Mercier and Sperber (M&S) offer a provocative view of argumentation as reasoning for social purposes. Human reasoning, they suggest, is not the same as classical inference in the sense that in reasoning, the rationale for conclusions is available for reflection and hence for communication and discussion. This is an important distinction, but there are also grounds for believing that reflective reasoning has general benefits for any cognitive agent, not just for social interaction.
A domain in which these benefits are evident is reasoning and decision making in medicine. I have a long-standing interest in the cognitive mechanisms that support decision making and other high-level cognitive processes that underpin human expertise, and argumentation has acquired a central role in our work. Early approaches based on logical and probabilistic simulations of cognitive processes yielded promising results (Fox Reference Fox1980), but extending either model to capture the flexible and adaptive character of human thinking proved difficult. Among the reasons for this were that there was no representation of the rationale on which to reflect – to question prior conclusions or the relevance of evidence, for example.
Subsequent work has sought to address this. This research programme has focused on artificial intelligence (AI) rather than psychology, so my comments should be taken as complementary to the M&S hypothesis rather than directly addressing it. However, I will suggest that a cognitive agent, whether human or artificial, derives major benefits from being able to reflect on its mental states; its goals, intentions, justifications for its beliefs and so on (Das et al. Reference Das, Fox, Elsdon and Hammond1997; Fox & Das Reference Fox and Das2000; Fox et al. Reference Fox, Clark, Glowinski and O'Neil1990). Metacognitive capabilities confer flexibility and robustness whether an agent is acting alone or in concert with others.
Mercier and Sperber's (M&S's) distinction between inference, which they call “intuitive,” and reasoning, which affords “reflection,” may perhaps be clarified by a formal perspective. A standard way of formalizing inference systems is to provide a “signature” that specifies how one set of sentences (e.g., propositions) is entailed by another set of sentences (e.g., a database of propositions and rules). This is a typical inference signature:
That is to say: Conclusion can be validly inferred from Database under the axioms of inference system L.
Complex cognitive tasks like decision making and planning require a more complex signature. To emulate human clinical decision making, we sought a reasoning model in which general medical knowledge is applied to specific patient data by arguing the pros and cons of alternative ways of achieving clinical goals. This is summarized by the following signature.
In contrast to the atomic conclusion of the inference signature, this formulation makes the structure of arguments explicit. In LA, a Logic of Argument (Fox et al. Reference Fox, Krause, Elvang-Goransson, Heckerman and Mamdani1993), the structure distinguishes three things: the Claim (a tentative conclusion), Grounds (justification), and Qualifier (the confidence in the Claim warranted by the argument. As in classical decision theory, but not classical logic, collections of arguments can be aggregated within the LA framework to yield an overall measure of confidence in competing claims. For example, an agent may have multiple lines of argument for and against competing diagnoses or treatments, each of which increases or decreases overall confidence.
LA was developed for cognitive tasks like situation assessment, decision making, and planning, which often involve uncertainty. Uncertainty is modelled explicitly by means of the Qualifier and therefore permits reflection. A qualifier may indicate that an argument “supports” or “opposes” a claim, for example. In The Uses of Argument the philosopher Stephen Toulmin has also pointed out that people routinely use linguistic qualifiers such as “presumably…,” “possibly…,” “probably…,” and their lexical and affixal negative forms; linguistic qualifiers can be formalised as conditions for accepting claims based on collections of arguments (Elvang-Goransson et al. Reference Elvang-Goransson, Krause, Fox, Clarke, Kruse and Moral1993). Quantitative schemes for expressing argument strength, such as Bayesian representations (e.g., Oaksford & Chater [Reference Oaksford and Chater2009] discussion in BBS vol. 32) can also be accommodated within the framework (Fox Reference Fox2003; Fox et al. Reference Fox, Krause, Elvang-Goransson, Heckerman and Mamdani1993).
It is a truism that the more supporting (opposing) arguments there are for a claim, the more (less) confidence we should have in it, which we have called the evidential mode (Fox, in press). Another mode, dialectical argumentation, exploits the observation that discussion and debate also commonly involves “attacks” which rebut or undercut the arguments of other agents. Researchers in AI and computational logic are giving substantial attention to argumentation for modelling interactions and dialogues between cognitive agents (Besnard & Hunter Reference Besnard and Hunter2008). Argumentation theory may therefore offer insights into the kinds of social interactions that M&S are investigating.
Formal argumentation theory has practical applications. LA is the foundation of PROforma, a language for modelling cognitive agents (Fox & Das Reference Fox and Das2000; Fox et al. Reference Fox, Beveridge and Glasspool2003); which has been used to develop many practical decision tools, notably in medicine (OpenClinical 2001–2006). Argumentation theory may also help to clarify the philosophical and theoretical nature of somewhat vague notions like evidence, as this term is commonly used in legal, medical, scientific, and other kinds of reasoning and in everyday decision-making and evidence-based discussions (OpenClinical 2001–2006).
These practical uses of argumentation theory do not directly address M&S's proposition that human cognition has evolved to support argument-based reasoning, but the practical power of argumentation techniques seems to corroborate their belief that there are good evolutionary reasons why humans argue. I do not intend to make psychological claims or predictions based on the specifics of LA, however. The logic was not developed as a descriptive model, and more recent argumentation systems do not all comply with its signature. However the general approach may offer a framework for articulating and comparing psychological theories of inference and reasoning and the cognitive processes that depend on them.
It appears that there are significant convergences between M&S's psychological proposals and developments in AI and computer science which have been driven by interests in artificial cognitive systems. Argumentation is a new and versatile reasoning paradigm that is being studied in many different fields; it may have wide implications for general theories of cognition.
Mercier and Sperber (M&S) offer a provocative view of argumentation as reasoning for social purposes. Human reasoning, they suggest, is not the same as classical inference in the sense that in reasoning, the rationale for conclusions is available for reflection and hence for communication and discussion. This is an important distinction, but there are also grounds for believing that reflective reasoning has general benefits for any cognitive agent, not just for social interaction.
A domain in which these benefits are evident is reasoning and decision making in medicine. I have a long-standing interest in the cognitive mechanisms that support decision making and other high-level cognitive processes that underpin human expertise, and argumentation has acquired a central role in our work. Early approaches based on logical and probabilistic simulations of cognitive processes yielded promising results (Fox Reference Fox1980), but extending either model to capture the flexible and adaptive character of human thinking proved difficult. Among the reasons for this were that there was no representation of the rationale on which to reflect – to question prior conclusions or the relevance of evidence, for example.
Subsequent work has sought to address this. This research programme has focused on artificial intelligence (AI) rather than psychology, so my comments should be taken as complementary to the M&S hypothesis rather than directly addressing it. However, I will suggest that a cognitive agent, whether human or artificial, derives major benefits from being able to reflect on its mental states; its goals, intentions, justifications for its beliefs and so on (Das et al. Reference Das, Fox, Elsdon and Hammond1997; Fox & Das Reference Fox and Das2000; Fox et al. Reference Fox, Clark, Glowinski and O'Neil1990). Metacognitive capabilities confer flexibility and robustness whether an agent is acting alone or in concert with others.
Mercier and Sperber's (M&S's) distinction between inference, which they call “intuitive,” and reasoning, which affords “reflection,” may perhaps be clarified by a formal perspective. A standard way of formalizing inference systems is to provide a “signature” that specifies how one set of sentences (e.g., propositions) is entailed by another set of sentences (e.g., a database of propositions and rules). This is a typical inference signature:
That is to say: Conclusion can be validly inferred from Database under the axioms of inference system L.
Complex cognitive tasks like decision making and planning require a more complex signature. To emulate human clinical decision making, we sought a reasoning model in which general medical knowledge is applied to specific patient data by arguing the pros and cons of alternative ways of achieving clinical goals. This is summarized by the following signature.
In contrast to the atomic conclusion of the inference signature, this formulation makes the structure of arguments explicit. In LA, a Logic of Argument (Fox et al. Reference Fox, Krause, Elvang-Goransson, Heckerman and Mamdani1993), the structure distinguishes three things: the Claim (a tentative conclusion), Grounds (justification), and Qualifier (the confidence in the Claim warranted by the argument. As in classical decision theory, but not classical logic, collections of arguments can be aggregated within the LA framework to yield an overall measure of confidence in competing claims. For example, an agent may have multiple lines of argument for and against competing diagnoses or treatments, each of which increases or decreases overall confidence.
LA was developed for cognitive tasks like situation assessment, decision making, and planning, which often involve uncertainty. Uncertainty is modelled explicitly by means of the Qualifier and therefore permits reflection. A qualifier may indicate that an argument “supports” or “opposes” a claim, for example. In The Uses of Argument the philosopher Stephen Toulmin has also pointed out that people routinely use linguistic qualifiers such as “presumably…,” “possibly…,” “probably…,” and their lexical and affixal negative forms; linguistic qualifiers can be formalised as conditions for accepting claims based on collections of arguments (Elvang-Goransson et al. Reference Elvang-Goransson, Krause, Fox, Clarke, Kruse and Moral1993). Quantitative schemes for expressing argument strength, such as Bayesian representations (e.g., Oaksford & Chater [Reference Oaksford and Chater2009] discussion in BBS vol. 32) can also be accommodated within the framework (Fox Reference Fox2003; Fox et al. Reference Fox, Krause, Elvang-Goransson, Heckerman and Mamdani1993).
It is a truism that the more supporting (opposing) arguments there are for a claim, the more (less) confidence we should have in it, which we have called the evidential mode (Fox, in press). Another mode, dialectical argumentation, exploits the observation that discussion and debate also commonly involves “attacks” which rebut or undercut the arguments of other agents. Researchers in AI and computational logic are giving substantial attention to argumentation for modelling interactions and dialogues between cognitive agents (Besnard & Hunter Reference Besnard and Hunter2008). Argumentation theory may therefore offer insights into the kinds of social interactions that M&S are investigating.
Formal argumentation theory has practical applications. LA is the foundation of PROforma, a language for modelling cognitive agents (Fox & Das Reference Fox and Das2000; Fox et al. Reference Fox, Beveridge and Glasspool2003); which has been used to develop many practical decision tools, notably in medicine (OpenClinical 2001–2006). Argumentation theory may also help to clarify the philosophical and theoretical nature of somewhat vague notions like evidence, as this term is commonly used in legal, medical, scientific, and other kinds of reasoning and in everyday decision-making and evidence-based discussions (OpenClinical 2001–2006).
These practical uses of argumentation theory do not directly address M&S's proposition that human cognition has evolved to support argument-based reasoning, but the practical power of argumentation techniques seems to corroborate their belief that there are good evolutionary reasons why humans argue. I do not intend to make psychological claims or predictions based on the specifics of LA, however. The logic was not developed as a descriptive model, and more recent argumentation systems do not all comply with its signature. However the general approach may offer a framework for articulating and comparing psychological theories of inference and reasoning and the cognitive processes that depend on them.
It appears that there are significant convergences between M&S's psychological proposals and developments in AI and computer science which have been driven by interests in artificial cognitive systems. Argumentation is a new and versatile reasoning paradigm that is being studied in many different fields; it may have wide implications for general theories of cognition.