Page 1

November 2007

Beckman Institute

With thanks to:

Collaborators: Ming-Wei Chang, Vasin Punyakanok, Lev Ratinov,

Mark Sammons, Scott Yih, Dav ZimakFunding: ARDA, under the AQUAINT program

NSF: ITR IIS-0085836, ITR IIS-0428472, ITR IIS- 0085980 A DOI grant under the Reflex program, DASH Optimization (Xpress-MP)

Natural Language ProcessingviaLearning and Global Inference with Constraints

Dan RothDepartment of Computer ScienceUniversity of Illinois at Urbana-Champaign

Page 2

Nice to Meet You

Page 3

Learning and Inference

Global decisions in which several local decisions play a role but there are mutual dependencies on their outcome.

(Learned) models/classifiers for different sub-problems

Incorporate classifiers’ information, along with constraints, in making coherent decisions – decisions that respect the local models as well as domain & context specific constraints.

Global inference for the best assignment to all variables of interest.

Page 4

Inference

Page 5

Comprehension

1. Christopher Robin was born in England. 2. Winnie the Pooh is a title of a book. 3. Christopher Robin’s dad was a magician. 4. Christopher Robin must be at least 65 now.

A process that maintains and updates a collection of propositions about the state of affairs.

(ENGLAND, June, 1989) - Christopher Robin is alive and well. He lives in England. He is the same person that you read about in the book, Winnie the Pooh. As a boy, Chris lived in a pretty home called Cotchfield Farm. When Chris was three years old, his father wrote a poem about him. The poem was printed in a magazine for others to read. Mr. Robin then wrote a book. He made up a fairy tale land where Chris lived. His friends were animals. There was a bear called Winnie the Pooh. There was also an owl and a young pig, called a piglet. All the animals were stuffed toys that Chris owned. Mr. Robin made them come to life with his words. The places in the story were all near Cotchfield Farm. Winnie the Pooh was written in 1925. Children still love to read about Christopher Robin and his animal friends. Most people don't know he is a real person who is grown now. He has written two books of his own. They tell what it is like to be famous.

Page 6

Illinois’ bored of education board

...Nissan Car and truck plant is ……divide life into plant and animal kingdom

(This Art) (can N) (will MD) (rust V) V,N,N

The dog bit the kid. He was taken to a veterinarian a hospital

What we Know: Stand Alone Ambiguity Resolution

Learn a function f: X Y that maps observations in a domain to one of several categories or <

Page 7

Theoretically: generalization bounds How many example does one need to see in order to

guarantee good behavior on previously unobserved examples.

Algorithmically: good learning algorithms for linear representations.

Can deal with very high dimensionality (106 features) Very efficient in terms of computation and # of examples.

On-line. Key issues remaining:

Learning protocols: how to minimize interaction (supervision); how to map domain/task information to supervision; semi-supervised learning; active learning; ranking; sequences; adaptation

What are the features? No good theoretical understanding here.

Programming systems that have multiple classifiers

Classification is Well Understood

Page 8

Comprehension

1. Christopher Robin was born in England. 2. Winnie the Pooh is a title of a book. 3. Christopher Robin’s dad was a magician. 4. Christopher Robin must be at least 65 now.

A process that maintains and updates a collection of propositions about the state of affairs.

(ENGLAND, June, 1989) - Christopher Robin is alive and well. He lives in England. He is the same person that you read about in the book, Winnie the Pooh. As a boy, Chris lived in a pretty home called Cotchfield Farm. When Chris was three years old, his father wrote a poem about him. The poem was printed in a magazine for others to read. Mr. Robin then wrote a book. He made up a fairy tale land where Chris lived. His friends were animals. There was a bear called Winnie the Pooh. There was also an owl and a young pig, called a piglet. All the animals were stuffed toys that Chris owned. Mr. Robin made them come to life with his words. The places in the story were all near Cotchfield Farm. Winnie the Pooh was written in 1925. Children still love to read about Christopher Robin and his animal friends. Most people don't know he is a real person who is grown now. He has written two books of his own. They tell what it is like to be famous.

This is an Inference Problem

Page 9

This Talk

Global Inference over Local Models/Classifiers + Expressive Constraints

Model Generality of the framework

Training Paradigms

Global vs. Local training Semi-Supervised Learning

Examples Semantic Parsing Information Extraction Pipeline processes

Page 10

Sequential Constrains Structure

Three models for sequential inference with classifiers[Punyakanok & Roth NIPS’01]

HMM; HMM with Classifiers Sufficient for easy problems

Conditional Models (PMM) Allows direct modeling of states as a function of input Classifiers may vary; SNoW (Winnow;Perceptron); MEMM: MaxEnt;

SVM based

Constraint Satisfaction Models (CSCL: more general constrains)

The inference problem is modeled as weighted 2-SAT With sequential constraints: shown to have efficient solution

Recent work – viewed as multi-class classification; emphasis on global training [Collins’02, CRFs,SVMs]; efficiency and performance issues

s1

o1

s2

o2

s3

o3

s4

o4

s5

o5

s6

o6

s1

o1

s2

o2

s3

o3

s4

o4

s5

o5

s6

o6By far, the most popular in applications

Allows for Dynamic Programming based

Inference

What if the structure of the problem/constraints is not sequential?

Page 11

Pipeline

Pipelining is a crude approximation; interactions occur across levels and down stream decisions often interact with previous decisions.

Leads to propagation of errors Occasionally, later stage problems are easier but

upstream mistakes will not be corrected.

POS Tagging

Phrases

Semantic Entities

Relations

Most problems are not single classification problems

Parsing

WSD Semantic Role Labeling

Raw Data

Looking for: Global inference over the outcomes of different (learned) predictors as a way to break away from this paradigm [between pipeline & fully global] Allows a flexible way to incorporate linguistic and structural constraints.

Page 12

Inference with General Constraint Structure

Dole ’s wife, Elizabeth , is a native of N.C.

E1 E2 E3

R12 R2

3

other 0.05per 0.85loc 0.10

other 0.05per 0.50loc 0.45

other 0.10per 0.60loc 0.30

irrelevant 0.10spouse_of 0.05born_in 0.85

irrelevant 0.05spouse_of 0.45born_in 0.50

irrelevant 0.05spouse_of 0.45born_in 0.50

other 0.05per 0.85loc 0.10

other 0.10per 0.60loc 0.30

other 0.05per 0.50loc 0.45

irrelevant 0.05spouse_of 0.45born_in 0.50

irrelevant 0.10spouse_of 0.05born_in 0.85

other 0.05per 0.50loc 0.45

Improvement over no inference: 2-5%

Page 13

Random Variables Y:

Conditional Distributions P (learned by models/classifiers) Constraints C– any Boolean function defined on partial assignments (possibly: + weights W )

Goal: Find the “best” assignment The assignment that achieves the highest global

performance. This is an Integer Programming Problem

Problem Setting

y7y4 y5 y6 y8

y1 y2 y3C(y1,y4)C(y2,y3,y6,y7,y8)

Y*=argmaxY PY subject to constraints C(+ WC)Other, more general ways to incorporate

soft constraints here [ACL’07]

observations

Page 14

A General Inference Setting

Essentially all complex models studied today can be viewed as optimizing a linear objective function: HMMs/CRFs [Roth’99; Collins’02;Laffarty et. al 02]

Linear objective functions can be derived from probabilistic perspective: Markov Random Field [standard; Kleinberg&Tardos] Optimization Problem (Metric

Labeling) [Chekuri et. al’01] Linear Programming Problems

Inference as Constrained Optimization [Yih&Roth CoNLL’04] [Punyakanok et. al

COLING’04]… The probabilistic perspective supports finding the most likely assignment

Not necessarily what we want Our Integer linear programming (ILP) formulation

Allows the incorporation of more general cost functions General (non-sequential) constraint structure Better exploitation (computationally) of hard constraints Can find the optimal solution if desired

Page 15

Formal Model

How to solve?

This is an Integer Linear Program

Solve using ILP packages gives an exact solution.

Search techniques are also possible

(Soft) constraints component

Weight Vector for “local” models

Penalty for violatingthe constraint.

How far away is y from a “legal” assignment

Subject to constraints

A collection of Classifiers; Log-linear models (HMM, CRF) or a combination

How to train?

Incorporating constraints in the learning process?

Page 16

Example: Semantic Role Labeling

I left my pearls to my daughter in my will .

[I]A0 left [my pearls]A1 [to my daughter]A2 [in my will]AM-LOC .

A0 Leaver A1 Things left A2 Benefactor AM-LOC Location I left my pearls to my daughter in my will .

Special Case (structure output problem): here, all the data is available at one time; in general, classifiers might be learned from different sources, at different times, at different contexts.

Implications on training paradigms

Overlapping arguments

If A2 is present, A1 must also be

present.

Who did what to whom, when, where, why,…

Page 17

PropBank [Palmer et. al. 05] provides a large human-annotated corpus of semantic verb-argument relations. It adds a layer of generic semantic labels to Penn Tree

Bank II. (Almost) all the labels are on the constituents of the

parse trees.

Core arguments: A0-A5 and AA different semantics for each verb specified in the PropBank Frame files

13 types of adjuncts labeled as AM-arg where arg specifies the adjunct type

Semantic Role Labeling (2/2)

Page 18

Algorithmic Approach

Identify argument candidates Pruning [Xue&Palmer, EMNLP’04] Argument Identifier

Binary classification (SNoW)

Classify argument candidates Argument Classifier

Multi-class classification (SNoW)

Inference Use the estimated probability

distribution given by the argument classifier

Use structural and linguistic constraints

Infer the optimal global output

I left my nice pearls to her

I left my nice pearls to her[ [ [ [ [ ] ] ] ] ]

I left my nice pearls to her[ [ [ [ [ ] ] ] ] ]

I left my nice pearls to her

I left my nice pearls to her

Identify Vocabulary

Inference over (old and new)

Vocabulary

candidate arguments

EASY

Page 19

InferenceI left my nice pearls to her

The output of the argument classifier often violates some constraints, especially when the sentence is long.

Finding the best legitimate output is formalized as an optimization problem and solved via Integer Linear Programming. [Punyakanok et. al 04, Roth & Yih 04;05]

Input: The probability estimation (by the argument classifier)

Structural and linguistic constraints

Allows incorporating expressive (non-sequential) constraints on the variables (the arguments types).

Page 20

Integer Linear Programming Inference

For each argument ai

Set up a Boolean variable: ai,t indicating whether ai is classified as t

Goal is to maximize i score(ai = t ) ai,t

Subject to the (linear) constraints Any Boolean constraints can be encoded as linear

constraint(s).

If score(ai = t ) = P(ai = t ), the objective is to find the assignment that maximizes the expected number of arguments that are correct and satisfies the constraints.

Page 21

No duplicate argument classes

a POTARG x{a = A0} 1

R-ARG

a2 POTARG , a POTARG x{a = A0} x{a2 = R-A0}

C-ARG a2 POTARG ,

(a POTARG) (a is before a2 ) x{a = A0} x{a2 = C-A0}

Many other possible constraints: Unique labels No overlapping or embedding Relations between number of arguments If verb is of type A, no argument of type B

Any Boolean rule can be encoded as a linear constraint.

If there is an R-ARG phrase, there is an ARG Phrase

If there is an C-ARG phrase, there is an ARG before it

Constraints

Joint inference can be used also to combine different SRL Systems.

Universally quantified rules

Page 22

ILP as a Unified Algorithmic Scheme

Consider a common model for sequential inference: HMM/CRF Inference in this model is done via the Viterbi Algorithm.

Viterbi is a special case of the Linear Programming based Inference. Viterbi is a shortest path problem, which is a LP, with a

canonical matrix that is totally unimodular. Therefore, you can get integrality constraints for free.

One can now incorporate non-sequential/expressive/declarative constraints by modifying this canonical matrix

The extension reduces to a polynomial scheme under some conditions (e.g., when constraints are sequential, when the solution space does not change, etc.)

Not necessarily increases complexity and very efficient in practice

[Roth&Yih, ICML’05]

y1 y2 y3 y4 y5y

x x1 x2 x3 x4 x5s

ABC

ABC

ABC

ABC

ABC

t

Learn a rather simple model; make decisions with a more expressive model

So far, shown the use of only (deterministic) constraints. Can be used with statistical constraints.

Page 23

Extracting Relations via Semantic AnalysisScreen shot from a CCG demo http://L2R.cs.uiuc.edu/~cogcomp

Semantic parsing reveals several relations in the sentence along with their arguments.

Top ranked system in CoNLL’05 shared task

Key difference is the Inference

This approach produces a very good semantic parser. F1~90%

Easy and fast: ~7 Sent/Sec (using Xpress-MP)

Page 24

This Talk

Global Inference over Local Models/Classifiers + Expressive Constraints

Model Generality of the framework

Training Paradigms

Global vs. Local training Semi-Supervised Learning

Examples Semantic Parsing Information Extraction Pipeline processes

Page 25

Given: Q: Who acquired Overture? Determine: A: Eyeing the huge market potential, currently led by Google, Yahoo took over search company Overture Services Inc last year.

Textual Entailment

Eyeing the huge market potential, currently led by Google, Yahoo took over search company Overture Services Inc. last year

Yahoo acquired Overture

Is it true that?(Textual Entailment)

Overture is a search company Google is a search company

……….

Google owns Overture

Phrasal verb paraphrasing [Connor&Roth’07]

Entity matching [Li et. al, AAAI’04, NAACL’04]

Semantic Role Labeling

Page 26

Training Paradigms that Support Global Inference

Incorporating general constraints (Algorithmic Approach)

Allow both statistical and expressive declarative constraints

Allow non-sequential constraints (generally difficult)

Coupling vs. Decoupling Training and Inference. Incorporating global constraints is important but Should it be done only at evaluation time or also at

training time? Issues related to:

modularity, efficiency and performance, availability of training data

May not be relevant in some problems.

Page 27

Input: o1 o2 o3 o4 o5 o6 o7 o8 o9 o10

Classifier 1:Classifier 2:

Infer:

Phrase Identification Problem

Use classifiers’ outcomes to identify phrases Final outcome determined by optimizing classifiers outcome

and constrains

[ [ [ []] ] ] ] ]

[ ] ][

Did this classifier make a

mistake? How to train it?

Page 28

Training in the presence of Constraints

General Training Paradigm: First Term: Learning from data Second Term: Guiding the model by constraints Can choose if constraints are included in training

or only in evaluation

Page 29

L+I: Learning plus Inference

IBT: Inference-based Training

Training w/o ConstraintsTesting: Inference with Constraints

x1

x6

x2

x5

x4

x3

x7

y1

y2

y5

y4

y3

f1(x)

f2(x)

f3(x)f4(x)

f5(x)

X

Y

Learning the components together!

Cartoon: each model can be

more complex and may have a view on a set of output

variables.

Page 30

-1 1 111Y’ Local Predictions

Perceptron-based Global Learning

x1

x6

x2

x5

x4

x3

x7

f1(x)

f2(x)

f3(x)f4(x)

f5(x)

X

Y

-1 1 1-1-1YTrue Global Labeling

-1 1 11-1Y’ Apply Constraints:

Which one is better? When and Why?

Page 31

Claims

When the local classification problems are “easy”, L+I outperforms IBT. In many applications, the components are identifiable

and easy to learn (e.g., argument, open-close, PER). Only when the local problems become difficult to solve in

isolation, IBT outperforms L+I, but needs a larger number of training examples. When data is scarce, problems are not easy and

constraints can be used, along with a “weak” model, to label unlabeled data and improve model.

Experimental results and theoretical intuition to support our claims.L+I: cheaper computationally; modular

IBT is better in the limit, and other extreme cases.

Combinations: L+I, and then IBT are possible

Page 32

opt=0.2opt=0.1opt=0

Bound Prediction

Local ≤ opt + ( ( d log m + log 1/ ) / m )1/2

Global ≤ 0 + ( ( cd log m + c2d + log 1/ ) / m )1/2

Bounds Simulated Data

L+I vs. IBT: the more identifiable individual problems are the better overall performance is with L+I

Indication for hardness of

problem

Page 33

Relative Merits: SRL

Difficulty of the learning problem

(# features)

L+I is better.

When the problem is artificially made harder, the tradeoff is clearer.

easyhard

In some cases problems are hard due to lack of training data.

Semi-supervised learning

Page 34

Prediction result of a trained HMM

Lars Ole Andersen . Program analysis and

specialization for the C Programming language

. PhD thesis .DIKU , University of Copenhagen , May1994 .

Information extraction with Background Knowledge (Constraints)

Lars Ole Andersen . Program analysis and specialization for the C Programming language. PhD thesis. DIKU , University of Copenhagen, May 1994 .

[AUTHOR] [TITLE] [EDITOR] [BOOKTITLE] [TECH-REPORT] [INSTITUTION] [DATE]

Violates lots of constraints!

Page 35

Examples of Constraints

Each field must be a consecutive list of words, and can appear at most once in a citation.

State transitions must occur on punctuation marks.

The citation can only start with AUTHOR or EDITOR.

The words pp., pages correspond to PAGE. Four digits starting with 20xx and 19xx are DATE. Quotations can appear only in TITLE …….

Page 36

Information Extraction with Constraints

Adding constraints, we get correct results!

[AUTHOR] Lars Ole Andersen . [TITLE] Program analysis and specialization

for the C Programming language .

[TECH-REPORT] PhD thesis .[INSTITUTION] DIKU , University of Copenhagen , [DATE] May, 1994 .

If incorporated into the training, better results mean Better Feedback!

Page 37

Semi-Supervised Learning with Constraints

Model

Decision Time Constraints

Un-labeled Data

Constraints

In traditional Semi-Supervised learning the model can drift away from the correct one.

Constraints can be used At decision time, to bias the objective function

towards favoring constraint satisfaction. At training to improve labeling of un-labled data (and

thus improve the model)

Page 38

=learn(Tr)

For N iterations do

T=

For each x in unlabeled dataset

y Inference(x, )

T=T {(x, y)}

Any supervised learning algorithm parametrized by

Augmenting the training set (feedback). Any inference algorithm (with constraints).

Inference(x,C, )

Constraint - Driven Learning (CODL) [Chang, Ratinov, Roth, ACL’07]

Page 39

Constraint - Driven Learning (CODL) [Chang, Ratinov, Roth, ACL’07]

=learn(Tr)

For N iterations do

T=

For each x in unlabeled dataset

{y1,…,yK} Top-K-Inference(x,C, )

T=T {(x, yi)}i=1…k

= +(1- )learn(T)Learn from new training data.Weight supervised and unsupervised model(Nigam2000*).

Augmenting the training set (feedback). Any inference algorithm (with constraints).

Any supervised learning algorithm parametrized by

Page 40

Token-based accuracy (inference with constraints)

60

65

70

75

80

85

90

95

100

5 10 15 20 25 300

Labeled set size

Ac

cu

rac

y

supervised Weighted EM CODL

Page 41

Objective function:

Semi-Supervised Learning with Constraints

# of available labeled examples

Learning w ConstraintsConstraints are used to Bootstrap a semi-supervised learner

Poor model + constraints are used to annotate unlabeled data, which in turn is used to keep training the model.

Learning w/o Constraints: 300 examples.

Page 42

Discussed a general paradigm for learning and inference in the context of natural language understanding tasks

A general Constraint Optimization approach for integration of learned models with additional (declarative or statistical) expressivity.

A paradigm for making Machine Learning practical – allow domain/task specific constraints.

How to train? Learn Locally and make use globally (via global

inference) [Punyakanok et. al IJCAI’05]

Ability to make use of domain & constraints to drive supervision

[Klementiev & Roth, ACL’06; Chang, Ratinov, Roth, ACL’07]

Conclusions

LBJ (Learning Based Java): http://L2R.cs.uiuc.edu/~cogcomp

A modeling language that supports programming along with building learned models and allows incorporating and

inference with constraints

Page 43

Random

Variables Y:

Conditional Distributions P (learned by models/classifiers) Constraints C– any Boolean function defined on partial assignments (possibly: + weights W ) Goal: Find the “best” assignment

The assignment that achieves the highest global accuracy.

This is an Integer Programming Problem

Problem Setting

y4 y5 y6 y7 y8

y1 y2 y3C(y1,y4)C(y2,y3,y6,y7,y8)

Y*=argmaxY PY subject to constraints C(+ WC)Other, more general ways to incorporate

soft constraints here [ACL’07]