Argumentation in Institutional Dialogues: Corpus Analysis

1
Argumentation in Institutional Dialogues Corpus
Analysis

• Mare Koit
• mare.koit_at_ut.ee

2
Outline

• Motivation
• Our argumentation model
• Empirical Material
• Results of Corpus Analysis
• Implementation
• Conclusion and Future Work

3
Motivation

• How do people argue?
• There are many dialogue corpora that contain
  human-human or human-computer dialogues. Most of
  them involve task-oriented dialogues in fairly
  simple domains
• Switchboard 2430 spontaneous conversations
  http//www.isip.msstate.edu/projects/switchboard
  
• HCRC Maptask 128 dialogues http//www.hcrc.ed.ac.u
  k/dialogue/maptask.html
• TRAINS corpus 98 problem solving dialogues
  http//www.cs.rochester.edu/research/trains/
• Monroe corpus 20 human-human mixed-initiative,
  task-oriented dialogues http//www.cs.rochester.ed
  u/research/cisd/resources/monroe/
• Linköping Dialogue Corpus 60 dialogues collected
  in Wizard of Oz-experiments http//www.ida.liu.se/
  nlplab/dialogues/corpora.html
• VERBMOBIL corpus http//verbmobil.dfki.de/
• Estonian Dialogue Corpus 600 human-human, 21
  Wizard of Oz dialogues http//math.ut.ee/koit/Dia
  loog/EDiC

4
Motivation (2)

• We have worked out an argumentation model which
  involves natural reasoning.
• Agreement negotiation conversations have been
  considered theoretically, without using of
  empirical material.
• Goal to investigate our argumentation model on
  Estonian human-human spoken dialogues.

5
Argumentation model

• Mare Koit and Haldur Õim. Argumentation in the
  Agreement Negotiation Process A Model that
  Involves Natural Reasoning. In Proc. of the
  Workshop W12 on Computational Models of Natural
  Argument. 16th European Conference on Artificial
  Intelligence, 5356, Valencia, Spain, 2004.
• We consider dialogues where the goal of A is to
  get partner B to carry out a certain action D
• exchange of arguments
• agreement negotiation dialogues
• Model of conversation agent
• reasoning processes that people supposedly go
  through when working out a decision (to do D or
  not)

6
Argumentation model (2)

• After A has requested B to do D then B can
  respond with agreement or rejection, depending on
  the result of his/her reasoning.
• a kind of BDI (belief-desire-intention) model (J.
  Allen)
• consists of two parts
• Part 1 a model of human motivational sphere
  (what is pleasant/unpleasant, useful/harmful etc.
  for a reasoning subject),
• Part 2 reasoning procedures.

7
Argumentation model (3)

• Part 1 Model of human motivational sphere
• Naïve theory of reasoning there are 3 basic
  factors in the human motivational sphere that
  regulate reasoning about D
• WISH-, NEEDED- and MUST-determinants.
• w ( w(are-resourcesD), w(pleasantD),
  w(unpleasantD), w(usefulD), w(harmfulD),
  w(obligatoryD), w(prohibitedD),
  w(punishment-for-doingD), w(punishment-for-not-doi
  ngD))
• numerical weights
• Part 2 Reasoning procedures
• Triggered by
• WISH-, NEEDED- or MUST-determinants.
• A reasoning procedure represents steps that the
  agent goes through in his/her reasoning process
  these consist in computing and comparing the
  weights of different aspects of D and the result
  is the decision to do or not to do D.

8
Argumentation model (4)

• Reasoning procedure triggered by
  NEEDED-determinant
• Input considerations w(useful) gt w(harmful)
• Are there enough resources for doing D?
• If not then do not do D.
• Is w(pleasant) gt w(unpleasant)?
• If not then go to 1.
• Is D prohibited?
• If not then do D.
• Is w(pleasant)w(useful) gt w(unpleasant)
  w(harmful) w(punishment-for-doing)?
• If yes then do D.
• Otherwise do not do D.
• 1 Is D obligatory?
• If not then do not do D.
• Is w(pleasant) w(useful) w(punishment-for-not-
  doing)gt w(unpleasant) w(harmful)?
• If yes then do D.
• Otherwise do not do D.

9
Argumentation model (5)

• A communicative strategy is an algorithm which is
  used by a participant of communication to achieve
  his/her communicative goal.
• The participant A having the goal that the
  partner B decides to do D can realize his/her
  communicative strategy in different ways (using
  different arguments for)
• stress pleasant aspects of D (i.e. entice B),
• stress usefulness of D for B (i.e. persuade B),
• stress punishment for not doing D if D is
  obligatory (threaten B).
• We call these concrete ways of realization of a
  communicative strategy communicative tactics.
• Communicative tactics are ways of argumentation.

10
Argumentation model (6)

• / Tactic of persuading /
• WHILE B is rejecting AND A is not giving up DO
• CASE Bs answer of
• no resources 
• present a counter-argument in order to point at
  the possibility to gain the resources, at the
  same time showing that the cost of gaining these
  resources is lower than the weight of the
  usefulness of D
• much harm 
• present a counter-argument to decrease the value
  of harmfulness in comparison with the weight of
  usefulness
• much unpleasant 
• present a counter-argument in order to downgrade
  the unpleasant aspects of D as compared to the
  useful aspects of D
• D is prohibited and the punishment is great 
• present a counter-argument in order to downgrade
  the weight of punishment as compared to the
  usefulness of D
• END CASE
• Present an argument to stress the usefulness of D.

11
Empirical Material

• 36 annotated dialogues were taken from Estonian
  Dialogue Corpus where a client calls a travel
  bureau having a goal to book a trip to a certain
  place (i.e. action D).
• A travel agent (A) could be interested in booking
  the trip by the client (B)
• gt we may expect that A tries to influence B in
  such a way that B would decide to book the trip
  immediately, in this bureau.
• Both of enticing and threatening can be
  eliminated because A is an official person and is
  obligated to communicate cooperatively,
  impersonally, friendly, peacefully (i.e. to stay
  in a fixed point of the communicative space). A
  only can persuade B.
• The general idea underlying the tactic of
  persuading is that A proposes arguments for
  usefulness of D trying to keep the weight of
  usefulness for B high enough and the possible
  negative values of other aspects brought out by B
  low enough so that the sum of weigths of positive
  aspects of D gt the sum of negative aspects gt B
  will decide to do D.

12
Empirical Material (2)

• /--/ utterance translated from Estonian
  dialogue act
• A an advantageous variant exists to fly with
  Finnair SA GIVING INFORMATION
• B yes FR CONTINUER
• A one must think about this price it is three
  thousand one hundred with taxes
• SA GIVING INFORMATION
• (.)
• B ahah FR CONTINUER
• A the next price class begins with three
  thousand eight hundred thirty plus taxes and
  these are at least one ton SA GIVING
  INFORMATION
• B mhmh FR CONTINUER
• A so the difference is quite big AI
  INFERENCE
• (.) from the next price class AI
  SPECIFICATION
• (0.5)
• B .hh clear FR BOUNDER
• A there are free places at the moment SA
  GIVING INFORMATION
• (0.5)
• B clear FR BOUNDER
• A but it must be bought off within three days
  AI EMPHAZISE
• B mhmh mhmh. FR ACKNOWLEDGEMENT
• /--/

13
Results of Corpus Analysis

• 36 dialogues
• 1896 utterances
• 11724 running words
• Number of dialogue act tags 2496
• 87 different acts
• The most frequent acts
• questions
• QUF WH-QUESTION (132)
• answers to questions
• QUS GIVING INFORMATION (298)
• free reactions
• FR CONTINUER (193)
• FR ACKNOWLEDGEMENT (141)

14
Results of Corpus Analysis (2)

• Argumentation a process of exchanging of
  arguments that begins with a proposal or offer by
  one participant to do an action and ends with an
  agreement or disagreement of the partner has
  only been found in 4 dialogues
• The typical chain of dialogue acts that form
  argumentation
• A DIF PROPOSAL/ DIF OFFER
• / information-sharing subdialogue /
• --gtA SA GIVING INFORMATION / AI ACCOUNT
• --gtB FR CONTINUER
• B DIS DISAGREEMENT

15
Implementation

• A simple dialogue system (travel agent)
  http//math.ut.ee/treumuth
• Next aim to include our argumentation model
  into the DS
• DS uses a vector of weights as a (standard)
  user model
• DSs goal is to achieve an agreement of a user to
  book a trip ( action D).
• DS assumes
• user has resources for doing D
  w(are-resources)1
• D is useful for her w(useful)1
• The values of the remaining aspects of D equal to
  0.
• The reasoning which is triggered by
  NEEDED-determinant yields the positive decision
  (to do D) on this model.
• If the client does not agree then persuasion
  starts.
• DS changes the user model depending on the users
  argument.

16
Conclusion

• Our initial goal was to verify our argumentation
  model on Estonian human-human spoken dialogues.
• Travel bureau dialogues were chosen from the
  Estonian dialogue corpus with the aim to find out
  communicative strategies and ways of
  argumentation which are used by a travel agent to
  force a client to book a trip.
• It turned out that there are very few dialogues
  where a travel agent actually tries to influence
  a client.
• The reason can be that the communicative goal of
  the client is only to get information, and not to
  book a trip during the conversation.
• Next aim to collect agreement negotiation
  dialogues
• calls to an education company where an agent
  proposes courses to a client