Argument Analysis Using Bayesian Networks

Transcription

1 24/3/2011 1/15 Argument Analysis Using Bayesian Networks Kevin B. Korb and Erik Nyberg Clayton School of IT Monash University

2 24/3/2011 2/15 Argument Analysis Seven Step Program Michael Scriven in Reasoning (1980): 1 Clarify meanings 2 Identify premises and conclusions 3 Graph the argument structure 4 Make it valid: counterexample, find missing premises 5 Counterargue: criticize premises 6 Introduce alternative arguments 7 Evaluate

3 24/3/2011 2/15 Argument Analysis Seven Step Program Michael Scriven in Reasoning (1980): 1 Clarify meanings 2 Identify premises and conclusions 3 Graph the argument structure 4 Make it valid: counterexample, find missing premises 5 Counterargue: criticize premises 6 Introduce alternative arguments 7 Evaluate

4 24/3/2011 2/15 Argument Analysis Seven Step Program Michael Scriven in Reasoning (1980): 1 Clarify meanings 2 Identify premises and conclusions 3 Graph the argument structure 4 Make it valid: counterexample, find missing premises 5 Counterargue: criticize premises 6 Introduce alternative arguments 7 Evaluate

5 24/3/2011 2/15 Argument Analysis Seven Step Program Michael Scriven in Reasoning (1980): 1 Clarify meanings 2 Identify premises and conclusions 3 Graph the argument structure 4 Make it valid: counterexample, find missing premises 5 Counterargue: criticize premises 6 Introduce alternative arguments 7 Evaluate

6 24/3/2011 2/15 Argument Analysis Seven Step Program Michael Scriven in Reasoning (1980): 1 Clarify meanings 2 Identify premises and conclusions 3 Graph the argument structure 4 Make it valid: counterexample, find missing premises 5 Counterargue: criticize premises 6 Introduce alternative arguments 7 Evaluate

7 24/3/2011 2/15 Argument Analysis Seven Step Program Michael Scriven in Reasoning (1980): 1 Clarify meanings 2 Identify premises and conclusions 3 Graph the argument structure 4 Make it valid: counterexample, find missing premises 5 Counterargue: criticize premises 6 Introduce alternative arguments 7 Evaluate

8 24/3/2011 2/15 Argument Analysis Seven Step Program Michael Scriven in Reasoning (1980): 1 Clarify meanings 2 Identify premises and conclusions 3 Graph the argument structure 4 Make it valid: counterexample, find missing premises 5 Counterargue: criticize premises 6 Introduce alternative arguments 7 Evaluate

9 24/3/2011 2/15 Argument Analysis Seven Step Program Michael Scriven in Reasoning (1980): 1 Clarify meanings 2 Identify premises and conclusions 3 Graph the argument structure 4 Make it valid: counterexample, find missing premises 5 Counterargue: criticize premises 6 Introduce alternative arguments 7 Evaluate

10 24/3/2011 3/15 Argument Mapping In the meantime, argument mapping software has become popular and is reasonably effective. E.g., Tim van Gelder s Rationale (

11 24/3/2011 4/15 Argument Mapping Problems All relations are qualitative Qualitative relations can be represented quantitatively, but not vice versa In particular, there is no concept of strength of belief, imposing cognitive load on the user No possibility of representing interaction effects. E.g., consider a defective car: You can get away with faulty electrics (e.g., some sparks flying) You can get away with a leaky fuel line You can t get away with both!

12 24/3/2011 4/15 Argument Mapping Problems All relations are qualitative Qualitative relations can be represented quantitatively, but not vice versa In particular, there is no concept of strength of belief, imposing cognitive load on the user No possibility of representing interaction effects. E.g., consider a defective car: You can get away with faulty electrics (e.g., some sparks flying) You can get away with a leaky fuel line You can t get away with both!

13 24/3/2011 4/15 Argument Mapping Problems All relations are qualitative Qualitative relations can be represented quantitatively, but not vice versa In particular, there is no concept of strength of belief, imposing cognitive load on the user No possibility of representing interaction effects. E.g., consider a defective car: You can get away with faulty electrics (e.g., some sparks flying) You can get away with a leaky fuel line You can t get away with both!

14 24/3/2011 4/15 Argument Mapping Problems All relations are qualitative Qualitative relations can be represented quantitatively, but not vice versa In particular, there is no concept of strength of belief, imposing cognitive load on the user No possibility of representing interaction effects. E.g., consider a defective car: You can get away with faulty electrics (e.g., some sparks flying) You can get away with a leaky fuel line You can t get away with both!

15 24/3/2011 4/15 Argument Mapping Problems All relations are qualitative Qualitative relations can be represented quantitatively, but not vice versa In particular, there is no concept of strength of belief, imposing cognitive load on the user No possibility of representing interaction effects. E.g., consider a defective car: You can get away with faulty electrics (e.g., some sparks flying) You can get away with a leaky fuel line You can t get away with both!

16 24/3/2011 4/15 Argument Mapping Problems All relations are qualitative Qualitative relations can be represented quantitatively, but not vice versa In particular, there is no concept of strength of belief, imposing cognitive load on the user No possibility of representing interaction effects. E.g., consider a defective car: You can get away with faulty electrics (e.g., some sparks flying) You can get away with a leaky fuel line You can t get away with both!

17 24/3/2011 4/15 Argument Mapping Problems All relations are qualitative Qualitative relations can be represented quantitatively, but not vice versa In particular, there is no concept of strength of belief, imposing cognitive load on the user No possibility of representing interaction effects. E.g., consider a defective car: You can get away with faulty electrics (e.g., some sparks flying) You can get away with a leaky fuel line You can t get away with both!

18 24/3/2011 5/15 A Naive Bayes Approach Recently, Peter Sturrock has championed Naive Bayes models in his scientific approach to reasoning (see

19 24/3/2011 6/15 A Naive Bayes Approach Some problems: Dependencies between evidence items go missing These are multiplied by virtual duplications of evidence Result: Sturrock endorses estimates that Shakespeare was from Stratford from about to Neil Thomason: If you chose a random human at the time, she or he d be more far likely to be the author than the guy from Stratford!!

20 24/3/2011 6/15 A Naive Bayes Approach Some problems: Dependencies between evidence items go missing These are multiplied by virtual duplications of evidence Result: Sturrock endorses estimates that Shakespeare was from Stratford from about to Neil Thomason: If you chose a random human at the time, she or he d be more far likely to be the author than the guy from Stratford!!

21 24/3/2011 6/15 A Naive Bayes Approach Some problems: Dependencies between evidence items go missing These are multiplied by virtual duplications of evidence Result: Sturrock endorses estimates that Shakespeare was from Stratford from about to Neil Thomason: If you chose a random human at the time, she or he d be more far likely to be the author than the guy from Stratford!!

22 24/3/2011 6/15 A Naive Bayes Approach Some problems: Dependencies between evidence items go missing These are multiplied by virtual duplications of evidence Result: Sturrock endorses estimates that Shakespeare was from Stratford from about to Neil Thomason: If you chose a random human at the time, she or he d be more far likely to be the author than the guy from Stratford!!

23 24/3/2011 7/15 A Naive Bayes Approach Inspired by Sturrock s work, Tim van Gelder is working on a GUI for interactive argument analysis. Here s screenshot from his prototype:

24 24/3/2011 8/15 A Bayesian Net Approach Advantages Can represent degrees of argument strength Provides graphical representation for: premises, missing premises, conclusions Supports interactions, dependencies BN tools allow for sensitivity analysis, hypothetical reasoning

25 24/3/2011 8/15 A Bayesian Net Approach Advantages Can represent degrees of argument strength Provides graphical representation for: premises, missing premises, conclusions Supports interactions, dependencies BN tools allow for sensitivity analysis, hypothetical reasoning

26 24/3/2011 8/15 A Bayesian Net Approach Advantages Can represent degrees of argument strength Provides graphical representation for: premises, missing premises, conclusions Supports interactions, dependencies BN tools allow for sensitivity analysis, hypothetical reasoning

27 24/3/2011 8/15 A Bayesian Net Approach Advantages Can represent degrees of argument strength Provides graphical representation for: premises, missing premises, conclusions Supports interactions, dependencies BN tools allow for sensitivity analysis, hypothetical reasoning

28 24/3/2011 9/15 A Bayesian Net Approach We propose using Bayesian nets to model arguments, instead. Here s a possible model of the breast cancer argument:

29 24/3/ /15 A Bayesian Net Approach The argument from background only runs against cancer:

30 24/3/ /15 A Bayesian Net Approach The argument from positive self-examination also runs against cancer:

31 24/3/ /15 A Bayesian Net Approach To get an argument for cancer, we have to go all the way:

32 24/3/ /15 A Bayesian Net Approach Disadvantages Many people are allergic to numbers The point isn t to find exact numbers, but numbers that convey appropriate strengths. Use sensitivity analysis to find reasonable ranges. Nodes need to express simple propositions. Avoid internal quantifiers, probability qualifiers. Abbott will lose the next election Not: Abbott will likely lose the next election Dense networks lead to parameter explosion. KISS. Best starting point likely to be a causal model. The best argument may look different.

33 24/3/ /15 A Bayesian Net Approach Disadvantages Many people are allergic to numbers The point isn t to find exact numbers, but numbers that convey appropriate strengths. Use sensitivity analysis to find reasonable ranges. Nodes need to express simple propositions. Avoid internal quantifiers, probability qualifiers. Abbott will lose the next election Not: Abbott will likely lose the next election Dense networks lead to parameter explosion. KISS. Best starting point likely to be a causal model. The best argument may look different.

34 24/3/ /15 A Bayesian Net Approach Disadvantages Many people are allergic to numbers The point isn t to find exact numbers, but numbers that convey appropriate strengths. Use sensitivity analysis to find reasonable ranges. Nodes need to express simple propositions. Avoid internal quantifiers, probability qualifiers. Abbott will lose the next election Not: Abbott will likely lose the next election Dense networks lead to parameter explosion. KISS. Best starting point likely to be a causal model. The best argument may look different.

35 24/3/ /15 A Bayesian Net Approach Disadvantages Many people are allergic to numbers The point isn t to find exact numbers, but numbers that convey appropriate strengths. Use sensitivity analysis to find reasonable ranges. Nodes need to express simple propositions. Avoid internal quantifiers, probability qualifiers. Abbott will lose the next election Not: Abbott will likely lose the next election Dense networks lead to parameter explosion. KISS. Best starting point likely to be a causal model. The best argument may look different.

36 24/3/ /15 A Bayesian Net Approach Disadvantages Many people are allergic to numbers The point isn t to find exact numbers, but numbers that convey appropriate strengths. Use sensitivity analysis to find reasonable ranges. Nodes need to express simple propositions. Avoid internal quantifiers, probability qualifiers. Abbott will lose the next election Not: Abbott will likely lose the next election Dense networks lead to parameter explosion. KISS. Best starting point likely to be a causal model. The best argument may look different.

37 24/3/ /15 A Bayesian Net Approach Disadvantages Many people are allergic to numbers The point isn t to find exact numbers, but numbers that convey appropriate strengths. Use sensitivity analysis to find reasonable ranges. Nodes need to express simple propositions. Avoid internal quantifiers, probability qualifiers. Abbott will lose the next election Not: Abbott will likely lose the next election Dense networks lead to parameter explosion. KISS. Best starting point likely to be a causal model. The best argument may look different.

38 24/3/ /15 A Bayesian Net Approach Disadvantages Many people are allergic to numbers The point isn t to find exact numbers, but numbers that convey appropriate strengths. Use sensitivity analysis to find reasonable ranges. Nodes need to express simple propositions. Avoid internal quantifiers, probability qualifiers. Abbott will lose the next election Not: Abbott will likely lose the next election Dense networks lead to parameter explosion. KISS. Best starting point likely to be a causal model. The best argument may look different.

39 24/3/ /15 A GUI for argument/decision making To suppress many of the complexities, we propose adapting the van Gelder s GUI. Instead of a one-size fits all GUI (possible with naive Bayes), we can taylor GUIs to arguments and decision problems.

40 24/3/ /15 Readings BayesianWatch Lewis Carroll. (1895). What the Tortoise Said to Achilles. Mind 4, Critical Thinking on the Web. A Fisher (1988). The Logic of Real Arguments. Cambridge. Thomas Gilovich, Dale Griffin, Daniel Kahneman (Eds.) (2002). Heuristics and Biases: The Psychology of Intuitive Judgment. Cambridge Univ. K Korb (2004). Bayesian Informal Logic and Fallacy. Informal Logic, 23, K. Popper. (1963). Conjectures and Refutations. London: Routledge. K. Popper. (1945). The Open Society and Its Enemies (2 Vols). London: Routledge. M Scriven (1980). Reasoning. McGraw Hill. P. Sturrock (2013). AKA Shakespeare: A Scientific Approach to the Authorship Question. Exoscience Publishing.