Page 1 April 2010 Carnegie Mellon University With thanks to: Collaborators: Ming-Wei Chang, James Clarke, Dan Goldwasser, Lev Ratinov, Vivek Srikumar, Many others Funding: NSF: ITR IIS , SoD-HCER , DHS; DARPA: Bootstrap Learning & Machine Reading Programs DASH Optimization (Xpress-MP) Constraints Driven Structured Learning with Indirect Supervision Dan Roth Department of Computer Science University of Illinois at Urbana-Champaign

Page 2 Nice to Meet You

Page 3 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 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 4 Coherency in Semantic Role Labeling Predicate-arguments generated should be consistent across phenomena The touchdown scored by Bettis cemented the victory of the Steelers. VerbNominalizationPreposition Predicate: score A0: Bettis (scorer) A1: The touchdown (points scored) Predicate: win A0: the Steelers (winner) Sense: 11(6) “the object of the preposition is the object of the underlying verb of the nominalization” Linguistic Constraints: A0: the Steelers  Sense(of): 11(6) A0:Bettis  Sense(by): 1(1)

Page 5 Semantic Parsing Successful interpretation involves multiple decisions  What entities appear in the interpretation?  “New York” refers to a state or a city?  How to compose fragments together? state(next_to()) >< next_to(state()) X :“What is the largest state that borders New York and Maryland ?" Y: largest( state( next_to( state(NY) AND next_to (state(MD))))

Page 6 Learning and Inference Natural Language Decisions are Structured  Global decisions in which several local decisions play a role but there are mutual dependencies on their outcome. It is essential to make coherent decisions in a way that takes the interdependencies into account. Joint, Global Inference. But: Learning structured models requires annotating structures.

Page 7 Constrained Conditional Models (aka ILP Inference) (Soft) constraints component Weight Vector for “local” models Penalty for violating the constraint. How far y is from a “legal” assignment Features, classifiers; log- linear models (HMM, CRF) or a combination CCMs can be viewed as a general interface to easily combine domain knowledge with data driven statistical models How to solve? This is an Integer Linear Program Solving using ILP packages gives an exact solution. Search techniques are also possible How to train? Training is learning the objective function How to exploit the structure to minimize supervision?

Page 8 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. A0Leaver A1Things left A2Benefactor AM-LOCLocation I left my pearls to my daughter in my will. Overlapping arguments If A2 is present, A1 must also be present. Who did what to whom, when, where, why,…

Page 9 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 10 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 Identify Vocabulary Inference over (old and new) Vocabulary candidate arguments

Page 11 Semantic Role Labeling (SRL) I left my pearls to my daughter in my will

Page 12 Semantic Role Labeling (SRL) I left my pearls to my daughter in my will

Page 13 Semantic Role Labeling (SRL) I left my pearls to my daughter in my will One inference problem for each verb predicate.

Page 14 Integer Linear Programming Inference For each argument a i  Set up a Boolean variable: a i,t indicating whether a i is classified as t Goal is to maximize   i score(a i = t ) a i,t  Subject to the (linear) constraints If score(a i = t ) = P(a i = t ), the objective is to find the assignment that maximizes the expected number of arguments that are correct and satisfies the constraints. The Constrained Conditional Model is completely decomposed during training

Page 15 No duplicate argument classes  a  P OT A RG x { a = A0 }  1 R-ARG  a2  P OT A RG,  a  P OT A RG x { a = A0 }  x { a2 = R-A0 } C-ARG  a2  P OT A RG,  (a  P OT A RG )  (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; order constraints If verb is of type A, no argument of type B Any Boolean rule can be encoded as a (collection of) linear constraints. 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 LBJ: allows a developer to encode constraints in FOL; these are compiled into linear inequalities automatically.

Page 16 Learning and Inference Natural Language Decisions are Structured  Global decisions in which several local decisions play a role but there are mutual dependencies on their outcome. It is essential to make coherent decisions in a way that takes the interdependencies into account. Joint, Global Inference. But: Learning structured models requires annotating structures.

Page 17 Information extraction without Prior Knowledge Prediction result of a trained HMM Lars Ole Andersen. Program analysis and specialization for the C Programming language. PhD thesis. DIKU, University of Copenhagen, May [AUTHOR] [TITLE] [EDITOR] [BOOKTITLE] [TECH-REPORT] [INSTITUTION] [DATE] Violates lots of natural constraints! Lars Ole Andersen. Program analysis and specialization for the C Programming language. PhD thesis. DIKU, University of Copenhagen, May 1994.

Page 18 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 ……. Easy to express pieces of “knowledge” Non Propositional; May use Quantifiers

Page 19 Adding constraints, we get correct results!  Without changing the model [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, Information Extraction with Constraints

Page 20 Guiding 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 to generate better training data  At training to improve labeling of un-labeled data (and thus improve the model)  At decision time, to bias the objective function towards favoring constraint satisfaction.

Page 21 Constraints Driven Learning (CoDL) (w 0, ½ 0 )= learn(L)‏ For N iterations do T=  For each x in unlabeled dataset h à argmax y w T Á (x,y) -  ½ k d C (x,y) T=T  {(x, h)} (w, ½ ) =  (w 0, ½ 0 ) + (1-  ) learn(T) [Chang, Ratinov, Roth, ACL’07;ICML’08,Long’10] Supervised learning algorithm parameterized by (w, ½ ). Learning can be justified as an optimization procedure for an objective function Inference with constraints: augment the training set Learn from new training data Weigh supervised & unsupervised models. Excellent Experimental Results showing the advantages of using constraints, especially with small amounts on labeled data [Chang et. al, Others]

Page 22 Objective function: Constraints Driven Learning (CODL) [Chang et. al 07,08; others] # of available labeled examples Learning w 10 Constraints Poor model + constraints Constraints are used to:  Bootstrap a semi-supervised learner  Correct weak models predictions on unlabeled data, which in turn are used to keep training the model. Learning w/o Constraints: 300 examples. Semi-Supervised Learning Paradigm that makes use of constraints to bootstrap from a small number of examples

Page 23 Learning and Inference Natural Language Decisions are Structured  Global decisions in which several local decisions play a role but there are mutual dependencies on their outcome. It is essential to make coherent decisions in a way that takes the interdependencies into account. Joint, Global Inference. But: Learning structured models requires annotating structures. Interdependencies among decision variables should be exploited in learning.  Goal: use minimal, indirect supervision  Amplify it using interdependencies among variables

Page 24 Two Ideas Idea1: Simple, easy to supervise, binary decisions often depend on the structure you care about. Learning to do well on the binary task can drive the structure learning. Idea2: Global Inference can be used to amplify the minimal supervision.  Idea 2 ½: There are several settings where a binary label can be used to replace a structured label. Perhaps the most intriguing is where you use the world response to the model’s actions.

Page 25 Outline Inference Semi-supervised Training Paradigms for structures  Constraints Driven Learning Indirect Supervision Training Paradigms for structure  Indirect Supervision Training with latent structure [NAACL’10] Transliteration; Textual Entailment; Paraphrasing  Training Structure Predictors by Inventing (easy to supervise) binary labels [ICML’10] POS, Information extraction tasks  Driving supervision signal from World’s Response [CoNLL’10] Semantic Parsing

Page 26 Former military specialist Carpenter took the helm at FictitiousCom Inc. after five years as press official at the United States embassy in the United Kingdom. Jim Carpenter worked for the US Government. Textual Entailment x1x1 x6x6 x2x2 x5x5 x4x4 x3x3 x7x7 x1x1 x2x2 x4x4 x3x3  Entailment Requires an Intermediate Representation  Alignment based Features  Given the intermediate features – learn a decision  Entail/ Does not Entail But only positive entailments are expected to have a meaningful intermediate representation

Page 27 Paraphrase Identification Consider the following sentences: S1: Druce will face murder charges, Conte said. S2: Conte said Druce will be charged with murder. Are S1 and S2 a paraphrase of each other? There is a need for an intermediate representation to justify this decision Given an input x 2 X Learn a model f : X ! {-1, 1} We need latent variables that explain: why this is a positive example. Given an input x 2 X Learn a model f : X ! H ! {-1, 1}

Page 28 Algorithms: Two Conceptual Approaches Two stage approach (typically used for TE and paraphrase identification)  Learn hidden variables; fix it Need supervision for the hidden layer (or heuristics)  For each example, extract features over x and (the fixed) h.  Learn a binary classier Proposed Approach: Joint Learning  Drive the learning of h from the binary labels  Find the best h(x)  An intermediate structure representation is good to the extent is supports better final prediction.  Algorithm?

Page 29 Learning with Constrained Latent Representation (LCLR): Intuition If x is positive  There must exist a good explanation (intermediate representation)  9 h, w T Á (x,h) ¸ 0  or, max h w T Á (x,h) ¸ 0 If x is negative  No explanation is good enough to support the answer  8 h, w T Á (x,h) · 0  or, max h w T Á (x,h) · 0 Altogether, this can be combined into an objective function: Min w ¸ /2 ||w|| 2 + C  i l(1-z i max h 2 C w T  {s} h s Á s (x i )) Why does inference help? New feature vector for the final decision. Chosen h selects a representation. Inference: best h subject to constraints C

Page 30 Optimization Non Convex, due to the maximization term inside the global minimization problem In each iteration:  Find the best feature representation h* for all positive examples (off- the shelf ILP solver)  Having fixed the representation for the positive examples, update w solving the convex optimization problem:  Not the standard SVM/LR: need inference Asymmetry: Only positive examples require a good intermediate representation that justify the positive label.  Consequently, the objective function decreases monotonically

Page 31  Formalized as Structured SVM + Constrained Hidden Structure  LCRL: Learning Constrained Latent Representation Iterative Objective Function Learning Inference for h subj. to C Prediction with inferred h Training w/r to binary decision label Initial Objective Function Generate features Update weight vector

Page 32 Learning with Constrained Latent Representation (LCLR): Framework LCLR provides a general inference formulation that allows that use of expressive constraints  Flexibly adapted for many tasks that require latent representations. Paraphrasing: Model input as graphs, V(G 1,2 ), E(G 1,2 )  Four Hidden variables: h v1,v2 – possible vertex mappings; h e1,e2 – possible edge mappings  Constraints: Each vertex in G 1 can be mapped to a single vertex in G 2 or to null Each edge in G 1 can be mapped to a single edge in G 2 or to null Edge mapping is active iff the corresponding node mappings are active LCLR ModelDeclarative model

Page 33 Experimental Results Transliteration: Recognizing Textual Entailment: Paraphrase Identification:*

Page 34 Outline Inference Semi-supervised Training Paradigms for structures  Constraints Driven Learning Indirect Supervision Training Paradigms for structure  Indirect Supervision Training with latent structure Transliteration; Textual Entailment; Paraphrasing  Training Structure Predictors by Inventing (easy to supervise) binary labels POS, Information extraction tasks  Driving supervision signal from World’s Response Semantic Parsing

Page 35 Structured Prediction Before, the structure was in the intermediate level  We cared about the structured representation only to the extent it helped the final binary decision  The binary decision variable was given as supervision What if we care about the structure?  Information Extraction; Relation Extraction; POS tagging, many others. Invent a companion binary decision problem!  Parse Citations: Lars Ole Andersen. Program analysis and specialization for the C Programming language. PhD thesis. DIKU, University of Copenhagen, May  Companion: Given a citation; does it have a legitimate parse?  POS Tagging  Companion: Given a word sequence, does it have a legitimate POS tagging sequence?

Page 36 Predicting phonetic alignment (For Transliteration) Target Task  Input: an English Named Entity and its Hebrew Transliteration  Output: Phonetic Alignment (character sequence mapping)  A structured output prediction task (many constraints), hard to label Companion Task  Input: an English Named Entity and an Hebrew Named Entity  Companion Output: Do they form a transliteration pair?  A binary output problem, easy to label  Negative Examples are FREE, given positive examples ylatI י ט לי הא Target Task Yes/No Why it is a companion task? Companion Task נ ל יי וא י I l l i n o i s

Page 37 Companion Task Binary Label as Indirect Supervision The two tasks are related just like the binary and structured tasks discussed earlier All positive examples must have a good structure Negative examples cannot have a good structure We are in the same setting as before  Binary labeled examples are easier to obtain  We can take advantage of this to help learning a structured model Here: combine binary learning and structured learning Positive transliteration pairs must have “good” phonetic alignments Negative transliteration pairs cannot have “good” phonetic alignments

Page 38 Learning Structure with Indirect Supervision In this case we care about the predicted structure Use both Structural learning and Binary learning The feasible structures of an example Correct Predicted Negative examples cannot have a good structure Negative examples restrict the space of hyperplanes supporting the decisions for x

Page 39 Joint Learning with Indirect Supervision (J-LIS) Joint learning : If available, make use of both supervision types ylatI י ט לי הא Target Task Yes/No Loss on Target TaskLoss on Companion Task Loss function: L B, as before; L S, Structural learning Key: the same parameter w for both components Companion Task נ ל יי וא י I l l i n o i s

Page 40 Experimental Result Very little direct (structured) supervision. (Almost free) Large amount binary indirect supervision

Page 41 Experimental Result Very little direct (structured) supervision. (Almost free) Large amount binary indirect supervision

Page 42 Relations to Other Frameworks B= Á, l=(squared) hinge loss: Structural SVM S= Á, LCLR  Related to Structural Latent SVM (Yu & Johachims) and to Felzenszwalb. If S= Á, Conceptually related to Contrastive Estimation  No “grouping” of good examples and bad neighbors  Max vs. Sum: we do not marginalize over the hidden structure space Allows for complex domain specific constraints Related to some Semi-Supervised approaches, but can use negative examples (Sheffer et. al)

Page 43 Outline Inference Semi-supervised Training Paradigms for structures  Constraints Driven Learning Indirect Supervision Training Paradigms for structure  Indirect Supervision Training with latent structure Transliteration; Textual Entailment; Paraphrasing  Training Structure Predictors by Inventing (easy to supervise) binary labels POS, Information extraction tasks  Driving supervision signal from World’s Response Semantic Parsing

Page 44 Connecting Language to the World Can I get a coffee with no sugar and just a bit of milk Can we rely on this interaction to provide supervision? MAKE(COFFEE,SUGAR=NO,MILK=LITTLE) Arggg Great! Semantic Parser

Page 45 Traditional approach: learn from logical forms and gold alignments EXPENSIVE! Semantic parsing is a structured prediction problem: identify mappings from text to a meaning representation Query Response: Supervision = Expected Response Check if Predicted response == Expected response Logical Query Real World Feedback Interactive Computer System Pennsylvania Query Response : r largest( state( next_to( const(NY)))) y “What is the largest state that borders NY?" NL Query x Train a structured predictor with this binary supervision ! Expected : Pennsylvania Predicted : NYC Negative Response Pennsylvania r Binary Supervision Expected : Pennsylvania Predicted : Pennsylvania Positive Response Our approach: use only the responses

Page 46 Learning Structures with a Binary Signal Protocol I: Direct learning with binary supervision  Uses predicted structures as examples for learning a binary decision Inference used to predict the query: (y,z) = argmax y,z w T Á (x,y,z) Positive feedback: add a positive binary example Negative feedback: add a negative binary example  Learned parameters form the objective function; iterate largest(state(next_to(const(NY)))) “What is the largest state that borders NY?" largest(state(next_to(const(NJ)))) “What is the largest state that borders NY?" - + state(next_to(const(NY)))) “What is the smallest state?" b + -

Page 47 Protocol II: Aggressive Learning with Binary supervision  Positive feedback IFF the structure is correct  (y,z) = argmax y,z w T Á (x,y,z) Train a structured predictor from these structures  Positive feedback: add a positive structured example  Iterate until no new structures are found Learning Structures with a Binary Signal Interactive Computer System Pennsylvania Correct Response! “What is the largest state that borders NY?" largest( state( next_to( const(NY)))) +

Page 48 Empirical Evaluation Key Question: Can we learn from this type of supervision? Algorithm# training structures Test set accuracy No Learning: Initial Objective Fn Binary signal: Protocol I % 69.2 % Binary signal: Protocol II073.2 % WM*2007 (fully supervised – uses gold structures) % *[WM] Y.-W. Wong and R. Mooney Learning synchronous grammars for semantic parsing with lambda calculus. ACL.

Page 49 Summary Constrained Conditional Model: Computation Framework for global interference and an vehicle for incorporating knowledge Direct supervision for structured NLP tasks is expensive  Indirect supervision is cheap and easy to obtain We suggested learning protocols for Indirect Supervision  Make use of simple, easy to get, binary supervision  Showed how to use it to learn structure  Done in the context of Constrained Conditional Models Inference is an essential part of propagating the simple supervision Learning Structures from Real World Feedback  Obtain binary supervision from “real world” interaction  Indirect supervision replaces direct supervision

Page 50 Features Versus Constraints Á i : X £ Y ! R; C i : X £ Y ! {0,1}; d : X £ Y ! R;  In principle, constraints and features can encode the same properties  In practice, they are very different Features  Local, short distance properties – to support tractable inference  Propositional (grounded):  E.g. True if “the followed by a Noun occurs in the sentence” Constraints  Global properties  Quantified, first order logic expressions  E.g.True iff “all y i s in the sequence y are assigned different values.” If Á (x,y) = Á (x) – constraints provide an easy way to introduce dependence on y Mathematically, soft constraints are features

Page 51 Constraints As a Way To Encode Prior Knowledge Consider encoding the knowledge that:  Entities of type A and B cannot occur simultaneously in a sentence The “Feature” Way  Requires larger models The Constraints Way  Keeps the model simple; add expressive constraints directly  A small set of constraints  Allows for decision time incorporation of constraints A effective way to inject knowledge Need more training data We can use constraints as a way to replace training data Allows one to learn simpler models