Argumentation, Dialogue Theory, and Probability Modeling: Alternative Frameworks for Argumentation Research in Education

Abstract

Toulmin's model of argumentation, developed in 1958, has guided much argumentation research in education. However, argumentation theory in philosophy and cognitive science has advanced considerably since 1958. There are currently several alternative frameworks of argumentation that can be useful for both research and practice in education. These frameworks include Walton's dialogue theory and Bayesian models of everyday arguments. This article reviews and evaluates these frameworks and shows how each can be applied instructionally (e.g., through the teaching of critical questions or probability modeling) and, from a research standpoint, in evaluating the content and quality of informal arguments. It is concluded that attention to these and other contemporary argumentation frameworks can help move the study of argumentation in education forward.

ACKNOWLEDGMENTS

I thank Ivan V. Ivanov for reviewing previous versions of the article.

Notes

¹ Duschl (2008) coded for several argument schemes during middle school science discussion (e.g., evidence to hypothesis) and found that students exposed to a science intervention focused on developing scientific reasoning used argumentation schemes more frequently than a comparison group. However, the rank order of how often each scheme was used was the same for both groups, suggesting that most students possess many of these schemes in a rudimentary form. According to Duschl (2008), instruction is needed so that students learn the “epistemic criteria” for applying the schemes and judging the strength of arguments. Duschl did not directly discuss the notion of critical questions, but such questions presumably reflect, at least in part, epistemic criteria.

² Royall (1997) conceived of the evidential likelihood more as the ratio of these two likelihoods when comparing two hypotheses, , which is the odds of H₁ when P(H ₂)=P(1−H ₁), that is, when P(H ₁)+P(H ₂)=1. In other words, when the two hypotheses are exhaustive and mutually exclusive.

³This analysis could be faulted, however, because it assumes that the various component probabilities are statistically independent. If the component probabilities are statistically dependent, meaning that the occurrence of one event makes another more or less likely, then the following formula must be used instead of simply multiplying the probabilities:

It is very plausible that an individual is more likely to “know something relevant” if they are a member of a terrorist organization. A more appropriate probability structure would involve the term “probability knows something relevant given that an individual is a member of a terrorist organization,” that is, P(B | A). The statistical dependency is likely to increase the joint probability somewhat.

   Modeling statistical dependencies will make the probability structure more complex. These dependencies can be modeled with graphical methods, specifically causal Bayesian networks (Pearl, 2000), although these networks can, given too many dependencies, become computationally intractable, unless very computationally expensive statistical techniques are used (Glymour, 2001). For analyzing arguments, it is likely not productive to consider all possible statistical dependencies, but if a plausible argument can be made that a particular dependency is strong and important, then an alternative to an independence model should be considered.

⁴Subjective and objective Bayesians will likely disagree on the merits of small sample analysis; a lot will depend on the confidence one can place in the likelihoods.