Models of Grammar Learning

CS 182 Lecture

April 24, 2008

2

What constitutes learning a language?

What are the sounds (Phonology) How to make words (Morphology) What do words mean (Semantics) How to put words together (Syntax) Social use of language (Pragmatics) Rules of conversations (Pragmatics)

3

Language Learning Problem

Prior knowledge Initial grammar G (set of ECG

constructions) Ontology (category relations) Language comprehension model

(analysis/resolution)

Hypothesis space: new ECG grammar G’ Search = processes for proposing new

constructions Relational Mapping, Merge, Compose

4

Language Learning Problem

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

learning…

Minimum Description Length:

)P( log)|P( log)|P( log

)P()|P()|P(

)P(

)P()|P()|P(

MMXXM

MMXXM

X

MMXXM

)P( log)|P( log)|P( log MMXXM

Probability of Model given Data:

5

Minimum Description Length

Choose grammar G to minimize cost(G|D): cost(G|D) = α • size(G) + β • complexity(D|G)

Approximates Bayesian learning; cost(G|D) ≈ posterior probability P(G|D)

Size of grammar = size(G) ≈ prior P(G) favor fewer/smaller constructions/roles; isomorphic

mappings

Complexity of data given grammar ≈ likelihood P(D|G) favor simpler analyses

(fewer, more likely constructions) based on derivation length + score of derivation

6

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

Gc

cG )size()size(

ce

cc emnc )length()size(

7

What do we know about language development?

(focusing mainly on first language acquisition of English-speaking, normal population)

8

Children are amazing learners

cooing

reduplicated babbling

first w

ord

0 mos 2 yr6 mos 3 yrs 4 yrs 5 yrs12 mos

two-word combinatio

ns

multi-word utte

rances

questions,

complex sentence

structures, c

onversatio

nal

princip

les

9

Phonology: Non-native contrasts

Werker and Tees (1984) Thompson: velar vs. uvular, /`ki/-/`qi/. Hindi: retroflex vs. dental, /t.a/-/ta/

0

2

4

6

8

10

12

14

16

18

20

6-8 months 8-10 months 10-12 months

yes

no

10

Finding words: Statistical learning

Saffran, Aslin and Newport (1996)

/bidaku/, /padoti/, /golabu/ /bidakupadotigolabubidaku/ 2 minutes of this continuous speech

stream By 8 months infants detect the words (vs

non-words and part-words)

pretty baby

11

Word order: agent and patient

Hirsch-Pasek and Golinkoff (1996) 1;4-1;7

mostly still in the one-word stage

Where is CM tickling BB?

12

Early syntax

agent + action ‘Daddy sit’ action + object ‘drive car’ agent + object ‘Mommy sock’ action + location ‘sit chair’ entity + location ‘toy floor’ possessor + possessed ‘my teddy’ entity + attribute ‘crayon big’ demonstrative + entity ‘this

telephone’

13

MOTHER: what are you doing?

NAOMI: I climbing up.

MOTHER: you’re climbing up?

2;0.18

FATHER: what’s the boy doing to the dog?

NAOMI: squeezing his neck.NAOMI: and the dog climbed

up the tree.NAOMI: now they’re both

safe.NAOMI: but he can climb

trees.4;9.3

FATHER: Nomi are you climbing up the books?

NAOMI: up.NAOMI: climbing.NAOMI: books.

1;11.3

Sachs corpus (CHILDES)

From Single Words To Complex Utterances

14

How Can Children Be So Good At Learning Language?

Gold’s Theorem:

No superfinite class of language is identifiable in the limit from positive data only

Principles & Parameters

Babies are born as blank slates but acquire language quickly (with noisy input and little correction) → Language must be innate:

Universal Grammar + parameter setting

But babies aren’t born as blank slates!

And they do not learn language in a vacuum!

15

Modifications of Gold’s Result

(Weakly) Ordered Examples, implicit negatives

Loosened Identification Conditions Complexity Measures, Best Fit

No Theorems will resolve these issues

16

Modeling the acquisition of grammar:

Theoretical assumptions

17

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 Even dependencies between non-adjacent items

(Saffran et al. 1996, Gomez 2002)

18

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)get ball

get bottle

get OBJECT

…

throw frisbee

throw ball

throw OBJECT…

this should be reminiscent of your

model merging assignment

19

Comprehensionis

partial.

(not just for dogs)

20

What children pick up from what they hear

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)

what did you throw it into?they’re throwing this in here.they’re throwing a ball.don’t throw it Nomi.

well you really shouldn’t throw things Nomi you know. remember how we told you you shouldn’t throw things.

what did you throw it into?they’re throwing this in here.they’re throwing a ball.don’t throw it Nomi.

well you really shouldn’t throw things Nomi you know. remember how we told you you shouldn’t throw things.

21

Language Learning Hypothesis

Children learn constructionsthat bridge the gap between

what they know from language

and

what they know from the rest of cognition

22

Modeling the acquisition of (early) grammar:

Comprehension-driven, usage-based

Natural Language Processing at Berkeley

Dan Klein

EECS Department

UC Berkeley

24

NLP: Motivation

It’d be great if machines could Read text and understand it Translate languages accurately Help us manage, summarize,

and aggregate information Use speech as a UI Talk to us / listen to us

But they can’t Language is complex Language is ambiguous Language is highly structured

25

Machine Translation

Syntactic MT Learn grammar

mappings between languages

Fully data-driven

26

Information Extraction

Unsupervised Coreference Resolution Take in lots of text Learn what the

entities are and how they corefer

Fully unsupervised, but gets supervised performance!

General research goal: unsupervised learning of meaning

27

Syntactic Learning

Grammar Induction Raw text in Learned grammars out Big result: this can be

done!

Grammar Refinement Coarse grammars in Detailed grammars

out Gives top parsing

systems

28

Influental members of the House Ways and Means Committee introduced legislation that would restrict how the new S&L bailout agency can raise capital, creating another potential obstacle to the government's sale of sick thrifts.

Syntactic Inference

Natural language is very ambiguous Grammars are huge Billions of parses

to consider Milliseconds to do it

30

Idea: Learn PCFGs with EM

Classic experiments on learning PCFGs with Expectation-Maximization [Lari and Young, 1990]

Full binary grammar over n symbols Parse uniformly/randomly at first Re-estimate rule expectations off of parses Repeat

Their conclusion: it doesn’t really work.

Xj

Xi

Xk{ X1 , X2 … Xn }

31

:( , , ) ( )

: ( )

( )

( | , , )( )

kT X i j yield T Sc

T yield T S

P T

P X i j SP T

2( | ) 1/a b cP X X X n

Re-estimation of PCFGs

Basic quantity needed for re-estimation with EM:

Can calculate in cubic time with the Inside-Outside algorithm. Consider an initial grammar where all productions have equal weight:

Then all trees have equal probability initially. Therefore, after one round of EM, the posterior over trees will (in the

absence of random perturbation) be approximately uniform over all trees, and symmetric over symbols.

32

Problem: “Uniform” Posteriors

Tree Uniform

Split Uniform

34

Overview: NLP at UCB

Lots of research and resources: Dan Klein: Statistical NLP / ML Marti Hearst: Stat NLP / HCI Jerry Feldman: Language and Mind Michael Jordan: Statistical Methods / ML Tom Griffiths: Statistical Learning / Psychology ICSI Speech and AI groups (Morgan, Stolcke,

Shriberg, Fillmore, Kay, Narayanan…) Great linguistics and stats departments!

No better place to solve the hard NLP problems!

35

Other Approaches

Evaluation: fraction of nodes in gold trees correctly posited in proposed trees (unlabeled recall)

Some recent work in learning constituency: [Adrians, 99] Language grammars aren’t general PCFGs [Clark, 01] Mutual-information filters detect constituents,

then an MDL-guided search assembles them [van Zaanen, 00] Finds low edit-distance sentence pairs

and extracts their differences

Adriaans, 1999 16.8

Clark, 2001 34.6

van Zaanen, 2000 35.6