herself - UMD Linguistics

Simulation of
Pronoun Acquisition
Naho Orita, Naomi Feldman, Jeffrey Lidz
Department of Linguistics
University of Maryland
Learning Problem
“Through the Looking-Glass, and What Alice Found There”, Lewis Carol
Learning Problem
Alice saw herself in the mirror.
Learning Problem
Alice saw herself in the mirror.
Learning Problem
Alice saw her in the mirror.
Learning Problem
Alice saw her in the mirror.
Learning Problem
Alice saw
herself in the
mirror.
Alice saw her
in the mirror.
Learning Problem
myself
ourself
Alice saw herself.
yourself
herself = Alice
herself
himself
themselves
me
us
Alice saw her.
you
her ≠ Alice
her
him
them
Learning Problem
myself
ourself
Reflexive
yourself
pronouns
herself
himself
themselves
me
us
Non-reflexive
you
pronouns
her
him
them
Learning Problem
myself
me
ourself
us
How
do
learners
acquire
yourself
you
this
knowledge?
herself
her
himself
him
themselves
them
Alice saw
Alice saw her
herself in the
in the mirror.
mirror.
Learning Model
Alice saw herself.
Learning Model
Alice saw herself.
Discourse Information
Learning Model
Alice saw herself.
Discourse Information
Pronoun Lexicon
me
himself
...
herself
yourself
she
...
Learning Model
Alice saw herself.
Discourse Information
Pronoun Lexicon
me
himself
...
herself
Syntax
Alice and herself are in the same clause.
Alice c-commands herself.
yourself
she
...
Learning Model
Alice saw herself.
Discourse Information
Pronoun Lexicon
me
himself
...
yourself
she
...
herself
“herself” is reflexive
Syntax
Alice and herself are in the same clause.
Alice c-commands herself.
Learning Model
Learning
Target
Alice saw herself.
Discourse Information
Pronoun Lexicon
me
himself
...
yourself
she
...
herself
“herself” is reflexive
Syntax
Alice and herself are in the same clause.
Alice c-commands herself.
Learning Problem
Alice saw
blick.
Learning Problem
Alice saw
blick.
Alice saw
Alice.
Learning Problem
Alice saw
blick.
Alice saw
Alice.
blick is
reflexive
Learning Problem
Alice saw
blick.
blick is
reflexive
Learning Problem
Alice saw
blick.
Alice saw
Alice.
blick is
reflexive
Learning Problem
Chicken-and-Egg Problem
Alice saw
blick.
Alice saw
Alice.
blick is
reflexive
Hypothesis
?
Alice saw
blick.
Hypothesis
Alice saw
blick.
Hypothesis
Discourse information
helps learners to learn
pronouns.
Alice saw
blick.
Hypothesis
How does discourse information help?
Alice saw splink.
Hypothesis
How does discourse information help?
Alice saw splink.
p(RedQueen)=0.8
p(Alice)=0.2
Hypothesis
How does discourse information help?
Alice saw splink.
p(RedQueen)=0.8
p(Alice)=0.2
Rest of the talk...
• What kind of discourse information is
available in the input?
• Can discourse information help an ideal
learner categorize pronouns?
Rest of the talk...
• What kind of discourse information is
available in the input?
• Can discourse information help an ideal
learner categorize pronouns?
Discourse Information
Alice saw splink.
p(RedQueen)=0.8
p(Alice)=0.2
Discourse Information
• McKeown (2012)
• Human Simulation experiment
Discourse Information
Mother:
Dry your back. Here.
Meghan:
I can’t.
Mother:
Come on now. Be a big girl.
Meghan:
I can’t.
Mother:
You can dry ___.
Meghan:
I can’t.
Discourse Information
Mother:
Dry your back. Here.
Meghan:
I can’t.
Mother:
Come on now. Be a big girl.
Meghan:
I can’t.
Mother:
You can dry ___.
Meghan:
I can’t.
Options
him, himself, me, Meghan, myself, ourselves, the bathmat,
the feet, the tub, them, themselves, us, you, your back,
yourself
Discourse Information
Mother:
Dry your back. Here.
Meghan:
I can’t.
Mother:
Come on now. Be a big girl.
Meghan:
I can’t.
Mother:
You can dry ___.
Meghan:
I can’t.
yourself? it?
Options
him, himself, me, Meghan, myself, ourselves, the bathmat,
the feet, the tub, them, themselves, us, you, your back,
yourself
Discourse Information
Results (N=20) in McKeown (2012)
Actual
category
Full Noun
Nonreflexive
Reflexive
% full Noun % non-refl
guesses
guesses
% refl
guesses
Discourse Information
Results (N=20) in McKeown (2012)
Actual
category
% full Noun % non-refl
guesses
guesses
% refl
guesses
Full Noun
71.8
23.2
5
Nonreflexive
23.4
73.4
3.2
Reflexive
15.8
16
68.2
Discourse Information
Mother:
Meghan:
Mother:
Meghan:
Mother:
Meghan:
Dry your back. Here.
I can’t.
Come on now. Be a big girl.
I can’t.
You can dry ___.
I can’t.
guesses of what entity is
likely to be referred to
p(no antecedent) = 0.15
Prior
p(You) = 0.7
p(else) = 0.15
guesses of the identity
of unknown words
p(full NP) = 0.15
p(reflexive) = 0.7
p(non-refl) = 0.15
Discussion
•
Adults have good guesses of what is referred to
using discourse information.
•
Given that adults can make good guesses,
suppose that children had equal access to those
guesses. Would that be enough information to
drive the correct categorization of the pronouns?
Rest of the talk...
• What kind of discourse information is
available in the input?
• Can discourse information help an ideal
learner categorize pronouns?
Learning Model
Learning
Target
Alice saw herself.
Discourse Information
Pronoun Lexicon
?
herself
“herself” is reflexive
Syntax
Alice and herself are in the same clause.
Alice c-commands herself.
Learning Model
Learning
Target
Alice saw herself.
Discourse Information
Pronoun Lexicon
?
Good
☺
herself
“herself” is reflexive
Syntax
Alice and herself are in the same clause.
Alice c-commands herself.
Learning Model
Learning
Target
Alice saw herself.
Discourse Information
Pronoun Lexicon
?
Good
☺
herself
“herself” is reflexive
Good
☺
Syntax
Alice and herself are in the same clause.
Alice c-commands herself.
Learning Model
Learning
Target
Alice saw herself.
Discourse Information
Pronoun Lexicon
?
Good
☺
herself
“herself” is reflexive
Good
☺
Syntax
Alice and herself are in the same clause.
Alice c-commands herself.
Good
☺
Learning Model
Learning
Target
Alice saw herself.
Discourse Information
Good
Pronoun Lexicon
category 1
category 2
?
?
two categories
☺
herself
“herself” is reflexive
Good
☺
Syntax
Alice and herself are in the same clause.
Alice c-commands herself.
Good
☺
Baseline Model
Learning
Target
Alice saw herself.
Discourse Information
Pronoun Lexicon
category 1
category 2
?
?
two categories
Bad
herself
“herself” is reflexive
Good
☺
Syntax
Alice and herself are in the same clause.
Alice c-commands herself.
Good
☺
Learning Model
Alice saw herself.
Prior
Discourse Information
Posterior
Pronoun Lexicon
myself
herself
...
...
me
her
...
...
herself
“herself” is reflexive
Built-in
Knowledge
Syntax
Alice and herself are in the same clause.
Alice c-commands herself.
Built-in
Knowledge
Generative Model
Red Queen said that Alice saw __
Generative Model
Red Queen said that Alice saw __
p(Alice)=0.8
Antecedent
p(Arefl) = 0.8
p(RedQueen)=0.2
Discourse
knowledge
Generative Model
Red Queen said that Alice saw __
p(Alice)=0.8
Dir(β)
Multi(ɸ)
Antecedent
p(x| Arefl, ɸ)
p(Arefl) = 0.8
p(RedQueen)=0.2
Discourse
knowledge
Generative Model
Red Queen said that Alice saw __
p(Alice)=0.8
Dir(β)
Multi(ɸ)
Pronoun
p(x| Arefl, ɸ)
herself
(Reflexive)
Antecedent
p(Arefl) = 0.8
p(RedQueen)=0.2
Discourse
knowledge
Baseline Model
Red Queen said that Alice saw __
Baseline Model
Red Queen said that Alice saw __
Uniformly
picks antecedent
p(RedQueen)=0.5
Antecedent
p(Alice)=0.5
p(Anon-refl) = 0.5
Baseline Model
Red Queen said that Alice saw __
Uniformly
picks antecedent
p(RedQueen)=0.5
Dir(β)
Multi(ɸ)
Antecedent
p(Alice)=0.5
p(x| Anon-refl, ɸ)
p(Anon-refl) = 0.5
Baseline Model
Red Queen said that Alice saw __
Uniformly
picks antecedent
p(RedQueen)=0.5
Dir(β)
Multi(ɸ)
p(x| Anon-refl, ɸ)
Pronoun
Antecedent
p(Alice)=0.5
her
(Non-reflexive)
p(Anon-refl) = 0.5
Model
Dir(β)
Pronoun
Multi(ɸ)
Discourse
knowledge
Antecedent
K
N
K = 2 (reflexive class and non-reflexive class)
N = 50 (from the experiment)
Model
Dir(β)
Pronoun
Multi(ɸ)
2
Discourse
knowledge
Antecedent
50
•
10 pronouns (myself, ourself, yourself, himself, themselves,
me, us, you, him, them), 50 sentences (25 reflexives, 25 nonreflexives) from McKeown (2012)
•
Used Gibbs sampling for inference.
Model
Posterior we want
Dir(β)
Pronoun
Multi(ɸ)
2
Category 1
Category 2
myself
ourself
yourself
himself
themselves
me
us
you
him
them
Discourse
knowledge
Antecedent
50
Results
Without discourse information
category 1
word
prob.
yourself
him
himself
me
them
themselves
us
you
myself
ourselves
0.29
0.21
0.21
0.12
0.09
0.09
0.0
0.0
0.0
0.0
category 2
word
prob.
you
0.625
myself
0.25
us
0.06
ourselves
0.06
me
0.0
him
0.0
them
0.0
yourself
0.0
himself
0.0
themselves 0.0
With discourse information
Results
Without discourse information
category 1
word
prob.
yourself
him
himself
me
them
themselves
us
you
myself
ourselves
0.29
0.21
0.21
0.12
0.09
0.09
0.0
0.0
0.0
0.0
category 2
word
prob.
you
0.625
myself
0.25
us
0.06
ourselves
0.06
me
0.0
him
0.0
them
0.0
yourself
0.0
himself
0.0
themselves 0.0
With discourse information
category 1
category 2
word
prob.
word
prob.
yourself
himself
myself
themselves
ourselves
you
him
me
them
us
0.4
0.28
0.16
0.12
0.04
0.0
0.0
0.0
0.0
0.0
you
him
me
them
us
yourself
himself
myself
themselves
ourselves
0.4
0.28
0.16
0.12
0.04
0.0
0.0
0.0
0.0
0.0
Discussion
Discourse information can help bootstrap the
chicken-and-egg problem.
Ongoing Work
In the previous model...
Pronoun Lexicon
The model doesn’t
have to learn syntax
relevant to pronoun
categories.
“herself” is reflexive
Syntax
Alice and herself are in the same clause.
Alice c-commands herself.
Good
☺
Ongoing Work
Pronoun Lexicon
But children have to learn
relevant syntactic
representations for
categorizing pronouns.
“herself” ?
Syntax
Alice and herself are in the same clause.
Alice c-commands herself.
??
Ongoing
Work
Discourse
knowledge
𝑥
locality
c-command
α
𝑧
θ
M
β
ɸ
w
K
N
Conclusion
Learning
Target
Alice saw herself.
Discourse Information
Pronoun Lexicon
myself
herself
...
...
☺
me
her
...
...
☺
herself
“herself” is reflexive
Syntax
Alice and herself are in the same clause.
Alice c-commands herself.
Acknowledgements
Thanks to:
Members of Computational Psycholinguistics Lab@UMD
Philip Resnik
Jordan Boyd-Graber
UMD’s Language Science community
http://languagescience.umd.edu
UMD’s NSF-IGERT DGE-0801465