The Explanation Mechanism Of The Causal Belief Revision

Belief revision is a process during which a rational agent change the belief state from inconsistency into consistency. Specifically, it is a process in which people first detect an inconsitency between new information and the original belief system, then decide which of the belief set to be revised after finding out a certain part of the belief conflicts with the new information, and eventually to accept the new information so that to adapt to the new enviroment.Concerning the mechansim of the belief revision, two basic principles had been proposed in philosophy and artificial intelligence:minimal change principle and entrenchment account. The former pointed out, people would make the minimal or the most conservative change of the original belief after detecting an inconsistency so that to maintain as much of the original belief as they can (Gardenfors,1988; Harman, 1986; James,1907).The latter claimed that some beliefs might be more entrenched than others in people's minds and thus more resistant to being abandoned (Gardenfors,1992). Both of the two accounts have been criticized for being lack of a clear operational definition, a proper measurement of the minimal change and entrechment of the belief and above all, they are in shortage of enough supporting data.Mental model theory has been so far the first psychological account of belief revision, which claimed that, people represent beliefs by constructing sets of mental models in which each model represents a possiblity. Belief revision is built on the representation of the beliefs (Johnson-Laird,& Byrne,2002; Johnson-Laird, Girotto, & Legrenzi,2004). The theory distinguished three main computations, in each of which one principle is supposed to be employed when people reason from inconsistency to consistency:first people must detect an inconsistency within a set of propositions by using the principle of modeling consistency; Second, people use the mismatch principle to decide which of the original belief to be revised. Third, people use the principle of causal knowledge to generate an explantion so that they can eventually resolve the inconsistency (Johnson-Laird, Girotto,& Legrenzi,2004).The problems of the theory lie in threefolds:it adopts an all-or-nothing manner in the representation of the belief as well as in the belief revision process; it overstates the role of rationality in belief revision, and in the meantime, it may underestimate the role of explanation in the process of belief revision.The explanation hypothesis has been another psychological account of belief revision. It suggested people resolve an inconsistency by explaining its origin. Explanations would be introduced as a new element into the causal structure after an inconsistency arises and people revise their causal beliefs on the basis of the new structure (Walsh,& Sloman,2004,2008; Walsh, Johnson-Laird,2009). The hypothesis fails to investigate the process of the belief revision, neither does it systematically build its theoretical structure on sufficient data.Based on the previous belief revision studies, this research focused on the mechanism of belief revision by answering two basic questions-how explanations determine belief revision response and when explanations are generated. Consisting of 5 experiments, the present research proceeded gradually to investigate the representation of beliefs, the computations people might have in the process of causal belief revision and the mechanism of causal belief revision. By testing the previous models of belief revision in our results, we intend to develop a psychological model of causal belief revision eventually.To tentatively investigate whether the availability of explanation influences the belief revision, we manipulated the causal structure and the presentation order of the new information and the categorical premises in a classical belief revision task in which participants were asked to choose to revise either the main premise or the categorical premise after detecting an inconsistency. The two independent variables were causal structure and the order of the categorical premise and facts. The former included positive structure (A causes B) and negative structure (A prevents B) while the latter included Modus Ponens order (the categorical premise was presented before the facts, A-not B) and Modus Tollens order (the categorical premise was presented after the facts, not B-A). The dependent variables were the scores of choices on revising the major premise. The results showed that people had a strong tendency to revise the main premise, which was even more obvious in the prevent structure than in the cause one. Inconsistent with what the mental model theory (Hasson,& Johnson-Laird,2003) predicted, the MP and MT orders didn't have an effect on people's belief revsion, and in the meantime, there was no interaction between the two variables. All the results are in favor of explanation hypothesis instead of mental model theory.To further investigate the decisive role of the availability of explanation in belief revision, we carried out two experiments to directly manipulate the availability of explanations in study two. Participants had to make one of the four choices after detecting an inconsistency:to revise the main premise, to revise the categorical premise, to revise both, to revise none of them. Scores for each choice were recorded. Hypothetical as well as familiar materials were employed in the two experiments. The independent variables were:the availability of the explanations (disablers) for different premises (three levels in experiment 2:explanations for the major premise, explanations for the minor premise and explanations for both; two levels in experiment 3:explanations for the major premise, explanations for the minor premise). Particpants had to give an explanation for the inconsistency after making the choice. The results showed that the manipuplation of the availability of the explanations for a certain premise had a direct effect on the belief revision results regardless of the nature of the materials and the type of the tasks, specifically, increasing the availability of disablers for a premise enhanced the tendence to revise the premise. Furthermore, the consistent response patterns in the hypothetical and familiar materials in the two experiments cast doubt over the entrenchment account.The previous two studies functionally prove the availability of explanations determine the way people revise their beliefs. In a causal relation, the cause must occur before the effect. So we have to find out the temporal order of the generation of explanation and belief revision response to finally determine the causal relation between explanations and belief revision responses. In study 3, we intended to examine the stage when people give an explanation.Two experiments were included. In experiment 4, the independent variables were: the strength of the major premises (Those included few disablers were recruited as high-plausible premises; those included many disablers were recruited as low-plausible premises) and the time to give an explanation (to explain before the presentation of the major premise; to explain after the presentation of the inconsistent new information). The dependent variables were:probability judgment of the major premise before and after generating an explanation:p1, p2. The results showed that, asking for explanation before the presentation of the major premise significantly lowered the probability judgment of the high-plausible premises but the same action only slightly brought down the probablity judgment of the low-plausible premises (pi). The inconsistent information given later generally brought down people's confidence of the major premise (p2) regardless of the strength of the major premises and the the time of generating an explanation. However, people lowered the probability more in the high-plausible premises than in the low-plausible premises.The findings in experiment 4 indicated that explanations might be included in the reprentation of the low-plausible premises while for the high-plausible premses, the availability of explanations is so much lower that people have to be forced to generate an explanation. Furthermore, it seems the process of evaluating the explanation is dynamic.In experiment 5, we used high-plausible premises (mostly daily used rules) with a common cause structure (A causes B; A causes C). People were asked to judge the probability of an effect given the presence of the cause and the presence or absence of another effect (P(C/A,-B) or P(C/A,B)). The independent variables were:consistency of the new information (new information that was consistent with the original propositions; new information that was inconsistent with the original propositions) and relevance of the explanation questions (those questions that asked for an explanation concerning the consistency of the new information were labled as relevant questions, otherwise they were labled as irrelevant questions). The dependent variables were:probability judgment of one effect given the presence of the cause (P(C/A)); probability judgment of the effect given the presence of the cause and the absence (or the presence) of another effect (P(C/A,-B) or P(C/A,B)); Reaction time to read the new information (RT1); Reaction time to judge the probability of P(C/A,-B) or P(C/A, B) (RT2). The reults showed that, people reduced the probability judgment when the new information was negative and increased the probablity judgment when the new information was positive. As for the reaction time, people spent more time reading the negative information than reading the positive information no matter the following explanation question was relevant or not. There was no signifant difference of the time to judge P(C/A,-B) when the new information was inconsistent between the two types of explanatin questions. But it takes longer for people to judge P(C/A,B) when the following question was relevant than when it was irrelevant. The findings in experiment 5 indicated that people would search for an explanation when they are faced with inconsistent information but they won't consider the causal mechanism until they are asked to judge P(C/A,B) when they are faced with positive information.In study 3, we found out that people either included disablers in the representation of the original beliefs or generated an explanation when they were reading the new information, thus we assume that explanations are usually generated before making a belief revision decision, eventually confirming the causal relation between the availability of explanations and the belief revision responses.The three stuides above couldn't be accounted by the previous models, we thus proposed the explanation hypothesis on the basis of causal model theory:the representation as well as the causal belief revision process is based on a probabilistic manner instead of an all-or-nothing manner; People hardly reject a premise in the face of some inconsistent information, rather, they choose to doubt a premise. Three computations might be involved in belief revision after inconsistent information arises: people first detect an inconsistency, and then they generate an explanation; finally they reassign the probability of the major premise based on the explanations they have in their mind. The availability of explanation detemines how people revise their beliefs. For the low-plausbility beliefs, explanations might be included in the representation of the major premise while for the high-plausbility beliefs, explanation are generated after they are confronted with inconsisten information.