Ontologies of Information Structure and Commonsense Psychology

1
Ontologies of Information Structure
andCommonsense Psychology

• Jerry R. Hobbs
• USC/ISI
• Marina del Rey, CA

2
Information Structure
3
Motivation
The best answer to a question is often a
diagram, a graph, a map, a photograph, a
video. What is the Krebs cycle? How has the
average height of adult American males
varied over the years? How did the Native
Americans get to America? What does Silvio
Berlusconi look like? What happened on September
11, 2001?
4
Grounding Symbols in Cognition
cause(perceive(a,x), cognize(a,c))
object state event process absence ...
concept (including propositions)
5
Grounding Symbols in Cognition
cause(perceive(a,smoke), cognize(a,fire)) cause(pe
rceive(a,cloud), cognize(a,dog)) cause(perceive(a,
bell), cognize(a,food)) No necessary causal
connection between x and c This schema makes
symbols possible cause(perceive(a,x),
cognize(a,concept-of(x))) but other things as
well
6
Grounding Symbols in Cognition
cause(present(b,x,a), perceive(a,x)) cause(percei
ve(a,x), cognize(a,c))
cause(car beeps, driver hears beep) cause(driver
hears beep, driver remembers seat belt)
7
Intention and Convention in Communication
Unintentional fidget --gt nervous
ouch --gt pain Presenter intends
concept, but not recognition of intent
my door is closed --gt Im not in
8
Recognizing Intent
know(b, cause(present(b,x,a), cognize(a,c))) goal
(b,cognize(a,c)) goal(b,g1) know(b,
cause(g2,g1)) etc --gt goal(b,g2) So b has
goal present(b,x,a), an executable action,
so he does it. a looks for causal explanation of
present(b,x,a) and comes up with exactly
this Intention is recognized.
9
Gricean Nonnatural Meaning
If an agent b has a goal g1 and g2 tends to cause
g1, then b may have as a goal that g2 cause
g1. If an agent b has as a goal that g2 cause
g1, then b has the goal g2. When a
recognizes this plan, he will recognize not only
bs goal to have a cognize c, but also bs
intention that a do so by virtue of the
causal relation between bs presenting x
and as cognizing c.
10
Mutual Belief and Convention
Mutual Belief mb(s,p) member(a,s) --gt
believe(a,p) mb(s,p) --gt mb(s,mb(s,p)) Struc
ture of a communicative convention mb(s,
cause(present(b,x,a), cognize(a,c)))
where member(a,s) and member(b,s) e.g. red flag
with white diagonal in community of boaters
means diver below
represent(x,c,s)
symbol
content
11
Composition in Symbol Systems
Symbol System
Content Domain
composite symbol
complex concept/ proposition
interpretation
composition operation
composition operation
basic concept/ proposition
atomic symbol
interpretation
12
Speech and Text(within sentences)
Symbol System
Content Domain
sentences a man works
complex proposition man(x) work(x)
interpretation
pred-arg rels, conjunction
concatenation
basic propositions man(x), work(y)
words a man, works
interpretation
13
Speech and Text(Discourse)
Symbol System
Content Domain
augmented discourse meaning
interpretation
discourse
coherence relations causality,
similarity, figure-ground
concatenation
sentence meanings
sentences
interpretation
14
Tables
R
b
a
R(a,b)
Spatial arrangement gt predicate-argument
relations
15
Beeps in Car
Atomic symbol beep gt somethings
wrong Composite symbols beep ..... beep
..... beep ..... gt fasten seat belt If car
is running beep beep beep beep gt door is
still open If car is off beep beep beep
beep gt lights are still on
16
Maps
Underlying regions of single color/pattern
meaningful region
gt
gt
icons
entities
overlay icons on field
location of entities
gt
gt
labels
names
icon and label adjacent
gt
name of entity
internal structure of icons
categories of entities
gt
17
Process Diagrams
(Futrelle, 1999)
adjacent groups w arrows between
state transitions
adjacent grouped icons
states
icons
entities
18
Documents
(Scott Powers, 2003)
Title Body Adjacent Paragraphs (mod Page, Col
breaks) Diagram near description
Conveys content of body Main detailed
content Read sequentially Coreference
Similarly, Web pages, PowerPoint presentations,
...
19
Face-to-Face Conversation
Atomic elements Speech, prosody Facial
expression Gaze direction Body
position Gestures w hands and
arms Composition operators Temporal
adjacency Temporal synchrony
Need to determine meaning/function of various
behaviors
20
Larger-Scale CommunicativePerformances
Lectures w PowerPoint slides Plays Demos ....
21
Coreference
Two noun phrases Icon and label Same icon in
two state groups in diagram Region of photo and
noun phrase in caption and phrase in
text Iconic gesture and phrase in speech
Useful for image search by keywords
22
Modalities and Media
(Hovy Arens, 1990 Allwood, 2002)
Channels of perception optical, acoustic,
chemical, pressure,

temperature, ...
Greatest opportunities for composition
Communication devices Primary speech,
gesture Secondary writing, drawing,
telephones, videotape,
computer terminals, .... Advantages and
disadvantages of each e.g., visual exploit
2-D structure to convey relations
23
Manifestations of Symbolic Entities
(Pease Niles, 2001)
We group together classes of symbolic entities
sharing same content and call them first
class entities. manifest(x1,x)
represent(x,c,s) --gt represent(x1,c,s) (defeasibl
y -- if P is in content of x then defeasibly it
is in content of x1)
A particular performance of Hamlet
A copy of that videotape
The performance of Hamlet
A videotape of that performance
Hamlet the play
The text of Hamlet
An edition of Hamlet
A copy of that edition
24
CommonsensePsychology
(work with Andrew Gordon, USC/ICT)
25
Methodology
(Gordon, 2000)
Agents plan, so to discover what agents know,
investigate strategies. Picked 10 planning
domains politics, warfare, personal
relationships, artistic performance, sales,
immunology, animal camoflage, ... Interviewed
experts to learn strategies Resulted in 372
strategies Rewrote strategies in controlled
vocabulary -- 988 terms Classified terms into 48
representational areas (space, time, ...)
18 general knowledge 30 commonsense
psychology Enrich each representational area by
text mining Formalize
26
Methodology
(Gordon, 2000)
Agents plan, so to discover what agents know,
investigate strategies. Picked 10 planning
domains politics, warfare, personal
relationships, artistic performance, sales,
immunology, animal camoflage, ... Interviewed
experts to learn strategies Resulted in 372
strategies Rewrote strategies in controlled
vocabulary -- 988 terms Classified terms into 48
representational areas (space, time, ...)
18 general knowledge 30 commonsense
psychology Enrich each representational area by
text mining Formalize
Machiavelli Sun Tzu his wife
27
Theories So Far
Memory Knowledge Management Envisioning
(Thinking) Goals and Planning
28
Why is Memory Important?
We plan to remember actions/information at the
appropriate time. We are responsible for
remembering. Why was Mary angry that John
forgot her birthday? But forgetting
is often a less serious breach than some
other reason. Why didnt you get me a
present? I forgot it was your
birthday. vs. I didnt want to.
29
Naive Model of Memory
Focus of Attention
concept
store
retrieve
concept
Memory
If in memory, then it was stored
30
Accessibility
Concepts in memory have varying accessibility.
concept-1
concept-2
concept-3
threshold
Concepts not retrievable
concept-4
31
Associations and Accessibility
concept-1
concept-0
concept-3
concept-2
concept-1
concept-3
concept-4
concept-2
concept-4
32
Associations and Accessability
Thinking of concepts makes associated
concepts more accessible. This give agents
partial control over memory
retrieval. Technique of memorization Rich
associations.
33
Remember and Forget
in memory above accessibility threshold --gt
remember retrieve --gt remember cause self to
retrieve --gt remember cause self to retrieve
after some effort --gt remember forget concept
lt--gt concept drops below accessibility
threshold
34
Remembering for a Time
We store concepts in memory until we need them
and then forget them. Where did I park my
car today? vs. Where did I park my car on
January 4? We use memory to satisfy knowledge
prerequisites for planned actions.
35
Knowledge ManagementBelief
Reify agents and propositions
believe(a,p) Reasoning is possible inside
belief believe(a,p) believe(a,p--gtq)
etc --gt believe(a,q) Perception causes belief
(seeing is believing) Communication tends to
cause belief BDI We act in ways that maximize
satisfaction of our goals, given our beliefs
36
Graded Belief
(Friedman Halpern, 2001)
0 ??gb(a,p) ?? 1 gb(a, pq) ??min(gb(a,p),
gb(a,q)) gb(a, p p--gtq) gb(a,q) gb(a,p)
1 lt--gt believe(a,p) The higher the graded
belief, the more likely agent is to act on
it
37
Knowledge Domain
Sentence set of propositions a claim
king(x,France)
bald(x) Knowledge domain Has a set of
characteristic predicates Is a set of
sentences all of whose claims have predicates
that are in the characteristic set Expert
Agent is defeasibly an expert in a knowledge
domain if agent knows sampling of
facts in the knowledge domain
(tests, inference from displays of knowledge)
propositional content
claim
38
Mutual Belief
mb(s,p) member(a,s) --gt believe(a,p) mb(s,p)
--gt mb(s,mb(s,p)) These rules are mutually
believed Can show that if a knows b is a member
of s and a knows s mutually believes
p, then a knows that b believes p Inference
of who knows what / who is an expert in what
from membership in communities
39
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. causally-involved(ei,e)
40
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, ....
41
Envisioning (Thinking)
Causal System
e1
e9
e4
e7
e2
e6
e10
e5
e8
e3
e11
es are causally involved
42
Envisioning
e4
Contiguous causal systems
e6
e5
e1
e4
e4
e7
e2
e6
e6
e5
e5
e8
e3
43
Envisioning
envisioned causal system slice
e1
e9
e4
e7
e2
e6
e10
e5
e8
e3
e11
Agent has this in focus
44
Envisioned Causal System
Prediction
Explanation
e1
e9
e4
e7
e2
e6
e10
e5
e8
e3
e11
A sequence of envisioned contiguous causal systems
45
Correspondence with Reality
If the events and states in the ECS are
believed, the ECS is the current world
understanding Need an account of how graded
belief is increased or decreased as
predictions and explanations are verified or
falsified.
46
Goals and Planning
Causal Knowledge (??e1,x)p(e1,x) --gt
(??e2)q(e2,x) cause(e1,e2) or, p
causes q (??e1,x)p(e1,x) --gt
(??e2)q(e2,x) cause(e1,e2) or, p
enables q where enable(e1,e2) lt--gt
cause(e1,e2) Planning Axioms
(??a,e1,e2)goal(a,e2) cause(e1,e2) etc --gt
goal(a,e1) (??a,e1,e2)goal(a,e2)
enable(e1,e2) --gt goal(a,e1)
subgoal(a,e1,e2)
47
Goals and Planning
Goals can be ... competitive
adversarial auxilliary .....
48
Collective Goals
Groups can have goals All agents in group
mutually believe the group has the goal All
agents have the individual goal that the group
achieves its goal Must bottom out
in individual agents actions Organizations are
such collective plans made concrete an
agents role in an organization is the actions
the agent carries out as a subgoal in the
collective plan
49
Where Do Goals Come From?
A False Mystery Stipulate goal(a,
thrive(a)) All else is causal knowledge/beliefs
about what causes thriving
50
Goal Themes
group of agents
set of possible eventualities
goal-theme(s,t) lt--gt (? a,e) member(a,s)
member(e,t) etc ---gt goal(a,e) From group
membership, we can infer beliefs and goals
and thus behavior (defeasibly) e.g., hes a
puritan / hedonist / geek / ....
51
Summary
Ontologies are important for communicating the
contents and capabilities of Web sites and
Web resources Most of this information is
now in the form of natural language We need
ontologies that are capable of expressing
the full range of content in Web sites, forming
the basis of a deeper lexical semantics I
have presented first cuts at some of the most
basic ontologies needed services, events,
time, space, information, human psychology