Deep Lexical Semantics 6' Similarity and Like

1
Deep Lexical Semantics6. Similarity and Like
Jerry R. Hobbs Information Sciences
Institute University of Southern
California 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
1. Similarity in Discourse, including VP
Ellipsis 2. Data on the Word like 3.
Inferences Required
4
Similarity
Properties are similar, if they are or imply
properties whose predicates are the same,
and whose arguments are coreferential
(identical) or similar. Similar p(e1,x1,
..., z1), p(e2,x2, ..., z2)
Coref(x1,...,x2,...) OR Similar(x1,x2)
.... Coref(z1,...,z2,...) OR
Similar(z1,z2) Arguments are similar, if
their other inferentially independent properties
are similar. Similar x1,x2
Similar p1(...,x1,...), p2(...,x2,...)
.... Similar q1(...,x1,...),
q2(...,x2,...) Mapping is preserved as
recursion progresses. Inferential Independence
K, P /gt Q K, Q /gt P
5
A Discourse ExampleParallelism
A ladder weighs 100 lb with its center of gravity
10 ft from the foot, and a 150 lb man is 10 ft
from the top.
force(w1,L,d1,x1) w1 lb(w1,100) L
ladder(L) d1 Down(d1) x1
distance(x1,f, 20 ft) f foot(f,L)
gt end(f,L) L
force(w2,y,d2,x2) w2 lb(w2,150) y gt
Coref(y,...,L,...) d2 Down(d2) x2
distance(x2,t, 10 ft) t top(t,z)
gt end(t,z) z gt
Coref(z,...,L,...)
6
VP Ellipsis and Strict and Sloppy Readings
(Hobbs Kehler, 1997)
John revised his paper before the teacher
did. gt ... before the teacher revised Johns
paper. ... before the teacher revised
the teachers paper.
Strict
Sloppy
7
The Problem of VP EllipsisMissing Readings
John revised his paper before the teacher did,
and Bill did too. J revised as paper before T
revised bs paper, and B revised cs paper
before T revised ds paper. JJJJ
JJBJ JJBB JTJT
JTBT JJJB
Why not this?
8
Other Phenomena Involving Parallelism
John revised his paper, and so did Bill. John
revised his paper, and Bill too. John revised his
paper before Bill did. John revised his paper,
and Bill did the same. John revised his paper,
and Bill followed suit. Adverbs
too Conjunctions and, but Discourse
adjacency Comparatives and Coordination Repairs C
ontrast, Exemplification, Generalization Interpre
tation of Tables.
9
The Simple Case
John revised his paper before the teacher did.
before(e11,e21) e11 revise(e11,j,p1)
j John(j) gt person(j)
p1 paper(p1) Poss(x1,p1)
x1 he(x1),
Coref(x1,...,j,...) e21 revise(e21,t,p2)
t teacher(t) gt person(t)
p2 paper(p2)
Poss(x2,p2) x2
Coref(x2,...,x1,...)
he(x2), Coref(x2,...,t,...)
Strict JJ
Sloppy JT
10
The Missing Reading
and
before
before
J ...
T ...
B ...
T ...
John revised his paper before the teacher did,
and Bill did too.
before(e11,e21) e11 revise(e11,j,p1)
j John(j) gt person(j)
p1 paper(p1) Poss(x1,p1)
x1 he(x1),
Coref(x1,...,j,...) e21 revise(e21,t,p2)
t teacher(t) gt person(t)
p2 paper(p2)
Poss(x2,p2) x2
Coref(x2,...,x1,...)
he(x2), Coref(x2,...,t,...)
before(e31,e41) e31 revise(e31,b,p3)
b Bill(b) gt person(b)
p3 paper(p3) Poss(x3,p3)
x3 Coref(x3,...,x1,...)
he(x3),
Coref(x3,...,b,...) e41 revise(e41,t,p4)
t teacher(t) p4
paper(p4) Poss(x4,p4)
x4 Coref(x4,...,x2,...)
he(x4),
Coref(x4,...,x3,...)
he(x4), Coref(x4,...,t,...)

?

JJJB reading would require choices marked by
11
Previous Approaches
Sag Williams JJJJ, JTBT Gawron Peters
JJJJ, JTBT, JJBB (Kamp) JJJJ,
JTBT, JJBB, JTJT
or all 6 Prüst? JJJJ,
JTBT Asher? all 6 Core
Dalrymple, Shieber, Pereira all 6 DSP
special linking relation the right 5
12
Outline
1. Similarity in Discourse, including VP
Ellipsis 2. Data on the Word like 3.
Inferences Required
13
(No Transcript)
14
Investigation

• Close examination of 108 examples of like
• in a variety of genres
• Carson McCullers Ballad of the Sad Cafe (20)
• 2. San Jose Mercury News business section (18)
• 3. Science articles on AIDS (4)
• 4. Shakespeares sonnets (24)
• 5. Transcripts of decision-making meetings (20)
• Lyrics of country n western songs (22)
• Makes both depth of analysis and generality
  possible

15
Ancient History of Like
IE lik body, form
Germanic likjan to please
OE lic body, form, appearance
ONorse gelic similar
OE lician to please
OE -lic having form of
ModE like
ME liken
Preposition
-like
ModE -ly
ModE like
Verb
16
Recent History of Like
(Romaine Lange, 1991)

• Like as hedge, complementizer (?)
• --gt Like as hedge on quotation
• She goes, Mom wants to talk to you.
  Its like Ha ha, youre
• about to get in trouble.
• Remember when we would get in pretend
  fights in the hall?
• And people would get like, Oh, my
  God.
• --gt Like as marker of reported speech
• She said, What are you doing here? And
  Im like, Nothing much.
• --gt Like as marker of what is thought but not
  said
• And I saw her coming, and Im like
  Noooooooooo.

17
The Syntax of Like and Related Words
like X is like Y in that P(X) P(Y) unlike X
is unlike Y in that P(X) P(Y) liker X is
liker to Z than Y is, in that
Del(P(X),P(Z)) lt Del(P(Y),P(Z)) alike X and Y
are alike in that P(X) P(Y) the like X and Y
and the like Z s.t. P(X) P(Y)
P(Z) likewise P(X). Likewise Y. gt
P(Y) The interpretation problem like poses
Find P
18
Wrong P
19
Four Cases

• The property P that X and Y have in common is
• stated explicitly in a syntactically
  related place or
• in an adjacent place.
• The property P must be inferred.
• X s.t. P1(X) like Y s.t. P2(Y) where P1 --gt
  P, P2 --gt P
• Complementizer on a sentient verb.
• 4. Hedge.

20
Three Subcases of Case 1

• The property P that X and Y have in common is
• stated explicitly in a syntactically
  related place or
• in an adjacent place.
• 1A. X is like a Y s.t. P(Y) Q(Y)
• P is the common property
• 1B. P(X) is like Y where P(Y) is known
• 1C. X is like Y. P(X) where P(Y) is known

P is an explicit property of Y
P is an explicit property of X syntactically
P is an explicit property of X after coreference
21
1A. P is Explicit Property of Y
X is like a Y s.t. P(Y) Q(Y)
Indeed at the mere mention of the words her face
would slowly darken with shame, and she would
stand there craning her neck against the collar
of her shirt, or rubbing her swamp boots
together, for all the world like a great,
shamed, dumb-tongued child.
Shared properties
Not a shared property
P great(x) shamed(x) dumb-tongued(x) Q
child(x) Contradictory
22
1A. P is Explicit Property of Y
X is like a Y s.t. P(Y) Q(Y)
I dont know how to handle the dream you left
behind. Its like a lighted candle burning up my
mind.
Shared properties
Not a shared property
P burn-up(x,m) mind(m,i) Metaphor
burn-up(x,m) --gt painful(x,i) destroy(x,m) Q
lighted(x) candle(x) Contradictory
Select the right inferences
23
1A. P is Explicit Property of Y
X is like a Y s.t. P(Y) Q(Y)
I ... look upon my self and curse my
fate, Wishing me like to one more rich in hope,
...
Shared properties
Not a shared property
String of properties from corecursion in
similarity
P more(o,i,er) rich(er,z,h) hope(h,z,w) Q
one(x) --gt Not Mutually Identifiable
Contradictory
24
1A. P is Explicit Property of Y
X is like a Y s.t. P(Y) Q(Y) Special case where
X Y
There is something a bit jarring about a car like
Buicks massive Roadmaster.
Shared properties
P Buick(x) massive(x) Roadmaster(x) Q --
25
1A. P is Explicit Property of YA Method
X is like a Y s.t. P(Y) Q(Y)
X tends to be unelaborated. Y was always
indefinite, except in special case. Q properties
were contradictory. P the other properties of
Y.
26
1B. P is Explicit Property of X
X s.t. Q(X) P(X) is like Y where generally
P(Y) is known
AsnH100a, which like GlyH100b is found in the
epsilon conformation in the native molecule, ....
P found-in(x,c) epsilon(e) nn(e,c)
conformation(c) in(c,m) native(m)
molecule(m) Q AsnH100a(x)
Shared property
Not shared used to identify X
27
1B. P is Explicit Property of X
X s.t. Q(X) P(X) is like Y where generally
P(Y) is known
The bluest eyes in Texas are haunting me
tonight. Like the stars that fill the midnight
sky, Her memory fills my mind.
P(Y) is actually stated
Infer sky(y) --gt contain(y,s) stars(s)
mind(x) --gt contain(x,m) thoughts(m)
memory(m) --gt thought(m) P fill(m,x)
contain(m,x) Q mind(x) memory(m)
Shared
Not shared
28
1B. P is Explicit Property of X
X s.t. Q(X) P(X) is like Y where generally
P(Y) is known
But when from highmost pitch with weary car, Like
feeble age he the sun reeleth from the day, ...
Saying p(x) to convey p(f(x))
Metonymy feeble age gt person with feeble
age Inference reel(x) --gt fall(x) P
reel(e,x) OR fall(e,x) Q sun(x) from(e,d)
day(d)
Shared
Not shared
29
1B. P is Explicit Property of X
X s.t. Q(X) P(X) is like Y where generally
P(Y) is known
Therefore are feasts so solemn and so rare, Since
seldom coming in that long year set, Like stones
of worth they thinly placed are, Or captain
jewels in the carcanet.
Metaphor year(t) --gt linear(t)
carcanet(c) --gt necklace(c) --gt linear(c) P
placed(e,x,t) thin(e) linear(t) Q(X)
feast(x) year(t) Q(Y) jewel(y) carcanet(c)
Shared
Not shared, but they line up
30
1B. Sloppy Readings
Sears, like other major department stores,
released a disappointing sales report for Sears
for last month.... Compare Sears released
a disappointing sales report for last
month, and other major department
stores did too. Like other Fab-antigen
complexes, there is a close complementarity
between the buried surfaces on the antigen and
Fab of this complex. Ah, if thou issueless
shalt hap to die, The world will wail thee like
a makeless wife mourns her mate, ...
widowed
31
1B. Likewise
Strict reading
When Miss Amelia had eaten her supper she wiped
her plate carefully with a slice of light
bread, and then poured her own clear, sweet syrup
over the bread. The hunchback did likewise --
except that he was more finicky and asked for a
new plate.
P when(e1,e2) eat(e1,x,s) Poss(x,s)
supper(x) and(e2,e3,e4) wipe(e3,x,p)
careful(e3) with(e3,b) pour(e4,x,r,b)
syrup(r) Q Miss-Amelia(x)
Shared
Not shared,
32
1B. Unlike
unlike X s.t. P(x) is unlike Y s.t. P(Y)
We always try to underpromise and overdeliver,
unlike others who do the opposite, Sim said.
Need logical form from morphology
under/over-Verb Verb(e,x,...)
under/over(e,z) Infer under(e,z) lt--gt
over(e,z) opposite(e1,e2)
p(e2,x,...) --gt imply(e1,e3) p(e3,x,...)
33
1B. P is Explicit Property of XA Boring Special
Case
X s.t. Q(X) P(X) is like Y where generally
P(Y) is known Y is subset of X Q is empty
Companies like Kraft, Procter Gamble and
Dowbrands have greatly expanded the number of
products that carry 800 numbers.
34
1B. P is Explicit Property of XAnother Boring
Special Case
X s.t. Q(X) P(X) is like Y where generally
P(Y) is known X is latest element in sequence Y
I came crawling home last night like many nights
before.
Shared property
Not a shared property
35
1B. P is Explicit Property of XA Method
X s.t. Q(X) P(X) is like Y where generally
P(Y) is known
Y tends to be unelaborated. P the explicit
properties of X that are known to be true
of Y Q the other properties of X.
36
1C. P(X) is Nearby
X is like Y. P(X) where generally P(Y) is known
I just wanted to see if you felt like me -- sorry
that we said goodbye.
Metonymy me gt I feel P Inference
sorry(e,x) --gt feeling(e,x)
37
1C. P(X) is Nearby
X is like Y. P(X) where generally P(Y) is
known Often P(X) is quite long -- thats why
the separate phrase or clause.
How like a winter hath my absence been From thee,
the pleasure of the fleeting year! What freezings
have I felt, what dark days seen! What old
Decembers bareness everywhere!
Metaphor freezing, dark, bare --gt induces
sadness P(X) during(e1e2e3,x) feel(e1,i,f)
freezing(f,i) see(e2,i,d)
dark(d) day(d) bare(e3,z)
everywhere(z) Q(X) absence(x,u) vs. winter(y)
38
P(X) is Nearby Alike
X is like Y. P(X) where generally P(Y) is
known alike X and Y are alike in that P(X)
P(Y)
Except for Reverend Willin, they were all alike
in many ways as has been said -- all having
taken pleasure from something or other, all
having wept and suffered in some way, most of
them tractable unless exasperated.
39
1C. P(X) is NearbyA Method
X is like Y. P(X) where generally P(Y) is known
X and Y tend to be unelaborated. P(X) the
properties expressed in an adjacent
phrase or clause known to be true of Y. Q(X)
the other properties of X.
40
2. P(X) is Inferred.
X s.t. P1(X) is like Y s.t. P2(Y), where P1 --gt P
and P2 --gt P Often P2 --gt P P3, and the text
implicates P3(X)
I see you drivin by just like a Phantom Jet
Metonymy Phantom Jet gt fly(Phantom Jet) P1
drive(u) --gt move(u) P2 fly(j) --gt move(e,j)
fast(e) P move(e,u) P3 fast(e)
41
2. P(X) is Inferred.
X s.t. P1(X) is like Y s.t. P2(Y), where P1 --gt P
and P2 --gt P Often P2 --gt P P3, and the text
implicates P3(X)
He might show up at 1130, something like that.
From show up at ___ syntactic context
Reference resolution that 1130 P1
something(x) --gt time(x) P2 1130(y) --gt
1130/-10min --gt time(y) P time(y) P3
1130/-10min
Half-order of magnitude fuzz around time
specified to half-hour precision
42
2. P(X) is Inferred Manner
X s.t. P1(X) is like Y s.t. P2(Y), where P1 --gt P
and P2 --gt P Often P2 --gt P P3, and the text
implicates P3(X)
The consumer is just making out like a bandit.
Metonymy bandit gt make-out(bandit) P1
consumer(x) --gt purchase(x,z) --gt
acquire(x,z) P2 bandit(y) --gt steal(y,z) --gt
acquire(y,z) low-cost(z,y) P acquire(x,z) P3
low-cost(z,x)
43
2. P(X) is Inferred Appearance
X s.t. P1(X) is like Y s.t. P2(Y), where P1 --gt P
and P2 --gt P Often P2 --gt P P3, and the text
implicates P3(X)
Wishing me like to one more rich in
hope, Featured like him, like him with friends
possessed,
Metonymy him gt features(f2,him) P1
features(f1,i) P2 features(f2,i) --gt
attractive(f2)
from coherence w rich in hope P
features(f1,i) P3 attractive(f1)
44
2. P(X) is Inferred Appearance
X s.t. P1(X) is like Y s.t. P2(Y), where P1 --gt P
and P2 --gt P Often P2 --gt P P3, and the text
implicates P3(X)
You still look like the day when I first met you.
Metonymy day gt what you looked like that
day P1 look(x,t2) P2 look(x,t1)
first-met(e,t1) --gt young(x,t1)
--gt
attractive(x,t1) P look(x,t2) P3
attractive(x,t2)
45
2. P(X) is Inferred the like
X s.t. P1(X) is like Y s.t. P2(Y), where P1 --gt P
and P2 --gt P the like Y1 and Y2 and the like
Z s.t. P(X) P(Y) P(Z)
It also prepared food for airlines and managed
food service for company cafeterias and the like.
P1 -- P2 cafeteria(y) --gt food-service(y)
institutional(y) P food-service(y)
institutional(y) P3 --
46
3. Complementizer on Sentient Verb
Contentful subjects Ichan seemed like the
most logical seller. It sounds like
something we could continue over lunch. I
feel like I can be flexible. Expletive it
subjects It seems like you might want to
block out some time... Looks like were two
of a kind. It feels like Ive gone out of
style.
Historical/Logical Analysis P
verb(z) Metonymy complement gt
verb(complement) P3 z complement
47
3. Complementizer on Sentient Verb
Contentful subjects Ichan seemed like the
most logical seller. It sounds like
something we could continue over lunch. I
feel like I can be flexible. Expletive it
subjects It seems like you might want to
block out some time... Looks like were two
of a kind. It feels like Ive gone out of
style.
Method In syntax, subcategorize verb,
e.g., feel subcategorized for
expletive it subject like
S complement
48
4. Hedges
You mean like if he gets here late. Okay, so
thats two demos, so thats like half an hour,
forty minutes. Therell be just like ten
minutes, or fifteen minutes of general
overview and then two demos.
Analysis Like as a hedge can be viewed as
ellipsis of something like, hence
derives from Case 2. P1 something(x)
P2 P2(Y) --gt P(Y) where P(Y) fuzzes out some
distinctions. Method In syntax, treat
like as an adverb as free in its placement
as only.
49
Some Numbers

• Novel BizNews
  Science Shakespr Meeting Lyrics Total
• 1. P 14 14
  4 20 1
  10 63
• 1A. P(Y) 4 1
  -- 1 --
  2 8
• 1B. P(X) 3 12
  4 17 --
  5 41
• 1C. like. P. 7 1
  -- 2 1
  3 14
• 2. Inference 6 3
  -- 4 3
  9 25
• 3. SentComp -- 1
  -- -- 8
  3 12
• 4. Hedge -- --
  -- -- 8
  -- 8
• Total 20 18
  4 24 20
  22 108

50
Outline
1. Similarity in Discourse, including VP
Ellipsis 2. Data on the Word like 3.
Inferences Required
51
Inferences Required for the Data
What knowledge was required to infer P or to
resolve metaphors and metonymy, or was a known
P(Y)?
What WordNet Contains
Definitions, Classifications, Attributes -
43 Other Axioms in Theories -
55 Specific Facts -
1 Total -
99
Theories Scalars - 1 Time -
3 Geography - 4 Material - 4 Weather
- 2 Plants - 2 Animals -
3 Human Body - 5
Mind Emotions - 8 Human Impact of
Things - 8 Human Action -
3 Human Life Course - 2 Social
Interaction - 3 Society -
3 Economics -
4
52
Theory of Mind and Emotions
(Work w Andrew Gordon, Barnden)
The mind is a container. A person uses
appearances to identify objects. A person feels
a general condition / emotional state. A person
grieves the loss of something good. A person
stores past events in memory. Memory of the loss
of something good causes sadness. Selection
involves judgment. A person who is crazy uses bad
judgment.
53
Theory of Human Impact of Objects and Events
Folk theory of physical sensations
Warmth is comfortable. Burning is painful. The
day is correlated with happiness. Appearance can
be attractive. Youth is attractive. Bright things
are attractive. Jewels are attractive. Unintended
spots are unattractive.
Folk theory of whats attractive
This may be a mishmash of random facts, but it is
just the sort of thing an OpenMind project could
collect (putting millions of Netizens to work).
54
Summary
Defined a concept of similarity in terms of
shared properties. It applies to a broad range
of linguistic phenomena, like discourse
structure, VP ellipsis, and lexical semantics
(like). For like, shared property is
usually explicit in nearby
discourse. Recent innovative meanings of like
arise naturally in this account. It is
plausible that the world knowledge / inferences
required for shared properties to be
discovered can be encoded.