Argumentation in Agent Systems Part 2: Dialogue

1
Argumentation in Agent SystemsPart 2Dialogue

• Henry Prakken
• EASSS-07
• 31-08-2007

2
Why study argumentation in agent technology?

• For internal reasoning of single agents
• Reasoning about beliefs, goals, intentions etc
  often is defeasible
• For interaction between multiple agents
• Information exchange involves explanation
• Collaboration and negotiation involve conflict of
  opinion and persuasion

3
Overview

• Recent trends in argumentation logics
• Argument schemes
• Epistemic vs. practical reasoning
• Argumentation in dialogue
• Dialogue game approach
• Types of dialogues
• How they involve argumentation
• The notion of commitment
• Some dialogue systems
• Agent behaviour in dialogues
• Research issues

4
Argument schemes general form

• The same as logical inference rules
• But also critical questions
• Pointers to undercutters

Premise 1, , Premise n Therefore (presumably),
conclusion
5
Statistical syllogism

• P and if P then usually Q is a reason to believe
  that Q
• Birds usually fly
• Critical question subproperty defeater?
• Conflicting generalisation about an exceptional
  class
• Penguins dont fly

6
Normative syllogism

• P and if P then as a rule Q is a reason to accept
  that Q
• Critical question are there exceptions?
• How does a lawyer argue for exceptions to a rule?
• Say legislation makes an exception
• Say it is motivated by the rules purpose
• Find an overruling principle
• Argue that rule application has bad consequences

7
Witness testimony

• Critical questions
• Is W sincere? (veracity)
• Did W really see P? (objectivity)
• Did P occur? (observational sensitivity)

Witness W says P Therefore (presumably), P
8
Temporal persistence

• Critical questions
• Was P known to be false between T1 and T2?
• Is the gap between T1 and T2 too long?

P is true at T1 and T2 gt T1 Therefore
(presumably), P is Still true at T2
9
Arguments from consequences

• Critical questions
• Does A also have bad consequences?
• Are there other ways to bring about the good
  consequences?

Action A brings about good consequences Therefore
(presumably), A should be done
10
Types of dialogues (Walton Krabbe)
11
Example

• P I offer you this Peugeot for 10000.
• P why do you reject my offer?
• P why are French cars no good?
• P why are French cars unsafe?
• P Meinwagen is biased since German car magazines
  usually are biased against French cars
• P why does Meinwagen have a very high
  reputation?.
• P OK, I accept your offer.

• O I reject your offer
• O since French cars are no good
• O since French cars are unsafe
• O since magazine Meinwagen says so
• O I concede that German car magazines usually
  are biased against French cars, but Meinwagen is
  not since it has a very high reputation.
• O OK, I retract that French cars are no good.
  Still I cannot pay 10.000 I offer 8.000.

12
Example (2)

• P I offer you this Peugeot for 10000.
• P why do you reject my offer?
• P why are French cars no good?
• P why are French cars unsafe?
• P Meinwagen is biased since German car magazines
  usually are biased against French cars
• P why does Meinwagen have a very high
  reputation?.
• P OK, I accept your offer.

• O I reject your offer
• O since French cars are no good
• O since French cars are unsafe
• O since magazine Meinwagen says so
• O I concede that German car magazines usually
  are biased against French cars, but Meinwagen is
  not since it has a very high reputation.
• O OK, I retract that French cars are no good.
  Still I cannot pay 10.000 I offer 8.000.

13
Example (3)

• P I offer you this Peugeot for 10000.
• P why do you reject my offer?
• P why are French cars no good?
• P why are French cars unsafe?
• P Meinwagen is biased since German car magazines
  usually are biased against French cars
• P why does Meinwagen have a very high
  reputation?.
• P OK, I accept your offer.

• O I reject your offer
• O since French cars are no good
• O since French cars are unsafe
• O since magazine Meinwagen says so
• O I concede that German car magazines usually
  are biased against French cars, but Meinwagen is
  not since it has a very high reputation.
• O OK, I retract that French cars are no good.
  Still I cannot pay 10.000 I offer 8.000.

14
Dialogue systems (according to Carlson 1983)

• Dialogue systems define the conditions under
  which an utterance is appropriate
• An utterance is appropriate if it furthers the
  goal of the dialogue in which it is made
• Appropriateness defined not at speech act level
  but at dialogue level
• Dialogue game approach

15
Dialogue game systems

• A dialogue purpose
• Participants (with roles)
• A communication language Lc
• With embedded topic language Lt and a logic for
  Lt
• A protocol for Lc
• Effect rules for Lc (commitment rules)
• Termination and outcome rules

16
Some history

• In philosophy formal dialectics
• (Hamblin 1970, MacKenzie 1979, Walton Krabbe
  1995, )
• Deductive setting
• In AI procedural defeasibility
• Loui (1998(1992)), Brewka (1994,2001)
• Adding counterarguments
• In AI Law dispute resolution
• (Gordon 1993, Bench-Capon 1998, Lodder 1999,
  Prakken 2000-2006, )
• Adding counterarguments and third parties
• In MAS agent interaction
• Parsons-Sierra-Jennings 1998, Amgoud-Maudet-Parson
  s 2000, McBurney-Parsons 2002,
• Adding agents

17
Persuasion

• Participants proponent (P) and opponent (O) of a
  dialogue topic t
• Dialogue goal resolve the conflict of opinion on
  t.
• Participants goals
• P wants O to accept t
• O wants P to give up t
• Typical speech acts
• Claim p, Concede p, retract p, Why p, p since S,
  

18
Information seeking

• Dialogue goal information exchange
• Agents goals learning(?)
• Typical speech acts
• Ask p, Tell p, Notell p,

19
Negotiation

• Dialogue goal agreement on reallocation of
  scarce resources
• Participants goals maximise individual gain
• Typical communication language
• Request p, Offer p, Accept p, Reject p,

20
Deliberation

• Participants any
• Dialogue goal resolve need for action
• Participants goals
• None initially
• Possible set of speech acts
• Propose, ask-justify, prefer, accept, reject,

21
Dialectical shifts to persuasion

• Information exchange explaining why something is
  the case or how I know it
• Persuasion over fact
• Negotiation explaining why offer is good for you
  or bad for me
• Persuasion over fact or action
• Deliberation explaining why proposal is good or
  bad for us
• Persuasion over fact or action

22
Commitment in dialogue

• Walton Krabbe (1995)
• General case commitment to action
• Special cases
• Commitment to action in dialogue (dialogical or
  propositional commitment)
• Commitment to action outside dialogue (social
  commitment)
• Negotiation and deliberation lead to social
  commitments
• Persuasion leads to dialogical commitments

23
Quality aspects of dialogue protocols

• Effectiveness does the protocol further the
  dialogue goal?
• Commitments
• Agents logical and dialogical consistency
• Efficiency (relevance, termination, ...)
• Fairness does the protocol respect the
  participants goals?
• Flexibility, opportunity,
• Public semantics can protocol compliance be
  externally observed?

24
Effectiveness vs fairness

• Relevance and efficiency moves should be related
  to the dialogue topic
• Relevance often enforced in rigid so efficient
  unique-move immediate response protocols
• But sometimes participants must have freedom to
  backtrack, to explore alternatives, to postpone
  responses,

25
Public semanticsCommitments in persuasion

• A participants publicly declared standpoints, so
  not the same as beliefs!
• Only commitments and dialogical behaviour should
  count for move legality
• Claim p is allowed only if you believe p
• vs.
• Claim p is allowed only if you are not committed
  to ?p and have not challenged p

26
Assertion/Acceptance attitudes

• Relative to speakers own knowledge!
• Confident/Thoughtful agent can assert/accept P
  iff he can construct an argument for P
• Careful/cautious agent can assert/accept P iff
  he can construct an argument for P and no
  stronger counterargument
• Thoughtful/skeptical agent can assert/accept P
  iff he can construct a justified argument for P
• If part of protocol, then protocol has no public
  semantics!

27
Two systems for persuasion dialogue

• Parsons, Wooldridge Amgoud
• Journal of Logic and Computation 13(2003)
• Prakken
• Journal of Logic and Computation 15(2005)

28
PWA languages, logic, agents

• Lc Claim p, Why p, Concede p, Claim S, Question
  p
• p ? Lt, S ? Lt
• Lt propositional
• Logic argumentation logic
• Arguments (S, p) such that
• S ? Lt, consistent
• S propositionally implies p
• Attack (S, p) attacks (S, p) iff
• ?p ? S and
• level(S) level(S)
• Semantics grounded
• Assumptions on agents
• Have a knowledge base KB ? Lt
• Have an assertion and acceptance attitude

29
PWA protocol

• W claims p
• B concedes if allowed, if not claims ?p if
  allowed or else challenges p
• If B claims ?p, then goto 2 with players roles
  reversed and ?p in place of p
• If B has challenged, then
• W claims S, an argument for p
• Goto 2 for each s ? S in turn.
• B concedes if allowed, or the dialogue
  terminates.
• Outcome do players agree at termination?

30
Example persuasion dialogue
P1 My car is safe. claim P2 Since it has an
airbag. argument P3 why does that not make my
car safe? challenge P4 Yes, that is what the
newspapers say, concession but that does not
prove anything, since newspapers are unreliable
sources of technological information
undercutter P5 OK, I was wrong that my car is
safe. retraction
O1 Why is your car safe? challenge O2 That is
true, concession but your car is still not safe
counterclaim O3 Since the newspapers recently
reported on airbags exploding without cause
rebuttal O4 Still your car is not safe, since
its maximum speed is very high. alternative
rebuttal
31
PWA example dialogue
P careful/cautious P1 claim safe P2 claim
airbag, airbag ? safe P3 claim airbag ?
safe
O thoughtful/cautious O1 why safe O2a concede
airbag O2b why airbag ? safe
P careful/cautious P1 claim safe. P2 why
?safe P3a concede newspaper P3b why newspaper
? ? safe
O confident/cautious O1 claim ?safe O2 claim
newspaper, newspaper ? ?
safe O3 claim newspaper ? ? safe
32
PWA characteristics

• Protocol
• multi-move
• (almost) unique-reply
• Deterministic in Lc
• Dialogues
• Short (no stepwise construction of arguments, no
  alternative replies)
• Only one side develops arguments
• Logic
• used for single agent check attitudes and
  construct argument

33
Prakken languages, logic, agents

• Lc Any, provided it has a reply structure
  (attacks surrenders)
• Lt any
• Logic argumentation logic
• Arguments trees of conclusive and/or defeasible
  inferences
• Attack depends on chosen logic
• Semantics grounded
• Assumptions on agents none.

34
Prakken example Lc (with reply structure)
35
Protocol variations

• Unique-vs multiple moves per turn
• Unique vs. multiple replies
• Immediate response or not

36
Prakken protocols (basic rules)

• Each noninitial move replies to some previous
  move of hearer
• Replying moves must be defined in Lc as a reply
  to their target
• Argue moves must respect underlying argumentation
  logic
• Termination if player to move has no legal moves
• Outcome what is dialogical status of initial
  move at termination?

37
Dialogical status of moves

• Each move in a dialogue is in or out
• A surrender is out,
• An attacker is
• in iff surrendered, else
• in iff all its attacking children are out

38
P1
O1-
P2-
P4
O2-
O3
P3
39
Functions of dialogical status

• Can determine winning
• Plaintiff wins iff at termination the initial
  claim is in defendant wins otherwise
• Can determine turntaking
• Turn shifts if dialogical status of initial move
  has changed
• Immediate response protocols (Loui 1998)
• Can be used in defining relevance

40
Relevant protocols

• A move must reply to a relevant target
• A target is relevant if changing its status
  changes the status of the initial claim
• Turn shifts if dialogical status of initial move
  has changed
• Immediate response protocols

41
P1
O1-
P2-
P4
O2-
O3
P3
42
P1
O1-
P2-
P4
O2
O3
P3-
O4
43
P1
O1-
P2-
P4
O2-
O3
P3
44
P1-
O1
P2-
P4-
O2-
O3
O4
P3
45
Prakken example dialogue
P1 claim safe
46
Prakken example dialogue
P1 claim safe
O1 why safe
47
Prakken example dialogue
P1 claim safe
O1 why safe
P2 safe since airbag, airbag ? safe
48
Prakken example dialogue
P1 claim safe
O1 why safe
P2 safe since airbag, airbag ? safe
O2a concede airbag
49
Prakken example dialogue
P1 claim safe
O1 why safe
P2 safe since airbag, airbag ? safe
O2a concede airbag
O2b ?safe since newspaper, newspaper ?
?safe
50
Prakken example dialogue
P1 claim safe
O1 why safe
P2 safe since airbag, airbag ? safe
O2a concede airbag
O2b ?safe since newspaper, newspaper ?
?safe
P3a concede newspaper
51
Prakken example dialogue
P1 claim safe
O1 why safe
P2 safe since airbag, airbag ? safe
O2a concede airbag
O2b ?safe since newspaper, newspaper ?
?safe
P3a concede newspaper
P3b so what since unreliable, unreliable
? so what
52
Prakken example dialogue
P1 claim safe
O1 why safe
P2 safe since airbag, airbag ? safe
O2a concede airbag
O2b ?safe since newspaper, newspaper ?
?safe
O3 ?safe since high speed, high speed ?
?safe
P3a concede newspaper
P3b so what since unreliable, unreliable
? so what
53
Prakken example dialogue
P1 claim safe
O1 why safe
P2 safe since airbag, airbag ? safe
P4 retract safe
O2a concede airbag
O2b ?safe since newspaper, newspaper ?
?safe
O3 ?safe since high speed, high speed ?
?safe
P3a concede newspaper
P3b so what since unreliable, unreliable
? so what
54
Argument graph

so what
unreliable
unreliable ? so what
55
Winning and logic

• A protocol should respect the underlying logic
• We want main claim is in iff it is implied by
  the exchanged information
• (except information that is disputed and not
  defended)
• Ensured in relevant protocols (under certain
  conditions)

56
Prakkens relevant protocols characteristics

• Protocol
• Multiple-move
• Multiple-reply
• Not deterministic in Lc
• Immediate-response
• Dialogues
• Can be long (stepwise construction of arguments,
  alternative replies
• Both sides can develop arguments
• Logic
• Used for single agent construct/attack arguments
• Used for outcome players jointly build
  dialectical graph

57
Filibustering

• Many two-party protocols allow obstructive
  behaviour
• P claim p
• O why p?
• P p since q
• O why q?
• P q since r
• O why r?
• ...

58
Possible sanctions

• Social sanctions
• I dont talk to you any more
• Shift of burden of proof by third party
• ...
• P q since r
• O why r?
• referee O, you must defend not-r!

59
Protocol design vs. agent design

• Can protocol designer rely on agent properties?
• Rationality
• Cooperativeness
• Social behaviour

60
Design of dialogical agents

• Assertion and acceptance attitudes (PWA)
• Model choice of move as planning / practical
  reasoning
• Amgoud 2006
• Apply game theory
• Roth 2007
• Much work remains to be done

61
Investigation of protocol properties(formal
proof of experimentation)

• Does protocol induce well-behaved dialogues?
  (is it fair and effective?)
• Do agent attitudes, external goals or social
  conventions induce well-behaved dialogues?
• If a claim is successfully defended, is it
  implied by
• The shared or joint commitments of all
  participants?
• The shared or joint beliefs of all participants?
• Do agent attitudes constrain or even predetermine
  the outcome?

62
Research issues

• Investigation of protocol properties
• Combinations of dialogue types
• Deliberation!
• Multi-party dialogues
• Protocol design vs agent design
• Embedding in social context
• A framework for dialogue games

63
Further reading

• Argumentation in logic
• H. Prakken G. Vreeswijk, Logics for defeasible
  argumentation (Handbook of Philosophical Logic,
  2nd edition)
• Mail henry_at_cs.uu.nl for a pdf copy
• Argumentation in dialogue
• H. Prakken, Formal systems for persuasion
  dialogue. The Knowledge Engineering Review
  21163-188, 2006.
• I. Rahwan et al., Argumentation-based
  negotiation, The Knowledge Engineering Review
  18343-375, 2003.
• For more resources see http//www.cs.uu.nl/people
  /henry/easss07.html