ITS: INTERACTIVE TRANSLATION SYSTEMAlan K. Melby, Melvin R. Smith, Jill PetersonBrigham Young UniversityProvo,At COLING78 we reported on an interactivetranslation system now called ITS, which useson-line man-machine interaction. This paperis an update on ITS with suggestions for futurework.Summary of ITSITS is a second-generation machine trans-lation system. Processing is divided intothree major steps: analysis, transfer, andsynthesis. Analysis is generally independentof the target language, and synthesis isnearly independent of the source language.The transfer step is dependent on both sourceand target languages. The intermediaterepresentation produced by analysis, adjustedby transfer, and processed by synthesis isdefined by Junction Grammar I and consists ofobjects called J-trees.The general characteristics of ITS,namely, modular design and a full syntacticanalysis as an intermediate representation,place ITS in the family of second generationsystems including the Montreal project and theGrenoble project.However, there are some further character-istics of ITS which must be mentioned to allowa better understanding of the system.The source texts for ITS are selected fromrather general English found in the publica-tions of the LDS Church. These publicationsare not at all restricted to theologicaldissertations. They include articles frommany authors in many styles on subjects rangingfrom gardening to short stories to descriptionsof foreign cultures to parent-child relations.The source texts for the Montreal project,on the other hand, consist of documents fromcarefully selected sub-language. Their firstcommercial system, the TAUM-METEO system,translates weather forecasts. Their currentproject, TAUM-AVIATION, is much more ambitiousbut is nevertheless restricted to manuals con-cerning the hydraulic systems of jet aircraft.aWhen a system is designed to translate aparticular sub-language, the system can takeadvantage of the properties of that sub-language.Returning to the general English textswhich are the specified input to ITS, we findthat there are no well-defined properties ofthe texts which can be used to reduce ambi-guity. This is why it was decided to useon-line man-machine interaction during theactual process of translation in ITS. ThisU.S.A.human interaction is, of course, expensive andso it was decided to share the burden of inter-action by analyzing the source text once andusing the result of analysis as the input toseveral transfer-synthesis pairs. This meansthat ITS is a one-to-many system. Beingone-to-many not only allows the overhead ofanalysis to be shared but also provides anelement of consistency in the output acrosslanguages. This is because all the targetlanguage texts are derived from the sameintermediate representation produced byanalysis.Thus, ITS is a second generation,interactive, one-to-many translation systemintended for general English text. Thecurrent target languages are Spanish,Portuguese, German, French, and (on a limitedbasis) Chinese.Status of ITSDuring the last two years, ITS has beendeveloped to the point where it has beentested for the possibility of commercial use.The dictionaries contain nearly thirty thousandwords, not counting idioms and other multi-wordunits. The system has recently translated aseries of test passages totaling over tenthousand words.The results of the tests were encouraging.The system was shown to be capable of trans-lating rather general text. Thanks to humaninteraction and an extensive English grammar,nearly all the sentences of the source textreceive a full Junction Grammar analysiswhich contains some semantic information beyondphrase structure; and the output, though farfrom perfect, is worth post-editing, and couldbe considerably improved with more dictionarywork.A major disappointment was the amount oftime spent on human interaction. It wasoriginally hoped that interaction could berestricted to analysis, since the overhead ofanalysis is shared by several target languages.However, it was found that some interaction wasalso required in transfer. The problem is thatinteraction in transfer is specific to onetarget language and is not shared. The resultis that the average amount of interaction perpage (~50 words) of text is currently aboutseven minutes for analysis (i.e. about thirty-five minutes shared by five languages), tenminutes for transfer interaction and fifteenminutes for post-editing. That makes a totalof about thirty minutes per page per language.These thirty minutes per page include a post-edit step by a human translator and so theUtah~424translation is roughly equivalent to a firstdraft translation by a human translator. Thetime spent per page by a human translatorvaries considerably but averages to approxi-mately one hour per page.Thus ITS appears to be somewhat fasterthan human translation and promises a degreeof consistency when a large document involvesseveral translators. However, an examinationof the total process of translation by manualmethods and by the current version of ITSresulted in a recent decision not to attempt acommercialization of ITS at this time. Somepoints to consider are the following:(I) Even though ITS may bethan manual translation, itenough to justify the effortand maintain the dictionariesof maintaining a computer to(2) The on-line interactiontrained operators.(3) Most translators do notmachine translation.(4) The scheduling of processing in a one-to-many system is rather tedious.Even though the current version of ITS isnot being commercialized, the authors remainoptimistic about the future of interactivemachine translation.Interactive translation will be the mostuseful in translating texts which are toogeneral or varied to be considered part of asub-language. The limits of fully-automatictranslation are well-known. To date, fullyautomatic translation has been shown to becommercially useful only when it is intendedto be merely indicative (e.g. Russian-EnglishMT at Rome Air Force Base) or when the systemis tailored to a sub-language (e.g. TAUM-METEO). If the need is for high-qualitytranslation of general text, the only possi-bilities seem to be (i) a highly successfullarge-scale AI approach, probably with aself-learning capability or (2) an interactiveapproach, with limited self-learning capabilityif possible.To be more specific on option (2), theauthors foresee a translator's work stationwhich would support two modes of operation.In one mode it would be a sophisticated buteasy to use word processor. Of course,multiple character sets would be available onthe video display and the typewriter qualityhardcopy device. The station would also con-tain a glossary which could be updated by thetranslator. The translator could even link upto a major terminology bank over the phone.slightly fasteris not yet fastrequired to buildand the expenserun ITS on.requires speciallyenjoy post-editingMartin Kay (Xerox) is currently working onsuch a station, with his own variations.In the second mode, the work stationwould be an interactive translation system.Source text could be entered directly from thekeyboard or the translator could insert adiskette containing the source text as it wasfirst entered on a word processor for publica-tion in the source language. Eventually, OCRmay be a practical input process for such astation. At any rate, after the source textis entered, the station would interactivelyresolve ambiguities and other problems in thetranslation. The interaction, to be attrac-tive, would have to average under ten minutesper page for a one-to-one translation and theoutput would have to be of such high qualitythat it could pass as a human translation withonly a few minor post-edit changes per page.The work station would also have to bereasonably priced (less than a compact auto-mobile).Finally, and most importantly, the workstation must work! That is, the station mustbe easy enough to use that the translator willwant to use it. The first mode (wordprocessor with dictionary lookup) must allowthe translator to produce a quality translationfaster than by manual methods. And the secondmode (interactive translation system) mustallow the translator to produce a qualitytranslation faster than by using a wordprocessor.When a work station such as the one justdescribed comes about, there may be fewtranslators who want to try it, even if itworks. Of course, if it does work, translatorswill have to have been involved in its develop-ment. But once a few translators venturevoluntarily to use it and find that it makesthem more productive and cost effective, thenthe pressure to use it will come from withinthe translator community, not from outside.This is one view of how the computer willbe used in translation. Rather than replacinghuman translators, computers will serve humantranslators. It agrees with Andreyewsky'sadvice to translators: "instead of fighting a'win/lose' battle with the machines, we mustwork toward developing an 'everybody wins'frame of reference. ''2The rest of this paper will consist ofseveral independent comments on some of thework on ITS the last two years which may beuseful or at least interesting to othersworking in machine translation. The commentswill concern word selection problems, inter-action, and linguistic programming languagesused in transfer..... 425-Word SelectionSelection of anthat is, replacementphrases is much moretext than in a sub-language. Several yearsago, it was hoped that interaction in analysisto select word senses would permit selectionof translation equivalents in transfer withoutfurther interaction. Instead it was foundthat the mapping between words in English andmultiple target languages is so variable thatno one set of English word senses willsatisfy all the target languages. To illus-trate this, two sample words will be considered:"run" and "line." Sample phrases taken froma corpus of LDS English will be given. Avariety of uses on a single word is manageablebut consider the problem of anticipating allpossible translations and how to distinguishthem on more than thirty thousand words.appropriateof sourcechallengingtranslation,words andin generalInteractionThe use of interaction is not a panacea.It is difficult to know when to ask a question.A major reason for the excessive interactionin the current version of ITS is the problem ofknowing when to ask a particular question. Howdoes one know if a particular ambiguity isgoing to make a difference in the translationor if the ambiguity will be preserved? Anotherproblem is that many sentences that are notambiguous to a human contain ambiguities whicha machine can resolve automatically only bysemantic tests. Semantic tests can be made insub-languages with some success, but it is noteasy. In general text, the increased varietymeans that semantic tests are much moredifficultexample,tem thatfollowingto make reliably. Consider, forthe problem of making a general sys-would naturally take care of theambiguities:Run"their provisions might run short""the train continued its run""I tried to run the pleasant thoughts of theday through my mind""she ran her fingers through the boy's hair""time had run out""someone opened the door and the dog ran out""sometimes chicken and sheep run in the street""he would wait until called to run an errand""how to run the organization""it ran its course""the bull ran its horns into the pumpkin""seeing the wave come, they ran into the trees""the bitterness ran deep""sometimes meetings ran late""confusion ran rampant""water runs faster"no running water""he took a running"shirt and runningLine"the line of march""line of communication""line of authority""front line of attack""over the end line but not between the goals""the next line or ancestor in question""thirty people stood in line""born of a royal line""the first line of a story""the bottom line (conclusion) is clear""draw a crazy (crooked) line on a piece ofpaper""they would be taught line upon line""the hard leathery lines of grandmother's face""those production lines"at the surface"start"shorts"(i)(2)(3)(4)Thebuta rule specific to "falling?"Hopefully, AI techniques will someday beable to resolve such ambiguities in large scalesystems as well as they now do in restrictedtests systems. Indeed, major advances in AImay well be essential to acceptable functioningof the work station described above, but thereis enough of Joseph Weizenbaum in me to believethat even in the long run human interactionwill be needed to produce high quality raw out-put in the machine translation of general text.A reasonable question would be: "why notguess on the unsure points during translationand let the post-editor clean them up?" Oneanswer is that raw output must be close tofinal form or the post-editor will either wantto start over or will not post-edit up to highquality. In other words, a post-editor willonly improve a text by a certain increment.426The fact that the man knew surprised us.(the fact which he knew vs. the fact ofhis knowing)I saw her shaking hands.(...her hands were shaking vs. she shookhands with someone)I made my brother a statue.(I made a statue for him vs. I made himinto a statue)Beware of falling victims to error.(beware of becoming victims or beware ofvictims that are falling, cf "beware offalling rocks")last example is not ambiguous to a humanhow does one resolve the ambiguity withoutThis has been determined in experience withmanual human translation being post-edited byanother human. Another answer is that inter-action may well point out ambiguities whichcould be missed by a human translator--particularly a native of the target language.An empirical method of determining theappropriate mix between interaction and post-editing is to have the program attach toeach potential interaction a guess with a con-fidence level. Then the translator sets aninteraction level. Then interactions willnot occur if the confidence level exceeds theinteraction level. This technique has beenpartially implemented in ITS and seems informa-tive.Transfer LanguasesFor those interested in comparing systems,we include two sample entries from the transferdictionary of ITS 3 and two sample entries fromthe transfer dictionary of the TAUM-AVIATIONsystem. 4ITS iADJOIN1 &s map2 ser conti/guo com \[l.v3 unir .v4 &w("l") c =pa;ser;conti/guo;com=PA1 &s p verb;pa;prep2. /* verb becomes copula with pred adjmodified by PP */3 &s *,f (,ad) verb;pa4 &s m (,ad)5 &s f,do (do6 &t *,t (ad,,,do) ;prepDESCRIPTION:(ADJOIN)1-3: Interact on the two translation choicesgiven.4: When "i" is returned from the MAP actioncode ("ser conti/guo eom" was chosen), calldictionary subroutine = PA, passing the 3parameters shown.(=PA)i: Associate the 3 names given with the para-meters passed.2: Comment3: Build a predicate adjective, setting theverb and predicate adjective to the values("ser" and "conti/guo") of the respectiveparameters.4: Move any insertions on the verb to thepredicate adjective.5: Find the direct object of the verb (ifany).6. If direct object is found then build aprepositional phrase modifying the predicate-427adjective, set the preposition, and make thedirect object the object of the preposition.J-TREE BEFORE:/½V + Nadjoin thebuildingJ-TREE AFTER:vserconti/guocom thebuildingITS 2AFTER1 &s f,do (,do2 &t &do/* preposition */3 &w(f,*,f,sxt,v, (do)) s depois/* obj isfull SV */4 &w(f,*,f,v (dO)) s depois de/* obj is PV*/5 &o map/* interact */6 depois de .w~time>7 atra/s de .w8 apo/s .w9 &endi0 &e map/* adverbial particle */ii depois .w12 atra/s .w~lace)DESCRIPTION:I: Find prepositional object.2: If found, then execute &do-group.3: When object is a noun clause, set "after"to "depois".4: When object is a gerund, set "after" to"depois de".5-8: Otherwise, interact on the three trans-lation choices given (utilizing disambigu-ation strings if present).9: End of &do-group.10-12: If object was not found, interact onthe two translation choices given (checkfor disambiguation strings).NOTE: This transfer is referential only andno structural manipulation is performed.TAUM I#FILTER# ==(*C.LAINE, LE 6 DECEMBER 1978. "FILTERED"EMPLOYE AVEC UN NOM DE PIECE DEVRAIT SETRADUIRE PAR "A FILTRE". POUR LE MOMENT NOUSALLONS LE RENDRE PAR "MUNI D'UN FILTRE"*)DEBUTSI NATURE (FC) EST V ALORSDEBUTSI PARCOURS/CV/GV GPREP CP P SEPO ALORSTRADUIRE EPO PAR #BOF#;SI PARCOURS/CV/GV/GOV/PH TELQUE TRAITS(ICI) =\[APRENOM\]$EPH ALORSDEBUTTRADUIRE FC PAR #MUNIR# \[Bi9\]:A1 := GP(GPREP(CP(P UT TRADUITE #DE#)),GN(CN(N UTTRADUITE # FILTRE#), DET(CART(ART UTTRADUITE #UN#)))):DEPLACER A1 EN OBJI\[i\] SOUS EPHFINSINONDEBUTTRADUIRE FC PAR #FILTRER#\[B6\];TRAITS EN FAC := TRAITS (FC)+\[TNOREF\]FINFINSINONTAUM 2#PRESENT#==(* S.L.,8.01.79. TEST: SI ADJ EST PRENONIMALTRADUIRE PAR "PRESENT", SINON TRADUIRE "S BEPRESENT" PAR "IL Y AVOIR X". CETTERESTRUCTURATION EST COMPLEXE. CF. LE CAHIERDES ADJECTIFS POUR UNE REPRESENTATION DE LARESTRUCTURATION. *)VAR Ai:UT; A2:ARBRELIBRE; A3:ARBRELIBRE FINDEBUTSI NATURE (FC) EST ADJ ALORSDEBUTSI PARCOURSE /CADJ /GV /GOV /PH TELQUETRAITS(ICI) =\[AREL\] /GN ALORS(* CE TEST SERT A SAVOIR SI L'ADJECTIF EST ENPOSITION PRENOMINALE. *)DEBUTSI PARCOURS /CADJ $ECADJ /CONJ GV $EGV/CONJ GOV (OPS CONJ (COPTELQUE (UT(ICI) =#BOF#) OU (UT(ICI)=#BE#)$ECOP)) $EGOV/CONJ PH $EPH ALORS(* ON NE TESTE PAS SI PH EST CONJOINT, CELAN'EST PAS PERTINENT; ON NE TESTE PAS SI GVDOMINE DES MODIFIEURS, PUISQUE CEUX-CI SONTLES MEMES POUR LES ADJECTIFS ET LES VERBES.*)DEBUTAi:=#V# #AVOIR#\[Bi\] ;A2:=CV (V UT TRADUITE AI) :REMONTER FC EN EGV ;DEPLACER A2 EN CV SOUS EGV ;EFFACER ECOP ;EFFACER ECADJ ;A3:=GP (GPREP (CP(P UT TRADUITE #BOF# ),EN SUBS GN(PRON UT TRADUITE #1L#)) ;SI PARCOURS EPH SUJ $ESUJ ALORSDEBUTSI PARCOURS ESUJ _ OBJD( ) $#OBJDALORSDEBUTINSERER CIRC(EN GP DEPLACE (EOBJD))EN CIRC\[i\] SOUS EPH;COPIER ESUJ EN OBJD SOUS EPH ;DEPLACER A3 EN SUJ SOUS EPH ;FINSINONDEBUTCOPIER ESUJ EN OBJD SOUS EPH ;DEPLACER A3 EN SUJ SOUS EPH ;FINFINSINONDEBUTTRADUIRE FC PAR UT(FC) ;ECRIRE SOMMET ;FINFINSINONDEBUTTRADUIRE FC PAR UT(FC) ;ECRIRE SOMMET ;FINFINSINONDEBUTTRADUIRE FC PAR #PR2ESENT#\[Fi,Pi\] ;FINFINSINON(*C. LAINE, LE 27 DECEMBRE 1978")SI NATURE (FC) EST V ALORSDEBUTTRADUIRE FC PAR #PR2ESENTER# \[B6\]FINSINONDEBUTTRADUIRE FC PAR UT(FC) ;FINFIN.(* FIN DE #PRESENT# *)428Footnotesi.Junction Grammar is described in:a. LYTLE, Eldon G.Structural Derivation in Russian1971 unpublished Ph.D. dissertationUniversity of Illinois at Champaign-UrbanaLYTLE, Eldon G.A Grammar of Subordinate Structuresin EnglishThe Hague: Mouton & Co., 1974LYTLE, Eldon G."Junction Grammar: Theory andApplication"in the 5th LACUS FORUM (1979)Hornbeam Press, Inc., 1980d. "Junction Grammar Bibliography"in JTA Fall 1977Translation Sciences Institute,Brigham Young University,Provo, Utah, U.S.A.ANDREYEWSKY, Alexander"Whither Automation and the Translator"in The ATA Chronicle, April/May 1980.RICHARDSON, Steven"Transfer Systems"M.A. thesisBrigham Young University, 1980Linguistics DepartmentTAUM ReportJune 1979Montrealb.e.2.3.4.--429--