CS460/626 : Natural Language Processing/Speech NLP and the WebProcessing/Speech, NLP and the Web

(Lecture 36syllabification and transliteration)transliteration)

Pushpak BhattacharyyaPushpak BhattacharyyaCSE Dept., IIT Bombay

11th A il 201111th April, 2011

Access from one language to another

Language L1 Language L2

T l ti T lit tiTranslation Transliteration

Gandhiji Boy

Cross Lingual Query

Query Translatio

n

Search

Birthplace of Gandhi Search

Engine

Ga d

English Document

Translate Document

Output

Sound based transliterationSound based transliteration

TransliterationScript Script ()(Gandhiji)

Transliteration using pronunciationConvert script to soundConvert sound to scriptConvert sound to script

Similar to interlingua based MT

MeaningSentence in source Language

Sentence in target LanguageLanguage Language

Two approaches to transliterationTwo approaches to transliterationSyllabification

Rule Machine Based Learning Based

Word SyllabifiedApple ap pleFundamental fun da men tal

Use GIZA++ Moses on parallel corpus

Phoneme-based approachPhoneme-based approach

Word in Word in P(w )Source language Target language

P( ps | ws) P ( wt | pt )

P(wt)

Pronunciationin

Source language

PronunciationIn

target languageP ( pt | ps )

Wt* = argmax (P (wt). P (wt | pt) . P (pt | ps) . P (ps | ws) )

Note: Phoneme is the smallest linguistically distinctive unit of sound.

Basics of syllablesBasics of syllablesSyllable is a unit of spoken languageSyllable is a unit of spoken languageconsisting of a single uninterrupted soundformed generally by a Vowel and preceded orfollowed by one or more consonants.

Vowels are the heart of a syllable (MostSonorous Element) (svayam raajate itisvaraH)Consonants act as sounds attached to

lvowels.

Syllable structure

A syllable consists of 3 major parts:-Onset (C)( )Nucleus (V)Coda (C)( )

Vowels sit in the Nucleus of a syllableConsonants may get attached as OnsetConsonants may get attached as Onset or Coda.Basic structure - CVBasic structure CV

Possible syllable structuresThe Nucleus isThe Nucleus is always presentOnset and CodaOnset and Coda may be absentPossiblePossible structures

VVCVVCCVC

Introduction to sonority theoryThe Sonority of a sound is its loudness

relative to other sounds with the same length stress and speech length, stress and speech.

Some sounds are more sonorousSome sounds are more sonorousWords in a language can be divided into syllablesySonority theory distinguishes syllables on the basis of sounds.

Sonority hierarchyDefined on the basis of amount of sound associatedThe sonority hierarchy is as follows:-

Vowels (a, e, i, o, u)Li id ( l )Liquids (y, r, l, v)Nasals (n, m)Fricatives (s z f sh th etc )Fricatives (s, z, f,..sh, th etc.)Affricates (ch, j)Stops (b, d, g, p, t, k)p ( , , g, p, , )

Sonority scaleObstruents canObstruents can be further classified into:-classified into:

FricativesAffricatesAffricatesStops

Sonority theory & syllablesA Syllable is a cluster of sonority, defined by a sonority peak acting as a structural magnet to the surrounding lower sonority elements the surrounding lower sonority elements.

Represented as waves of sonority orRepresented as waves of sonority or Sonority Profile of that syllable

Nucleus

Onset Coda

Sonority sequencing principle

The Sonority Profile of a syllable must rise until its Peak(Nucleus), and then fall.( ),

PeakPeak(Nucleus)

Onset Coda

Maximal onset principleThe Intervocalic consonants are maximally

assigned to the Onsets of syllables inf it ith U i l d Lconformity with Universal and Language-

Specific Conditions.Determines underlying syllable divisionDetermines underlying syllable divisionExample

DIPLOMADIPLOMADIP LO MA & DI PLOMAMA

Syllable Structure: a more detailed look

Count of no. of syllables in a word is roughly/intuitively the no. of vocalic segments in a word.Thus, presence of a vowel is an obligatory element in the structure of a syllable. This vowel is called nucleus.Basic Configuration: (C)V(C).Part of syllable preceding the nucleus is called the onset.Elements coming after the nucleus are called the codaElements coming after the nucleus are called the coda.Nucleus and coda together are referred to as the rhyme.

S Syllable, O OnsetR Rhyme, N NucleusCo Coda

Syllable Structure: ExamplesSyllable Structure: Examples

word

sprint

Syllable Structure: ExamplesSyllable Structure: Examples

may

No Coda.

opt

No Onset.No Onset.

air

No Coda, No Onset.

Syllable StructureSyllable StructureOpen Syllable: ends in vowelClosed syllable: ends in consonant or consonant cluster

Light Syllable: A syllable which is open and ends in a shortLight Syllable: A syllable which is open and ends in a short vowel

General Description CV.Example, air.Example, air .

Heavy Syllable: Closed syllables or syllables ending in diphthong

Example: optExample: optExample, may

Syllabification: Determining Syllable Boundaries

Given a string of syllables (word), what is the coda of one and the onset of another?

In a sequence such as VCV, where V is any vowel and C is any consonant, is the medial C the coda of the first syllable (VC.V) or the onset of the second syllable (V.CV)? ( ) y ( )

To determine the correct groupings, there are some rules, two of them being the most important and significant:two of them being the most important and significant:

Maximal Onset Principle, Sonority Hierarchy

Rule Based SyllabificationRule Based Syllabification

(This and statistical syllabification and application to Transliteration: DDP work by

Ankit Aggrwal 2009)Ankit Aggrwal, 2009)

PhonotacticsPhonotactics

Guidesconstructionofonsetsandcodas

Restrictscombinationofphonemes.

Ingeneral,rulesoperatearoundthesonorityhierarchy.E.g.,Fricative/s/isloweronthesonorityhierarchythanthelateral/l/,sothecombination /sl/ is permitted in onsets and /ls/ is permitted in codas Oppositecombination/sl/ispermittedinonsetsand/ls/ispermittedincodas.Oppositeisnotallowed.

Thus,slipsandpulsearepossibleEnglishwords.

lsipsandpuslarenotpossible.

Constraints on Onsets1

Oneconsonant:Only//cannotoccurinsyllableinitialposition.

Twoconsonant:Werefertothescaleofsonority.Sequencernisruledoutsincethereisadecreaseofsonority.

MinimalSonorityDistance:Distanceinsonoritybetweenthefirstandthesecondelementintheonsetmustbeofatleast2degrees.

Thus,onlyalimitedno.ofpossibletwoconsonantclusters.

Threeconsonant:Restrictedtolicensedtwoconsonantonsetsprecededby/s/.

Also,/s/canonlybefollowedbyavoicelesssound.

Therefore,only/spl/,/spr/,/str/,/skr/,/spj/,/stj/,/skj/,/skw/,/skl/,/ j/ ill b ll d ( li )/smj/willbeallowed.(splinter,spray,strongetc.)

While/sbl/,/sbr/,/sdr/,/sgr/,/sr/willberuledout.

Constraints on Onsets2

Possible2consonatclustersinanOnset

Constraints on Coda1Constraints on Coda

Constraints on Coda2Constraints on Coda

Other ConstraintsOther Constraints

Nucleus:Thefollowingcanoccurasnucleus:Allvowelsounds(monophthongs aswellasdiphthongs).

Syllabic:yBoththeonsetandthecodaareoptional(asseenpreviously).

/j/attheendofanonset(/pj/,/bj/,/tj/,/dj/,/kj/,/fj/,/vj/,/j/,/sj/,/zj/,/hj/,/mj/,/nj/,/lj/,/spj/,/stj/,/skj/)mustbefollowedby/u/or//.

Longvowelsanddiphthongsarenotfollowedby//.//israreinsyllableinitialposition.

Stop+/w/before/u,,,a/areexcluded.

Implementation1p

1. Identify the first nucleus (single or multiple consecutive vowels) in y ( g p )the word.

2. Parse all the preceding consonants as the onset of the first syllable.3. Find the next nucleus in the word. If unsuccessful in finding another

l 'll i l th t t th i ht f thnucleus, we'll simply parse the consonants to the right of the current nucleus as the coda of the first syllable, else we will move to the next step. Now the consonant cluster between these two nuclei have to be divided in two parts, one as coda of 1st syllable and the other onset of 2nd syllable.

4. If the no. of consonants in the cluster is one, it will be the onset of the second nucleus as per the Maximal Onset Principle and Constrains on OnsetConstrains on Onset.

5. If it is two, check whether both of these can go to the onset of the second syllable as per the allowable onsets.

Implementation2p6. If not , the first consonant will be the coda of the first syllable and

the second consonant will be the onset of the second syllable.the second consonant will be the onset of the second syllable.

7. If the no. of consonants in the cluster is three, check if they can b th t f th d ll bl if t h k f th l tbe the onset of the second syllable, if not, check for the last two, if not, parse only the last consonant as the onset of the second syllable.

8. If the no. of consonants in the cluster is more than three, except the last three consonants, parse all the consonants as the coda of the first syllable. With the remaining 3 consonants, apply step 7step 7.

9. Continue the procedure for the rest of the string.

AccuracyAccuracy=(no.ofwordscorrectlysyllabifiedx100) %

no.ofwords

10,000easywordswerechosentoestablishproofofconcept

91wordswerefoundtobeincorrectlysyllabified.

Accuracy:99.09%.

Highprecision,lowrecall(acharacteristicofknowledgebasedAI)

Statistical syllabificationStatistical syllabification

Statistical Machine Syllabification

DataSourcesofdata

1. ElectionCommissionofIndia(ECI)NameList:ThiswebsourceprovidesnativeIndiannameswritteninbothEnglishandHindi.

( )2. DelhiUniversity(DU)StudentList:ThiswebsourcesprovidesnativeIndiannameswritteninEnglishonly.Thesenamesweremanuallytransliteratedforthepurposesoftrainingdata.

3. IndianInstituteofTechnology,Bombay(IITB)StudentList:TheAcademicd a s u e o ec o ogy, o bay ( ) S ude sOfficeofIITBprovidedthisdataofstudentswhograduatedintheyear2007.

4. NamedEntitiesWorkshop(NEWS)2009EnglishHindiparallelnames:Ali t f i d b t E li h d Hi di f i 11k i id dlistofpairednamesbetweenEnglishandHindiofsize11kisprovided.

Westartwith8,000randomlychosennamesfromtheECINameli d ll bif h lllistandsyllabifythemmanually.

Statistical Machine Syllabification

EffectofTrainingFormatSyllableseparatedFormat

Source Targetsudakar sudakarchhagan chhaganj i t h ji t hjitesh jiteshnarayan narayanshiv shivmadhav madhavmohammad mohammadj a y a n t e e d e v i ja yan tee de vi

SyllablemarkedFormat

jayanteedevi jayanteedevi

Source Targetsudakar su_da_karchhagan chha_ganjitesh ji_teshnarayan na_ra_yanshiv shivmadhav ma_dhavmohammad mo_ham_madjayanteedevi ja_yan_tee_de_vi

Statistical Machine Syllabification

Syllableseparated Syllablemarked

80%85%90%95%100%

veAccuracy

y p y

60%65%70%75%80%

1 2 3 4 5

Cumulativ

AccuracyLevel

Top1Accuracy:71.8%(Syllableseparated);80.5%(Syllablemarked)

T 5 A 83 4% (S ll bl d) 90 4% (S ll blTop5Accuracy:83.4%(Syllableseparated);90.4%(Syllablemarked)

Statistical Machine Syllabification

EffectofDataSize4differentexperimentswereperformed:

1. 8k:NamesfromtheECINamelist.

2. 12k:Anadditional4knamesweresyllabifiedtoincreasethedatasize.

3. 18k:IITBStudentListandtheDUStudentListwasincludedandsyllabified.

4. 23k:SomemorenamesfromECINameListandDUStudentListweresyllabifiedandthisdataactsasthefinaldata forus.

93.8%97.5% 98.3% 98.5% 98.6%

90.0%

95.0%

100.0%

ccuracy

8k 12k 18k 23k

70.0%

75.0%

80.0%

85.0%

CumulativeA

1 2 3 4 5

AccuracyLevel

Statistical Machine Syllabification

EffectofLanguageModelngramOrder

99.0%

3gram 4gram 5gram 6gram 7gram

95.0%

97.0%

ccuracy

89.0%

91.0%

93.0%

CumulativeAc

85.0%

87.0%

1 2 3 4 5

AccuracyLevel

Eff t f S ll bifi tiEffect of Syllabification on TransliterationTransliteration

Transliteration using syllabification

Input String

Moses

Syllabified data

(sourceside)Automatic S llabifieMoses Syllabifier

S ll bifi d St i

Learnt syllabifier

Syllabified StringLearnt

transliterator

Moses Automatic TransliteratorSyllabified

data(parallel)(parallel)

Transliterated String

Statistical Machine Transliteration

100%

Syllableseparated Syllablemarked

65%70%75%80%85%90%95%100%

tive

Accuracy

45%50%55%60%65%

1 2 3 4 5 6

Cumulat

A L l

Top1Accuracy:60.1%(Syllableseparated);50.2%(Syllablemarked)

Top 5 Accuracy: 87 2% (Syllable separated); 79 3% (Syllable marked)

AccuracyLevel

Top5Accuracy:87.2%(Syllableseparated);79.3%(Syllablemarked)

No syllabification: Baseline Performance

Topn CorrectCorrect%age

Cumulative%age

Additional Additional

1 1,868 41.5% 41.5%2 520 11.6% 53.1%3 246 5.5% 58.5%4 119 2 6% 61 2%4 119 2.6% 61.2%5 81 1.8% 63.0%

Below5 1,666 37.0% 100.0%4,500

TheTop5Accuracyisjust63.0%whichismuchlowerthanrequired.

40

Manual, i.e., ideal syllabification: , , yskyline Performance

Syllableseparated Syllablemarked

80%85%90%95%100%

veAccuracy

y p y

60%65%70%75%80%

1 2 3 4 5

Cumulativ

AccuracyLevel

Top1Accuracy:71.8%(Syllableseparated);80.5%(Syllablemarked)

T 5 A 87 4% (S ll bl d) 90 4% (S ll blTop5Accuracy:87.4%(Syllableseparated);90.4%(Syllablemarked)

41

Obsevations

Transliteration likely to be helped by syllabificationsyllabificationRule based syllabification: high precision low recallprecision low recallStatistical syllabification: reasonable accuracy at top 5accuracy at top-5

ReferencesReferences1. NasreenAbdulJaleelandLeahS.Larkey.Statisticaltransliterationforenglish

arabiccrosslanguageinformationretrieval.InConferenceonInformationandKnowledgeManagement,pages139146,2003.

2. AnnK.FarmerAndrianAkmajian,RichardM.DemersandRobertM.Harnish.Linguistics:AnIntroductiontoLanguageandCommunication.MITPress,5thgu st cs t oduct o to a guage a d Co u cat o ess, 5tedition,2001.

3. AssociationforComputerLinguistics.CollapsedConsonantandVowelModels:NewApproachesforEnglishPersianTransliterationandBackTransliteration,20072007.

4. SlavenBilacandHozumiTanaka.Directcombinationofspellingandpronunciationinformationforrobustbacktransliteration.InConferencesonComputationalLinguisticsandIntelligentTextProcessing,pages413424,2005.

5. IanLaneBingZhao,NguyenBachandStephanVogel.Aloglinearblocktransliterationmodelbasedonbistreamhmms.HLT/NAACL2007,2007.

References6. H.L.JinandK.F.Wong.AChinesedictionaryconstructionalgorithmfor

informationretrieval.InACMTransactionsonAsianLanguageInformationProcessing,pages281296,December2002.

7. K.KnightandJ.Graehl.Machinetransliteration.InComputationalLinguistics,pages24(4):599612,Dec.1998.pages ( ) 599 6 , ec 998

8. LeeFengChienLongJiang,MingZhouandChenNiu.Namedentitytranslationwithwebminingandtransliteration.InInternationalJointConferenceonArtificialIntelligence(IJCAL07),pages16291634,2007.

9. DanMateescu.EnglishPhoneticsandPhonologicalTheory.2003.

10. DellaPietraP.BrownandR.Mercer.Themathematicsofstatisticalmachinetranslation:Parameterestimation.InComputationalLinguistics,page19(2):263U 311, 1990.19(2):263U 311,1990.

11. GanapathirajuMadhaviPrahalladLavanyaandPrahalladKishore.AsimpleapproachforbuildingtransliterationeditorsforIndianlanguages.ZhejiangUniversitySCIENCE2005,2005.