Issuu on Google+

Derivations, Features and Models Synchronous Derivation

Product of Experts Model Models that factor over rules

lo haré de muy buen grado .

! r

P (er |fr ) P (fr |er )

λ3

λ2

...

Language model factors over n-grams

Grammar X → 〈 lo haré X . ; I will do it X . 〉 X → 〈 de muy buen grado ; gladly 〉

I !

i=1

P (ei |ei−1 , ..., e1 )

λ1

We want to estimate these models


Derivations, Features and Models Synchronous Derivation X lo haré de muy buen grado .

Product of Experts Model Models that factor over rules

! r

X gladly Grammar X → 〈 lo haré X . ; I will do it X . 〉 X → 〈 de muy buen grado ; gladly 〉

P (er |fr ) P (fr |er )

λ3

λ2

...

Language model factors over n-grams I !

i=1

P (ei |ei−1 , ..., e1 )

λ1

We want to estimate these models


Derivations, Features and Models Synchronous Derivation X X lo haré de muy buen grado .

Product of Experts Model Models that factor over rules

! r

X X I will do it gladly . Grammar X → 〈 lo haré X . ; I will do it X . 〉 X → 〈 de muy buen grado ; gladly 〉

P (er |fr ) P (fr |er )

λ3

λ2

...

Language model factors over n-grams I !

i=1

P (ei |ei−1 , ..., e1 )

λ1

We want to estimate these models


Richer Rules Allow Richer Models In 1993, we aligned words Yo lo harĂŠ maĂąana I will do it tomorrow


Richer Rules Allow Richer Models In 1993, we aligned words Yo lo haré mañana I will do it tomorrow

English (E)

P( E | mañana )

tomorrow morning

0.7 0.3


Richer Rules Allow Richer Models In 1993, we aligned words Yo lo haré mañana I will do it tomorrow

In 1999, we aligned phrases Yo lo haré mañana I will do it tomorrow

English (E)

P( E | mañana )

tomorrow morning

0.7 0.3


Richer Rules Allow Richer Models In 1993, we aligned words Yo lo haré mañana I will do it tomorrow

English (E)

P( E | mañana )

tomorrow morning

0.7 0.3

English (E)

P( E | lo haré )

will do it will do so

0.8 0.2

In 1999, we aligned phrases Yo lo haré mañana I will do it tomorrow


Richer Rules Allow Richer Models In 1993, we aligned words Yo lo haré mañana I will do it tomorrow

English (E)

P( E | mañana )

tomorrow morning

0.7 0.3

English (E)

P( E | lo haré )

will do it will do so

0.8 0.2

In 1999, we aligned phrases Yo lo haré mañana I will do it tomorrow

In 2004, we aligned trees Yo lo haré mañana I will do it tomorrow


Richer Rules Allow Richer Models In 1993, we aligned words Yo lo haré mañana I will do it tomorrow

English (E)

P( E | mañana )

tomorrow morning

0.7 0.3

English (E)

P( E | lo haré )

will do it will do so

0.8 0.2

In 1999, we aligned phrases Yo lo haré mañana I will do it tomorrow

In 2004, we aligned trees Yo lo haré mañana I will do it tomorrow NP VP


Richer Rules Allow Richer Models In 1993, we aligned words Yo lo haré mañana I will do it tomorrow

English (E)

P( E | mañana )

tomorrow morning

0.7 0.3

English (E)

P( E | lo haré )

will do it will do so

0.8 0.2

In 1999, we aligned phrases Yo lo haré mañana I will do it tomorrow

In 2004, we aligned trees Yo lo haré mañana I will do it tomorrow NP VP


Richer Rules Allow Richer Models In 1993, we aligned words Yo lo haré mañana I will do it tomorrow

English (E)

P( E | mañana )

tomorrow morning

0.7 0.3

English (E)

P( E | lo haré )

will do it will do so

0.8 0.2

In 1999, we aligned phrases Yo lo haré mañana I will do it tomorrow

In 2004, we aligned trees VP

Yo lo haré mañana I will do it tomorrow NP VP

P(

MD

VP VB

will do

PRN NP

it

VP

lo haré

NP

) = 0.8


Aligning Structural Components Today, we actually still align words

The Dark Secrets of

MT Revealed


Aligning Structural Components Today, we actually still align words

1

The Dark Secrets of

MT Revealed

Align words with a probabilistic model


Aligning Structural Components Today, we actually still align words

1

The Dark

Align words with a probabilistic model

Secrets of

Yo lo harĂŠ maĂąana

MT

I will do it tomorrow

Revealed


Aligning Structural Components Today, we actually still align words

1

The Dark Secrets of

MT Revealed

2

Align words with a probabilistic model Infer presence of larger structures from this alignment

Yo lo harĂŠ maĂąana I will do it tomorrow


Aligning Structural Components Today, we actually still align words

1

The Dark Secrets of

MT Revealed

2

Align words with a probabilistic model Infer presence of larger structures from this alignment

Yo lo harĂŠ maĂąana I will do it tomorrow


Aligning Structural Components Today, we actually still align words

1

The Dark Secrets of

MT Revealed

2

3

Align words with a probabilistic model Infer presence of larger structures from this alignment Translate with the larger structures

Yo lo harĂŠ maĂąana I will do it tomorrow


Statistical Word Alignment E:

Thank you

,

I

shall

F:

Gracias

lo

harĂŠ

de

,

do

so

gladly

.

muy buen grado

.


Statistical Word Alignment 1

E:

2

3

4

5

6

7

8

9

Thank you

,

I

shall

do

so

gladly

.

Gracias

lo

harĂŠ

de

muy buen grado

.

A:

F:

,

Model Parameters Emissions: P( F1 = Gracias | EA1 = Thank )

Transitions: P( A2 = 3 | A1 = 1)


Statistical Word Alignment 1

E:

2

Thank you

A:

1

F:

Gracias

,

3

4

5

6

7

8

9

,

I

shall

do

so

gladly

.

lo

harĂŠ

de

muy buen grado

.

Model Parameters Emissions: P( F1 = Gracias | EA1 = Thank )

Transitions: P( A2 = 3 | A1 = 1)


Statistical Word Alignment 1

E:

2

Thank you

A:

1

3

F:

Gracias

,

3

4

5

6

7

8

9

,

I

shall

do

so

gladly

.

lo

harĂŠ

de

muy buen grado

.

Model Parameters Emissions: P( F1 = Gracias | EA1 = Thank )

Transitions: P( A2 = 3 | A1 = 1)


Statistical Word Alignment 1

E:

2

Thank you

3

4

5

6

7

8

9

,

I

shall

do

so

gladly

.

8

8

8

9

A:

1

3

7

6

8

F:

Gracias

,

lo

harĂŠ

de

muy buen grado

.

Model Parameters Emissions: P( F1 = Gracias | EA1 = Thank )

Transitions: P( A2 = 3 | A1 = 1)


Syntax-Sensitive Word Alignments

1

E:

2

Thank you

A:

1

F:

Gracias

,

3

4

5

6

7

8

9

,

I

shall

do

so

gladly

.

lo

harĂŠ

de

muy buen grado

.

Model Parameters Emissions: P( F1 = Gracias | EA1 = Thank )

Transitions: P( A2 = 3 | A1 = 1)


Syntax-Sensitive Word Alignments

1

E:

2

Thank you

A:

1

3

F:

Gracias

,

3

4

5

6

7

8

9

,

I

shall

do

so

gladly

.

lo

harĂŠ

de

muy buen grado

.

Model Parameters Emissions: P( F1 = Gracias | EA1 = Thank )

Transitions: P( A2 = 3 | A1 = 1)


Syntax-Sensitive Word Alignments S

,

S VP VB 1

E:

S VP

NP

Thank you

A:

1

3

F:

Gracias

,

VP

MD

NP PRP 2

.

PRP

VB

PRP

ADV

3

4

5

6

7

8

9

,

I

shall

do

so

gladly

.

lo

harĂŠ

de

muy buen grado

.

Model Parameters Emissions: P( F1 = Gracias | EA1 = Thank )

Transitions: P( A2 = 3 | A1 = 1)


Syntax-Sensitive Word Alignments S

,

S VP VB 1

E:

S VP

NP

Thank you

A:

1

3

F:

Gracias

,

VP

MD

NP PRP 2

.

PRP

VB

PRP

ADV

3

4

5

6

7

8

9

,

I

shall

do

so

gladly

.

lo

harĂŠ

de

muy buen grado

.

Model Parameters Emissions: P( F1 = Gracias | EA1 = Thank )

Transitions: P( A2 = 3 | A1 = 1)


A Problem with Idioms

Machine translation system: Je vois un chat

Model of translation


A Problem with Idioms

Machine translation system: Je vois un chat

Model of translation

I see a spade


A Problem with Idioms Sentence-aligned parallel corpus: ... appelez un chat un chat ...

... ... call a spade a spade

Machine translation system: Je vois un chat

Model of translation

I see a spade


A Generative Phrase Alignment Model Process for generating a sentence pair:


A Generative Phrase Alignment Model Process for generating a sentence pair: Choose number of phrase pairs


A Generative Phrase Alignment Model Process for generating a sentence pair: Choose number of phrase pairs Generate each phrase pair

日本 冻结

提供 援助

向 俄

Japan to freeze

aid

to Russia


A Generative Phrase Alignment Model Process for generating a sentence pair: Choose number of phrase pairs Generate each phrase pair Keep the English order

日本 冻结

提供 援助

向 俄

Japan to freeze

aid

to Russia

Japan to freeze

aid to Russia


A Generative Phrase Alignment Model Process for generating a sentence pair: Choose number of phrase pairs Generate each phrase pair Keep the English order Reorder the Chinese phrases

日本 冻结

提供 援助

向 俄

Japan to freeze

aid

to Russia

Japan to freeze

aid to Russia


A Generative Phrase Alignment Model Process for generating a sentence pair: Choose number of phrase pairs Generate each phrase pair Keep the English order Reorder the Chinese phrases

日本 冻结

提供 援助

向 俄

Japan to freeze

aid

to Russia

日本 冻结

向 俄

Japan to freeze

提供 援助

aid to Russia


A Generative Phrase Alignment Model Process for generating a sentence pair: Choose number of phrase pairs Generate each phrase pair Keep the English order Reorder the Chinese phrases

日本 冻结

提供 援助

向 俄

Japan to freeze

aid

to Russia

日本 冻结

向 俄

Japan to freeze

提供 援助

aid to Russia 日本 冻结 向 俄 提供 援助

A phrase-aligned sentence pair (a list of phrase pairs and a permutation)


GLT Test