Evaluating Explanations in Law, Science, and Everyday Life

This article reviews a theory of explanatory coherence that provides a psychologically plausible account of how people evaluate competing explanations. The theory is implemented in a computational model that uses simple artificial neural networks to simulate many important cases of scientific and legal reasoning. Current research directions include extensions to emotional thinking and implementation in more biologically realistic neural networks. KEYWORDS—explanation; coherence; neural networks; legal reasoning; emotion In CSI and other television crime shows, investigators collect evidence in order to determine the causes of a crime. For example, if a young woman ismurdered, the policemay consider as suspects the woman’s boyfriend and her father. Inferences about who is the most likely culprit will be based on which hypothesis—that the boyfriend did it or that the father did it—fits best with all the available evidence. These hypotheses provide possible explanations of the evidence; for example the hypothesis that the boyfriend was the murderer may explain why his fingerprints are on the murder weapon. Conclusions about who the actual criminal was and who was innocent depend on evaluating competing explanations of the evidence. This kind of explanatory inference is ubiquitous in human thinking, ranging frommechanical repair tomedical diagnosis to scientific theorizing. When your car fails to start, you consider alternative explanations such as that it is out of gas or that the battery is dead. In medicine, a physician considers possible diseases that would explain a patient’s symptoms and bases a treatment plan on what he or she thinks is the most plausible diagnosis. Psychologists publishing theoretical papers often offer sets of hypotheses that they contend provide better explanations of the results of experiments than alternative theories do. Explanation evaluation is amental process that is important in many areas of psychology. Cognitive psychologists have investigated causal reasoning, which often requires a person to determine the most likely cause of a surprising event. Social psychologists have studied how people explain the behavior of others. Clinical psychologists are sometimes interested in the emotion-laden reasoning by which people construct explanations of their own situations. In all these kinds of cases, people’s thinking involves evaluating competing explanations of what they observe. But explanation evaluation is not simply a matter of determining which of two or more competing hypotheses fits best with the evidence. We may also need to consider how hypotheses fit with each other, particularly when one hypothesis provides an explanation of another. This layering of hypotheses is particularly evident in legal reasoning when questions of motive are salient. Crime investigators considering whether the boyfriend or the father is the more likely murderer will naturally consider possible motives that might explain why one of them would have wanted to kill the young woman. Hence the cognitive process of explanation evaluation must consider the fit of hypotheses with each other as well as with the evidence, so that inference involves coming up with the overall most coherent picture of what happened. This article reviews a theory of explanatory coherence that provides a psychologically plausible account of how people evaluate competing explanations. After sketching the theory, I describe how it is implemented in a computational model that uses a simple artificial neural network to evaluate competing explanations. This model has been applied to many important cases of scientific and legal reasoning. Finally, I describe current directions in the development and application of the theory of explanatory coherence, including connections with emotional thinking and implementation in more biologically realistic neural networks. EXPLANATORY COHERENCE: THE THEORY Table 1 lists seven principles that concisely state the theory of explanatory coherence (Thagard, 1989, 1992, 2000). These C D I R 4 2 4 B Dispatch: 8.5.06 Journal: CDIR CE: Blackwell Journal Name Manuscript No. Author Received: No. of pages: 5 PE: Sarvanan/Mini Address correspondence to Paul Thagard, Philosophy Department, University of Waterloo, Waterloo, ON N2L 3G1, Canada; e-mail: [email protected]. CURRENT DIRECTIONS IN PSYCHOLOGICAL SCIENCE Volume 15—Number 3 141 Copyright r 2006 Association for Psychological Science CDIR 424 (b w us C D IR 4 24 1 41 ..1 45 .P D F 5/ 8/ 20 06 1 :1 1: 13 P M 8 20 15 B yt es 5 P A G ES O pe ra to r= ) principles are rather abstract, so let me explain them in terms of the legal example already introduced. The hypothesis that the boyfriend killed the woman explains the evidence that the woman is dead, so the hypothesis and the evidence cohere with each other, in accord with principle E2, Explanation. Although the relation between the hypothesis and evidence is asymmetrical, with the former explaining the latter and not vice versa, the coherence relation between them is symmetrical: They hang together equally, as indicated by principle E1, Symmetry. Principle E2 also allows the possibility of hypotheses explaining each other, as when the hypothesis that the boyfriend is the murderer is explained by the motive that he was jealous. Explanation can involve multiple hypotheses—for example that the boyfriend was both jealous and angry—that then cohere with each other. However, E2 includes a simplicity principle in clause (c), so that hypotheses that involve many hypotheses will have less coherence. For example, the theory that the womanwas killed by space aliens who arrived from Alpha Centauri and singled her out for execution because of her hair color requires multiple hypotheses that lack simplicity as well as independent support. Simplicity is a matter of explaining a lot with few assumptions. The theory of explanatory coherence is neutral about what constitutes an explanation, but I have argued independently that good explanations are based on causal mechanisms (Thagard, 1999). According to principle E3, Analogy, explanations can gain coherence by virtue of being analogous to ones already accepted. For example, if the boyfriend had a past history of being jealously angry with girlfriends and assaulting them, then these cases provide analogies that make the hypothesis that the boyfriend did it more plausible in the current case. Principle E4, Data Priority, says that observational evidence gets a degree of coherence on its own, providing a degree of priority to such observations as that the woman is dead and the boyfriend’s fingerprints are on a knife found near the body. This principle does not require that observations be indubitable, but leaves open the possibility that observations could be found to be erroneous despite their initial degree of coherence. Principles E5 and E6 deal with competing hypotheses that are incoherent with each other. E5, Contradiction, handles the most straightforward case in which two hypotheses are logically contradictory; but typically the relation between competitors is looser, as captured in E6, Competition. Normally, we treat the hypothesis that the boyfriend was the murderer as competing with the hypothesis that the father did it, even though these are not contradictory: It is logically possible that the boyfriend and the father together killed the woman. But if there is reason to suspect that the boyfriend and the father acted together in a conspiracy, then the two hypotheses—the boyfriend did it and the father did it—are explanatorily connected, so they should be treated as coherent with each other rather than incoherent. Ordinarily, however, two hypotheses that independently explain evidence will be treated as competitors that are incoherent with each other. Finally, principle E7, Acceptance, states that we should accept and reject propositions on the basis of their overall coherence with each other. Because hypotheses and evidence can be coherent and incoherent with each other in many ways, E7 makes inference a highly complex and nonlinear process. We cannot simply accept the evidence and then accept a hypothesis and then reject its competitors, because evidence and competing hypotheses must all be evaluated together with respect to how they fit with each other. This makes explanation evaluation sound like a very mysterious holistic process, but I will now describe how a simple artificial neural network can perform the required computation. EXPLANATORY COHERENCE: THE MODEL The first step in implementing explanatory coherence computationally is to represent each proposition by a unit, a highly simplified artificial neuron that is connected to other units by excitatory and inhibitory links. As in real neurons, an excitatory link is one that enables one neuron to increase the firing of another, whereas an inhibitory link decreases firing. In the crime example, the hypothesis that the boyfriend is the murderer can TABLE 1 Principles of Explanatory Coherence