Upload
kerry-williamson
View
223
Download
2
Tags:
Embed Size (px)
Citation preview
Borovets, sept. 2003
Discourse theories and technologies
Dan Cristea
“Al. I. Cuza” University of Iasi,
Faculty of Computer Science
and
Romanian Academy,
Institute of Theoretical Computer Science
Borovets, sept. 2003
ContentIntroduction
– What is discourse? Text and discourse. Coherence and cohesion.
Theories– attentional state theory– rhetorical structure theory– centering theory– veins theory
Technologies– segmentation of discourse– Marcu’s parser – VT parser
Related issues on discourse – anaphora resolution, summarisation, information extraction
Borovets, sept. 2003
What is discourse?
Longman: 1. a serious speech or piece or writing on a particular subject: Professor Grant delivered a long discourse on aspects of moral theology. 2. serious conversation between people: You can’t expect meaningful discourse when you two disagree so violently. 3. the language used in particular kinds of speech or writing: scientific discourse.
Borovets, sept. 2003
Text and discourse
Syntactically – a discourse is more than a single sentence.
A text is not a discourse!
But it becomes a discourse the very moment it is read or listen by a human... or a machine.
Borovets, sept. 2003
Time and discourse
Discourse has a dynamic nature
Time axesreal time
discourse time
story time
1 2
2 11000 1030800 920
1 2
Borovets, sept. 2003
Cohesion and coherence
A text has cohesion when its parts closely correlate.
A text is coherent when it makes sense, with respect to an accepted setting, real or virtual.
Borovets, sept. 2003
Interpretation of discourse
discourse interpretation
text knowledge base
knowledge about the language
knowledge about the world
knowledge about the situation
knowledge about the author
Borovets, sept. 2003
Discourse phenomena: interruptions and flash-backs
E: Now attach the pull rope to the top of the engine.
By the way, did you buy gasoline today?
A: Yes. I got some when I bought the new lawnmower wheel.
I forgot to take the gas with me, so I bought a new one.
E: Did it cost much?
A: No, and we could use another anyway to keep with the tractor.
E: OK, how far have you got?
Did you get it attached?
from [Allen, 1987]
Borovets, sept. 2003
Discourse phenomena: pop-overs
E: Now attach the pull rope to the top of the engine.
By the way, did you buy gasoline today?
A: Yes. I got some when I bought the new lawnmower wheel.
I forgot to take the gas with me, so I bought a new one.
E: Did it cost much?
A: No, and we could use another anyway to keep with the tractor.
E: OK, how far have you got?
Did you get it attached?
from [Allen, 1987]
Borovets, sept. 2003
Discourse phenomena: inference load and pronoun use
Why is it that some discourses seem more difficult to understand than others?
Why do we use the pronouns as we do?
Borovets, sept. 2003
Discourse theories?
Sub-domain of Computational Linguistics: searching for the laws that govern the discourse and the models making possible an automated analysis, representation and generation of the discourse.
Borovets, sept. 2003
Discourse theories
• atentional state theory• rhetorical state theory• centering theory• veins theory
Borovets, sept. 2003
Attentional state theory (AST)Barbara Grosz & Candence Sidner, 1987
Models the linguistic structure of discourse
Gives an account on intentions and how are they combined
Explains the shift of attention during discourse interpretation and referentiality in terms of discourse structure
3 components
Borovets, sept. 2003
AST: 1st component
• a linguistic structure: – more sentences are aggregated in the same
segment– segments display a recursive structure
Borovets, sept. 2003
AST: 2nd component• an intentional structure:
– a segment communicates an intention, it has a goal to accomplish in the reader;
– the goals of the component segments contribute to the realisation of the goal of the overall segment;
– two type of relations between segment goals: dominance and satisfaction-precedence
Borovets, sept. 2003
AST: 2nd componentRelations: dominance
DSP A dominates DSP AA: the intention associated with DSP AA contributes to the satisfaction of the intention associated with DSP A
A
AA AB AC
AAA AAB ABA ABB
Borovets, sept. 2003
AST: 2nd componentRelations: satisfaction-precedenceDSP AA satisfaction-precedes DSP AB: DSP AA must be satisfied before
DSP AB
A
AA AB AC
AAA AAB ABA ABB
Borovets, sept. 2003
AST: 3rd component• an attentional state
– to each segment corresponds a space of entities under focus
– these spaces have the dynamics of a stack
A
AA AB AC
AAA AAB ABA ABB
A
SA
Borovets, sept. 2003
AST: 3rd component• an attentional state
– to each segment corresponds a space of entities under focus
– these spaces have the dynamics of a stack
A
AB AC
AAA AAB ABA ABB
A
SA
AASAA
Borovets, sept. 2003
AST: 3rd component• an attentional state
– to each segment corresponds a space of entities under focus
– these spaces have the dynamics of a stack
A
AB AC
AAB ABA ABB
A
SA
AASAA
AAA
SAAA
Borovets, sept. 2003
AST: 3rd component• an attentional state
– to each segment corresponds a space of entities under focus
– these spaces have the dynamics of a stack
A
AB AC
ABA ABB
A
SA
AASAA
AAA AAB
SAAB
Borovets, sept. 2003
AST: 3rd component• an attentional state
– to each segment corresponds a space of entities under focus
– these spaces have the dynamics of a stack
A
AB AC
ABA ABB
A
SA
AASAB
AAA AAB
Borovets, sept. 2003
AST: 3rd component• an attentional state
– to each segment corresponds a space of entities under focus
– these spaces have the dynamics of a stack
A
AB AC
ABA ABB
A
SA
AASAB
AAA AAB
SABA
Borovets, sept. 2003
AST: 3rd component• an attentional state
– to each segment corresponds a space of entities under focus
– these spaces have the dynamics of a stack
A
AB AC
ABA ABB
A
SA
AASAB
AAA AAB
SABB
Borovets, sept. 2003
AST: 3rd component• an attentional state
– to each segment corresponds a space of entities under focus
– these spaces have the dynamics of a stack
A
AB AC
ABA ABB
A
SA
AASAC
AAA AAB
Borovets, sept. 2003
AST: 3rd component• an attentional state
– accessibility modeled by the top-down access in the stack
A
AA AB AC
AAA AAB ABA ABB SA
SAB
SABB
Borovets, sept. 2003
AST: pluses
• Discourse structure: – a proposal for discourse structure (an example)– stack behavior models hierarchical relationships
among text segments
• Reference: accounted for by accessibility in stack
• Interruptions
• Flash-backs
Borovets, sept. 2003
AST: interruptionsE: Now attach the pull rope to the top of the engine.
By the way, did you buy gasoline today?
A: Yes. I got some when I bought the new lawnmower wheel.
I forgot to take the gas with me, so I bought a new one.
E: Did it cost much?
A: No, and we could use another anyway to keep with the tractor.
E: OK, how far have you got?
Did you get it attached?
from [Allen, 1987]
An interruption is a discourse segment whose DSP is not dominated nor satisfaction-preceded by the DSP of the immediately proceeding segment.
Borovets, sept. 2003
AST: interruptionsE: Now attach the pull rope to the top of the engine.
…
…
…
…
By the way, did you buy gasoline today?
A: Yes. I got some when I bought the new lawnmower wheel.I forgot to take the gas with me, so I bought a new one.
E: Did it cost much?
A: No, and we could use another anyway to keep with the tractor.
E: OK, how far have you got?
Did you get it attached?
Borovets, sept. 2003
AST: flashbacksSinit …
SABC …
SBill OK. Now how do I say that Bill is...Whoops I forgot about ABC.I need an individual concept for the company ABC. …
SBill Now back to Bill.How do I say that Bill is an employee of ABC?
From [Grosz & Sidner, 1987]
A flashback is a particular kind of interruption whose DSP satisfaction-precedes the interrupted segment or a segment that dominates the interrupted segment.
SBill
Sinit
SFB
Sinit
SABC
SBill
SFB
SBill
SFB
Sinit
SABC SBill
Sinit
SFB
Borovets, sept. 2003
AST: minuses
Stack mechanism fails for certain dominant/dominated segment configurations when granularity is sufficiently fine
Does not accommodate left satellites
Borovets, sept. 2003
AST doesn‘t accommodate left satellites
a. Jack and Sue went to buy a new lawn mowerb. since their old one was stolen.c. Sue had seen the men who took it andd. had chased them down the street,e. but they'd driven away in a truck. f. After looking in the store g. they realized they couldn't afford a new one.h. By the way, Jack lost his job last monthi. so he's been short of cash recently.j. He has been looking for a new one,k. but so far hasn't had any luck. l. Anyway, they finally found a used one at a garage sale.
Allen, 1993
Borovets, sept. 2003
AST doesn‘t accommodate left satellites
a. Jack and Sue went to buy a new lawn mower b. since their old one was stolen.
f. After looking in the store
g. they realized they couldn't afford a new one.
l. Anyway, they finally found a used one at a garage sale.
c. Sue had seen the men who took it and
d. had chased them down the street,
e. but they'd driven away in a truck.
h. By the way, Jack lost his job last month
i. so he's been short of cash recently. j. He has been looking for a new one, k. but so far hasn't had any luck.
Borovets, sept. 2003
Attentional state stack
a
a. Jack and Sue went to buy a new lawn mower
a,b
a. Jack and Sue went to buy a new lawn mower b. since their old one was stolen.
c,d,e
a. Jack and Sue went to buy a new lawn mower b. since their old one was stolen.
c. Sue had seen the men who took it and
d. had chased them down the street,
e. but they'd driven away in a truck.
a. Jack and Sue went to buy a new lawn mower b. since their old one was stolen.
f. After looking in the storeg. they realized they couldn't afford a new one.
c. Sue had seen the men who took it and
d. had chased them down the street,
e. but they'd driven away in a truck.
a,b,f,g
a. Jack and Sue went to buy a new lawn mower b. since their old one was stolen.
f. After looking in the storeg. they realized they couldn't afford a new one.
c. Sue had seen the men who took it and
d. had chased them down the street,
e. but they'd driven away in a truck.
h. By the way, Jack lost his job last month
i. so he's been short of cash recently. j. He has been looking for a new one, k. but so far hasn't had any luck.
h,i,j,k
a,b,f,g,l
a. Jack and Sue went to buy a new lawn mower b. since their old one was stolen.
f. After looking in the storeg. they realized they couldn't afford a new one.
l. Anyway, they finally found a used one at a garage sale.
c. Sue had seen the men who took it and
d. had chased them down the street,
e. but they'd driven away in a truck.
h. By the way, Jack lost his job last month
i. so he's been short of cash recently. j. He has been looking for a new one, k. but so far hasn't had any luck.
Borovets, sept. 2003
Problem:a finer granularity
f. After looking in the store
a. Jack and Sue went to buy a new lawn mower
g. they realized they couldn't afford a new one.
l. Anyway, they finally found a used one at a garage sale.
b. since their old one was stolen.
Borovets, sept. 2003
Problem
a a a
b
a
b
c,d,e
a, g
f
Borovets, sept. 2003
Rhetorical structure theory
Basics• text span: un uninterrupted linear interval of text• relation: holds between two non-overlapping spans, called
nucleus and satellite– a nucleus is more important than a satellite (deletion and
substitution tests)– relations: hypotactic (nucleus + satellite) and paratactic (2 nuclei)
• scheme: integrates by a relation two or more text spans (like grammar rules)
• RST analysis are trees• they reflect a judge interpretation (therefore could be
subjective)
William Mann & Sandra Thompson, 1987
Borovets, sept. 2003
RST schemes
relation
text span: nucleus
text span: satellite
relation
text span: nucleus
text span: nucleus
Borovets, sept. 2003
RST schemes: equivalences
relation1 relation2
relation1
relation2
relation1
relation2
Borovets, sept. 2003
RST schemes: equivalences
relation relation
relation
relation
relation
relation
Borovets, sept. 2003
EVIDENCE relation
1. The program as published for calendar year 1980 really works.
2. In only a few minutes, I entered all the figures from my 1980 tax return
3. and got a result which agreed with my hand calculations to the penny.
EVIDENCE
1-3
2-31
EVIDENCE
constraint on N: R might not believe N to a degree satisfactory to W
constraint on S: R believes S or finds it credible
effect: R’s belief of N is increased
Borovets, sept. 2003
CONCESSION relationCONCESSION
constraint on N: W has positive regard to the situation presented in N
constraint on S: W is not claiming that the situation presented in S doesn’t hold
constraint on the combination N+S: W acknowledges a potential incompatibility between the situations presented in N and S; W regards the situation presented in N and S as compatible
effect: R’s positive regard for the situation presented in N is increased
1. Although Dioxin is toxic to certain animals,2. evidence is lacking that it has any serious
long-term effects on human beings.
CONCESSION
1-2
21
Borovets, sept. 2003
CIRCUMSTANCE relationCIRCUMSTANCE
constraint on N: none
constraint on S: S presents a situation
constraint on the combination N+S: S sets a framework (spatial or temporal) within which R is intended to interpret the situation presented in N
effect: R recognizes that the situation presented in S provides the framework for interpreting N
1. Probably the most extreme case of Visitors Fever I ever witnessed was a few summers ago
2. when I visited relatives in Midwest.
CIRCUMSTANCE
1-2
21
Borovets, sept. 2003
A more complex example1. Farmington Police had to help control traffic recently
2. when hundreds of people lined up to be among the first applying for jobs at the yet-to-open Marriot Hotel.
3. The hotel’s help-wanted announcement – for 300 openings – was a rare opportunity for many unemployed.
4. The people waiting in line carried a message of claims that the jobless could be employed if only they showed enough moxie.
5. Every rule has exceptions,
6. but the tragic and too-common tableaux of hundreds of people snake-lining up for any task with a paycheck illustrates a lack of jobs,
7 not laziness.
circumstance
32
2-3
volitional result
1-3
4
evidence
5
6
antithesis
7
6-7
concession
5-7
4-7
background
1-7
Borovets, sept. 2003
RST relationsSubject matter
(informational)
ElaborationCircumstanceSolutionhoodVolitional CauseVolitional ResultNon-Volitional CauseNon-Volitional ResultPurposeConditionOtherwiseInterpretationEvaluationRestatementSummarySequenceContrast
Presentational (intentional)
Motivation
Antithesis
Background
Enablement
Evidence
Justify
Concession
Borovets, sept. 2003
Problem: multiple interpretations[Moore & Polack, 1992]
1
2
motivation
3
motivation
Intentional level
3
1
condition
2
condition
Motivational level
1. Come back at 5:00.2. Then we can go to the hardware store before it closes.3. This way we can finish the bookshelves tonight.
Borovets, sept. 2003
Any other complains?
• no indication on referentiality
• how many relations?
• how relations are discovered?
• ...
Borovets, sept. 2003
How distant are AST & RST?
• Mosser&Moore (1996) and Marcu (1997):– granularity: AST - undefined, RST - fine
(clause level)– structure: trees– internal nodes: relations (AST:2, RST: 28,
Hobbs, Knott: hierarchy of relations)
Borovets, sept. 2003
Centering - a theory of local discourse coherence• Joshi,A.K. and Weinstein,S., 1981: “Control of Inference: Role
of Some Aspects of Discourse-Structure Centering“• Grosz,B.; Joshi,A.K. and Weinstein,S., 1986: “Towards a
computational theory of discourse interpretation”• Brennan,S.E.; Friedman,M.W. and Pollard,C.J., 1987: “A
Centering approach to pronouns“ • Grosz,B.; Joshi,A.K. and Weinstein,S, 1995: “Centering: A
framework for modeling the local coherence of discourse”
• Strube,M. and Hahn,U., 1996: “Functional Centering“• Walker,M.A.; Joshi,A.K. and Prince,E.F. (eds.), 1997:
“Centering in Discourse“• Kameyama,M., 1997: “Intrasentential Centering: A Case
Study“
Borovets, sept. 2003
Goals of the theory
• explains why certain texts are more difficult to process than others
• explains why we use the pronouns as we use them
• anchors a practical approach for anaphora resolution
Borovets, sept. 2003
A smooth discourse
from [Walker, Joshi and Prince, 1997]
a. Jeff1 helped Dick2 wash the car.
b. He1 washed the windows as Dick2 waxed the car.
c. He1 soaped a pane.
He in c. is Jeff because soaping can only be related to
the washing event
Borovets, sept. 2003
A ... smooth discourse
a. Jeff1 helped Dick2 wash the car.
b. He1 washed the windows as Dick2 waxed the car.
c. He2 buffed the hood.
He in c. is Dick, because buffing can only be related to the waxing event.
less
Borovets, sept. 2003
Focus helps to disambiguate pronominal anaphora
1. Susan1 is a fine friend.2. She1 gives people the most wonderful
presents.3. She1 just gave Betsy2 a wonderful bottle of
wine.4. She1 told her2 it was quite rare.5. She1 knows a lot about wine.
from Grosz, Joshi&Weinstein, 1995
Borovets, sept. 2003
CT explains the normality or oddness of some utterances
a. Terry really goofs sometimes. b. Yesterday was a beautiful day and he was
excited about trying out his new sailboat. c. He wanted Tony to join on a sailing
expedition. d. He called him at 6 A.M.e. He was sick and furious at being woken up
so early. from Grosz, Joshi&Weinstein, 1995
Borovets, sept. 2003
This text repairs the proceeding, but...a. Terry really goofs sometimes. b. Yesterday was a beautiful day and he was excited
about trying out his new sailboat. c. He wanted Tony to join on a sailing expedition. d. He called him at 6 A.M.e. Tony was sick and furious at being woken up so
early. f. He told Terry to get lost and hung up.g. Of, course he hadn’t intended to upset Tony.
Borovets, sept. 2003
Finally, we are done!
a. Terry really goofs sometimes. b. Yesterday was a beautiful day and he was excited
about trying out his new sailboat. c. He wanted Tony to join on a sailing expedition. d. He called him at 6 A.M.e. Tony was sick and furious at being woken up so
early. f. He told Terry to get lost and hung up.g. Of, course Terry hadn’t intended to upset Tony.
Borovets, sept. 2003
The main lines
• Applies to one segment of discourse– refers to Grosz&Sidner´s Attentional State Theory– Hahn&Strube applied centering for finding
discourse segments
• Sees the segment made of adjacent utterances (sentences)
• Kameyama, proposes intrasentential centering (1997)
Borovets, sept. 2003
Centers
John gave Mary a flower.
person1type = personname = Johngender = masc
person2 type = personname = Marygender = fem
flower1type = flowernumber = sg
the realisation relation
Borovets, sept. 2003
What is a center?
John went to Mary‘s house
person1 type = personname = Marygender = fem
house1type = housenumber = sg
the realisation relation
He met her down the street.
He found it down the street.
Borovets, sept. 2003
For each utterance, compute:
- a list of forward-looking centers
Cf(u) = <e1, e2, ... ek>
- a backward-looking center
Cb(u) = ei
- a prefered center
Cp(u) = e1
Borovets, sept. 2003
Rule 1: pronoun realisation
• If some element of Cf(ui-1) is realised as a pronoun in ui, than so is Cb(ui)– it captures the intuition that pronominalisation is
one way to indicate discourse salience– if there are multiple pronouns in a sentence,
realising discourse entities from the previous utterance, than Cb must be one of them
– if there is just one pronoun, then the pronoun must be the Cb
Borovets, sept. 2003
Rule 1 obeyeda. Terry really goofs sometimes.
Cf = ([Terry])b. Yesterday was a beatiful day and he was excited
about trying out his new sailboat.
Cf = (he=his=[Terry], [the sailboat])
Cb = [Terry]c. He wanted Tony to join on a sailing expedition.
Cf = (he=[Terry], [Tony], [the expedition])
Cb = [Terry]d. He called him at 6 A.M.
Cf = (he=[Terry], him=[Tony])
Cb = [Terry]
Borovets, sept. 2003
Rule 1 still obeyeda. Terry really goofs sometimes.
Cf = ([Terry])b. Yesterday was a beatiful day and he was excited
about trying out his new sailboat.
Cf = (he=his=[Terry], [the sailboat])
Cb = [Terry]c. He wanted Tony to join on a sailing expedition.
Cf = (he=[Terry], [Tony], [the expedition])
Cb = [Terry]d. He called Tony at 6 A.M.
Cf = (he=[Terry], [Tony])
Cb = [Terry]
Borovets, sept. 2003
Rule 1 disobeyeda. Terry really goofs sometimes.
Cf = ([Terry])b. Yesterday was a beatiful day and he was excited
about trying out his new sailboat.
Cf = (he=his=[Terry], [the sailboat])
Cb = [Terry]c. He wanted Tony to join on a sailing expedition.
Cf = (he=[Terry], [Tony], [the expedition])
Cb = [Terry]d. Terry called him at 6 A.M.
Cf = ([Terry], him=[Tony])
Cb = [Terry]
Borovets, sept. 2003
Rule 2: transitions
Cb(u) = Cb(u-1) Cb(u) Cb(u-1)
Cb(u) = Cp(u)
Cb(u) Cp(u)
CONTINUING SMOOTH SHIFT
RETAINING ABRUPT SHIFT
CON > RET > SSH > ASH
Borovets, sept. 2003
Rule 2 disobeyeda. Terry really goofs sometimes.
Cf = ([Terry])b. Yesterday was a beatiful day and he was excited about trying
out his new sailboat.
Cf = (he=his=[Terry], [the sailboat])
Cb = [Terry]c. He wanted Tony to join on a sailing expedition.
Cf = (he=[Terry], [Tony], [the expedition])
Cb = [Terry]d. He called Tony at 6 A.M.
Cf = (he=[Terry], [Tony])
Cb = [Terry]e. He was sick and furious at being woken up so early.
Cf = (he=[Tony])
Cb = [Tony]
CONTINUING
CONTINUING
CONTINUING
SMOOTH SHIFT
Borovets, sept. 2003
Rule 1 disobeyed again...
d. He called Tony at 6 A.M.
Cf = (he=[Terry], [Tony])
Cb = [Terry]
e. Tony was sick and furious at being woken up so early.
Cf = ([Tony])
Cb = [Tony]
f. He told Terry to get lost and hung up.
Cf = (he=[Tony], [Terry])
Cb = [Tony]
g. Of, course he hadn’t intended to upset Tony.
Cf = (he=[Terry], [Tony])
Cb = [Tony]
CONTINUING
SMOOTH SHIFT
CONTINUING
RETAINING
Borovets, sept. 2003
The discourse repaired...d. He called Tony at 6 A.M.
Cf = (he=[Terry], [Tony])
Cb = [Terry]e. Tony was sick and furious at being woken up so early.
Cf = ([Tony])
Cb = [Tony]f. He told Terry to get lost and hung up.
Cf = (he=[Tony], [Terry])
Cb = [Tony]g. Of, course Terry hadn’t intended to upset Tony.
Cf = ([Terry], [Tony])
Cb = [Tony]
CONTINUING
SMOOTH SHIFT
CONTINUING
RETAINING
Borovets, sept. 2003
Centering hints on pronominal anaphoraa. I haven’t seen Jeff for several days.
b. Carl thinks he’s studying for his exams.
c. I think he? went to the Cape with Linda.
from [Grosz, Joshi & Weinstein, 1983]
Cf = (I=[I], [Jeff])
Cb = [I]
Cf = ([Carl], he=[Jeff], [Jeff´s exams])
Cb = [Jeff]
Borovets, sept. 2003
Centering explains why we understand he in unit c as Jeff
b. Carl thinks he’s studying for his exams.
c. I think he? went to the Cape with Linda.
Cf = ([Carl], he=[Jeff], [Jeff´s exams])
Cb = [Jeff]
RETAINING
ABRUPT SHIFT
Cf = (I=[I], he=[Jeff], [the Cape], [Linda])
Cb = [Jeff]Cf = (I=[I], he=[Carl], [the Cape], [Linda])
Cb = [Carl]
Jeff
Borovets, sept. 2003
Centering: any problems?
• applies inside a segment– BUT what a segment is?
• allignment of Cf ellements– on what criteria (surface-order, syntactic role,
functional [Strube&Hahn, 1996]) language dependant
• null pronouns – Italian, Japanese, Romanian
parameterised centering [Poesio et al., 2000]
Borovets, sept. 2003
Veins theory (VT)
• to offer a better understanding of the relationship between discourse structure and referentiality
• to accommodate the left satellite problem• to be independent of granularity• to generalise Centering from local to global• to allow summarization and, especially,
focused summaries• to assist anaphora resolution processes• to guide discourse parsing
Cristea, Ide & Romary, 1998
Borovets, sept. 2003
VT intuitions
A satellite or a nucleus can refer a nuclear sibling on the left: in combinations u1
n R u2s, or u1
n R u2n, u2
can refer u1.
Ex. 11. John left without an umbrella 2. although he watched the TV meteo anouncing raining.
Borovets, sept. 2003
VT intuitions
A nucleus can refer a left satellite: in combinations u1
s R u2n, u2 can refer u1.
Ex. 2
1. Although John saw the TV whether forecast announcing raining
2. he left home without an umbrella.
Borovets, sept. 2003
VT intuitionsA right satellite of a nucleus u is not accessible
from another right sibling of u, nuclear or satellite: in combinations (u1
n R1 u2s)n R2 u3
n or (u1n R1 u2
s)n R2 u3s, u3
can refer u1 but not u2.
Ex. 31. John told Mary that he loves her.2. He was never married 3. and lived until 40 with his mother. 4. She, on the contrary, was married twice.
Borovets, sept. 2003
VT intuitionsA nucleus blocks the accessibility from a right
satellite towards a left satellite of it: in combinations (u1
s R1 u2n)n R2 u3
s, u3 can refer u2
but not u1.
Ex. 41. With one year before finishing his mandate of president of the company2. Mr. W. Ross has begun to manipulate its bankrupt. *3. Gossips claimed that he has obtained it by fraud.
Borovets, sept. 2003
VT intuitionsPossible repairs of Ex. 4: Ex. 5
1. Mr. W. Ross has begun to manipulate the bankrupt of his company2. with one year before finishing his mandate of president.3. Gossips claimed that he has obtained it by fraud. Ex. 61. With one year before finishing his mandate of president of the company2. Mr. W. Ross has begun to manipulate its bankrupt. 3. Gossips claimed that he has been elected by fraud.
Borovets, sept. 2003
VT´s basics
• a discourse structure can be represented by a (binary) tree
• a terminal node (unit) – a span of text, syntactically a sentence or clause
• an intermediate level node (relation) – a span of text that has a structure of its own
• nodes are polarised: nuclei (most important) and satellites (less important)
• relation names are ignored
Borovets, sept. 2003
1 2 3 4
5
6 7 8
9
10
11 12
13-??
??-??
H = 1 9 *V = 1 9 *
H = 1V = 1 9 *
H = 9V = 1 9 *
H = 1V = 1 9 *
H = 5V = 1 5 9 *
H = 1V = 1 9 *
H = 3V = 1 3 5 9 *
H = 6 7V = 1 5 6 7 9 *
H = 9V = 1 9 *
H = 9V = 1 9 *
H = 9V = 1 (8) 9 *
H = 10V = 1 9 10 *
H = 11V = 1 9 10 11 *H = 3
V = 1 3 5 9DRA = 1 3 H = 9
V = 1 (8) 9DRA = 1 8 9
Trees as in RST
1
4
2 3
relations
units
nuclear
Borovets, sept. 2003
The definitions of VT: heads
• Head expression of a node: the sequence of the most important units within the corresponding span of text:– the head of a terminal node: its label– the head of a non-terminal node: the
concatenation of the head expressions of the nuclear children
• the important units are projected up to the level where the corresponding span is seen as a satellite
Borovets, sept. 2003
Head expressions are computed bottom-up
ba
c
H=aH=b
H=a H=c
H=cH=a c
Borovets, sept. 2003
Veins
to understand a piece of text in the context of the whole discourse one needs the significant units within the span together with other surrounding units
Vein expression of a node: the sequence of units that are required to understand the span of text covered by the node, in the context of the whole discourse
Borovets, sept. 2003
Veins
• Vein expressions are computed top-down starting with the root – the vein expression of the root is its head
expression
Borovets, sept. 2003
Vein expressions
• A nuclear node with no satellites to the left
V=v
V=v
V=v
Borovets, sept. 2003
Vein expressions
• A satellite node on the right (simplified)
V=seq(h, v)
V=v
H=h
Borovets, sept. 2003
Consider the text:
1. John sold his bicycle
2. although Bill would have wanted it.
3. He obtained a good price for it,
4. which Bill could not have afforded.
5. Therefore he decided to use the money to go on a trip.
Borovets, sept. 2003
Computing the veins ofJohn and Bill
H=1 3
H=1
H=3
H=1
H=2H=3
H=4
H=5
H=1 3 5
1 2 3 4
5V=1 3 5
V=1 3 5
V=1 3 5
V=1 3 5
V=1 3 5
V=1 2 3 5
V=1 3 5
V=1 3 5
V=1 3 4 5
Borovets, sept. 2003
Computing the veins ofJohn and Bill
1 2 3 4
5V=1 3 5
V=1 3 5
V=1 2 3 5 V=1 3 5
V=1 3 4 5
Borovets, sept. 2003
The account of VT on cohession• Grounds
– a three layered approach to referential representation
– no need for right reference (cataphora)!– vein expressions are contexts in which to
look for referents
Borovets, sept. 2003
A three layered approach to anaphora
the text layer ………………..……………………….…………….
the semantic layer …………………………..…centera
a b
the projections layer ………………………………………………fsa
RE a projects fsa
fsa proposes centerafsb evokes centera
fsb
RE b projects fsb
Cristea&Dima, 2001; Cristea et al., 2002
Borovets, sept. 2003
A three layered approach to anaphora
1. John sold his bicycle
2. although Bill would have wanted it.
the text layer ………………………………………………………
the semantic layer ……………………...………
it his bicycle
the projections layer ………………………………………………
evokesproposes
no = sgsem=bicycledet = yes
projects
no = sgsem=bicycledet = yes
projects
no = sgsem=¬human
Borovets, sept. 2003
No need for right references!
1. Although Bill would have wanted it,
2. John sold his bicycle to somebody else.
the text layer ………………………………………………………
the semantic layer ……………………………...
it his bicycle
the projections layer ………………………………………………
projects
no = sgsem=¬human
projects
no = sgsem=bicycledet = yes
evokesproposes
no = sgsem=¬humanno = sgsem=bicycledet = yes
Borovets, sept. 2003
Domain of evocative accessibility (DEA)
DEA(u) = pref(u, vein(u))
Reminder! The vein expression of a terminal node (discourse unit): the sequence of units that are required to understand just that unit, in the context of the whole discourse.
(simplified)
Borovets, sept. 2003
Direct referencesIf A and B are units in this textual order, A belongs to the DRA of B
and aA is linearly the most recent (to B) RE that realizes the same center as bB, we say that b directly co-refers the center evoked by a.
text layer ………………………………………………………………………
semantic layer ………..………………………………………………….
A B
centera
a b
a. direct co-reference b. functional direct co-reference
A B
centera
a b
centerb
role
Borovets, sept. 2003
Indirect referencesIf A, B and C are units in this order, bB is linearly the most recent
(to C) RE that realises the same center as cC, B is not on the DRA of C, A is linearly the most recent (to B) unit that is both on the DRA of B and of C, and it contains a RE aA such that bB realises the same center as a, we say that c indirectly co-refers the center realised by a.
A similar definition applies for indirect functional references.
units and DRAs …………………….. A B
text layer…..………………….. …………………
semantic layer……...……………….centera
b
C
c a
Borovets, sept. 2003
Inferential references
We call any reference that is not direct or indirect – a inferential reference: there is no intersection between the backward looking chain of co-references’ units and the DRA of the anaphor's unit.
A particular category of inference references - pragmatic references, or pseudo references: REs that practically do not refer back in the discourse although an entity identically realised was already introduced.
Borovets, sept. 2003
units and their DRAs REs projection restrictions evocative and situations processes post-evocative processes
bicycle1type = bicycleowner = [John]gender = humannum = sg
his bicycle
he
Johntype = personname = Johngender = mascnum = sg
John
Bill
which
pos = def nountype = bicycleowner = [John]gender = humannum = sg
pos = proper nountype = personname = Johngender = mascnum = sg
pos = pron type = humangender = mascnum = sg
it
price
pos = proper nountype = personname = Billgender = mascnum = sg
Billtype = personname = Billgender = mascnum = sg
pos = pron gender = humannum = sg
pos = undef noun type = pricenum = sg
itpos = pron gender = humannum = sg
Bill
pos = proper nountype = personname = Billgender = mascnum = sg
pos = pron
he
pos = pron type = humangender = mascnum = sg
the moneypos = def noun type = moneynum = pl
a trippos = undef noun num = sg
2
1
price1type = pricegender = humannum = sg
money1type = moneynum = pl
trip1type = tripnum = sg
direct references
indirect reference
inferential reference
3
4
5
Borovets, sept. 2003
The first conjecture (of cohesion)
Antecedents of anaphora are found mainly in the DEA given by the vein expression of the unit the anaphor belongs to.
Only if they are not found there, the rest of the units must be searched (Extended-DEA) – inferential:
Unit Vein DEA E-DEA
4 1 3 4 5 1 3 4 2 1 3 4
Borovets, sept. 2003
direct references >
indirect references >
inferential references
“>” = “more frequent”, “easier to process”
The first conjecture (of cohesion)
Borovets, sept. 2003
Experimental results
Source No. of units
Total no. of refs
Direct Indirect Inferential
English 62 97 77.3% 14.4% 8.3%
French 48 110 89.1% 10.0% 0.9%
Romanian 66 111 93.7% 1.8% 4.5%
Total 176 318 87.1% 8.5% 4.4%
Borovets, sept. 2003
Two classes of models
• Linear-k models– the antecedent of an RE can be found in the k
discourse units that immediately precede it.• Linear-0: intra-unit model.• Linear-k: most anaphora and co-reference resolution systems.
• Discourse-k models– the antecedent of an RE can be found in the k
discourse units that hierarchically precede it.• Use Veins Theory to compute the units that hierarchically
precede a unit.
Borovets, sept. 2003
Experiments
• Took 30 texts from the MUC7 corpus (news stories about changes in executive management personnel)
• The texts were annotated for co-reference relations of identity [Hirschman and Chinchor,97]
• The texts were annotated for rhetorical structure [Marcu, Amorrortu, and Romera,99]
• Fused the annotations; computed the veins; computed the DEAs and E-DEAs
Cristea et al., 2000
Borovets, sept. 2003
Experiment 1
• Estimated the potential of Linear-k and Discourse-VT-k models to establish correct co-referential links.– For each k, each re, and each model M
(Linear or VT)• p(M-k,re,EDEAk) =
• p(M-k,Corpus) = re Corpus p(M-k,re,EDEAk)
1, re can be resolved to antecedents in EDEAk
0, otherwise.{
Borovets, sept. 2003
Potentials (2)
70,00%
75,00%
80,00%
85,00%
90,00%
95,00%
0 1 2 3 4 5 6 7 8 9EDRA size
VT-k Linear-k
Borovets, sept. 2003
Direct comparison: VT-k/Linear-k
0,96
0,98
1
1,02
1,04
1,06
1,08
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20
EDRA sizeVT-k/Lin-k
Borovets, sept. 2003
Experiment 2
• Estimated the effort required to establish correct co-referential links when using Linear-k and Discourse-VT-k models.– For each k, each re, and each model M
(Linear or VT)• e(M-k,re,EDEAk) =
• e(M-k,Corpus) = re Corpus e(M-k,re,EDEAk)
d, the distance between re and the closestantecedent in E-DEAk
k, if no such antecedent exists.{
Borovets, sept. 2003
Example1. Michael D. Casey, a top Johnson&Johnson manager, moved to Genetic
Therapy Inc., a small biotechnology concern here,
2. to become its president and chief operating officer.
3. Mr. Casey, 46 years old, was president of J&J's McNeil
Pharmaceutical subsidiary,
4. which was merged with another J&J unit, Ortho Pharmaceutical Corp.,
this year in a cost-cutting move.
5. Mr. Casey succeeds M. James Barrett, 50, as president of Genetic
Therapy.
6. Mr. Barrett remains chief executive officer
7. and becomes chairman.
8. Mr. Casey said
9. he made the move to the smaller company
VT: DEA(9) = <1, 8, 9>
Linear VT-EDEA
K=0 9 9 K=1 9,8 9,8 K=2 9,8,7 9,8,1 K=3 9,8,7,6 9,8,1,7 K=4 9,8,7,6,5 9,8,1,7,6
Borovets, sept. 2003
Example
Michael D. Casey
Genetic Therapy Inc.
10 11 122 3 4
Mr. Casey
Genetic Therapy Inc.
5 6 7
Mr. Casey
81
the smaller company
9
Borovets, sept. 2003
Results
0
1000
2000
3000
4000
5000
6000
7000
8000
1 3 5 7 9 11 13 15 17 19 25 35 45 55 65 75 85 95
EDEA sizeVT Process Lin Process
Borovets, sept. 2003
The account of VT on coherence
• Veins give a natural way to generalize Centering from local to global
Borovets, sept. 2003
1 2 3 4
5 V=1 3 5
V=1 3 5
V=1 2 3 5 V=1 3 5V=1 3 4 5
Vein expressions give „lines of argumentation“
1. John sold his bicycle
1. John sold his bicycle
3. He obtained a good price for it,
5. Therefore he decided to use the money to go on a trip.
Borovets, sept. 2003
1 2 3 4
5 V=1 3 5
V=1 3 5
V=1 2 3 5 V=1 3 5V=1 3 4 5
Lines of argumentation
3. He obtained a good price for it,
1. John sold his bicycle
3. He obtained a good price for it,
5. Therefore he decided to use the money to go on a trip.
Borovets, sept. 2003
1 2 3 4
5 V=1 3 5
V=1 3 5
V=1 2 3 5 V=1 3 5V=1 3 4 5
Lines of argumentation
5. Therefore he decided to use the money to go on a trip.
1. John sold his bicycle
3. He obtained a good price for it,
5. Therefore he decided to use the money to go on a trip.
Borovets, sept. 2003
Evaluating the coherence of a discourse• A smoothness score:
– CONTINUING = 4– RETAINING = 3– SMOOTH SHIFT =2– ABRUPT SHIFT = 1– NO Cb = 0
• A global smoothness score: summing up the score of all units
Borovets, sept. 2003
The second conjecture (on coherence)• The global smoothness score of a
discourse when computed following VT is at least as high as the score computed following CT.
• That is, we claim that long-distance transitions computed using VT are systematically smoother than accidental transitions at segment boundaries.
Borovets, sept. 2003
Transitions and scores on a linear adjacency metricJ = [John], b = [John's bicycle], B = [Bill], p = [price], m = [the money], t = [a trip])
1 2 3 4 5
Cf J, b B, b J, p, b p, B J, m, t
Cb J b b p -
Trans ASH RET SSH No Cb
Score 1 3 2 0
Global 6/4 = 1.5
Borovets, sept. 2003
Transitions and scores on a hierarchical adjacency metric
1 2
Cf J, b B, b
Cb J B
Trans ASH
Score 1
Global
1 3 4
J, b J, p, b p, B
J J p
CON SSH
4 2
1 3 5
J, b J, p, b p, m, t
J J J
CON
4
11/4=2.75
Borovets, sept. 2003
Verifying the second conjecture
Source No. of transitions
CT score Average CT score
per transition
VT score Average VT score
per transition
English 59 76 1.25 84 1.38
French 47 109 2.35 116 2.47
Romanian 65 142 2.18 152 2.34
Total 173 327 1.89 352 2.03
Borovets, sept. 2003
VT vs. Linear and Stack-based
Length of DEA
Effort
stack-based
VT linear
Borovets, sept. 2003
Exceptions
• Cases where neither VT nor the stack-based model finds an antecedent
• Four types:– pragmatic
• the Senate, the key in lock them up and throw away the key, our in our streets
– proper nouns• Mr. Gerstner, Senator Biden
– common nouns• the steelmaker, the proceeds, the top job
– pronouns
Borovets, sept. 2003
Exceptions
decreasing evoking
power
Type of RE VT
Stack-based
pragmatic 56.3% 0.0%
proper nouns 22.7% 26.1%
common nouns 16.0% 39.1%
pronouns 5.0% 34.8%
Borovets, sept. 2003
Interpretation
• alignment between evoking power and percentage of exceptions– VT predictions wrt DEA are fundamentally
correct
• pronoun exceptions linked to the RST attribution relation – integrate X said with the attributed
quotation?
Borovets, sept. 2003
Fix points of most discourse theories• structure seen as trees (AST, RST, Hobbs,
Polanyi, Marcu, etc.)• sub-trees of the whole tree make up
discourse segments• sequentiality (with some exceptions in Hobbs)• compositionality• a view on the relationship between
referentiality and structure
Borovets, sept. 2003
Parsing discourse
• Marcu’s parser
• expectations-driven incremental parsing
• VT parsing
Borovets, sept. 2003
Elements of Marcu’s constructions• a set of rhetorical relations between elementary
discourse units• consider a guesser (oracle) giving all RR between
elementary or extended text spans• each span:
– status (nuclearity)– type (RR)– promotion set
• Strong compositionality criteria: if a relation R holds between two textual spans, than R also holds between the most important units of the constituent spans.
Borovets, sept. 2003
Example[Animals heal 1][but trees compartmentalize. 2][They
endure a lifetime of injury and infection by setting boundaries that resist the spread of the invading microorganisms. 3]
1-2
21
3
1-3
S=satT= leafPS=[3]
S=nuc-satT= ELABORATION
PS=[1, 2]
S=nucT= leafPS=[2]
S=nucT= leafPS=[1]
S=nucT= CONTRAST
PS=[1, 2]
Borovets, sept. 2003
Constraint-satisfaction parsing
• Two ways to derive valid discourse structures:– a model-theoretic approach: apply axioms that
characterize the general constraints of a text structure to the text under scrutiny declarative specification of the constraints that characterize a valid structure
– a theorem-proving approach: specify rewriting rules that map a sequence of textual units into a a valid text structure control the process of derivation
Daniel Marcu, 1997, 2000
Borovets, sept. 2003
Example of constraints
• text spans do not overlap• at most one rhetorical relation can connect
two adjacent discourse spans• there exists a discourse segment (the root)
that spans over the entire text• the status, type and promotion set that are
associated with a discourse segment reflect the strong compositionality criteria
Borovets, sept. 2003
Marcu’s constraint propagation algorithm
INPUT – N textual units each characterized by 3 variables
(S, T, PS)– the set of RR that hold between units and spans
OUTPUT– 3*N*(N+1)/2 variables– all potential structures
Performance: exponential, for texts longer than 4 units does not finish in less than 3 hours (on a Sparc Ultra 2-2170)
Borovets, sept. 2003
Conversion into a propositional logic proverINPUT
– a sequence of N textual units– a set of RR that hold between units and spans
OUTPUT– all valid discourse structures
Performance (using Davis-Putnam and greedy derivation methods): still exponential, but much better (for a text with 19 units - 250 trees computed in 50 sec.)
Borovets, sept. 2003
The proof-theoretic approach
INPUT – a sequence of N textual units– a set of RR that hold between units and
spans
OUTPUT– all valid discourse structures
Performance: 19 units aprox. 3 hours and > 24000 trees
Borovets, sept. 2003
No surprise for the poor performanceFinding solutions of constraint-satisfaction
problems and finding models of theories of propositional formulas are NP-complete problems
Garey&Johnson, 1979
Parsing phrase-structure trees in the presence of functional constraints can be exponential in the worst case
Barton et al., 1985
Borovets, sept. 2003
Compiling grammars in Chomsky normal formSet of constraints applied to the input set RR to
compile a CNF grammarThen apply a top-down derivation of the grammar
rulesThe compiled grammars shown to be sound and
completePerformance: the algorithm generates O(N6)
grammar rules in O(N6) steps. Then the CKY algorithm derives the text structures in O(N3) steps (the 19 units text processed in 5.2 sec.)
Borovets, sept. 2003
Example of all possible derivations[Animals heal 1][but trees compartmentalize. 2][They
endure a lifetime of injury and infection by setting boundaries that resist the spread of the invading microorganisms. 3]
1-2
21
3
1-3
S=satT= leafPS=[3]
S=nuc-satT= ELABORATION
PS=[1, 2]
S=nucT= leafPS=[2]
S=nucT= leafPS=[1]
S=nucT= CONTRAST
PS=[1, 2]
2-3
32
1
1-3
S=satT= leafPS=[3]
S=satT= ELABORATION
PS=[2]
S=nucT= leafPS=[2]
S=nucT= leafPS=[1]
S=nuc-satT= CONTRAST
PS=[1, 2]
Borovets, sept. 2003
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
Borovets, sept. 2003
Incremental discourse parsing - a TAG inspired approach
Adjoining to the right frontier
a1
0
0
1
a
a*
’a
a1
7
Borovets, sept. 2003
Substitution
k+1
k. Although Bill would have wanted it,
k
k+1. John sold his bicycle to somebody else.
k
’
k+1
Borovets, sept. 2003
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&Webber, 1997
Borovets, sept. 2003
a. Clinton is bound to win the elections.
a
b
EVIDENCE
*
b. He is a natural born campaigner.
9
Expectations-driven incremental parsing
Borovets, sept. 2003
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
Borovets, sept. 2003
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
Borovets, sept. 2003
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
Borovets, sept. 2003
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
Borovets, sept. 2003
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
Borovets, sept. 2003
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
Borovets, sept. 2003
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
Borovets, sept. 2003
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
Borovets, sept. 2003
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
Borovets, sept. 2003
The non-monotonicity of incremental parsing[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]
From [Cristea&Webber,1998]
1
EXAMPLE
2 3
CIRCUMSTANCE
4
EVIDENCE
1
-, - for example
2 3
whenever -, -
4
because -, -
Borovets, sept. 2003
Marcu’s shallow approach
• use markers to identify: – discourse segments– rhetorical relations– discourse structure
• markers are three-ways ambiguous:– sentential versus discourse usage– rhetorical relation they signal– the size of the textual span they connect
• corpus investigation of cue phrases
Borovets, sept. 2003
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 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
Borovets, sept. 2003
What cue-phrases can 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>
Borovets, sept. 2003
What cue-phrases can 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
Borovets, sept. 2003
What cue-phrases can 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
-, - 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
Borovets, sept. 2003
Marcu’s algorithm for rhetorical parsing of free texts• determine the set D of all cue phrases in T with a
discourse function• segment T in sections, paragraphs, sentences and
clauses using the corpus (C)• for each section, paragraph and sentence:
– use C to hypothesize rhetorical relations signaled by markers in D
– use cohesion to hypothesize rhetorical relations not signaled by markers
– find all the valid text trees that correspond to that level– select the best tree based on assigned weights
• merge the resulted trees
Borovets, sept. 2003
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)cDRA=ac
a
b c
V=(a)(b)cDRA=abcrightwrong
Borovets, sept. 2003
VT guides an incremental discourse parsing
The tree resulted after the parsing process 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
Borovets, sept. 2003
The architecture
blackboard
shallow parser
POS tagger
reference resolution expert
discourse parser
transitions expert
cue-words expert
Borovets, sept. 2003
reference expert
transitions expert
The discourse parser implements a „beam search“
N
N
Borovets, sept. 2003
Processing the text:first 2 steps
1. John sold his bicycle
[John‘s bike]
[John]
1 2
concession
2. although Bill would have wanted it.
2
concession
*
E-DEA=12
[Bill]
it
REF score
1 direct 2
2
Borovets, sept. 2003
[Bill]
[John]
1. John sold his bicycle
1 2
concession
2. although Bill would have wanted it.
ABRUPT SHIFT
CT score
1
1
[bike]
[bike]
Processing the text:first 2 steps
Borovets, sept. 2003
3
unknown
*
Processing the text:step 3
1. John sold his bicycle
1 2
concession
2. although Bill would have wanted it.
3. He obtained a good price for it,
1
2
concession
3
unknown
He
[price]
it
[Bill]
[bike]
REF score
1 direct 2
2
42 direct
E-DEA=123
Borovets, sept. 2003
Processing the text:step 3
1. John sold his bicycle
1 2
concession
2. although Bill would have wanted it.
3. He obtained a good price for it,
1
2
concession
3
unknown
[price]
CT score
1.5
[bike]
[Bill]
ABRUPT SHIFT 1
[Bill]
[bike]
[John]
[bike]
2SMOOTH SHIFT
Borovets, sept. 2003
4
elaboration
*
Processing the text:step 4
1. John sold his bicycle
2. although Bill would have wanted it.
3. He obtained a good price for it,
4. which Bill could not have afforded.
1
2
concession
3
unknown1
2
concession
3
unknown
4
elaboration
E-DEA=1234which
[Bill]
[Bill]
[price]
[bike]
REF score
1 direct 2
1.4
42 direct
21 direct
-11 direct: prop-n/pron
Borovets, sept. 2003
Processing the text:step 4
1. John sold his bicycle
2. although Bill would have wanted it.
3. He obtained a good price for it,
4. which Bill could not have afforded.
1
2
concession
3
unknown1
2
concession
3
unknown
4
elaboration
CT score
2
3RETAINING
ABRUPT SHIFT 1
2SMOOTH SHIFT
[bike]
[John]
[Bill]
[bike] [price]
[Bill]
[bike]
[bike]
[Bill]
Borovets, sept. 2003
Processing the text:step 3, variant
1. John sold his bicycle
2. although Bill would have wanted it.
3. He obtained a good price for it,
3
unknown
*
REF score
1 direct 2
2
42 direct
1 2
concession
1 2
concession 3
unknown
E-DEA=213
He
[price]
it
[bike]
[John]
Borovets, sept. 2003
Processing the text:step 3, variant
1. John sold his bicycle
2. although Bill would have wanted it.
3. He obtained a good price for it,
1 2
concession 3
unknown
CT score
2.5
CONTINUING 4
ABRUPT SHIFT 1
[bike]
[John]
[price]
[John]
[bike]
[Bill]
[bike]
Borovets, sept. 2003
Processing the text:step 4, variant
4
elaboration
*
1. John sold his bicycle
2. although Bill would have wanted it.
3. He obtained a good price for it,
4. which Bill could not have afforded.
1 2
concession 3
unknown
1 2
concession
4
elaboration
3
unknown
[price]
[bike]
[John]
[Bill]
[bike]
REF score
1 direct 2
1.6
42 direct
21 direct
01 inferential
which
[Bill]
E-DEA=2134
Borovets, sept. 2003
Processing the text:step 4, variant
1. John sold his bicycle
2. although Bill would have wanted it.
3. He obtained a good price for it,
4. which Bill could not have afforded.
1 2
concession 3
unknown
1 2
concession
4
elaboration
3
unknown
[bike]
[John]
[price]
[John]
[bike]
[Bill]
[bike]
[bike]
[Bill]
CT score
2.3
CONTINUING 4
ABRUPT SHIFT 1
2SMOOTH SHIFT
Borovets, sept. 2003
4
conclusion
*
Processing the text:step 5
1. John sold his bicycle
2. although Bill would have wanted it.
3. He obtained a good price for it,
4. which Bill could not have afforded.
1 2
concession 3
unknown
1 2
concession
4
elaboration
3
unknown
5. Therefore he decided to use the money for going in a trip.
1 2
concession
4 5
conclusion3
elaboration
unknown
he
[money]
[trip]
REF score
1 direct 2
1.7
42 direct
21 direct
01 inferential
21 direct
REF score
[bike]
[Bill]E-DEA=21345
Borovets, sept. 2003
Processing the text:step 5
1. John sold his bicycle
2. although Bill would have wanted it.
3. He obtained a good price for it,
4. which Bill could not have afforded.
1 2
concession 3
unknown
1 2
concession
4
elaboration
3
unknown
REF score5. Therefore he decided to use the money for going in a trip.
1 2
concession
4 5
conclusion3
elaboration
unknown
[bike]
[John]
[price]
[John]
[bike]
[Bill]
[bike]
[bike]
[Bill][money]
[trip]
[Bill]
CT score
2.25
CONTINUING 4
ABRUPT SHIFT 1
2SMOOTH SHIFT
2SMOOTH SHIFT
Borovets, sept. 2003
4
conclusion
*
Processing the text:step 5, variant 2
1. John sold his bicycle
2. although Bill would have wanted it.
3. He obtained a good price for it,
4. which Bill could not have afforded.
1 2
concession
4
elaboration
3
unknown
5. Therefore he decided to use the money for going in a trip.
REF score
1 direct 2
1.7
42 direct
21 direct
01 inferential
21 direct
REF score
1 2
concession
4
5
conclusion
3
elaboration
unknown
[price]
[bike]
[John] E-DEA=24135
he
[money]
[trip]
Borovets, sept. 2003
Processing the text:step 5, variant 2
1. John sold his bicycle
2. although Bill would have wanted it.
3. He obtained a good price for it,
4. which Bill could not have afforded.
1 2
concession 3
unknown
1 2
concession
4
elaboration
3
unknown
5. Therefore he decided to use the money for going in a trip.
1 2
concession
4
5
conclusion
3
elaboration
unknown
[bike]
[John]
[price]
[John]
[bike]
[Bill]
[bike]
[bike]
[Bill][money]
[trip]
[John]
CT score
2.75
CONTINUING 4
ABRUPT SHIFT 1
2SMOOTH SHIFT
CONTINUING 4
Borovets, sept. 2003
4
conclusion
*
Processing the text:step 5, variant 3
1. John sold his bicycle
2. although Bill would have wanted it.
3. He obtained a good price for it,
4. which Bill could not have afforded.
1 2
concession
4
elaboration
3
unknown
5. Therefore he decided to use the money for going in a trip.
1 2
concession
4
5
conclusion
3
elaboration
unknown
[bike]
[John] E-DEA=23415
he
[money]
[trip]
REF score
1 direct 2
1.7
42 direct
21 direct
01 inferential
21 direct
REF score
Borovets, sept. 2003
Processing the text:step 5, variant 3
1. John sold his bicycle
2. although Bill would have wanted it.
3. He obtained a good price for it,
4. which Bill could not have afforded.
1 2
concession 3
unknown
1 2
concession
4
elaboration
3
unknown
5. Therefore he decided to use the money for going in a trip.
1 2
concession
4
5
conclusion
3
elaboration
unknown
CT score
2.75
CONTINUING 4
ABRUPT SHIFT 1
2SMOOTH SHIFT
CONTINUING 4
[bike]
[John]
[price]
[John]
[bike]
[Bill]
[bike]
[bike]
[Bill]
[money]
[trip]
[John]
Borovets, sept. 2003
Processing the text:results
1. John sold his bicycle
2. although Bill would have wanted it.
3. He obtained a good price for it,
4. which Bill could not have afforded.
5. Therefore he decided to use the money for going in a trip.
1 2
concession
4 5
conclusion3
elaboration
unknown
1 2
concession
4
5
conclusion
3
elaboration
unknown
1 2
concession
4
5
conclusion
3
elaboration
unknown
Ref=1.7 HCT=2.25
Ref=1.7 HCT=2.75 Ref=1.7 HCT=2.75
Borovets, sept. 2003
Conclusions
• a versatile architecture that allows for a „detached“ behaviour
• the more experts you have and the more accurate knowledge they have the more finer results you get
• adequate for an on-line processing of discourse
• real-time abstracting of typed text, IE, real-time translation, deep understanding...
Borovets, sept. 2003
Related work
• towards an integrated discourse parsing – anaphora resolution algorithm
• we need corpora resources annotated to fit discourse needs
• anaphora resolution guided by discourse structure• “reverse” anaphora: correct discourse structure• focused summarization • in search for an underspecified representation