Deep Lexical Semantics or Commonsense Knowledge and Lexical Semantics I' Introduction

1
Deep Lexical SemanticsorCommonsense Knowledge
and Lexical SemanticsI. Introduction
Jerry R. Hobbs Information Sciences
Institute University of Southern
California Marina del Rey, CA
2
Overview of This Module

• 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

How do we encode theories of these?
How do we define words like these?
3
This Lecture

• Motivation
• Core theories and the lexicon
• Composite entities
• Figure-ground
• Scales

4
A Problem in Discourse UnderstandingPronoun
Resolution
The plain was reduced by erosion _ to its present
level.
1. What is eroding? 2. What does its refer
to?
5
Pronoun Resolution
The plain was reduced by erosion to its
present level. LF reduce'(e1,e2,p,l)
plain(p) erode'(e2,x) present(e3)
level'(e3,l,y) KB decrease(p,l,s)
vertical(s) etc1(p,l,s) --gt reduce'(e,z,p,l)
A decrease on a vertical scale is a
reduction. landform(p) flat(p)
etc2(p) --gt plain(p) A flat
landform is a plain. at'(e,y,l)
on(l,s) vertical(s) flat(y) etc3(e,y,l,s)
---gt level'(e,l,y) If
a flat thing y is at l on a vertical scale, then
l is the level of y. decrease'(x,l,s)
landform(x) altitude(s) etc4(x,l,s)
---gt erode'(e,x) A landform
decreasing in altitude is erosion.
vertical(s) etc5(s) --gt altitude(s)
One kind of vertical scale is altitude.
6
The Pronoun Resolved
The plain was reduced by erosion to its
present level. LF reduce'(e1, e2,p,l)
plain(p) erode'(e2,x) present(e3)
level'(e3,l,y) KB decrease(p,l,s)
vertical(s) etc1(p,l,s) --gt reduce'(e,z,p,l)
landform(p) flat(p) etc2(p) --gt
plain(p) at'(e,y,l)
on(l,s) vertical(s) flat(y) etc3(e,y,l,s)
---gt level'(e,l,y)
decrease'(x,l,s) landform(x)
altitude(s) etc4(x,l,s) ---gt
erode'(e,x)
vertical(s) etc5(s) --gt altitude(s)

yp
xp
7
But.
Where does all this knowledge come from?
The plain was reduced by erosion to its
present level. LF reduce'(e1, e2, p,l)
plain(p) erode'(e2,x) present(e3)
level'(e3,l,y) KB decrease(p,l,s)
vertical(s) etc1(p,l,s) --gt reduce'(e,z,p,l)
A decrease on a vertical scale is a
reduction. landform(p) flat(p)
etc2(p) --gt plain(p) A flat
landform is a plain. at'(e,y,l)
on(l,s) vertical(s) flat(y) etc3(e,y,l,s)
---gt level'(e,l,y) If
a flat thing y is at l on a vertical scale, then
l is the level of y. decrease'(x,l,s)
landform(x) altitude(s) etc4(x,l,s)
---gt erode'(e,x) A landform
decreasing in altitude is erosion.
vertical(s) etc5(s) --gt altitude(s)
One kind of vertical scale is altitude.
8
Is it lexical knowledge or world knowledge?
Does it matter?
KB decrease(p,l,s) vertical(s)
etc1(p,l,s) --gt reduce'(e,z,p,l) A
decrease on a vertical scale is a reduction.
landform(p) flat(p) etc2(p) --gt
plain(p) A flat landform is a
plain. at'(e,y,l) on(l,s)
vertical(s) flat(y) etc3(e,y,l,s)
---gt level'(e,l,y) If a flat
thing y is at l on a vertical scale, then l is
the level of y. decrease'(x,l,s)
landform(x) altitude(s) etc4(x,l,s)
---gt erode'(e,x) A landform
decreasing in altitude is erosion.
vertical(s) etc5(s) --gt altitude(s)
One kind of vertical scale is altitude.
9
A Warm-Up Exercise

• 1. Drive the demons out of her and teach her to
  stay away from my husband!!
• 2. Shortly before nine I drove my jalopy to the
  street facing the Lake and parked the car in
  shadows.
• 3. He drove carefully in the direction of the
  brief tour they had taken earlier.
• 4. Her scream split up the silence of the car,
  accompanied by the rattling of the freight, and
  then Cappy came off the floor, his legs driving
  him hard.
• 5. With an untrained local labor pool, many
  experts believe, that policy could drive
  businesses from the city.
• 6. Treasury Undersecretary David Mulford
  defended the Treasurys efforts this fall to
  drive down the value of the dollar.
• 7. Even today range riders will come upon
  mummified bodies of men who attempted nothing
  more difficult than a twenty-mile hike and slowly
  lost direction, were tortured by the heat, driven
  mad by the constant and unfulfilled promise of
  the landscape, and who finally died.
• 8. Cows were kept in backyard barns, and boys
  were hired to drive them to and from the pasture
  on the edge of town.
• 9. He had to drive the hammer really hard to
  get the nail into that plank!
• 10. She learned to drive a bulldozer from her
  uncle, who was a road maker.
• 11. I used to drive a taxi (for work) before I
  went to night school.
• 12. BewareRalph drives a hard bargain you
  will probably lose all your money.

10
My Answers
Sense 1 1, 4, 5, 6, 7, 8, 9 Sense 2 2, 3, 10,
11, maybe 8 (11 is habitual(drive2) ) Sense
3 12
x drives y
x forcefully causes y to move
x causes y to move in particular direction
x forcefully causes there to be y hard bargain
move (at) interpreted in many ways
physical 2, 3, 10, 11 4, 5?, 8, 9
psychological (in mind) 1, 7 location of
value on scale 6 existence in social space
12, 5?
11
This Lecture

• Motivation
• Core theories and the lexicon
• Composite entities
• Figure-ground
• Scales

12
Core Theories and the Lexicon
LEXICON
drive
Economy
Social Groups
Cognition
Physical Space
cause
force
move
CORE THEORIES
13
Core Theories and the Lexicon
Define (Characterize) words in terms supplied by
the core theories. range(x,y,z) lt--gt
scale(s) subscale(s1,s) bottom(y,s1)
top(z,s1) in(u1,x) at(u1,y)
in(u2,x) at(u2,z) (??u ? x)(? v ?
s1) at(u,v)
s
s1
y
z
v
x u1 . . . . . . . u . . . . . u2
Axiomatize core theories with richly explicated
core predicates Core Theory of Scales scale,
lt, subscale, top, bottom, at
14
Generative Semantics Revisited
Lexical Decomposition kill(x,y) --gt
cause(x, become(not(alive(y))))
This reasoning is needed in My roommate
killed all my plants. He didnt water them
once while I was gone.
15
Levels of Processing
S --gt NP VP
Syntax and Compositional Semantics
Syn(w1,x,N,-,-,-,-) Syn(w2,e,V,x,N,-,-,-)
--gt Syn(w1 w2,e,V,-,-,-,-) Lexical
Axioms kill(e,x,y) living-thing(y) --gt
Syn(kill,e,V,x,N,y,N) Lexical
Decomposition cause(e,x,become(not(alive(
y)))) --gt kill(e,x,y) Core Theories
water --gt nourish enable(nourish,alive)
pred-arg structure (incl word sense)
selectional constraint
spelling or phonology
category
subcategorization
16
A Sentence Interpreted
Syntax
Syn(My roommate killed my plant.,e,-,-)
He didnt water them.
Syn(killed my plant.,e,x,-)
Syn(My roommate,x,-,-)
Syn(my plant.,y,-,-)
Syn(killed,e,x,y)
roommate(x,i)
Lexical Axioms
plant(y)
kill(e,x,y)
Lexical Decomposition
cause(x,become(not(alive(y))))
cause(not(nourish(x,y)),not(alive(y)))
Core Theories
cause(not(water(x,y)),not(nourish(x,y)))
17
On the Nature of Primitives
The Old View Primitives are undefined
predicates that everything else is
defined in terms of. The New View
So-called Primitives are predicates that are
central to the core theories and common
in lexical decompositions. Almost
nothing is defined, but axioms constrain
the interpretations of predicates. In that
sense, almost everything is a primitive
whose meaning is tightly constrained.
18
Some of the Core Theories
Structure of Information
Goal-Directed Behavior
Material
Cognition
Force Dynamics
Time
Space
Change
Scales
Figure-Ground
Causality
Stuff
Composite Entities
Sets
19
This Lecture

• Motivation
• Core theories and the lexicon
• Composite entities
• Figure-ground
• Scales

20
Composite Entities
Composite Entity Something made of other
things 1. A set of components 2.
Properties of the components 3. Relations
between components (the structure) 4.
Properties of the whole 5. Relations
between the whole and entities in the
environment (incl. function). Structure-
Function Articulations How are functional
relations implemented in the structure
of the elements.
.
.
.
.
.
.
.
.
.
21
A Composite Entity
Book 1. pages, cover, binding, content
2. paper(pages), cardboard(cover), ... 3.
enclose(cover,pages), convey(pages,content), ...
4. physobj, 8in 5.
write(person,book),
function(book,read(person,book)), ...
Where to put a book At home physobj,
8in gt on shelf In library content
fiction, writer McMurty gt ... In
bookstore ..., paper(cover) gt ...
22
Composite Entities
For every composite entity A coarse
granularity, where components and
their properties and relations are
ignored. A fine granularity, where
components reasoned about.
23
Complex Events
A complex event is a composite entity, where
among the components are the participants
and the subevents among the relations
are agent-of, etc. relations between
subevents and participants among the
relations are causal relations among
the subevents.
24
Characterizing Some Words
system(x) lt--gt composite-entity(x)
assemble(x,y) lt--gt (E s1,s2)composite-entity
(y) cause(x,change(s1,s2))
not(subset(relations-of(y),s1))
subset(relations-of(y),s2)
complicated(x) The more components, the
more complicated. The more relations, the
more complicated. The more complicated the
components, the more complicated the
whole. The more complicated the relations as
composite eventualities, the more
complicated the whole.
25
This Lecture

• Motivation
• Core theories and the lexicon
• Composite entities
• Figure-ground
• Scales

26
The Figure-Ground Relation
The predicate at relates an external entity
to an element in a composite entity.
at(x,y,s)
.
.
X
S
.
y
.
.
.
Requirement All the elements of S must be
similar in the sense of sharing some property
that allows a consistent interpretation of
at. Pages of a book Okay John is at
p. 45. Pages, cover, content of book No.
27
Aside The Nature of Radial Categories
word1
word3
word4
word2
Incremental changes in axioms q1 (x) q2(x) --gt
p(x) gt q1 (x) q3(x) --gt p(x)
Characterized by a set of axioms q1 (x) q2(x)
--gt p(x)

e.g., radial categories of drive
28
Radial Categories for at
Person _at_ Org
at least, at most
Entity _at_ Predication
PhysObj _at_ Time
PhysObj _at_ PhysLoc
Entity _at_ Scale
Event _at_ Time
Event _at_ PhysLoc
Directed-Action-toward
Event _at_ Event
TextEvent _at_ TextLoc
Causality
29
Associated Inferences
PhysObj _at_ PhysLoc at part --gt in whole at
X --gt not at Y at enables some event
trapped at the airport not at office, unable
to do work Person _at_ Org substitute
structure of organization for spatial structure
analyst at Schwab enable belief in
statement Event _at_ PhysLoc p(X) at Y --gt X
at Y the men met at the airport
Entity _at_ Predication X at predicate p in
system of predicates gt p(X) my heart
and mind are at war warring(heart,mind)
Entity _at_ Scale X is at Y on S
predicate-for(S)(x,y) the plane is at
an altitude of 36,000 feet altitude(plane,
36000ft) Event _at_ Time meeting at 3
oclock Event _at_ Event he came in at
daylight Event1 _at_ time-of(Event2)
Causality he knew at a glance
cause(glance,know)
30
The Ontological Ascent
cause(change)
Actions
change(at,at)
Events
Coarsening of Granularity
at(ent,ent,sys)
Properties
system of entities
Sets, Processes
Gruber, Jackendoff, Lakoff, Herskovits, Talmy,
Croft
entity
cause(x,change(at(y,z),at(y,w)))
31
The Ontological Ascent
cause(change)
In a stunning reversal for one of
Silicon Valley's fastest-growing companies, Media
Vision Technology Inc. said Thursday it will
report a sharp decline in sales and "a
substantial loss" in the quarter ending March 31
--- a jolt that cut its stock price in half.
change(at,at)
at(ent,ent,sys)
system of entities
change( at(x,y,Soc), at(x,a,Soc))
Coarsening of Granularity
entity
32
The Ontological Ascent
cause(change)
In a stunning reversal for one of
Silicon Valley's fastest-growing companies, Media
Vision Technology Inc. said Thursday it will
report a sharp decline in sales and "a
substantial loss" in the quarter ending March 31
--- a jolt that cut its stock price in half.
change(at,at)
at(ent,ent,sys)
system of entities
system of entities
change( at(sales,x,Q), at(sales,y,Q))
Coarsening of Granularity
entity
33
The Ontological Ascent
cause(change)
In a stunning reversal for one of
Silicon Valley's fastest-growing companies, Media
Vision Technology Inc. said Thursday it will
report a sharp decline in sales and "a
substantial loss" in the quarter ending March 31
--- a jolt that cut its stock price in half.
change(at,at)
at(ent,ent,sys)
system of entities
at( decline, x, Qchange)
Coarsening of Granularity
entity
34
The Ontological Ascent
cause(change)
In a stunning reversal for one of
Silicon Valley's fastest-growing companies, Media
Vision Technology Inc. said Thursday it will
report a sharp decline in sales and "a
substantial loss" in the quarter ending March 31
--- a jolt that cut its stock price in half.
change(at,at)
at(ent,ent,sys)
at(ent,ent,sys)
system of entities
cause(itMVT, change( at(declineloss-in,
x, CommMinds),
at(declineloss-in, y, CommMinds)))
Coarsening of Granularity
entity
35
A FrameNet Frame
(Fillmore)
cause(e,x, change(at(y,w), at(z,w))) --gt
replace(e, x, y, z) --gt
Syn(replace, e,V, x,N, y,N, z,P.with) --gt
substitute(e, x, z, y) --gt
Syn(substitute, e,V, x,N, z,N, y,P.for)
36
FrameNet Frame Blending
Transfer cause(x,change(at(y,x),at(y,z)))
Speech cognize(x,y) --gt
at(y,x) Assailing at(y,z) gt at(y,z)
injure(y,z) Economic Transfer own(x,y)
--gt at(y,x)
Reciprocity cause(x,change(at(y1,x),at(y1,z
))) cause(z,change(at(y2,z),at(y2,x))) Con
versation cognize(z,y) --gt
at(y,z) Fighting at(y,w) gt at(y,w)
injure(y,w) Exchange own(z,y) --gt at(y,z)
37
This Lecture

• Motivation
• Core theories and the lexicon
• Composite entities
• Figure-ground
• Scales

38
Scales
Set of elements with a partial ordering lt Can
define subscale, total ordering, dense, top,
bottom, reverse, Allen's relations among
subscales, adjacency, connected, interval
(connected subscale) Examples distance, time,
happiness, damage, preference, ... Scales
may or may not be totally ordered
39
Levels of Structure on Scales
not okay
okay
0
--

qualitative amounts
Md
Lo
Hi
orders of magnitude
half orders of magnitude
integers
real numbers
40
Some Multiple Choice Questions
1. About how many children are there in the
average family? a) 1
c) 10 e) 100 2. About how
many children are there in the average
classroom? a) 1 c) 10
e) 100
41
Some Multiple Choice Questions
1. About how many children are there in the
average family? a) 1 b) 3 c)
10 d) 30 e) 100 2. About how many
children are there in the average
classroom? a) 1 b) 3 c) 10
d) 30 e) 100 3. About how many oranges
are there in a basket full of oranges?
a) 1 b) 3 c) 10 d) 30 e)
100
Often the most appropriate estimate of a
quantity is to a half order of magnitude.
42
"about", "approximately", "nearly"Implicit
Precision
There were 920 people at the meeting. Are the
following true or false? a) There were about
1000 people at the meeting. -- TRUE b)
There were about 900 people at the meeting.
-- TRUE c) There were about 980 people at
the meeting. -- FALSE a) Implicit
precision 200, 250, or 500. b) Implicit
precision 100 c) Implicit precision 10
43
What about Means
X is about N N n g, for some integer n
and some HOM g. g is the implicit
precision. N - .5g lt X lt N .5g The
HOM between 1 and 10 is usually 5 or 2. (3
lacks good divisibility properties.) The HOM
between 10 and 100 is often 25, because it
is close to 101.5 and has good divisibility
properties.
44
The Examples Explained
b) about 900 n 9, g 100, 850 lt X lt
950 c) about 980 n 98, g 10, 975 lt X lt
985 a) about 1000 n 2, g
500, 750 lt X lt 1250 n 4, g 250,
875 lt X lt 1125 n 5, g 200, 900 lt
X lt 1100 n 10, g 100, 950 lt X lt 1050
45
Natural HOMs
Linear extent Examples 6 feet
person, door, chair, table,
desk
can be moved by one person,
can accommodate one
person 2 feet TV
set, dog, basket, watermelon, sack
can be held in two
arms 8 inches book,
football, cantelope
can be held in one hand,
manipulated
with difficulty in one hand 3
inches pen, mouse,
hamburger,orange, cup
can be held with the fingers
1 inch french fry,
eraser, peppermint candy
can be bitten, can be
manipulated
easily with two fingers and thumb
1/4 inch MM, thumb tack,
diamond
handled with care between two fingers
46
Natural HOMs
Linear extent Examples 6 feet
person, door, chair, table,
desk
can be moved by one person,
can accommodate
one person 18 feet
office, room
one person can move around
can
accommodate several people 20 yards
house, restaurant, small yard,
class 60 yards
commercial building, large yard 200 yards
small factory, field 600
yards large factory, large
bridge, dam 1 mile
town, airport 3 miles
small city 10 miles
large city, small county 30 miles
large county 100 miles
small state 300 miles
large state, small nation 1000
miles typical large European
nation 3000 miles the
United States, China
47
where
How is this word used in a corpus of news,
novels, poetry, etc.? farms where corn is
grown Where corn is grown, farmers prosper.
The Midwest is where corn is grown.
Where is corn grown? Examined 74 examples.
Figure at Ground
PhysObj at Phys Loc
7 Where are you? Prop of
PhysObj at Prop of PhysLoc 61
Where corn is grown, farmers prosper
Abstraction at Abstraction
6 I dont know where to put these
examples.
48
Relative Size of Figure and Ground
Ground is same HOM as Figure 36
Right here beside me is where you belong.
Ground is one HOM larger than Figure 13
the counter where slabs of meat were kept
The front room was where Marvin stayed. Ground
is two HOMs larger than Figure 5 the
houses where the workers live In 54 of 68
cases, HOM(Figure) ? HOM(Ground) ?
HOM(Figure) 2
49
Relative Sizes of Figure and Ground
11 cases where Ground is more than 2 HOMs
larger than Figure 10 cases long-term
activities of mobile entities the laws
of New York, where the business is based 1
case a treasure chest where a jewel is
hidden 3 cases where Figure is larger than
Ground poetic My heart is where you belong
50
"Several"
Does "several" mean 1 HOM? several Ns Ns ?
S If S ? 10, then Ns ? 2-5. If S ? 30-100,
then Ns ? 3-8.
2-5 13 of 25 Several women walked
into the cafe. 3-8 11 of 25 About
80,000 people lost their long-distance service
and several communities lost their 911
emergency phone. 3-12 1 of 25 ...
criminal investigation of GE and several of its
employees.
51
Hi, Md, and Lo Regions of a Scale
Lo
Hi
Md
Provides a useful, coarse-grained structure for
scales top(Hi(s)) top(s), bottom(Lo(s))
bottom(s) x in ? Lo(s) y ? Hi(s) --gt x lt
y Absolute form of adjectives tall
Hi(Height-Scale) short Lo(Height-Scale) Ca
n be iterated very
52
Qualitative Amountsand Functionality
Qualitative amounts should be related to
functionality x ? Hi(s) lt--gt
(??a,e)goal(a,e) enable(x ? Hi(s),e) x
? Lo(s) lt--gt (??a,e)goal(a,e) prevent(x ?
Lo(s),e)
53
Qualitative Amounts and Distributions
Qualitative amounts should be related to
distributions A naive theory of
distributions -ile(x,s,t) n lt--gt
x at y on t x1 x1
at y1 on t y1lty on t / s n
-ile(x,s,t) gt .8 --gt x ? Hi(t) -ile(x,s,t)
lt .6 --gt x ? Hi(t) But -- "Most towns in
America are small." -- HOMs?
1K
3K
10K
30K
100K
300K
1M
3M
10M
small
large
54
Composite Scales
s2
s is composite of s1 and s2 s ltx,ygt
x in s1, y in s2 x1 lt x2 y1 lt y2 ---gt
ltx1,y1gt lt ltx2,y2gt but normally not
lt--- Allows analysis of complex scales,
such as area, happiness, damage, ... Damage
degree of loss of functionality
X cost of repair
s1
55
Whats Next?

• 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

How do we encode theories of these?
How do we define words like these?