C SC 620Advanced Topics in NaturalLanguage ProcessingLecture 194/6Reading List• Readings in Machine Translation, Eds. Nirenburg, S. et al. MIT Press2003.• Reading list:– 12. Correlational Analysis and Mechanical Translation. Ceccato, S.– 13. Automatic Translation: Some Theoretical Aspects and the Design of aTranslation System. Kulagina, O. and I. Mel’cuk– 16. Automatic Translation and the Concept of Sublanguage. Lehrberger, J.– 17. The Proper Place of Men and Machines in Language Translation.Kay, M.Paper 17. The Proper Place of Men andMachine in Language Translation. M. Kay• Introduction– Fully Automatic High Quality Translation (FAHQT) not realistic– Main point:• There is a great deal that computer scientists and linguists couldcontribute to the practical problem of producing translations, but, intheir own interests as well as those of their customers, they shouldnever be asked to provide an engineering solution to a problem thatthey only dimly understand• By doing only what can be done with absolute surety and reliabilitynow and by going forward from there in short, carefully measuredsteps, very considerable gains can be virtually guaranteed to allconcernedPaper 17. The Proper Place of Men andMachine in Language Translation. M. Kay• Machine Translation and Linguists– Translation for a pronoun– Target sentence pair:• Since the dictionary is constructed on the basis of the text that is beingprocessed, it need refer to only a small amount of context to resolveambiguities• Since the dictionary is constructed by a native speaker of the language,he need refer to only a small amount of context to resolve ambiguities• Il (French) -> {it, he}– Pronoun ambiguity• Ad hoc solution (case-by-case, statistical)• Not solved– Great number of ad hoc solutions built into existing MT systemsand any enhancement in the future will require more and more ofthe samePaper 17. The Proper Place of Men andMachine in Language Translation. M. Kay• Machine Translation and Computer Science– Linguistic perspective: Unlikely to be adequate engineering wherewe know there is no adequate science– Programming style and techinque• Efficiency important - large quantities of text to be handled• Low level programming languages used (1980)– More likely to be ad hoc– Efficiency gains at most linear• To be continually improvable, program needs to be perspicuous (nodoubt about the role that each of its parts play) and robust (can bechanged in important ways without fear of damage)Paper 17. The Proper Place of Men andMachine in Language Translation. M. Kay• The Statistical Defense– Linguistics requires of its practitioners remarkable virtuosity inconstructing examples of problems such as no existing or proposedcomputer system could possibly solve– The claim is that we do not have to solve them so long as they donot crop up very often• May not have an algorithm that will identify the antecedent of apronoun whenever a human reader could, but if it can devise a methodthat will identify it most of the time, that will be good enough– Algorithm that works most of the time is of very little use unlessthere is some automatic way of deciding when it is and when it isnot working• If algorithm could draw the proofreader’s attention to all cases ofpronominal reference that were in doubt, and if a high proportion ofthe cases were known to be correctly handled, then the utility of thetechnique is clearPaper 17. The Proper Place of Men andMachine in Language Translation. M. Kay• The Statistical Defense– If program works 99%, 90% , 80% or 50% of the time but we cannot tellwhich cases, the amount of work left to the repairman is essentially thesame.– Real situation is worse: many decisions of the same difficulty must bemade in the course of translating a single sentence. If there is reason toexpect each of them to be correct 90% of the time, there need only beseven of them in a stretch of text to reduce the expectation to .97=.48• The Sorcerer’s-Apprentice Defense– Incomplete theories is a worse base to build systems than no theories– The man looked at the girl with the telescope– French admits parallel ambiguities in PP attachment– The man looked at the girl with penetrating eyes– aux yeux or de ses yeux, avec yeux not acceptablePaper 17. The Proper Place of Men andMachine in Language Translation. M. Kay• The Sorcerer’s-Apprentice Defense– Ils signeront le document pourvu que leur gouvernement accepte– (I)• They {will, are going to} sign the document{ supplied, furnished, provided}that their government accepts– (II)• They {will, are going to} sign the document{provided that, on condition that,only if} that their government accepts– Intersect (I) and (II)• The Translator’s Amanuensis– transcriber• Text Editing• Translation Aids– Dictionary– Idioms (frequently occurring items)Paper 17. The Proper Place of Men andMachine in Language Translation. M. Kay• Machine Translation– Translator under the tight control of a human translator– Human/machine collaboration– Consult the user– It is far better that the labor and ingenuity spent ondeveloping the machine’s ability to make bad guessesshould be employed more productively– Cascading errors common in machine-only systems– Keep track of words and phrases used in some specialwayReadings• Readings in Machine Translation– 19. Montague Grammar and Machine Translation. Landsbergen, J.– 20. Dialogue Translation vs. Text Translation – InterpretationBased 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 inMultilingual Machine Translation. Boitet, C.– 31. A Framework of a Mechanical Translation between Japaneseand English by Analogy Principle. Nagao, M.– 32. A Statistical Approach to Machine Translation. Brown, P. F. etal.• Papers from recent MT Summit