DTIC ADA457732: Learning Domain-Specific Transfer Rules: An Experiment with Korean to English Translation

8 Pages·0.04 MB·English

by  Defense Technical Information Center

Learning Domain-Specific Transfer Rules:
An Experiment with Korean to English Translation
Benoit Lavoie,MichaelWhite,andTanyaKorelsky
CoGenTex,Inc.
840HanshawRoad
Ithaca,NY14850,USA
benoit,mike,[email protected]
Abstract sisting of any source and target tree sub-patterns
matching asubset oftheparse features. Amorede-
We describe the design of an MT system that em-
tailed description ofthe differences can befound in
ploys transfer rules induced from parsed bitexts
(Lavoie et.al.,2001).
and present evaluation results. The system learns
lexico-structural transfer rules using syntactic pat-
2 OverallRuntime SystemDesign
tern matching, statistical co-occurrence and error-
OurKoreantoEnglishMTruntimesystemrelieson
driven filtering. In an experiment with domain-
thefollowing off-the-shelf softwarecomponents:
specificKoreantoEnglishtranslation, theapproach
yielded substantial improvements over three base-
Koreanparser Forparsing,weusedthewidecov-
linesystems.
erage syntactic dependency parser for Korean
developed by (Yoon et al., 1997). The parser
1 Introduction
wasnottrained onourcorpus.
In this paper, we describe the design of an MT
Transfercomponent For transfer of the Korean
system that employs transfer rules induced from
parses to English structures, we used the same
parsedbitextsandpresentevaluationresultsforKo-
lexico-structuraltransferframeworkas(Lavoie
rean to English translation. Our approach is based
etal.,2000).
on lexico-structural transfer (Nasr et. al., 1997),
and extends recent work reported in (Han et al., Realizer For surface realization of the transferred
2000)aboutKoreantoEnglishtransferinparticular. English syntactic structures, we used the Re-
Whereas Han et al. focus on high quality domain- alPro English realizer (Lavoie and Rambow,
specifictranslation usinghandcrafted transfer rules, 1997).
inthis workweinstead focus onautomating theac-
Thetrainingofthesystemisdescribedinthenext
quisition ofsuch rules.
twosections.
The proposed approach is inspired by example-
based machine translation (EBMT; Nagao, 1984;
3 DataPreparation
Sato and Nagao, 1990; Maruyama and Watanabe,
3.1 ParsesfortheBitexts
1992) and is similar to the recent works of (Mey-
ersetal.,1998)and(Richardsonetal.,2001)where In our experiments, we used a parallel corpus de-
transfer rules are also derived after aligning the rived from bilingual training manuals provided by
source and target nodes of corresponding parses. the U.S. Defense Language Institute. The corpus
However,while(Meyersetal.,1998)and(Richard- consistsofaKoreandialogof4,183sentencesabout
son et al., 2001) only consider parses and rules battlescenario messagetrafficandtheirEnglishhu-
withlexicallabelsandsyntactic roles,ourapproach mantranslations.
uses parses containing any syntactic information The parses for the Korean sentences were ob-
provided by parsers (lexical labels, syntactic roles, tained using Yoon’s parser, as in the runtime sys-
tense, number, person, etc.), and derives rules con- tem. The parses for the English human transla-
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
2002 2. REPORT TYPE 00-00-2002 to 00-00-2002
4. TITLE AND SUBTITLE 5a. CONTRACT NUMBER
Learning Domain-Specific Transfer Rules: An Experiment with Korean
5b. GRANT NUMBER
to English Translation
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
CoGenTex Inc,840 Hanshaw Road,Ithaca,NY,14850 REPORT NUMBER
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 7
unclassified unclassified unclassified
Standard Form 298 (Rev. 8-98)
Prescribed by ANSI Std Z39-18
Korean English
(S1) {i} {Ci-To-Reul} {Ta-Si} {Po-Ra}. Avg.sentencesize 6.43 9.36
this + map-accusative + again + look-imp Avg.parsesize 6.43 7.34
(D1) {po} [class=vbma ente={ra}] (
s1 {ci-to} [class=nnin2 ppca={reul}] (
s1 {i} [class=ande] ) Table 2: Average sizes for sentences and parses in
s1 {ta-si} [class=adco2] ) training set
(S2) Look at the map again.
(D2) look [class=verb mood=imp] (
Korean English
attr at [class=preposition] (
ii map [class=common_noun article=def] ) Avg. sentence size 7.04 10.58
attr again [class=adverb] ) Avg. parsesize 7.02 8.56
Table 3: Average sizes for sentences and parses in
Figure 1: Syntactic dependency representations for
testset
corresponding KoreanandEnglishsentences
Korean English the parses containing intra-sentential punctuation
Avg.sentencesize 9.08 13.26 marksotherthancommashadincorrectdependency
Avg.parsesize 8.96 10.77 assignments, due to the limitations of our conver-
sion grammars. Consequently, in our experiments
Table 1: Average sizes for sentences and parses in we have primarily focused on a higher quality sub-
corpus set of 1,483 sentence pairs, automatically selected
byeliminatingfromthecorpusallparsepairswhere
tions were derived from an English Tree Bank de- one of the parses contained more than 11 content
veloped in (Han et al., 2000). To enable the sur- nodesorinvolvedproblematicintra-sententialpunc-
face realization of the English parses via RealPro, tuation.
we automatically converted the phrase structures Wedividedthishigherqualitysubsetintotraining
of the English Tree Bank into deep-syntactic de- and test sets. Forthe test set, werandomly selected
pendencystructures(DSyntSs)oftheMeaning-Text 50 parse pairs containing at least 5nodes each. For
Theory (MTT) (Mel’cˇuk, 1988) using Xia’s con- the training set, we used the remaining 1,433 parse
verter (Xia and Palmer, 2001) and our own conver- pairs. Theaverage sentence lengths and parse sizes
sion grammars. The realization results of the re- for the training and test sets are represented in Ta-
sulting DSyntSs for our training corpus yielded a bles2and3.
unigram and bigram accuracy (f-score) of approxi-
3.3 CreatingtheBaselineTransferDictionary
mately95%and90%,respectively.
In our system, transfer dictionaries contain Ko-
A DSyntS is an unordered tree where all nodes
rean to English lexico-structural transfer rules de-
aremeaning-bearing andlexicalized. Sincetheout-
fined using the formalism described in (Nasr et.
putoftheYoonparserisquitesimilar,wehaveused
al., 1997), extended to include log likelihood ra-
itsoutput asis. Thesyntactic dependency represen-
tios (Manning and Schutze, 1999: 172-175). Sam-
tations for two corresponding Korean1 and English
ple transfer rules are illustrated in Section 4. The
sentences areshowninFigure1.
simplesttransfer rulesconsistofdirectlexicalmap-
3.2 TrainingandTestSetsofParsePairs pings, while the most complex may contain source
The average sentence lengths (in words) and parse and target syntactic patterns composed of multiple
nodesdefinedwithlexicaland/orsyntactic features.
sizes (in nodes) for the 4,183 Korean and English
Each transfer rule is assigned a log likelihood ratio
sentences inourcorpus aregiveninTable1.
calculated using thetraining parse set.
In examining the Korean parses, we found that
To create the baseline transfer dictionary for our
many of the larger parses, especially those contain-
experiments, we had three bilingual dictionary re-
ing intra-sentential punctuation, had incorrect de-
sources atourdisposal:
pendency assignments, incomplete lemmatization
or were incomplete parses. In examining the En-
Acorpus-basedhandcrafteddictionary: This
glish converted parses, we found that many of
dictionary wasmanuallyassembledby(Hanet
1Koreanisrepresentedinromanizedformatinthispaper. al.,2000)forthesamecorpususedhere. Note,
Korean English
Lexicalcoverage 92.18% 90.17% @KOREAN:
{po} [class=vbma] (
s1 $X [ppca={reul}] )
Table4: Concurrent lexical coverage oftraining set @ENGLISH:
bybaseline dictionary look [class=verb] (
attr at [class=preposition] (
ii $X ))
however, that it was developed for different @-2xLOG_LIKELIHOOD: 12.77
parse representations, and with an emphasis
primarily onthe lexical coverage ofthe source
Figure 2: Transfer rule for English lexicalization
parses, rather than the source and target parse
andpreposition insertion
pairs.
Acorpus-basedextracted dictionary: This dic-
4 Transfer RuleInduction
tionary was automatically created from our
corpus bytheRALIgroupfromtheUniversity The transfer rule induction process has the follow-
of Montreal. Since the extraction heuristics ing steps described below (additional details can
did not handle the rich morphological suffixes alsobefound in(Lavoie et.al.,2001)):
of Korean, the extraction results contained
inflected wordsratherthan lexemes. Nodes of the corresponding source and target
(cid:1)
Awidecoverage dictionary: This dictionary of parses are aligned using the baseline transfer
70,300entries wascreated bySystran,without dictionary and some heuristics based on the
regardtoourcorpus. similarity of part-of-speech and syntactic con-
text.
We processed and combined these resources as
Transfer rule candidates are generated based
follows: (cid:1)
on the sub-patterns that contain the corre-
First,wereplaced the inflected wordswithun- spondingalignednodesinthesourceandtarget
(cid:1) inflected lexemes using Yoon’s morphological parses.
analyzerandawidecoverageEnglishmorpho-
The transfer rule candidates are ordered based
logical database (KarpandSchabes, 1992). (cid:1)
ontheir likelihood ratios.
Second, we merged all morphologically an-
(cid:1) Thetransfer rulecandidates arefiltered,oneat
alyzed entries after removing all non-lexical (cid:1)
a time, in the order of the likelihood ratios, by
features, since these features generally did not
removingthoserulecandidatesthatdonotpro-
matchthosefound intheparses.
duceanoverallimprovementintheaccuracyof
Third,wematchedtheresultingtransferdictio- thetransferred parses.
(cid:1)
naryentrieswiththetrainingparseset,inorder
todetermineforeachentryallpossiblepart-of- Figures 2 and 3 show two sample induced rules.
speechinstantiations anddependency relation- The rule formalism uses notation similar to the
ships. For each distinct instantiation, we cal- syntactic dependency notation shown in Figure 1,
culated aloglikelihood ratio. augmented with variable arguments prefixed with
Finally, we created a baseline dictionary us- $ characters. These two lexico-structural rules can
(cid:1) ingtheinstantiatedruleswhosesourcepatterns beusedtotransferaKoreansyntacticrepresentation
hadthebestloglikelihood ratios. forci-to-reulpo-ratoanEnglishsyntacticrepresen-
tation for look at the map. The first rule lexicalizes
Table4illustratestheconcurrentlexicalcoverage the English predicate and inserts the corresponding
ofthetrainingsetusingtheresultingbaselinedictio- prepositionwhilethesecondruleinsertstheEnglish
nary, i.e. the percentage of nodes covered by rules imperative attribute.
whose source and target patterns both match. Note
4.1 AligningtheParseNodes
that since the baseline dictionary contained some
noise, weallowed induced rules to override ones in To align the nodes in the source and target parse
thebaseline dictionary whereapplicable. trees, we devised a new dynamic programming
@KOREAN: @KOREAN:
$X [class=vbma ente={ra}] $X1 [aid=$1] (
@ENGLISH: $R1 $X2 [aid=$2] )
$X [class=verb mood=imp] @ENGLISH:
@-2xLOG_LIKELIHOOD: 33.37 $Y1 [aid=$1] (
$R2 $Y2 (
$R3 $Y3 [aid=$2] ) )
Figure3: Transferruleforimperative forms
Figure4: Alignment constraint
alignmentalgorithmthatperformsatop-down,bidi-
rectionalbeamsearchfortheleastcostmappingbe- Each node of a candidate transfer rule must have
tween these nodes. The algorithm is parameterized its relation attribute (relationship with its governor)
by the costs of (1) aligning two nodes whose lex- specifiedifitisaninternalnode,otherwisethisrelation
emes are not found in the baseline transfer dictio- mustnotbespecified:
nary; (2) aligning two nodes with differing parts of
e.g.
speech;(3)deletingorinsertinganodeinthesource
$X1 ( $R $X2 )
ortargettree;and(4)aligningtwonodeswhoserel- (cid:1)
ativelocations differ.
Figure5: Independent attribute constraint
To determine an appropriate part of speech cost
measure,wefirstextractedasmallsetofparsepairs
thatcouldbereliablyalignedusinglexicalmatching necessary forKoreantoEnglish transfer sofar).
alone, and then based the cost measure on the co-
4.2.2 Attributeconstraints
occurrence counts of the observed parts of speech
Attribute constraints are used to limit the space of
pairings. Theremaining costs weresetbyhand.
possible transfer rule candidates that can be gen-
Asaresultofthealignmentprocess,alignmentid
erated from the sub-trees satisfying the alignment
attributes (aid) are added to the nodes of the parse
constraints. Candidatetransferrulesmustsatisfyall
pairs. Some nodes may be in alignment with no
oftheattributeconstraints. Attributeconstraintscan
other node, such as English prepositions not found
bedivided intotwotypes:
intheKoreanDSyntS.
independentattributeconstraints,whosescope
4.2 GeneratingRuleCandidates
(cid:1)
coversonlyonepartofacandidatetransferrule
Candidate transfer rulesaregenerated byextracting
and which are the same for the source and tar-
source and target treesub-patterns from thealigned
getparts;
parsepairsusingthetwosetofconstraintsdescribed
below. concurrent attribute constraints, whose scope
(cid:1)
extends toboththesource andtarget parts ofa
4.2.1 Alignmentconstraints
candidate transfer rule.
Figure 4 shows an example alignment constraint.
This constraint, which matches the structural pat- Figures 5 and 6 give examples of an indepen-
ternsofthetransferruleillustrated inFigure2,uses dentattributeconstraintandofaconcurrentattribute
theaidalignment attribute toindicate thatinaKo- constraint. As with the alignment constraints, we
rean and English parse pair, any source and target suggest that a relatively small number of attribute
sub-treesmatchingthisalignmentconstraint(where constraints is necessary to generate most of the de-
$X1and$Y1arealigned,i.e.havethesameaidat- siredrules foragivenlanguage pair.
tribute values,andwhere$X2and$Y3arealigned)
4.3 OrderingRuleCandidates
can be used as a point of departure for generat-
ing transfer rule candidates. We suggest that align- Inthenextstep,transferrulecandidates areordered
mentconstraints such asthisonecanbeusedtode- asfollows. First,theyaresortedbytheirdecreasing
fine most of the possible syntactic divergences be- log likelihood ratios. Second, if two or more can-
tweenlanguages(Dorr,1994),andthatonlyahand- didate transfer rules have the same log likelihood
ful of them are necessary for two given languages ratio, ties are broken by a specificity heuristic, with
(wehaveidentified11generalalignmentconstraints the result that more general rules are ordered ahead
Ina candidatetransferrule, inclusionofthelexemesof
twoalignednodesmustbedoneconcurrently: @KOREAN:
{po} [class=vbma ente={ra}] (
s1 $X [ppca={reul}] )
e.g.
@ENGLISH:
$X [aid=$1] look [class=verb mood=imp] (
and attr at [class=preposition] (
$Y [aid=$1] ii $X ) )
@-2xLOG_LIKELIHOOD: 11.40
e.g.
[aid=$1]
and (cid:1)
[aid=$1] Figure 7: Transfer rule for English imperative with
(cid:1) lexicalization andpreposition insertion
Figure6: Concurrent attribute constraint
tribute constraints, an initial list of 801,674 trans-
of more specific ones. The specificity of a rule is
fer rule candidates was then generated from these
defined to be the following sum: the number of at-
sub-tree pairs. The initial list was subsequently re-
tributes foundinthesourceandtargetpatterns,plus
duced to 32,877 unique transfer rule candidates by
1foreachforeachlexemeattributeandforeachde-
removing duplicates and by eliminating candidates
pendency relationship. In our initial experiments,
that had the same source pattern as another candi-
thissimpleheuristic hasbeensatisfactory.
date with a better log likelihood ratio. After filter-
4.4 FilteringRuleCandidates ing,2,133ofthesetransfer rulecandidates wereac-
cepted. Weexpectthatthenumberofacceptedrules
Oncethecandidatetransferruleshavebeenordered,
perparsepairwilldecreasewithlargertrainingsets,
error-drivenfilteringisusedtoselectthosethatyield
though thisremains tobeverified.
improvements over the baseline transfer dictionary.
Thealgorithmworksasfollows. First,intheinitial- The rule illustrated in Figure 3 was accepted as
ization step, the set of accepted transfer rules is set the 65th best transfer rule with a log likelihood ra-
to just those appearing in the baseline transfer dic- tioof 33.37, andthe rule illustrated inFigure 2was
tionary. Thecurrenterrorrateisalsoestablished,by accepted as the 189th best transfer rule candidate
applying these transfer rules to all the source struc- with a log likelihood ratio of 12.77. An example
tures andcalculating theoverall difference between of a candidate transfer rule that was not accepted is
the resulting transferred structures and the target the one that combines the features of the two rules
parses, using a tree accuracy recall and precision mentionedabove,illustratedinFigure7. Thistrans-
measure(determinedbycomparingthefeaturesand ferrulecandidatehadalowerloglikelihoodratioof
dependency relationships in the transferred parses 11.40; consequently, it is only considered after the
and corresponding target parses). Then, in a sin- tworulesmentionedabove,andsinceitprovidesno
glepassthroughtheordered listofcandidates, each further improvement upon these two rules, it is fil-
transfer rule candidate is tested to see if it reduces teredout.
the error rate. During each iteration, the candidate
Inaninformalinspection ofthetop100accepted
transferruleisprovisionally addedtothecurrentset
transfer rules, we found that most of them appear
of accepted rules and the updated set is applied to
to be fairly general rules that would normally be
allthesourcestructures. Iftheoveralldifferencebe-
found in a general syntactic-based transfer dictio-
tweenthetransferredstructuresandthetargetparses
nary. In looking at the remaining rules, we found
is lower than the current error rate, then the can-
thattherulestendedtobecomeincreasinglycorpus-
didate is accepted and the current error rate is up-
specific.
dated; otherwise, the candidate is rejected and re-
movedfromthecurrent set. The induction results were obtained using a Java
implementation of the induction component. Mi-
4.5 DiscussionofInducedRules
crosoft SQLServer was used to count and dedupli-
In our experiments, the alignment constraints cate the rule candidates. The data preparation and
yielded22,881sourceandtargetsub-treepairsfrom induction processes took about 12 hours on a 300
the training set of 1,433 parse pairs. Using the at- MHzPCwith256MBRAM.
5 Evaluation System 1-gPrec 2-gPrec 3-gPrec 4-gPrec
Babelfish 0.3814 0.1207 0.0467 0.0193
5.1 SystemsCompared
GIZA++/RW 0.1894 0.0173 0.0 0.0
Babelfish As a first baseline, we used Babelfish LexOnly 0.4234 0.1252 0.0450 0.0145
fromSystran,acommerciallargecoverageMT Lex+Induced 0.4725 0.1618 0.0577 0.0185
system supporting Korean to English transla-
tion. This system was not trained on our cor- Table5: BleuN-gramprecision scores
pus.
GIZA++/RW As a second baseline, we used an
System BleuScore
off-the-shelf statistical MT system, consisting
Babelfish 0.0802
of the ISI ReWrite Decoder (Germann et al., GIZA++/RW NA
2001) together with a translation model pro- LexOnly 0.0767
duced by GIZA++ (Och and Ney, 2000) and Lex+Induced 0.0950
a language model produced by the CMU Sta-
tistical Language Modeling Toolkit (Clarkson Table6: Bleuoverallprecision scores
andRosenfeld,1997). Thissystemwastrained
onourcorpus only.
LexOnly As a third baseline, we used our system 5.3 HumanEvaluationResults
withthebaselinetransferdictionaryasthesole
transfer resource. For the human evaluation, we asked two English
native speakers to rank the quality of the transla-
Lex+Induced Wecomparedthethreebaselinesys-
tion results produced by the Babelfish, Lex Only
tems against our complete system, using the
and Lex+Induced, with preference given to fidelity
baseline transfer dictionary augmented with
over fluency. (The translation results of the statisti-
theinduced transfer rules.
cal system were not yet available when the evalua-
We ran each of the four systems on the test set tion was performed.) A rank of 1 was assigned to
of 50 Korean sentences described in Section 3.2 thebesttranslation, arankoftwotothesecond best
and compared the resulting translations using the andarankof3tothethird,withtiesallowed.
automatic evaluation and the human evaluation de-
Table 7 shows the pairwise comparisons of the
scribed below.
three systems. The top section indicates that the
5.2 AutomaticEvaluationResults Babelfish and Lex Only baseline systems are es-
sentially tied, with neither system preferred more
Fortheautomaticevaluation,weusedtheBleumet-
frequently than the other. In contrast, the middle
ric from IBM (Papineni et al., 2001). The Bleu
andbottomsections showthatoursystemimproves
metriccombinesseveralmodifiedN-gramprecision
substantiallyoverbothbaselinesystems;moststrik-
measures (N=1to4), and uses brevity penalties to
ingly, our system (Lex+Induced) was preferred al-
penalize translations that are shorter than the refer-
most 20% morefrequently than the Babelfish base-
encesentences.
line(46%to27%,withties27%ofthetime).
Table 5 shows the Bleu N-gram precision scores
for each of the four systems. Our system
(Lex+Induced)hadbetterprecisionscoresthaneach
SystemPairComparison Result
of the baseline systems, except in the case of 4-
BabelfishbetterthanLexOnly 37%
grams, where it slightly trailed Babelfish. The sta-
LexOnlybetterthanBabelfish 36%
tistical baseline system (GIZA++/RW) performed
BabelfishsameasLexOnly 27%
poorly,asmighthavebeenexpectedgiventhesmall
BabelfishbetterthanLex+Induced 27%
amountoftraining data.
Lex+InducedbetterthanBabelfish 46%
Table 6 shows the Bleu overall precision scores. BabelfishsameasLex+Induced 27%
Our system (Lex+Induced) improved substantially LexOnlybetterthanLex+Induced 18%
over both the Lex Only and Babelfish baseline sys- Lex+InducedbetterthanLexOnly 41%
tems. The score for the statistical baseline system LexOnlysameasLex+Induced 41%
(GIZA++/RW) is not meaningful, due to the ab-
sence of3-gramand4-gram matches. Table7: Humanevaluation results
6 Conclusionand Future Work Benoit Lavoie and Owen Rambow. 1997. RealPro – A Fast,
Portable Sentence Realizer. InProceedings of the Confer-
InthispaperwehavedescribedthedesignofanMT
enceonAppliedNaturalLanguage Processing(ANLP’97),
system based on lexico-structural transfer rules in-
Washington,DC.
duced from parsed bitexts. In a small scale exper- BenoitLavoie,RichardKittredge, TanyaKorelsky, andOwen
iment with Korean to English translation, we have Rambow. 2000. A framework for MT and multilingual
demonstrated a substantial improvement over three NLG systems based on uniform lexico-structural process-
ing. InProceedingsofANLP/NAACL2000, Seattle,Wash-
baseline systems, including a nearly 20% improve-
ington.
ment in the preference rate for our system over Ba-
BenoitLavoie,MichaelWhite,andTanyaKorelsky. 2001. In-
belfish (which was not trained on our corpus). Al- ducing Lexico-Structural Transfer Rules from Parsed Bi-
though our experimentation was aimed at Korean texts. In Proceedings of the ACL 2001 Workshop on
Data-driven Machine Translation, pages 17–24, Toulouse,
toEnglish translation, webelieve that our approach
France.
canbereadily applied toother language pairs.
Christopher D. Manning and Hinrich Schutze. 1999. Foun-
It remains for future work to explore how well dations of Statistical Natural Language Processing. MIT
the approach would fare with a much larger train- Press.
ing corpus. One foreseeable problem concerns the H. Maruyama and H. Watanabe. 1992. Tree cover Search
AlgorithmforExample-BasedTranslation. InProceedings
treatment of lengthy training sentences: since the
ofthe Fourth International Conference on Theoretical and
number of transfer rule candidates generated grows
Methodological Issues in Machine Translation (TMI’92),
exponentially with the size of the parse tree pairs, pages173–184.
refinements will be necessary in order to make use Yuji Matsumoto, Hiroyuki Hishimoto, and Takehito Utsuro.
of complex sentences; one option might be to auto- 1993. StructuralMatchingofParallelTexts. InProceedings
ofthe31stAnnualMeetingsoftheAssociationforCompu-
maticallychunklongersentences intosmallerunits.
tationalLinguistics(ACL’93),pages23–30.
IgorMel’cˇuk. 1988. Dependency Syntax. StateUniversityof
Acknowledgements
NewYorkPress,Albany,NY.
WethankRichardKittredgeforhelpfuldiscussion,DarylMc- AdamMeyers,RomanYangarber,RalphGrishman,Catherine
CulloughandTedCaldwellfortheirhelpwithevaluationand Macleod, and Antonio Moreno-Sandoval. 1998. Deriving
Fei Xia for her assistance with the automatic conversion of TransferRulesfromDominance-PreservingAlignments. In
the phrase structure parses to syntactic dependency represen- ProceedingsofCOLING-ACL’98,pages843–847.
tations.WealsothankChung-hyeHan,ChulwooPark,Martha Makoto Nagao. 1984. Aframework of a mechanical transla-
Palmer, and Joseph Rosenzweig for the handcrafted Korean- tionbetweenJapeneseandEnglishbyanalogyprinciple. In
Englishtransferdictionary,andGrahamRussellforthecorpus- A. Elithorn and R. Banerji, editors, Artificial and Human
based extracted transfer dictionary. This work has been par- Intelligence.NATOPublications.
tiallysupportedbyDARPATIDEScontractno.N66001-00-C- Alexis Nasr, Owen Rambow, Martha Palmer, and Joseph
8009. Rosenzweig. 1997. Enriching lexical transfer with cross-
linguisticsemanticfeatures. InProceedingsoftheInterlin-
References
guaWorkshopattheMTSummit,SanDiego,California.
PhilipClarksonandRonaldRosenfeld. 1997. StatisticalLan- FranzJosefOchandHermannNey. 2000. ImprovedStatistical
guage Modeling Using the CMU-Cambridge Toolkit. In AlignmentModels. InProceedingsofACL’00,pages440–
ProceedingsofEurospeech’97. 447,Hongkong,China.
Bonnie Dorr. 1994. Machine translation divergences: A for- Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing
maldescriptionandproposedsolution. ComputationalLin- Zhu. 2001. Bleu: a Method for Automatic Evaluation of
guistics,20(4):597–635. MachineTranslation. InIBMtechnicalreport,RC22176.
Ulrich Germann, Michael Jahr, Kevin Knight, Daniel Marcu StephenD.Richardson,WilliamB.Dolan,ArulMenezes,and
andKenjiYamada. 2001. FastDecoding andOptimalDe- MonicaCorston-Oliver. 2001. OvercomingtheCustomiza-
codingforMachineTranslation. InProceedingsofACL’01, tion Bottleneck using Example-based MT. InProceedings
Toulouse,France. oftheACL2001WorkshoponData-drivenMachineTrans-
Chung hye Han, Benoit Lavoie, Martha Palmer, Owen lation,pages9–16,Toulouse,France.
Rambow, Richard Kittredge, Tanya Korelsky, Nari Kim, Satoshi Sato and Makoto Nagao. 1990. Toward Memory-
and Myunghee Kim. 2000. Handling Structural Diver- basedTranslation. InProceedingsofthe13thInternational
gences and Recovering Dropped Arguments in a Korean- Conference on Computational Linguistics (COLING’90),
EnglishMachineTranslationSystem. InProceedingsofthe pages247–252.
FourthConferenceonMachineTranslationintheAmericas Fei Xia and Martha Palmer. 2001. Converting Dependency
(AMTA’00),MisionDelSol,Mexico. StructurestoPhraseStructures. InNotesoftheFirstHuman
Daniel Karp and Yves Schabes. 1992. A Wide Coverage LanguageTechnologyConference,SanDiego,California.
Public Domain Morphological Analyzer for English. In Juntae Yoon, Seonho Kim, and Mansuk Song. 1997. New
Proceedings of the Fifteenth International Conference on parsingmethodusingglobalassociationtable. InProceed-
Computational Linguistics (COLING’92), pages 950–955, ingsofthe5thInternationalWorkshoponParsingTechnol-
Nantes,France. ogy.