learning when (and when not) to omit objects in english: the role of verb semantic selectivity...

37
Learning When (and When Not) to Omit Objects in English: The Role of Verb Semantic Selectivity Tamara Nicol Medina IRCS, University of Pennsylvania Collaborators: Barbara Landau, Johns Hopkins University Philip Resnik, University of Maryland

Upload: marilynn-holmes

Post on 16-Dec-2015

214 views

Category:

Documents


0 download

TRANSCRIPT

Learning When (and When Not)to Omit Objects in English:

The Role of Verb Semantic Selectivity

Tamara Nicol Medina

IRCS, University of Pennsylvania

Collaborators:Barbara Landau, Johns Hopkins University

Philip Resnik, University of Maryland

The (Indefinite) Implicit Object Construction (in English)

John is eatingJohn is reading

• Verb selects for an object, but none is overtly specified.

• Interpretation is of an indefinite and non-specific object.

(something / some food). (something / written material).

* John is reading (War and Peace).

• Grammaticality varies across verbs.

* John is pushing.* John is opening.

Verb Semantic Selectivity

Overview

1. Measures of Verb Semantic Selectivity

• Selectional Preference Strength (Resnik, 1996)

• Object Similarity

2. Children’s knowledge of Verb Semantic Selectivity

3. Implicit objects in spontaneous speech

• Young child• Mother

Verb Semantic Selectivity• The omitted object tends to be

recoverable from the verb.

John is eating (some food) / drinking (a beverage) / singing (a song).

• Verbs that select for a wide variety of semantic complements, and therefore there is no one recoverable interpretation, tend to resist implicit objects.

John is bringing *(something) / making *(something) / hanging *(something).

Indefinite implicit objects are allowed to the extent that they are recoverable.

Selectional Preference Strength (SPS) (Resnik, 1996)

Don’t push your brother.Move that chair.Do you want an apple?

“like”

Tony likes that girl.I don’t like this couch.I really like bananas.

People Furniture FoodsPeople Furniture Foods

“eat”

Eat your lunch.He’s eating cereal.She always eats avocados.

People Furniture Foods

cvc

vcvSPS i

cii Pr

PrlogPr

An information-theoretic model of verbs’ strength of semantic preferences. Calculates the strength of a verb’s selection for the semantic argument classes from which its complements (or objects) are drawn.

For all argument classes (c), PRIOR, Pr(c) – the overall distribution of argument classes POSTERIOR, Pr(c|vi) – the distribution of argument classes, given a particular verbThe greater the difference between Pr(c) and Pr(c|vi), the higher SPS will be.(Argument classes were those listed in WordNet.)

SPS and Implicit Objects

Relative SPS is correlated with the relative frequency of an implicit object.Brown corpus of American English (Francis and Kučera, 1982 )

0%

10%

20%

30%

40%

50%

60%

70%

80%

90%

100%

po

ur

dri

nk

pa

cksi

ng

ste

al

ea

th

an

gw

ea

ro

pe

np

ush

say

pu

lllik

ew

rite

pla

yh

itca

tch

exp

lare

ad

wa

tc do

he

ar

call

wa

nt

sho

wb

rin

gp

ut

see

find

take ge

tg

ive

ma

keVerb

% Implicit Objects

4.80

0.72

SPS% Implicit Objects

SPS

r = 0.48, p < 0.05

Object Similarity

A psychological measure of the semantic selectivity of a verb for its objects. Calculated as the average of similarity judgments given for pairs of objects that occur with a verb.

“cereal” “bacon” 4

“clothes” “shirt” 5

“juice” “ladybug” 1

“a pencil” “a summer toy” 1

EAT

PACK

BRING

WANT

Similarity judgments made over the actual complements, not the argument classes.

Children’s Knowledge ofVerbs’ Semantic Selectivity

Does children’s knowledge of verbs’ semantic selectional preferences correspond to their mothers’?

• Comprehension: Accurate interpretation of omitted object

• Production: Appropriate restriction of omitted objects

Children’s Knowledge ofVerbs’ Semantic Selectivity

Subjects Children

• 2;6-3;0 yrs, n=20• 3;6-4;0 yrs, n=20

Mothers• earlier period, n=10• later period, n=10

Stimuli

30 verbs from Resnik (1996)

Children’s Knowledge ofVerbs’ Semantic Selectivity

Procedure Children

• Question and answer game with puppet: “What are some things you (could)…

eat / bring / catch / etc. ?”• 15 verbs per child• 3 responses per verb• Stickers as reward

Mothers• Take home questionnaire• Instructed to “use the kind of words you would use in conversation with your child”.

SPS and OS:Comparable Measures

Mothers of Older Children

1.00

1.50

2.00

2.50

3.00

3.50

4.00

4.50

5.00

5.50

6.00

1.00 1.50 2.00 2.50 3.00

SPS

OS

r = 0.72, p < 0.05

Correlated. Some high,

some low, some in-between.

SPS and OS:Comparable Measures

Mothers of Younger Children

r = 0.60, p < 0.05

1.00

1.50

2.00

2.50

3.00

3.50

4.00

4.50

5.00

5.50

6.00

1.00 1.50 2.00 2.50 3.00 3.50

SPS

OS

Correlated. Some high,

some low, some in-between.

Older Children and Mothers

1.00

1.50

2.00

2.50

3.00

3.50

4.00

1.00 1.50 2.00 2.50 3.00 3.50 4.00

Mothers of Older Children

Old

er C

hil

dre

n

SPSr = 0.79, p < 0.05

t(1,29) = -2.39, p < 0.05

1.00

1.50

2.00

2.50

3.00

3.50

4.00

4.50

5.00

5.50

6.00

1.00 2.00 3.00 4.00 5.00 6.00

Mothers of Older Children

Old

er C

hil

dre

n

OSr = 0.67, p < 0.05

t(1,29) = -2.38, p < 0.05

Correlated, but older children’s SPS and OS are lower(broader semantic selectivity).

Younger Children and Mothers

SPSr = 0.75, p < 0.05

t(1,29) = -3.69, p < 0.05

OSr = 0.67, p < 0.05

t(1,29) = -2.33, p < 0.05

Correlated, but younger children’s SPS and OS are lower(broader semantic selectivity).

1.00

1.50

2.00

2.50

3.00

3.50

4.00

1.00 1.50 2.00 2.50 3.00 3.50 4.00

Mothers of Younger Children

Yo

un

ger

Ch

ild

ren

1.00

1.50

2.00

2.50

3.00

3.50

4.00

4.50

5.00

5.50

6.00

1.00 2.00 3.00 4.00 5.00 6.00

Mothers of Younger Children

Yo

un

ger

Ch

ild

ren

Younger and Older Children

SPSr = 0.85, p < 0.05

t(1,29) = 1.27, p > 0.05

OSr = 0.74, p < 0.05

t(1,29) = -0.41, p > 0.05

Correlated. Younger children’s SPS and OS are NOT lower than older children’s (equivalent semantic selectivity).

1.00

1.50

2.00

2.50

3.00

3.50

4.00

1.00 1.50 2.00 2.50 3.00 3.50 4.00

Older Children

Yo

un

ge

r C

hil

dre

n

1.00

1.50

2.00

2.50

3.00

3.50

4.00

4.50

5.00

5.50

6.00

1.00 2.00 3.00 4.00 5.00 6.00

Mothers of Younger Children

Yo

un

ge

r C

hil

dre

n

Summary and Discussion

Correlations between children’s and mothers’ SPS/OS:Verbs that are higher in SPS/OS for mothers are also higher in SPS/OS for children.

- Could recover meaning of implicit objects- Could notice that only the high SPS/OS verbs occur with implicit objects

At both age periods, children’s SPS/OS is lower overall than their mothers’.

Does this really mean that children’s selection of semantic arguments is broader than their mothers’?

- Maybe- Could be children’s approach to the task

Implicit Objectsin Spontaneous Speech

Does the use of indefinite implicit objects correspond to Verb Semantic Selectivity…

• In child-directed speech?• In the child’s own productions?

Implicit Objectsin Spontaneous Speech

Corpus

Sarah corpus (Brown, 1973), set of 29 verbs Sarah

• 2;6-3;0 yrs, 455 utterances• 3;6-4;0 yrs, 559 utterances

Sarah’s Mother• earlier period, 836 utterances• later period, 706 utterances

Children’s Knowledge ofVerbs’ Semantic Selectivity

Coding• Presence or absence of complement

Definite Implicit Object = referrent available in previous 4-5 utterances and/or physically present

Indefinite Implicit Object = referrent NOT available in previous 4-5 utterances and/or physically present

• Grammaticality• Presence of absence of subject

Measures• SPS / OS from Elicited Objects Task

Rate ofIndefinite Implicit Objects

0%

2%

4%

6%

8%

Sarah Mother Sarah Mother

2;6 - 3;0 3;6 - 4;0

% Im

plic

it O

bje

cts

.

7.5% (n = 34)

11 verbs

71% grammatical

4.5% (n = 38)

7 verbs

100% grammatical

*

4.8% (n = 27)

9 verbs

81% grammatical

3.4% (n = 24)

6 verbs

100% grammatical

Possible Reasons forObject Omission

• Verb Semantic Selectivity• “Matching” the input• Memory overload

Indefinite Implicit Objectsand Verb Semantic Selectivity

Does the rate of indefinite implicit objects increase as a function of SPS/OS?Linear Regression – Does one continuous variable (SPS/OS) predict another (rate of implicit objects)?

0.00

0.10

0.20

0.30

0.40

0.50

0.60

0.70

0.80

0.90

1.00

0.50 1.50 2.50 3.50 4.50

SPS

Rat

e o

f Im

plic

it O

bje

cts

F(1,23) = 0.33, p > 0.05

0.00

0.10

0.20

0.30

0.40

0.50

0.60

0.70

0.80

0.90

1.00

1.00 3.00 5.00 7.00

OS

Rat

e o

f Im

plic

it O

bje

cts

F(1,23) = 0.88, p > 0.05

Sarah’s Mother: Older Age Period

Indefinite Implicit Objectsand Verb Semantic Selectivity

Does the rate of indefinite implicit objects increase as a function of SPS/OS?Linear Regression – Does one continuous variable (SPS/OS) predict another (rate of implicit objects)?

F(1,23) = 3.04, p > 0.05 F(1,23) = 3.62, p > 0.05

Sarah’s Mother: Younger Age Period

0.00

0.10

0.20

0.30

0.40

0.50

0.60

0.70

0.80

0.90

1.00

0.50 1.50 2.50 3.50 4.50

SPS

Rat

e o

f Im

plic

it O

bje

cts

0.00

0.10

0.20

0.30

0.40

0.50

0.60

0.70

0.80

0.90

1.00

1.00 3.00 5.00 7.00

OS

Rat

e o

f Im

plic

it O

bje

cts

Indefinite Implicit Objectsand Verb Semantic Selectivity

Does the rate of indefinite implicit objects increase as a function of SPS/OS?Linear Regression – Does one continuous variable (SPS/OS) predict another (rate of implicit objects)?

F(1,23) = 3.06, p > 0.05 F(1,23) = 3.62, p > 0.05

Sarah: Older Age Period

0.00

0.10

0.20

0.30

0.40

0.50

0.60

0.70

0.80

0.90

1.00

0.50 1.50 2.50 3.50 4.50

SPS

Rat

e o

f Im

plic

it O

bje

cts

0.00

0.10

0.20

0.30

0.40

0.50

0.60

0.70

0.80

0.90

1.00

1.00 3.00 5.00 7.00

OS

Rat

e o

f Im

plic

it O

bje

cts

Indefinite Implicit Objectsand Verb Semantic Selectivity

Does the rate of indefinite implicit objects increase as a function of SPS/OS?Linear Regression – Does one continuous variable (SPS/OS) predict another (rate of implicit objects)?

F(1,23) = 3.49, p > 0.05 F(1,23) = 8.23, p < 0.05

Sarah: Younger Age Period

0.00

0.10

0.20

0.30

0.40

0.50

0.60

0.70

0.80

0.90

1.00

0.50 1.50 2.50 3.50 4.50

SPS

Rat

e o

f Im

plic

it O

bje

cts

0.00

0.10

0.20

0.30

0.40

0.50

0.60

0.70

0.80

0.90

1.00

1.00 3.00 5.00 7.00

OS

Rat

e o

f Im

plic

it O

bje

cts

Indefinite Implicit Objectsand Verb Semantic Selectivity

Does the rate of indefinite implicit objects increase as a function of SPS/OS?

No.

(Except for Sarah with regard to OS during the younger age period.)

Indefinite Implicit Objectsand Verb Semantic Selectivity

Is a verb more likely to occur with an implicit object given higher SPS/OS?Logistic Regression – Does one continuous variable (SPS/OS) predict a binary variable (whether a verb ever occurred with an implicit object)?

2 = 2.24, p < 0.05

Sarah’s Mother: Older Age Period

0.00

0.10

0.20

0.30

0.40

0.50

0.60

0.70

0.80

0.90

1.00

0.50 1.50 2.50 3.50 4.50

SPS

Est

. Pr.

of

Imp

licit

Ob

ject

s

2 = 4.38, p < 0.05

0.00

0.10

0.20

0.30

0.40

0.50

0.60

0.70

0.80

0.90

1.00

1.00 3.00 5.00 7.00

OS

Est

. Pr.

of

Imp

licit

Ob

ject

s

Indefinite Implicit Objectsand Verb Semantic Selectivity

Is a verb more likely to occur with an implicit object given higher SPS/OS?Logistic Regression – Does one continuous variable (SPS/OS) predict a binary variable (whether a verb ever occurred with an implicit object)?

2 = 7.19, p < 0.05

Sarah’s Mother: Younger Age Period

2 = 8.10, p < 0.05

0.00

0.10

0.20

0.30

0.40

0.50

0.60

0.70

0.80

0.90

1.00

0.50 1.50 2.50 3.50 4.50

SPS

Est

. Pr.

of

Imp

licit

Ob

ject

s

0.00

0.10

0.20

0.30

0.40

0.50

0.60

0.70

0.80

0.90

1.00

1.00 3.00 5.00 7.00

OS

Est

. Pr.

of

Imp

licit

Ob

ject

s

Indefinite Implicit Objectsand Verb Semantic Selectivity

Is a verb more likely to occur with an implicit object given higher SPS/OS?Logistic Regression – Does one continuous variable (SPS/OS) predict a binary variable (whether a verb ever occurred with an implicit object)?

2 = 3.14, p = 0.07

Sarah: Older Age Period

2 = 2.51, p = 0.11

0.00

0.10

0.20

0.30

0.40

0.50

0.60

0.70

0.80

0.90

1.00

0.50 1.50 2.50 3.50 4.50

SPS

Est

. Pr.

of

Imp

licit

Ob

ject

s

0.00

0.10

0.20

0.30

0.40

0.50

0.60

0.70

0.80

0.90

1.00

1.00 3.00 5.00 7.00

OS

Est

. Pr.

of

Imp

licit

Ob

ject

s

Indefinite Implicit Objectsand Verb Semantic Selectivity

Is a verb more likely to occur with an implicit object given higher SPS/OS?Logistic Regression – Does one continuous variable (SPS/OS) predict a binary variable (whether a verb ever occurred with an implicit object)?

2 = 0.39, p > 0.05

Sarah: Younger Age Period

2 = 0.06, p > 0.05

0.00

0.10

0.20

0.30

0.40

0.50

0.60

0.70

0.80

0.90

1.00

0.50 1.50 2.50 3.50 4.50

SPS

Est

. Pr.

of

Imp

licit

Ob

ject

s

0.00

0.10

0.20

0.30

0.40

0.50

0.60

0.70

0.80

0.90

1.00

1.00 3.00 5.00 7.00

OS

Est

. Pr.

of

Imp

licit

Ob

ject

s

Indefinite Implicit Objectsand Verb Semantic Selectivity

Is a verb more likely to occur with an implicit object given higher SPS/OS?

Yes, for Sarah’s mother.

No, for Sarah. But there appears to be improvement by the older age period.

Verbs Used withImplicit Objects

Does Sarah use implicit objects with the same verbs as her mother?

0

5

10

15

20

25

30

35

40

Both Sarah Only Both Sarah Only

Younger Age Period Older Age Period

Ave

. Rat

e o

f Im

plic

it O

bje

ct P

er V

erb

6 verbs16.7 - 100%

5 verbs1.3 - 20%

5 verbs11.1 – 50.%

4 verbs6.3 – 33%

37% 3.4% 34% 9.8%

0

2

4

6

8

10

Overt Subject Null Subject

Rat

e of

Inde

finite

Impl

icit

Obj

ects

A Performance Explanation?

Could memory overload be contributing to Sarah’s higher rate of indefinite implicit objects at the younger age period?

23 implicit objects (7.1%) 11 implicit objects (8.4%)

2 = 0.23, p > 0.05

Review of Findings:Verb Semantic Selectivity

• At both age periods, children’s verb semantic selectivity appeared somewhat broader than their mothers’.

– May be due to the nature of the task.

• Even so, verbs that are high/low in SPS/OS for mothers were similarly high/low for children.

– Comprehend indefinite implicit objects in the child-directed input.

– Recognize the systematicity with which implicit objects occur in the child-directed input.

Review of Findings:Indefinite Implicit Objects

Sarah’s Mother• Higher SPS/OS predicted which verbs she used indefinite

implicit objects with.– Did not use a higher rate of implicit objects as a function of

increasing SPS/OS.

Sarah• Higher SPS/OS did not predict which verbs she used implicit

objects with.– However, she appears to be getting closer by the

older age period.

• Used implicit objects with some verbs that her mother did not.

– However, at both age periods, the majority of her implicit objects are with the same verbs her mother used with implicit objects.

• Not due to memory overload.

Discussion• Not a conservative start.

– Child doesn’t wait to hear which verbs occur with implicit objects, and then start dropping with these verbs herself.

• How will she arrive at the adult target grammar?– By paying attention to the relationship between a

verb’s SPS/OS and whether it occurs with an overt or implicit indefinite object. Younger age period: Mother used 18 overt indefinite

objects with 5 low SPS/OS verbs. Older age period: Mother used 7 overt indefinite objects

with 7 low SPS/OS verbs.

– Adjust parameters of grammar to result in overt objects for low SPS/OS verbs.

Selectional Preference Strength (SPS) (Resnik, 1996)

cvc

vcvSPS i

cii Pr

PrlogPr

c iii pccpcpSPS

Pr

1log

Pr

1logPr

nfreq

nclassesNc

cwordsn

11

Pr

nvfreq

nclassesNvc i

cwordsni ,

11Pr