1 Semi-Supervised Approaches for Learning to Parse Natural Languages Slides are from Rebecca Hwa, Ray Mooney

2 The Role of Parsing in Language Applications… As a stand-alone application –Grammar checking As a pre-processing step –Question Answering –Information extraction As an integral part of a model –Speech Recognition –Machine Translation

3 Parsing Parsers provide syntactic analyses of sentences VP saw PN her VB NP S PN I NP Input: I saw her

4 Challenges in Building Parsers Disambiguation –Lexical disambiguation –Structural disambiguation Rule Exceptions –Many lexical dependencies Manual Grammar Construction –Limited coverage –Difficult to maintain

5 Meeting these Challenges: Statistical Parsing Disambiguation? –Resolve local ambiguities with global likelihood Rule Exceptions? –Lexicalized representation Manual Grammar Construction? –Automatic induction from large corpora –A new challenge: how to obtain training corpora? –Make better use of unlabeled data with machine learning techniques and linguistic knowledge

6 Roadmap Parsing as a learning problem Semi-supervised approaches –Sample selection –Co-training –Corrected Co-training Conclusion and further directions

7 Parsing Ambiguities T1:T1:T2:T2: VP saw her N duck PNS VB NP S PN I PP with PNP a NDET NP telescope VP saw her N duck PNS VB NP S PN I PP with PNP a NDET NP telescope Input: “I saw her duck with a telescope”

8 Disambiguation with Statistical Parsing T1:T1:T2:T2: VP saw her N duck PNS VB NP S PN I PP with PNP a NDET NP telescope VP saw her N duck PNS VB NP S PN I PP with PNP a NDET NP telescope W = “I saw her duck with a telescope”

9 A Statistical Parsing Model Probabilistic Context-Free Grammar (PCFG) Associate probabilities with production rules Likelihood of the parse is computed from the rules used Learn rule probabilities from training data 0.3 NP PN 0.5 DET a 0.1 DET an the 0.4 DET Example of PCFG rules: NP DET N

10 Sentence Probability Assume productions for each node are chosen independently. Probability of derivation is the product of the probabilities of its productions. P(D 1 ) = 0.1 x 0.5 x 0.5 x 0.6 x 0.6 x 0.5 x 0.3 x 1.0 x 0.2 x 0.2 x 0.5 x 0.8 = D1D1 S VP Verb NP Det Nominal Nominal PP book Prep NP through Houston Proper-Noun the flight Noun

Syntactic Disambiguation Resolve ambiguity by picking most probable parse tree. 11 D2D2 VP Verb NP Det Nominal book Prep NP through Houston Proper-Noun the flight Noun S VP 0.1 PP 0.3 P(D 2 ) = 0.1 x 0.3 x 0.5 x 0.6 x 0.5 x 0.6 x 0.3 x 1.0 x 0.5 x 0.2 x 0.2 x 0.8 =

Sentence Probability Probability of a sentence is the sum of the probabilities of all of its derivations. 12 P(“book the flight through Houston”) = P(D 1 ) + P(D 2 ) = =

13 PCFG: Supervised Training If parse trees are provided for training sentences, a grammar and its parameters can be can all be estimated directly from counts accumulated from the tree-bank (with appropriate smoothing) Tree Bank Supervised PCFG Training S → NP VP S → VP NP → Det A N NP → NP PP NP → PropN A → ε A → Adj A PP → Prep NP VP → V NP VP → VP PP English S NP VP John V NP PP put the dog in the pen S NP VP John V NP PP put the dog in the pen

Estimating Production Probabilities Set of production rules can be taken directly from the set of rewrites in the treebank. Parameters can be directly estimated from frequency counts in the treebank. 14

15 Vanilla PCFG Limitations Since probabilities of productions do not rely on specific words or concepts, only general structural disambiguation is possible (e.g. prefer to attach PPs to Nominals). Consequently, vanilla PCFGs cannot resolve syntactic ambiguities that require semantics to resolve, e.g. ate with fork vs. meatballs. In order to work well, PCFGs must be lexicalized, i.e. productions must be specialized to specific words by including their head-word in their LHS non-terminals (e.g. VP-ate).

Example of Importance of Lexicalization A general preference for attaching PPs to NPs rather than VPs can be learned by a vanilla PCFG. But the desired preference can depend on specific words. 16 S → NP VP S → VP NP → Det A N NP → NP PP NP → PropN A → ε A → Adj A PP → Prep NP VP → V NP VP → VP PP English PCFG Parser S NP VP John V NP PP put the dog in the pen John put the dog in the pen.

17 Example of Importance of Lexicalization A general preference for attaching PPs to NPs rather than VPs can be learned by a vanilla PCFG. But the desired preference can depend on specific words. S → NP VP S → VP NP → Det A N NP → NP PP NP → PropN A → ε A → Adj A PP → Prep NP VP → V NP VP → VP PP English PCFG Parser S NP VP John V NP put the dog in the pen X John put the dog in the pen.

Head Words Syntactic phrases usually have a word in them that is most “central” to the phrase. Linguists have defined the concept of a lexical head of a phrase. Simple rules can identify the head of any phrase by percolating head words up the parse tree. –Head of a VP is the main verb –Head of an NP is the main noun –Head of a PP is the preposition –Head of a sentence is the head of its VP

Lexicalized Productions Specialized productions can be generated by including the head word and its POS of each non- terminal as part of that non-terminal’s symbol. S VP VBD NP DT Nominal Nominal PP liked IN NP in the dog NN DT Nominal NNthe pen NNP NP John pen-NN in-IN dog-NN liked-VBD John-NNP Nominal dog-NN → Nominal dog-NN PP in-IN

Lexicalized Productions S VP VP PP DT Nominal put IN NP in the dog NN DT Nominal NNthe pen NNP NP John pen-NN in-IN dog-NN put-VBD John-NNP NPVBD put-VBD VP put-VBD → VP put-VBD PP in-IN

Parameterizing Lexicalized Productions Accurately estimating parameters on such a large number of very specialized productions could require enormous amounts of treebank data. Need some way of estimating parameters for lexicalized productions that makes reasonable independence assumptions so that accurate probabilities for very specific rules can be learned.

Collins’ Parser Collins’ (1999) parser assumes a simple generative model of lexicalized productions. Models productions based on context to the left and the right of the head daughter. –LHS → L n L n  1 …L 1 H R 1 …R m  1 R m First generate the head (H) and then repeatedly generate left (L i ) and right (R i ) context symbols until the symbol STOP is generated.

Sample Production Generation VP put-VBD → VBD put-VBD NP dog-NN PP in-IN VP put-VBD →VBD put-VBD NP dog- NN HL1L1 STOPPP in-IN STOP R1R1 R2R2 R3R3 P L (STOP | VP put-VBD) * P H (VBD | VP put-VBD )* P R (NP dog-NN | VP put-VBD ) * P R (PP in-IN | VP put-VBD ) * P R (STOP | VPput-VBD)

Count(PPin-IN right of head in a VPput-VBD production ) Estimating Production Generation Parameters Estimate P H, P L, and P R parameters from treebank data. P R (PP in-IN | VP put-VBD ) = Count(symbol right of head in a VP put-VBD ) Count(NPdog-NN right of head in a VPput-VBD production ) P R (NP dog-NN | VP put-VBD ) = Smooth estimates by linearly interpolating with simpler models conditioned on just POS tag or no lexical info. smP R (PP in-IN | VP put-VBD ) = 1 P R (PP in-IN | VP put-VBD ) + (1  1 ) ( 2 P R (PP in-IN | VP VBD ) + (1  2 ) P R (PP in-IN | VP)) Count(symbol right of head in a VP put-VBD )

25 Parsing Evaluation Metrics PARSEVAL metrics measure the fraction of the constituents that match between the computed and human parse trees. If P is the system’s parse tree and T is the human parse tree (the “gold standard”): –Recall = (# correct constituents in P) / (# constituents in T) –Precision = (# correct constituents in P) / (# constituents in P) Labeled Precision and labeled recall require getting the non-terminal label on the constituent node correct to count as correct. F 1 is the harmonic mean of precision and recall.

26 Treebank Results Results of current state-of-the-art systems on the English Penn WSJ treebank are slightly greater than 90% labeled precision and recall.

27 Supervised Learning Avoids Manual Construction Training examples are pairs of problems and answers Training examples for parsing: a collection of sentence, parse tree pairs (Treebank) –From the treebank, get maximum likelihood estimates for the parsing model New challenge: treebanks are difficult to obtain –Needs human experts –Takes years to complete

28 Learning to ClassifyLearning to Parse Train a model to decide: should a prepositional phrase modify the verb before it or the noun? Train a model to decide: what is the most likely parse for a sentence W? Training examples: (v, saw, duck, with, telescope) (n, saw, duck, with, feathers) (v, saw, stars, with, telescope) (n, saw, stars, with, Oscars) … [S [NP-SBJ [NNP Ford] [NNP Motor] [NNP Co.]] [VP [VBD acquired] [NP [NP [CD 5] [NN %]] [PP [IN of] [NP [NP [DT the] [NNS shares]] [PP [IN in] [NP [NNP Jaguar] [NNP PLC]]]]]]]. ] [S [NP-SBJ [NNP Pierre] [NNP Vinken]] [VP [MD will] [VP [VB join] [NP [DT the] [NN board]] [PP [IN as] [NP [DT a] [NN director]]].] …

29 Hwa’s Approach Sample selection –Reduce the amount of training data by picking more useful examples Co-training –Improve parsing performance from unlabeled data Corrected Co-training –Combine ideas from both sample selection and co-training

30 Roadmap Parsing as a learning problem Semi-supervised approaches –Sample selection Overview Scoring functions Evaluation –Co-training –Corrected Co-training Conclusion and further directions

31 Sample Selection Assumption –Have lots of unlabeled data (cheap resource) –Have a human annotator (expensive resource) Iterative training session –Learner selects sentences to learn from –Annotator labels these sentences Goal: Predict the benefit of annotation –Learner selects sentences with the highest Training Utility Values (TUVs) –Key issue: scoring function to estimate TUV

32 Algorithm Initialize Train the parser on a small treebank (seed data) to get the initial parameter values. Repeat Create candidate set by randomly sample the unlabeled pool. Estimate the TUV of each sentence in the candidate set with a scoring function, f. Pick the n sentences with the highest score (according to f). Human labels these n sentences and add them to training set. Re-train the parser with the updated training set. Until (no more data).

33 Scoring Function Approximate the TUV of each sentence –True TUVs are not known Need relative ranking Ranking criteria –Knowledge about the domain e.g., sentence clusters, sentence length, … –Output of the hypothesis e.g., error-rate of the parse, uncertainty of the parse, … ….

34 Proposed Scoring Functions Using domain knowledge – long sentences tend to be complex Uncertainty about the output of the parser – tree entropy Minimize mistakes made by the parser – use an oracle scoring function find sentences with the most parsing inaccuracies f error f te f len

35 Entropy Measure of uncertainty in a distribution –Uniform distribution very uncertain –Spike distribution very certain Expected number of bits for encoding a probability distribution, X

36 Tree Entropy Scoring Function Distribution over parse trees for sentence W: Tree entropy: uncertainty of the parse distribution Scoring function: ratio of actual parse tree entropy to that of a uniform distribution

37 Oracle Scoring Function 1 - the accuracy rate of the most-likely parse Parse accuracy metric: f-score Precision = # of correctly labeled constituents # of constituents generated Recall = # of correctly labeled constituents # of constituents in correct answer f-score = harmonic mean of precision and recall f error

38 Experimental Setup Parsing model: –Collins Model 2 Candidate pool –WSJ sec 02-21, with the annotation stripped Initial labeled examples: 500 sentences Per iteration: add 100 sentences Testing metric: f-score (precision/recall) Test data: –~2000 unseen sentences (from WSJ sec 00) Baseline –Annotate data in sequential order

39 Training Examples Vs. Parsing Performance

40 Parsing Performance Vs. Constituents Labeled

41 Co-Training [ Blum and Mitchell, 1998 ] Assumptions –Have a small treebank –No further human assistance –Have two different kinds of parsers A subset of each parser’s output becomes new training data for the other Goal: –select sentences that are labeled with confidence by one parser but labeled with uncertainty by the other parser.

42 Algorithm Initialize Train two parsers on a small treebank (seed data) to get the initial models. Repeat Create candidate set by randomly sample the unlabeled pool. Each parser labels the candidate set and estimates the accuracy of its output with scoring function, f. Choose examples according to some selection method, S (using the scores from f). Add them to the parsers’ training sets. Re-train parsers with the updated training sets. Until (no more data).

43 Scoring Functions Evaluates the quality of each parser’s output Ideally, function measures accuracy –Oracle f F-score combined prec./rec. of the parse Practical scoring functions –Conditional probability f cprob Prob(parse | sentence) –Others (joint probability, entropy, etc.)

44 Selection Methods Above-n: S above-n –The score of the teacher’s parse is greater than n Difference: S diff-n –The score of the teacher’s parse is greater than that of the student’s parse by n Intersection: S int-n –The score of the teacher’s parse is one of its n% highest while the score of the student’s parse for the same sentence is one of the student’s n% lowest

45 Experimental Setup Co-training parsers: –Lexicalized Tree Adjoining Grammar parser [Sarkar, 2002] –Lexicalized Context Free Grammar parser [Collins, 1997] Seed data: 1000 parsed sentences from WSJ sec02 Unlabeled pool: rest of the WSJ sec02-21, stripped Consider 500 unlabeled sentences per iteration Development set: WSJ sec00 Test set: WSJ sec23 Results: graphs for the Collins parser

46 Selection Methods and Co-Training Two scoring functions: f F-score (oracle), f cprob Multiple view selection vs. one view selection –Three selection methods: S above-n, S diff-n, S int-n Maximizing utility vs. minimizing error –For f F-score, we vary n to control accuracy rate of the training data –Loose control More sentences (avg. F-score = 85%) –Tight control Fewer sentences (avg. F-score = 95%)

47 Co-Training using f F-score with Loose Control

48 Co-Training using f F-score with Tight Control

49 Co-Training using f cprob

50 Roadmap Parsing as a learning problem Semi-supervised approaches –Sample selection –Co-training –Corrected Co-training Conclusion and further directions

51 Corrected Co-Training Human reviews and corrects the machine outputs before they are added to the training set Can be seen as a variant of sample selection [cf. Muslea et al., 2000 ] Applied to Base NP detection [ Pierce & Cardie, 2001 ]

52 Algorithm Initialize: Train two parsers on a small treebank (seed data) to get the initial models. Repeat Create candidate set by randomly sample the unlabeled pool. Each parser labels the candidate set and estimates the accuracy of its output with scoring function, f. Choose examples according to some selection method, S (using the scores from f). Human reviews and corrects the chosen examples. Add them to the parsers’ training sets. Re-train parsers with the updated training sets. Until (no more data).

53 Selection Methods and Corrected Co-Training Two scoring functions: f F-score, f cprob Three selection methods: S above-n, S diff-n, S int-n

54 Corrected Co-Training using f F-score (Reviews)

55 Corrected Co-Training using f F-score (Corrections)

56 Corrected Co-Training using f cprob (Reviews)

57 Corrected Co-Training using f cprob (Corrections)