Differential Possessor Expression in English: Reevaluating Animacy and Topicality Effects Catherine

1
Differential Possessor Expression in English
Re-evaluating Animacy and Topicality
EffectsCatherine O'Connor Boston University
Arto Anttila NYU Vivienne Fong NYU Joan Maling
Brandeis University Annual Meeting of the
Linguistic Society of AmericaJanuary 9 - 11,
2004Boston, Massachusetts
2
The question
What are the factors that drive the English
alternation between the "Saxon genitive" and the
"Of genitive"?
X'S Spec
The man's widow
The widow of the man
OF-X Comp
3
Hypothesis 1
4
This is a statistical tendency, at best
Hemispheres Magazine, 2001
5
Hypothesis 2
6
This is also a statistical tendency
7
Hypothesis 3
8
An analytical problem
These three hypotheses are seriously confounded

His advocacy of betting on the ponies
Pronouns are light
Humans are often topical
Topics are often expressed as pronouns
9
Another analytical problem
Which examples can legitimately be expected to
alternate between Of-X and X's?
There are many, many such distractors
10
Our plan of inquiry
1. Secure a large number of OF-X and X'S
tokens in the Brown Corpus.
2. Exclude tokens of non-reversible types.
3. Code remaining tokens for weight, animacy and
discourse status.
4. Control for confounds where possible, and try
to model the statistical findings within an OT
grammar, following work by Aissen and others.
11
1. Cleaning the sample
12
First exclude non-nominals.
Using F.Karlsson's part-of-speech tagged version
of the Brown corpus (1995), we excluded all
irrelevant Of-NP and NP's tokens. A few
examples
Verbal OF-X He thought of her.
Adjectival OF-X bald and afraid of women.
Contraction X'S Kate's all right.
13
"All NP" sample, after removal of non-nominal
examples
14
Second exclude all tokens of non-reversible
constructions A few examples
Partitives half of his
stirrup guard
Measure and a drop of
liquor container phrases two saucers
of water
Classifier phrases a grove of trees
a flight
of wooden steps
Configuration and strips of skin
constitutive phrases a...castle of pine
boughs
15
Second exclude all tokens of non-reversible
constructions
'Sort' phrases the crassest kind of
materialism
Headless OF-X that of a frustrated gnome
Nominal dog-eared men's
magazines compounds
and many more
a man of brooding suspicions the concept of the
white-suited big-daddy colonel the notion of
philosophy as Queen Bee . . .
16
Partially clean sample after removal of 'strict'
non-reversibles
17
Third exclude tokens where reversal
substantially alters meaning--'soft'
non-reversibles

• Idioms, fixed phrases, and titles

bachelor of science science's bachelor
Satan's L'il Lamb the L'il Lamb of
Satan
(b) Deverbal nominals with argument constraints
(see handout for more examples)
18
Cleaner sample after removal of 'soft'
non-reversibles
19
2. Coding the sample
20
For each token,
we coded the head,
and the modifier.
Each was coded for animacy, definiteness, NP
form, and weight.
21
CODING for ANIMACY
Human(oid)sAnimals
ANIMATE
Human organizations
ORG
Concrete objectsAbstract entitiesLocations
Temporal entities
INANIMATE
22
CODING for WEIGHT
Arnold et al., Wasow, and J. Hawkins assert that
the orthographic word is a reasonable measure
of weight for most purposes.
It is also easily automated.
Each head and modifier were coded for weight in
words, from 1 through 20.
23
How to code for Discourse Status?
Even simple codes such as 'New', 'Inferrable',
and 'Old' are quite time-consuming, although they
are clearly desirable.
With thousands of tokens, we chose instead to
exploit certain robust relationships between NP
form and discourse status / accessibility.
Relying on previous research of Prince, Gundel et
al., Ariel, i.a., we coded modifiers and heads
for NP form and for morphosyntactic definiteness.
24
Coding for NP Form and Definiteness
Most accessible, most topical, discourse-old...
Pronoun
Proper Noun
Common Noun (definite)
Least accessible, least topical, discourse-new...
Common Noun(indefinite)
25
3. Controlling weights
26
After we coded our clean sample for weight, we
noticed that 99 of our X'S examples had
possessive modifiers that were 1, 2, or 3 words
in weight.
his only attack on the Republicansthe taxpayers'
pocketsSpeaker Sam Rayburn's forces
We controlled for weight effects by limiting OF-X
tokens to those of 1, 2, or 3 words in weight.
the invasion of Cubathe rapid growth of juvenile
delinquencythe 9th precinct of the 23rd ward
27
Cleanest sample, after removal of modifiers
greater than 3 words in weight
28
3. Generalizations
29
We decided to convert the raw numbers of X'S and
Of-X tokens into ratios.
For example, of 4177 animate tokens X'S
Of-X
3909
268
30
Let's compare the inanimate tokens
1359 inanimate tokens X'S
Of-X
357
1002
31
Ratio of X'S to OF-X by Animacy categoryin
Cleanest sample (n6034)
15 1
log scale
1.3 1
1 3
(N3937)
(N498)
(N1359)
32
Ratio of X'S to OF-X by NP form type in
'Cleanest' sample (n6034)
2971
1.85 1
1 2.3
1 7.7
(N3577)
(N971)
(N947)
(N539)
33
Both animacy and discourse status seem to have a
large effect. How about weight? Do we find
effects of similar magnitude when we examine our
possessive modifiers by our three weight
values--1, 2 and 3 words?
Yes here, the results span two orders of
magnitude.
34
Ratio of X'S to OF-X by Weight in Cleanest
sample (n6034)
10 1
1 2
1 5
(N4443)
(N1174)
(N417)
35
Animacy, discourse status, and weight all show
strong effects. If we hold one factor constant,
do the other factors disappear?
First we will hold animacy constant and look at
the effects of discourse status, through the
proxy of NP form.
36
Ratio of X's to OF-X by NP form, controlling
Animacy (n6034)
(N4177)
(N498)
(N1359)
37
If we hold NP form constant, do the Animacy
rankings hold up?
38
Ratio of X's to OF-X by Animacy category,
controlling NP form (n6034)
(n3577)
(n971)
(n947)
(n539)
39
Do the animacy and discourse status ratios hold
the same relative values when we control for
weight?
Yes. If we repeat the process for all tokens
with modifiers of weight 1, 2, and 3, the
relative ranking of ratios stays the same, and
the magnitude of the differences persists.
40
Just how robust are these results?
The animacy and discourse status effects remained
intact no matter what we controlled for. The
relative ranking of the animacy and NP form
categories was unchanged, although the ratios
themselves differed somewhat in magnitude.
What would happen if we computed the same
ratios on our original sample of 'All NPs'
(n10,006)? Did all our laborious extractions
and exclusions really make a difference?
41
Return to initial sample, "All NPs"
X'S 4744 47
OF-X 5263 53
42
Ratio of X'S to OF-X by NP form type Comparison
of Cleanest (n6034) vs. All NPs (n9963)
297
28
1.85
0.96
0.44
0.15
0.13
0.04
43
Ratio of X'S to OF-X by Animacy Comparison of
Cleanest (n6140) vs. All NPs (n9963)
14.59
5.53
1.29
0.59
0.11
0.36
44
Interpreting the results
45
Recall our goal
4. Control for confounds where possible, and try
to model the statistical findings within an OT
grammar, following work by Aissen and others.
This is in progress, with very good results. A
set of three binary constraints fits the data
from our corpus study, and makes predictions that
can be tested cross-linguistically.
46
OT Analysis
In this preliminary phase, we classify possessors
in terms of three binary features
animatehuman pronoun
47

• Input
• anim, hum, pron she
• anim, hum, pron butler
• anim, hum, pron it (organization)
• anim, hum, pron it (animal)
• anim, hum, pron it (other)
• anim, hum, pron dog
• anim, hum, pron government
• anim, hum, pron table

48
OT constraints for the Complement (the X in Of-X)
(a) P/C No pron in Comp. P-NP/C No
pron in Comp. (b) A/C No anim in
Comp. A-I/C No anim in Comp. (c) H/C
No hum in Comp. H-NH/C No humin Comp.

49
OT constraints for the Specifier(the X in X'S)

• (a) NP/S No -pron in Spec.
• NP-P/S No pron in Spec.
• (b) I/S No -anim in Spec.
• I-A/S No anim in Spec.
• (c) NH/S No -hum in Spec.
• NH-H/S No hum in Spec.

50
19 predicted languages (out of 256 logically
possible ones)
51

• The OT grammar imposes a partial ordering on
  possessor types in terms of their propensity to
  appear in the Specifier and Complement positions.

52
Aissen Lattice

53
Empirical support?

• Do the predicted rankings of affinity for Spec
  (X'S) hold in our corpus?
• We examined the percentages of the relevant NPs
  in Spec vs. Comp

54
Aissen Lattice
99.9
100
100
75

37
97
50
9
55
Confirmation

• The predictions hold nearly perfectly
• in the 'Cleanest' sample.

Will they hold for the 'All NP' sample?
56
Aissen Lattice
98
99
100
31

18
76
35
3
57
Conclusions
We conducted a large corpus study in which many
features were controlled and investigated
(relationality) We found robust effects for
animacy, discourse status and weight difficult
to disaggregate We modelled these using three
binary features and six pairs of constraints we
made predictions about crosslinguistic factorial
typology.
58
Conclusions
However, the inevitable question "So what?"
"Using OT you can probably model a guy frying an
egg." (Arto Anttila)
Moreover, our OT model treated discourse status
("pronoun") and humanness and animacy as
independent factors. This is probably wrong, and
there is nothing explanatory about it.
59
Conclusions
The question we would like to answer is a
metatheoretical question
What does it mean for the theory of grammar that
these animacy and discourse and weight effects
are so robust, and yet so inextricably combined?
Is it possible to demonstrate in a principled way
whether one derives from another, at least with
respect to the alternation between the X'S and
OF-X constructions? We haven't answered this
question.
60
Conclusions
So now that we have spent several years looking
closely at thousands of examples of the two
constructions, and now that we have shown that
some effects are so robust that you can get them
with really unfiltered data, we would like to
invite anyone who is interested to try their hand
at these questions. We will be putting our coded
corpus on the web soon and we invite you to
investigate it yourself...
(See acknowledgements section of handout for URL
and email details)
61
AcknowledgementsThis research was supported by
NSF grant BCS-008037, "Optimal Typology of
Determiner Phrases". The support of the NSF
Linguistics Program is gratefully acknowledged.
No endorsement of this research is implied.Many
thanks to our graduate research
assistantsGregory GarretsonMarj HoganBarbora
Skarabelaand our undergraduate research
assistantsAmy Rose DealJohn MannaMany thanks
also to Joan Bresnan, Annie Zaenen, and Tom
Wasow, for discussions of animacy.Thanks to
Boston University students in LS 751, Spring
2002, for discussions of some of this material.
62
A conjecture
Claim The prenominal position favors
accessible, topical referents. (Many
clause-level constructions tend towards 'old
first, new last,' and this can be observed in the
NP as well.)
Observation Discourse topics tend to be human
1st and 2nd persons are among the most accessible
entities in any speech situation.
As speakers discuss events involving animate
actors, inanimate NPs introduce background
objects, properties, and arguments of many
predicates.
63
A conjecture
How does this well-worn observation buy us
anything?
If Animates are highly topical, and thus highly
accessible entities, thus mentioned frequently,
they will be expressed with pronouns lighter
entities are favored in initial position. So are
older entities. The preference of Animates
(particularly Humans) for the initial position is
a by-product of their topicality and concomitant
length.
64
A conjecture
If inanimates are usually mentioned only once or
twice, they will predominantly get expressed as
definite and indefinite common nouns that must be
fully informative, thus longer.
This may explain the redundancy between
inanimates, common nouns, higher weights, and the
OF-X position.
Inanimates are not usually highly topical, and
thus they do not favor the X'S slot. Their
appearance in the OF-X construction is a
by-product of their non-topical status and
length, not a fact about animacy per se.
65
Hypothesis 4
LEXICAL SEMANTICS Some have claimed that
features of the head noun (e.g. relationality) or
the semantic relation between the head and
modifier account for most of the variation
(Stefanowitsch, Taylor, Barker, i.a.)
66
Relational semantics
Recall the lexical semantics hypothesis. We have
tried to control for effects of nominal head
semantics by excluding as many strict and soft
non-reversibles as we can find. However, there
is one more issue to deal with. Barker,
Stefanowitsch, and others have claimed that only
relational heads are truly reversible. Yet our
X'S sample includes many examples of possession
of non-relational nouns.
Kim's truck --
??the truck of Kim
These are said to be irreversible because truck
is not relational.
67
Relational semantics
If we limited our X'S and OF-X sample to only
relational heads, would the animacy and discourse
status effects disappear or persist?
To test this, we selected a large sample of
relational heads from our Clean sample (which
excludes all strict non-reversible constructions,
e.g. partitives etc.) We disaggregated all the
examples that had kinship or body part heads his
cousin or the feet of Fred Astaire. This
included 934 tokens.
68
Ratio of X'S to OF-X by NP form type. "Clean"
sample Relational Heads only (n934)
350
9.0
1.89
.22
69
Relational semantics
Then we took all the tokens that had Concrete
Inanimate heads (excluding body parts, which are
relational.) Not all of these examples are
non-relational, but the prototypical examples of
non-relational nouns often include concrete
objects that may be possessed by humans but do
not have any discernible argument structure of
their own.
This included 489 tokens.
70
Ratio of X'S to OF-X by NP form type. Clean
sample Relational Heads (n934) vs. Concrete
Inanimate Heads (n489)
350
327
9.0
1.89
1.9
.72
.22
.52
71
Text analysis (Brown Corpus) an excerpt from a
Western novel.
Some facts 2000 words Approximately 650 coded NP
'mentions' Approximately 150 distinct 'referents'
Dan Morgan his dry lips night his plans and
dreams sleep a wife ...as fickle as Ann
72
Core participantsthose with the most
mentionsDan Morgan 165 mentionsThe visitors
63 mentionsSharon Jones 87 mentionsBilly
Jones 43 mentions
Peripheral participants non-core a trick Al
Budd had thought up 2 mentionsa pathetic,
woebegone expression 1 mentionan idiot 1
mentionordinary years 1 mention
73
Core participantsDan Morgan 165 mentionsThe
visitors 63 mentionsSharon Jones 87
mentionsBilly Jones 43 mentions
How are core participants realized?More than
half of the NPs in the text refer to one of these
four "Core" participants. Of these 358 mentions,
83 are pronouns.
74
Peripheral elementsa trick Al Budd had thought
up 2 mentionsa pathetic, woebegone expression
1 mentionan idiot 1 mentionordinary years 1
mention
How are peripheral participants realized?Over
75 of the 308 "Peripheral" participants are
mentioned only once or twice.Of these 308 NPs,
over 85 are common nouns.