Discourse Structure and Anaphoric Accessibility

1
Discourse Structure and Anaphoric Accessibility

• Massimo Poesio and Barbara Di Eugenio
• with help from Gerard Keohane

2
Content

• Empirical Investigations of Discourse Structure
• Grosz and Sidners theory of the Global Focus
• Relational Discourse Analysis
• How we used RDA to study GS
• Results
• Discussion

3
Empirical Investigations of Discourse Structure
A new opportunity

• Original proposals concerning effect of discourse
  structure on accessibility (Reichman, 1985 Fox,
  1987 Grosz and Sidner, 1986) based on
  unsystematic analysis of data
• These days we know more about reliable studies of
  discourse phenomena (Passonneau and Litman, 1993
  Carletta et al, 1997)
• These new resources already used to propose new
  theories of anaphora and discourse structure such
  as Veins Theory (Cristea, Ide, Marcu, et al,
  1998, 1999, 2000)
• The goal of this project use a reliably
  annotated corpus (the Sherlock corpus from the
  University of Pittsburgh, Moser and Moore, 1996
  Di Eugenio et al, 1997) to study claims of GS

4
Grosz and Sidners Theory of the Global Focus

• The structure of a discourse is determined by the
  intentions utterances are meant to convey
  (DISCOURSE SEGMENT PURPOSES)
• INTENTIONAL STRUCTURE DOMINANCE and SAT-PRECEDES
  relations between DSPs
• ATTENTIONAL STRUCTURE a stack of FOCUS SPACES
• Focus spaces on the stack contain accessible
  discourse entities
• Presence on the stack reflects intentional
  structure
• The problem how to identify DSPs in a discourse

5
Relational Discourse Analysis (RDA)

• Moore and Pollack, 1992 Moser and Moore, 1996
• Combines ideas from RST and Grosz and Sidners
  theory
• From Grosz and Sidner discourse structure is
  determined by intentional structure
• RDA-SEGMENT a segment expressing an intentional
  relation
• From RST segments have internal structure
• CORE (cfr. NUCLEUS)
• CONTRIBUTOR (cfr. SATELLITE)
• Both INTENTIONAL and INFORMATIONAL relations
• A fixed number of intentional relations
• Has been proven to be usable for reliable analysis

6
RDA Analysis of an excerpt from a tutorial

• 1.1 Before troubleshooting inside the text
  station,
• 1.2 Its always best to eliminate both the UUT
  and the TP
• 2.1 Since the test package is moved frequently
• 2.2 It is prone to damage
• 3.1 Also, testing the test package is much easier
  and faster
• 3.2 than opening up test station drawers.

CONVINCE
CONVINCE
ENABLE
Prescribed-act Wrong-act
Causeeffect
step1step2
1.1
1.2
2.2
2.1
3.2
3.1
7
Moser and Moore mapping between RST relations
and GS

• Basic principles
• Every DSP must be associated with a core
• Constituents of the RDA structure that do not
  include cores such as clusters do not
  introduce DSPs
• Consequences for attentional state
• A new focus space only pushed when a segment is
  open
• Information relations do not affect the
  attentional state

8
Mapping RDA into Attentional State

• 1.1 Before troubleshooting inside the text
  station,
• 1.2 Its always best to eliminate both the UUT
  and the TP
• 2.1 Since the test package is moved frequently
• 2.2 It is prone to damage
• 3.1 Also, testing the test package is much easier
  and faster
• 3.2 than opening up test station drawers.

DSP1
CONVINCE
CONVINCE
DSP 2
ENABLE
Prescribed-act Wrong-act
Causeeffect
step1step2
1.1
1.2
2.2
2.1
3.2
3.1
9
Using an RDA-annotated corpus to study anaphoric
accessibility

• The data the SHERLOCK corpus, already annotated
  according to RDA instructions (Moser, 1996)
• Added anaphoric annotation according to GNOME
  instructions (Poesio, 2000) derived from MATE
  scheme (Poesio Bruneseaux and Romary, 1999)
• Use RDA analysis to drive focus space
  construction
• Measure
• Accessibility
• Perplexity

10
The Data the SHERLOCK corpus

• 17 tutorial dialogues collected within the
  Sherlock project (Lesgold et al, 1992)
• Students solve electronic troubleshooting problem
• 313 turns, 1333 clauses
• RDA annotation Moser and Moore, 1996
• Reliability verified at different levels
• Intentional relations CONCEDE, CONVINCE, ENABLE,
  JOINT

11
An example of Sherlock dialogue

• STUDENT
• 1.1 Why isn't measurement signal path green
  during good test
• readings (steps)?
• TUTOR
• 2.1 For each step that passed,
• 2.2 you know the measurement path is good.
• 2.3 You also know that one of the measurement
  paths is bad.
• 2.4 Showing the UUT, Test Package, and
  measurement section as
• unknown is correct
• 2.5 because, you know when you get your fail
  that
• something was wrong,
• 2.6 but you didn't know exactly what.
• 2.7 The DMM is green
• 2.8 because it has been working all along.
• 2.9 The stimulus section is green
• 2.10 because it was not used
• 2.11 and is assumed to be good.

12
Anaphoric Annotation

• The GNOME scheme (Poesio, 2000)
• Mark up all NPs as NE element, with a variety of
  attributes
• About 3000 NEs
• Use separate ANTE element to mark up anaphoric
  relations (including bridges)
• In this annotation only direct anaphoric
  relations
• (About 1500 total)

13
Evaluation

• A PERL script simulates focus space construction
  and computes accessibility and perplexity
• Accessibility whether antecedent is in focus
  stack
• Perplexity Sum 1/d(xi ) m(xi) (where m(xi) 1
  if xi matches anaphor, 0 otherwise)
• Parameters for focus space construction
• PUSHING
• Whenever relation is encountered (either
  informational or intentional)
• Only when intentional
• POPPING
• As soon as associated constituent is completed
• Immediate popping of contributors, delayed
  popping of cores
• Delayed popping of contributors

14
Evaluation I Intentional vs Informational
Accessibility
OK NO Out of AP PN
All 199 74 63 158
Intentional (immediate popping) 280 20 63 131
Perplexity All 0.83, Intentional 1.23
15
Complications
ENABLE

• 24.13a Since S52 puts a return (0 VDC) on its
  outputs
• 24.13b when they are active,
• 24.14 the inactive state must be some other
  voltage.
• 24.15 So even though you may not know what the
  other voltage is,
• 24.16 You can test to ensure that
• 24.17a the active pins are 0 VDC
• 24.17b and all the inactive pins are not 0 VDC.

DSP 1
CONCEDE
ENABLE
24.14
24.16
24.15
Effectcause
24.13a
24.13b
Contrast1 contrast2
24.17a
24.17b
16
Complications
ENABLE

• 24.13a Since S52 puts a return (0 VDC) on its
  outputs
• 24.13b when they are active,
• 24.14 the inactive state must be some other
  voltage.
• 24.15 So even though you may not know what the
  other voltage is,
• 24.16 You can test to ensure that
• 24.17a the active pins are 0 VDC
• 24.17b and all the inactive pins are not 0 VDC.

DSP 1
CONCEDE
ENABLE
24.14
24.16
24.15
Effectcause
24.13a
24.13b
Contrast1 contrast2
24.17a
24.17b
17
Evaluation II Delayed Popping
Accessibility
OK NO
Immediate popping 280 20
Delay pop of cores 287 16
Delay pop of contributors 310 8

• Average perplexity with immediate popping 1.23
• Delayed popping of cores 1.3
• Delayed popping of contributors 1.33

Perplexity
18
Discussion

• Accessibility
• Intentional vs. informational distinction makes
  sense
• Cfr. Fox
• Want to keep contributors as well as cores on
  stack
• cfr. Veins Theory
• An evaluation of Grosz and Sidners framework
• The most direct implementation makes quite a few
  discourse entities unaccessible
• Difficult to interpret more complex operations in
  terms of intentional structure
• Alternative a cache model (cfr. Guindon 1985,
  Walker 1996, 1998)
• Version 1 (conservative) cache of focus spaces
• Version 2 cache of forward looking centers

19
Cache-based global focus a conservative proposal

• Cache elements are FOCUS SPACES
• Cache elements are RANKED
• Current focus space lt other constituents of
  same segment lt dominating segments lt focus
  spaces of contributors to closed spaces(Cfr.
  Reichman 85)
• Search algorithm follow ranking
• Cache replacement algorithm
• Opening RDA segment open new focus space,
  replace lowest-ranked element of cache, assign it
  highest rank
• Closing RDA segment Assign lowest rank to
  embedded contributors