts.

3. Apathy

	This is a kind of  ab,  because an apathetic person will not
perform actions that he normally would perform.  It can be inferred
by failure to perform some actions and also leads to inferring that some
other actions also won't be performed.  Causality may be chemical
in blood, e.g. from fatigue or illness, but apathy also arises from some kinds
of experiences.

4. Contempt

	It is associated with a belief that a certain person $P2$ doesn't
count for some purpose.  P2 may have done something contemptible; in this
case  P2  is excluded from the class of people who have to be respected.
Perhaps contempt is the opposite of respect.  Facts about  P2  are regarded
as unimportant.  This has to tie in with more general axioms about decisions
that certain classes of facts are unimportant which leads to their not
being retrieved in deciding on action.  At first I thought that contempt
would require some specific axioms relating it to action, but now it doesn't
seem so.

5. Awe

	I mention awe in order to remark that it makes sense only in more
sophisticated contexts than we need immediately consider.  Also it's a
rather vague concept.  We could put a few axioms about it in CYC's database
regarding it as intense respect, but we should also put in something so
that CYC will be modest about its own understanding of it and suspicious
that the term isn't being used in any precise way by an interlocutor.

6. Admiration

	This is still vague but easier than awe.  To admire  x  is
to think it is a good whatever it is.  If  x  is a person, then one
wishes one had some of his good qualities.  What action this might
lead to is relegated to the axioms about implementing wishes.  If
x  is an object, perhaps one wishes one possessed such an object
and admires in some respect either the possessor or the maker of the
object.

7. Disappointment

	disappointment causes unhappiness

	When it becomes believed that an expected favorable event won't
occur, disappointment results.

Advice for programs: 

	Don't disappoint people.

8. Contentment

	This is a goal.  Moreover, it may be caused by achieving other
goals.  It is incompatible with certain negative emotions.  It may take
an object.  He is content with the way his son's education is going.
It is not readily predicted, i.e. CYC (or another person) may not be able to
predict whether a person will be content if certain events occur.