Curs 11: Aşteptări şi satisfacerea lor în parsarea discursului Dan Cristea Selecţie de sliduri.

33
Curs 11: Aşteptări şi satisfacerea lor în parsarea discursului Dan Cristea Selecţie de sliduri

Transcript of Curs 11: Aşteptări şi satisfacerea lor în parsarea discursului Dan Cristea Selecţie de sliduri.

Curs 11: Aşteptări şi satisfacerea lor în parsarea discursului

Dan Cristea

Selecţie de sliduri

Expectations and Incremental Discourse Parsing

Cristea and Webber, EACL-1998Cristea et al, Venice-2003

Cristea et al, CICLING-2005

Incremental discourse parsing

The principle of sequenciality

• A left to right reading of the terminal frontier of the tree associated with a discourse must correspond to the span of text it analyses in the same left-to-right order.

6

Incremental discourse parsing - a TAG inspired approach

Adjoining to the right frontier

a1

0

0

1

a

a*

’a

a1

7

(Polanyi, 1988)

Substitution in case of free expectations

k+1

k. Although Bill would have wanted it,

k

k+1. John sold his bicycle to somebody else.

k

k+1

Expectations-driven incremental parsing

a. Clinton is bound to win the elections.

b. He is a natural born campaigner.

c. If you hold some position on an issue,

d. then if Clinton wants to get your vote,

e. he will assure you with great sincerity that he holds that position too.

a

b

EVIDENCE

c

d e

EVIDENCE

ANT-CONS

ANT-CONS

8

(Cristea and Webber, 1998)

a. Clinton is bound to win the elections.

a

b

EVIDENCE

*

b. He is a natural born campaigner.

9

Expectations-driven incremental parsing

a. Clinton is bound to win the elections.b. He is a natural born campaigner.

a b

EVIDENCE

c. If you hold some position on an issue,

c

EVIDENCE

*

10

Expectations-driven incremental parsing

a. Clinton is bound to win the elections.b. He is a natural born campaigner.

c. If you hold some position on an issue,

a

b

EVIDENCE

c

EVIDENCE

11

Expectations-driven incremental parsing

a. Clinton is bound to win the elections.b. He is a natural born campaigner.c. If you hold some position on an issue,

a

b

EVIDENCE

EVIDENCE

c

ANT-CONS

?

12

Expectations-driven incremental parsing

a. Clinton is bound to win the elections.b. He is a natural born campaigner.

a b

EVIDENCE

c. If you hold some position on an issue,

EVIDENCE

*c

ANT-CONS

?

13

Expectations-driven incremental parsing

a. Clinton is bound to win the elections.b. He is a natural born campaigner.c. If you hold some position on an issue,

a

b

EVIDENCE

EVIDENCE

c

ANT-CONS

?

d. he will assure you with great sincerity that he holds that position too.

14

Expectations-driven incremental parsing

a. Clinton is bound to win the elections.

b. He is a natural born campaigner.

c. If you hold some position on an issue,

d. he will assure you with great sincerity that he holds that position too.

a

b

EVIDENCE

EVIDENCE

c

ANT-CONS

d

15

Expectations-driven incremental parsing

a. Clinton is bound to win the elections.

b. He is a natural born campaigner.

c. If you hold some position on an issue,d. then if Clinton wants to get your vote,

d

ANT-CONS

?

a

b

EVIDENCE

EVIDENCE

c

ANT-CONS

?

16

Expectations-driven incremental parsing

a. Clinton is bound to win the elections.

b. He is a natural born campaigner.

c. If you hold some position on an issue,d. then if Clinton wants to get your vote,e. he will assure you with great sincerity that he holds that position too.

d

ANT-CONS

a

b

EVIDENCE

EVIDENCE

c

ANT-CONS

?

17

Expectations-driven incremental parsing

a. Clinton is bound to win the elections.

b. He is a natural born campaigner.

c. If you hold some position on an issue,d. then if Clinton wants to get your vote,e. he will assure you with great sincerity that he holds that position too.

d

ANT-CONS

a

b

EVIDENCE

EVIDENCE

c

ANT-CONS

e

18

Expectations-driven incremental parsing

What can cue-phrases tell us about structure?

[Because John is such a generous man 1] [– whenever he is asked for money, 2] [he will give whatever he has, for example 3] [– he deserves the “Citizen of the year” award. 4]

Reproduced from (Cristea and Webber,1998)

1 2 3 4

because -, -1 2 3 4

1 2 3

1 2

1 3 42

1 32

1 42 3

because <something>, <something>

(Marcu, 1997, 2000; Cristea et al., 2003, 2005)

What can cue-phrases tell us about structure?

[Because John is such a generous man 1] [– whenever he is asked for money, 2] [he will give whatever he has, for example 3] [– he deserves the “Citizen of the year” award. 4]

2 3

whenever -, -2 3 4

2 3

2 43

whenever <something>, <something>

What can cue-phrases tell us about structure?

[Because John is such a generous man 1] [– whenever he is asked for money, 2] [he will give whatever he has, for example 3] [– he deserves the “Citizen of the year” award. 4]

1 2 3

-, - for example1 2 3

2 3

1 32

<something>, <something> for example

What can cue-phrases tell us about structure?

[Because John is such a generous man 1] [– whenever he is asked for money, 2] [he will give whatever he has, for example 3] [– he deserves the “Citizen of the year” award. 4]

There are only two trees that can be obtained after considering all constraints:

1

-, - for example

2 3

whenever -, -

4

because -, -

1

2 3

4

because -, -

-, - for example

whenever -, -

because < >, < >1 2 3 4

whenever < >, < >2 3

< >, < > for example1 2 3

because < >, < >1 2 3 4

whenever < >, < >2 3

< >, < > for example2 3

4

The incremental generation of the first interpretation

[Because John is such a generous man 1]

1

because -, -

[– whenever he is asked for money, 2]

2

whenever -, -

?

*

1

?

because -, -

[Because John is such a generous man 1]

[– whenever he is asked for money, 2]

2

whenever -, -3

[he will give whatever he has, for example 3]

The incremental generation of the first interpretation

1

?

3

because -, -

[Because John is such a generous man 1]

[– whenever he is asked for money, 2]

2

whenever -, -

4

[he will give whatever he has, for example 3][– he deserves the “Citizen of the year” award. 4]

The incremental generation of the first interpretation

1

?

3

4

because -, -

[Because John is such a generous man 1]

[– whenever he is asked for money, 2]

2

whenever -, -

[he will give whatever he has, for example 3][– he deserves the “Citizen of the year” award. 4]

The incremental generation of the first interpretation

[Because John is such a generous man 1]

1

because -, -

2

whenever -, -

[– whenever he is asked for money, 2]

The incremental generation of the second interpretation

[Because John is such a generous man 1]

1

because -, -

2

whenever -, -

[– whenever he is asked for money, 2][he will give whatever he has, for example 3]

3

-, - for example

*

The incremental generation of the second interpretation

[Because John is such a generous man 1]

[– whenever he is asked for money, 2][he will give whatever he has, for example 3]

1

because -, -

2

whenever -, -

3

-, - for example

[– he deserves the “Citizen of the year” award. 4]

4

The incremental generation of the second interpretation

[Because John is such a generous man 1]

[– whenever he is asked for money, 2][he will give whatever he has, for example 3]

1

because -, -

2

whenever -, -

3

-, - for example 4

[– he deserves the “Citizen of the year” award. 4]

The incremental generation of the second interpretation

[Because John is such a generous man 1] [– whenever he is asked for money, 2] [he will give whatever he has, for example 3] [– he deserves the “Citizen of the year” award. 4]

1

-, - for example

2 3

whenever -, -

4

because -, -

1

2 3

4

because -, -

-, - for example

whenever -, -

How can references help in discovering the structure?

DEA=1 4

DEA=1 2 3DEA=1 2

DEA=2 4

DEA=2 3DEA=1 2

wrong

right

How can references help in discovering the structure?

a. Because Mary was upset,

b. even if John agreed,

c. they didn’t speak to one another for several days.

c

a b V=(a)cDEA=ac

a

b c

V=(a)(b)cDEA=abcrightwrong

Incremental parallel processing

NP-chunker AR-engine

segmentator edts-builder

disc. parser summarizer

(Cristea et al., 2005)

VT guides an incremental discourse parsing

The tree resulted after parsing is the one which manifests:– the more natural overall references over the

discourse structure – the smoothest overall CT transitions on veins

(Cristea, 2000; Cristea et al., 2003, 2005)

coherence

cohesionreference score

transitions score

The discourse parser implements a „beam search“

N

N

trees observing markers‘ well-

formedness constraints