BIC

Learning and Inference in Natural Language From Stand Alone Learning Tasks to StructuredRepresentations: 

Learning and Inference in Natural Language From Stand Alone Learning Tasks to Structured Representations Dan Roth Department of Computer Science University of Illinois at Urbana-Champaign Joint work with my students: Vasin Punyakanok, Wen-tau Yih, Dav Zimak Biologically Inspired Computing Sendai, Japan, Nov. 2004

Slide2: 

We have concentrated on developing the theoretical basis within which to address some of the obstacles and on developing an experimental paradigm so that realistic experiments can be performed to validate the theoretical basis. The emphasis is on large-scale real-world problems in natural language understanding and visual recognition The group develops theories and systems pertaining to intelligent behavior using a unified methodology. At the heart of the approach is the idea that learning has a central role in intelligence. Cognitive Computation Group

Cognitive Computation Group: 

Cognitive Computation Group Foundations Learning Theory: Classification; Multi-Class Classification; Ranking Knowledge Representation: Relational Representations, Relational Kernels Inference approaches: structural mappings Intelligent Information Access Information Extraction Named Entities and Relations Matching Entities Mentions within and across documents and data bases Natural Language Processing Semantic Role Labeling Question answering Semantics Software Basic tools development: SNoW, FEX; shallow parser, pos tagger, semantic parser, … Some of our work on understanding the role of learning in supporting reasoning in the natural language domain

Comprehension: 

Comprehension (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. 1. Who is Christopher Robin? 2. When was Winnie the Pooh written? 3. What did Mr. Robin do when Chris was three years old? 4. Where did young Chris live? 5. Why did Chris write two books of his own?

What we Know: Stand Alone Ambiguity Resolution: 

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 Tiger was in Washington for the PGA Tour What we Know: Stand Alone Ambiguity Resolution

Comprehension: 

Comprehension (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. 1. Who is Christopher Robin? 2. When was Winnie the Pooh written? 3. What did Mr. Robin do when Chris was three years old? 4. Where did young Chris live? 5. Why did Chris write two books of his own?

Inference: 

Inference

Inference with Classifiers: 

Global decisions in which several local decisions play a role but there are mutual dependencies on their outcome. Learned classifiers for different sub-problems Incorporate classifiers’ information, along with constraints, in making coherent decisions – decisions that respect the local classifiers as well as domain & context specific constraints. Global inference for the best assignment to all variables of interest. Inference with Classifiers

Overview: 

Overview Stand Alone Learning Modeling Representational Issues. Computational Issues Inference Making Decisions under General Constraints Semantic Role Labeling How to train Components of Global Decisions Tradeoff that depends on easiness of learning components. Feedback to learning is (indirectly) given by the reasoning stage. There may not be a need (or even a possibility) to learn exactly, but only to the extent that is supports Reasoning.

Structured Input  Feature Mapping  Learning Structured Output: 

Structured Input  Feature Mapping  Learning Structured Output

Stand Alone Ambiguity Resolution: 

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 Tiger was in Washington for the PGA Tour Stand Alone Ambiguity Resolution

Disambiguation Problems: 

Disambiguation Problems Middle Eastern ____ are known for their sweetness Task: Decide which of { deserts , desserts } is more likely in the given context. Ambiguity: modeled as confusion sets (class labels C ) C={ deserts, desserts} C={ Noun,Adj.., Verb…} C={ topic=Finance, topic=Computing} C={ NE=Person, NE=location}

Learning to Disambiguate: 

Learning to Disambiguate Given a confusion set C={ deserts, desserts} sentence (s) Middle Eastern ____ are known for their sweetness Map into a feature based representation  : S  { 1(s), 2(s), …} Learn a function FC that determines which of C={ deserts, desserts} is more likely in a given context. FC (x)= w ¢  (x) Evaluate the function on future C sentences

Example: Representation: 

S= I don’t know whether to laugh or cry [x x x x] Consider words, pos tags, relative location in window Generate binary features representing presence of: a word/pos within window around target word don’t within +/-3 know within +/-3 Verb at -1 to within +/- 3 laugh within +/-3 to a +1 conjunctions of size 2, within window of size 3 words: know__to; ___to laugh pos+words: Verb__to; ____to Verb Example: Representation The sentence is represented as a set of its active features S= (don’t at -2 , know within +/-3,… ____to Verb,...) Hope: S=I don’t care whether to laugh or cry has almost the same representation

Structured Input: Features can be Complex: 

Structured Input: Features can be Complex [NP Which type] [PP of ] [NP submarine] [VP was bought ] [ADVPrecently ] [PP by ] [NP South Korea ] (. ?) S = John will join the board as a director Can be an involved process; builds on previous learners; computationally hard; some algorithms (perceptron) support implicit mapping

Notes on Representation: 

A feature is a function over sentences, which maps a sentence to a set of properties of the sentence.  : S  {0,1} or [0,1] There is a huge number of potential features (~105); Out of these – only a small number is actually active in each example. Representation: List only features that are active (non zero) in example When the number of features is fixed, the collection of examples is {1(s), 2(s), … n(s)} = {0,1}n. No need to fix number of features (on-line algorithms). infinite attribute domain { 1(s), 2(s), …} = {0,1}1 Some algorithms can make use of variable size input. Notes on Representation

Slide17: 

New discriminator in functionally simpler Embedding

Natural Language: Domain Characteristics: 

The number of potential features is very large The instance space is sparse Decisions depend on a small set of features (sparse) Want to learn from a number of examples that is small relative to the dimensionality Natural Language: Domain Characteristics

Algorithm Descriptions: 

Focus: Two families of on-line algorithms Examples x 2 {0,1}n; Hypothesis w 2 Rn; Prediction: sgn{w ¢ x - } Additive weight update algorithm (Perceptron, Rosenblatt, 1958. Variations exist) Multiplicative weight update algorithm (Winnow, Littlestone, 1988. Variations exist) Algorithm Descriptions

Generalization: 

Dominated by the sparseness of the function space Most features are irrelevant  advantage to multiplicative # of examples required by multiplicative algorithms depends mostly on # of relevant features Generalization bounds depend on ||w||. Lesser issue: Sparseness of features space Very few active features  advantage to additive. Generalization depend on ||x|| [Kivinen/Warmuth 95] Generalization

Slide21: 

Function: At least 10 out of fixed 100 variables are active Dimensionality is n Perceptron,SVMs n: Total # of Variables (Dimensionality) Winnow Mistakes bounds for 10 of 100 of n # of mistakes to convergence

Multiclass Classification in NLP: 

Multiclass Classification in NLP Name/Entity Recognition Label people, locations, and organizations in a sentence [PER Sam Houston],[born in] [LOC Virginia], [was a member of the] [ORG US Congress]. Decompose into sub-problems Sam Houston, born in Virginia...  (PER,LOC,ORG,?)  PER (1) Sam Houston, born in Virginia...  (PER,LOC,ORG,?)  None (0) Sam Houston, born in Virginia...  (PER,LOC,ORG,?)  LOC (2) Input : {0,1}d or Rd Output: {0,1,2,3,...,k}

Solving Multi-Class via Binary Learning: 

Solving Multi-Class via Binary Learning Decompose; use Winner-Take-All y = argmax wi ¢ x + ti wi, x  Rn , ti  R (Pairwise classification also possible) Key issue: how to train the binary classifiers wi Via Kessler Construction - comparative training - allows learning voroni diagrams. Equivalently: learn in nk-dimension 1-vs-all: not expressive enough

Detour – Basic Classifier: SNoW: 

Detour – Basic Classifier: SNoW A learning architecture that supports several linear update rules (Winnow, Perceptron, naïve Bayes) Allows regularization; pruning; many options True multi-class classification [Har-Peled, Roth, Zimak, NIPS 2003] Variable size examples; very good support for large scale domains like NLP both in terms of number of examples and number of features. Very efficient (1-2 order of magnitude faster than SVMs) Integrated with an expressive Feature EXtraction Language (FEX) [Dowload from: http://L2R.cs.uiuc.edu/~cogcomp ]

Summary: Stand Alone Classification: 

Summary: Stand Alone Classification Theory is well understood Generalization bounds Practical issues Essentially all work is done with linear representations Features: generated explicitly or implicitly (Kernels) Tradeoff here is relatively understood Success on a large number of large scale classification problems Key issues: Features How to decide what are good features How to compute/extract features (intermediate representations) Supervision: learning protocol

Overview: 

Overview Stand Alone Learning Modeling Representational Issues. Computational Issues Inference Making Decisions under General Constraints Semantic Role Labeling How to train Components of Global Decisions Tradeoff that depends on easiness of learning components. Feedback to learning is (indirectly) given by the reasoning stage. There may not be a need (or even a possibility) to learn exactly, but only to the extent that is supports Reasoning.

Comprehension: 

Comprehension (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. 1. Who is Christopher Robin? 2. When was Winnie the Pooh written? 3. What did Mr. Robin do when Chris was three years old? 4. Where did young Chris live? 5. Why did Chris write two books of his own?

Identifying Phrase Structure: 

Identifying Phrase Structure Classifiers Recognizing “The beginning of NP” Recognizing “The end of NP” (or: word based classifiers: BIO representation) Also for other kinds of phrases… Some Constraints Phrases do not overlap Order of phrases Length of phrases Use classifiers to infer a coherent set of phrases He reckons the current account deficit will narrow to only # 1.8 billion in September [NP He ] [VP reckons ] [NP the current account deficit ] [VP will narrow ] [PP to ] [NP only # 1.8 billion ] [PP in ] [NP September ]

Constrains Structure : 

Constrains Structure Sequential Constraints Three models for sequential inference with classifiers [Punyakanok&Roth NIPS’01,JMLR05] HMM; HMM with Classifiers Conditional Models Constraint Satisfaction Models (CSCL: more general constrains) Other models have been proposed that can deal with sequential structures. Conditional models (other classifiers); CRF, StructurePerceptron [later] Many Applications: Shallow Parsing, Named Entity; Biological Sequences General Constraints Structure An Integer/Linear Programming formulation [Roth&Yih ‘02,’03,’04] Allow for Dynamic Programming based Inference No Dynamic Programming.

Identifying Entities and Relations: 

Identifying Entities and Relations J.V. Oswald was murdered at JFK after his assassin, K. F. Johns… Identify: J.V. Oswald was murdered at JFK after his assassin, K. F. Johns… location Kill (X, Y) Identify named entities Identify relations between entities Exploit mutual dependencies between named entities and relation to yield a coherent global detection. Some knowledge (classifiers) may be known in advance Some constraints may be available only at decision time

Inference with Classifiers: 

Inference with Classifiers Scenario: Global decisions in which several local decisions / components play a role, but there are mutual dependencies on their outcome. Assume: Learned classifiers for different sub-problems Constraints on classifiers’ labels (known during training or only at evaluation time). Goal: Incorporate classifiers’ predictions, along with the constraints, in making coherent decisions – that respect the classifiers as well as domain/context specific constrains. Formally: Global inference for the best assignment to all variables of interest.

Setting: 

Setting Inference with classifiers is not a new idea. On sequential constraint structure: HMM, PMM [Punyakanok&Roth], CRF[Lafferty et al.], CSCL[Punyakanok&Roth] On general structure: Heuristic search Attempts to use Bayesian Networks [Roth&Yih’02] have problems The Proposed Integer linear programming (ILP) formulation General: works on non-sequential constraint structure Expressive: can represent many types of constraints Optimal: finds the optimal solution Fast: commercial packages are able to quickly solve very large problems (hundreds of variables and constraints)

Problem Setting (1/2): 

Problem Setting (1/2) Random Variables X: Conditional Distributions P (learned by classifiers) Constraints C– any Boolean function defined on partial assignments (possible weights W on constraints) Goal: Find the “best” assignment The assignment that achieves the highest global accuracy. This is an Integer Programming Problem X*=argmaxX PX subject to constraints C Everything is Linear

Integer Linear Programming : 

Integer Linear Programming A set of binary variables, x = (x1,…, xd) A cost vector p Rd, Cost matrices C1RdRt ; C2RdRr, t, r: # of (inequality, equality) constraints; d - # of variables. The ILP solution x* is the vector that maximizes the cost function, x* = argmax x {0,1}d px Subject to C2x> b1; and C1x = b2, where b1, b2Rd and x{0,1}d

Problem Setting (2/2): 

Problem Setting (2/2) Very general formalism; Connections to a large number of well studied optimisation problems and a variety of applications. Justification: direct argument for the appropriate “best assignment” Relations to Markov Random Fields (but better computationally) Significant modelling and computational advantages

Semantic Role Labeling: 

Semantic Role Labeling For each verb in a sentence identify all constituents that fill a semantic role determine their roles Agent, Patient or Instrument, … Their adjuncts, e.g., Locative, Temporal or Manner PropBank project [Kingsbury & Palmer02] provides a large human-annotated corpus of semantic verb-argument relations. Experiment: CoNLL-2004 shared task [Carreras & Marquez 04] No parsed data in the input

Example: 

Example A0 represents the leaver, A1 represents the thing left, A2 represents the benefactor, AM-LOC is an adjunct indicating the location of the action, V determines the verb.

Argument Types: 

Argument Types A0-A5 and AA have different semantics for each verb as specified in the PropBank Frame files. 13 types of adjuncts labeled as AM-XXX where ARG specifies the adjunct type. C-ARG is used to specify the continuity of the argument ARG. In some cases, the actual agent is labeled as the appropriate argument type, ARG, while the relative pronoun is instead labeled as R-ARG.

Examples: 

Examples C-ARG R-ARG

Algorithm: 

Algorithm I. Find potential argument candidates (Filtering) II. Classify arguments to types III. Inference for Argument Structure Cost Function Constraints Integer linear programming (ILP)

I. Find Potential Arguments: 

I. Find Potential Arguments An argument can be any set of consecutive words Restrict potential arguments Classify BEGIN(word) BEGIN(word) = 1  “word begins argument” Classify END(word) END(word) = 1  “word ends argument” Argument (wi,...,wj) is a potential argument iff BEGIN(wi) = 1 and END(wj) = 1 Reduce set of potential arguments (PotArg)

II. Arguments Type Likelihood: 

II. Arguments Type Likelihood Assign type-likelihood How likely is it that arg a is type t? For all a  POTARG , t  T P (argument a = type t ) 0.3 0.2 0.2 0.3 0.6 0.0 0.0 0.4 A0 C-A1 A1 Ø

Details – Phrase-level Classifier: 

Details – Phrase-level Classifier Learn a classifier (SNoW) ARGTYPE(arg) P(arg)  {A0,A1,...,C-A0,...,AM-LOC,...} argmaxt{A0,A1,...,C-A0,...,LOC,...} wt P(arg) Estimate Probabilities Softmax over SNoW activations P(a = t) = exp(wt P(a)) / Z

What is a Good Assignment?: 

What is a Good Assignment? Likelihood of being correct P(Arg a = Type t) if t is the correct type for argument a For a set of arguments a1, a2, ..., an Expected number of arguments that are correct  i P( ai = ti ) The solution is the assignment with the maximum expected number of correct arguments.

Inference: 

Inference Maximize expected number correct T* = argmaxT  i P( ai = ti ) Subject to some constraints Structural and Linguistic (R-A1A1) I left my nice pearls to her I left my nice pearls to her

LP Formulation – Linear Cost: 

LP Formulation – Linear Cost Cost function a  POTARG P(a=t) = a  POTARG , t  T P(a=t) x{a=t} Indicator variables x{a1=A0}, x{a1= A1}, …, x{a4= AM-LOC}, x{P4=}  {0,1} Total Cost = p(a1= A0)· x(a1= A1) + p(a1= )· x(a1= ) +… + p(a4= )· x(a4= ) Corresponds to Maximizing expected number of correct phrases

Linear Constraints (1/2): 

Binary values  a  POTARG , t  T , x{a = t}  {0,1} Unique labels  a  POTARG ,  t  T x{a = t} = 1 No overlapping or embedding a1 and a2 overlap  x{a1=Ø} + x{a2=Ø}  1 Linear Constraints (1/2)

Linear Constraints (2/2): 

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: Exactly one argument of type Z If verb is of type A, no argument of type B Linear Constraints (2/2) Any Boolean rule can be encoded as a linear constraint. If the is an R-ARG phrase, there is an ARG Phrase If the is an C-ARG phrase, there is an ARG before it

Discussion: 

Discussion Inference approach used also for simultaneous named entities and relation identification (CoNLL’04) A few other problems in progress Global inference helps ! All constraints vs. only non-overlapping constraints: error reduction > 5% ; > 1% absolute F1 A lot of room for improvement (additional constraints) Easy and fast: 70-80 Sentences/Second (using Xpress-MP) Modeling and Implementation details: http://L2R.cs.uiuc.edu/~cogcomp http://www.scottyih.org/html/publications.html#ILP

Overview: 

Overview Stand Alone Learning Modeling Representational Issues. Computational Issues Inference Semantic Role Labeling Making Decisions under General Constraints How to train Components of Global Decisions Tradeoff that depends on easiness of learning components. Feedback to learning is (indirectly) given by the reasoning stage. There may not be a need (or even a possibility) to learn exactly, but only to the extent that is supports Reasoning.

Phrase Identification Problem: 

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 Output: s1 s2 s3 s4 s5 s6 s7 s8 s9 s10 Did this classifier make a mistake? How to train it?

Learning Structured Output: 

Learning Structured Output Input variables, x = (x1,…, xd) 2 X; Output variables y = (y1,…, yd) 2 Y A set of constrains C(Y) µ Y A cost function f(x,y) that assigns a score to each possible output. The cost function is linear in the components of y = (y1,…, yd): f(x, (y1,…, yd) ) = i fi(x, y) Each scoring function (classifier) is linear over some feature space fi(x,y) = wi  (x,y) Therefore the overall cost function is linear We seek a solution y* that maximizes the cost function, Subject to the constrain s C(Y) y* = argmax C(y) i  (x,y)

Learning and Inference Structured Output: 

Learning and Inference Structured Output Inference is the task of determining an optimal assignment y given an assignment x. For sequential structure of constraints, polynomial-time algorithms such as Viterbi or CSCL [Punyakanok&Roth, NIPS’01] can be used. For general structure of constraints, we proposed a formalism that uses Integer Linear Programming (ILP). Irrespective of the inference chosen, there are several ways to learn the scoring function parameters. These differ in whether or not the structure-based inference process is leveraged during training. Learning Local Classifiers: Decouple Learning from Inference. Learning Global Classifiers: Interleave inference with learning.

Learning Local and Global Classifiers: 

Learning Local and Global Classifiers Learning Local Classifiers: No knowledge of the inference used during learning. For each example (x, y) ∈ D, the learning algorithm must ensure that each component of y produces the correct output. Global constraint are enforced only at evaluation time. Learning Global Classifiers: Train to produce correct global output. Feedback from the inference process determines which classifiers to provide feedback to; together, the classifiers and the inference yield the desired result. At each step a subset of the classifiers are updated according to inference feedback. Conceptually similar to CRF and Collin’s Perceptron; we provide an online algorithm with a more general inference procedure.,

Relative Merits: 

Relative Merits Learning Local Classifiers = L+I Learning Global Classifier = Inference based Training (IBT) Claim: With a fixed number of examples: 1. If the local classification tasks are separable, then L+I outperforms IBT. 2. If the task is globally separable, but not locally separable then IBT outperforms L+I only with sufficient examples. This number correlates with the degree of the separability of the local classifiers. (The more strict the constrains are, the larger IBT’s example is)

Relative Merits: 

Relative Merits Learning Local Classifiers = L+I Learning Global Classifier = Inference based Training (IBT) LO: learning a stand alone component Results on the SRL task: In order to get to the region In which IBT is good, we reduced the number of features used by the individuals classifiers

Relative Merits (2): 

Relative Merits (2) Simulation results in which we compare different learning strategies in various degrees of difficulties of the local classifiers. κ = 0 implies locally linearly separability. Higher κ indicates harder local classification. K=0.15 K=1 K=0

Summary: 

Summary Stand alone learning Learning itself is well understood Learning & Inference with General Global Constraints Many problems in NLP involve several interdependent components and requires applying inference to obtain the best global solution Need to incorporate linguistics and other constraints into NLP reasoning A general paradigm for inference over learned components, based on (ILP) over general learners and expressive constraints. Preliminary understanding of relative merits of different training approaches. In practical situations, decoupling training from inference is advantageous. Features: how to map to a high dimensional space Learning protocol (weaker forms of supervision) What are the components? What else do we do wrong: Developmental Issues

Questions?: 

Questions? Thank you