Running Head: A Relevance Theory of Induction

A framework theory, organized around the principle of relevance, is proposed for category-based reasoning. According to the relevance principle, people assume that the premises are informative with respect to the conclusions. This idea leads to the prediction that people will use causal scenarios and property reinforcement strategies in inductive reasoning. These predictions are contrasted with both existing models and normative logic. Judgments of argument strength were gathered in three different countries and the results showed the importance of both causal scenarios and property reinforcement in category-based inferences. The relation between the relevance framework and existing models of category-based inductive reasoning is discussed in the light of these findings. A Relevance Theory of Induction 3 A Relevance Theory of Induction Introduction One of the central functions of categorization is to support reasoning. Having categorized some entity as a bird, one may predict with reasonable confidence that it builds a nest, sings and can fly, though none of these inferences is certain. In addition, between-category relations may guide reasoning. For example, from the knowledge that robins have some enzyme in their blood one is likely to be more confident that sparrows also have this enzyme than that raccoons have this enzyme. The basis for this confidence may be that robins are more similar to sparrows than to raccoons or that robins and sparrows share a lower rank superordinate category (birds versus vertebrates) than do robins and raccoons. Recently, researchers have developed specific models for category-based reasoning and generated a range of distinctive reasoning phenomena (see Heit, 2000, for a review). These phenomena are quite robust when American college students are the research participants but at least some of them do not generalize well to other populations. To address these limitations, we will offer not so much a specific model but rather a framework theory organized around the principle of relevance. This theory is more abstract than many of its predecessors and one might imagine a number of implementations consistent with the relevance framework. Nonetheless, we will see that the relevance theory has testable implications. The rest of the paper is organized as follows. First, we briefly review two of the most influential models for induction, the Osherson, Smith, Wilkie, Lopez, and Shafir (1990) category-based induction model and Sloman’s (1993) feature-based induction model. Next, we turn to the question of the generality of reasoning phenomena and describe two, more abstract A Relevance Theory of Induction 4 approaches that may be able to address the question of generality. Then we offer a theory at an intermediate level of abstraction, the so-called relevance theory, and describe some tests of its implications. Finally, we summarize and argue that there are benefits from approaching induction from a number of levels of analysis. The similarity-coverage model (SCM). The Osherson, et al. (1990) model of induction is driven by two related notions, similarity and coverage. Similarity refers to the assumption that, all else being equal, people are more likely to extend a predicate from a base premise to a target premise to the extent that the target category is similar to the base category. Given the premise that dogs have sesamoid bones, we are more likely to think that wolves have sesamoid bones than that cows do. The SCM also assumes that judgments may be partially based on the similarity of the premise category to examples of the lowest level superordinate category that spans the premise and conclusion categories. Consider, for example, the premise “bears have sesamoid bones” and the conclusion “Therefore, all mammals have sesamoid bones”. According to the SCM, to evaluate this argument people would generate examples of the mammal category (e.g., dog, cow, wolves, horse, lion) and compute their similarity to the premise category, bear. In this example, the coverage would be the sum of the similarities of retrieved instances to bear. If the premise were that whales have sesamoid bones and the conclusion that all mammals do, then the same instance retrieval and similarity calculation process is assumed to operate. In this case the summed similarities or coverage would be less, because whales are atypical mammals and less similar on average to other mammals than are bears. This example illustrates that the SCM predicts typicality effects in reasoning because typical examples have better coverage than atypical examples. A Relevance Theory of Induction 5 Two of the best-studied phenomena associated with the SCM are two phenomena that rely on the notion of coverage, typicality and diversity. As we have seen, typicality effects in reasoning follow directly from the definition of typicality in terms of similarity to other category members (Rosch and Mervis, 1975). Diversity concerns coverage associated with multiple premise arguments. Consider, for example, the relative strength of the premises that crows and blackbirds have property X versus the premise that crows and ducks have property X for the conclusion that all birds have property X. In the SCM coverage is based on the average maximal similarity that examples of the category have to the premise examples. Crows and blackbirds are quite similar and the coverage provided by each of them will be redundant to that provided by the other. In contrast, ducks are different from crows and will have substantially greater similarity to a number of birds than will crows (e.g., geese, swans, loons, pelicans, gulls). This will produce better overall coverage. In short, the SCM predicts that two diverse premises will have greater induction strength for a category than two similar premises (Note, however, that two very different but atypical examples of a category, such as penguins and hummingbirds, may have poor overall coverage and therefore, coverage cannot be equated with dissimilarity of premises, see Osherson et al., 1990, pp. 199-200). The SCM is deceptively simple. It has only a single parameter reflecting the relative weight given to the similarity and coverage components. Given a set of category similarities, it can be used to generate a variety of both intuitive and counter-intuitive predictions that have received considerable support (see Osherson, et al., 1990). Feature-based Induction Model (FBIM). Sloman’s (1993) feature-based induction model also relies on the notions of (featural) similarity and (featural) coverage. The central idea is that A Relevance Theory of Induction 6 similarity is driven by matching and mismatching features and that an argument is strong to the extent that the premise and conclusion categories share features. A distinctive property of the FBIM is that is does not use category information in the sense that it does not distinguish between different levels of categorization. Instead, it assumes that all categories are represented in terms of features and that argument strength is based on feature overlap. It may seem that the FBIM is just the SCM with the notion of similarity being decomposed into featural matches and mismatches. But the FBIM has no notion of generating category examples and the fact that it treats a category as just a feature set leads to some unique predictions, predictions that have received support (e.g., Sloman, 1993, 1998). Although FBIM and SCM are distinct, for the present purposes we will treat them as providing more or less comparable accounts of typicality and diversity effects, phenomena to which we now turn. Typicality and diversity are very robust phenomena in American undergraduate study populations. But these results do not generalize well to other groups. López, Atran, Coley, Medin, and Smith (1997) used local mammals as stimuli to study induction among the Itza’ Maya. University of Michigan undergraduates’ reasoning about mammals of Michigan provided a control or comparison condition. The Itza’ Maya showed reliable typicality effects but either no diversity effects or below chance diversity effects. Undergraduates displayed strong typicality and diversity effects. Proffitt, Coley, and Medin (2000) studied different types of tree experts’ reasoning about trees. None of the groups showed typicality effects. Taxonomists showed reliable diversity effects but parks maintenance workers responded below chance on diversity probes. Bailenson, Shum, Atran, Medin, and Coley (2002) studied Itza’ Maya, USA bird experts’ and Northwestern University undergraduates’ categorizing and reasoning about birds of Illinois and birds of Guatemala. The Itza and the bird experts were not reliably above chance on either A Relevance Theory of Induction 7 typicality or diversity probes. Undergraduate responses and justifications strongly conformed to both typicality and diversity. In short, typicality and diversity effects are far from common in populations that have considerable knowledge concerning the domain of categories under study. Why don’t experts and Itza’ (who are themselves biological experts) produce clear typicality and diversity responding? The most salient reason is that often they are instead employing causal and ecological reasoning about the kinds in question. For example, Proffitt et al. (2000) found that tree experts often reason about arguments involving novel tree diseases in terms of how widely planted different kinds of trees are, their susceptibility to disease and so on. Lopez et al. (1997) noted very similar reasoning strategies among the Itza’ Maya. One might argue that these informants were not treating these predicates as truly blank properties but to take this stance artificially limits the potential scope of models of induction and risks a certain circularity (e.g., the SCM should apply only where the responses match its predictions). Furthermore, diversity-based reasoning isn’t absent in these populations but rather seems to be one of several strategies employed. An alternative approach to capturing this range of results is to broaden the scope of induction models. We now turn to two models for induction that do just that. Hypothesis-Based Induction. McDonald, Samuels and Rispoli (1996) proposed what they refer to as a hypothesis-based model of induction. They argue that induction may be guided more by theories or explanations rather than similarity itself. On this view, inductive strength may be based in part on whether the premises suggest alternative categories or hypotheses to the conclusion category given (these act as competing explanations; see also Sloman, 1994). They provide support for their framework by asking people to generate hypotheses or explanations and A Relevance Theory of Induction 8 showing that inductive confidence decreases when there are competing hypotheses (candidates for a conclusion category). A Bayesian model. Heit (1998, 2000) has taken a Bayesian approach to category-based induction. The idea is that people have expectations about the distributions of properties or features and that their judgments are based on these subjective distributions. Consider, for example, typicality effects in reasoning. The reasoner is assumed to consider those features that are unique to a premise, those that might hold for the premise and categories which are subordinate to or overlap with the conclusion category, and those features that match the conclusion category. The advantage that a typical premise has over an atypical premise is that it may have relatively fewer distinctive features and fewer features shared with overlapping or subordinate categories. The Bayesian model is similar in spirit to Sloman’s (1993) feature-based induction model, though they are far from equivalent (see also Griffiths & Tenenbaum, 2000). A nice feature of the Bayesian approach (and the hypothesis-based model as well) is that it provides for more flexibility in induction. People’s knowledge may lead them to have different expectations about the features relevant for induction. For example, Heit and Rubinstein’s (1994) finding that physiological features or predicates trigger different patterns of induction than behavioral predicates follows naturally from this framework. Depending on expectations about feature distributions, the Bayesian framework, like Sloman’s feature-based model, may provide an account for when diversity effects are or are not obtained. The Bayesian model may be evaluated in a manner analogous to the hypothesis-based induction model. For a set of premise categories and predicates one might obtain people’s judgments about common and distinctive features and then use these distributions to make predictions about reasoning phenomena. A close correspondence supports the model. Lack of A Relevance Theory of Induction 9 correspondence suggests either that the model is flawed or that the feature elicitation procedure is faulty. Analysis. We think that each of the models for induction so far proposed contains valuable insights. All of the models constrain their predictions by obtaining predictor measures (e.g., similarity judgments, featural distributions, hypotheses) and using them to predict patterns of reasoning. The SCM is perhaps the most constrained in that the similarity judgments may be collected in a task remote from the reasoning task. To the extent that relevant features and hypotheses are thought to depend on the predicates and specific combinations of premises and conclusions, then the predictor variables must be collected in a context very close to the actual reasoning task. Very likely there is a tradeoff---the closer the predictor task is to the predicted, the more accurate predictions should be. But it is also true that the closer the tasks are, the more open the framework is to the criticism that its account has a circular flavor. There seems to be something of a continuum. At one end, the SCM makes strong predictions but fails to capture some of the dynamic aspects of how people reason about categories. Bayesian and hypothesis-based models can address many of the more contextualized aspects of reasoning but are less able to make a priori predictions. In this paper we offer an intermediate level framework---our goal is to provide an account of the dynamic and contextdependent components of category-based reasoning by postulating some processing principles that fall under the broad umbrella of relevance. We turn to that now. Relevance Theory – An Overview The lack of generality of typicality and diversity effects beyond undergraduate populations represents a serious limitation of most current models of induction, which generally predict that these phenomena will be more robust than they are. One of our test sessions with a A Relevance Theory of Induction 10 tree expert provided the impetus for a shift toward a different framework theory. The expert was being given typicality probes such as the following: “Suppose we know that river birch get Disease X and that white oaks get Disease Y, which disease do you think is more likely to affect all trees?” In this case, the expert said Disease X, noting that river birches are very susceptible to disease; so, “if one gets it they all get it.” The very next probe involved the gingko tree and the expert choose the disease associated with it as more likely to affect all trees on the grounds that “Gingkos are so resistant to disease that if they get it, it must be a very powerful disease.” He then said that he felt as if he had just contradicted himself, but that nonetheless these seemed like the right answers. Normatively, this expert’s answers do not represent a contradiction. Instead, he appeared to be using the information that was most salient and accessible to guide his reasoning (on spontaneous feature listing tasks experts indicate that birches are notoriously susceptible to, and gingkos notoriously resistant to, diseases). Simply put, the expert was using the knowledge that he considered most relevant. We believe that Sperber and Wilson’s (1996) relevance theory provides a good framework for understanding category-based induction. Furthermore, it leads to a number of novel predictions that contrast with those of other models of induction. In relevance theory, relevance is seen as a property of inputs to cognitive processes: “An input is relevant to an individual at a certain time if processing this input yields cognitive effects. Examples of cognitive effects are the revision of previous beliefs, or the derivation of contextual conclusions, that is, conclusions that follow from the input taken together with previously available information. Such revisions or conclusions are A Relevance Theory of Induction 11 particularly relevant when they answer questions that the individual had in mind (or in an experimental situation, was presented with).” Van der Herst, Politzer &Sperber, (in press) In the Proffitt et al (2000) studies background knowledge about properties of trees and diseases presumably provides that basis for the sorts contextual conclusions mentioned by our tree expert. Van der Herst, et al, (in press) further elaborate: “Everything else being equal, the greater the cognitive effects achieved by processing an input, the greater its relevance. On the other hand, the greater the effort involved in processing an input, the lower the relevance.... One implication of the definition of relevance in terms of effect and effort is that salient information, everything else being equal, has greater relevance, given that accessing it requires less effort.” Potentially, there are two problems with relevance theory that may limit its applicability to studies of induction. One is that it is not possible to maximize two functions at once. In general, more effort should lead to more effect so it isn’t obvious how to trade off one for the other in determining relevance. The second, related problem is that relevance theory appears to subject to the same circularity criticism that we have raised with respect to Bayesian and hypothesis-based models. Although it is not possible to simultaneously maximize (least) effort and (greatest) effect, one can experimentally manipulate effort and effect to determine whether they have the sorts of consequences predicted by relevance theory. In the present paper our focus is on undergraduates. They generally have little background knowledge to bring to bear on the sorts of reasoning tasks we have used. Consequently, it is not surprising that they rely heavily on more abstract reasoning strategies. However, it may be possible to select probes related to the limited biological knowledge they have in order to vary what Sperber et al call effect. As we shall see, it is also A Relevance Theory of Induction 12 easy to experimentally manipulate effort. In the next few paragraphs we will outline how relevance theory may apply to category-based induction and then develop specific predictions for our studies. Relevance in category-based induction. The general idea is that the premises are assumed to be relevant to the conclusion(s). One motivation for this view is the fact that experiments take place in a social context and participants reasonably infer that the experimenter is being relevant and informative with respect to the inductive argument forms (cf. Grice, 1975). We also believe, however, that people may generally assume something like a principle of relevance or informativeness regardless of the source of observations. How does the principle of relevance constrain induction? We suggest that when a blank property or predicate is associated with some premise category, people tend to associate that property with the most distinctive or informative features or categories associated with the premise. For example, immediate superordinate categories generally should be more salient and relevant than more remote super-ordinates, because immediate superordinates are more unusual (have lower base rates) and are therefore more informative (in an information-theoretic sense) than remote superordinates. Note that informativeness in this case follows a principle of parsimony and that it is concordant with the Osherson et al.'s (1990) SCM in assuming that the lowest level super-ordinate capturing premise and conclusion categories is activated. Another way of thinking about relevance is to suggest that, when given an argument to evaluate, participants ask themselves why this particular premise (and not some other one) is given for the particular conclusion under consideration. For example, suppose one is given the premise that “Skunks have Property X.” According to the relevance framework, good candidates for what Property X might be related to are features that are distinctive of skunks; that is. A Relevance Theory of Induction 13 features that skunks have that similar mammals such as squirrels or muskrats do not have. Two possibilities that immediately come to mind are that they are striped and that they can create a very strong odor. The conclusion category may act as a further important constraint on assumed relevance. For example, if the conclusion is that “Zebras also have Property X,” then it becomes plausible that Property X is related to being striped and that the argument should be considered to be at least moderately strong. If the conclusion were instead, “Onions have property X,” a participant who assumes that the experimenter is following a relevance principle should be more likely to assume that Property X refers to odor rather than stripedness. Note also, that the argument going from zebras to skunks may be stronger that one going from skunks to zebras because skunks have two salient features and zebras may have only one, being striped (though perhaps being an African mammal is another one). Summing up so far, the relevance framework suggests two processing principles. One is that distinctive properties of premise categories are candidates for providing the relevant basis for induction. (To be sure, particular predicates can support or undermine candidates for relevance; if the premise were “Skunks weigh more than Martians”, then stripedness and odor clearly would be irrelevant.) The second idea is that comparing the premise and conclusion categories acts as a further constraint on relevance by either reinforcing or undermining candidates for relevance based on the premise categories considered by themselves. We further suggest that this same comparison process is used (for related ideas on the importance of comparison processes, see also Hahn & Chater, 1997; Medin, Goldstone & Gentner, 1993) if there is more than one premise category (finding out that both skunks and onions have Property X might make one fairly sure that Property X is linked to having a strong odor) or even more than one conclusion category. A Relevance Theory of Induction 14 Although relevance often may involve categories, unlike the SCM, relevance theory is not restricted to them. Instead, non-taxonomic categories, properties, and even thematic relationships may form the basis for categorical induction. For example, a premise statement that kangaroos have some property may trigger mammals as the relevant category but it may also lead to Australian animals or mammals with pouches as the relevant superordinate. Another difference from the SCM is that, for a given rank or level, some superordinates may be more informative (salient) than others. To continue the prior example, kangaroo should be more likely to activate Australian animals than muskrat should be to activate North American animals (at least for participants from universities in the United States). That is, the fact that Australia is more distinctive with respect to the animals that inhabit it should make it more likely that Australian animals would be seen as a relevant category for induction. A third difference from the SCM (and the feature-based and Bayesian approaches as well) is that premises and conclusions may be linked through causal reasoning. Shortly we will amplify this point. The principle that premises are compared with each other to determine relevant categories and properties is similar in spirit to the McDonald et al. hypothesis assessment model (see also Gentner and Medin, 1998 and Blok & Gentner , 2000, for related ideas concerning premise comparison). Their efforts and experiments were directed at linking category-based induction with other research in the hypothesis-testing tradition. They view premises of arguments as triggering hypotheses that fix the scope for induction. Our goals are tied more directly to manipulating effect and effort, in most instances through comparison processes used to fix relevance. Although one can certainly cast the outcome of such comparison processes as hypotheses, the relevance framework leads to a new set of predicted induction phenomena and a A Relevance Theory of Induction 15 different slant on the effects described by Osherson et al. (1990). Before bringing out these predictions, we first turn to the role of causal reasoning in induction. Causal relations. Consider the following inductive argument: “Grass has enzyme X, therefore cows have enzyme X”. The Osherson et al. SCM would assess this argument in terms of the similarity of grass to cows and the coverage of grass in the lowest level superordinate category that includes cows and grass (living things). Consequently, the argument strength should be low, according to the SCM. As mentioned earlier, our relevance framework employs a notion of similarity constrained by comparison processes and allows for thematic or causal relations to affect induction. For this example, people are likely to retrieve a linkage between cows and grass, namely that cows eat grass. This knowledge invites the causal inference that enzyme X might be transmitted from grass to cows by ingestion. Consequently, the argument about grass and cows should seem to be strong (and relevant). In brief, by selecting categories (and properties) about which undergraduates may have relevant background information we may be able to vary what Van der Herst et al (2000) call effect. Biological experts or Itza’ Maya have a great deal of background knowledge such that arguments involving biological categories will naturally produce large effects, often in terms of causal relations Manipulating effort. The relevance framework suggests some straightforward ways of varying effort to affect inductive confidence. First, with respect to comparison processes, additional premise (and conclusion) categories can be used to reinforce or undermine the ease and likelihood of seeing some property as relevant. Consider again an argument going from skunks to zebras. Adding the premise that striped bass also have the property in question should work to make it easier to conclude that the property in question is linked to having stripes and therefore applies to zebras. In fact, relevance may even over-ride normative considerations. A Relevance Theory of Induction 16 Suppose we compare an argument going from skunks to zebras with an argument going from skunks to striped bass and zebras. It’s possible that the comparisons of conclusions and premise will so boost confidence that the relevant basis for induction has been identified that the argument with the conjunctive conclusion will be seen as stronger than the one with a single conclusion category. A similar contrast involving effort is readily available for causal scenarios. Consider the argument that Grass has Enzyme X and therefore Humans also have Enzyme X. A potential causal linkage may be less transparent than for the case with the same premise but where the conclusion is that therefore Cows and Humans have Enzyme X. The addition of Cows (and the accessible knowledge that humans drink the milk of cows) may make it easy to create or retrieve a causal linkage from grass to humans and lead to the conjunctive conclusion being evaluated as stronger than the single conclusion (obviously this prediction has to be evaluated in a between participants design). Work on the availability heuristic in relation to causal schemas (e.g. Tversky and Kahneman, 1974) also suggests that causal relations will more readily affect inductive confidence when the cause is the premise and the effect the conclusion than for the reverse order. In short, the relevance framework leads to a number of novel and in some cases non-normative predictions. A Note on Blank versus nonblank properties. Osherson et al. (1990) define "blank" properties as those for which participants have few beliefs and are unlikely to evoke beliefs that cause one argument to have more strength than another. For example, most people have no a priori opinion about whether robins or ostriches "require biotin for protein synthesis." The SCM works best in explaining induction phenomena that involve blank properties. Indeed, in order to account for arguments with nonblank predicates, Smith, Shafir and Osherson (1993) showed that A Relevance Theory of Induction 17 a number of additional processing assumptions needed to be added to the similarity coverage framework. The distinction between blank and nonblank properties, however, is not always clearcut. Heit and Rubinstein (1994), for example, showed that undergraduates generalized abstract behavioral properties in a different pattern than abstract physiological properties (behavioral similarity had a greater effect in the former condition). The relevance framework suggests that interactions between premise and conclusion categories or between premise categories may evoke beliefs about even the blankest of blank properties. Suppose we modify our earlier argument to the more abstract form, "Grass has some property X, therefore cows have property X." It seems likely that people will still entertain the idea that X may be something that can be transmitted from grass to cows. Even an isolated premise may evoke certain beliefs. For example, the premise, "Penguins have property Y" is likely to trigger expectations about property Y that render Penguins a relevant, informative, premise category. In this case people might expect that property Y is an adaptation to an Antarctic environment or linked to swimming and waddling rather than flying or even the abstract belief that the property must be unusual because penguins are unusual birds. The Osherson et al. strategy of using abstract, unfamiliar properties is very effective for seeing what other information people bring to the task to determine relevance and draw inferences. In addition, the absence of a strong borderline between blank and nonblank properties suggests that we should be able to develop models of induction that address a range of specificity and familiarity of predicates. One advantage of the relevance framework is that it does not require blank predicates (nor do the feature-based, Bayesian or hypothesis models). A Relevance Theory of Induction 18 Summary of Relevance framework predictions. The specific assumptions we have been developing can be seen as implementing the general principle of relevance for the case of category-based induction. The main ideas are that induction involves a search for relevance and that candidates for relevance are salient properties and (causal) relations. Most important for present purposes is the idea that effect and effort can be manipulated by use of undergraduates’ background knowledge and by introducing additional premises and/or conclusions that increase or decrease effort. The key experimental manipulations in our studies are: 1. strengthening and weakening of candidates for relevance via property reinforcement, and 2. scenario (causal) instantiation and manipulations of effort designed to increase or decrease access to causal associations. So far we have kept our descriptions at a general level rather than adopt a specific, quantitative model. This is in large measure because the framework leads to a number of qualitative predictions and would be consistent with a large set of specific instantiations. In addition, the determination of relevance may require fairly flexible processing principles. For example, rather than assuming a fixed order of comparisons, comparisons may be guided by the strength of correspondence between the representations associated with a comparison (as Goldstone & Medin, 1994, assume), which could alter the comparisons themselves. Consider the following argument: “Polar bears have CO3 and walruses have CO3, therefore polar bears have CO3”. In this case the excellent correspondence between the first premise and the conclusion (namely, identity) might well preempt the comparison of premises to each other and lead directly to the inference that the argument is perfectly strong. For the present we restrict ourselves to these general ideas about property weakening and strengthening. In the next section we present predictions/phenomena that are tied to the relevance framework. A Relevance Theory of Induction 19 Predictions of the Relevance Framework. In this paper we focus on phenomena associated with effort and effect by varying causal scenarios and property reinforcement. In general, our strategy is to develop items for which relevance-based reasoning makes predictions that either run counter to other models (e.g., nondiversity) or that contradict normative judgments (e.g., conclusion conjunction fallacy). We examine five phenomena involving causal scenarios. The first of these, causal asymmetry, predicts that an inference from A to B will be rated as stronger than an inference from B to A when a relevant causal scenario about the transmission of a property from A to B is salient. For example, GAZELLES / LIONS should be stronger than LIONS / GAZELLES because it is easier to imagine a property being transmitted from gazelles to lions via the food chain than vice versa. The motivating idea is that premise order affects the effort needed to activate a causal scenario. Next, causal violation of similarity and causal non-diversity pit causal relations directly against predictions derived from other models. We examine whether causal relations might override similarity by strengthening inferences between dissimilar premise and conclusion categories (e.g., the GRASS / COWS example discussed above), and whether causal relations might override diversity by weakening otherwise diverse premises. For example, ROBINS + WORMS / GOLDFISH may be stronger than ROBINS + IGUANAS / GOLDFISH in terms of sheer coverage of the inclusive category animals, but the salient fact that robins eat worms might make it plausible that they would share a property not generally shared, and therefore weaken the former argument. These sorts of effects of causal scenarios have been frequently in expert populations (e.g. Proffitt, et al., 2000) and the main new contribution is to show that these effects can be predicted in advance and that they can be demonstrated in A Relevance Theory of Induction 20 undergraduates, as long as relevant background knowledge is pinpointed (in the relevance framework one is selecting items where background knowledge will produce larger effects). The fourth causal phenomenon involves a case in which salient causal scenarios lead to logically non-normative judgments: causal conjunction fallacy. Normatively, adding a conclusion category to an argument should never strengthen it. In contrast, the causal conjunction fallacy predicts that adding a conclusion category which strengthens a causal link between premise and conclusion might strengthen the argument. For example, GRAIN / MICE + HAWKS might be considered stronger than GRAIN / HAWKS (a logical fallacy) because it may foster a causal link from grain to mice to hawks. In terms of the relevance framework the addition of the MICE premises reduces the effort needed to develop the causal linkage from grain to hawks. Finally, causal non-monotonicity predicts that adding a premise category might weaken an argument if it highlights a causal relation between premise categories not shared by the conclusion category. For example, HUMANS / OAKS might be considered stronger than HUMANS + MOSQUITOES / OAKS, because mosquitoes might plausibly transmit a property to humans but not to oaks. The SCM only allows for weakening if the additional premise increases the abstractness of the lowest level super-ordinate category that covers premises and conclusion and the (original form of the) FBIM does not allow for non-monotonicities at all. We also investigate three phenomena involving property reinforcement which parallel the phenomena presented above for causal scenarios. Non-diversity via property reinforcement suggests that if an otherwise diverse set of premises share a salient property not shared by the conclusion category, the reinforcement of the property might weaken that argument relative to a related argument with less diverse premises. This is not unlike Tversky’s well-known (1977) A Relevance Theory of Induction 21 diagnosity principle for similarity judgments. For instance, in the SCM framework, the argument PIGS + CHICKENS / COBRAS is assumed to be stronger via coverage than the argument PIGS + WHALES / COBRAS (because pigs are mammals and chickens are birds, and therefore cover the inclusive category animal better than pigs and whales, two mammals). However, pigs and chickens--unlike cobras--are farm animals and raised for food; these properties might weaken the argument. Conclusion conjunction fallacy via property reinforcement predicts that adding a conclusion category that reinforces a property shared by premise and conclusion might strengthen the argument. For example, DRAFT HORSES / RACE HORSES + PONIES might be considered stronger than DRAFT HORSES / PONIES. As we shall see, the significance of this effect is less that it is non-normative than that it contrasts the relevance framework with alternative models. Non-monotonicity via property reinforcement predicts that adding premise categories might weaken an argument if the added categories reinforce a property shared by all premise categories but not by the conclusion category. For example, BROWN BEARS / BUFFALO might be considered stronger than BROWN BEARS + POLAR BEARS + GRIZZLY BEARS / BUFFALO because in the latter case participants may be thinking that the relevant conclusion categories is bears. In general, our strategy is to present participants with sets of arguments in which relevance theory makes predictions that run counter to one or more other models. To the degree that relevance-based arguments are rated as stronger, the relevance framework is supported. A Relevance Theory of Induction 22