From monologue to dialogue: mapping discourse to dialogue structure

1
From monologue to dialogue mapping discourse to
dialogue structure

• Paul Piwek
• Centre for Research in Computing

Dialogue Dynamics The scaling up challenge,
London 2008
2
Collaborators

• Helmut Prendinger
• Hugo Hernault
• Mitsuru Ishizuka

3
Aim

• Provide an outline of how work on mapping text in
  monologue form to text in dialogue form ? i.e.,
  fictive dialogue ? might help address the
  knowledge acquisition bottle neck for dialogue
  (question-answering) systems.

4
Plan

• Background the generation of fictive dialogue.
  (what, why and how)
• Relevance to dialogue systems
• Description of the T2D text-to-dialogue system
• Conclusions

5
What is fictive dialogue?

• Historical precedent Plato, Erasmus, Galileo, ,
  Hofstadter
• Common on Radio, TV, Theatre, Games,

6
Why Fictive Dialogue?

• A means for presenting information
• which complements monologue
• diagrams and pictures.
• successful for entertainment

7
Why Fictive Dialogue?

• A means for presenting information
• which complements monologue
• diagrams and pictures.
• successful for entertainment
• allows an author to introduce different points of
  view

8
Why Fictive Dialogue?

• A means for presenting information
• which complements monologue
• diagrams and pictures.
• successful for entertainment
• allows an author to introduce different points of
  view
• can be effective in education and persuasion

9
Why Fictive Dialogue?

• Students write more in free recall test (Craig et
  al., 2000)
• Students ask more and deeper questions in a
  transfer task (Craig et al., 2000)
• There is more discussion amongst students and
  less irrelevant banter (Lee et al., 1999)
• Student learning is at least as good as in
  monologue condition (Cox et al., 1999)
• Team of two agents having a conversation more
  persuasive than a single agent directly
  addressing a user (Suzuki Yamada, 2004)

10
Generating Fictive Dialogue
11
Generating Fictive Dialogue
12
Generating Fictive Dialogue
13
Generating Fictive Dialogue
14
Generating Fictive Dialogue
15
Generating Fictive Dialogue

• Approaches
• From data to script (Piwek Van Deemter, 2007
  RLaC Van Deemter et al., in press AIJ)

16
Database (Java JAM)

• FACT attribute "car-1" "horsepower "80hp"
• FACT impact "car-1" "horsepower "sportiness"
  "pos"
• FACT importance "horsepower" "sportiness" "high"
• FACT role "Ritchie" "seller"
• FACT role "Tina" "buyer"
• FACT trait "Ritchie" "politeness" "impolite"

17
Generating Fictive Dialogue

• Approaches
• From data to script (Piwek Van Deemter, 2007
  RLaC Van Deemter et al., in press AIJ)

18
Generating Fictive Dialogue

• Approaches
• From data to script (Piwek Van Deemter, 2007
  RLaC Van Deemter et al., in press AIJ)
• From text to script (T2D) Piwek et al. 2007 IVA07

Patient information leaflets
19
Relevance to Dialogue Systems

• A by-product of the T2D text-to-dialogue system
  is DialogueNet ? a repository of question-answer
  pairs annotated with underlying coherence
  relations.

20
Relevance to Dialogue Systems

• Shallow Question-Answering
• Repository set QA pairs.
• Answering (Similarity) matching of user
  questions with questions in the repository
• Repository Question patterns with answer
  patterns.
• Answering Matching of user question with
  question patterns in the repository
• Repository Knowledge base.
• Answering Inferring answers from KB to fully
  interpreted user questions.
• Deep Question-Answering

21
Relevance to Dialogue Systems

• In all these approaches to QA, an essential
  component is the repository. Often there is a
  scaling up problem, because it has to be
  populated manually.
• How does T2D fit in? automatic construction of
  a collection of annotated question-answer pairs
  from text.

22
The T2D System

• Automatic generation of (multimodal) dialogues
  from text (cf. Isard et al., 2003)
• and other available online sources (ontologies,
  Wikipedia)
• Text-to-dialogue transformation is
  meaning-preserving (faithful) (cf. Nadamoto and
  Tanaka 2005)
• Basis of transformation is discourse/text rather
  than individual sentences
• Supports coherence of output dialogue
• Purpose anyone including non-experts can
  produce highly attractive content
• The automatically generated dialogue structures
  (DialogueNet) are intended to support
• Multimodal information presentation
• But, possibly also Interactive Question-Answering

23
T2D System Architecture
Slide design inspired by J. Cassell on BEAT
24
T2D Input

• Patient Information Leaflet (from PIL corpus)
• Do not take Klaricid tablets if you are
  allergic to clarithromycin. Klaricid does not
  interact with oral contraceptives.

25
T2D System Architecture
26
Rhetorical Structure Theory

• Mann, W.C., Thompson, S.A. (1988). Rhetorical
  structure theory Toward a functional theory of
  text organization. Text, 8(3)243-281.
• RST A theory of text organization
• Segments text into non-overlapping, semantically
  independent units
• Identifies rhetorical (discourse/coherence)
  relations between text segment Elaboration,
  Contrast, Circumstance, Concession, Condition,
  Motivation,
• Nucleus and satellite are distinguished
  (nucleus describes relatively more important
  information than satellite)

27
Rhetorical Structure Theory

• Example To eat Japanese food, use chopsticks.
• Definition of MEANS A means satellite specifies
  a method, mechanism, instrument, channel or
  conduit for accomplishing some goal.

28
Rhetorical Structure Theory

• Mononuclear relations nucleus and satellite are
  distinguished.
• Multinuclear relations no distinguished nucleus.

29
T2D Input

• Patient Information Leaflet (from PILS corpus)
• Do not take Klaricid tablets if you are
  allergic to clarithromycin. Klaricid does not
  interact with oral contraceptives.

30
RST DAS Analyzer (Le and Abeysinghe, 2003)

• 1) Text segmentation
• Do not take Klaricid tablets
• You are allergic to clarithromycin
• Klaricid does not interact with oral
  contraceptives
• 2) Identify rhetorical relations between
  discourse segments

31
RST DAS Analyzer
Do not take Klaricid tablets if you are allergic
to clarithromycin or other macrolide antibiotics
such as erythromycin or azithromycin. If you
have any liver or kidney problems consult your
doctor before taking these tablets. Klaricid does
not interact with oral contraceptives.
32
T2D System Architecture
33
RST Tree to DialogueNet
Question
Answer
ANSWER (MEANS)
Use chopsticks.
How can I eat Japanese food?
34
RST Tree to DialogueNet
35
RST Tree to DialogueNet

• Abstract mapping information preserving

• MEANS(P,Q) ?
• Layman How can I inf_mark_rmv(P)?
• Expert Q.

(The function inf-mark_rmv removes the infinite
mark of the main verb).
36
RST Tree to DialogueNet
1) Nucleus in Imperative Form (Take Klaricid
tablets / Do not take Klaricid
tablets) CONDITION(P,Q) imperative(P) ?
Layman Under what circumstances should I P?
Expert If Q. 2) Nucleus in Declarative Form
with Modal Auxiliary (You should take Klaracid
tablets) CONDITION(P,Q) declarative-modal-aux(
P) ? Layman Under what circumstances
flip(P)? Expert If Q. 3) Alternative
Mapping CONDITION(P,Q) ? Layman What if
Q. Expert Then P. P is
PIyou,youI,myyour,yourmy,mineyours,your
smine flip(X) inverses subject and auxiliary.
37
RST Tree to DialogueNet

• If P, then Q condition(P,Q)
• What if P? Q qa(What if P?,Q)

38
RST Tree to DialogueNet

• If P, then Q condition(P,Q)
• What if P? Q qa(What if P?,Q)
• condition(P,Q) ? qa(What if P?,Q)

39
RST Tree to DialogueNet

• If P, then Q condition(P,Q)
• What if P? Q qa(What if P?,Q)
• condition(P,Q) ? qa(What if P?,Q)
• What if P? Q qa(lambda(x,condition(P,x)),Q)
• for qa function appl gtgt condition(P,Q)

40
RST Tree to DialogueNet

• P ? qa(lambda(x,P),E) with P PEx
• If it rains, the tiles get wet. ? What if it
  rains?
• If it rains, the tiles get wet. ? Under what
  circumstances do the tiles get wet? If it rains.
• John is at home because I saw his car outside ?
  What is the relation between ..? The latter is
  evidence for the former.
• John saw Mary ?Who did John see? M.

41
RST Tree to DialogueNet
42
T2D System Architecture
(Not done yet)
43
T2D System Architecture
(Not done yet)
44
MPML3D

• ltMPML3D version"1.0"gt
• ltHeadgtlt/Headgt
• ltBody startImmediately"Demo1"gt
• - ltTask name"Demo1" priority"0"gt
• - ltSequentialgt
• - ltParallelgt
•   ltAction name"yuukiSpeak"gtyuuki.speak(Under
  what circumstances shouldnt I take Klaricid
  tablets? ")lt/Actiongt
•   ltAction minor"true" startOn"yuukiSpeak1.
  begin" stopOn"yuukiSpeak9.end"gtyuuki.turnHead(-
  10,0.2,10,0.3)lt/Actiongt
•   ltAction minor"true" startOn"yuukiSpeak1.
  end" stopOn"yuukiSpeak9.end"gtken.turnHead(10,0.
  2,10,0.2)lt/Actiongt
•   lt/Parallelgt
•   ltActiongtken.turnHead(10,0.2,0.3,0.2)lt/Actiongt
  
• - ltParallelgt
•   ltAction name"kenSpeak"gtken.speak( If you
  are allergic to clarithromycin. ")lt/Actiongt
•   ltAction minor"true" startOn"kenSpeak1.en
  d" stopOn"kenSpeak6.end"gtken.turnHead(10,0.2,10
  ,0.2)lt/Actiongt
•   ltAction minor"true" startOn"kenSpeak2.be
  gin"gtken.gesture("BEAT_SINGLE", 0.2,
  0.6)lt/Actiongt
•   ltAction minor"true" startOn"kenSpeak2.be
  gin"gtyuuki.gesture("breath")lt/Actiongt
•   lt/Parallelgt
• lt/Sequentialgt
•   lt/Taskgt

45
Preliminary Evaluation

• Ongoing work 9 relations implemented so far.
• DAS evaluation
• Random sample of 100 conditionals
• correct 61, failure 39
• Details failure 19 OCR error 19 No mapping
  22 DAS crashes 40 incorrect analysis (15.6
  overall).
• Mapping evaluation
• Fresh random sample of 100 conditionals
• Manually annotated in terms of RST
• Mapping 92 correct 8 incorrect (4 machinese
  syntax error 4 mapping rules did not cover a
  specific case).

46
Limitations and Open Questions

• Discourse parser (DAS)
• Reliability of output is an issue
• Is there a formal method to decide the quality of
  DAS output?
• Alternative SPADE (Soricut Marcu, 2003), but
  only within sentence relations approx. 80
  accuracy
• QA pairs
• Variation of mappings is currently limited
• Doesnt give you indirect answers
• Mappings manually created, but content-independent
  
• Only now starting to investigate other forms of
  text input (so far only Patient Information
  Leaflets corpus studied). Currently,
  investigating processing of Penn Discourse
  Treebank input (WSJ).

47
Conclusions

• A novel method for automatically generating
  question-answer pairs/dialogue from text based on
  coherence relations
• Possibly useful by-product for QA systems
  DialogueNet (QA pairs annotated with coherence
  relations)

48
Conclusions

• A novel method for automatically generating
  question-answer pairs/dialogue from text based on
  coherence relations
• Possibly useful by-product for QA systems
  DialogueNet (QA pairs annotated with coherence
  relations)

Thank you!