Enriched translation model using morphology in MT
Luong Minh ThangWING group meeting – 07 July, 2009
04/11/23 1
Overview
• Brief recap on SMT & morphological analysis• Motivation• Enriched translation model
– Twin phrase-table construction– Merging phrase tables
• Experiments• Conclusion
04/11/23 2
SMT overview – alignment
• Parallel data
04/11/23
These are , first and foremost , messages of concern at the economic and social problems that we are experiencing , in spite of a period of sustained growth stemming from years of efforts by all our fellow citizens .
Ensinnäkin kohtaamiemme taloudellisten ja sosiaalisten vaikeuksien vuoksi on havaittavissa huolestumista , vaikka kasvu on kestävällä pohjalla ja tulosta vuosien ponnisteluista , kaikkien kansalaistemme taholta .
• Alignment: one-to-many (1-M)
Maria no daba una botefada a la bruja verde
NULL Mary did not slap the green witchSource
Target3
SMT overview – translation model
• Intersect alignment 1-M + M-1 M – M• Extracting phrases from M-M alignment translation model (phrase table).
04/11/23
problems ||| ongelmat ||| 0.372611 0.597858 0.114146 0.13882 2.718 problems ||| ongelmasta ||| 0.352941 0.423077 0.000836237 0.0012435 2.718…problems ||| vaikeuksista ||| 0.0696946 0.105991 0.0124042 0.0130002 2.718problems ||| vaikeuksien ||| 0.0410959 0.062069 0.000836237 0.0010174 2.718
Phrase penalty
Translation probabilitiesEnglish eForeign f
4
Lexical probabilities
Recap - Morphological analysis
• Morpheme: minimal meaning-bearing unitEnglish: machine + s, present + ed, etc.Finnish: oppositio + kansa + n + edusta + ja
= opposition of parliament member
• Morfessor (Creutz & Lagus, 2007): segment words, unsupervised manner un/PRE + fortunate/STM + ly/SUF
04/11/23 5
Motivation• Problem:
– Multiple word forms in morphology-complex language, e.g. ongelmat, ongelmasta, etc.
– Rare words often occur and are hard to align incorrect entries in normal (word-align) phrase table.
• Solution:– Construct morpheme-align phrase table (PT) to
aggregate better statistics for rare words.– Combine word- and morpheme-align PTs to produce
even better translation model in a proper way.04/11/23 6
Overview
• Brief recap on SMT & morphological analysis• Motivation• Enriched translation model
– Twin phrase-table construction– Merging phrase tables
• Experiments• Conclusion
04/11/23 7
Twin phrase-table (PT) construction
04/11/23
GIZA++
Decoding
Word alignment
Morpheme alignment
Word Morpheme
PTm
PTwm
Phrase Extraction
PTw
Morphological segmentation
Phrase Extraction
GIZA++
PT merging
problem/STM+ s/SUF ||| ongelma/STM+ t/SUF
problem/STM+ s/SUF ||| vaikeu/STM+ ksi/SUF+ sta/SUF
problems ||| vaikeuksista
8
Existing PT-merging methods
• Add-feature - (Nakov, 2008; Chen et. al. 2009):
F1 = F2 = F3 =
heuristic-driven
• Interpolation - (Wu & Wang, 2007) :– tran(f|e) = α * tran1(f|e) + (1- α) * tran2(f|e)
– lex(f|e) = β * lex1(f|e) + (1- β) * lex2(f|e)
not consider score “meaning”04/11/23
1 if from 1st PT
0.5 otherwise
1 if from 2nd PT
0.5 otherwise
1 if from both PTs0.5 otherwise
9
Our merging method – normalizing translation probabilities
tran1(e|f) =count1(e, f) / ∑e count1(e, f) tran2(e|f) =count2(e, f) / ∑e count2(e, f)
04/11/23 10
problem + s ||| vaikeu + ksi + staproblem + s ||| ongelma + sta
problem + s ||| ongelma + tproblem + s ||| ongelma + tproblem + s ||| ongelma + tproblem + s ||| vaikeu + ksi + sta
PTwm PTm
MLE
Our merging method – normalizing translation probabilities
tran(vaikeuksista | problems) =1/2=0.5tran(ongelmasta | problems) =1/2=0.5
tran(ongelmat | problems) = 3/4 = 0.75tran(vaikeuksista | problems) = 1/4 = 0.25
Undesired translation!
tran(vaikeuksista | problems) = (0.5 + 0.25)/2 = 0.375 tran(ongelmat | problems) = (0 + 0.75)/2 = 0.375tran(ongelmasta | problems) = (0.5 + 0)/2 = 0.25
Interpolation(ratio = 0.5)
04/11/23 11
problem + s ||| vaikeu + ksi + staproblem + s ||| ongelma + sta
problem + s ||| ongelma + tproblem + s ||| ongelma + tproblem + s ||| ongelma + tproblem + s ||| vaikeu + ksi + sta
PTwm PTm
MLE
Our merging method – normalizing translation probabilities
tran1(e|f) =count1(e, f) / ∑e count1(e, f) tran2(e|f) =count2(e, f) / ∑e count2(e, f)
04/11/23 12
Normalization
tran(e|f) =[ count1(e, f) + count2(e, f)] / [ ∑e count1(e, f) + ∑e count2(e, f) ]
problem + s ||| vaikeu + ksi + staproblem + s ||| ongelma + sta
problem + s ||| ongelma + tproblem + s ||| ongelma + tproblem + s ||| ongelma + tproblem + s ||| vaikeu + ksi + sta
PTwm PTm
MLE
Our merging method – normalizing translation probabilities
tran(vaikeuksista | problems) =1/2=0.5tran(ongelmasta | problems) =1/2=0.5
tran(ongelmat | problems) = 3/4 = 0.75tran(vaikeuksista | problems) = 1/4 = 0.25
tran(vaikeuksista | problems) = (1 + 1)/(2+4) = 0.33 tran(ongelmat | problems) = (0 + 3)/(2 + 4) = 0.5tran(ongelmasta | problems) = (1 + 0)/(2 + 4) = 0.17
Desired translation!
Normalization
04/11/23 13
problem + s ||| vaikeu + ksi + staproblem + s ||| ongelma + sta
problem + s ||| ongelma + tproblem + s ||| ongelma + tproblem + s ||| ongelma + tproblem + s ||| vaikeu + ksi + sta
PTwm PTm
MLE
Our merging method – full lexical probability interpolation
lex(vaikeuksista | problems) = w1
lex(ongelmasta | problems) = w2
lex(vaikeu + ksi + sta | problem + s) = m1
lex(ongelma + t | problem + s) = m3
lex(vaikeuksista | problems) = (w1 + m1)/2 lex(ongelmat | problems) = (w2 + 0)/2 lex(ongelmasta | problems) = (0 + m3) /2
NormalInterpolation(ratio = 0.5) Missing
interpolated probabilities !
PTm lexical model
P(vaikeuksista|problems)P(ongelmasta|problems)
P(vaikeu|problem), P(ongelma|problem), P(t|s), P(ksi|s),P(sta|s)
04/11/23 14
PTw lexical model
• Estimate lex(ongelma + sta | problem + s) using PTm lexical model m2
• Estimate lex(ongelmat | problems) using PTw lexical model w3
FullInterpolation
Overview
• Brief recap on SMT & morphological analysis• Motivation• Enriched translation model
– Twin phrase-table construction– Merging phrase tables
• Experiments• Conclusion
04/11/23 15
Experiments – dataset
• 2005 ACL shared task (Koehn & Monz, 2005)
04/11/23 16
Experiments – baselines
• w-system: uses PTw translate at word-level
• m-system: uses PTm translate at morpheme-level
• m-BLEU: BLEU where each token unit is a morpheme04/11/23 17
Experiments – our system
• Improvements over m-system and w-system are statistically significant using sign test by (Collins et al. 2005)
04/11/23 18
Conclusion
Our contributions:• Enrich the translation model without using
additional data.• Propose a principal way to merge phrase
tables generated at different granularities.
04/11/23 19
Q & A
•Thank you !!!
04/11/23 20