C SC 620 Advanced Topics in Natural Language ProcessingReading ListPowerPoint PresentationSlide 4Slide 5Slide 6Slide 7Slide 8Slide 9Slide 10Slide 11Slide 12Paper 22. Pros and Cons of the Pivot and Transfer Approaches in Multilingual Machine Translation. Boitet, C.Slide 14Slide 15Slide 16Slide 17Slide 18Slide 19Slide 20Slide 21C SC 620Advanced Topics in Natural Language ProcessingLecture 224/15Reading List•Readings in Machine Translation, Eds. Nirenburg, S. et al. MIT Press 2003.–19. Montague Grammar and Machine Translation. Landsbergen, J.–20. Dialogue Translation vs. Text Translation – Interpretation Based Approach. Tsujii, J.-I. And M. Nagao–21. Translation by Structural Correspondences. Kaplan, R. et al.–22. Pros and Cons of the Pivot and Transfer Approaches in Multilingual Machine Translation. Boitet, C.–31. A Framework of a Mechanical Translation between Japanese and English by Analogy Principle. Nagao, M.–32. A Statistical Approach to Machine Translation. Brown, P. F. et al.•Similar to the PhraselatorPaper 22. Pros and Cons of the Pivot and Transfer Approaches in Multilingual Machine Translation. Boitet, C.•Time: 90s•Introduction: Why is the Pivot Approach Not Universally Used?–Pivot (interlingua): O(n) parsers/analyzers–Transfer: O(n2) parsers/analyzers–n = number of languages–Pivot dictionaries: monolingual–Transfer dictionaries: bilingualPaper 22. Pros and Cons of the Pivot and Transfer Approaches in Multilingual Machine Translation. Boitet, C.•Pure Pivot Approaches–Independent pivot lexicon–Universal notation for determination, quantification, actualization (time/modality/aspect), thematization, etc.•I.1 Pure Pivot Lexicons are Challenging …•1.1 … But Specific of a Domain (Interpretation Language)–May be possible to define a completely artificial language for a fixed and restricted domain–TITUS system: textile domain•1.2 … Or Specific of a Language Group (Standard Language)–Standard Language: e.g. English•Double translations for all pairs of languages not containing the pivot•No implementation known•“Idiosyncratic gap” between language familiesPaper 22. Pros and Cons of the Pivot and Transfer Approaches in Multilingual Machine Translation. Boitet, C.•1.2 … Or Specific of a Language Group (Standard Language)–Artificial Language: e.g. Esperanto•BSO project•Double translations for all pairs of languages•Lack of sufficient technical vocabulary–need about 50,000 terms in any typical technical domain–Esperanto too small•“Idiosyncratic gap” still exists–Esperanto borrows from several language families–but unavoidable that many distinctions and ways of expression are left out–mur (French) - wall–muro (Italian, seen from outside), parete (seen from inside)Paper 22. Pros and Cons of the Pivot and Transfer Approaches in Multilingual Machine Translation. Boitet, C.•1.3 … And Always Very Difficult to Construct (Conceptual Decomposition/Enumeration)–Define small number of conceptual primitives and decompose all lexical items in terms of them–Conceptual dependency graphs will be huge–Use “subroutines” - conceptual enumeration–Japanese CICC project: 250,000 concepts–Construction process is non-montonic•new concept, revise dictionary for all languages–Difficult to see if concept already exists if its name is difficult to guess•“pros and cons” translated into another languagePaper 22. Pros and Cons of the Pivot and Transfer Approaches in Multilingual Machine Translation. Boitet, C.•I.2 Pure Pivot Structure Loses Information …–Extremely rare that two different terms or constructions of a language are completely synonymous–Unavoidable information useful for quality translation will be lost•2.1 … At the Lexical Level–wall -> wall seen from outside -> muro–wall (seen from outside) -> ???–muro -> wall –parete -> wall (distinction lost)•2.2 … At the Lower Interpretation Levels (Style)–One obtains paraphrases•Impossible to parallel styles as all trace of the source expression is lost•2.3 … At Non-Universal Grammatical Levels–“All or nothing” problemPaper 22. Pros and Cons of the Pivot and Transfer Approaches in Multilingual Machine Translation. Boitet, C.•II. Transfer Approaches–Avoid Pivot difficulties–1 -> many or many -> 1 situations•II.1 The Hybrid Approaches May Be Worse, Because the Square Problem Remains …–Lexical language-specific–Grammatical and relational symbols are universal–Big transfer dictionary needed•1.1 … If the Lexicons are Only Monolingual (CETA)–Grenoble group (CETA)–Hybrid pivot approachPaper 22. Pros and Cons of the Pivot and Transfer Approaches in Multilingual Machine Translation. Boitet, C.•1.2 … And Even If Some Part Becomes Universal (EUROTRA)–EUROTRA (1983) –9 languages–linguistic development scattered across 11 countries–transfer approach–part number approach for technical terms•II.2 Transfer Architectures Using m-Structures–Sequential or –Integrated approach using a multilevel structural descriptor•2.1 … Allow to Reach a Higher Quality–no universal notation for tense/aspect/modality–source language specific•2.2 … May be Preferable in 1->m Situations–Big firms - documentation produced in one languagePaper 22. Pros and Cons of the Pivot and Transfer Approaches in Multilingual Machine Translation. Boitet, C.•III. Both Approaches for the Future?•III.1 Pivot•1.1 Domain-Specific Pivots: New Applications?–CAD/CAM and expert systems: generation from knowledge base•1.2 Conceptual Decomposition/Enumeration a Challenge–EDR–Multilingual conceptual database (EuroWordNet?)Paper 22. Pros and Cons of the Pivot and Transfer Approaches in Multilingual Machine Translation. Boitet, C.•III.2 Transfer•2.1 Conversion from First to Second Generation–SYSTRAN (used in babelfish.altavista)–1G to 2G (?), see comments on CETA (pg.276)–Concepts dictionaries•2.2 Composition in n<->n Situations: The Structured Language Approach–Relay translation•4 Romance languages•4 Germanic