Constructing Grammar: a computational model of the acquisition of early constructions

1
Constructing Grammar a computational model of
the acquisition of early constructions

• CS 182 Lecture
• April 27, 2004

2
Acknowledgements
All of the work in this set of slides comes
from Nancy Chang nchang_at_icsi.berkeley.edu UC
Berkeley / International Computer Science
Institute
3
From Single Words To Complex Utterances
FATHER Nomi are you climbing up the
books? NAOMI up. NAOMI climbing. NAOMI books. 1
11.3
FATHER whats the boy doing to the
dog? NAOMI squeezing his neck. NAOMI and the
dog climbed up the tree. NAOMI now theyre both
safe. NAOMI but he can climb trees. 49.3
MOTHER what are you doing? NAOMI I climbing
up. MOTHER youre climbing up? 20.18
Sachs corpus (CHILDES)
4
Development Of Throw
18.0 throw throw off 110.28 I throwded it.
( I fell) I throwded. ( I fell) 111.3 I
throw it. I throw it ice. ( I throw the
ice) throwing in. throwing.
12.9 dont throw the bear. 110.11 dont throw
them on the ground. 111.3 Nomi dont throw
the books down. what do you throw it
into? what did you throw it into? 111.9 they
re throwing this in here. throwing the thing.
5
Development Of Throw (contd)
can I throw it? I throwed Georgie. could I
throw that? 20.5 throw it? you throw
that? 20.18 gonna throw that? 21.17 throw it
in the garbage. throw in there. 25.0 throw it
in that. 211.12 I throwed it in the diaper pail.
20.3 dont throw it Nomi. Nomi stop
throwing. well you really shouldnt throw
things Nomi you know. remember how we told you
you shouldnt throw things.
6
How Can Kids Be So Good At This?

• Golds Theorem
• No superfinite class of language is identifiable
  in the limit from positive data only
• Ah-Ha!
• Kids are born as blank slates. And they soon
  become competent speakers of language(s).
• Language must be innate
• Universal Grammar parameter setting

• But Kids arent born as blank slates!
• And they do not learn language in a vacuum!

7
Language Acquisition

• Opulence of the substrate
• Prelinguistic children already have rich
  sensorimotor representations and sophisticated
  social knowledge
• intention inference, reference resolution
• language-specific event conceptualizations
• (Bloom 2000, Tomasello 1995, Bowerman Choi,
  Slobin, et al.)
• Children are sensitive to statistical information
• Phonological transitional probabilities
• Most frequent items in adult input learned
  earliest
• (Saffran et al. 1998, Tomasello 2000)

8
Language Acquisition

• Basic Scenes
• Simple clause constructions are associated
  directly with scenes basic to human experience
• (Goldberg 1995, Slobin 1985)
• Verb Island Hypothesis
• Children learn their earliest constructions
  (arguments, syntactic marking) on a
  verb-specific basis
• (Tomasello 1992)

throw frisbee
get ball
this should be reminiscent of a6
throw ball
get bottle


get OBJECT
throw OBJECT
9

• Comprehensionispartial.
• (not just for dogs)

10
What We Say To Kids and What They Hear
what did you throw it into? theyre throwing this
in here. theyre throwing a ball. dont throw it
Nomi. well you really shouldnt throw things
Nomi you know. remember how we told you you
shouldnt throw things.
what did you throw it into? theyre throwing this
in here. theyre throwing a ball. dont throw it
Nomi. well you really shouldnt throw things
Nomi you know. remember how we told you you
shouldnt throw things.

• Children use rich situational context / cues to
  fill in the gaps
• They also have at their disposal embodied
  knowledge and statistical correlations (i.e.
  experience)

11
Language Learning Hypothesis

• Children learn constructionsthat bridge the gap
  between
• what they know from language
• and
• what they know from the rest of cognition

12
Lecture Outline

• Introduction
• Computational Bridge
• Learning Algorithm
• Computational Model
• Results

13
Embodied Construction Grammar(Bergen and Chang
2002)

• Embodied representations
• active perceptual and motor schemas (image
  schemas, x-schemas, frames, etc.)
• situational and discourse context
• Construction Grammar
• Linguistic units relate form and
  meaning/function.
• Both constituency and (lexical) dependencies
  allowed.
• Constraint-based
• based on feature structure unification
• Diverse factors can flexibly interact.

14
Usage-based Language Learning
Reorganize
Hypothesize
Partial
Acquisition
15
Words Are Tied To Embodied Meaning
FORM (sound)
MEANING (stuff)
lexical constructions
you
you
Human
throw
throw
ball
ball
Object
block
block
16
In ECG Notation
FORM (sound)
MEANING (stuff)
lexical constructions
schema Addressee subcase of Human
you
you
schema Throw roles thrower throwee
throw
throw
ball
ball
schema Ball subcase of Object
schema Block subcase of Object
block
block
17
Verb Island Construction
FORM (sound)
MEANING (stuff)
Thrower-Throw-Object
t1 before t2 t2 before t3
t2.thrower ? t1 t2.throwee ? t3
schema Addressee subcase of Human
you
you
schema Throw roles thrower throwee
throw
throw
the
schema Ball subcase of Object
ball
ball
schema Block subcase of Object
block
block
18
Analyzing You Throw THe Ball
FORM (sound)
MEANING (stuff)
Thrower-Throw-Object
t1 before t2 t2 before t3
t2.thrower ? t1 t2.throwee ? t3
schema Addressee subcase of Human
Addressee
you
you
schema Throw roles thrower throwee
Throw thrower throwee
throw
throw
the
schema Ball subcase of Object
Ball
ball
ball
schema Block subcase of Object
block
block
19
Verb Island Construction
FORM (sound)
MEANING (stuff)
Thrower-Throw-Object
t1 before t2 t2 before t3
t2.thrower ? t1 t2.throwee ? t3
Addressee
you
you
Throw thrower throwee
throw
throw
the
Ball
ball
ball
schema Block subcase of Object
block
block
20
THe Corresponding Construction
construction THROWER-THROW-OBJECT constructional
constituents t1 REF-EXPRESSION t2
THROW t3 OBJECT-REF form t1f before
t2f t2f before t3f meaning t2m.thrower ?
t1m t2m.throwee ? t3m
role-filler bindings
21
Lecture Outline

• Introduction
• Computational Bridge
• Learning Algorithm
• Relational Mapping
• Merging
• Composing
• Computational Model
• Results

22
Learning-Analysis Cycle
Reorganize

• Learner updates CxnSet

a. Hypothesize new map.
b. Reorganize CxnSet (merge or compose).
c. Reinforce (based on usage).
23
Overview of Learning Processes

• Relational mapping
• throw the ball
• Merging
• throw the block
• throwing the ball
• Composing
• throw the ball
• ball off
• you throw the ball off

24
Initial Single-Word Stage
FORM (sound)
MEANING (stuff)
lexical constructions
schema Addressee subcase of Human
you
you
schema Throw roles thrower throwee
throw
throw
ball
ball
schema Ball subcase of Object
schema Block subcase of Object
block
block
25
New Data You Throw The Ball
FORM
MEANING
SITUATION
Self
schema Addressee subcase of Human
you
Addressee
Addressee
you
schema Throw roles thrower throwee
Throw thrower throwee
Throw thrower throwee
throw
throw
the
schema Ball subcase of Object
ball
Ball
Ball
ball
schema Block subcase of Object
block
block
26
New Data You Throw The Ball
FORM
MEANING
SITUATION
Self
schema Addressee subcase of Human
you
Addressee
you
schema Throw roles thrower throwee
Throw thrower throwee
throw
throw
the
schema Ball subcase of Object
ball
Ball
ball
schema Block subcase of Object
block
block
27
New Data You Throw The Ball
FORM
MEANING
SITUATION
Self
you
Addressee
you
Throw thrower throwee
throw
throw
the
ball
Ball
ball
schema Block subcase of Object
block
block
28
New Data You Throw The Ball
FORM
MEANING
SITUATION
Self
you
Addressee
you
Throw thrower throwee
throw
throw
the
ball
Ball
ball
schema Block subcase of Object
block
block
29
New Data You Throw The Ball
FORM
MEANING
SITUATION
Self
you
Addressee
you
Throw thrower throwee
throw
throw
the
ball
Ball
ball
schema Block subcase of Object
block
block
30
New Construction Hypothesized
construction THROW-BALL constructional
constituents t THROW b
BALL form tf before bf meaning tm.throwee
? bm
31
Relational Mapping

• Creating a new construction with constituents A
  and B is equivalent to learning a mapping between
  form relation (e.g. word order) over Af and Bf
  and meaning relation over Am and Bm

Throw thrower throwee
throw
throw
the
ball
Ball
ball
32
Meaning Relations

• strictly isomorphic
• Bm is a role-filler of Am (or vice versa)
• (Am.r1 ? Bm)
• shared role-filler
• Am and Bm each have a role filled by the same
  entity
• (Am.r1 ? Bm.r2)
• sibling role-fillers
• Am and Bm fill roles of the same schema
• (Y.r1 ? Am, Y.r2 ? Bm)

33
When Can We Hypothesize A New Map

• strictly isomorphic
• Bm is a role-filler of Am (or vice versa)
• (Am.r1 ? Bm)

Af
Am
A
form-relation
role-filler
Bf
Bm
B
throw ball throw.throwee ? ball
34
When Can We Hypothesize A New Map

• shared role-filler
• Am and Bm each have a role filled by the same
  entity
• (Am.r1 ? Bm.r2)

role-filler
Af
Am
A
form-relation
X
role-filler
Bf
Bm
B
put ball down put.mover ? ball down.tr ?
ball
35
When Can We Hypothesize A New Map

• sibling role-fillers
• Am and Bm fill roles of the same schema
• (Y.r1 ? Am, Y.r2 ? Bm)

role-filler
Af
Am
A
form-relation
Y
role-filler
Bf
Bm
B
Nomi ball possession.possessor ?
Nomi possession.possessed ? ball
36
Overview of Learning Processes

• Relational mapping
• throw the ball
• Merging
• throw the block
• throwing the ball
• Composing
• throw the ball
• ball off
• you throw the ball off

37
Merging Similar Constructions
Throw.throwee Block
throw before block
throw before ball
Throw.throwee Ball
Throw.aspect ongoing
throw before-s ing
38
Resulting Construction
construction THROW-OBJECT constructional
constituents t THROW o
OBJECT form tf before of meaning
tm.throwee ? om
39
Or Merge Similar Suffix
40
Overview of Learning Processes

• Relational mapping
• throw the ball
• Merging
• throw the block
• throwing the ball
• Composing
• throw the ball
• ball off
• you throw the ball off

41
Composing Co-occurring Constructions
42
Resulting Construction
construction THROW-BALL-OFF constructional
constituents t THROW b BALL o
OFF form tf before bf bf before
of meaning evokes MOTION as m tm.throwee ?
bm m.mover ? bm m.path ? om
43
Lecture Outline

• Introduction
• Computational Bridge
• Learning Algorithm
• Computational Model
• Results

44
Language Learning Problem

• Prior knowledge
• Initial grammar G (set of ECG constructions)
• Ontology (category relations)
• Language comprehension model (analysis/resolution)

45
Language Learning Problem

• Training data set of (utterance, situation)
  pairs
• utterance sequence of familiar and novel
  segments
• situation entities relational bindings

46
Language Learning Problem

• Hypothesis space new ECG grammar G
• Search processes for proposing new constructions

47
Language Learning Problem

• Performance measure
• Goal Comprehension should improve with training
• Criterion need some objective function to guide
  learning

Likelihood of Model given Data
Minimum Description Length
48
Minimum Description Length

• Choose grammar G to minimize cost(GD)
• cost(GD) a size(G) ß complexity(DG)
• Approximates Bayesian learning cost(GD)
  posterior probability P(GD)
• Size of grammar size(G) prior P(G)
• favor fewer/smaller constructions/roles
  isomorphic mappings
• Complexity of data given grammar likelihood
  P(DG)
• favor simpler analyses(fewer, more likely
  constructions)
• based on derivation length score of derivation

49
Size Of Grammar

• Size of the grammar G is the sum of the size of
  each construction
• Size of each construction c is
• where
• nc number of constituents in c,
• mc number of constraints in c,
• length(e) slot chain length of element
  reference e

50
Example The Throw-Ball Cxn
construction THROW-BALL constructional
constituents t THROW b
BALL form tf before bf meaning tm.throwee
? bm
size(THROW-BALL) 2 2 (2 3) 9
51
Complexity of Data Given Grammar

• Complexity of the data D given grammar G is the
  sum of the analysis score of each input token d
• Analysis score of each input token d is
• where
• c is a construction used in the analysis of d
• weightc relative frequency of c,
• typer number of ontology items of type r
  used,
• heightd height of the derivation graph,
• semfitd semantic fit provide by the analyzer

52
Lecture Outline

• Introduction
• Computational Bridge
• Learning Algorithm
• Computational Model
• Results

53
Experiment Learning Verb Islands

• Question
• Can the proposed construction learning model
  acquire English item-based motion constructions?
• Given initial lexicon and ontology
• Data child-directed language annotated with
  contextual information.

54
Experiment Learning Verb Islands

• Subset of the CHILDES database of parent-child
  interactions (MacWhinney 1991 Slobin et al.)
• coded by developmental psychologists for
• form particles, deictics, pronouns, locative
  phrases, etc.
• meaning temporality, person, pragmatic
  function,type of motion (self-movement vs.
  caused movement animate being vs. inanimate
  object, etc.)
• crosslinguistic (English, French, Italian,
  Spanish)
• English motion utterances 829 parent, 690 child
  utterances
• English all utterances 3160 adult, 5408 child
• age span is 12 to 26

55
Learning Throw-Constructions
56
Learning Results
57
Experiment Learning Verb Islands

• Item-based constructions are learned by the
  model.
• Basic processes produce constructions similar to
  those learned by the child.
• Generalization beyond encountered data is
  available.
• Differences in verb learning lend support to verb
  island hypothesis.
• Future directions
• full English corpus
• Crosslinguistic data Russian (case marking),
  Mandarin Chinese (directional particles, aspect
  markers)

58
Summary

• Cognitively plausible situated learning processes
• What do kids start with?
• perceptual, motor, social, world knowledge
• meanings of single words
• What kind of input drives acquisition?
• Social-pragmatic knowledge
• Statistical properties of linguistic input
• What is the learning loop?
• Use existing linguistic knowledge to analyze
  input
• Use social-pragmatic knowledge to understand
  situation
• Hypothesize new constructions to bridge the gap