Machine Translation implementation Classical and Statistical - PowerPoint PPT Presentation

Session 3: Constraint-based Transfer � Lab1: Syntactic Transfer – a Prolog Machine Translation implementation – Classical and Statistical Approaches � Background on Lexical-Functional Grammar (LFG) Session 3: Constraint-based Transfer � Constraint-based Transfer � Kaplan et al. 1989: Translation by Structural Jonas Kuhn Universität des Saarlandes, Saarbrücken Correspondence The University of Texas at Austin jonask@coli.uni-sb.de DGfS/CL Fall School 2005, Ruhr-Universität Bochum, September 19-30, 2005 Jonas Kuhn: MT 2 Lab1: Syntactic Transfer Lab1: Suggested exercises � A Prolog implementation 1. Specify simple rules required for some of the English �� Spanish divergence examples from � Analysis � Transfer � Generation Trujillo chapter � Each part: processing engine + declarative rules Change the lexicon: distinguish a surface form and 2. � Variants of this lab exercise depending on an underlying citation form Prolog background: The transfer rules can be expressed more generally � � Use given Prolog implementation of the processing systems + concentrate on specifying The language-specific grammars should constrain � rules possible combinations of morphological variants � Use given parser and generator, but specify You can try to add certain ideas from interlingua- � Prolog predicate for transfer based translation (language-independent citation � Implement all parts from scratch (not forms, e.g., for definite articles) recommended due to time constraits) Add a new language pair (e.g., 3. English �� German) Jonas Kuhn: MT 3 Jonas Kuhn: MT 4

Some additional background on lab Given code for the lab � The given parser and generator do not use the built-in DCG rule � The Prolog predicates for processing are format, but a slightly different format defined in a file engine.pl , separate from the � The parse-tree building argument is automatically added in a compilation step grammar rules � Each rule has an argument for the language it belongs to (this could be exploited for language-independent rules, � The engine.pl file is consulted automatically using “_”) when the grammar file is loaded � Each category is represented as a Prolog list: the first element is the main category label, the remaining elments � This is triggered by the following line at the can be used for linguistic features beginning of the grammar file: rule(en, [np,Agr] ---> [ [det,Agr], [n1,Agr] ] ). rule(en, [n1,Agr] ---> [ [n,Agr] ] ). :- ensure_loaded('engine.pl'). rule(en, [n1,Agr] ---> [ [adj,Agr], [n,Agr] ] ). word(en, [det,_]/the ---> [the]). word(en, [n, agr(sg)]/soup ---> [soup]). word(en, [adj, agr(sg)]/delicious ---> [delicious]). Jonas Kuhn: MT 5 Jonas Kuhn: MT 6 Given code for the lab Given code for the lab � When the grammar rules are consulted, they � Parsing predicate are automatically compiled into a special parse(Lang,StartSymbol,String,Tree) internal format � Example: � This is triggered by the following line at the ?- parse(en, [np|_]/_, [the,soup], T). end: � Short parse predicate: :- compile_grammars. p(String) � You can look at the effect of the compilation � The language and start symbol are “guessed” on tree representations by parsing a string � The resulting tree is printed using the (using the given parse predicates) pretty_print predicate Jonas Kuhn: MT 7 Jonas Kuhn: MT 8

Given code for the lab Given code for the lab � Transfer predicate � Generation predicate translate(SourceLang,TargetLang, generate(Lang,Tree,String) SourceString,TargetString) � Example: � The translation predicates � calls the parsing predicate (guessing the start symbol) ?- parse(en, [np|_]/_, [the,soup], T), � prints the resulting source tree generate(en,T,X). � recursively traverses the source tree, applying � Short generation predicate: transformation rules � prints the resulting (underspecified) target tree g(Tree) � calls the generation predicate � Again, the language is “guessed” � returns the resulting string as TargetString � Example: � The resulting string is printed ?- translate(en,es,[the,delicious,soup],S). � Short translation predicate: t(SourceString) Jonas Kuhn: MT 9 Jonas Kuhn: MT 10 Transformation rules Lexical-Functional Grammar � Grammar formalism originally proposed by Joan � The rules will be applied top-down in the Bresnan and Ron Kaplan in the early 1980s order they are specified � Non-derivational � A cut (“ ! ”) is used to exclude backtracking � Lexicalist � Parallel levels of representation � The more specific rules should go to the top! � C(onstituent)-structure [Phrase structure tree repr.] � F(unctional)-structure [Complex feature structure repr.] � (Lexicosemantic) A(rgument)-structure � Semantic Structure � Grammar specification as constraint schemata describing projection functions across the levels of representation Jonas Kuhn: MT 11 Jonas Kuhn: MT 12

Lexical-Functional Grammar Lexical-Functional Grammar � Constraints describing the relation between � Computational perspective: various levels of representation � One of the best high-level grammar development system (with a very fast parser/generator) is based on an LFG implementation C-Structure F-Structure � XLE – developed over c. 15 years by Ron Kaplan, John Maxwell and colleagues at the Palo Alto Research Center (formerly Xerox PARC) � Broad-coverage LFG grammars for a growing number of languages have been developed in the Parallel Grammar development project (ParGram) � English, German, French, Japanese, Norwegian, Korean, Arabic, … ( φ (n 1 ) SUBJ) = φ (n 2 ) Jonas Kuhn: MT 13 Jonas Kuhn: MT 14 Lexical-Functional Grammar LFG-based Machine Translation � F-Structure as a syntactic representation abstracting � Additional “tau projection” to describe cross- away from most language-specific aspects of linguistic relation realization � Constraints on F-Structure are specified as annotations in (C-Structure) rewrite rules S � NP VP ( ↑ SUBJ)= ↓ ↑ = ↓ Read: (1) the f-structure projected from the NP node ( ↓ ) is the same as the f-structure embedded under the feature SUBJ in the f-structure projected from the S node ( ↑ ) ; (2) the f-structure projected from the VP node ( ↓ ) is identical to the f-structure projected from the S node ( ↑ ) Jonas Kuhn: MT 15 Jonas Kuhn: MT 16

Tau projection: structural divergence Tau projection: structural divergence S � NP VP ( ↑ SUBJ)= ↓ ↑ = ↓ � Example VP � V NP ↑ ↑ = ↓ ( ↑ OBJ)= ↓ SUBJ � German: Der Student beantwortet die Frage φ S ( ↑ SUBJ) � French: L’étudiant répond à la question VP OBJ ( ↑ OBJ) NP NP V DET N DET N Student beantwortet die der Frage � Source language information from the verb’s lexical entry: Jonas Kuhn: MT 17 Jonas Kuhn: MT 18 Tau projection: structural divergence Tau projection: head switching τ τ ↑ � Example: SUBJ ↑ τ ( ↑ SUBJ) � English: The baby just fell � French: Le bébé vient de tomber AOBJ OBJ τ ( ↑ OBJ) � Assuming that at f-structure adverbs like just introduce a grammatical function that embeds φ the f-structure of the rest of the sentence Information from the transfer lexicon entry for the verb beantworten : � beantwortet Jonas Kuhn: MT 19 Jonas Kuhn: MT 20

Tau projection: head switching Tau projection: head switching φ φ S S NP NP ADV ADV VP VP DET DET N N Transfer entry for adverb: baby just baby just the the fell fell Jonas Kuhn: MT 21 Jonas Kuhn: MT 22 Tau projection: head switching Issues in the tau projection approach τ τ ↑ � Kaplan et al. 1989 paper is the beginning of a debate: � Kaplan et al. 1989 (original proposal) � Sadler/Thompson 1991 (criticism: problem with head-switching in embedded contexts) � Kaplan/Wedekind 1993 (reply: extension of LFG formalism by the restriction operator to deal with mismatches between syntactic and Transfer entry for adverb: semantic heads) Jonas Kuhn: MT 23 Jonas Kuhn: MT 24

LFG-based Machine Translation � Conclusion: Flexible approach taking advantage of linguistic generalizations at various levels � Modularization of transfer information (separate from language-specific grammatical information) is somewhat problematic Jonas Kuhn: MT 25