pragmatics and translation - sepln

22
Page 1 In this paper, we present an approach to machine translation rooted in pragmatic 1 . It is an approach which relies on: the use of context to interpret source language utterances and to produce target language correlates, a context of the utterance which crucially includes nested beliefs environments which are constructed and modified through ascription during processing, knowledge of the world which is accessed for constructing or modifying the context, context-sensitive (non-monotonic) inferencing within the context to resolve ambiguities during interpretation or to select expressions during production. In Section 1, we will begin the discussion by motivating the need for a pragmatics- based approach which takes into account the beliefs of the participants in the translation process by focussing on one example. In Section 2, we show that such examples are not isolated instances but pervade translations. We show that there are patterns of differences in multiple translations of the same text which are related to different global interpretations of what the text itself is about or to differing world views on the part of the translators. In Section 3, we present a number of concepts related to translation from a pragmatics-based perspective and suggest a possible computational framework for implementing such an approach. In the final section, we conclude by discussing the implications of a pragmatics-based approach for translation analysis, translation evaluation and for future directions in machine translation research. This discussion represents a summary of work presented in more detail in Farwell and Helmreich (1993, 1995, 1996, 1997 and 1998). Section 1 In this section, we argue for a pragmatics- based approach to machine translation. The need for pragmatics in Natural Language Processing and Machine Translation has long been recognized (Bar-Hillel 1960, Wilks 1975, Nirenburg et al. 1992). Yet pragmatics (as the study of language in context) is often difficult to distinguish from semantics (the study of the connection between the language sign system and the world it represents) 2 . The history of Machine Translation shows a progression towards systems containing more and more knowledge in order to represent the meaning of the source text. We take as a 1. Work on this project was supported under DOD grant: MDA904-92-C-5189. 2. Many theorists such as Pustejovsky (1991), Lakoff (Lakoff & Johnson 1980, Lakoff 1987), or Nirenburg (1987) appear to us to coalesce language meaning with language use and thus are interested in semantics. Other theorists such as Fauconnier (1985), Green (1984), or Nunberg (1978), place emphasis on the issues of lan- guage use and appear to us to attempt to sub- sume semantics within pragmatics. Pragmatics and Translation David Farwell Stephen Helmreich Computing Research Laboratory New Mexico State University Las Cruces, New Mexico

Upload: others

Post on 22-Feb-2022

2 views

Category:

Documents


0 download

TRANSCRIPT

Page 1

In this paper,we presentan approachto

machine translation rooted in pragmatic1.It is an approach which relies on:

• the use of context to interpret sourcelanguage utterances and to producetarget language correlates,

• a context of the utterance whichcrucially includes nested beliefsenvironmentswhichareconstructedandmodified through ascription duringprocessing,

• knowledge of the world which isaccessed for constructing or modifyingthe context,

• context-sensitive (non-monotonic)inferencingwithin thecontextto resolveambiguities during interpretation or toselect expressions during production.

In Section1, we will begin the discussionby motivating the need for a pragmatics-basedapproachwhich takes into accountthe beliefs of the participants in thetranslation process by focussing on oneexample.In Section2, we show that suchexamples are not isolated instancesbutpervadetranslations.We show that thereare patterns of differences in multipletranslationsof the same text which arerelatedto different global interpretationsofwhat the text itself is aboutor to differingworld views on the part of the translators.In Section 3, we present a number ofconcepts related to translation from apragmatics-basedperspectiveandsuggesta

possible computational framework forimplementingsuchanapproach.In thefinalsection, we conclude by discussing theimplications of a pragmatics-basedapproach for translation analysis,translation evaluation and for futuredirections in machine translation research.

This discussionrepresentsa summary ofwork presentedin more detail in FarwellandHelmreich(1993,1995,1996,1997and1998).

Section 1

In this section,we arguefor a pragmatics-basedapproachto machinetranslation.Theneed for pragmaticsin Natural LanguageProcessingand Machine Translation haslong been recognized (Bar-Hillel 1960,Wilks 1975, Nirenburg et al. 1992). Yetpragmatics(as the study of languageincontext) is often difficult to distinguishfrom semantics(thestudyof theconnectionbetweenthe languagesign systemand the

world it represents)2. The history ofMachine Translationshows a progressiontowardssystemscontainingmoreandmoreknowledge in order to represent themeaningof the sourcetext. We take as a

1. Work on thisprojectwassupportedunderDODgrant: MDA904-92-C-5189.

2. Many theorists such as Pustejovsky (1991),Lakoff (Lakoff & Johnson 1980, Lakoff 1987),or Nirenburg (1987) appear to us to coalescelanguage meaning with language use and thusareinterestedin semantics.OthertheoristssuchasFauconnier(1985),Green(1984),or Nunberg(1978), place emphasis on the issues of lan-guage use and appear to us to attempt to sub-sume semantics within pragmatics.

Pragmatics and Translation

David FarwellStephen Helmreich

Computing Research LaboratoryNew Mexico State University

Las Cruces, New Mexico

Page 2

point of departure the Pangloss/Mikrokosmosproject which representsthemost consistent, large-scale attempt todevelop a knowledge-based,interlingualmachine translation system to date(Pangloss 1994).

As might be expected,various pragmaticissues have been addressedwithin thePanglosssystem. The Mikrokosmos MTengine, for example, monitors and usesinformation about speakersand hearers,discoursestructure,and world knowledgein constructing its interlingualrepresentationof an input text (Attardo1994, Carlson & Nirenburg 1990). It canhandle metonymic constructions andmetaphors (Onyshkevych & Nirenburg1994).However,we suggestthat it is still asemantics-basedsystem. Even thoughcontextual information and world-knowledgeare both representedand usedinferentially, the goal is to produce asemantic representation of the text.

Language,however,is usednot simply toreporteventsin theworld. It is alsousedtoconvey the rich mental models thatindividualsandculturesbring to bearon thecommunicationprocess.It is the claim of apragmatics-basedapproachthat textsdo nothavemeanings,but ratherthat in producingtexts,peopleintendmeanings.Thus,a textcan only be approached through aninterpretation.That is to say, the translatorattemptsto understandtheauthor'sintent increating the source text for the originalaudienceand then recreates,to the extentpossible,that intent for the targetaudience

using the target language3.

This pragmatics-basedapproach, then,focuseson providing interpretationsof atext that insure a coherentaccountof theintent behind the text. Such an approachrelies heavily on representationsof thebeliefs and other mentalattitudes(suchasexpectations,hopes,likes, and dislikes) ofthe participants in the communicativeprocess:the author,the translator,andtheir

respectiveaudiences(the addressees).4 Inparticular, then, we require a systemthatcan model:

• the translator's beliefs,• thetranslator'sbeliefsabouttheauthor's

beliefs,• thetranslator'sbeliefsabouttheauthor's

beliefs about the (source language)addressee's beliefs,

• the translator's beliefs about the (targetlanguage) addressee's beliefs.

Such systemmodels can be found in thework on ViewGen(Ballim & Wilks 1991),a systemwhich constructsbeliefsspacesforany numberof agentsaboutany numberoftopics.In addition,we find a computationalbasis for the required inferencing in thework associated the ATT-Meta system(Barndenet al. 1994).This systemincludesa default inference mechanism, anepistemic logic involving four possiblepositive truth conditions (certain, default,probable,andpossible),a truth maintenancesystem,metaphoricalpretensecocoons(fortreatingmetaphorsas true), and simulativereasoning(for inferencingwithin embeddedbelief states).

Motivating Pragmatics-basedTranslation

For the sakeof expositionwithin a limitedspace, we motivate the need for apragmatics-basedapproachby way of acase involving the translationof a single

3. Wedonotsuggestthisasapsychologicalmodelof translation. Translation, we believe, is an artor skill and not a faculty (in the sense of Fodor1984) and can therefore be approached in anynumber of different ways. We suggest only thatthis is a plausible and appropriate model forcomputational representation.As is obvious here and throughout this article,the influence of the work of Grice (1975) isclear. Implicaturesandinferencesfrom "what issaid" are as fully part of the communication asthe semantic content of the utterance itself.

4. In this paper we will use the term “belief” torefer to any epistemic state. We do not distin-guish (except by degree of commitment)between beliefs, knowledge, hypotheses, andthoughts. For us, they are all beliefs, though ofcourse, held with different degrees of tenacity.

Page 3

Frenchnounphrase,takenfrom a subtitledmotion picture, Jesusof Montreal (Arcand1989). We havedevisedall the alternativetranslationsexceptfor the original subtitle.We discuss three alternative translationswhich stemfrom (a) alternativeviewsof thebeliefs of the film's audience (theaddresseesof the translation) and (b)alternativeviews of the translator'sbeliefsabout the world.

We will simplify mattersby assumingthatthe translator is sitting beside usinterpreting the film as it develops,eachcharacteran independentagent.Thus, forany given utterance,thereare four relevantparticipants:theactorwho speaks,theactorwho is addressed,the translator and theaudience(theaddresseesof the translation).We also ignore the additionalcomplexitiesinvolved in thefact thatbothactorsarealsospeakingin some senseon behalf of theauthor of the screen play to the film’soriginal audience.

As background(i.e., thatpartof thecontextof which theprotagonists,thetranslatorandthe film's audience are all aware), weprovide the following synopsis.

A priest at a shrine outsideMontreal hasbeen sponsoringa religious drama everysummer for 35 years. Since the text hasbecome somewhat outdated, he asksCoulombe,a youngactorwho hasrecentlyreturned from an extended sojourn, tomodernizethe script andto play the part ofJesus.He agreesandimmediatelysetsaboutlooking for collaborators. The priestsuggeststhat Constance,an old friend ofCoulombe’s,would be a good personforCoulombeto enlist in his endeavorand sohe seeksher out. Sheagreesto work withhim, and, in passing,invites him to stay ather apartment. He agrees.

In a later scene,the sceneof our attention,Coulombereturnsearlierthanusualto whatheassumesis anemptyapartment.He startsto makehimself comfortable,makingsomenoisein the process.At this point he hearssomeone moving about in Constance's

bedroomand, suddenly,sheemergesfromwithin, closing the door behind her. Shesays,T'esdéjà là, toi? (Back already?)andthen,coughingsignificantly,saysto herself,Bon... (Okay...). At this point Coulomberealizesthat theremay be someoneelseinthe bedroomand whispers,Tu veuxque jem'enaille? (ShouldI go?).Sheshakesherheadno, laughsnervously,opensthe doorand says to whomever is inside, Ben,écoutes,sors (Come on out), On va pasjouer une scènede Feydeau(This isn't abedroom farce).

It is this last utteranceand its subtitle thatwe wish focus on. The translator whoprovided the subtitles for the film hasglossed On va pas jouer une scène deFeydeauasThis isn't a bedroomfarce. Thisis not the only possible translation, ofcourse,andwe look at alternativesso astodraw out the underlying assumptionsthatdetermined the translator's choices.

At the time of Constance'sutterance,theprotagonists,thetranslatorandtheaudiencehave the following beliefs (among others).

Coulombe is living in Constance'sapartment.

They are clearly close friends andcolleagues.

Coulombe has entered the apartmentunexpectedly early.

It is still mid-afternoon.He accidentally makes a loud noise.Constance emerges from her bedroom

dressed in a nightgown and closes thedoor behind her.

She is somewhat flustered byCoulombe's unexpected presence.

Coulombe believes there is someoneelse in Constance's bedroom that hehas caught them in a compromisingsituation.

Coulombe believes that Constance andthe other person might prefer someprivacy.

Coulombe believes that if Constance andthe other person might wish to keepthe identity of the other secret.

Constance believes that Coulombebelieves that she and the other have

Page 4

been caught in a compromisingsituation.

Constance believes that Coulombebelieves that they might prefer someprivacy.

Constance wishes to changeCoulombe's belief.

Constance tells Coulombe not to leave.Constance tells the person in the

bedroom to come out and show him/herself.

Without going into the details, the analysisof the utterance begins by establishing, onlinguistic grounds, that Constance is usingon va pas jouer to express to the person inthe bedroom that she does not wish thecurrent situation (such as Coulombe'sdiscovery of her and the unknown personalone together in her bedroom) to beunderstood as being of a to-be-determinedtype of play (i.e., we'renot playing..., we'renot going to play ...). The next step is toassign an interpretation to un scène deFeydeau. Again, on linguistic grounds,coupled in this case with knowledge of theworld, we establish that Constance is usingun scènede Feydeauto refer to the type ofsituation that might be used as a scene in aplay by the 19th century French playwrightFeydeau who wrote bedroom farces (i.e., ascenefrom a bedroomfarce). To arrive atthis interpretation, it must be the case that:

Constance believes Feydeau is aplaywright and that Feydeau wrotebedroom farces.

Constance believes the person in herbedroom believes Feydeau is aplaywright and that Feydeau wrotebedroom farces.

The interpretation is completed byconfirming that the situation underdiscussion (i.e., Constance and someonealone together in her bedroom) is indeedone that Feydeau may have written about.This becomes especially plausible when itturns out that the man in Constance's roomis a priest, in fact, the very same priest whohired Coulombe to update the play.

Having arrived at an interpretation, thetranslator now needs to provide anequivalent expression for an Englishspeaking audience. To express that somecurrent situation is not of some type, he/sheselects the expression This is no ... or Thisisn't ... or some such English equivalent. Asfor a situation typical of a bedroom farce ofthe sort that Feydeau might write about,e.g., two people getting caught in acompromising position by a significant

other,5 the translator checks his/her beliefsabout the audience. If it is assumed that theaddressee of the translation would nottypically believe that Feydeau is aplaywright or that Feydeau wrote bedroomfarces, quite possible for those unfamiliarwith French culture or with the theater, thenreference to Feydeau will fail to have theintended effect and some alternativeexpression must be chosen, e.g., a bedroomfarce or a scene from a bedroom farce.

This leads to the first case of variation intranslation stemming from variations in thetranslator's beliefs, namely, those based onvariations in the beliefs the translatorattributes to the addressee of the translation,the non-French speaking audience of thefilm. In the event that the translator assumesthat the film's audience has the same beliefsabout Feydeau as the speaker (Constance)and the addressee (the unknown person inthe bedroom), he/she would most likelytake advantage of those beliefs to provide atranslation that more closely approximatesthe source language utterance in form andcontent, relying on the addressees of the

5. There appear to be two scenarios that could bedrawn from a bedroom farce. One involves, assuggested above, being caught in a compromis-ing situation by a jealous husband or lover andthus enduring an unpleasant scene. The otherinvolves being apprehended by a moral or socialarbiter, so that some unacceptable behavior isexposed. These two results: jealous rage orexposure, usually result (at least temporarily inbedroom farces) in public shame and humilia-tion, and in possible long-term negative socialconsequences such as divorce or loss of posi-tion.

Page 5

translationto usethosebeliefsappropriatelyto interpretConstance'sutterance.That is, ifthe translatorassumesthe addresseesof thetranslationbelievethat Feydeauis a Frenchplaywrightwho wrotebedroomfarces,thenhe/shewould most likely haveglossedthe

utterance,on the basisof the rhetic act,6 asThis is not a scenefrom Feydeau. In fact, itappearsthat the translatorassumesthat thenon-Frenchspeakingaudienceof the filmhasbeliefsaboutFeydeaudifferentfrom thebeliefs of Constanceand the unknownpersonin the bedroom,or, perhaps,hasnobeliefs about Feydeau at all. Thus, thetranslatoravoidsany expressionthat wouldrely on suchbeliefsfor the interpretationofthe translationandglossesthe utterance,onthe basis of the illocutionary act, as Thisisn’t a bedroom farce.

A secondcaseof translationvariationbasedon variationsin thebeliefs-contextconcernsvariationsin the beliefsof translatorsaboutthe world. It is possible,for instance,thatthe translator did not have the necessarybeliefs about Feydeau to work out theintendedinterpretation.This is, of course,the baneof translators:lack of the relevantknowledgeof the world, and it is far morecommonthanappreciatedevenin extremelymundanediscoursesuch as generalnewsarticles. It need not, however, detertranslation, even appropriate translation.When there is a lack of knowledge, thetranslatormust fall back to an even moreabstractlevel of interpretationasa basisforthe translation,that of the functionaleffectof the utterance. (Or, of course, thetranslatorcould fall backto a translationofthe rhetic act, hoping that his addresseeswill beableto interpretwhathe cannot.)Inour case, Constance is informing theunknown person in the bedroom thatCoulombewill not be scandalizedby theirliaison and that the unknownpersonin thebedroomcansafelyshowhimself.This maybe accomplishedby glossing On va pasjouer unescènedeFeydeau, on thebasisofthe intended perlocutionaryeffect of theutterance as There's nothing to worryabout; it's safe. This strategy,however,canbe ratherdangeroussincethereis little datato aid in identifying the specific functionaleffect.

Summary

We believethat this exampledemonstartesthat an approach to language rooted inpragmatics offers significant advantagesover an approachthat seespragmaticsassimply an additional component of thelanguagesystem.In applyingthis approachto thefield of languagetranslation,we haveshown how such an approachcan provideexplanations of the many possibletranslations for the same text.

6. We distinguish three levels of interpretation,whichweidentify looselywith threeof Austin’sspeech acts (Austin 1962). At the first level, theintent of the speaker is to utter words of a lan-guagewith amoreor lessdefinitesenseor refer-ence (p. 95). This is the rhetic act, and it isinterpreted correctly if the hearer can identifythe senses and references intended by thespeaker. This is essentially what is normallycalled the semantic content of the utterance.At a secondlevel of interpretation,the hearerattemptsto understandthe actualmessagethatthespeaker wishesto communicate,which mayor, morelikely, is not identicalwith thesemanticcomtent,but rathertheresultof the interferencefrom this content and other premisesdrawnfrom assumedknowledgeor context. We iden-tify this loosely with the illocutionary act, thecommunicative intent of the utterance.Finally, at thebroadestlevel, thereis anintenttoproducea changein the hearer. This changeistheperlocutionaryeffect of theutterance,andisan interpretationof what thehearerunderstandsto bethepurposeor goalof thespeaker in mak-ing the utterance.In the utteranceat hand, the semanticcontent(rhetic act) is that we are not playing a scenefromFeydeau.Thecommunicative content(illo-cutionary act) is to tell the addresseethat thecurrentsituationis notonein whichheneedfeardisclosureof his identity. The perlocutionaryintent of the utterance is to encouragetheaddresseeto comeout of thebedroomby assur-ing him that it is safe to do so.

Page 6

Section 2

In this section we present certain conceptsrelated to a pragmatics-based approach toMachine Translation and sketch out aninformal processing model. We begin bydeveloping notions of discourse context andutterance context and then outline a two-stage translation process involvinginterpretation and translation. Next, weintroduce the notion of "user-friendlytranslation" and, finally, conclude with adiscussion of translation equivalence.

Background

As a context for the discussion, we considertwo differing translations into English of eltercer piso and el segundopiso in thefollowing Spanish sentence taken from anews article about the Moscow real estatemarket in the early 1990’s.

... los 300 metros cuadrados del tercerpiso estaban disponibles pero fueronaquilados ..., sólo quedael segundopiso ....

While one translator has rendered theseexpressions as the third floor and thesecond floor respectively, another hasrendered them as the fourth floor and thethird floor. Although these two translationsare clearly different, they are, in fact, bothaccurate and they are not necessarilylogically inconsistent. The reason resides inthe differing beliefs the translators haveabout the beliefs of the author andaddressees of the Spanish text, and of theaudience of the English translation.

We hypothesize that first translator assumesthat the author of the text shares thetranslator's floor naming convention (say,using ground floor, first floor, etc. for thelevels of a building as opposed to first floor,secondfloor, etc.) and that the addresseesof the translation also share the translator'sfloor naming convention (though it is alsopossible that the author and addresseesshare a convention which the translatordoes not). Thus, the first translator refers to

the fourth level above ground as the thirdfloor and the third level above ground asthe secondfloor. If those assumptions arecorrect (and we do not have the crucialinformation to determine this), then the firsttranslator’s translation is equivalent at boththe level of the rhetic (semantic) andillocutionary (communicative) actst.Otherwise, the translation will be lessequivalent since the addressees of thetranslation will have to access informationabout alternative floor naming conventionsand make the appropriate inferences inorder to arrive at the author's intendedmeaning.

We hypothesize that the second translatorassumes that either the author does notshare the translator's floor namingconvention (say, using first floor, secondfloor, etc. vs ground floor, first floor, etc.)or, alternatively, the addressees of thetranslation do not share the author’s floornaming convention. Thus, the secondtranslator refers to the fourth level aboveground as the fourth floor and the thirdlevel above ground as the third floor. Ifeither of those sets of assumptions is correct(again, we do not have the crucialinformation to determine whether they are),then the second translator's translation isequivalent at least at the illocutionary levelof communicative content. Otherwise, thetranslation will be less equivalent since theaddressees of the translation will have toaccess information about alternative floornaming conventions and make theappropriate inferences in order to arrive atthe author's intended meaning.

Beliefs and Inferencing

As a framework for modeling thetranslation process, we adopt a beliefsascription mechanism (such as that ofBallim & Wilks 1991) for constructing therelevant, recursively embedded beliefsspaces of the participants in the translation,ontologies (such as those of Nirenburg etal. 1995) for representing the beliefs withinthese spaces, and a default inferencing

Page 7

engine (such as that of Barnden et al. 1994)for carrying out the inferencing over thesebeliefs within these spaces.

Following Ballim & Wilks 1991, beliefsabout a topic are represented as a boxlabeled in the upper left-hand corner.

Beliefs of an agent about a topic arerepresented as a box labeled in the lowercenter which, in turn, contains a topic box.

The basic default rule for ascribing beliefsis:

unless there is specific evidence to thecontrary, agent1 ascribes its beliefs toagent2.

Evidence to the contrary consists ofpreexisting beliefs of the target agent thatare contradictory to or inconsistent with thebeliefs being ascribed. We indicate theapplication of this rule by an arrow pointingfrom the source agent’s environment to thetarget agent’s environment. In Figure 1,

Agent1’s beliefs about a topic have beenascribed to Agent2.

Following Barnden et al. 1994, we furtherassume the existence of default (defeasible)inferencing engine of the sort used in ATT-Meta. This reasoning mechanism, whengiven a goal to prove, will evaluate allevidence chains that it can find for both thegoal and its negation. Each step in the chainis given an evidence status (such as certain,default, possible). The evidence for both thegoal and its negation receives such a statusand a resolution procedure then determinesthe evidentiary status of the goal. In thefollowing example, the rule has a defaultstatus, while the fact is certain. The resultof combining the default rule with thecertain fact is a default status conclusion.

At the same time, ATT-Meta examinesevidence for the goal ¬fly(tweety), such as,for example, that Tweety is a penguin orhas a broken wing.

Discourse Context

The discourse context consists of beliefsabout particular people, places, events, etc.(e.g. author and addressee), aboutontological classes of people, objects,events, etc., about language use and

topic

topic

agent default inference rule: bird(x)→fly(x)actual fact: bird(tweety)default conclusion: fly(tweety)

Figure 1: Agent1’s beliefs about a Topic and about Agent2’s beliefs about the Topic

Agent2

Agent2

Agent1

Topic

Topic

Page 8

communication and about social andcultural conventions.

With respect to the specific example underconsideration, the discourse context wouldinclude such beliefs as:

in Europe (and elsewhere), people referto the ground level of a multi-storybuilding as the ground floor, the nextlevel up as the first floor, and so on,

in the USA (and elsewhere), people referto the ground level of a multi-storybuilding as the first floor, the nextlevel up as the second floor, and soon.

It would also include such beliefs as:

the author is a Spanish speaker,the addressees are Spanish speakers,Spanish speakers may be Spaniards,Spaniards are Europeans,

This knowledge may be accessed by thetranslator, ascribed to the author oraddressees of the source language text or tothe audience of the translation in order toestablish contextually coherentinterpretations or to select contextuallyrelevant expressions. We might representthe contribution of the discourse context forprocessing the example as in Figure 2below.

european (x) → refer-to-as (x, ground level, ground floor )american (x) → refer-to-as (x, ground-level, fir st floor )spanish (x) → european (x).

spanish (author)european (author)*refer-to-as (author, ground-level, “ground floor”)*

spanish (addressee)european (addressee)*refer-to-as (author, ground-level, “ground floor”)*

* belief resulting from default inferencing using ascribed rules of inference

author

translator

addressee

addressee

author

{

Figure 2: Discourse Context

Page 9

Utterance Context

The utterancecontext consistsof beliefsabout the objectsand eventsmentionedorimplied during the discourse, thecommunicativestate of the discourseandthe“open issues”(i.e., theobjectsor eventswhoseconnectionsto the contexthaveyetto be established).

With respectto the specificexampleunderconsideration,the utterancecontext wouldinclude such beliefs as:

the commercial real estate market inMoscow is expanding rapidly,

properties are renting at the equivalent of

$700 to $800/m2/year,properties are renting at the third highest

rates in the world (behind Tokyo andHong Kong),

properties are in high demand,the market is dominated by poverty, legal

uncertainty, the principle of the “richget richer”.

This knowledgemay also be accessedbythe translator, ascribed to the author or

addresseesof the sourcelanguagetext orthe audienceof the translationin order toestablish contextually coherentinterpretationsor to produce contextuallyrelevant expressions.We might representthecontributionof theutterancecontextforprocessing the example as in Figure 3below.

Interpretation

The processof translationbeginswith thetranslatoridentifying of the intentionof theauthor of the source languagetext. Thetranslator starts with the assumptionthatwhat the translator knows about sourcelanguageconventions,about cultural andsocial conventions of source languagespeakers and about people's presumedknowledgeof the world is the sameasthatwhich the author knows exceptwhere thetranslator, explicitly or by implication,knows that the author'sknowledgediffers.Similarly the translator assumesthat theauthor's knowledge about the source

expanding (commercial-real-estate-market, in Moscow)

rental price (commercial-real-estate, $700-800/m2/year)relative price (commercial-real-estate, third-highest-in-world)dominate (market, poverty & legal-uncertainty & law-of-jungle){ author

translator

addressee

addressee

author

Figure 3: Utterance Context

Page 10

language addressees is the same as thetranslator's knowledge of the addresseesexcept where there is explicit orimplicational evidence to the contrary. Thisknowledge, the translator's beliefs modifiedby that which the translator knows theauthor believes in contradiction to thetranslator’s beliefs, constitutes the discoursecontext or background context.

The utterance context provides foregroundcontext for interpretation. The initialcontext for processing our example mightinclude some of the translator’s beliefsabout the participants in the sourcelanguage interaction or about the objectsand events described in the text thus far(such as the belief that the Moscow referredto is in Russia, etc.). These are representedbelow as b1, b2, and b3. It will also includebeliefs the translator assumes the author hasabout the objects and events described inthe text thus far (such as the belief that thereal estate market in Moscow is booming,etc.). These are represented as b4 and b5.

The initial task of the translator, then, is toassign to the author, for each expression Eiuttered, an intention based on the form andsemantics of the expression and the beliefspresent in the utterance context. That is, theexpression, Ei, is associated with a semanticrepresentation, p(a,b), that needs to beinterpreted within the utterance context. Forour example, the expressions el tercer pisoand el segundo piso would be provided withsemantic representations akin to ιx| floor(x)∧ third(x) and ιx| floor(x) ∧ second(x).

The interpretation involves inferring abelief (represented alternatively as b6, b8 orb10) that is informative and compatible withthe utterance context by inferencing frombeliefs in the utterance and discoursecontexts (represented alternatively as b7, b9or b11). In other words, b6 can be inferredfrom p(a,b) and b7 as represented in Figure4, b8 from p(a,b) and b9 as represented inFigure 5, and so on. For our example, thetranslator might infer that the author isreferring to the fourth and third levels of the

b4b5

author

translator

addressee

addressee

author

b1b2b3b7

{b6

Figure 4: SL interpretation 1 (SL—i1)

Page 11

building (b6) given that the author andaddressees of the source language text areSpanish speakers (b2), many Spanishspeakers are European (b3), and Europeansrefer to the ground level of a multi-story

building as the ground floor, the first levelup as the first floor and so on (b7). Otherinterpretations (e.g., b8 or b10) are possiblegiven alternative assumptions about whobelieves what (e.g., b9 or b11). Finally, from

Figure 5: SL interpretation 2 (SL — i2)

b4b5

author

translator

addressee

addressee

author

b1b2b3b9

{b8

Figure 6: SL interpretation 3 (SL—i3)

b4b5

author

translator

addressee

addressee

author

b1b2b3b11

{b10

Page 12

thesepossibleinterpretations,the translatorselects one as the author’s intention.Assumein this casethatinterpretationis theone describedabove (b6) and representedbelow as SL—i1.

Translation

The second step is for the translator toexpress in the target language,with itsdifferentsetof linguistic conventions,to theTL audience,with its different social andcultural conventionsand,perhaps,differentconventionalknowledgeof the world, theSL author's intention (b6). The translatormust producean expressionwhich permitsat least that intention to be identified and,perhaps, any other possible intentionssupported by the original SL text.

To begin this process,the SL utterancecontext prior to the analysis of Ei issubtractedfrom the author’s intention, SL—i1. What remains represents theinformation added in processing Ei,including the beliefs neededto infer theauthor’s intention from p(a,b).

The next stepis to substitutethe discoursecontext of the SL interaction with thediscoursecontextof the TL interactionandto substitutetheutterancecontextof theSLauthor for the utterance context of the

translator(asTL author).This mayresultina changein the beliefs in the TL utterancecontextandcertainlyentailschangesin theTL discoursecontext.For instance,supposethat as a result of swapping out the SLdiscourse context for the TL discoursecontext, b7, the European floor namingconventionis blockedin the TL audience’sbeliefs spaceby the preexistingb9, that is,the American floor naming convention.Thatis to say,thetranslatorrealizesthatthefloor namingconventionof the audienceofthe translationis different from that of theaddressees of the SL text.

The next step is to generatean utterance

Ei∗, with semanticrepresentationp∗(a∗,b∗)

such that one of its interpretationsin thenewcontextis equivalentto theinformationderivedfrom Ei, in the original SL context,i.e., b6. Assumingthat the intention of thetranslator(as TL author) is to refer to thefourth and third levels above ground andassumingthat the floor namingconventionof theTL audiencein theAmericanone,therelevant expressionswill have semanticrepresentationsakin to ιx| floor(x) ∧fourth(x) andιx| floor(x) ∧ third(x), that is,the fourth floor and the third floorrespectively.This result is representedinFigure 7.

b1b2b3b9 SL/TL substitutions

translator

addressee

{

Figure 7: TL interpretation 1 (TL—i1)

addressee

b6

Page 13

The final step is to provide the possible

alternative interpretations of Ei∗, given the

initial TL discourse context, and TLutterance context.

Alternative Beliefs Affect Readings

There are many potential scenarios beyondthose presented above which are based onthe observation that the author used Spanishto addressee the readers of the originalarticle. For instance, with respect tointerpretation, the translator might consideras significant the fact that the buildingreferred to is in Moscow. This, coupledwith the assumption that Russians appear tofollow the US floor-naming convention,implies that the author was referring to thethird and second levels above ground in theoriginal text. Alternatively, the translatormight consider as significant the fact thatthe real estate agent that was quoted wasspeaking Russian. This, again coupled withthe assumption that Russian speakersappear to follow the US floor-namingconvention, implies that the author(indirectly through the real estate agent)was referring to the third and second levelsabove ground.

With respect to translation, the translatormight take as relevant the fact the text isintended for an audience of real-estateagents and thus the need to identify theexact level above ground is important. Inthat case the translator might have producedthe third or fourth story and the second orthird story as possible translations. Thenagain, the translator might take as relevantthe fact the audience of the translation isuninterested in the precise story referred tobut rather in the general availability ofspace. In that case the translator mightappropriately translate the expressions asone floor and the floor below respectively.We refer to variations in translation whichare derived from attempting toaccommodate the beliefs of the audience ofthe translation as "user-friendly"translation.

Thus, those facts identified by the translatoras relevant to establishing the author’sintention or to expressing that intention to agiven audience are crucial to the process. Inaddition, since most such assumptions aresimply not evaluable empirically, it is notobvious that there is any clear notion of"correct" translation.

Note also that an author's intention is notnecessarily to describe an event or state ofaffairs or to present their thoughts on sometopic. It might just as well be tocommunicate a mood or emotion, somesensory input, and so on. Language as aresource for communication provides theauthor with form as well as meaning whichcan be manipulated to such ends. Thus,form cannot be discounted a priori intranslation in the process.

Finally, a "context", while actually veryspecific in comparison with all possiblecontexts for all possible expressions forachieving all possible purposes, can berather large, abstract, and/or vague (e.g.,providing coherence with respect to aMarxist or Freudian or Catholic frameworkor some combination of such frameworks).That is, if a translator can identify one oranother such frameworks as an organizingprinciple to the written or spoken discourse,especially if it can be expressly attributed tothe author, then it too may have to be takeninto account.

Equivalence

A central objective of the approach is tosupport a notion of translation equivalencewhich is determined by the degree ofoverlap between information that is statedexplicitly and that is inferred from thecontext in the corresponding source andtarget language utterances or texts. That isto say, we assume that, because ofinformation gain and information lossduring translation, corresponding sourceand target language utterances or texts areto a greater or lesser degree "equivalent"

Page 14

depending on the total informationconveyed.

Theprocessingschemadescribedabovecanbe used to formally define translationequivalencenot on thebasisof themeaningof a text, but with respect to theinterpretation of the text. We take theinterpretationof the sourcetext to be thefinal context of sourceaddressees'beliefsstatelessthe initial state.This is essentiallythe elementsexplicitly communicatedbythe utterance plus the assumptionsandinferences used to support theinterpretation.

By comparing the interpretation (corestatementplus requisite context) of thesource language utterance with thecorrespondinginterpretationof the targetlanguage translation, we can determine,beliefsenvironmentby beliefsenvironment,which beliefs are sharedand which beliefsare not shared. Thus, for this simpleexamplecase,we end up with an orderedquadrupleconsistingof (a) beliefsin sourceaddressees’environment that are not intargetaddresses'environment;(b) beliefsintargetaddressees'environmentthat are notin the sourceaddressees’environment;(c)beliefs in both environmentswith possiblyvaryinglevelsof confidence;(d) confidencelevels for the interpretations as a whole.

This approachallows us to developa farmore elaborate notion of translationequivalencethan hasbeenproposedin thepast.First, we can distinguishbetweenthecore equivalenceof expressionsin sourceand target language texts and overallequivalence. Core equivalence is thesimilarity of speaker’s intention for thecorrespondinginterpretationsin sourceandtarget languages(e.g., SL—i1 and TL—i1or the levelsabovegroundand the beliefs/inferences required to identify it). Theoverall equivalenceis the total similarity(and difference)of the correspondingsetsof interpretations(e.g.,SL—i2, SL—i3, ...and TL—i2, TL—i3, ...).

It also allows us to develop bothquantitative and qualitative notions ofequivalence. Quantitative notions ofequivalencetake into accountthe numberof beliefs/inferencesneededto establishthespeaker’sintention (or, more broadly, thefull set of interpretations). Qualitativenotions of equivalencecan be defined onthe basis of the beliefs/inferencesused;their “currency”; the “simplicity” of theirconnection and so on.

Given this information-rich structure,it isclear that many different kinds of``translationequivalencerelations''couldbedefined. It should also be clear from theforegoing that in very few caseswill theinterpretations be identical.

Section 3

We have shown in previous sectionsthatbeliefs influencetranslationandthat this isreflected in the construction of aninterpretationof a text. The interpretationisconstructedfrom the semanticsof the textin conjunctionwith additionalpropositionsneededfor integrating that semanticsintothe utterancecontextin a coherentmanner.It is this interpretation,we claim, whichservesas the basis for human translationand which should serve, as well, as thebasis for automatic translation.

In this section we examine two furthercaseswhich we believeshowsthat it is justsucha coherentinterpretationthatunderliestranslation. We look at two differenttranslationsof eachof two texts.In the firstcase, patterns of differences in thetranslationscan be relatedto the differentglobal assumptionsof what the story itselfis about. In the secondcase,patternsofdifferencesin thetranslationscanberelatedto differing world views on the part of thetranslators.

Background

The texts and translationsused for thisdiscussion are drawn from a corpusprepared for the DARPA Machine

Page 15

Translation Evaluation in 1994, as reportedin White et al. (1994). The these materialsare available on-line at:

http://ursula.georgetown.edu/mt_web/index.html.

For this evaluation 300 newspaper texts(100 each in Spanish, French, andJapanese) were selected. There was norestriction as to subject domain. Twotranslations of each text into English werecommissioned from two differenttranslation houses. The translators weregiven strict instructions to neither add norremove information and to follow a specificstyle sheet.

The evaluation was conducted entirely bynative speakers of English, using one of theEnglish translations as the standard againstwhich the machine translations wereevaluated. In addition, the secondtranslation was also evaluated alongside themachine translations.

Translations were rated for fluency,informativeness, and adequacy.Interestingly, when ranked against the firsthuman translation, the second humantranslation (though better than any of themachine translations) still scored less thanperfect in informativeness and adequacy.

We have examined in detail two of thearticles drawn from the Spanish corpusalong with their accompanying translations.While these articles were selected preciselybecause they seemed to offer interestingvariations in the translations, we do notbelieve that they differ significantly fromthe other articles.

Methodology

Our methodology of analysis was asfollows. We first processed the Spanish textusing the initial modules of CRL'sPanglyzer Spanish analysis system (Farwellet al., 1994). This involved first breakingthe text into words and sentences andtagging each word and item of punctuationwith an identifying part-of-speech tag along

with relevant morphological information.Next, the output is processed by thePanglyzer’s phrase recognizer. This groupswords in the input into small chunks thatare semantically and syntactically cohesiveand unambiguous.

Then the two translations were cut up andaligned with these chunks. In some caseswhere two Spanish chunks were translatedby one indivisible English chunk, the twoSpanish chunks were combined into one. Inother cases, a number of aligned Englishchunks contained more than one differencebetween them. These multiple differenceswere broken out. We also examined the twotranslations sentence by sentence to lookfor differences in syntactic structure thatwere not reflected within any particularchunk.

Finally, we identified and classified eachdivergent translation unit (and eachsyntactic difference).

Classification

We had three basic categories ofdifferences: errors, free variation, andbelief-based. That is, we felt that anydifference in translation either reflected anerror or misunderstanding of the text,reflected an arbitrary choice of expressionhaving no impact on the meaning or effectof the text, or reflected a difference in themeaning or effect of the texts due todifferent interpretations of the sourcelanguage text. This last category, it shouldbe emphasized, is quite distinct from theerror category in that we believe that bothof the resulting translations, althoughdifferent, are licensed by the sourcelanguage text.

There were also some differences that werefer to as "derivative" in that a differencein one place resulted in a difference inanother place. For instance, if one translatorused a plural subject where the other used asingular, that could result in a derivativedifference in the morphology of the verb.

Page 16

Note that this classification is a based onour understanding of the translator's intentin choosing a particular translation and notsimply on the exterior form of thetranslation. We also note that, as a result, itmight be possible in any particular case toargue for a different classification, based onone's own understanding of the text and ofthe translation. However, the currentfindings are the result of agreementbetween the two authors and so have acertain degree of reliability.

Errors. The error category was the smallestof the three classes of differences. This wasto be expected since both translations weredone by qualified translators. We identifiedthree types of errors. First there are errorsthat are unintentional or accidental innature. Spelling errors, for example,generally fall into this category. The secondtype are errors or extreme awkwardness inthe translation due to interference from thesource language text. The third class oferror consists of intentional errors in that itappears that the translator did not justoverlook something but rather got it wrong.That is, the source language text not onlydid not provide a basis for the translationbut rather appeared to provide a basis for acontradictory translation.

Free Variation. The second category weexamined was free variation. To someextent this is a flexible category. That is, atthe strictest level, there were cases wherewe could see absolutely no differences insemantic content, connotations, style,register, or invited inferences. These wouldinclude such differences as the use (or non-use) of a definite article with plural nouns,writing out numbers versus using numerals,or the use of that versus which as a relativepronoun. At a less restrictive level, weincluded cases where there were differencesbetween the lexical choices, but it was notclear that they could be related directly todiffering beliefs about the text. An examplehere is the translation of sectores insectores costeros as either coastal area orcoastal sector. Sector has a slightly more

military or formal feel than area, but notenough for us to classify this as a belief-based difference.

Belief-based Differences. In this categorywe placed all translation differences that wefelt communicated substantially differentinformation, enough so that the readers ofthe differing translations would havedifferent ideas as to the nature of the sourcetext or the events described in that text.These differences, however, were not suchthat one could, on the basis of the sourcetext, identify them as wrong or incorrect.

We identified two subtypes of belief-baseddifferences. The first consisted primarily ofadditions or alterations of information thatare related to the beliefs of the translatorabout the target language audience. That is,information was added or altered if it wasfelt necessary to communicate the sourcelanguage author’s intent to the targetaudience properly. Similarly, informationmay have been deleted if it was redundantand could be recovered from context.

The second subtype of belief-baseddifferences include those that are related tothe beliefs of the translators about theevents recounted in the source text or on theattitudes of the participants (including thesource language author) about these events.Of the two subtypes, these are easily themost common.

Results

The results of our analysis for both texts arepresented in Table 1. As can be seen, theresults are substantially in agreement.About 40% of the segments in each textshowed differences in the translations. Ofthese differences, there are few outrighterrors, as one might expect from qualified,professional translators, working understrict guidelines. Of the remainingdifferences, they are split about half andhalf between free variation and belief-baseddifferences.

Page 17

We expected to find a certain number ofcases where the different beliefs of thetranslators about the world (and thereforeabout the events described in the sourcetext) influenced their translations. And wedid. What we did not necessarily expectwas that the large majority of thesedifferences patterned on the basis of eitherdiffering, yet internally consistent, overallinterpretations of the source languagearticle or differing global views of thetranslators. To demonstrate this, weexamine each text in more detail.

Text 1: Earthquakes in Chile

The news event which triggered the originalarticle was an unusual increase in thepurchases of living necessities in a Chileanport city, following newspaper and radioreports about the possibility of anearthquake striking the city. Thegovernmental Office of Emergencies(ONEMI) is a major source for this portion(the first half) of the article. The secondhalf of the article is devoted to a moregeneral discussion, supported by quotesfrom appropriate scientists, about the actual

likelihood of an earthquake and itspredictability.

A key factor in forming the translatorsinterpretations of the article appears to betheir understanding of the goals andreliability of the Office of Emergency(ONEMI) source. If the goals of ONEMIare understood as helpful, then they may berelied upon as a source of correctinformation. One may take at face valuetheir criticism of the news media forblowing the earthquake reports out ofproportion and encouraging irrationalbehavior among the populace. The secondhalf of the article validates this view byshowing that from a scientific perspectivethere is both little cause for alarm and notmuch that one can do anyway inpreparation.

On the other hand, if one suspects that thegoals of the ONEMI source may be self-serving, then their information may bebiased. It could be that the ONEMI sourceis trying to save face for the organizationwhich currently has egg on it for not havingencouraged reasonable preparations for animpending disaster and for not being ready

Text 1 Text 2Data

Number of words 403 392Number of phrases 170 192Number of phrases with differences 66 87Percentage of phrases with differences 39% 45%

Differences

Errors 13 14Unintentional (6) (3)Interference (6) (7)Wrong interpretation (1) (4)

Free Variation 32 38Belief-based 36 39

Source Text Related (30) (37)Target Audience Related (6) (2)

Derivatives: (12)=== ===

Total: 81 91

Table 1: Quantitative Results

Page 18

to deal with the naturally upset citizens.One way to do so is to blame others (in thiscase, irresponsible media). This view isconfirmed by the second half of the article,showing that from a scientific perspectivethere have been advances in predictingearthquakes and that sensing devices are inplace.

These two different evaluations then hookup with different understandings of whatkind of story is being written. If one takesthe first analysis, then the story fits nicelyinto a "blame the media" framework inwhich the accuracy or quality of mediareporting of an event itself becomes thesubject of another news media story.

On the other hand, if one takes the secondanalysis, this fits in nicely with a "blame thegovernment" framework in whichincompetence by government bureaucrats isthe key element of the story, oftenaccompanied by an attempt at covering upthe incompetence.

These two broad interpretations of thesource language author’s intent areencapsulated in the translations of theheadline shown below:

Accumulación de víveres por anunciossísmicos in Chile.

Hoarding Caused by Earthquake Pre-dictions in Chile.

STOCKPILING OF PROVISIONSBECAUSE OF PREDICTED EARTH-QUAKES IN CHILE.

In this one phrase there are threedifferences between the two translations:(1) Hoarding versus STOCKPILING OFPROISIONSfor Accumulaciónde víveres;(2) Causedby versus BECAUSEOF for porand (3) Earthquake Predictions versusPREDICTED EARTHQUAKES foranunciossísmicos. In the first translation,irrational, selfish behavior (hoarding) isrelated by a causal chain (caused by) to anirresponsible speech act (prediction) about a

hypothetical future event. In the secondheadline, on the other hand, rational activity(stockpiling) is based rationally (becauseof) on expectations of a future catastrophe(predicted earthquakes). In the firsttranslated headline, it is the predictions ofearthquakes that are at the center of thestory, while in the second, it is theearthquakes themselves that form the focusfor the story.

Another clue to this difference ofinterpretation is in the translations of:

La posibilidad de un remezón desas-troso...tuvountratamientoinadecuadoen recientes versiones periodísticas,...criticó la repartición.

Here the second translator keeps thestructure of the Spanish sentence(grammatically incorrect in English) inwhich criticar (to criticize) takes as itsdirect object the criticism itself:

The possibility of a disastroustremor...received inadequate treatmentin recent newspaper versions,...criti-cized the office.

The first translator, however, in keepingwith an understanding that the mediaexaggeration is at the heart of the story,chose to use the verb say, but then toindicate explicitly that the report wascritical of the media:

The possibility of a disastroustremor...was inadequately treated inrecent news stories,...said the release,which was critical of the media.

The first translator, then, understands thereto be an entire report (repartición, generalsense) issued by the Office for Emergenciesas critical of the media, while the secondtranslator envisions only a general commentfrom the office (repartición, usage inChile), perhaps an attempt on the part of agovernment bureaucrat to deflect the toughquestioning of the reporter.

Page 19

If these two differing interpretations of theevents in the story do, in fact, underlie thetwo different translation, we would expectto find certain other patterns as well. Wewould expect, for instance, that the firsttranslator, focussing on the local responseto the media misunderstanding, shouldchoose translations which play up theirrational behavior of the local inhabitants,translations that emphasize theunpredictability of earthquakes (and so theimplausibility of suddenly preparing forthem), and play down the seriousness ofsuch an earthquake, should it occur. Thesecond translator, focussing on the ill-prepared ONEMI, should do just theopposite: play up the rationality of theresponse of the people to an earthquake thatcan be accurately predicted to occur soon,with very serious consequences. And,indeed, this is what we find. Without goinginto all the details, we found that 21 of the30 source-text related, belief-baseddifferences were supportive of the generalhypothesis and its corollaries.

Text 2: Trial in Amazonia

The second text was a news story about thebeginning of the trial of a former policemanaccused of murdering two brothers, whowere instrumental in a labor union in ruralAmazonia. The accusation was that thepoliceman had been hired to commit thekilling, and the article goes into some detailabout the background of labor violence inthe province.

In this article, the results of our study werevery similar to those of the first article.There were not many errors and themajority of differences were split betweenfree variants and belief-related differences.The large majority of the belief-relateddifferences (30 out of 36) were related tothe source text interpretation, while asmaller number were related to (assumed)beliefs of the target language audience.

As with the first article, we also found thatthe large majority of the belief-related

differences reflected a broad consistentpattern of differences between the twotranslations. However, in this case thepatterns reflected the two translators’different general world views, particularlythose relating to economics and politics.Briefly put, the first translator views thearticle from what would generally be calleda left-wing perspective, while the secondtranslator seems to translate from a moreconservative position.

Hovy (Hovy 1988) has shown that therhetorical goals of speakers influence howtheir communications of the same event arerealized differently in different situations bydifferent speakers and has modeled this fora natural language generator, PAULINE.What we find here is that even translationsof a single text describing a particular eventdiffer according to the rhetorical stance ofthe translator and, in this case, according tothe translator’s political and economicunderstanding of the world.

So, for instance, the first (left-wing)translator translates sindicalista(describingthe murdered brothers) as labor leaderwhile the second (right-wing) translatorsimply uses union member. In translatingthe Spanish asesino and asesinado, thepoliceman on trial is described by the firsttranslator as an assassin who hasassassinatedthe victims, while the secondtranslator describes him as a killer who hasmurdered the victims. The first translatortranslates campesino as peasant andterrateniente as landholder while thesecond translates these terms as smallfarmer and landowner. Similarly, thesecond translator shows more respect bytranslation policía as police officer, justiciaas justice, and juezaas theCourt in contrastto the first translator’s policeman, law andjudge, respectively.

In short, the first translator translates thearticle from the perspective of a leftistpolitical analysis, where the economically-deprived working classes are in conflictwith the upper-class wealthy landowners,

Page 20

who can use the apparatus of thegovernment to enforce their class interests,and, in the face of strong opposition, mayeven resort to extra-legal means to maintaintheir position.

The second translator sees the article from amore conservative perspective. The trial isbringing the killer to justice, despite anatmosphere of violence brought on by thelarge scale union organizing.

In all, 30 out of the 37 source text-related,beliefs-based differences reflected thisunderlying difference in world view.

Summary

Overall the important conclusions we drawfrom these analyses are as follows. (1) Evengiven strict instructions, competenttranslators will frequently produce differenttranslations, even of simple newspapertexts. (2) Many of these differences resultin translations that are not synonymous.They may differ in actual content,emotional effect, or impliedpresuppositions. Nonetheless bothtranslations reflect responsibleinterpretations of the source text. (3) In thetwo articles examined, most of thesemeaningful translation differences could berelated to an underlying difference ininterpretation or understanding of the textas a whole.

Conclusions

With respect to translation, a pragmatics-based approach provides a much moreexplicit framework for reasoning about themany choices that translators must make inproducing a translation. But the centralassumption of the approach is that languageis vague and texts radically underspecifythe interpretation. This is why translatorsmust interpret utterances against a contextof beliefs about the world, about theelements of the utterance context, and aboutthe topic and related individuals and states-of-affairs.

The primary implication of this approach totranslation is that the beliefs of thetranslator and the way in which thetranslator reasons with them account for theeventual form of the translation. Thosebeliefs and that reasoning should thereforebe the focus of any critical analysis oftranslation quality. We need to ask:

• which beliefs are supported by the text(and how),

• which beliefs are supported by worldknowledge (and how),

• which beliefs are supported bylinguistic conventions (and how),

• which beliefs are supported by socio-cultural conventions (and how).

If the answers to these questions aresatisfactory, the translation is satisfactory.If they are less than satisfactory, then so toois the translation.

From the perspective of a pragmatics-basedmodel of translation and in view of thewidespread and significant translationvariants to be expected from both humanand machine translation systems, it shouldbe clear that the focus of evaluation shouldbe on (1) the similarity and differencebetween the beliefs of the participants andthe inferences performed during the sourceand target language interactions, and (2) onthe naturalness of expression of the targetlanguage text. It should also be clear thatthere is a wide range of potentiallyappropriate translations for a giveninteraction. Since variations in translationarise from differences in participants’beliefs and since each of the participants(translator, author, reader and audience) hasa different and incomplete knowledge ofthe individuals, objects, situations andevents referred to in a communicativeinteraction, the potential for variation isquite large.

Finally, we feel that a pragmatics-basedapproach to machine translation offers theonly direct assault on the issues raised byBar-Hillel as early as 1959 (Bar-Hillel,1960). It is not simply that MT systems

Page 21

need knowledge, they need to be able tocreate complex structures of beliefs and tobe able to reason within those structures inorder to arrive at an appropriateinterpretation in spite of incomplete orpossibly inconsistent knowledge. To ignorethis fact is to delay progress on boththeoretical and applied MT. Form-basedtranslation cannot work: It must besupported by inferencing from knowledgewithin a context.

References

Arcand, D. (1989) Jesusof Montreal. Mira-max: Orion Home Video.

Attardo, D. (1994) MIKROKOSMOS TMRBuilder's Guide. Internal Memo, Comput-ing Research Laboratory, New MexicoState University, Las Cruces, NM.

Austin, J. L. (1962). How To Do ThingsWith Words. Cambridge, MA: HarvardUniversity Press.

Ballim, A. and Wilks, Y. (1991). ArtificialBelievers: TheAscriptionof Belief. Hills-dale, NJ: Lawrence Erlbaum.

Bar-Hillel, Y. (1960). The Present Status ofAutomatic Translation of Languages. InF.L. Alt (ed.) Advances in Comput-ers,1:91-163. New York: Academic Press.

Barnden, J.A., Helmreich, S., Iverson, E.and Stein, G.C. (1994). An IntegratedImplementation of Simulative, Uncertainand Metaphorical Reasoning about Men-tal States. In J. Doyle, E. Sandewall and P.Torasso (Eds.), Principles of KnowledgeRepresentationand Reasoning:Proceed-ings of the Fourth International Confer-ence, pp.27-38. San Mateo, CA: MorganKaufmann.

Carlson, L. and Nirenburg, S. (1990). WorldModeling for NLP. Technical Report90-121. Center for Machine Translation, Car-negie Mellon University, Pittsburgh, PA.

Farwell, D., and Helmreich, S. (1993).Insights into Communicative Intent fromPragmatics-based Translation Equiva-lence. Paper presented at The Seventh

Annual SouthwesternCognitive ScienceConference. March 6-7, 1993. Tucson,AZ.

Farwell, D., and Helmreich, S. (1995). Thisis not a Bedroom Farce: Pragmatics andTranslation. Proceedingsof the FourthInternational Colloquium on CognitiveScience(ICCS-95), pp.73-82. Donostia,Spain.

Farwell, D., and Helmreich, S. (1996).Ground Floor? First Floor? Going Up.Presented at the Fifth InternationalPrag-maticsConference. July 4-9, 1996. Mex-ico City, Mexico.

Farwell, D., and Helmreich, S. (1997).Assassins or Murderers: Translation ofPolitically Sensitive Materials. Presentedat the 26thAnnualMeetingof theLinguis-tics Associationof theSouthwestInterna-tional PragmaticsConference. October 4-9, 1997. Los Angeles, CA.

Farwell, D. and Helmreich, S. (1998).Translation Differences and Pragmatics-Based Machine Translation. MachineTranslation 13: 17-39.

Farwell, D., Helmreich, S., Jin, W., Casper,M., Hargrave, J., Molina-Salgado, H., andWeng, F. (1994). Panglyzer: Spanish Lan-guage Analysis System. ProceedingsoftheFirstConferenceof theAssociationforMachine Translation in the Americas,pp.56-65.

Fauconnier, G. (1985). Mental Spaces:Aspectsof MeaningConstructionin Natu-ral Language. Cambridge, MA: MITPress.

Fodor, J. A. (1984). The Modularity ofMind. Cambridge, MA: MIT Press.

Green, G. M. (1984) Some remarks on howwords mean. Indiana University Linguis-tics Club, Indiana University, Blooming-ton, IN.

Grice, H. (1975). Logic and conversation. InP. Cole and J. Morgan, (Eds.) Syntaxandsemantics, 3: Speech acts, pp. 41-58. NewYork: Academic Press.

Page 22

Hovy, E. (1988). Generating Natural Lan-guage Under Pragmatic Constraints.Hillsdale, NJ: Lawrence Erlbaum Associ-ates.

Lakoff, G. (1987). Women,fire, anddanger-ous things. Chicago: University of Chi-cago Press.

Lakoff, G. and Johnson, M. (1980). Meta-phors we live by. Chicago: University ofChicago Press.

Nirenburg, S. (Ed.) (1987). Machine trans-lation: Theoretical and methodologicalissues.Studiesin Natural Language Pro-cessing. Cambridge: Cambridge Univer-sity Press.

Nirenburg, S., Carbonell, J., Tomita, M., andGoodman, K. (1992). Machine Transla-tion: A Knowledge-BasedApproach. SanMateo, CA: Morgan Kaufmann.

Nunberg, G. 1978. ThePragmaticsof Refer-ence. Indiana University Linguistics Club,Indiana University, Bloomington, IN.

Onyshkevych, B. and Nirenburg, S. (1994).The Lexicon in the Scheme of KBMTThings. Memoranda in Cognitive andComputerScience:MCCS94-277, Com-puting Research Laboratory, New MexicoState University, Las Cruces, NM.

The PANGLOSS Mark III Machine Transla-tion System (1994). A Joint TechnicalReport by NMSU CRL, USC ISI andCMU CMT. Carnegie Mellon University,Pittsburgh, PA.

Pustejovsky, J. (1991). The Generative Lex-icon. Computational Linguistics, 17:4.

White, J., O'Connell, T., and O’Mara, F.(1994). The ARPA MT Evaluation Meth-odologies: Evolution, Lessons, and FutureApproaches. Proceedings of the FirstConference of the Association forMachine Translation in the Americas,pp.193-206, Columbia, MD.

Wilks, Y. (1975). A Preferential Pattern-Seeking Semantics for Natural LanguageInference. Artificial Intelligence, 6:53-74.