Machine Translation Zoo - Univerzita Karlovaufal.mff.cuni.cz/~popel/papers/2013_05_06_zoo.pdf ·...
Transcript of Machine Translation Zoo - Univerzita Karlovaufal.mff.cuni.cz/~popel/papers/2013_05_06_zoo.pdf ·...
![Page 1: Machine Translation Zoo - Univerzita Karlovaufal.mff.cuni.cz/~popel/papers/2013_05_06_zoo.pdf · Machine Translation Zoo Tree-to-tree transfer and Discriminative learning Martin Popel](https://reader030.fdocuments.us/reader030/viewer/2022011900/5f01d7387e708231d4014c16/html5/thumbnails/1.jpg)
Machine Translation ZooTree-to-tree transfer and Discriminative learning
Martin Popel
UFAL (Institute of Formal and Applied Linguistics)Charles University in Prague
May 5th 2013, Seminar of Formal Linguistics, Prague
![Page 2: Machine Translation Zoo - Univerzita Karlovaufal.mff.cuni.cz/~popel/papers/2013_05_06_zoo.pdf · Machine Translation Zoo Tree-to-tree transfer and Discriminative learning Martin Popel](https://reader030.fdocuments.us/reader030/viewer/2022011900/5f01d7387e708231d4014c16/html5/thumbnails/2.jpg)
Today’s Menu
1 MT IntroTaxonomyHybrids
2 Online LearningPerceptronStructured Prediction
3 Guided Learning
4 Back to MTEasy-First Decoding in MTGuided Learning in MT
![Page 3: Machine Translation Zoo - Univerzita Karlovaufal.mff.cuni.cz/~popel/papers/2013_05_06_zoo.pdf · Machine Translation Zoo Tree-to-tree transfer and Discriminative learning Martin Popel](https://reader030.fdocuments.us/reader030/viewer/2022011900/5f01d7387e708231d4014c16/html5/thumbnails/3.jpg)
Today’s Menu
1 MT IntroTaxonomyHybrids
2 Online LearningPerceptronStructured Prediction
3 Guided Learning
4 Back to MTEasy-First Decoding in MTGuided Learning in MT
![Page 4: Machine Translation Zoo - Univerzita Karlovaufal.mff.cuni.cz/~popel/papers/2013_05_06_zoo.pdf · Machine Translation Zoo Tree-to-tree transfer and Discriminative learning Martin Popel](https://reader030.fdocuments.us/reader030/viewer/2022011900/5f01d7387e708231d4014c16/html5/thumbnails/4.jpg)
MT Intro Online Learning Guided Learning Back to MT
Phrase-based MT (Moses)
Training
word-alignment(Giza++ & symmetrization)
phrase extraction
tune parameters (MERT)
Decoding
get all matching rules
find one derivationwith a maximum score (beam search)
![Page 5: Machine Translation Zoo - Univerzita Karlovaufal.mff.cuni.cz/~popel/papers/2013_05_06_zoo.pdf · Machine Translation Zoo Tree-to-tree transfer and Discriminative learning Martin Popel](https://reader030.fdocuments.us/reader030/viewer/2022011900/5f01d7387e708231d4014c16/html5/thumbnails/5.jpg)
MT Intro Online Learning Guided Learning Back to MT
TectoMT
Training
analyze CzEng to t-layer
t-node alignment
learn one MaxEnt model for each source lemma and formeme
Decoding
get all translation variants for each lemma and formeme
find a labeling with a maximum score (HMTM)
![Page 6: Machine Translation Zoo - Univerzita Karlovaufal.mff.cuni.cz/~popel/papers/2013_05_06_zoo.pdf · Machine Translation Zoo Tree-to-tree transfer and Discriminative learning Martin Popel](https://reader030.fdocuments.us/reader030/viewer/2022011900/5f01d7387e708231d4014c16/html5/thumbnails/6.jpg)
MT Intro Online Learning Guided Learning Back to MT
TectoMT – MaxEnt Model
ANALYSIS
TRANSFER
SYNTHESIS
He agreed with the unions to cut all overtime. Dohodl se s odbory na zrušení všech přesčasů.
agree / v:fintense=past, voice=activenegation=0, sempos=v
he / n:subj union /n:with+X
overtime / n:obj
cut / v:inf,has_left_child=0, sempos=v,has_right_child=1, tag=VB,position=right, named_entity=0
all / adj:attr
chop, saw, trim, shorten, lumber, hew,
lower, delete, cropabolish, cancel,...
![Page 7: Machine Translation Zoo - Univerzita Karlovaufal.mff.cuni.cz/~popel/papers/2013_05_06_zoo.pdf · Machine Translation Zoo Tree-to-tree transfer and Discriminative learning Martin Popel](https://reader030.fdocuments.us/reader030/viewer/2022011900/5f01d7387e708231d4014c16/html5/thumbnails/7.jpg)
MT Intro Online Learning Guided Learning Back to MT
Machine Translation Taxonomy
Level of transfer:
surface, shallow syntax, tectogrammatical
Base translation unit (BTU):
word, phrase, phrase with gaps, treelet, node
Extract more segmentations in training?
Try (search) more segmentations in decoding?
Use more segmentations in the output translation?
What is the context X in P(BTUtarget |BTUsource ,X )?
Considering just Translation Model:
(Brown et al., 1993)word-based
![Page 8: Machine Translation Zoo - Univerzita Karlovaufal.mff.cuni.cz/~popel/papers/2013_05_06_zoo.pdf · Machine Translation Zoo Tree-to-tree transfer and Discriminative learning Martin Popel](https://reader030.fdocuments.us/reader030/viewer/2022011900/5f01d7387e708231d4014c16/html5/thumbnails/8.jpg)
MT Intro Online Learning Guided Learning Back to MT
Machine Translation Taxonomy
Level of transfer: surface
, shallow syntax, tectogrammatical
Base translation unit (BTU):word
, phrase, phrase with gaps, treelet, node
Extract more segmentations in training? no
Try (search) more segmentations in decoding? no
Use more segmentations in the output translation? no
What is the context X in P(BTUtarget |BTUsource ,X )?Considering just Translation Model: nothing
(Brown et al., 1993)word-based
![Page 9: Machine Translation Zoo - Univerzita Karlovaufal.mff.cuni.cz/~popel/papers/2013_05_06_zoo.pdf · Machine Translation Zoo Tree-to-tree transfer and Discriminative learning Martin Popel](https://reader030.fdocuments.us/reader030/viewer/2022011900/5f01d7387e708231d4014c16/html5/thumbnails/9.jpg)
MT Intro Online Learning Guided Learning Back to MT
Machine Translation Taxonomy
Level of transfer: surface
, shallow syntax, tectogrammatical
Base translation unit (BTU):word, phrase
, phrase with gaps, treelet, node
Extract more segmentations in training? yes
Try (search) more segmentations in decoding? yes
Use more segmentations in the output translation? no
What is the context X in P(BTUtarget |BTUsource ,X )?Considering just Translation Model: nothing
(Brown et al., 1993)word-based
(Koehn et al., 2003)phrase-based
![Page 10: Machine Translation Zoo - Univerzita Karlovaufal.mff.cuni.cz/~popel/papers/2013_05_06_zoo.pdf · Machine Translation Zoo Tree-to-tree transfer and Discriminative learning Martin Popel](https://reader030.fdocuments.us/reader030/viewer/2022011900/5f01d7387e708231d4014c16/html5/thumbnails/10.jpg)
MT Intro Online Learning Guided Learning Back to MT
Machine Translation Taxonomy
Level of transfer: surface
, shallow syntax, tectogrammatical
Base translation unit (BTU):word, phrase, phrase with gaps
, treelet, node
Extract more segmentations in training? yes
Try (search) more segmentations in decoding? yes
Use more segmentations in the output translation? no
What is the context X in P(BTUtarget |BTUsource ,X )?Considering just Translation Model: nothing
(Brown et al., 1993)word-based
(Koehn et al., 2003)phrase-based
(Chiang, 2005)hierarchical
![Page 11: Machine Translation Zoo - Univerzita Karlovaufal.mff.cuni.cz/~popel/papers/2013_05_06_zoo.pdf · Machine Translation Zoo Tree-to-tree transfer and Discriminative learning Martin Popel](https://reader030.fdocuments.us/reader030/viewer/2022011900/5f01d7387e708231d4014c16/html5/thumbnails/11.jpg)
MT Intro Online Learning Guided Learning Back to MT
Machine Translation Taxonomy
Level of transfer: surface, shallow syntax
, tectogrammatical
Base translation unit (BTU):word, phrase, phrase with gaps, treelet
, node
Extract more segmentations in training? no
Try (search) more segmentations in decoding? no
Use more segmentations in the output translation? no
What is the context X in P(BTUtarget |BTUsource ,X )?Considering just Translation Model: neighboring treelets
(Brown et al., 1993)word-based
(Koehn et al., 2003)phrase-based
(Chiang, 2005)hierarchical
(Quirk and Menezes, 2006)dep. treelet to string
![Page 12: Machine Translation Zoo - Univerzita Karlovaufal.mff.cuni.cz/~popel/papers/2013_05_06_zoo.pdf · Machine Translation Zoo Tree-to-tree transfer and Discriminative learning Martin Popel](https://reader030.fdocuments.us/reader030/viewer/2022011900/5f01d7387e708231d4014c16/html5/thumbnails/12.jpg)
MT Intro Online Learning Guided Learning Back to MT
Machine Translation Taxonomy
Level of transfer: surface, shallow syntax, tectogrammatical
Base translation unit (BTU):word, phrase, phrase with gaps, treelet, node
Extract more segmentations in training? no
Try (search) more segmentations in decoding? no
Use more segmentations in the output translation? no
What is the context X in P(BTUtarget |BTUsource ,X )?Considering just Translation Model: neighboring nodes
(Brown et al., 1993)word-based
(Koehn et al., 2003)phrase-based
(Chiang, 2005)hierarchical
(Quirk and Menezes, 2006)dep. treelet to string
(Marecek et al., 2010)TectoMT
![Page 13: Machine Translation Zoo - Univerzita Karlovaufal.mff.cuni.cz/~popel/papers/2013_05_06_zoo.pdf · Machine Translation Zoo Tree-to-tree transfer and Discriminative learning Martin Popel](https://reader030.fdocuments.us/reader030/viewer/2022011900/5f01d7387e708231d4014c16/html5/thumbnails/13.jpg)
MT Intro Online Learning Guided Learning Back to MT
Machine Translation Taxonomy
Level of transfer: surface, shallow syntax, tectogrammatical
Base translation unit (BTU):word, phrase, phrase with gaps, treelet, node
Extract more segmentations in training? yes
Try (search) more segmentations in decoding? yes
Use more segmentations in the output translation? yes
What is the context X in P(BTUtarget |BTUsource ,X )?Considering just Translation Model: nothing
(Brown et al., 1993)word-based
(Koehn et al., 2003)phrase-based
(Chiang, 2005)hierarchical
(Quirk and Menezes, 2006)dep. treelet to string
(Marecek et al., 2010)TectoMT
(Arun, 2011)Monte Carlo
![Page 14: Machine Translation Zoo - Univerzita Karlovaufal.mff.cuni.cz/~popel/papers/2013_05_06_zoo.pdf · Machine Translation Zoo Tree-to-tree transfer and Discriminative learning Martin Popel](https://reader030.fdocuments.us/reader030/viewer/2022011900/5f01d7387e708231d4014c16/html5/thumbnails/14.jpg)
MT Intro Online Learning Guided Learning Back to MT
Hybrids: TectoMoses
Linearize source t-trees (two factors: lemma and formeme), translatewith Moses, project dependencies and use TectoMT synthesis.
& blocks
source language (English) target language (Czech)
morphological layer
analytical layer
tectogramatical layer
a-layer
m-layer
w-layer
ANALYSIS TRANSFER SYNTHESIS
t-layer
tokenizationlemmatizationtagger (Morce)
parser (McDonald's MST)analytical functions
mark edges to contract
build t-tree
fill formems grammatemes useHMTM
querydictionary
fill morphological categories
impose agreement
add functional words
generatewordforms
concatenate
rule based statistical
segmentation
![Page 15: Machine Translation Zoo - Univerzita Karlovaufal.mff.cuni.cz/~popel/papers/2013_05_06_zoo.pdf · Machine Translation Zoo Tree-to-tree transfer and Discriminative learning Martin Popel](https://reader030.fdocuments.us/reader030/viewer/2022011900/5f01d7387e708231d4014c16/html5/thumbnails/15.jpg)
MT Intro Online Learning Guided Learning Back to MT
Hybrids: TectoMoses
Linearize source t-trees (two factors: lemma and formeme), translatewith Moses, project dependencies and use TectoMT synthesis.
& blocks
source language (English) target language (Czech)
morphological layer
analytical layer
tectogramatical layer
a-layer
m-layer
w-layer
ANALYSIS TRANSFER SYNTHESIS
t-layer
tokenizationlemmatizationtagger (Morce)
parser (McDonald's MST)analytical functions
mark edges to contract
build t-tree
fill formems grammatemes useHMTM
querydictionary
fill morphological categories
impose agreement
add functional words
generatewordforms
concatenate
rule based statistical
segmentation
![Page 16: Machine Translation Zoo - Univerzita Karlovaufal.mff.cuni.cz/~popel/papers/2013_05_06_zoo.pdf · Machine Translation Zoo Tree-to-tree transfer and Discriminative learning Martin Popel](https://reader030.fdocuments.us/reader030/viewer/2022011900/5f01d7387e708231d4014c16/html5/thumbnails/16.jpg)
MT Intro Online Learning Guided Learning Back to MT
Hybrids: TectoMoses
Linearize source t-trees (two factors: lemma and formeme), translatewith Moses, project dependencies and use TectoMT synthesis.
& blocks
source language (English) target language (Czech)
morphological layer
analytical layer
tectogramatical layer
a-layer
m-layer
w-layer
ANALYSIS TRANSFER SYNTHESIS
t-layer
tokenizationlemmatizationtagger (Morce)
parser (McDonald's MST)analytical functions
mark edges to contract
build t-tree
fill formems grammatemes useHMTM
querydictionary
fill morphological categories
impose agreement
add functional words
generatewordforms
concatenate
rule based statistical
segmentation
![Page 17: Machine Translation Zoo - Univerzita Karlovaufal.mff.cuni.cz/~popel/papers/2013_05_06_zoo.pdf · Machine Translation Zoo Tree-to-tree transfer and Discriminative learning Martin Popel](https://reader030.fdocuments.us/reader030/viewer/2022011900/5f01d7387e708231d4014c16/html5/thumbnails/17.jpg)
MT Intro Online Learning Guided Learning Back to MT
Hybrids: PhraseFix
Done for WMT 2013 by Petra Galuscakova:
Post-edit TectoMT output using Moses
trained on cs-tectomt → cs-reference (whole CzEng).
How to post-edit only when confident?
filter phrase tableadd “confidence” feature for MERTimprove alignment (monolingual)boost phrase table (e.g. with identities)
Future work:
use also source (English) sentences ⇒ multi-source translation
project only content words (using TectoMT)
factored translation with non-synchronous(overlapping) factors
![Page 18: Machine Translation Zoo - Univerzita Karlovaufal.mff.cuni.cz/~popel/papers/2013_05_06_zoo.pdf · Machine Translation Zoo Tree-to-tree transfer and Discriminative learning Martin Popel](https://reader030.fdocuments.us/reader030/viewer/2022011900/5f01d7387e708231d4014c16/html5/thumbnails/18.jpg)
MT Intro Online Learning Guided Learning Back to MT
Hybrids: PhraseFix
Done for WMT 2013 by Petra Galuscakova:
Post-edit TectoMT output using Moses
trained on cs-tectomt → cs-reference (whole CzEng).
How to post-edit only when confident?
filter phrase tableadd “confidence” feature for MERTimprove alignment (monolingual)boost phrase table (e.g. with identities)
Future work:
use also source (English) sentences ⇒ multi-source translation
project only content words (using TectoMT)
factored translation with non-synchronous(overlapping) factors
![Page 19: Machine Translation Zoo - Univerzita Karlovaufal.mff.cuni.cz/~popel/papers/2013_05_06_zoo.pdf · Machine Translation Zoo Tree-to-tree transfer and Discriminative learning Martin Popel](https://reader030.fdocuments.us/reader030/viewer/2022011900/5f01d7387e708231d4014c16/html5/thumbnails/19.jpg)
MT Intro Online Learning Guided Learning Back to MT
Even More Hybrids: DepFix, AddToTrain, Chimera
DepFix (Rosa et al., 2012)
post-edit SMT using syntactic analysis and rules
exploit also the source sentences, robust parsing
AddToTrain (Bojar, Galuscakova)
translate monolingual news (or WMT devsets) with TectoMT
add this to Moses parallel training data
Chimera
post-edit AddToTrain output with DepFix
sent to WMT 2013 in attempt to beat Google
![Page 20: Machine Translation Zoo - Univerzita Karlovaufal.mff.cuni.cz/~popel/papers/2013_05_06_zoo.pdf · Machine Translation Zoo Tree-to-tree transfer and Discriminative learning Martin Popel](https://reader030.fdocuments.us/reader030/viewer/2022011900/5f01d7387e708231d4014c16/html5/thumbnails/20.jpg)
MT Intro Online Learning Guided Learning Back to MT
Even More Hybrids: DepFix, AddToTrain, Chimera
DepFix (Rosa et al., 2012)
post-edit SMT using syntactic analysis and rules
exploit also the source sentences, robust parsing
AddToTrain (Bojar, Galuscakova)
translate monolingual news (or WMT devsets) with TectoMT
add this to Moses parallel training data
Chimera
post-edit AddToTrain output with DepFix
sent to WMT 2013 in attempt to beat Google
![Page 21: Machine Translation Zoo - Univerzita Karlovaufal.mff.cuni.cz/~popel/papers/2013_05_06_zoo.pdf · Machine Translation Zoo Tree-to-tree transfer and Discriminative learning Martin Popel](https://reader030.fdocuments.us/reader030/viewer/2022011900/5f01d7387e708231d4014c16/html5/thumbnails/21.jpg)
Today’s Menu
1 MT IntroTaxonomyHybrids
2 Online LearningPerceptronStructured Prediction
3 Guided Learning
4 Back to MTEasy-First Decoding in MTGuided Learning in MT
![Page 22: Machine Translation Zoo - Univerzita Karlovaufal.mff.cuni.cz/~popel/papers/2013_05_06_zoo.pdf · Machine Translation Zoo Tree-to-tree transfer and Discriminative learning Martin Popel](https://reader030.fdocuments.us/reader030/viewer/2022011900/5f01d7387e708231d4014c16/html5/thumbnails/22.jpg)
MT Intro Online Learning Guided Learning Back to MT
General Algorithm for Online Learning
w := 0while (x, ygold) := get new data()
ypred := prediction(w, x)w += update(x, ygold , ypred)
Output: w
![Page 23: Machine Translation Zoo - Univerzita Karlovaufal.mff.cuni.cz/~popel/papers/2013_05_06_zoo.pdf · Machine Translation Zoo Tree-to-tree transfer and Discriminative learning Martin Popel](https://reader030.fdocuments.us/reader030/viewer/2022011900/5f01d7387e708231d4014c16/html5/thumbnails/23.jpg)
MT Intro Online Learning Guided Learning Back to MT
General Algorithm for Online Learning
w := 0while (x, ygold) := get new data()
ypred := prediction(w, x)w += update(x, ygold , ypred)
Output: w
initialize all weights to zero
for each instance (observation)1. get its features x2. do the prediction ypred3. get the correct label ygold4. update the weights
![Page 24: Machine Translation Zoo - Univerzita Karlovaufal.mff.cuni.cz/~popel/papers/2013_05_06_zoo.pdf · Machine Translation Zoo Tree-to-tree transfer and Discriminative learning Martin Popel](https://reader030.fdocuments.us/reader030/viewer/2022011900/5f01d7387e708231d4014c16/html5/thumbnails/24.jpg)
MT Intro Online Learning Guided Learning Back to MT
General Algorithm for Online Learning
w := 0while (x, ygold) := get new data()
ypred := prediction(w, x)w += update(x, ygold , ypred)
Output: w
initialize all weights to zerofor each instance (observation)
1. get its features x2. do the prediction ypred3. get the correct label ygold4. update the weights
![Page 25: Machine Translation Zoo - Univerzita Karlovaufal.mff.cuni.cz/~popel/papers/2013_05_06_zoo.pdf · Machine Translation Zoo Tree-to-tree transfer and Discriminative learning Martin Popel](https://reader030.fdocuments.us/reader030/viewer/2022011900/5f01d7387e708231d4014c16/html5/thumbnails/25.jpg)
MT Intro Online Learning Guided Learning Back to MT
General Algorithm for Online Learning
w := 0while (x, ygold) := get new data()
ypred := prediction(w, x)w += update(x, ygold , ypred)
Output: w
initialize all weights to zerofor each instance (observation)1. get its features x
2. do the prediction ypred3. get the correct label ygold4. update the weights
![Page 26: Machine Translation Zoo - Univerzita Karlovaufal.mff.cuni.cz/~popel/papers/2013_05_06_zoo.pdf · Machine Translation Zoo Tree-to-tree transfer and Discriminative learning Martin Popel](https://reader030.fdocuments.us/reader030/viewer/2022011900/5f01d7387e708231d4014c16/html5/thumbnails/26.jpg)
MT Intro Online Learning Guided Learning Back to MT
General Algorithm for Online Learning
w := 0while (x, ygold) := get new data()
ypred := prediction(w, x)w += update(x, ygold , ypred)
Output: w
initialize all weights to zerofor each instance (observation)1. get its features x2. do the prediction ypred
3. get the correct label ygold4. update the weights
![Page 27: Machine Translation Zoo - Univerzita Karlovaufal.mff.cuni.cz/~popel/papers/2013_05_06_zoo.pdf · Machine Translation Zoo Tree-to-tree transfer and Discriminative learning Martin Popel](https://reader030.fdocuments.us/reader030/viewer/2022011900/5f01d7387e708231d4014c16/html5/thumbnails/27.jpg)
MT Intro Online Learning Guided Learning Back to MT
General Algorithm for Online Learning
w := 0while (x, ygold) := get new data()
ypred := prediction(w, x)w += update(x, ygold , ypred)
Output: w
initialize all weights to zerofor each instance (observation)1. get its features x2. do the prediction ypred3. get the correct label ygold
4. update the weights
![Page 28: Machine Translation Zoo - Univerzita Karlovaufal.mff.cuni.cz/~popel/papers/2013_05_06_zoo.pdf · Machine Translation Zoo Tree-to-tree transfer and Discriminative learning Martin Popel](https://reader030.fdocuments.us/reader030/viewer/2022011900/5f01d7387e708231d4014c16/html5/thumbnails/28.jpg)
MT Intro Online Learning Guided Learning Back to MT
General Algorithm for Online Learning
w := 0while (x, ygold) := get new data()
ypred := prediction(w, x)w += update(x, ygold , ypred)
Output: w
initialize all weights to zerofor each instance (observation)1. get its features x2. do the prediction ypred3. get the correct label ygold4. update the weights
![Page 29: Machine Translation Zoo - Univerzita Karlovaufal.mff.cuni.cz/~popel/papers/2013_05_06_zoo.pdf · Machine Translation Zoo Tree-to-tree transfer and Discriminative learning Martin Popel](https://reader030.fdocuments.us/reader030/viewer/2022011900/5f01d7387e708231d4014c16/html5/thumbnails/29.jpg)
MT Intro Online Learning Guided Learning Back to MT
General Algorithm for Online Learning
w := 0while (x, ygold) := get new data()
ypred := prediction(w, x)w += update(x, ygold , ypred)
Output: w
initialize all weights to zerofor each instance (observation)1. get its features x2. do the prediction ypred3. get the correct label ygold4. update the weights
![Page 30: Machine Translation Zoo - Univerzita Karlovaufal.mff.cuni.cz/~popel/papers/2013_05_06_zoo.pdf · Machine Translation Zoo Tree-to-tree transfer and Discriminative learning Martin Popel](https://reader030.fdocuments.us/reader030/viewer/2022011900/5f01d7387e708231d4014c16/html5/thumbnails/30.jpg)
MT Intro Online Learning Guided Learning Back to MT
General Algorithm for Online Learning
w := 0while (x, ygold) := get new data()
ypred := prediction(w, x)w += update(x, ygold , ypred)
Output: w
initialize all weights to zerofor each instance (observation)1. get its features x2. do the prediction ypred3. get the correct label ygold4. update the weights
Definition: conservative online learning
no error ⇒ no updatei.e., if ypred = ygold then update(x, ygold , ypred) = 0
![Page 31: Machine Translation Zoo - Univerzita Karlovaufal.mff.cuni.cz/~popel/papers/2013_05_06_zoo.pdf · Machine Translation Zoo Tree-to-tree transfer and Discriminative learning Martin Popel](https://reader030.fdocuments.us/reader030/viewer/2022011900/5f01d7387e708231d4014c16/html5/thumbnails/31.jpg)
MT Intro Online Learning Guided Learning Back to MT
General Algorithm for Online Learning
w := 0while (x, ygold) := get new data()
ypred := prediction(w, x)w += update(x, ygold , ypred)//prediction(w, x) = ygold
Output: w
initialize all weights to zerofor each instance (observation)1. get its features x2. do the prediction ypred3. get the correct label ygold4. update the weights
Definition: conservative online learning
no error ⇒ no updatei.e., if ypred = ygold then update(x, ygold , ypred) = 0
Definition: aggressive online learning
after the update, the instance would be classified correctly
![Page 32: Machine Translation Zoo - Univerzita Karlovaufal.mff.cuni.cz/~popel/papers/2013_05_06_zoo.pdf · Machine Translation Zoo Tree-to-tree transfer and Discriminative learning Martin Popel](https://reader030.fdocuments.us/reader030/viewer/2022011900/5f01d7387e708231d4014c16/html5/thumbnails/32.jpg)
MT Intro Online Learning Guided Learning Back to MT
Perceptron
w := 0while (x, ygold) := get new data()
ypred := prediction(w, x)w += update(x, ygold , ypred)
Output: w
prediction(w, x)def=
update(x, ygold , ypred)def=
Binary Perceptron
[w · x > 0]
α(ygold − ypred) · x
Multi-class Perceptron
arg maxy w · f(x, y)
α(f(x, ygold)− f(x, ypred)
)
![Page 33: Machine Translation Zoo - Univerzita Karlovaufal.mff.cuni.cz/~popel/papers/2013_05_06_zoo.pdf · Machine Translation Zoo Tree-to-tree transfer and Discriminative learning Martin Popel](https://reader030.fdocuments.us/reader030/viewer/2022011900/5f01d7387e708231d4014c16/html5/thumbnails/33.jpg)
MT Intro Online Learning Guided Learning Back to MT
Perceptron
w := 0while (x, ygold) := get new data()
ypred := prediction(w, x)w += update(x, ygold , ypred)
Output: w
dot product (similarity score)of weights and featuresw · x =
∑i wixi
prediction(w, x)def=
update(x, ygold , ypred)def=
Binary Perceptron
[w · x > 0]
α(ygold − ypred) · x
Multi-class Perceptron
arg maxy w · f(x, y)
α(f(x, ygold)− f(x, ypred)
)
![Page 34: Machine Translation Zoo - Univerzita Karlovaufal.mff.cuni.cz/~popel/papers/2013_05_06_zoo.pdf · Machine Translation Zoo Tree-to-tree transfer and Discriminative learning Martin Popel](https://reader030.fdocuments.us/reader030/viewer/2022011900/5f01d7387e708231d4014c16/html5/thumbnails/34.jpg)
MT Intro Online Learning Guided Learning Back to MT
Perceptron
w := 0while (x, ygold) := get new data()
ypred := prediction(w, x)w += update(x, ygold , ypred)
Output: w
dot product (similarity score)of weights and featuresw · x =
∑i wixi
Iverson bracket
[P] =
{1 if P is true;
0 otherwise.
prediction(w, x)def=
update(x, ygold , ypred)def=
Binary Perceptron
[w · x > 0]
α(ygold − ypred) · x
Multi-class Perceptron
arg maxy w · f(x, y)
α(f(x, ygold)− f(x, ypred)
)
![Page 35: Machine Translation Zoo - Univerzita Karlovaufal.mff.cuni.cz/~popel/papers/2013_05_06_zoo.pdf · Machine Translation Zoo Tree-to-tree transfer and Discriminative learning Martin Popel](https://reader030.fdocuments.us/reader030/viewer/2022011900/5f01d7387e708231d4014c16/html5/thumbnails/35.jpg)
MT Intro Online Learning Guided Learning Back to MT
Perceptron
w := 0while (x, ygold) := get new data()
ypred := prediction(w, x)w += update(x, ygold , ypred)
Output: w
dot product (similarity score)of weights and featuresw · x =
∑i wixi
Iverson bracket
[P] =
{1 if P is true;
0 otherwise.
prediction(w, x)def=
update(x, ygold , ypred)def=
Binary Perceptron
[w · x > 0]
α(ygold − ypred) · x
Multi-class Perceptron
arg maxy w · f(x, y)
α(f(x, ygold)− f(x, ypred)
)
learning rate (step size) α > 0
![Page 36: Machine Translation Zoo - Univerzita Karlovaufal.mff.cuni.cz/~popel/papers/2013_05_06_zoo.pdf · Machine Translation Zoo Tree-to-tree transfer and Discriminative learning Martin Popel](https://reader030.fdocuments.us/reader030/viewer/2022011900/5f01d7387e708231d4014c16/html5/thumbnails/36.jpg)
MT Intro Online Learning Guided Learning Back to MT
Perceptron
w := 0while (x, ygold) := get new data()
ypred := prediction(w, x)w += update(x, ygold , ypred)
Output: w
prediction(w, x)def=
update(x, ygold , ypred)def=
Binary Perceptron
[w · x > 0]
α(ygold − ypred) · x
Multi-class Perceptron
arg maxy w · f(x, y)
α(f(x, ygold)− f(x, ypred)
)learning rate (step size) α > 0
![Page 37: Machine Translation Zoo - Univerzita Karlovaufal.mff.cuni.cz/~popel/papers/2013_05_06_zoo.pdf · Machine Translation Zoo Tree-to-tree transfer and Discriminative learning Martin Popel](https://reader030.fdocuments.us/reader030/viewer/2022011900/5f01d7387e708231d4014c16/html5/thumbnails/37.jpg)
MT Intro Online Learning Guided Learning Back to MT
Perceptron
w := 0while (x, ygold) := get new data()
ypred := prediction(w, x)w += update(x, ygold , ypred)
Output: w
Special case:multi-prototype features
f(x, y)def= [y = class1] · x ,
[y = class2] · x ,· · ·[y = classC ] · x
prediction(w, x)def=
update(x, ygold , ypred)def=
Binary Perceptron
[w · x > 0]
α(ygold − ypred) · x
Multi-class Perceptron
arg maxy w · f(x, y)
α(f(x, ygold)− f(x, ypred)
)w := w + αf(x, ygold)− αf(x, ypred)
![Page 38: Machine Translation Zoo - Univerzita Karlovaufal.mff.cuni.cz/~popel/papers/2013_05_06_zoo.pdf · Machine Translation Zoo Tree-to-tree transfer and Discriminative learning Martin Popel](https://reader030.fdocuments.us/reader030/viewer/2022011900/5f01d7387e708231d4014c16/html5/thumbnails/38.jpg)
MT Intro Online Learning Guided Learning Back to MT
Perceptron
w := 0while (x, ygold) := get new data()
ypred := prediction(w, x)w += update(x, ygold , ypred)
Output: w
General case:any label-dependent features, e.g.
f101(x, y)def= [(y=NNP or y=NNPS)
and x capitalized ]
prediction(w, x)def=
update(x, ygold , ypred)def=
Binary Perceptron
[w · x > 0]
α(ygold − ypred) · x
Multi-class Perceptron
arg maxy w · f(x, y)
α(f(x, ygold)− f(x, ypred)
)
![Page 39: Machine Translation Zoo - Univerzita Karlovaufal.mff.cuni.cz/~popel/papers/2013_05_06_zoo.pdf · Machine Translation Zoo Tree-to-tree transfer and Discriminative learning Martin Popel](https://reader030.fdocuments.us/reader030/viewer/2022011900/5f01d7387e708231d4014c16/html5/thumbnails/39.jpg)
MT Intro Online Learning Guided Learning Back to MT
Structured Prediction
the number of possible labels is huge
labels y have a structure (graph, tree, sequence,. . . )
usually can be decomposed (factorized) into subproblems
local features
fi (x, y, j) can use whole x, but only such yk where k is “near” j
f101(x, y, j)def= [ (yj=NNP or yj=NNPS) and word xj capitalized ]
f102(x, y, j)def= [ yj=NNP and yj−1=NNP and |x| ≤ 6 ]
global features
Fi (x, y)def=∑
j fi (x, y, j)F101 . . . number of capitalized words with tag NNP or NNPSF102 . . . number of NNP followed by NNP
or 0 if the sentence is longer than six wordsWe can define also features that cannot be decomposed
![Page 40: Machine Translation Zoo - Univerzita Karlovaufal.mff.cuni.cz/~popel/papers/2013_05_06_zoo.pdf · Machine Translation Zoo Tree-to-tree transfer and Discriminative learning Martin Popel](https://reader030.fdocuments.us/reader030/viewer/2022011900/5f01d7387e708231d4014c16/html5/thumbnails/40.jpg)
MT Intro Online Learning Guided Learning Back to MT
Structured Prediction using Online Learning
1 local approach
update after each local decisionoutput of previous decisions used in local featurese.g. Structured Perceptron (Collins, 2002)ypred = arg maxy
∑i wi fi (x, yj , yj−1, ...)
2 global approach
generate n-best list (lattice) of outputs y for the whole xcompute global features, do update for each x (sentence)we are re-ranking the n-best liste.g. MIRA (Crammer and Singer, 2003)ypred = arg maxy
∑i wiFi (x, y)
![Page 41: Machine Translation Zoo - Univerzita Karlovaufal.mff.cuni.cz/~popel/papers/2013_05_06_zoo.pdf · Machine Translation Zoo Tree-to-tree transfer and Discriminative learning Martin Popel](https://reader030.fdocuments.us/reader030/viewer/2022011900/5f01d7387e708231d4014c16/html5/thumbnails/41.jpg)
MT Intro Online Learning Guided Learning Back to MT
Margin-based Online Learning
Definitions
score(y) = w · f(x, y)
margin(y) =score(ygold)− score(y)
margin > 0⇒ no error|margin| ∼ confidence
hinge loss(y) =max
(0, 1−margin(y)
)Online Prediction and Update
ypreddef= arg maxw · f(x, y)
w += α(f(x, ygold)− f(x, ypred)
)
Perceptron
αPerc
def= 1 (or any fixed value > 0)
Passive Aggressive (PA)
αPA
def=
hinge loss(ypred )||f(x,ygold )−f(x,ypred )||2
Passive Aggressive I
αPA-I
def= min {C , αPA}
Passive Aggressive II
αPA-II
def=
hinge loss(ypred )
||f(x,ygold )−f(x,ypred )||2+ 12C
![Page 42: Machine Translation Zoo - Univerzita Karlovaufal.mff.cuni.cz/~popel/papers/2013_05_06_zoo.pdf · Machine Translation Zoo Tree-to-tree transfer and Discriminative learning Martin Popel](https://reader030.fdocuments.us/reader030/viewer/2022011900/5f01d7387e708231d4014c16/html5/thumbnails/42.jpg)
MT Intro Online Learning Guided Learning Back to MT
Margin-based Online Learning
Definitions
score(y) = w · f(x, y)
margin(y) =score(ygold)− score(y)
margin > 0⇒ no error|margin| ∼ confidence
hinge loss(y) =max
(0, 1−margin(y)
)Online Prediction and Update
ypreddef= arg maxw · f(x, y)
w += α(f(x, ygold)− f(x, ypred)
)
Perceptron
αPerc
def= 1 (or any fixed value > 0)
Passive Aggressive (PA)
αPA
def=
hinge loss(ypred )||f(x,ygold )−f(x,ypred )||2
Passive Aggressive I
αPA-I
def= min {C , αPA}
Passive Aggressive II
αPA-II
def=
hinge loss(ypred )
||f(x,ygold )−f(x,ypred )||2+ 12C
![Page 43: Machine Translation Zoo - Univerzita Karlovaufal.mff.cuni.cz/~popel/papers/2013_05_06_zoo.pdf · Machine Translation Zoo Tree-to-tree transfer and Discriminative learning Martin Popel](https://reader030.fdocuments.us/reader030/viewer/2022011900/5f01d7387e708231d4014c16/html5/thumbnails/43.jpg)
MT Intro Online Learning Guided Learning Back to MT
Margin-based Online Learning
Definitions
score(y) = w · f(x, y)
margin(y) =score(ygold)− score(y)
margin > 0⇒ no error|margin| ∼ confidence
hinge loss(y) =max
(0, 1−margin(y)
)Online Prediction and Update
ypreddef= arg maxw · f(x, y)
w += α(f(x, ygold)− f(x, ypred)
)
Perceptron
αPerc
def= 1 (or any fixed value > 0)
Passive Aggressive (PA)
αPA
def=
hinge loss(ypred )||f(x,ygold )−f(x,ypred )||2
Passive Aggressive I
αPA-I
def= min {C , αPA}
Passive Aggressive II
αPA-II
def=
hinge loss(ypred )
||f(x,ygold )−f(x,ypred )||2+ 12C
![Page 44: Machine Translation Zoo - Univerzita Karlovaufal.mff.cuni.cz/~popel/papers/2013_05_06_zoo.pdf · Machine Translation Zoo Tree-to-tree transfer and Discriminative learning Martin Popel](https://reader030.fdocuments.us/reader030/viewer/2022011900/5f01d7387e708231d4014c16/html5/thumbnails/44.jpg)
MT Intro Online Learning Guided Learning Back to MT
Margin-based Online Learning
Definitions
score(y) = w · f(x, y)
margin(y) =score(ygold)− score(y)
margin > 0⇒ no error|margin| ∼ confidence
hinge loss(y) =max
(0, 1−margin(y)
)Online Prediction and Update
ypreddef= arg maxw · f(x, y)
w += α(f(x, ygold)− f(x, ypred)
)
Perceptron
αPerc
def= 1 (or any fixed value > 0)
Passive Aggressive (PA)
αPA
def=
hinge loss(ypred )||f(x,ygold )−f(x,ypred )||2
Passive Aggressive I
αPA-I
def= min {C , αPA}
Passive Aggressive II
αPA-II
def=
hinge loss(ypred )
||f(x,ygold )−f(x,ypred )||2+ 12C
![Page 45: Machine Translation Zoo - Univerzita Karlovaufal.mff.cuni.cz/~popel/papers/2013_05_06_zoo.pdf · Machine Translation Zoo Tree-to-tree transfer and Discriminative learning Martin Popel](https://reader030.fdocuments.us/reader030/viewer/2022011900/5f01d7387e708231d4014c16/html5/thumbnails/45.jpg)
MT Intro Online Learning Guided Learning Back to MT
Margin-based Online Learning
Definitions
score(y) = w · f(x, y)
margin(y) =score(ygold)− score(y)
margin > 0⇒ no error|margin| ∼ confidence
hinge loss(y) =max
(0, 1−margin(y)
)Online Prediction and Update
ypreddef= arg maxw · f(x, y)
w += α(f(x, ygold)− f(x, ypred)
)
Perceptron
αPerc
def= 1 (or any fixed value > 0)
Passive Aggressive (PA)
αPA
def=
hinge loss(ypred )||f(x,ygold )−f(x,ypred )||2
Passive Aggressive I
αPA-I
def= min {C , αPA}
Passive Aggressive II
αPA-II
def=
hinge loss(ypred )
||f(x,ygold )−f(x,ypred )||2+ 12C
![Page 46: Machine Translation Zoo - Univerzita Karlovaufal.mff.cuni.cz/~popel/papers/2013_05_06_zoo.pdf · Machine Translation Zoo Tree-to-tree transfer and Discriminative learning Martin Popel](https://reader030.fdocuments.us/reader030/viewer/2022011900/5f01d7387e708231d4014c16/html5/thumbnails/46.jpg)
MT Intro Online Learning Guided Learning Back to MT
Cost-sensitive Online Learning
Definitions
cost(y) = external error metric (non-negative)e.g. 1 - similarity of y and ygold
hinge loss(y) = max(0, cost(y)−margin(y)
)Hope and Fear
w += α(f(x, ygold)− f(x, ypred)
)min-cost yhope
def= arg maxy −cost(y)
max-score yfeardef= arg maxy score(y)
cost-diminished yhopedef= arg maxy score(y)− cost(y)
cost-augmented yfeardef= arg maxy score(y) + cost(y)
max-cost yfeardef= arg maxy cost(y)
![Page 47: Machine Translation Zoo - Univerzita Karlovaufal.mff.cuni.cz/~popel/papers/2013_05_06_zoo.pdf · Machine Translation Zoo Tree-to-tree transfer and Discriminative learning Martin Popel](https://reader030.fdocuments.us/reader030/viewer/2022011900/5f01d7387e708231d4014c16/html5/thumbnails/47.jpg)
MT Intro Online Learning Guided Learning Back to MT
Cost-sensitive Online Learning
Definitions
cost(y) = external error metric (non-negative)e.g. 1 - similarity of y and ygold
hinge loss(y) = max(0, cost(y)−margin(y)
)Hope and Fear
w += α(f(x, ygold)− f(x, ypred)
)min-cost yhope
def= arg maxy −cost(y)
max-score yfeardef= arg maxy score(y)
cost-diminished yhopedef= arg maxy score(y)− cost(y)
cost-augmented yfeardef= arg maxy score(y) + cost(y)
max-cost yfeardef= arg maxy cost(y)
![Page 48: Machine Translation Zoo - Univerzita Karlovaufal.mff.cuni.cz/~popel/papers/2013_05_06_zoo.pdf · Machine Translation Zoo Tree-to-tree transfer and Discriminative learning Martin Popel](https://reader030.fdocuments.us/reader030/viewer/2022011900/5f01d7387e708231d4014c16/html5/thumbnails/48.jpg)
MT Intro Online Learning Guided Learning Back to MT
Cost-sensitive Online Learning
Definitions
cost(y) = external error metric (non-negative)e.g. 1 - similarity of y and ygold
hinge loss(y) = max(0, cost(y)−margin(y)
)Hope and Fear
w += α(f(x, yhope)− f(x, yfear )
)min-cost yhope
def= arg maxy −cost(y)
max-score yfeardef= arg maxy score(y)
cost-diminished yhopedef= arg maxy score(y)− cost(y)
cost-augmented yfeardef= arg maxy score(y) + cost(y)
max-cost yfeardef= arg maxy cost(y)
![Page 49: Machine Translation Zoo - Univerzita Karlovaufal.mff.cuni.cz/~popel/papers/2013_05_06_zoo.pdf · Machine Translation Zoo Tree-to-tree transfer and Discriminative learning Martin Popel](https://reader030.fdocuments.us/reader030/viewer/2022011900/5f01d7387e708231d4014c16/html5/thumbnails/49.jpg)
MT Intro Online Learning Guided Learning Back to MT
Cost-sensitive Online Learning
Hypothesis Selection -cost
score
min-cost(max-BLEU) hope
max-score fear (ypred)
cost-diminished hope
cost-augmented fear
n-best listmax-cost fear
![Page 50: Machine Translation Zoo - Univerzita Karlovaufal.mff.cuni.cz/~popel/papers/2013_05_06_zoo.pdf · Machine Translation Zoo Tree-to-tree transfer and Discriminative learning Martin Popel](https://reader030.fdocuments.us/reader030/viewer/2022011900/5f01d7387e708231d4014c16/html5/thumbnails/50.jpg)
MT Intro Online Learning Guided Learning Back to MT
Cost-sensitive Online Learning
Hypothesis Selection -cost
score
min-cost(max-BLEU) hope
max-score fear (ypred)
cost-diminished hope
cost-augmented fear
n-best listmax-cost fear
![Page 51: Machine Translation Zoo - Univerzita Karlovaufal.mff.cuni.cz/~popel/papers/2013_05_06_zoo.pdf · Machine Translation Zoo Tree-to-tree transfer and Discriminative learning Martin Popel](https://reader030.fdocuments.us/reader030/viewer/2022011900/5f01d7387e708231d4014c16/html5/thumbnails/51.jpg)
MT Intro Online Learning Guided Learning Back to MT
Cost-sensitive Online Learning
Hypothesis Selection -cost
score
min-cost(max-BLEU) hope
max-score fear (ypred)
cost-diminished hope
cost-augmented fear
n-best listmax-cost fear
![Page 52: Machine Translation Zoo - Univerzita Karlovaufal.mff.cuni.cz/~popel/papers/2013_05_06_zoo.pdf · Machine Translation Zoo Tree-to-tree transfer and Discriminative learning Martin Popel](https://reader030.fdocuments.us/reader030/viewer/2022011900/5f01d7387e708231d4014c16/html5/thumbnails/52.jpg)
MT Intro Online Learning Guided Learning Back to MT
Cost-sensitive Online Learning
Hypothesis Selection -cost
score
min-cost(max-BLEU) hope
max-score fear (ypred)
cost-diminished hope
cost-augmented fear
n-best listmax-cost fear
![Page 53: Machine Translation Zoo - Univerzita Karlovaufal.mff.cuni.cz/~popel/papers/2013_05_06_zoo.pdf · Machine Translation Zoo Tree-to-tree transfer and Discriminative learning Martin Popel](https://reader030.fdocuments.us/reader030/viewer/2022011900/5f01d7387e708231d4014c16/html5/thumbnails/53.jpg)
MT Intro Online Learning Guided Learning Back to MT
Cost-sensitive Online Learning
Hypothesis Selection -cost
score
min-cost(max-BLEU) hope
max-score fear (ypred)
cost-diminished hope
cost-augmented fear
n-best listmax-cost fear
![Page 54: Machine Translation Zoo - Univerzita Karlovaufal.mff.cuni.cz/~popel/papers/2013_05_06_zoo.pdf · Machine Translation Zoo Tree-to-tree transfer and Discriminative learning Martin Popel](https://reader030.fdocuments.us/reader030/viewer/2022011900/5f01d7387e708231d4014c16/html5/thumbnails/54.jpg)
MT Intro Online Learning Guided Learning Back to MT
Cost-sensitive Online Learning
Hypothesis Selection -cost
score
min-cost(max-BLEU) hope
max-score fear (ypred)
cost-diminished hope
cost-augmented fear
n-best list
max-cost fear
![Page 55: Machine Translation Zoo - Univerzita Karlovaufal.mff.cuni.cz/~popel/papers/2013_05_06_zoo.pdf · Machine Translation Zoo Tree-to-tree transfer and Discriminative learning Martin Popel](https://reader030.fdocuments.us/reader030/viewer/2022011900/5f01d7387e708231d4014c16/html5/thumbnails/55.jpg)
MT Intro Online Learning Guided Learning Back to MT
Cost-sensitive Online Learning
Hypothesis Selection -cost
score
min-cost(max-BLEU) hope
max-score fear (ypred)
cost-diminished hope
cost-augmented fear
n-best listmax-cost fear
![Page 56: Machine Translation Zoo - Univerzita Karlovaufal.mff.cuni.cz/~popel/papers/2013_05_06_zoo.pdf · Machine Translation Zoo Tree-to-tree transfer and Discriminative learning Martin Popel](https://reader030.fdocuments.us/reader030/viewer/2022011900/5f01d7387e708231d4014c16/html5/thumbnails/56.jpg)
MT Intro Online Learning Guided Learning Back to MT
Application to MT
x = source sentenceygold = its reference translation
more references sometimes available
reference may be unreachable
we score derivations (which include latent variables)one translation may have more derivations
![Page 57: Machine Translation Zoo - Univerzita Karlovaufal.mff.cuni.cz/~popel/papers/2013_05_06_zoo.pdf · Machine Translation Zoo Tree-to-tree transfer and Discriminative learning Martin Popel](https://reader030.fdocuments.us/reader030/viewer/2022011900/5f01d7387e708231d4014c16/html5/thumbnails/57.jpg)
Today’s Menu
1 MT IntroTaxonomyHybrids
2 Online LearningPerceptronStructured Prediction
3 Guided Learning
4 Back to MTEasy-First Decoding in MTGuided Learning in MT
![Page 58: Machine Translation Zoo - Univerzita Karlovaufal.mff.cuni.cz/~popel/papers/2013_05_06_zoo.pdf · Machine Translation Zoo Tree-to-tree transfer and Discriminative learning Martin Popel](https://reader030.fdocuments.us/reader030/viewer/2022011900/5f01d7387e708231d4014c16/html5/thumbnails/58.jpg)
MT Intro Online Learning Guided Learning Back to MT
Easy-First Decoding (PoS Tagging)
score
Agatha found that book interesting
NN
VB
JJ
DT
DT
f123def= [yj = DT , yj+1 = NN]
IN
RB
f124def= [yj = RB, yj+1 = NN]
VBN
VBD
NNP
yfear
yhope
(Shen, Satta and Joshi, 2007)
![Page 59: Machine Translation Zoo - Univerzita Karlovaufal.mff.cuni.cz/~popel/papers/2013_05_06_zoo.pdf · Machine Translation Zoo Tree-to-tree transfer and Discriminative learning Martin Popel](https://reader030.fdocuments.us/reader030/viewer/2022011900/5f01d7387e708231d4014c16/html5/thumbnails/59.jpg)
MT Intro Online Learning Guided Learning Back to MT
Easy-First Decoding (PoS Tagging)
score
Agatha found that book interesting
NN
VB
JJ
DT
DT
f123def= [yj = DT , yj+1 = NN]
IN
RB
f124def= [yj = RB, yj+1 = NN]
VBN
VBD
NNP
yfear
yhope
(Shen, Satta and Joshi, 2007)
![Page 60: Machine Translation Zoo - Univerzita Karlovaufal.mff.cuni.cz/~popel/papers/2013_05_06_zoo.pdf · Machine Translation Zoo Tree-to-tree transfer and Discriminative learning Martin Popel](https://reader030.fdocuments.us/reader030/viewer/2022011900/5f01d7387e708231d4014c16/html5/thumbnails/60.jpg)
MT Intro Online Learning Guided Learning Back to MT
Easy-First Decoding (PoS Tagging)
score
Agatha found that book interesting
NN
VB
JJ
DT
DT
f123def= [yj = DT , yj+1 = NN]
IN
RB
f124def= [yj = RB, yj+1 = NN]
VBN
VBD
NNP
yfear
yhope
(Shen, Satta and Joshi, 2007)
![Page 61: Machine Translation Zoo - Univerzita Karlovaufal.mff.cuni.cz/~popel/papers/2013_05_06_zoo.pdf · Machine Translation Zoo Tree-to-tree transfer and Discriminative learning Martin Popel](https://reader030.fdocuments.us/reader030/viewer/2022011900/5f01d7387e708231d4014c16/html5/thumbnails/61.jpg)
MT Intro Online Learning Guided Learning Back to MT
Easy-First Decoding (PoS Tagging)
score
Agatha found that book interesting
NN
VB
JJ
DT
DT
f123def= [yj = DT , yj+1 = NN]
IN
RB
f124def= [yj = RB, yj+1 = NN]
VBN
VBD
NNP
yfear
yhope
(Shen, Satta and Joshi, 2007)
![Page 62: Machine Translation Zoo - Univerzita Karlovaufal.mff.cuni.cz/~popel/papers/2013_05_06_zoo.pdf · Machine Translation Zoo Tree-to-tree transfer and Discriminative learning Martin Popel](https://reader030.fdocuments.us/reader030/viewer/2022011900/5f01d7387e708231d4014c16/html5/thumbnails/62.jpg)
MT Intro Online Learning Guided Learning Back to MT
Easy-First Decoding (PoS Tagging)
score
Agatha found that book interesting
NN
VB
JJ
DT
DT
f123def= [yj = DT , yj+1 = NN]
IN
RB
f124def= [yj = RB, yj+1 = NN]
VBN
VBD
NNP
yfear
yhope
(Shen, Satta and Joshi, 2007)
![Page 63: Machine Translation Zoo - Univerzita Karlovaufal.mff.cuni.cz/~popel/papers/2013_05_06_zoo.pdf · Machine Translation Zoo Tree-to-tree transfer and Discriminative learning Martin Popel](https://reader030.fdocuments.us/reader030/viewer/2022011900/5f01d7387e708231d4014c16/html5/thumbnails/63.jpg)
MT Intro Online Learning Guided Learning Back to MT
Easy-First Decoding (PoS Tagging)
score
Agatha found that book interesting
NN
VB
JJ
DT
DT
f123def= [yj = DT , yj+1 = NN]
IN
RB
f124def= [yj = RB, yj+1 = NN]
VBN
VBD
NNP
yfear
yhope
(Shen, Satta and Joshi, 2007)
![Page 64: Machine Translation Zoo - Univerzita Karlovaufal.mff.cuni.cz/~popel/papers/2013_05_06_zoo.pdf · Machine Translation Zoo Tree-to-tree transfer and Discriminative learning Martin Popel](https://reader030.fdocuments.us/reader030/viewer/2022011900/5f01d7387e708231d4014c16/html5/thumbnails/64.jpg)
MT Intro Online Learning Guided Learning Back to MT
Easy-First Decoding (PoS Tagging)
score
Agatha found that book interesting
NN
VB
JJ
DT
DT
f123def= [yj = DT , yj+1 = NN]
IN
RB
f124def= [yj = RB, yj+1 = NN]
VBN
VBD
NNP
yfear
yhope
(Shen, Satta and Joshi, 2007)
![Page 65: Machine Translation Zoo - Univerzita Karlovaufal.mff.cuni.cz/~popel/papers/2013_05_06_zoo.pdf · Machine Translation Zoo Tree-to-tree transfer and Discriminative learning Martin Popel](https://reader030.fdocuments.us/reader030/viewer/2022011900/5f01d7387e708231d4014c16/html5/thumbnails/65.jpg)
MT Intro Online Learning Guided Learning Back to MT
Easy-First Decoding (PoS Tagging)
score
Agatha found that book interesting
NN
VB
JJDT
DT
f123def= [yj = DT , yj+1 = NN]
IN
RB
f124def= [yj = RB, yj+1 = NN]
VBN
VBD
NNP
yfear
yhope
(Shen, Satta and Joshi, 2007)
![Page 66: Machine Translation Zoo - Univerzita Karlovaufal.mff.cuni.cz/~popel/papers/2013_05_06_zoo.pdf · Machine Translation Zoo Tree-to-tree transfer and Discriminative learning Martin Popel](https://reader030.fdocuments.us/reader030/viewer/2022011900/5f01d7387e708231d4014c16/html5/thumbnails/66.jpg)
MT Intro Online Learning Guided Learning Back to MT
Easy-First Decoding (PoS Tagging)
score
Agatha found that book interesting
NN
VB
JJDT
DT
f123def= [yj = DT , yj+1 = NN]
IN
RB
f124def= [yj = RB, yj+1 = NN]
VBN
VBD
NNP
yfear
yhope
(Shen, Satta and Joshi, 2007)
![Page 67: Machine Translation Zoo - Univerzita Karlovaufal.mff.cuni.cz/~popel/papers/2013_05_06_zoo.pdf · Machine Translation Zoo Tree-to-tree transfer and Discriminative learning Martin Popel](https://reader030.fdocuments.us/reader030/viewer/2022011900/5f01d7387e708231d4014c16/html5/thumbnails/67.jpg)
MT Intro Online Learning Guided Learning Back to MT
Guided Learning (PoS Tagging)
score
Agatha found that book interesting
NN
VB
JJDT
DT
f123def= [yj = DT , yj+1 = NN]
IN
RB
f124def= [yj = RB, yj+1 = NN]
VBN
VBD
NNP
yfear
yhope
(Shen, Satta and Joshi, 2007)
![Page 68: Machine Translation Zoo - Univerzita Karlovaufal.mff.cuni.cz/~popel/papers/2013_05_06_zoo.pdf · Machine Translation Zoo Tree-to-tree transfer and Discriminative learning Martin Popel](https://reader030.fdocuments.us/reader030/viewer/2022011900/5f01d7387e708231d4014c16/html5/thumbnails/68.jpg)
MT Intro Online Learning Guided Learning Back to MT
Easy-First Decoding (Dependency Parsing)
(Goldberg and Elhadad, 2010)
![Page 69: Machine Translation Zoo - Univerzita Karlovaufal.mff.cuni.cz/~popel/papers/2013_05_06_zoo.pdf · Machine Translation Zoo Tree-to-tree transfer and Discriminative learning Martin Popel](https://reader030.fdocuments.us/reader030/viewer/2022011900/5f01d7387e708231d4014c16/html5/thumbnails/69.jpg)
MT Intro Online Learning Guided Learning Back to MT
Easy-First Decoding (Dependency Parsing)
(Goldberg and Elhadad, 2010)
![Page 70: Machine Translation Zoo - Univerzita Karlovaufal.mff.cuni.cz/~popel/papers/2013_05_06_zoo.pdf · Machine Translation Zoo Tree-to-tree transfer and Discriminative learning Martin Popel](https://reader030.fdocuments.us/reader030/viewer/2022011900/5f01d7387e708231d4014c16/html5/thumbnails/70.jpg)
MT Intro Online Learning Guided Learning Back to MT
Easy-First Decoding (Dependency Parsing)
(Goldberg and Elhadad, 2010)
![Page 71: Machine Translation Zoo - Univerzita Karlovaufal.mff.cuni.cz/~popel/papers/2013_05_06_zoo.pdf · Machine Translation Zoo Tree-to-tree transfer and Discriminative learning Martin Popel](https://reader030.fdocuments.us/reader030/viewer/2022011900/5f01d7387e708231d4014c16/html5/thumbnails/71.jpg)
MT Intro Online Learning Guided Learning Back to MT
Easy-First Decoding (Dependency Parsing)
(Goldberg and Elhadad, 2010)
![Page 72: Machine Translation Zoo - Univerzita Karlovaufal.mff.cuni.cz/~popel/papers/2013_05_06_zoo.pdf · Machine Translation Zoo Tree-to-tree transfer and Discriminative learning Martin Popel](https://reader030.fdocuments.us/reader030/viewer/2022011900/5f01d7387e708231d4014c16/html5/thumbnails/72.jpg)
Today’s Menu
1 MT IntroTaxonomyHybrids
2 Online LearningPerceptronStructured Prediction
3 Guided Learning
4 Back to MTEasy-First Decoding in MTGuided Learning in MT
![Page 73: Machine Translation Zoo - Univerzita Karlovaufal.mff.cuni.cz/~popel/papers/2013_05_06_zoo.pdf · Machine Translation Zoo Tree-to-tree transfer and Discriminative learning Martin Popel](https://reader030.fdocuments.us/reader030/viewer/2022011900/5f01d7387e708231d4014c16/html5/thumbnails/73.jpg)
MT Intro Online Learning Guided Learning Back to MT
Easy-First Decoding (Phrase-Based MT)
score
Agatha
Agate
found
prisla
that
ta
book
kniha
interesting
zajımava
![Page 74: Machine Translation Zoo - Univerzita Karlovaufal.mff.cuni.cz/~popel/papers/2013_05_06_zoo.pdf · Machine Translation Zoo Tree-to-tree transfer and Discriminative learning Martin Popel](https://reader030.fdocuments.us/reader030/viewer/2022011900/5f01d7387e708231d4014c16/html5/thumbnails/74.jpg)
MT Intro Online Learning Guided Learning Back to MT
Easy-First Decoding (Phrase-Based MT)
score
Agatha
Agate
found
prisla
that
ta
book
kniha
interesting
zajımava
Agatha
prisla
nasel
zjistil, ze
ze
ta
ze rezervovat
ta kniha
ze kniha
kniha
rezervovat
kniha zajımava
kniha zajımavy
zajımavy
zajımava
![Page 75: Machine Translation Zoo - Univerzita Karlovaufal.mff.cuni.cz/~popel/papers/2013_05_06_zoo.pdf · Machine Translation Zoo Tree-to-tree transfer and Discriminative learning Martin Popel](https://reader030.fdocuments.us/reader030/viewer/2022011900/5f01d7387e708231d4014c16/html5/thumbnails/75.jpg)
MT Intro Online Learning Guided Learning Back to MT
Easy-First Decoding (Phrase-Based MT)
score
Agatha
Agate
found
prisla
that
ta
book
kniha
interesting
zajımava
ze rezervovat
ta kniha
ze kniha
kniha
rezervovat
kniha zajımava
kniha zajımavy
![Page 76: Machine Translation Zoo - Univerzita Karlovaufal.mff.cuni.cz/~popel/papers/2013_05_06_zoo.pdf · Machine Translation Zoo Tree-to-tree transfer and Discriminative learning Martin Popel](https://reader030.fdocuments.us/reader030/viewer/2022011900/5f01d7387e708231d4014c16/html5/thumbnails/76.jpg)
MT Intro Online Learning Guided Learning Back to MT
Easy-First Decoding (Phrase-Based MT)
score
Agatha
Agate
found
prisla
that
ta
book
kniha
interesting
zajımava
ze rezervovat
ta kniha
ze kniha
kniha
rezervovat
kniha zajımava
kniha zajımavy
![Page 77: Machine Translation Zoo - Univerzita Karlovaufal.mff.cuni.cz/~popel/papers/2013_05_06_zoo.pdf · Machine Translation Zoo Tree-to-tree transfer and Discriminative learning Martin Popel](https://reader030.fdocuments.us/reader030/viewer/2022011900/5f01d7387e708231d4014c16/html5/thumbnails/77.jpg)
MT Intro Online Learning Guided Learning Back to MT
Easy-First Decoding (Phrase-Based MT)
score
Agatha
Agate
found
prisla
that
ta
book
kniha
interesting
zajımava
kniha zajımava
kniha zajımavy
zajımavy
zajımava
language model
![Page 78: Machine Translation Zoo - Univerzita Karlovaufal.mff.cuni.cz/~popel/papers/2013_05_06_zoo.pdf · Machine Translation Zoo Tree-to-tree transfer and Discriminative learning Martin Popel](https://reader030.fdocuments.us/reader030/viewer/2022011900/5f01d7387e708231d4014c16/html5/thumbnails/78.jpg)
MT Intro Online Learning Guided Learning Back to MT
Easy-First Decoding (Phrase-Based MT)
score
Agatha
Agate
found
prisla
that
ta
book
kniha
interesting
zajımava
kniha zajımava
kniha zajımavy
zajımavy
zajımava
language model
![Page 79: Machine Translation Zoo - Univerzita Karlovaufal.mff.cuni.cz/~popel/papers/2013_05_06_zoo.pdf · Machine Translation Zoo Tree-to-tree transfer and Discriminative learning Martin Popel](https://reader030.fdocuments.us/reader030/viewer/2022011900/5f01d7387e708231d4014c16/html5/thumbnails/79.jpg)
MT Intro Online Learning Guided Learning Back to MT
Easy-First Decoding (Phrase-Based MT)
score
Agatha
Agate
found
prisla
that
ta
book
kniha
interesting
zajımava
kniha zajımava
kniha zajımavy
zajımavy
zajımava
![Page 80: Machine Translation Zoo - Univerzita Karlovaufal.mff.cuni.cz/~popel/papers/2013_05_06_zoo.pdf · Machine Translation Zoo Tree-to-tree transfer and Discriminative learning Martin Popel](https://reader030.fdocuments.us/reader030/viewer/2022011900/5f01d7387e708231d4014c16/html5/thumbnails/80.jpg)
MT Intro Online Learning Guided Learning Back to MT
Easy-First Decoding (Phrase-Based MT)
score
Agatha
Agate
found
prisla
that
ta
book
kniha
interesting
zajımava
![Page 81: Machine Translation Zoo - Univerzita Karlovaufal.mff.cuni.cz/~popel/papers/2013_05_06_zoo.pdf · Machine Translation Zoo Tree-to-tree transfer and Discriminative learning Martin Popel](https://reader030.fdocuments.us/reader030/viewer/2022011900/5f01d7387e708231d4014c16/html5/thumbnails/81.jpg)
MT Intro Online Learning Guided Learning Back to MT
Easy-First Decoding (Phrase-Based MT)
score
Agatha
Agate
found
prisla
that
ta
book
kniha
interesting
zajımava
prisla
nasel
zjistil, ze
ze
ta
![Page 82: Machine Translation Zoo - Univerzita Karlovaufal.mff.cuni.cz/~popel/papers/2013_05_06_zoo.pdf · Machine Translation Zoo Tree-to-tree transfer and Discriminative learning Martin Popel](https://reader030.fdocuments.us/reader030/viewer/2022011900/5f01d7387e708231d4014c16/html5/thumbnails/82.jpg)
MT Intro Online Learning Guided Learning Back to MT
Features for Guided Learning in MT
Source Segment Features
segment size (number of words)
entropy P(target|source) = −∑
i P(src , trgi ) · logP(trgi |src)
log count(source)
source language model: log P(source)
word identity, e.g. f42def= [src=found that]
PoS identity, e.g. f43def= [src pos=VBD IN]
Target-dependent Features
log P(trg|src)
target language model: log P(target | previous segment)
log count(target)?
identity, e.g. f142def= [src=found that & trg=zjistil]
Combinations and Quantizations
[size(src) = 3] · logP(trg |src)[size(src) = 3 &− 3 < logP(trg |src) < −2]etc.
![Page 83: Machine Translation Zoo - Univerzita Karlovaufal.mff.cuni.cz/~popel/papers/2013_05_06_zoo.pdf · Machine Translation Zoo Tree-to-tree transfer and Discriminative learning Martin Popel](https://reader030.fdocuments.us/reader030/viewer/2022011900/5f01d7387e708231d4014c16/html5/thumbnails/83.jpg)
MT Intro Online Learning Guided Learning Back to MT
Features for Guided Learning in MT
Source Segment Features
segment size (number of words)
entropy P(target|source) = −∑
i P(src , trgi ) · logP(trgi |src)
log count(source)
source language model: log P(source)
word identity, e.g. f42def= [src=found that]
PoS identity, e.g. f43def= [src pos=VBD IN]
Target-dependent Features
log P(trg|src)
target language model: log P(target | previous segment)
log count(target)?
identity, e.g. f142def= [src=found that & trg=zjistil]
Combinations and Quantizations
[size(src) = 3] · logP(trg |src)[size(src) = 3 &− 3 < logP(trg |src) < −2]etc.
![Page 84: Machine Translation Zoo - Univerzita Karlovaufal.mff.cuni.cz/~popel/papers/2013_05_06_zoo.pdf · Machine Translation Zoo Tree-to-tree transfer and Discriminative learning Martin Popel](https://reader030.fdocuments.us/reader030/viewer/2022011900/5f01d7387e708231d4014c16/html5/thumbnails/84.jpg)
MT Intro Online Learning Guided Learning Back to MT
Application to Tecto Trees
findv:fin
Agathan:subj
bookn:obj
thisadj:attr
interestingadj:compl
![Page 85: Machine Translation Zoo - Univerzita Karlovaufal.mff.cuni.cz/~popel/papers/2013_05_06_zoo.pdf · Machine Translation Zoo Tree-to-tree transfer and Discriminative learning Martin Popel](https://reader030.fdocuments.us/reader030/viewer/2022011900/5f01d7387e708231d4014c16/html5/thumbnails/85.jpg)
MT Intro Online Learning Guided Learning Back to MT
Application to Tecto Trees
v:fin
find
n:subj
Agatha
n:obj
book
adj:attr
this
adj:compl
interesting
![Page 86: Machine Translation Zoo - Univerzita Karlovaufal.mff.cuni.cz/~popel/papers/2013_05_06_zoo.pdf · Machine Translation Zoo Tree-to-tree transfer and Discriminative learning Martin Popel](https://reader030.fdocuments.us/reader030/viewer/2022011900/5f01d7387e708231d4014c16/html5/thumbnails/86.jpg)
MT Intro Online Learning Guided Learning Back to MT
What have you seen in the Zoo
![Page 87: Machine Translation Zoo - Univerzita Karlovaufal.mff.cuni.cz/~popel/papers/2013_05_06_zoo.pdf · Machine Translation Zoo Tree-to-tree transfer and Discriminative learning Martin Popel](https://reader030.fdocuments.us/reader030/viewer/2022011900/5f01d7387e708231d4014c16/html5/thumbnails/87.jpg)
MT Intro Online Learning Guided Learning Back to MT
Predictions?
Hope and Fear
http://syd.korpus.cz/
http://books.google.com/ngrams/
![Page 88: Machine Translation Zoo - Univerzita Karlovaufal.mff.cuni.cz/~popel/papers/2013_05_06_zoo.pdf · Machine Translation Zoo Tree-to-tree transfer and Discriminative learning Martin Popel](https://reader030.fdocuments.us/reader030/viewer/2022011900/5f01d7387e708231d4014c16/html5/thumbnails/88.jpg)
MT Intro Online Learning Guided Learning Back to MT
Predictions?
Hope and Fear
http://syd.korpus.cz/
http://books.google.com/ngrams/