Causality and Modality

1
Causality and Modality
Jerry R. Hobbs USC Information Sciences
Institute Marina del Rey, CA
2
Where Are We?

• Introduction Core theories of commonsense
• knowledge and their relation to the lexicon
• Framework Logic and abduction
• Cognition and the cognitive lexicon
• Time and now
• Causality and modality
• Similarity and like

3
Outline

• A Theory of Causality
• Causal complexes
• Some classical issues
• The predicate cause
• The causal lexicon
• Modality The case of would
• Examples
• Data

4
Aims
1. A rigorous, monotonic account of
causality, as complex as necessary
(frontload complexity) 2. A simple and
serviceable treatment of causality,
amenable to fast inference, grounded in 1
5
A Principle in EncodingCommonsense Knowledge
Most interesting concepts cant be defined
with necessary and sufficient conditions. The
most we can hope for is to characterize
them with necessary conditions and
sufficient conditions.
No rules of form cause(e1,e2) lt--gt
.... Lots of rules of form cause(e1,e2) --gt
.... and .... --gt cause(e1,e2)
6
Scope
Want an account of causality that works for
physical causation folk psychological
causation social causation political
causation .... Causes and effects can be
actions, agentless events, and states.
His not signing the contract caused it to be
invalid.
7
Explanation in Discourse
The police prohibited the women from
demonstrating. They feared violence. Logical
Form prohibit'(p1,p,d) demonstrate'(d,w)
CoherenceRel(p1,f,p1) fear'(f,y,v)
violent'(v,z)

cause(f,p1) Knowledge Base
fear'(f,p,v) --gt diswant'(d2,x,v)
cause(f,d2) demonstrate'(d,w) --gt
cause(d,v) violent'(v,z) cause(d,v)
diswant'(d2,p,v) --gt diswant'(d1,p,d)

cause(d2,d1) diswant'(d1,p,d)
authority(p) --gt prohibit'(p1,p,d)

cause(d1,p1) cause(e1,e2)
cause(e2,e3) --gt cause(e1,e3)
8
Explanation
The police prohibited the women from
demonstrating. They feared violence. Logical
Form prohibit'(p1,p,d) demonstrate'(d,w)
CoherenceRel(p1,f,p1) fear'(f,y,v)
violent'(v,z)

cause(f,p1) Knowledge Base
fear'(f,p,v) --gt diswant'(d2,x,v)
cause(f,d2) demonstrate'(d,w) --gt
cause(d,v) violent'(v,z) cause(d,v)
diswant'(d2,p,v) --gt diswant'(d1,p,d)

cause(d2,d1) diswant'(d1,p,d)
authority(p) --gt prohibit'(p1,p,d)

cause(d1,p1) cause(e1,e2)
cause(e2,e3) --gt cause(e1,e3)
9
Notational PreliminariesOntological Promiscuity
tall(x) x is tall tall(e,x) e is the
eventuality of xs being tall holds(e,w) e
holds/happens/obtains in possible world
w Rexists(e) e holds/happens/obtains in the
real world (A x) p(x) lt--gt (E e) p(e,x)
Rexists(e) and(e,e1,e2) e is the eventuality
of both e1 and e2 holding not(e1,e) e1 is the
eventuality of e not holding
(abbreviate e1 as e) Temporal facts are
properties of eventualities
before(e1,e2), at-time(e,t) Both type and token
eventualities (type typical element of set
of token eventualities)
10
Another Principle of EncodingCommonsense
Knowledge
When you focus on a special phenomenon, use
a special-purpose logic that highlights the
phenomenon. When the research is part of a
larger enterprise of encoding commonsense
knowledge for everyday planning, natural
language understanding, etc., use a simple
and uniform logic.
11
Causal Complex
causal complex
s
causal-complex(s,e) e1 ? s, ....
e1
e2
e3
e
e4
....
effect
When every event or state in s happens or holds,
then e happens or holds. All eventualities
in s are relevant.
12
Closest World
w1
w2
e1
closest-world(w2,w1,e1,C) lt--gt holds(e1,w1)
holds(e1,w2) (A e2 ?? w2-w1) (w1 ?
w2) ? C ? e1 --gt e2 (w1
? w2) ? C -/-gt e2 w2 is a closest world to w1
when e1 is added to w1, under constraints
C. C a --gt b e1 a
w1 a, b, c, d w2 a,
b, c, d
All changes are precipitated by e1
13
Change Relevance
change-relevant(e1,e) lt--gt (E w1,w2)
closest-world(w2,w1,e1,C)
holds(e,w1) lt--gt holds(e,w2) Changing e1
changes e. causal-complex(s,e) e1 ?? s
--gt change-relevant(e1,e) Eventualities in a
causal complex are relevant to the effect.
14
Example Switches
e
e1
w2 is w1 except that e1 is toggled. If light
toggles, e1 is change-relevant to e.
15
Example Dominos
gt
e1
e
e1
e
w1 all up
w2 all down
So e1 is change-relevant to e.
gt
e1
e1
e
e
w1 all down
w2 e1 up
So e1 is not change-relevant to e.
No way to lift e by lifting e1.
16
Outline

• A Theory of Causality
• Causal complexes
• Some classical issues
• The predicate cause
• The causal lexicon
• Modality The case of would
• Examples
• Data

17
Causality and Time
Greek wall, Keramikos, Athens, Greece, 400s BC
Inca wall, Cuzco, Peru, 1400s AD
Proof of the influence of the Incas on Greek
architecture
18
Causal FlowGun Example
Constraints
Bang
Fire
Death
Is Bang change-relevant to Death? Under
current definitions Yes.
19
Causal Flow and Time
causally-involved(e1,e) lt--gt (E s)
causal-complex(s,e) e1 ?? s causally-involved(e
1,e) --gt before(e,e1) causally-involved(e1,e2)
causally-involved(e2,e1) --gt
causally-prior(e1,e2) causally-involved(e1,e)
before(e1,e) --gt causally-prior(e1,e)
Bayesian nets typically respect causal
flow. Ortiz LP rules for prediction
LE rules for explanation
favor these in computing closest world
20
Loosening Constraints
Constraints in DNF Fire v Bang
Bang v Fire
....
If negative instance of e1 occurs in a constraint
c with some positive instance of a causally
prior e2, then disjoin abc to the constraint.
Bang v ab1 v Fire
Call this Defeas(C,e1)
21
Accommodating Causal Priority
closest-world(w2,w1,e1,C) lt--gt w1 ? Con(S,C)
w2 ? Con(S,Defeas(C,e1)) holds(e1,w1)
holds(e1,w2) (A e0 ? w2)
causally-prior(e0,e1) --gt e0 ? w1 (A
e2 ?? w2-w1) (w1 ? w2) ? Defeas(C,e1) ? e1 --gt
e2 (w1 ? w2) ?
Defeas(C,e1) -/-gt e2 change-relevant(e1,e) lt--gt
(E w1 ? Con(S,C) ,w2 ? Con(S,Defeas(C,e1)))
closest-world(w2,w1,e1,C)
holds(e,w1) lt--gt holds(e,w2)
22
Accommodating Causal Priority
Turning on e1 does not change its causal priors
closest-world(w2,w1,e1,C) lt--gt w1 ? Con(S,C)
w2 ? Con(S,Defeas(C,e1)) holds(e1,w1)
holds(e1,w2) (A e0 ? w2)
causally-prior(e0,e1) --gt e0 ? w1 (A
e2 ?? w2-w1) (w1 ? w2) ? Defeas(C,e1) ? e1 --gt
e2 (w1 ? w2) ?
Defeas(C,e1) -/-gt e2 change-relevant(e1,e) lt--gt
(E w1 ? Con(S,C) ,w2 ? Con(S,Defeas(C,e1)))
closest-world(w2,w1,e1,C)
holds(e,w1) lt--gt holds(e,w2)
23
Accommodating Causal Priority
Turning on e1 does not change its causal priors
closest-world(w2,w1,e1,C) lt--gt w1 ? Con(S,C)
w2 ? Con(S,Defeas(C,e1)) holds(e1,w1)
holds(e1,w2) (A e0 ? w2)
causally-prior(e0,e1) --gt e0 ? w1 (A
e2 ?? w2-w1) (w1 ? w2) ? Defeas(C,e1) ? e1 --gt
e2 (w1 ? w2) ?
Defeas(C,e1) -/-gt e2 change-relevant(e1,e) lt--gt
(E w1 ? Con(S,C) ,w2 ? Con(S,Defeas(C,e1)))
closest-world(w2,w1,e1,C)
holds(e,w1) lt--gt holds(e,w2)
This loosening allows that
24
Example Gun Again
Constraints Fire --gt Bang Bang ab1 --gt
Fire ....
causally-prior(Fire,Bang), causally-prior(Fire,De
ath)
Closest world w1 Fire, Bang,
Death, ?ab1 w2 Fire, Bang,
Death, ab1 So Bang is no longer
change-relevant to Death.
25
Structure in Causal ComplexesExample Vase
L I let go of vase F vase falls B vase
breaks
Four Causal Complexes
L F B
F B
L B
L F
So,
L F B
26
Composing Causal Complexes
s1
s2
e2
e1
if consistent
The union is a causal complex for e2.
s1?s2
e1
e2
27
Counterfactuals
If John werent a millionaire, he wouldnt have
retired.
cause(millionaire(J), retire(J))
Lewis, Ortiz, Pearl Counterfactuals are
priveleged evidence
about causation. Pearl A way to capture
notion of closest world. Me We have good
intuitions about how to figure out what
is in or out of a causal complex.
Develop theory of causality. Use it to
characterize words like if, would,
subjunctive. Use that to interpret
counterfactuals.
28
Causality and Implication
No good characterization of the difference
between causality and implication.
chimpanzee(x) --gt monkey(x)
implies, not causes
Counterfactuals are of no help If Bonzo
were a monkey, he wouldnt be a chimpanzee.
Possibilities 1. Implication is a kind of
washed out causality, in the
informational domain 2. Metonymy
implication causes beliefs to propagate
imply(p,q) lt--gt cause(bel(a,p), bel(a,q))
29
Pre-emption
Problem for Counterfactual Account
P A poisons Cs canteen H B drills hole in
Cs canteen E Cs canteen is empty D C dies
If B hadnt drilled a hole in Cs
canteen, then C wouldnt have died.
But the causal complex account is clear
H
E
D
"Gosford Park"
E
P
30
Causality and Probability
s1 is a causal complex for e with probability
p.
s
s1
e
s1 ? s, causal-complex(s,e), and s-s1
occurs with probability p.
Probabilistic models of causality provide a good
intermediate level between causal complexes
and .....
31
Outline

• A Theory of Causality
• Causal complexes
• Some classical issues
• The predicate cause
• The causal lexicon
• Modality The case of would
• Examples
• Data

32
Cause
In a causal complex, some eventualities are
distinguished as causes.
Causes are the focus of planning,
prediction, explanation, interpreting discourse (b
ut not diagnosis)
presumable
power on
finger in socket
shock
cause
What is presumable depends on task,
context, knowledge base, ....
33
Presumable
Why an eventuality e might be presumable e
almost always holds es holding is the norm
(statistical or telic) e is known to hold
e holds in all cases of current interest Why
an eventuality e may be labelled a cause e
often doesnt hold e is not known to hold
e is an action by the agent e is the
last event in a complex of events e is
required more frequently
34
Cause vs. Enabling Condition
(Cheng Novick, 1991)
Focal set S of situations is determined by
context. Effect occurs in subset S1 of S,
e true in (almost) all situations in S
gt e is enabling condition e true in
(almost) all situations in S1, but (almost)
no situations in S-S1 gt e is
cause Turn this around Pick e's that will
(almost) always be true Focal set is a set
of situations where these e's hold. Enabling
conditions are stipulated. Context what
eventualities are defeasibly true.
35
Presumable
These are reasons e might almost always be true
Why an eventuality e might be presumable e
almost always holds es holding is the norm
(statistical or telic) e is known to hold
e holds in all cases of current interest Why
an eventuality e may be labelled a cause e
often doesnt hold e is not known to hold
e is an action by the agent e is the
last event in a complex of events e is
required more frequently
36
Causes in Causal Complexes
cause(e1,e) --gt (E s) causal-complex(s,e) e1 ?
s (A e2 ?
s-e1) presumable(e2)
v
trigger(e1,s,e2)
s
e2
e
e1
e3
37
Expressing Causal Knowledge
Not (A s) ... some long characterization
of s ... --gt (E e) q(e,...)
causal-complex(s,e) but (A e1,x) p(e1,x)
--gt (E e) q(e,x) cause(e1,e) or in diagnosis
(A e,x) q(e,x) --gt (E s,e1)
causal-complex(s,e)
p(e1,x) e1 ? s
(e.g., Winograd example)
38
Defeasibility of Cause
Axioms like (A e1,x) p(e1,x) --gt (E e)
q(e,x) cause(e1,e) are really (A e1,x)
p(e1,x) abi(e1,x) --gt (E e) q(e,x)
cause(e1,e) where abi(e1,x) lt--gt (A e,s)
causal-complex(s,e) e1 in ? s
q(e,x)
presumable(s-e1)
--gt (E e2) e2 ? s-e1 Rexists(e2)
(something presumable doesn't hold)
39
General Properties of Cause
Causes imply effects (causal modus ponens)
causal-complex(s,e) Rexists(s) --gt
Rexists(e) So defeasibly cause(e1,e)
Rexists(e1) --gt Rexists(e) Events have
causes Rexists(e) eventuality(e)
--gt (E e1) Rexists(e1) cause(e1,e) Cause
and time/causal flow cause(e1,e) --gt
before(e,e1) cause(e1,e) --gt
causally-prior(e,e1)
40
Transitivity of Cause
cause(e1,e2) cause(e2,e3) ab2(e1,e2,e3) --gt
cause(e1,e3)
something presumable in ccf(e1,e2) or ccf(e2,e3)
is not true or ccf (e1,e2) U ccf(e2,e3) is
inconsistent
where ccf(e1,e2) is causal complex by
virtue of which e1 causes e2.
41
Transitivity of Cause?
Hart and Honoré The cold cause the road
to ice over. The icy road caused the
accident. The cold caused the accident. NY
Daily News 20 DIE IN HEAT WAVE Moens and
Steedman Johns leaving caused Sue to
cry. Sues crying caused her mother to be
upset. Johns leaving caused Sues mother to
be upset. Sues Mother Look what he did to
me!
42
Antisymmetric and Antireflexive?
Book A leaning slightly to right
Feedback loops
causes
Book B leaning slightly to left
43
Causality and Planning
Planning is exploiting causal knowledge to
achieve things in the world.
s
e2
e1
e3 e4
Operator s Body e1 Preconditions e2 Add
e3 Delete e4
Effects e3 e4
44
Enablement
Power --gt Touch causes Shock Power causes
Shock
P
gt
S
P
S
T
There is a world in which toggling to P toggles
S. gt P is change-relevant to S. gt It is
consistent for P to be in a causal complex for
S. P is presumable gt P is not
presumable gt It is consistent that
cause(P,S)
enable(e1,e2) lt--gt cause(e1,e2)
45
Outline

• A Theory of Causality
• Causal complexes
• Some classical issues
• The predicate cause
• The causal lexicon
• Modality The case of would
• Examples
• Data

46
Lexicon of Cause
Compare definitions of prevent,
enable, let, help, maintain in Ortiz's theory of
causality (1999) and in mine.
47
Ortizs Theory of Causality
Three relations among events
counterfactual implication
accommodates selective lifting of constraints
and the frame problem strongest
antecedent condition also handles
pre-emption cause also excludes
noncausal implicational relations
e.g., method-of Shies away from causation of
negative eventualities.
48
Prevent
Me prevent(e1,e2) lt--gt cause(e1,e2) Orti
z prevent(e1,e2) occurs(e1,t1)
--SAC--gt holds(possible(occurs(e2,t
2)),t2) SAC vs cause Johns sitting
prevented him from standing up. Ortiz should
have used "cause"
49
Why Possibility in Prevention
Ortiz His closing the barn door prevented
the horse from escaping. gt ? His closing the
barn door caused the horse not to escape. gt
If he hadn't cosed the barn door, the horse
would have escaped.
(Maybe the horse wouldn't have tried) But this
makes the horse trying to escape play a causal
role in its not escaping. Possibility is a
property not of prevention
but of discourse about
prevention.
50
What is Possible?
cause(e1,e2) vs. cause(1,possible(e2))
possible(e2) lt--gt cause(C,e2)
So if cause(e1,possible(e2)) then e1 brings
about a change from C to C where cause(C,
e2) cause(C, e2) e1 must be in or
implied by C cause(CUe1, e2)
51
Enable
Me enable(e1,e2) lt--gt cause(e1,e2) Orti
z enable(e1,e2) occurs(e1,t1)
--SAC--gt holds(possible(occurs(e2,t2)
),t2) If e1 had not occurred, then e2
would not have been possible. SAC vs
cause Not stopping on the way home
enabled him to keep driving. Should
have used "cause"
52
Let
Me let(e1,e2) lt--gt cause(e1,e2) Ortiz
let(e1,e2) prevent(e1,e2), manner
refraining Does e2 have to be possible?
Dan Rather is letting me give this
talk. Constraints on arguments of cause that
e and e are possible.
53
Help
help1(a,b,e) lt--gt (E s1,s2) agent(a,s1)
agent(b,s2)
help(s1,e) help(s2,e)
goal(b,e) Me help(s1,e)
lt--gt (E s) subset(s1,s)
causal-complex(s,e) Ortiz
help(s1,e) s1 enables e to occur with a
limited use of
resources (No theory of resources) s1
set of eventualities of utilizing bits of
resources hinder(s1,e) lt--gt help(s1,e)
54
Maintain
Me maintain(e1,e) lt--gt cause(e1,change(e,
e)) maintain1 (a,e) lt--gt (E e1)
agent(a,e1) maintain(e1,e) Agents
causal role cause Ortiz maintain1(a,e)
a successfully hinders e Agents causal
role expend resources
55
Outline

• A Theory of Causality
• Causal complexes
• Some classical issues
• The predicate cause
• The causal lexicon
• Modality The case of would
• Examples
• Data

56
CounterfactualsKnowledge Base
drink'(d,j) --gt sick'(s,j) cause(d,j) had8(e)
--gt not(e) cause(d,s) --gt if(d,s) cause(d,s)
--gt would(s) would(s) --gt not(s) goal'(g1,j,s)
cause(d,s) --gt goal'(g2,j,d)
cause(g1,g2) cause(d,s) etci(d,s) --gt
imply(s,d) imply(e1,e2) --gt if(e1,e2)
57
Counterfactuals Example 1
If John had drunk too much, he would have got
sick.
if(d,s) had8(d) drink'(d,j) would(s)
sick'(s,j)
cause(d,s)
58
Counterfactuals Example 2
John didn't drink too much. He would have got
sick.
holds(d) drink'(d,j) CoRel(d,s)
would'(w,s) sick'(s,j)
cause(g1,g2)
goal'(g1,j,d)
cause(d,s)
goal'(g2,j,s)
59
Counterfactuals Example 3
If John had not drunk too much, he would not have
got sick.
if(d,s) had8(d) drink'(d,j) would(s)
sick'(s,j)
imply(d,s)
cause(d,s)
etci(d,s)
60
Causality and Modality
(Frank Kamp, 1997)
I dont own a TV set. I would watch it all the
time.
Axioms If c causes e4, then e4 would happen
given constraints c. cause(e3,c,e4)
--gt would(e3,e4,c) If doing something is
bad for you, that causes you not to do it.
cause(e3,e2,e4) bad-for(e4,i) p(e4,i)
--gt cause(e3,e1) not(e1,e2)
Owning something caused you to use it.
own(e2,i,t) --gt cause(e3,e2,e4)
use(e4,i,t) Watching TV is bad for you.
watch(e4,i,t) tv(t) --gt bad-for(e4,t)
Using a TV is watching it. tv(t)
use(e4,i,t) --gt watch(e4,i,t) Discourse
segments can be linked by causal explanation.
cause(e3,e1) --gt CoherenceRel(e1,e3)
61
Causality and Modality
I dont own a TV set. I would watch it all the
time.
LF
Rexists(e1)
own(e2,i,t)
CoherenceRel(e1,e3)
Rexists(e3)
tv(t)
would(e3,e4,c)
watch(e4,i,x)
not(e1,e2)
Proof
cause(e3,e1)
ce2
xt
cause(e3,e2,e4)
bad-for(e4,i)
watch(e4,i,t)
tv(t)
use(e4,i,t)
own(e2,i,t)
62
Outline

• A Theory of Causality
• Causal complexes
• Some classical issues
• The predicate cause
• The causal lexicon
• Modality The case of would
• Examples
• Data

63
Uses of would
The data Carson McCullers novel 26
SJMN business articles 20 Science
articles on AIDS 9 Shakespeares
sonnets 23 Meeting transcripts
30 Country western lyrics 23 Total
131
11 examples, all from Shakespeare, had old sense
of would as want to Tired
with all these, from these I would be gone.
Question Is there an explicit cause mentioned
near the occurrence of would in the other
120?
64
Causes in Clauses
Cause is syntactically related to would-event
66 Cause is subject of would-event 27
... the standard would allow intelligence
agencies to spy on private
companies and individuals. Often these are
structure causes function ... to
provide generic products wouldnt serve the
clientele. Often these are evaluative
It would obviously be nice to have data from
more patients. Cause in main clause, would in
subordinate 1 ... the London routes sale
is another major stumbling block before
the combination would assume control of TWA.
65
Causes in Clauses
Cause is syntactically related to would-event
66 Cause is in adjunct 38 Cause is in
subordinate clause with because, every time,
if, once, were, or when 20
Were he reading this article, he would have
finished by now. Cause is in PP with at,
by, for, from, in, with 12
At the mere mention of the words, her face would
darken with shame. Cause is
in infinitival adjunct 5 Volitional
Reversal If A causes B, then
wanting B causes wanting A Miss Amelia
would not leave him by himself to suffer with
this fright. Source of
want to meaning? Cause is in gerund 1
Miss Amelia, being rich, would not go out
of her way to murder a vagabond
for a few trifles of junk.
66
Causes in Discourse
Cause is in clause or sentence adjacent to
effect 11/120 Possible - would 5
Amino acid substitutions in B cell epitopes can
abrogate immune reactions. These
escape variants would have a survival advantage.
Desire - would 2 I wanna make you
mine forever. Theres nothin on this earth
I would not do. (Volitional Reverse)
Contrastive Violated Expectation pattern 4
There were those who would have courted her, but
Miss Amelia cared nothing for the love
of men....
67
Cognitive, Communicative, andHabitual Contexts
would in cognitive or communicative contexts
11/120 Ive seen you smile more than I
thought you would. He said Kerkorian would
probably discuss the offer with Ichan
this weekend. The cause is the cognizers or
speakers cognitive state. would in habitual
contexts 9/120 Whatever happened to
old-fashioned love, The kind that would see
you through? Some aspect or property of the
time period is the cause.
68
Hedges
Causal meaning of would is washed out into a
hedge 14 would with like, rather, be
inclined, not mind 8 I have five
different subtasks Id like my people to
describe. The request is mitigated because
the preference is the effect of some
implicit cause. would with verb of cognition
6 I would have thought youd want to show
him the stuff he likes best first, get
him in a good mood. would mitigates a
disagreement by hinting at causal factors
beyond the speakers control. The
rearrangement of the H3 loop would seem to be
primarily a result of accommodating
TyrP105 of the peptide. The claim is
mitigated because although there is obvious
structural cause, theres no experimental
verification.
69
would and Truth
would is often used with unreal situations, but
this is not necessarily the case He
said he had no idea why the investigator would
ask for the arbitration documents. The
abstract causal nature of would makes is
useful in hypothetical contexts.
70
What Causes What?
What general categories of causality are there?
Monotonic scale-scale causality
Structure causes function Volitional
reverse Intermediate causes, transmitted
causality But .... There are just a lot
of specific facts
71
Cause as MonotonicScale-to-Scale Function
cause(change(at(x,y), at(x,z)), change(at(w,u),
at(w,v))), y lt z, u lt v "The more ....,
the more ...." Rules "The more you press
on the accellerator, the faster you go."
.
.
.
.
The more, the merrier. The bigger they are, the
harder they fall. The higher, the fewer. The more
axioms, the fewer models.
72
Cause as MonotonicScale-to-Scale Function
"The more ...., the more ...." Rules in the
would data The more information we have,
the better analysis we can do. The older one
is, the less love one finds. The more for
you, the less for me. The more the effort
thats required, the higher the value has
to be. The easier the task, the greater
ones willingness to do it. The worse the
problem, the higher ones priority for solving
it. .....
73
Transmitted Causality
cause is transitive cause(e1,e2)
cause(e2,e3) --gt cause(e1,e3) Using this rules
backwards decomposes a causal chain into
its parts e2 transmits causality from e1 to
e3. Causal connections via "vector bosons"
cause(p(x),q(y)) lt-- cause(p(x),
move(z,x,y))
cause(move(z,x,y),q(y)) e.g., photon,
raindrop, virus, utterance Allows
decomposition of events into causal chains.
74
Event Structure in Language
cause Agent ---------gt
change(Instr)
cause
---------gt change(Object)
cause
---------gt
change(Beneficiary) John pounded the
nail with a hammer for George.
75
Lots of Specific Facts about Causes
In the domain of cognition in the would data
Being able to control X causes one to prefer
X. A theory causes an expectation.
Wanting X causes one to try to achieve X. X
happens causes one to want to explain X.
Knowledge causes feelings. Good/bad feelings
can cause good/bad actions. Perception
causes knowledge. An obligation causes one
to accept it as a goal. Love causes
happiness. When axiomatizing a domain, we simply
need to encode its causal relations.
76
Summary
Causal complexes can be reasoned about
monotonically, but we rarely have complete
knowledge of them. Cause provides a useful
notion for most causal reasoning, but is
nonmonotonic. Some general properties of
cause. Some common lexical items defined in
terms of cause, including modals. Language
can be used to discover our causal
knowledge.