Learning Transfer Rules for Machine Translation with Limited Data Thesis Defense Katharina Probst Committee: Alon Lavie (Chair) Jaime Carbonell Lori Levin Bonnie Dorr, University of Maryland Introduction (I) • Why has Machine Translation been applied only to few language pairs? – Bilingual corpora available only for few language pairs (English-French, Japanese-English, etc.) – Natural Language Processing tools available only for few language (English, German, Spanish, Japanese, etc.) – Scaling to other languages often difficult, timeconsuming, and knowledge-intensive • What can we do to change this? 2 Introduction (II) • This thesis presents a framework for automatic inference of transfer rules • Transfer rules capture syntactic and morphological mappings between languages • Learned from small, word-aligned training corpus • Rules are learned for unbalanced language pairs, where more data and tools are available for one language (L1) than for the other (L2) 3 Training Data Example Setting the Stage Rule Learning Experimental Results Conclusions SL: the widespread interest NP in the election [the interest the widespread in the election] DET ADJ N PP TL: h &niin h rxb b h bxirwt Alignment:((1,1),(1,3),(2,4),(3,2),(4,5 the widespread ),(5,6),(6,7)) interest Type: NP Parse: (<NP> (DET the-1) PREP NP (ADJ widespread-2) (N interest-3) (<PP> (PREP in-4) (<NP> (DET the-5) in DET N (N election-6)))) the election 4 Transfer Rule Formalism ;;L2: h &niin h rxb b h bxirwt ;;L1: the widespread interest in the election NP::NP [“h” N “h” Adj PP] -> [“the” Adj N PP] ((X1::Y1)(X2::Y3) (X3::Y1)(X4::Y2) (X5::Y4) ((Y3 num) = (X2 num)) ((X2 num) = sg) ((X2 gen) = m)) Training example Rule type Component sequences Component alignments Agreement constraints Value constraints 5 Research Goals (I) 1. Develop a framework for learning transfer rules from bilingual data • Training corpus: set of sentences/phrases in one language with translation into other language (= bilingual corpus), word-aligned • Rules include a) a context-free backbone and b) unification constraints 2. Improve of the grammaticality of MT output by automatically learned rules • Learned rules improve translation quality in run-time system 6 Research Goals (II) 3. Learn rules in the absence of a parser for one of the languages • Infer syntactic knowledge about minor language using a) projection from major language, b) analysis of word alignments, c) morphology information, and d) bilingual dictionary 4. Combine a set of different knowledge sources in a meaningful way • Resources (parser, morphology modules, dictionary, etc.) often disagree • Combine conflicting knowledge sources 7 Research Goals (III) Address limited-data scenarios with `frugal‘ techniques • “Unbalanced” language pairs with little or no bilingual data • Training corpus is small (~120 sentences and phrases), but carefully designed 6. Pushing MT research in the direction of incorporating syntax into statistical-based systems • Infer highly involved linguistic information, incorporate with statistical decoder in hybrid system 5. 8 Thesis Statement (I) • Given bilingual, word-aligned data, and given a parser for one of the languages in the translation pair, we can learn a set of syntactic transfer rules for MT. • The rules consist of a context-free backbone and unification constraints, learned in two separate stages. • The resulting rules form a syntactic translation grammar for the language pair and are used in a statistical transfer system to translate unseen examples. 9 Thesis Statement (II) • The translation quality of a run-time system that uses the learned rules is – superior to a system that does not use the learned rules – comparable to the performance using a small manual grammar written by an expert on Hebrew->English and Hindi->English translation tasks. • The thesis presents a new approach to learning transfer rules for Machine Translation in that the system learns syntactic models from text in a novel way and in a rich hypothesis space, aiming at emulating a human grammar writer. 10 Talk Overview • • • • Setting the Stage: related work, system overview, training data Rule Learning – Step 1: Seed Generation – Step 2: Compositionality – Step 3: Unification Constraints Experimental Results Conclusion 11 Depth of Analysis Related Work: MT overview Source Language Semantics-based MT Syntax-based MT Statistical MT, EBMT Setting the Stage Rule Learning Experimental Results Conclusions Analyze meaning Analyze structure Analyze sequence Target Language 12 Related Work (I) Setting the Stage Rule Learning Experimental Results Conclusions • Traditional transfer-based MT: analysis, transfer, generation (Hutchins and Somers 1992, Senellart et al. 2001) • Data-driven MT: – EBMT: store database of examples, possibly generalized (Sato and Nagao 1990, Brown 1997) – SMT: usually noisy channel model: translation model + target language model (Vogel et al. 2003, Och and Ney 2002, Brown 2004) • Hybrid (Knight et al. 1995, Habash and Dorr 2002) 13 Related Work (II) Setting the Stage Rule Learning Experimental Results Conclusions • Structure/syntax for MT – EBMT (Alshawi et al. 2000, Watanabe et al. 2002) – SMT (Yamada and Knight 2001, Wu 1997) – Other approaches (Habash and Dorr 2002, Menezes and Richardson 2001) • Learning from elicited data / small datasets (Nirenburg 1998, McShane et al 2003, Jones and Havrilla 1998) 14 Training Data Example Setting the Stage Rule Learning Experimental Results Conclusions SL: the widespread interest NP in the election [the interest the widespread in the election] DET ADJ N PP TL: h &niin h rxb b h bxirwt Alignment:((1,1),(1,3),(2,4),(3,2),(4,5 the widespread ),(5,6),(6,7)) interest Type: NP Parse: (<NP> (DET the-1) PREP NP (ADJ widespread-2) (N interest-3) (<PP> (PREP in-4) (<NP> (DET the-5) in DET N (N election-6)))) the election 15 Transfer Rule Formalism ;;L2: h &niin h rxb b h bxirwt ;;[the interest the widespread in the election] ;;L1: the widespread interest in the election NP::NP [“h” N “h” Adj PP] -> [“the” Adj N PP] ((X1::Y1)(X2::Y3) (X3::Y1)(X4::Y2) (X5::Y4) ((Y3 num) = (X2 num)) ((X2 num) = sg) ((X2 gen) = m)) Setting the Stage Rule Learning Experimental Results Conclusions Training example Rule type Component sequences Component alignments Agreement constraints Value constraints 16 Training Data Collection • • Setting the Stage Rule Learning Experimental Results Conclusions Elicitation Corpora – Generally designed to cover major linguistic phenomena – Bilingual user translates and word aligns Structural Elicitation Corpus – Designed to cover a wide variety of structural phenomena (Probst and Lavie 2004) – 120 sentences and phrases – Targeting specific constituent types: AdvP, AdjP, NP, PP, SBAR, S with subtypes 17 – Translated into Hebrew, Hindi Resources Setting the Stage Rule Learning Experimental Results Conclusions • L1 parses: Either from statistical parser (Charniak 1999), or use data from Penn Treebank • L1 morphology: Can be obtained or created (I created one for English) • L1 language model: Trained on a large amount of monolingual data • L2 morphology: If available, use morphology module. If not, use automated techniques, such as (Goldsmith 2001) or (Probst 2003). • Bilingual lexicon: gives word-level correspondences, created from training data or previously existing 18 Development and Testing Environment Setting the Stage Rule Learning Experimental Results Conclusions • Syntactic transfer engine: takes rules and lexicon and produces all possible partial translations • Statistical decoder: uses word-to-word probabilities and TL language model to extract best combination of partial translations (Vogel et al. 2003) 19 Setting the Stage Rule Learning Experimental Results Conclusions System Overview Bilingual training data Training time Rule Learner L1 parses & morphology L2 morphology Bilingual Lexicon Run time Learned Rules Transfer Engine Lattice L1 Language Model Statistical Decoder L2 test data Final Translation 20 Overview of Learning Phases 1. 2. 3. Setting the Stage Rule Learning Experimental Results Conclusions Seed Generation: create initial guesses at rules based on specific training examples Compositionality: add context-free structure to rules, rules can combine Constraint learning: learn appropriate unification constraints 21 Seed Generation Setting the Stage Rule Learning Experimental Results Conclusions • “Training example in rule format” • Produce rules that closely reflect training examples • But: generalize to POS level when words are 1-1 aligned • Rules are fully functional, but little generalization • Seed rules are intended as input for later two learning phases 22 Seed Generation – Sample Learned rule Setting the Stage Rule Learning Experimental Results Conclusions ;;L2: TKNIT H @IPWL H HTNDBWTIT ;;[ plan the care the voluntary] ;;L1: THE VOLUNTARY CARE PLAN ;;C-Structure:(<NP> (DET the-1) (<ADJP> (ADJ voluntary-2)) (N care-3)(N plan-4)) NP::NP [N "H" N "H" ADJ] -> ["THE" ADJ N N] ( (X1::Y4) (X3::Y3) (X5::Y2) ) 23 Seed Generation Algorithm Setting the Stage Rule Learning Experimental Results Conclusions • For a given training example, produce a seed rule • For all 1-1 aligned words, enter the POS tag (e.g. “N”) into component sequences – Get POS tags from morphology module and parse – Hypothesis: on unseen data, any words of this POS can fill this slot • For all not 1-1 aligned words, put actual words in component sequences • L2 and L1 type are parse’s root label • Derive alignments from training example 24 Compositionality Setting the Stage Rule Learning Experimental Results Conclusions • Generalize seed rules to reflect structure • Infer a partial constituent grammar for L2 • Rules map mixture of – Lexical items (LIT) – Parts of speech (PT) – Constituents (NT) • Analyze L1 parse to find generalizations • Produced rules are context-free 25 Compositionality Example Setting the Stage Rule Learning Experimental Results Conclusions ;;L2: $ BTWK H M&@PH HIH $M ;;[ that inside the envelope was name] ;;L1: THAT INSIDE THE ENVELOPE WAS A NAME ;;C-Structure:(<SBAR> (SUBORD that-1) (<SINV> (<PP> (PREP inside-2) (<NP> (DET the-3)(N envelope-4))) (<VP> (V was-5)) (<NP> (DET a-6)(N name-7)))) SBAR::SBAR [SUBORD PP V NP] -> [SUBORD PP V NP] ( (X1::Y1) (X2::Y2) (X3::Y3) (X4::Y4) ) 26 Basic Compositionality Algorithm Setting the Stage Rule Learning Experimental Results Conclusions • Traverse parse tree in order to partition sentence • For each sub-tree, if there is previously learned rule that can account for the subtree and its translation, introduce compositional element • Compositional element: subtree’s root label for both L1 and L2 • Adjust alignments • Note: preference for maximum generalization, because tree traversed from top 27 Maximum Compositionality Setting the Stage Rule Learning Experimental Results Conclusions • Assume that lower-level rules exist Assumption is correct if training data is completely compositional • Introduce compositional elements for direct children of parse root node • Results in higher level of compositionality, thus higher generalization power • Can overgeneralize, but because of strong decoder generally preferable 28 Other Advanced Compositionality Techniques Setting the Stage Rule Learning Experimental Results Conclusions • Techniques that allow you to generalize to POS not 1-1 aligned words • Techniques that enhance the dictionary based on training data • Techniques that deal with noun compounds • Rule filters to ensure that no learned rules violate axioms 29 Constraint Learning Setting the Stage Rule Learning Experimental Results Conclusions • Annotate context-free compositional rules with unification constraints a) limit applicability of rules to certain contexts (thereby limiting parsing ambiguity) b) ensure the passing of a feature value from source to target language (thereby limiting transfer ambiguity) c) disallow certain target language outputs (thereby limiting generation ambiguity) • Value constraints and agreement constraints are learned separately 30 Constraint Learning Overview 1. 2. 3. Setting the Stage Rule Learning Experimental Results Conclusions Introduce basic constraints: use morphology module(s) and parses to introduce constraints for words in training example Create agreement constraints (where appropriate) by merging basic constraints Retain appropriate value constraints: help in constricting a rule to some contexts or restricting output 31 Constraint Learning – Agreement Constraints (I) Setting the Stage Rule Learning Experimental Results Conclusions • For example: In an NP, do the adjective and the noun agree in number? • in Hebrew the good boys: • Correct: H ILDIM @WBIM the.det.def boy.pl.m good.pl.m “the good boys” • Incorrect: H ILDIM @WB the.det.def boy.pl.m good.sg.m “the good boys” 32 Constraint Learning – Agreement Constraints (II) Setting the Stage Rule Learning Experimental Results Conclusions • E.g. number in a determiner and the corresponding noun • Use a likelihood ratio test to determine what value constraints can be merged into agreement constraints • The log-likelihood ratio is defined by proposing distributions that could have given rise to the data: – Null Hypothesis: The values are independently distributed. – Alternative Hypothesis: The values are not independently distributed. • For sparse data, use heuristic test: if more evidence for than against agreement constraint 33 Constraint Learning – Agreement Constraints (III) Setting the Stage Rule Learning Experimental Results Conclusions • Collect all instances in the training data where an adjective and a noun mark for number • Count how often the feature value is the same, how often different • Feature values are distributed by – Two multinomial distributions (if they’re independent, e.g. Null hypothesis) – One multinomial distribution (if they should agree, e.g. Alternate hypothesis) • Compute log-likelihood under each scenario and perform LL ratio or heuristic test • Generalize to cross-lingual case 34 Constraint Learning – Value Constraints ;;L2: ild @wb ;;[ boy good] ;;L1: a good boy NP::NP [N ADJ] -> [``A'' ADJ N] (... ((X1 NUM) = SG) ((X2 NUM) = SG) ...) Setting the Stage Rule Learning Experimental Results Conclusions ;;L2: ildim t@wbim ;;[ boys good] ;;L1: good boys NP::NP [N ADJ] -> [ADJ N] (... ((X1 NUM) = PL) ((X2 NUM) = PL) ...) Retain value constraints to distinguish 35 Constraint Learning – Value Constraints Setting the Stage Rule Learning Experimental Results Conclusions • Retain those value constraints that determine the structure of the L2 translation • If you have two rules with – different L2 component sequences – same L1 component sequence – they differ in only a value constraint • Retain the value constraint to distinguish 36 Constraint Learning – Sample Learned Rule ;;L2: ANI AIN@LIGN@I ;;[ I intelligent] ;;L1: I AM INTELLIGENT S::S [NP ADJP] -> [NP “AM” ADJP] ( (X1::Y1) (X2::Y3) ((X1 NUM) = (X2 NUM)) ((Y1 NUM) = (X1 NUM)) ((Y1 PER) = (X1 PER)) (Y0 = Y2) ) Setting the Stage Rule Learning Experimental Results Conclusions 37 Dimensions of Evaluation Setting the Stage Rule Learning Experimental Results Conclusions • Learning Phases / Settings: default, Seed Generation only, Compositionality, Constraint Learning • Evaluation: rule-based evaluation + pruning • Test Corpora: TestSet, TestSuite • Run-time Settings: Lengthlimit • Portability: Hindi→English translation 38 Test Corpora Setting the Stage Rule Learning Experimental Results Conclusions • Test corpora: 1. Test Corpus: Newspaper text (Haaretz): 65 sentences, 1 reference translation 2. Test Suite: specific phenomena: 138 sentences, 1 reference translation 3. Hindi: 245 sentences, 4 reference translations • Compare: statistical system only, system with manually written grammar, system with learned grammar • Manually written grammar: written by expert within about a month (both Hebrew and Hindi) 39 Test Corpus Evaluation, Default Settings (I) Setting the Stage Rule Learning Experimental Results Conclusions Grammar BLEU METEOR No Grammar 0.0565 0.3019 Manual Grammar 0.0817 0.3241 Learned Grammar (With Constraints) 0.078 0.3293 40 Test Corpus Evaluation, Default Settings (II) Setting the Stage Rule Learning Experimental Results Conclusions Learned grammar performs statistically significantly better than baseline • Performed one-tailed paired t-test • BLEU with resampling: t-value: 81.98, p-value:0 (df=999) → Significant at 100% confidence level Median of differences: -0.0217 with 95% confidence interval [-0.0383,-0.0056] • METEOR: t-value: 1.73, p-value: 0.044 (df=61) → Significant at higher than 95% confidence level 41 Test Corpus Evaluation, Default Settings (III) Setting the Stage Rule Learning Experimental Results Conclusions 42 Test Corpus Evaluation, Different Settings (I) Setting the Stage Rule Learning Experimental Results Conclusions Grammar BLEU METEOR No Grammar 0.0565 0.3019 Manual Grammar 0.0817 0.3241 Learned Grammar (Seed Generation) 0.0741 0.3239 Learned Grammar (Compositionality) 0.0777 0.3360 Learned Grammar (With Constraints) 0.078 0.3293 43 Test Corpus Evaluation, Different Settings (II) Setting the Stage Rule Learning Experimental Results Conclusions System times in seconds, lattice sizes: Grammar Learned Grammar (Compositionality) Learned Grammar (With Constraints) Transfer Engine (in seconds system time) Lattice size (in mb) Decoder (in seconds system time) 54.98 187 3123.38 33.28 140 2287.47 → ~ 20% reduction in lattice size! 44 Evaluation with Rule Scoring (I) • • • • • • • Setting the Stage Rule Learning Experimental Results Conclusions Estimate translation power of the rules Use training data: most training examples are actually unseen data for a given rule Match arc against the reference translation A rule’s score is the average of all its arcs’ scores Order the rules by precision score, prune Goal of rule scoring: limit run-time Note trade-off with decoder power 45 Setting the Stage Rule Learning Experimental Results Conclusions Evaluation with Rule Scoring (II) Grammar BLEU ModBLEU METEOR No Grammar 0.0565 0.1362 0.3019 Manual Grammar 0.0817 0.1546 0.3241 Learned Grammar (25%) 0.0565 0.1362 0.3019 Learned Grammar (50%) 0.0592 0.1389 0.3075 Learned Grammar (75%) 0.0800 0.1533 0.3296 Learned Grammar (full) 0.078 0.1524 0.3293 46 Evaluation with Rule Scoring (III) Setting the Stage Rule Learning Experimental Results Conclusions Grammar TrEngine LatticeSize Decoder Learned Grammar (25%) 1.02 330342 22.55 Learned Grammar (50%) 1.81 13431206 189.89 Learned Grammar (75%) 5.91 29242597 397.06 Learned Grammar (full) 33.28 149713589 2287.47 47 Test Suite Evaluation (I) Setting the Stage Rule Learning Experimental Results Conclusions • Test suite designed to target specific constructions – Conjunctions of PPs – Adverb phrases – Reordering of adjectives and nouns – AdjP embedded in NP – Possessives –… • Designed in English, translated into Hebrew • 138 sentences, one reference translation 48 Test Suite Evaluation (II) Grammar BLEU Setting the Stage Rule Learning Experimental Results Conclusions METEOR Baseline 0.0746 0.4146 Manual grammar 0.1179 0.4471 Learned Grammar 0.1199 0.4655 49 Test Suite Evaluation (III) Setting the Stage Rule Learning Experimental Results Conclusions Learned grammar performs statistically significantly better than baseline • Performed one-tailed paired t-test • BLEU with resampling: t-value: 122.53, p-value:0 (df=999) → Statistically significantly better at 100% confidence level Median of differences: -0.0462 with 95% confidence interval [-0.0245,-0.0721] • METEOR: t-value: 47.20, p-value: 0.0 (df=137) → Statistically significantly better at 100% confidence level 50 Test Suite Evaluation (IV) Setting the Stage Rule Learning Experimental Results Conclusions 51 Hindi-English Portability Test (I) Grammar BLEU Setting the Stage Rule Learning Experimental Results Conclusions METEOR Baseline 0.1003 0.3659 Manual grammar 0.1052 0.3676 Learned Grammar 0.1033 0.3685 52 Hindi-English Portability Test (II) Setting the Stage Rule Learning Experimental Results Conclusions Learned grammar performs statistically significantly better than baseline • Performed one-tailed paired t-test • BLEU with resampling: t-value: 37.20, p-value:0 (df=999) → Statistically significantly better at 100% confidence level Median of differences: -0.0024 with 80% confidence interval [-0.0052,0.0001] • METEOR: t-value: 1.72, p-value: 0.043 (df=244) → Statistically significantly better at higher than 95% 53 confidence level Hindi-English Portability Test (III) Setting the Stage Rule Learning Experimental Results Conclusions 54 Discussion of Results • • • • • Setting the Stage Rule Learning Experimental Results Conclusions Performance superior to standard SMT system Learned grammar comparable to manual grammar Learned grammar: higher METEOR score, indicating that it is more general Constraints: slightly lower performance in exchange for higher run-time efficiency Pruning: slightly lower performance in exchange for higher run-time efficiency 55 Conclusions and Contributions 1. 2. 3. 4. 5. 6. Setting the Stage Rule Learning Experimental Results Conclusions Framework for learning transfer rules from bilingual data Improvement of translation output in hybrid transfer and statistical system Addressing limited-data scenarios with ‘frugal’ techniques Combining different knowledge sources in a meaningful way Pushing MT research in the direction of incorporating syntax into statistical-based systems Human-readable rules that can be improved by an expert 56 Summary Setting the Stage Rule Learning Experimental Results Conclusions “Take a bilingual word-aligned corpus, and learn transfer rules with constituent transfer and unification constraints.” “Is it a big corpus?” “Ahem. No.” “Do I have a parser for both languages?” “No, just for one.” “… So I can use a dictionary, morphology modules, a parser … But these are all imperfect resources. How do I combine them?” “We can do it!” “Ok.” 57 Thank you! 58 Additional Slides 59 References (I) Ayan, Fazil, Bonnie J. Dorr, and Nizar Habash. Application of Alignment to Real-World Data: Combining Linguistic and Statistical Techniques for Adaptable MT. Proceedings of AMTA-2004. Baldwin, Timothy and Aline Villavicencio. 2002. Extracting the Unextractable: A case study on verb-particles. Proceedings of CoNLL-2002. Brown, Ralf D., A Modified Burrows-Wheeler Transform for Highly-Scalable Example-Based Translation, Proceedings of AMTA-2004. Charniak, Eugene, Kevin Knight and Kenji Yamada. 2003. Syntax-based Language Models for Statistical Machine Translation. Proceedings of MT-Summit IX. 60 References (II) Hutchins, John W. and Harold L. Somers. 1992. An Introduction to Machine Translation. Academic Press, London. Jones, Douglas and R. Havrilla. Twisted Pair Grammar: Support for Rapid Development of Machine Translation for Low Density Languages. Proceedings of AMTA-98. Menezes, Arul and Stephen D. Richardson. A best-first alignment algorithm for automatic extraction of transfer mappings from bilingual corpora. Proceedings of the Workshop on Data-driven Machine Translation at ACL2001. Nirenburg, Sergei. Project Boas: A Linguist in the Box as a Multi-Purpose Language Resource. Proceedings of LREC-98. 61 References (III) Orasan, Constantin and Richard Evans. 2001. Learning to identify animate references. Proceedings of CoNLL2001. Probst, Katharina. 2003. Using ‘smart’ bilingual projection to feature-tag a monolingual dictionary. Proceedings of CoNLL-2003. Probst, Katharina and Alon Lavie. A Structurally Diverse Minimal Corpus for Eliciting Structural Mappings between Languages. Proceedings of AMTA-04. Probst, Katharina and Lori Levin. 2002. Challenges in Automated Elicitation of a Controlled Bilingual Corpus. Proceedings of TMI-02. 62 References (IV) Senellart, Jean, Mirko Plitt, Christophe Bailly, and Francoise Cardoso. 2001. Resource Alignment and Implicit Transfer. Proceedings of MT-Summit VIII. Vogel, Stephan and Alicia Tribble. 2002. Improving Statistical Machine Translation for a Speech-to-Speech Translation Task. Proceedings of ICSLP-2002. Watanabe, Hideo, Sadao Kurohashi, and Eiji Aramaki. 2000. Finding Structural Correspondences from Bilingual Parsed Corpus for Corpus-based Translation. Proceedings of COLING-2000. 63 Log-likelihood test for agreement constraints (I) Setting the Stage Rule Learning Experimental Results Conclusions Future Work • create list of all possible index pairs that should be considered for an agreement constraint: • L1 only constraints: – list of all head-head pairs that ever occur with the same feature (not necessarily same value), and all head-nonheads in the same constituent that occur with the same feature (not necessarily same value). – For example, possible agreement constraint: Num agreement between Det and N in a NP where the Det is a dependent of N • L2 only constraints: same as L1 only constraints above. • L2→L1 constraints: all situations where two aligned indices mark the same feature 64 Log-likelihood test for agreement constraints (II) Setting the Stage Rule Learning Experimental Results Conclusions Future Work • Hypothesis 0: The values are independently distributed. • Hypothesis 1: The values are not independently distributed. • Under the null hypothesis: • Under the alternative hypothesis: where ind is 1 if vxi1 = vxi2 and 0 otherwise. 65 Log-likelihood test for agreement constraints (III) Setting the Stage Rule Learning Experimental Results Conclusions Future Work • i1 and i2 drawn from a multinomial distribution. where cvi is the number of times the value vi was encountered for the given feature (e.g. PERS), and k is the number of possible values for the feature (e.g. 1st, 2nd, 3rd). • If strong evidence for Hypothesis 0, introduce agreement constraint • For cases where there is not enough evidence either 66 way (n<10), use heuristic test Lexicon Enhancement for Hebrew Adverbs (I) • • • • • Example 1: “B” “$MX” “happily” Example 2: “IWTR” “GBWH” “taller” These are not necessarily in the dictionary Both processes are productive How can we add these and similar entries to lexicon? Automatically? 67 Lexicon Enhancement for Hebrew Adverbs (II) For all 1-2 (L1-L2) alignments in training data{ 1. extract all cases of at least 2 instances where one word is constant (constant word: wL2c, non-constant word wL2v, non-constant word wL1v) 2. For each word wL2v{ 2.1. Get all L1 translations 2.2. Find the closest match wL1match to wL1v 2.3. Learn replacement rule wL1match->wL1v } 3. For each word wL2POS of same POS as wL2c{ 3.1. For each possible translations wL1POS { 3.1.1. Apply all replacement rules possible wL1POS->wL1POSmod 3.1.2. For each applied replacement rule, insert into lexicon entry: [“wc” wL2POS] -> [wL1POSmod] } } 68 Lexicon Enhancement for Hebrew Adverbs (III) • Example: B $MX -> happily • Possible translations of $MX: – joy – happiness • Use edit distance to find that happiness is wL1match for happily • Learn replacement rule ness->ly 69 Lexicon Enhancement for Hebrew Adverbs (IV) • For all L2 Nouns in the dictionary, get all possible L1 translations, and apply the replacement rule • If replacement rule can be applied, add lexicon entry • Examples of new adverbs added to lexicon: ADV::ADV |: ["B" "$APTNWT"] -> ["AMBITIOUSLY"] ADV::ADV |: ["B" "$BIRWT"] -> ["BRITTLELY"] ADV::ADV |: ["B" "$G&WN"] -> ["MADLY"] ADV::ADV |: ["B" "$I@TIWT"] -> ["METHODICALLY"]70 Lexicon Enhancement for Hebrew Comparatives • Same process as for adverbs • Examples of new comparatives added to lexicon: ADJ::ADJ |: ["IWTR" "MLA"] -> ["FULLER"] ADJ::ADJ |: ["IWTR" "MPGR"] -> ["SLOWER"] ADJ::ADJ |: ["IWTR" "MQCH"] -> ["HEATER"] • All words are checked in the BNC • Comment: automatic process, thus far from perfect 71 Some notation • • • • • • • • • Setting the Stage Rule Learning Experimental Results Conclusions Future Work SL: Source Language, language to be translated from TL: Target Language, language to be translated into L1: language for which abundant information is available L2: language for which less information is available (Here:) SL = L2 = Hebrew, Hindi (Here:) TL = L1 = English POS: part of speech, e.g. noun, adjective, verb Parse: structural (tree) analysis of sentence Lattice: list of partial translations, arranged by length and start index 72 Training Data Example Setting the Stage Rule Learning Experimental Results Conclusions Future Work SL: the widespread interest in the election TL: h &niin h rxb b h bxirwt Alignment:((1,1),(1,3),(2,4),(3,2),(4,5),(5,6),(6,7)) Type: NP Parse: (<NP> (DET the-1)(ADJ widespread-2)(N interest-3) (<PP> (PREP in-4) (<NP> (DET the-5)(N election-6)))) 73 Seed Generation Algorithm Setting the Stage Rule Learning Experimental Results Conclusions Future Work for all training examples { for all 1-1 aligned words { get the L1 POS tag from the parse get the L2 POS tag from the morphology module and the dictionary if the L1 POS and the L2 POS tags are not the same, leave both words lexicalized } for all other words { leave the words lexicalized } create rule word alignments from training example set L2 type and L1 type to be the parse root’s label } 74 Taxonomy of Structural Mappings (I) Setting the Stage Rule Learning Experimental Results Conclusions Future Work • Non-terminals (NT): – used in two rule parts: • type definition of a rule (both for SLand TL, meaning X0 and Y0), • constituent sequences for both languages. – any label that can be the type of a rule – describe higher-level structures such as sentences (S), noun phrases (NP), or prepositional phrases(PP). – can be filled with more than one word: filled by other rules. 75 Taxonomy of Structural Mappings (II) Setting the Stage Rule Learning Experimental Results Conclusions Future Work • Pre-terminals (PT): – used only in the constituent sequences of the rules, not as X0 or Y0 types. – filled with only one word, except phrasal lexicon entries: filled by lexical entries, not by other grammar rules. • Terminals (LIT): – lexicalized entries in the constituent sequences – can be used on both the x- and the y-side – can only be filled by the specified terminal itself. 76 Taxonomy of Structural Mappings (III) Setting the Stage Rule Learning Experimental Results Conclusions Future Work • NTs must not be aligned 1-0 or 0-1 • PTs must not be aligned 1-0 or 0-1. • Any word in the bilingual training pair must participate in exactly one LIT, PT, or NT. • An L1 NT is assumed to translate into the same NT inL2. 77 Taxonomy of Structural Mappings (IV) Setting the Stage Rule Learning Experimental Results Conclusions Future Work • Transformation I (SL type into SL component sequence). NT → (NT | PT | LIT)+ • Transformation II (SL type into TL type). NTi → NTi (same type of NT) • Transformation III (TL type into TL component sequence). NT → (NT | PT | LIT)+ • Transformation IV (SL components into TL components). NTi → NTi+ (same type of NT) PT → PT+ LIT → ε ε → LIT 78 Basic Compositionality Pseudocode Setting the Stage Rule Learning Experimental Results Conclusions Future Work traverse parse top-down for each node i in parse { extract the subtree rooted at i extract the L1 chunk cL1 rooted at i and the corresponding L2 chunk cL2 (using alignments) if transfer engine can translate cL1 into cL2 using previously learned rules { introduce compositional element: replace POS sequence for cL1 and cL2 with label of node i adjust alignments } do not traverse already covered subtree } } 79 Co-Embedding Resolution, Iterative Type Learning Setting the Stage Rule Learning Experimental Results Conclusions Future Work • Problem: looking for previously learned rules – Must determine optimal learning ordering • Co-Embedding Resolution: – Tag each training example with depth of tree, i.e. how many embedded elements – Then learn lowest to highest • Iterative Type Learning: – Some types (e.g. PPs) are frequently embedded in others (e.g. NP) – Pre-determine the order in which types are learned 80 Compositionality – Sample Learned Rules (II) Setting the Stage Rule Learning Experimental Results Conclusions ;;L2: RQ AM H RKBT TGI& ;;L1: ONLY IF THE TRAIN ARRIVES ;;C-Structure:(<SBAR> (<ADVP> (ADV only-1)) (SUBORD if-2) (<S> (<NP> (DET the-3)(N train-4)) (<VP> (V arrives-5)))) SBAR::SBAR [ADVP SUBORD S] -> [ADVP SUBORD S] ( (X1::Y1) (X2::Y2) (X3::Y3) ) 81 Taxonomy of Constraints (I) Setting the Stage Rule Learning Experimental Results Conclusions Future Work Parameter Possible Values value or agreement value, agreement level POS, constituent, POS/constituent L2, L1, L2→L1 language constrains head head, non-head, head+nonhead 82 Co-Embedding Resolution, Iterative Type Learning Setting the Stage Rule Learning Experimental Results Conclusions Future Work find highest co-embedding score in training data find the number of types to learn, ntypes for (i = 0; i < maxco-embedding; i++) { for (j = 0; j < ntypes; j++) { for all training examples with co-embedding score i and of type j { perform Seed Generation perform Compositionality Learning } } } 83 Taxonomy of Constraints (II) Setting the Stage Rule Learning Experimental Results Conclusions Future Work Subtype Value/ Agreement Language Level Comment 1 value x POS Group1-2 2 value x const Group1-2 3 value x POS/const can't exist 4 value y POS Group4 5 value y const Group5 6 value y POS/const can't exist 7 value xy POS can't exist 8 value xy const can't exist 9 value xy POS/const can't exist 10 agreement x POS Group10-12 11 agreement x const Group10-12 12 agreement x POS/const Group10-12 13 agreement y POS 14 agreement y const 15 agreement y POS/const Group13-15 16 agreement xy POS Group16-18 17 agreement xy const Group16-18 18 agreement xy POS/const Group16-18 Group13-15 Group13-15 84 Taxonomy of Constraints (III) Subtype Value/ Agr Language Setting the Stage Rule Learning Experimental Results Conclusions Future Work Level 1,2 value x POS or const 4,5 value y POS or const 10,11,12 agreement x POS or const or POS/const 13,14,15 agreement y POS or const or POS/const 16,17,18 agreement xy POS or const or POS/const 85 Constraint Learning – Sample Learned Rules (II) Setting the Stage Rule Learning Experimental Results Conclusions ;;L2: H ILD AKL KI HWA HIH R&B ;;L1: THE BOY ATE BECAUSE HE WAS HUNGRY S::S [NP V SBAR] -> [NP V SBAR] ( (X1::Y1) (X2::Y2) (X3::Y3) (X0 = X2) ((X1 GEN) = (X2 GEN)) ((X1 NUM) = (X2 NUM)) ((Y1 NUM) = (X1 NUM)) ((Y2 TENSE) = (X2 TENSE)) ((Y3 NUM) = (X3 NUM)) ((Y3 TENSE) = (X3 TENSE)) (Y0 = Y2)) 86 Evaluation with Different Length Limits (I) Grammar 1 2 No Grammar 0.171 0.2962 0.3016 0.3012 0.3019 0.3019 Manual Grammar Learned Grammar 0.1744 0.297 0.171 3 4 Setting the Stage Rule Learning Experimental Results Conclusions Future Work 5 6 0.3141 0.3182 0.3232 0.3241 0.2995 0.3072 0.3252 0.3282 0.3293 87 Evaluation with Different Length Limits (II) (METEOR score) Setting the Stage Rule Learning Experimental Results Conclusions Future Work 88 Discussion of Results: Comparison of Translations Setting the Stage Rule Learning Experimental Results Conclusions Future Work (back to Hebrew-English) No grammar: the doctor helps to patients his Learned grammar: the doctor helps to his patients Reference translation: The doctor helps his patients No grammar: the soldier writes many letters to the family of he Learned grammar: the soldier writes many letters to his family Reference translation: The soldier writes many letters to his family 89 Time Complexity of Algorithms • • • • Setting the Stage Rule Learning Experimental Results Conclusions Seed Generation: O(n) Compositionality: –Basic: O(n*max(tree_depth)) –Maximum Compositionality: O(n*max(num_children)) Constraint Learning: O(n*max(num_basic_constraints)) Practically: no issue 90 If I had 6 more months… • • • Setting the Stage Rule Learning Experimental Results Conclusions Future Work Application to larger datasets – Training data enhancement to obtain training examples at different levels (NPs, PPs, etc.) – More emphasis on rule scoring (more noise) – More emphasis on context learning: constraints Constraint learning as version space learning problem Integrate rules into statistical system more directly, without producing full lattice 91