DTIC ADA460337: Building a Statistical Machine Translation System from Scratch: How Much Bang for the Buck Can We Expect

9 Pages·0.09 MB·English

by  Defense Technical Information Center

Building a Statistical Machine Translation System from Scratch:
How Much Bang for the Buck Can We Expect?
UlrichGermann
USC InformationSciences Institute
4676AdmiraltyWay, Suite1001
MarinadelRey, CA 90292
[email protected]
Abstract 2 Data CollectionandPreparation
Wereportonourexperiencewithbuildinga 2.1 ObtainingTamilData
statistical MT system from scratch, includ-
Tamil data is not very difficult to find on the web.
ing the creation of a small parallel Tamil-
There are several Tamil newspapers and magazines
English corpus, and the results of a task-
withonlineeditions,andthelargeinternationalTamil
basedpilotevaluationofstatisticalMTsys-
communityfostersthe use of the Internetforthe dis-
tems trained on sets of ca. 1300 and ca.
semination of information. After initial investigation
5000parallelsentencesofTamilandEnglish
ofseveralwebsiteswedecidedtodownloadourexper-
data. Ourresultsshowthatevenwithappar-
imental corpus from www.tamilnet.com, a news
ently incomprehensible system output, hu-
site that provides local news on Sri Lanka in both
mans without any knowledge of Tamil can
TamilandEnglish. TheTamilandEnglishnewstexts
achieve performance rates as high as 86%
onthissitedonotseemtobetranslationsofeachother.
accuracyfortopicidentification,93%recall
The availability of a fairly large in-domaincorpusof
for document retrieval, and 64% recall on
local news on Sri Lanka in English (over 2 million
questionanswering(plusanadditional14%
words) allowed us to train an in-domainEnglish lan-
partiallycorrectanswers).
guagemodelofSriLankannews.
1 Introduction 2.2 EncodingandTokenization
Crisesanddisastersfrequentlyattractinternationalat- Tamil is written in a phonematic, non-Latin script.
tention to regions of the world that have previously Several encoding schemes exist in parallel. Even
been largelyignoredby the internationalcommunity. thoughthe Unicodestandardincludesa set of glyphs
While it is possible to stock up on emergency relief forTamil,itisnotwidelyusedinpractice. Mostweb
supplies and, for the worst case, weapons, regardless sitesthatofferTamillanguagematerialassumeLatin-1
ofwhereexactlytheyareeventuallygoingtobeused, encodingandrelyonspecialtruetypefonts,whichof-
thiscannotbedonewithmultilingualinformationpro- ten are also offered for free download at those sites.
cessingtechnology.Thistechnologywilloftenhaveto Tamil text is therefore fairly easy to identify on web
be developedafter the fact in a quick responseto the sites via the faceattribute ofthe HTML fonttag. All
given situation. Multilingual data resources for sta- thatisnecessaryisalistofTamilfontnamesusedby
tistical approaches, such as parallelcorpora, may not thedifferentsites,andknowledgeaboutwhichencod-
alwaysbeavailable. ings these fonts implement. While we could restrict
In the fall of 2000, we decided to put the current ourselvestoonedatasourceandencodingforourex-
stateofthearttothetestwithrespecttotherapidcon- periment, any large-scale system would have to take
structionofamachinetranslationsystemfromscratch. this into account. In order to make the source text
Withinonemonth,wewould recognizable to humans who have no knowledge of
hiretranslators; Tamil,wedecidedtoworkwithtransliteratedtext1.
(cid:15)
translateasmuchtextaspossible;and
(cid:15) 2.3 TranslatingtheCorpus
trainastatisticalMTsystemonthedatathuscre-
(cid:15)
ated. Originallywehopedtobeabletocreateaparallelcor-
pusof about100,000wordson the Tamilside within
ThelanguageofchoicewasTamil,whichisspoken
one month, using several translators. Professional
inSriLankaandinthesouthernpartofIndia.Tamilis
ahead-lastlanguagewithaveryrichmorphologyand 1Translations, however, were produced from the original
thereforequitedifferentfromEnglish. Tamil.
Report Documentation Page Form Approved
OMB No. 0704-0188
Public reporting burden for the collection of information is estimated to average 1 hour per response, including the time for reviewing instructions, searching existing data sources, gathering and
maintaining the data needed, and completing and reviewing the collection of information. Send comments regarding this burden estimate or any other aspect of this collection of information,
including suggestions for reducing this burden, to Washington Headquarters Services, Directorate for Information Operations and Reports, 1215 Jefferson Davis Highway, Suite 1204, Arlington
VA 22202-4302. Respondents should be aware that notwithstanding any other provision of law, no person shall be subject to a penalty for failing to comply with a collection of information if it

does not display a currently valid OMB control number.
1. REPORT DATE 3. DATES COVERED
2001 2. REPORT TYPE 00-00-2001 to 00-00-2001
4. TITLE AND SUBTITLE 5a. CONTRACT NUMBER
Building a Statistical Machine Translation System from Scratch: How
5b. GRANT NUMBER
Much Bang for the Buck Can We Expect?
5c. PROGRAM ELEMENT NUMBER
6. AUTHOR(S) 5d. PROJECT NUMBER
5e. TASK NUMBER
5f. WORK UNIT NUMBER
7. PERFORMING ORGANIZATION NAME(S) AND ADDRESS(ES) 8. PERFORMING ORGANIZATION
Information Sciences Institute,University of Southern California,4676 REPORT NUMBER
Admirality Way Suite 1001,Marina del Rey,CA,90292
9. SPONSORING/MONITORING AGENCY NAME(S) AND ADDRESS(ES) 10. SPONSOR/MONITOR’S ACRONYM(S)
11. SPONSOR/MONITOR’S REPORT
NUMBER(S)
12. DISTRIBUTION/AVAILABILITY STATEMENT
Approved for public release; distribution unlimited
13. SUPPLEMENTARY NOTES
14. ABSTRACT
15. SUBJECT TERMS
16. SECURITY CLASSIFICATION OF: 17. LIMITATION OF 18. NUMBER 19a. NAME OF
ABSTRACT OF PAGES RESPONSIBLE PERSON

a. REPORT b. ABSTRACT c. THIS PAGE 8
unclassified unclassified unclassified
Standard Form 298 (Rev. 8-98)
Prescribed by ANSI Std Z39-18
translationservicesintheUScurrentlychargeratesof estimated2.5personmonths.However,agoodpartof
about30centsperEnglishwordfortranslationsfrom this effort led to resources that can also be used for
Tamil into English. Given that the English transla- otherpurposes.
tion of a Tamil text usually contains about 1.2 times
2.4 LessonsLearnedforFutureProjects
asmanywordsastheTamiloriginal,thetranslationof
acorpusof100,000Tamilwordswouldcostapproxi- If we were to giveadvicefor future,similar projects,
matelyUSD36,000.Thiswasfarbeyondourbudget. wewouldemphasizeandrecommendthefollowing:
InIndia, bycomparison,rawtranslationsmaycost
2.4.1 Goodtranslatorsarenoteasytofind
as little as one cent per Tamil word2. However, out-
It is difficult to find good translators for a short-
sourcingthetranslationworkabroadwasnotfeasible
termcommitment. Unlessoneiswillingtopayapre-
forus, sincewehadneithertheadministrativeinfras-
mium price, it is unlikely that one will find profes-
tructurenorthe time to managesuch an effort. Also,
sionaltranslatorswhoarewillingto commitmuchof
workingwith partnersso remote would have made it
theirtimeforalimitedperiodoftimeandonshortno-
very difficult to communicate our exact needs and to
tice.
implementproperqualitycontrol.
We finallydecidedtohireastranslatorsfourenter- 2.4.2 Makethetranslationjobattractive
ing and second-year graduate students in the depart-
Asforeignstudents,ourtranslatorswouldeachhave
ment of engineering whose native language is Tamil
been allowed to work up to twenty hours per week.
and who had responded to an ad posted in the local
None of them did, because the work was frustrating
mailinglistforstudentsfromIndia.
and boring, and because they found more attractive,
In order to manage the corpus translation process,
long term employment on campus. Our translators’
we set up a web interface through which the transla- frustrationmayhavebeenfosteredbyseveralfactors:
torscouldretrievesourcetextsanduploadtheirtrans-
lations, post-editors could post-edit text online, the thedifferencesbetweenSriLankanTamil(theva-
(cid:15)
projectprogresscouldbemonitored,andallincoming rietyusedinourcorpus)andtheTamilspokenin
textwasavailabletootherprojectmembersassoonas SouthernIndia(thenativelanguageofourtrans-
itwassubmitted. lators),whichmadetranslating,accordingtoour
We originally assumed that translators would be translators,verydifficult;
abletotranslateabout500wordsperhourifwewere
the lack of translation experienceof our transla-
content with raw translations and hardly any format- (cid:15)
tors;and
ting, and if we allowed them to skip difficult words
or sentences. This estimate was based on an inter- our high expectations. We originally told our
(cid:15)
nalevaluation,inwhichmultilingualmembersofour translators that since they were not working on
group translated sample documents from their native site, we would expect the translation of 500
language (Arabic, German, Romanian) into English wordsperhourreported.Whenwelaterswitched
andkepttrackofthetimetheyspentonthis. to hourly pay regardless of translation volume,
Itturnedoutthatourexpectationswereverymuch thetranslationvolumepickedupslightly.
exaggeratedwithbothrespecttotranslationspeedand
2.4.3 Bepreparedtopost-edit
thequalityoftranslation. Theactualtranslationspeed
forTamilvariedbetween156and247wordsperhour In professional translating, translators typically
withanaverageof170wordsperhour.In139hoursof translateintotheirnativelanguageonly. Onemaynot
reportedtranslationtime(overaperiodofeventually6 be able to find translators with English as their na-
weeks),about24,000words/1,300sentencesofTamil tive language for low density or “small” languages,
text were translated, at an effective cost of ca. 10.8 so it may be necessary to have the translations post-
cents per Tamilword (translators’ compensationplus editedbypeoplewith greaterlanguageproficiencyin
administrativeoverhead).Thisfiguredoesnotinclude English.
theeffortformanuallypost-editingthetranslationsby
2.4.4 Havetranslatorsandpost-editorsworkon
anativespeakerofEnglish(12-16personhours).
site
Theoverallorganizationoftheproject(sourcedata
Itisbettertohavetranslatorsandpost-editorswork
retrieval, hiring and management of the translators,
on site and ideally as teams, so that they can resolve
design and implementation of the web interface for
ambiguitiesandmisunderstandingsimmediatelywith-
managingtheprojectviatheInternet,developmentof
out the delays of communicating indirectly, be it by
transliteratorandstemmer,etc.)requiredanadditional
email or other means. A post-editor who does not
2PersonalcommunicationswithThomasMalten,University knowthesourcelanguagemaymisinterpretthetrans-
ofCologne. lator,asthefollowingcasefromourcorpusillustrates:
100 100
80 80
%) %)
n n
e (i 60 e (i 60
g g
a a
er er
v v
o 40 o 40
c c
xt xt
e e
t t
20 20
0 0
0 20 40 60 80 100 120 0 20 40 60 80 100 120
corpus size (in thousand tokens) corpus size (in thousand tokens)
English Tamil
Figure 1: Textcoverageon previouslyunseen text for English (left) and Tamil (right). The upperline in each
graphshowsthecoveragebytokensthathavebeenseenatleastonce,thelowerlineshowsthecoveragebytokens
thathavebeenseenatleast5times. Theerrorbarsindicatestandarddeviation.
ments of 1000 tokens each. For each segment ,
Raw translation: Information about the schools in we computed how many of the tokens had been pnrek-
whichpeoplewhomigratedtoKudaanaduarestaying viously seen in the segments . The upper
isbeinggathered. lineinthegraphsshowsthepenrc1e:n:ta:gnek(cid:0)of1tokensin
Post-editedversion:Informationabouttheschoolsin thathadoccurredatleastoncebeforeinthesegmenntks
(sic!) which immigrantsto Kudaanaduare attending ,thelowerlineshowsthepercentageofto-
isbeinggathered. kne1n:s:t:hnakt(cid:0)h1adbeenseenatleastfivetimesbefore.
ForthepurposeofstatisticalNLP,itseemsreason-
Inthiscase, thepost-editorclearlymisinterpretedthe abletoassumethatthelowercurvegivesabetterindi-
translator. What the translator meant to and actually cationofhowmanypercentofpreviouslyunseentext
did say is that information was being gathered about we can expect to be “known” to a statistical model
the schools in which migrants/war refugees who had trainedonacorpusof tokens.
m
arrived in Kudaanadu had found shelter. However, Atacorpussizeof24,000tokens,whichisapprox-
the post-editorinterpretedthe phrase people who mi- imatelythesizeoftheparallelcorpuswewereableto
grated to Kudaana as describing immigrants and as- create during our experiment, about 28% of all word
sumedthatinformationwasbeinggatheredabouttheir formsinpreviouslyunseenSriLankanTamiltextcan-
educationratherthantheirhousing. not be found in the corpus, and 50% have been seen
lessthan5times. Inotherwords,ifwetrainasystem
3 EvaluationExperiments on this data, we can expect it to stumble over every
other word! At a corpus size of 100,000 tokens, the
3.1 APrioriConsiderations
numbersare17%and33%.
Thericherthemorphologyofalanguageis,thegreater ForEnglish,thenumbersare9%/23%foracorpus
isthetotalnumberofdistinctwordformsthatagiven of 24K tokens and 0%/8% for a corpus of 100K to-
corpus consists of, and the smaller is the probability kens.
thata certain wordform actuallyoccursin anygiven Inordertoboostthetextcoveragewebuiltasimple
textsegment. Figure1showsthepercentageofword textstemmerforTamil, basedontheTamilinflection
formsin unseentextthathaveoccurredin previously tables in Steever(1990) and some additional inspec-
seen text as a function of the amount of previously tionofourparallelcorpus. Thestemmerusesregular
seen text. Thegraphon the leftshowsthe curvesfor expressionmatchingtocutoffinflectionalendingsand
English,theoneontherightthecurvesforSriLankan introduce some extra tokens for negation and certain
Tamil. The graphsshow the averagesof 100 runson case markings (such as locative and genitive), which
different text fragments; the error bars indicate stan- areallmarkedmorphologicallyinTamil. Itshouldbe
darddeviation. notedthatthestemmerisfarfromperfectandwasonly
Thenumberswerecomputedinthefollowingman- intended to be an interim solution. The performance
ner: A corpus of 120,000 tokens was split into seg- increasesaredisplayedinFigure2. Foracorpussize
100 methods,namelyglossing(replacingeachTamilword
bythemostlikelycandidatefromthetranslationtables
80 createdwiththeEGYPTtoolkit)andModel4decoding
(Brownetal.,1995;Germannetal.,2001).
%)
n Figure 3 shows the outputof the differentsystems
e (i 60 incomparisonwiththehumantranslation.
g
a Wethenconductedthefollowingexperiments.
er
v
o 40
xt c 3.2.1 DocumentClassificationTask
e
t Seven human subjects without any knowledge of
20
Tamil were given translationsof a set of 15 texts (all
from the Berkeley corpus) and asked to categorize
0 themaccordingtothefollowingtopichierarchy:
0 20 40 60 80 100 120
corpus size (in thousand tokens)
NewsaboutSriLanka
(cid:15)
Reports about clashes between the Sri
Figure 2: Text coverage increase by stemming for
(cid:15)
LankanarmyandtheLiberationTigers
Tamil. The solid lines indicate text coverage for un-
Sri Lankan security-related news (arrests,
stemmeddata(seenatleastonceandatleastfivetimes,
(cid:15)
armsdeals,etc.)
respectively), the dashed lines the text coverage for
SriLankanpoliticalnews(strikes,transport,
stemmeddata.
(cid:15)
telecom)
concernsSriLankabutdoesn’tfitanyofthe
of24Ktokens,thepercentagesofunknownitemsdrop (cid:15) above
to19%(from28%;neverseenbefore)and36%(from
NewsaboutPakistan/India
50%; seen less than 5 times). For a training corpus (cid:15)
Nuclear tests in Pakistan and India, includ-
of100Ktokens,thenumbersare12%and23%(from (cid:15)
ing their aftermath (international reactions,
17%/33%).
etc.)
3.2 Task-BasedPilotEvaluation Corruption investigation against Benazir
(cid:15)
Buto
Giventhesenumbers,itisobviousthatonecannotex-
News about Pakistan/India but none of the
pect much performance from a system that relies on (cid:15)
above
modelstrainedononly24Ktokensofdata. Asamat-
teroffact,itisclosetoimpossibletomakeanysense Internationalnews
(cid:15)
whatsoeveroftheoutputofsuchasystem(cf.Fig. 3). Disasters,accidents
(cid:15)
To get an estimate of the performance with more NelsonMandela’sbirthday
(cid:15)
trainingdata, we augmentedour corpuswith a paral- Otherinternationalnews
(cid:15)
lel corpusof internationalnews texts in Southern In- Impossibletotell
dian Tamil which was made available to us by Fred (cid:15)
GeyoftheUniversityofCaliforniaatBerkeley(hence- Exceptforoneduplicateset,eachsubjectreceiveda
forth: Berkeley corpus). This corpus contains ca. differentsetoftranslations. Thesetsdifferedintrain-
3,800 sentence pairs with 75,800 Tamil tokens after ing parameters and the translation method used. Ta-
stemming (before stemming: 60,000; the difference ble 1 shows the results of this evaluation. The dif-
is due to the introduction of additional markers dur- ferencebetweenthesubjects5aand5b,whoreceived
ing stemming). Some of the parallel data was with- thesamesetoftranslations,suggeststhattheindivid-
held for system evaluation; the augmented training ualclassifiers’accuracyinfluencestheresultssomuch
corpus (Berkeley and TamilNet corpus; short B+TN) as to blur the effect of the other parameters. There
hadasizeof85KtokensontheTamilside. Theaug- seemstobeatendencyforglossingtoworkbetterthan
mented training corpus had a text coverage of 81% Model4decoding.Glossing,inoursystem,isasimple
(seen at least once; 75% without augmentation), and baselinealgorithmthatprovidesthemostlikelyword
67% (seen at least 5 times; 60% without augmenta- translation for each word of input. Translation can-
tion), respectively, for Sri Lankan Tamil. We trained didates and their probabilities are retrieved from the
IBMTranslationModel4(Brownetal.,1993)bothon translationtable,whichispartofthetranslationmodel
ourcorpusaloneandontheaugmentedcorpus,using trainedontheparallelcorpus.
theEGYPTtoolkit(Knightetal.,1999;Al-Onaizanet The document classification test is foremost and
al., 1999), and then translated a number of texts us- aboveallameasureofthequalityofthetranslationta-
ing differenttranslation modelsand differenttransfer bleforfrequentlyoccurringwords. Inpractice,actual
SmallCorpus AugmentedCorpus
24Ktokens(Tamil)oftrainingdata 85Ktokens(Tamil)oftrainingdata
Humantranslation Gloss Model4decodinga Gloss Model4decoding
Government-United vanniunitednational governmentunited
childrenunited thegovernmentwith
NationalParty partykalloya nationalpartymeet
nationalparty unitednationalparty
meetingwillnottake tomorrow day
tomorrow. meetsday.
placetomorrow. naTaipeRamaaTTa naTaipeRamaaTTa
Theproposed fridayprogressiveit throughout
meetingtomorrow, andunitednational progressive
Thursday,between samekaTcikkumiTai governmentthe
thePeoplesFront tomorrowthursday onfridayhealth unitednationalin throughoutthe
Governmentandthe naTaipeRavirunew mothersunited kaTcikkumiTaiday governmentandthe
UnitedNational politicalvaappu nationaltomorrowon thursdaylenonew unitednationalparty
Partyregardingthe informationkalloya thursdaynew politicalvaappu .daythursdayleno
NewPolitical piRitoru politicalfoundon foundmeetpiRitoru newpoliticalfound
Documentwas tin2attiRkupa information. tin2attiRkupa meetmovies.
announcedas pin2pooTappaTT- pin2pooTappaTT-
postponedtoalater uLLataakafound uLLataakamovies
date. is .
oflankanpresident
Presidential oflankantoand
chandrikasankari
Secretariatsources freedomtamilian
freedomnow thesrilankanforces
saythatthismeeting now ’slankatoadvisor
veLinaaTTukkuc ofpresident
waspostponedasthe veLinaaTTukkuc freedomofthe
cen2Riruppat chandrikasankari
SriLankanPresident cen2Riruppat tamilianteamto
iccantippu freedompresently
Chandrika iccantippu secretariatiscalled
piRpooTappaTT- secretariatcircles
Bandaranayakkahas piRpooTappaTT- minister.
uLLataakapresident reported.
gonetoaforeign uLLataakato
secretariatcircles
country. secretariatin*is
reported.
Atthesametimeitis
tobenotedthatthe
meetingofthesub
hecountries
committee thataboutreturning
returningthe
examiningthe andiccantippu heraboutthemilitary hiscountryandthe
iccantippu
politicaldocument naTaipeRumen2a andtheywere returningtoincrease
naTaipeRumen2athe
betweentheUnited andivvaTTaaram announced. theradio.
ivvaTTaarammore
NationalPartyand moreand*is
thereported.
theGovernmentwas
heldyesterday
Tuesdayevening
aBrownetal.(1993);Brownetal.(1995)
Figure3: Sampleoutputofvarioussystems
Table1: ResultsoftheDocumentClassificationTask. Testsubjectswereaskedtoclassifythetranslationsof15
documentsinto4majorand11minorcategories.
partially
input pegginga? transfer correct incorrect
correctb
1 raw no M4decodingc 7 4 4
2 stemmed yes M4decoding 8 3 4
3 stemmed no M4decoding 13 2 0
4 raw no gloss 13 1 1
5a stemmed yes gloss 8 3 4
5b stemmed yes gloss 12 2 1
6 stemmed no gloss 11 2 2
apeggingcausesthetrainingalgorithmtoconsideralargersearchspace
bcorrecttoplevelcategorybutincorrectsub-category
ctranslationbymaximizingtheIBMModel4probabilityofthesource/translationpair(Brown
etal.,1993;Brownetal.,1995)
classification might be performed by automatic pro- nificant manner. We were surprised that even with
ceduresrather than humans. If we dare to accept the the poor translation performance of our system, re-
top performances of our human subjects as the ten- calls as high as 93% at a precision of 88% could be
tative upper bound of what can be achieved with the achieved. Secondly, the data shows that gapsare not
current system using a translation model trained on randomlydistributedoverthedata,butthatsomeques-
85K tokens of Tamil text and the corresponding En- tionsclearlyseemtohavebeenmoredifficultthanoth-
glishtranslations,wecanconcludethattheclassifica- ers. One of the particulardifficultaspects ofthe task
tionaccuracycanexceed86%(13/15)forfine-grained was the spelling of names. Question 11, for exam-
classificationandreach100%forcoarse-grainedclas- ple,askedWhathappenedtoChandraKumarAbayas-
sification. However,giventheextremelysmallsample ingh?. Inthetranslations,however,itwasrenderedin
sizeinthisevaluation,theevidenceshouldnotbecon- simple transliteration: cantirakumaara apayacingka.
sideredconclusive. Itrequiresaconsiderabledegreeoftenacityandimag-
inationtofindthisconnection.
3.2.2 DocumentRetrievalTask
The document retrieval task and the question an- 3.2.3 QuestionAnsweringTask
sweringtask(seebelow)werecombinedintoonetask. In order to measure the performance in the ques-
Thesubjectsreceived14texts(fromtheTamilNetcor- tion answering partof this evaluation, we considered
pus)and15lead questionsplus13 additionalfollow- onlyquestionsrelevantto the documentsthatthe test
up questions. Their task was to identify the docu- subjectshadidentifiedcorrectly. Becauseofthediffi-
ment(s)thatcontain(s)theanswertothequestionand cultyofthetask,wewerelenienttosomedegreeinthe
toanswerthequestionsasked. Typicalleadquestions evaluation.Forexample,ifthecorrectanswerwasthe
were questionssuch asWhat is the security situation formerpresidentoftheteacher’sunionandtheanswer
in Trinconmalee?, or Who is S. Thivakarasa?; typi- given was an official of the teacher’s union, we still
calfollow-upquestionswereWhoisincontrolofthe countedthis as “close enough”and thereforecorrect.
situation?, What happened to him/her?, or How did Inaddition,wealsoallowedpartiallycorrectanswers,
(other)peoplereacttowhathappened?.Asinthepre- that is, answers that went into the right direction but
viousexperiment,eachsubjectreceivedtheoutputof werenotquitecorrect. Forexample,ifthecorrectan-
adifferentsystem. swer was The army imposed a curfew on fishing, we
Table 2 shows the result of the document retrieval countedtheanswerthearmyisstoppingfishingboats
task. Again, the sample size was too small to draw as partially correct. All in all, it was very difficultto
anyfinalconclusions,butourresultsseem tosuggest evaluatethis section of the task, becauseit was often
the following. Firstly, the test subject in the group close to impossible to determine whether the answer
dealing with output of systems trained on the bigger was just an educated guess or actually based on the
trainingsettendtoperformbetterthantheonesdeal- text. Thereweresomecaseswhereanswerswerepar-
ingwiththeresultsoftrainingonlessdata. Thissug- tiallyorevenfullycorrecteventhoughthecorrectdoc-
gests thatthe jumpfrom24Kto 85Ktokensoftrain- umenthad notbeen identified. In retrospectwe con-
ing data mightimprovesystem performancein a sig- clude that it would have been better to have the test
Table2:Recallandprecisiononthedocumentretrievaltask.Testsubjectswereaskedtoidentifythedocument(s)
containingtheanswersto15leadquestions.Blackdotsindicatesuccessfulidentificationofatleastonedocument
containingtheanswer.
tcroarinpiunsg mtreatnhsol.d 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 recall precision
1 TNa glossing 67% 79%
(cid:15) (cid:15) (cid:15) (cid:15) (cid:15) (cid:15) (cid:15) (cid:15) (cid:15) (cid:15)
2 TN M4dec.b 53% 80%
(cid:15) (cid:15) (cid:15) (cid:15) (cid:15) (cid:15) (cid:15) (cid:15)
3 TN bothc 60% 48%
(cid:15) (cid:15) (cid:15) (cid:15) (cid:15) (cid:15) (cid:15) (cid:15) (cid:15)
4 TNd both 67% 79%
(cid:15) (cid:15) (cid:15) (cid:15) (cid:15) (cid:15) (cid:15) (cid:15) (cid:15) (cid:15)
5 B+TN.1 glossing 80% 87%
(cid:15) (cid:15) (cid:15) (cid:15) (cid:15) (cid:15) (cid:15) (cid:15) (cid:15) (cid:15) (cid:15) (cid:15)
6 B+TN.1 M4dec. 67% 86%
(cid:15) (cid:15) (cid:15) (cid:15) (cid:15) (cid:15) (cid:15) (cid:15) (cid:15) (cid:15)
7 B+TN.1 both 80% 86%
(cid:15) (cid:15) (cid:15) (cid:15) (cid:15) (cid:15) (cid:15) (cid:15) (cid:15) (cid:15) (cid:15) (cid:15)
8 B+TN.1f both 60% 73%
(cid:15) (cid:15) (cid:15) (cid:15) (cid:15) (cid:15) (cid:15) (cid:15)
9 B+TN.1g both 93% 88%
(cid:15) (cid:15) (cid:15) (cid:15) (cid:15) (cid:15) (cid:15) (cid:15) (cid:15) (cid:15) (cid:15) (cid:15) (cid:15) (cid:15)
10 B+TN.2h both 87% 75%
(cid:15) (cid:15) (cid:15) (cid:15) (cid:15) (cid:15) (cid:15) (cid:15) (cid:15) (cid:15) (cid:15) (cid:15) (cid:15)
HumanTranslations 100% 100%
(cid:15) (cid:15) (cid:15) (cid:15) (cid:15) (cid:15) (cid:15) (cid:15) (cid:15) (cid:15) (cid:15) (cid:15) (cid:15) (cid:15) (cid:15)
aTamilNetcorpusonly;stemmed;1291alignedtextchunks;23,359tokensonTamilside;1000trainingiterations.
bIBMModel4decoding.
cBothglossingandIBMModel4decodingwereavailabletothetestsubject.
dsameasabove,buttrainedwithpeggingoption(morethoroughsearchduringtraining);10trainingiterations.
eBerkeleyandTamilNetcorpora;5069alignedtextchunks;85421tokensonTamilside;100trainingiterations.
fsameasabove;10trainingiterations.
gsameasabove,trainedwithpeggingoption;10trainingiterations.
hBerkeleyandTamilNetCorpora,raw(unstemmed);64439tokensonTamilside,50trainingiterations.
subjects mark up those text passages in the text that
Table 3: Accuracy on question answering. The test
justifytheiranswers.
sets are the same as in Table 2. Only questions con-
Again,thedatasuggeststhatthedifferenceintrain-
cerningdocumentsthatwereidentifiedcorrectlywere
ingcorpussizedoesaffecttheamountofinformation
consideredinthisevaluation.
thatisavailablefromthesystemoutput. Subjectsus-
ing output of a system based on a translation model
test training No.ofrele- partially
correct
thatwastrainedononlytheTamilNetdatatendtoper- set corpus vantquestions correct
formworse than subjectsusingoutputfroma system 1 TN 17 2=12% 0= 0%
basedonatranslationmodeltrainedonthelargercor- 2 TN 16 3=19% 4= 25%
pus. The poor performance on test set No. 6 may 3 TN 18 2=11% 0= 0%
suggest that for this task and at this level of transla- 4 TN 17 7=41% 5= 19%
tion quality, glossing provides more informative out- 4 B+TN.1 23 14=61% 6= 26%
putthan Model4 decoding. This resultis notpartic- 6 B+TN.1 19 5=26% 3= 16%
ularly surprising, since we noticed that Model 4 de- 7 B+TN.1 22 14=64% 3= 14%
codingtendstoleaveoutmorewordsthanacceptable. 8 B+TN.1 19 12=63% 1= 5%
Clearly, this is one area where the translation model 9 B+TN.1 26 14=54% 5= 19%
hastobeimproved. 10 B+TN.2 24 8=33% 9= 38%
Test set 10 is the only set produced by a system human 28 24=86% 2= 7%
using a translation modeltrained on raw, unstemmed
data. It is unclear whether the poor performance on
questionansweringforthistest setisdueto a princi-
pallyworsetranslationqualityorthe(lackof)tenacity 14% partially correct answers!). However, using a
and willingness of the test subject to work her way system such as the one discussed in the paper is not
throughthesystemoutput. an option for actual information processing. Espe-
Allinall,wewereastonishedbytheamountofin- cially those subjects that had to deal with the output
formation that our test subjects were able to retrieve of systems trained on the smaller corpusexperienced
fromthematerialtheyreceived(thetoprecallforthe thetaskasutterlyfrustratingandwouldnotwanttodo
question answering task is 64%, plus an additional itagain.
4 Conclusions performanceimprovementwhentheMTsystems
are scaled up? One approach may show rapid
We have reported on our experience with rapidly improvements initially but also reach a plateau
buildingastatisticalMTsystemfromscratch. Within quickly, whereas another may show slow but
ca. 140translatorhours,wewereabletocreateapar- steadyimprovements.
allelcorpusofabout1300sentencepairswith24,000
Is there any potential for bootstrapping the re-
tokensontheTamilside,atanaveragetranslationrate (cid:15)
sourcecreationprocessbyusingknowledgethat
ofapproximately170Tamilwordsperhour.
canbeextractedfromlittleandpoordatatospeed
Veryclearly,theeffortneededtocreateparalleldata upthecreationofmoreandbetterdata?
isoneofthebiggestobstaclestotherapiddevelopment
These are some of the the questionsthat will need
ofstatisticalMTsystemsfornewlanguages.
tobeaddressedinfutureresearchonQuickMT.
Withtheoutputofasystemwhichusesatranslation
modeltrainedonthesmallamountofparalleldatathat 5 Acknowledgments
we created during the course of our experiment, hu-
man test subjects achieved a recall of over 50% on ThisresearchhasbeenfundedbytheDARPATIDES
the document retrieval task but generally performed program under grant No. N66001-00-1-8914. We
poorlyonquestionanswering(lessthan20%). would like to thank our translators as well as Fred
Gey of the University of California at Berkeley and
The addition of an additionalcorpusof 3,800 sen-
ThomasMaltenoftheUniversityofColognefortheir
tence pairs allowed us to estimate the benefits of in-
kindsupportofourinvestigationsintoTamil,andour
creasing the overall corpus size by roughly 300%.
testsubjectsfortheirtenacityandpatienceduringthis
BasedonourexperiencewithtranslatingtheTamilNet
experiment.
corpus, this additional effort would require an addi-
tional450translatorand36to48post-editorhours.
With the additional training data, we were able to References
produceoutputthatincreasedtheperformanceonour
Yaser Al-Onaizan, David Purdy, Jan Curin, Michael
evaluationtasks (documentretrievalandquestionan-
Jahr, Kevin Knight, John Lafferty, Dan Melamed,
swering)toupto93%fordocumentretrievaland64%
Noah A. Smith, Franz Josef Och, and David
forquestionanswering. Yarowsky. 1999. Statistical machine translation.
With respect to the scenario of “MT in a month”, Finalreport,CenterforLanguageandSpeechPro-
wecannowmakethefollowingcalculation: Ifweas- cessing,JohnHopkinsUniversity.
sumethattheaveragetranslatortranslatesatarateof
Peter F. Brown, Stephen A. Della Pietra, Vincent J.
170words/hourandisabletospend6-7hoursperday
Della Pietra, and Robert L. Mercer. 1993. The
on actual translations, then a translator can translate mathematicsof statistical machine translation: Pa-
about1000-1200wordsperday.Inordertotranslatea rameter estimation. Computational Linguistics,
corpusof100,000wordswithinonemonth(assuming 19(2):263–311.
afive-dayworkweek),wethereforeneedfourtofive
Peter Brown, John Cocke, Stephen Della Pietra, Vin-
full time translators. For this effort, we can expect
cent Della Pietra, Frederick Jelinek, Jennifer Lai,
atranslationsystemwhoseperformanceresemblesthe
andRobertMercer. 1995. Methodandsystem for
oneshowninourevaluation.
naturallanguagetranslation. U.S.Patent5,477,451,
This, of course, raises the following questions, Dec19.
whichweareonlyabletoaskbutnottoansweratthis
UlrichGermann,MichaelJahr, KevinKnight, Daniel
point.
Marcu, and Kenji Yamada. 2001. Fast decoding
Canthetranslationmodelandthealgorithmsfor andoptimaldecodingforMT. In39thAnnualMeet-
(cid:15)
statistical training be improved so that they re- ing of the Association for Computational Linguis-
quirelessdatatoproduceacceptableresults? tics(ACL-2001).
Aretheremoreefficientusesofscarceresources
Kevin Knight, Yaser Al-Onaizan, David Purdy, Jan
(cid:15)
(such as language experts and translators) for Curin, MichaelJahr, John Lafferty, Dan Melamed,
building a statistical (or any other) MT system NoahSmith,FranzJosefOch,andDavidYarowsky.
quickly,forexamplethecreationoflessbutmore 1999. EGYPT: a statistical machine translation
informative data, e.g. a parallel corpus with toolkit. http://www.clsp.jhu.edu/ws99/
alignmentsonthewordlevel,orthecompilation projects/mt/.
of a glossary/dictionary of the most frequently
Sanford B. Steever. 1990. Tamil and the Dravidian
usedterms?
languages. InBernardComrie,editor,TheWorld’s
Howdothevariousapproachescomparewithre- MajorLanguages,pages725–746.OxfordUniver-
(cid:15)
spect to the ratio of construction effort versus sityPress,NewYork,Oxford.