Invited Lecture
CS 4705: Introduction to Natural Language Processing
Fall 2004
Machine Translation:
Challenges and Approaches
Nizar Habash
Post-doctoral Fellow
Center for Computational Learning Systems
Columbia University
Sounds like Faulkner?
It lay on the table a candle
burning at each corner upon
the envelope tied in a soiled
pink garter two artificial
flowers. Not hit a man in
glasses.
 Faulkner
William
Faulkner,
"The sound and
 Machine
Translation
the fury“
It was once a shade, which
was in all beautiful weather
under a tree and varied like
the branches in the wind.
Es war einmal ein Schatten, der lag
bei jedem schönen Wetter unter
einem Baum und schwankte wie die
 Faulkner
Zweige
im Wind.
Helmut
Wördemann,
"Der
 Machine
Translation
unzufriedene Schatten“
(Translated by Systran)
http://www.ee.ucla.edu/~simkin/sounds_like_faulkner.html
Progress in MT
Statistical MT example
2002
2003
Human Translation
Egyptair Has Tomorrow to Egypt Air May Resume its
Resume Its Flights to
Flights to Libya Tomorrow
Libya
Cairo, April 6 (AFP) - An
Cairo
4-6
(AFP)
said
an
Egypt Air official
Cairo 6-4 ( AFP ) - an official at the Egyptian
official announced today Aviation Company today
announced, on Tuesday,
in the Egyptian lines
that the company egyptair that Egypt Air will resume
company for flying
may resume as of
its flights to Libya as of
Tuesday is a company " tomorrow, Wednesday its tomorrow, Wednesday,
insistent for flying " may flights to Libya after the
after the UN Security
resumed a consideration International Security
Council had announced the
Council resolution to the
of a day Wednesday
suspension of the
suspension of the embargo
tomorrow her trips to
embargo imposed on
imposed on Libya.
Libya of Security Council Libya.
decision trace
international the imposed
ban comment .
insistent Wednesday
may recurred her trips to
Libya tomorrow for flying
Form a talk by Charles Wayne, DARPA
Road Map
•
•
•
•
Why Machine Translation (MT)?
Multilingual Challenges for MT
MT Approaches
MT Evaluation
Why (Machine) Translation?
Languages in the world
• 6,800 living languages
• 600 with written tradition
• 95% of world population
speaks 100 languages
Translation Market
• $8 Billion Global Market
• Doubling every five years
(Donald Barabé, invited talk, MT Summit 2003)
Why Machine Translation?
• Full Translation
– Domain specific
• Weather reports
• Machine-aided Translation
– Translation dictionaries
– Translation memories
– Requires post-editing
• Cross-lingual NLP applications
– Cross-language IR
– Cross-language Summarization
Road Map
• Why Machine Translation (MT)?
• Multilingual Challenges for MT
–
–
–
–
Orthographic variations
Lexical ambiguity
Morphological variations
Translation divergences
• MT Paradigms
• MT Evaluation
Multilingual Challenges
• Orthographic Variations
– Ambiguous spelling
•
‫األوال ُد اش َعارا كتب االوالد اشعارا‬
َ َ‫َكت‬
ْ ‫ب‬
– Ambiguous word boundaries
•
• Lexical Ambiguity
– Bank  ‫( بنك‬financial) vs. ‫( ضفة‬river)
– Eat  essen (human) vs. fressen (animal)
Multilingual Challenges
Morphological Variations
• Affixation vs. Root+Pattern
write
kill
 written
 killed
‫كتب‬
‫قتل‬


‫مكتوب‬
‫مقتول‬
do
 done
‫فعل‬

‫مفعول‬
• Tokenization
conj
noun
article
plural
And the cars 
‫ والسيارات‬
Et les voitures 
and the cars
w Al SyArAt
et le voitures
Multilingual Challenges
Translation Divergences
• How languages map semantics to syntax
• 35% of sentences in TREC El Norte Corpus (Dorr et al 2002)
• Divergence Types
– Categorial (X tener hambre  X be hungry)
[98%]
– Conflational (X dar puñaladas a Z  X stab Z) [83%]
– Structural (X entrar en Y  X enter Y)
[35%]
– Head Swapping
(X cruzar Y nadando  X swim across Y)
– Thematic (X gustar a Y  Y like X)
[8%]
[6%]
Translation Divergences
conflation
‫ليس‬
‫ا نا‬
be
‫هنا‬
‫لست هنا‬
I-am-not here
I
not
etre
here
I am not here
Je
ne pas
ici
Je ne suis pas ici
I not be not here
Translation Divergences
categorial, thematic and structural
*
‫ا نا‬
be
‫بردان‬
I
*
tener
cold
Yo
frio
‫קר‬
‫ל‬
‫אני‬
‫انا بردان‬
I cold
I am cold
tengo frio
I-have cold
‫קר לי‬
cold for-me
Translation Divergences
head swap and categorial
‫اسرع‬
swim
I
across quickly
river
‫انا‬
‫عبور‬
‫سباحة‬
‫نهر‬
I swam across the river quickly ‫اسرعت عبور النهر سباحة‬
I-sped crossing the-river swimming
Translation Divergences
head swap and categorial
swim
I
across quickly
river
‫חצה‬
‫אני‬
‫את‬
‫ב‬
‫ב‬
‫נהר‬
‫שחיה‬
‫מהירות‬
I swam across the river quickly ‫חציתי את הנהר בשחיה במהירות‬
I-crossed obj river in-swim speedily
Translation Divergences
head swap and categorial
‫انا‬
‫اسرع‬
ver
b
‫عبور‬
‫سباحة‬
ver
b
‫אני‬
‫نهر‬
ver
b
I
swim
across quickly
river
‫חצה‬
‫את‬
‫ב‬
‫ב‬
‫נהר‬
‫שחיה‬
‫מהירות‬
Translation Divergences
Orthography+Morphology+Syntax
mom’s car
car
possessed-by
mom
妈妈的车
mama
de che
‫سيارة ماما‬
sayyArat mama
la voiture de maman
Road Map
• Why Machine Translation (MT)?
• Multilingual Challenges for MT
• MT Approaches
– Gisting / Transfer / Interlingua
– Statistical / Symbolic / Hybrid
– Practical Considerations
• MT Evaluation
MT Approaches
MT Pyramid
Source meaning
Target meaning
Source syntax
Source word
Analysis
Target syntax
Gisting
Target word
Generation
MT Approaches
Gisting Example
Sobre la base de dichas experiencias se estableció en 1988 una metodología.
Envelope her basis out speak experiences them settle at 1988 one methodology.
On the basis of these experiences, a methodology was arrived at in 1988.
MT Approaches
MT Pyramid
Source meaning
Source syntax
Source word
Analysis
Target meaning
Transfer
Gisting
Target syntax
Target word
Generation
MT Approaches
Transfer Example
• Transfer Lexicon
– Map SL structure to TL structure
poner
:subj
butter
:mod
:obj
X
mantequilla

en
:subj
X
:obj
Y
:obj
Y
X puso mantequilla en Y
X buttered Y
MT Approaches
MT Pyramid
Source meaning
Source syntax
Source word
Analysis
Interlingua
Transfer
Gisting
Target meaning
Target syntax
Target word
Generation
MT Approaches
Interlingua Example: Lexical Conceptual Structure
(Dorr, 1993)
MT Approaches
MT Pyramid
Source meaning
Source syntax
Source word
Analysis
Interlingua
Transfer
Gisting
Target meaning
Target syntax
Target word
Generation
MT Approaches
MT Pyramid
Source meaning Interlingual Lexicons
Source syntax
Source word
Analysis
Transfer Lexicons
Target meaning
Target syntax
Dictionaries/Parallel Corpora
Target word
Generation
MT Approaches
Statistical vs. Symbolic
Source meaning
Source syntax
Source word
Analysis
Target meaning
Target syntax
Target word
Generation
MT Approaches
Noisy Channel Model
Portions from http://www.clsp.jhu.edu/ws03/preworkshop/lecture_yamada.pdf
MT Approaches
IBM Model (Word-based Model)
http://www.clsp.jhu.edu/ws03/preworkshop/lecture_yamada.pdf
MT Approaches
Statistical vs. Symbolic vs. Hybrid
Source meaning
Source syntax
Source word
Analysis
Target meaning
Target syntax
Target word
Generation
MT Approaches
Statistical vs. Symbolic vs. Hybrid
Source meaning
Source syntax
Source word
Analysis
Target meaning
Target syntax
Target word
Generation
MT Approaches
Hybrid Example: GHMT
• Generation-Heavy Hybrid Machine Transaltion
• Lexical transfer but NO structural transfer
Maria puso la mantequilla en el pan.
lay locate place put
render set stand
poner
:subj
:mod
:obj
Maria
mantequilla

en
:obj
:subj
:mod
:obj
Maria
butter bilberry
on in
into at
:obj
pan
bread loaf
MT Approaches
Hybrid Example: GHMT
• LCS-driven Expansion
• Conflation Example
PUTV
Agent
MARIA
Theme
BUTTERN
[CAUSE GO]
Goal
BREAD
BUTTERV
[CAUSE GO]
Agent
Goal
MARIA
BREAD
CategorialVariation
MT Approaches
Hybrid Example: GHMT
• Structural Overgeneration
put
Maria
butter
lay
on
Maria
bread
butter
Maria
bread
at
loaf
bread
Maria
butter
render
butter
…
Maria
butter
into
loaf
MT Approaches
Hybrid Example: GHMT
Target Statistical Resources
• Structural N-gram Model
– Long-distance
– Lexemes
• Surface N-gram Model
– Local
– Surface-forms
buy
John
car
a
red
John bought a red car
MT Approaches
Hybrid Example: GHMT
Linearization &Ranking
Maria
Maria
Maria
Maria
Maria
Maria
Maria
buttered the bread
butters the bread
breaded the butter
breads the butter
buttered the loaf
butters the loaf
put the butter on bread
-47.0841
-47.2994
-48.7334
-48.835
-51.3784
-51.5937
-54.128
MT Approaches
Practical Considerations
• Resources Availability
– Parsers and Generators
• Input/Output compatability
– Translation Lexicons
• Word-based vs. Transfer/Interlingua
– Parallel Corpora
• Domain of interest
• Bigger is better
• Time Availability
– Statistical training, resource building
MT Approaches
Resource Poverty
No Parser?
No Translation Dictionary?
Parallel Corpus
• Align with rich language
• Extract dictionary
•Parse rich side
•Infer parses
•Build a statistical parser
Road Map
•
•
•
•
Why Machine Translation (MT)?
Multilingual Challenges for MT
MT Approaches
MT Evaluation
MT Evaluation
• More art than science
• Wide range of Metrics/Techniques
– interface, …, scalability, …, faithfulness, ...
space/time complexity, … etc.
• Automatic vs. Human-based
– Dumb Machines vs. Slow Humans
MT Evaluation Metrics
(Church and Hovy 1993)
• System-based Metrics
Count internal resources: size of lexicon,
number of grammar rules, etc.
– easy to measure
– not comparable across systems
– not necessarily related to utility
MT Evaluation Metrics
• Text-based Metrics
– Sentence-based Metrics
• Quality: Accuracy, Fluency, Coherence, etc.
• 3-point scale to 100-point scale
– Comprehensibility Metrics
•
•
•
•
Comprehension, Informativeness,
x-point scales, questionnaires
most related to utility
hard to measure
MT Evaluation Metrics
• Text-based Metrics (cont’d)
– Amount of Post-Editing
• number of keystrokes per page
• not necessarily related to utility
• Cost-based Metrics
– Cost per page
– Time per page
Human-based Evaluation Example
Accuracy Criteria
5
4
3
2
1
co n ten ts o f o rig in al sen ten ce co n v ey e d (m ig h t n eed m in o r
co rrectio n s)
co n ten ts o f o rig in al sen ten ce co n v ey e d B U T e rro rs in w o rd o rd er
co n ten ts o f o rig in al sen ten ce g en erally co n v ey ed B U T e rro rs in
relatio n sh ip b etw een p h rases, ten se, sin g u lar/p lu ra l, etc.
co n ten ts o f o rig in al sen ten ce n o t ad eq u ately co n v ey ed , p o rtio n s
o f o rig in al sen ten ce in co rrectly tran slated , m issin g m o d ifiers
co n ten ts o f o rig in al sen ten ce n o t co n v ey ed , m issin g v erb s,
su b jects, o b jects, p h rases o r clau ses
Human-based Evaluation Example
Fluency Criteria
5
4
3
2
1
clear m ean in g , g o o d g ram m ar, term in o lo g y an d sen ten ce
stru ctu re
clear m ean in g B U T b ad g ra m m ar, b ad term in o lo g y o r b ad
sen ten ce stru ctu re
m ean in g g rasp ab le B U T am b ig u ities d u e to b ad g ram m ar, b ad
term in o lo g y o r b ad sen ten ce stru ctu re
m ean in g u n clear B U T in ferab le
m ean in g ab so lu tely u n clear
Fluency vs. Accuracy
FAHQ
MT
conMT
Prof.
MT
Info.
MT
Fluency
Accuracy
Automatic Evaluation Example
Bleu Metric
• Bleu
–
–
–
–
–
BiLingual Evaluation Understudy (Papineni et al 2001)
Modified n-gram precision with length penalty
Quick, inexpensive and language independent
Correlates highly with human evaluation
Bias against synonyms and inflectional variations
Automatic Evaluation Example
Bleu Metric
Test Sentence
Gold Standard References
colorless green ideas sleep furiously
all dull jade ideas sleep irately
drab emerald concepts sleep furiously
colorless immature thoughts nap angrily
Automatic Evaluation Example
Bleu Metric
Test Sentence
Gold Standard References
colorless green ideas sleep furiously
all dull jade ideas sleep irately
drab emerald concepts sleep furiously
colorless immature thoughts nap angrily
Unigram precision = 4/5
Automatic Evaluation Example
Bleu Metric
Test Sentence
Gold Standard References
colorless green ideas sleep furiously
colorless green ideas sleep furiously
colorless green ideas sleep furiously
colorless green ideas sleep furiously
all dull jade ideas sleep irately
drab emerald concepts sleep furiously
colorless immature thoughts nap angrily
Unigram precision = 4 / 5 = 0.8
Bigram precision = 2 / 4 = 0.5
Bleu Score = (a1 a2 …an)1/n
= (0.8 ╳ 0.5)½ = 0.6325  63.25
Descargar

Semitic Linguistic Phenomena