Translation Divergence LING 580MT Fei Xia 1/10/06 Papers • Bonnie Dorr (1994): Machine Translation Divergences: a Formal Description and Proposed Solution Outline • • • • Formal definition of translation divergence Seven types of divergence Discussion Remaining questions Formal definition of translation divergence Distinction between the source and target languages Two categories (Bernett et. al., 1991): • Translation divergence: same information, different structures • Translation mismatches: different information important, but outside of the scope of the paper How to define translation divergence formally? Define the language-to-language divergence via language-to-interlingua divergence: Interlingua: lexical conceptual structure (LCS) Language-to-interlingua: mapping from syntactic form to LCS Lexical conceptual structure (LCS) [ T ( X ') X ' ([ T (W ') W ' ], ' [ T ( Z ' ) Z 1 ]...[ 1 ' ' T (Zn ) Zn] ' [ T ( Q ' ) Q 1 ]...[ 1 ' ' T (Qm ) Q m ])] X’ T(X’) W’ T(W’) Z1’ T(Z1’) … Zn’ T(Zn’) Q 1’ T(Q1’) … Q m’ T(Qm’) X’: logical head W’: logical subject Z1’…Zn’: logical argument Q1’…Qm’: logical modifiers T(Φ’) is the logical type (Event, Path, ….) of the primitive Φ’ (CAUSE, LET, GO, …) Root LCS (RLCS) • A RLCS is an un-instantiated LCS that is associated with a word definition in the lexicon (i.e., a LCS with unfilled variable position) • LCSs are recursively defined. RLCS representation for go GOLoc Event X Thing TOLoc Path ATLoc Position X Thing Z Location It is different from dependency structure Composed LCS (CLCS) • A CLCS is an instantiated LCS that is the result of combining two or more RLCSs by means of unification (roughly). • This is the interlingua form that serves as the pivot between the source and target languages. CLCS representation for “John went happily to school” GOLoc Event John Thing TOLoc Path Happily Manner ATLoc Position John Thing School Location The operations of combining are not defined in this paper. Syntactic phrase X: syntactic head W: external argument Z-MAX i: internal arguments Q-MAXi: syntactic adjuncts Similar to X-bar theory, GB theory, etc. An example Mapping between LCS and syntactic form • Generalized linking routine (GLR): – – – – X’ X W’ W Z’ Z Q’ Q (logical head syntactic head) (logical subject external argument) (logical argument internal argument) (logical modifiers syntactic adjunct) • Canonical syntactic realization (CSR) – Relate T(Φ’) to CAT(Φ): (logical type syntactic category) Ex: THING N, EVENT V Divergence problem • Translation divergences occur when there is an exception either to the GLR or to the CSR (or to both) in one language, but not in the other. Outline • • • • Formal definition of translation divergence Seven types of divergence Discussion Remaining questions T1: Thematic divergence • The repositioning of arguments w.r.t. a head. • GLR: W’ Z and Z’W • Example: I like Mary Maria me gusta :INT and :EXT General Solution T2: Promotional Divergence • Promoting a logical modifier into a main verb position (or vice versa) • GLR: X’Z and Q’X • Ex: John usually goes home Juan suele ir a casa :PROMOTE General Solution T3: Demotional Divergence • Demoting a logical head into an internal argument (adjunct?) position (or vice versa). • GLR: X’Q and Z’X • Ex: I like to eat Ich gern esse :DEMOTE General Solution T4: Structural divergence • It does not alter the positions used in GLR mapping • But it changes the nature of the relation between different positions (i.e., the “” correspondence) • Ex: John entered the house Juan entro en la casa * marker Marker forces logical constituents to be realized compositionally at different levels General solution T5: Conflational Divergence • The suppression of a CLCS constituent (or the inverse of the process) • GLR: correspondence of step (3) or (4) is changed. Example I stabbed John Yo le di punaladas a Juan :CONFLATED General solution T6: Categorical divergence • CAT(Φ) is different from CSR(T(Φ’)). • Ex: I am hungry Ich hunger habe :CAT General solution T7: Lexical divergence • As a side effect of other divergences. • Ex: John broke into the room Juan forzo la entrada al cuarto Summary of seven types • Repositioning (GLR mappings): thematic, promotional, demotional divergences • Changing correspondence: structural, conflational divergences • Category: categorical divergence • ??: Lexical divergence Discussion Discussion • Limits on Repositioning Divergences • Promotional vs. Demotional Divergences • Lexical Selection: Full Coverage Constraint • Interacting Divergence Types Limits on Repositioning divergences • Three types to cover all repositioning divergences: – Thematic: W’Z, Z’W – Promotional: X’Z, Q’X – Demotional: X’Q, Z’X • (X, W, Z, Q) (X’, W’, Z’, Q’) – W has a special status: 44=256 33=27 – a CLCS must contain exactly one head: 33=2712 Limits on Repositioning Divergences (cont) • Z can never be associated with Q’, and Q can never be associated with Z’: 12 5 • Modifying relation cannot be reversed: 54 (Q’X, X’Q, Z’Z) • Argument relation cannot be reversed: 4 3 (Z’X, X’Z, Q’Q) • Canonical positions: 3 2 Promotional vs. Demotional Divergences • Promotion is triggered by a main verb (e.g., soler in soler-usually) • Demotion is triggered by an adverb (e.g., gern in like-gern) Interacting Divergence Types • Promotional and thematic divergence: S: Leer libros le suele gustar a Juan ‘reading books (him) tends to please (to) John’ E: John usually likes reading books Remaining questions Remaining questions: Interlingua • How to build RLCS? – What are logical head, subject, arguments and modifiers? Ex: like likingly – How to represent a verb: stab CAUSE GOPoss KNIFE-WOUND • How are RLCSs combined to form CLCSs? – Unification = substitution? • Are CLCSs really sufficient to handle all the languages? Remaining issues: divergences • Are the seven types really sufficient to cover all the convergences? – Is the “proof” for limits on repositioning divergences convincing? – “Translation divergences occur when there is an exception to GLR/CSR in one language, but not the other”: what if there are exceptions in both languages? – Can a dependent of X become a dependent of Y? Remaining issues: MT • How to build a real MT system with this approach?