Squib-Deflationary account of invited inferences

A deflationary account of invited inferences Ira Noveck, Mathilde Bonnefond & Jean-Baptiste Van der Henst This squib reconsiders Geis & Zwicky’s influential proposal on Invited Inference, according to which conditionals are regularly “perfected” to biconditionals. We first show that the “regularity” assumption attached to conditional perfection is doubtful in light of established experimental findings concerning other logical terms, such as Some and or and the conjunction and. We then review existing conditional data with the aim of making them cohere with these other experimental findings. We argue that (a) the process that leaves the impression of a biconditional reading (the acceptance of a fallacious argument such as the Affirmation of the Consequent) arises only after all participants detect a violation on-line from what is essentially a surprising minor premise and that; (b) some participants make an effort to adjust to such unexpected violations at a relatively small cognitive cost in order to accept invalid arguments while others persist in rejecting whatever follows and at a greater cognitive cost. Both of these features of conditional processing undermine claims from Geis & Zwicky’s proposal. To those who study reasoning and pragmatics, the best known squib is arguably Geis and Zwicky’s (1971) On invited inferences. It was and remains a valuable contribution to the cognitive science literature and especially to those who have long aimed to better understandseek explanations for the way the meanings meaning of logical terms, such as if and other logical connectives and or, are enriched pragmatically. Its main contribution was presented as a principle known as Conditional Perfection -- “a sentence of the form X  Y invites an inference of the form non-X  non-Y.” The authors explain it further by writing: This principle asserts a connection between linguistic form and a tendency of the human mind ‘to perfect conditionals to biconditionals’ in words suggested to us by Lauri Kartunnen. This tendency is manifested in two classical logical fallacies, Affirming the Consequent (concluding X from X  Y and Y) and Denying the Antecedent (concluding ~ Y from X  Y and ~X). Another admirable aspect of their squib is that it refers to invited inferences naturalistically as “underbrush that needs to be cleared before investigations of semantics can thrive.” This sort of description is much more attractive than the artifactual way it is usual put – as the “wastebasket of semantics” (see Mey 2001 for an historical account of this expression). The proposal and its comment capture well what many take as a given, which is that conditionals can be readily understood as biconditionals through a pragmatic enrichment of some sort that is practically automatic. As they put it, the “association of inferences with semantic content can be highly regular.” Their account of conditional enrichment has remained influential even after forty years of continued linguistic research into conditionals (for a review, see Van der Auwera, 1997) and dozens of psychological experiments (for a review, see Politzer, 2007). The goal of the present squib is to reconsider the accepted notion of invited inference in light of findings from the conditional reasoning literature and, more pertinently, current experimental pragmatic research pertaining to other logical terms, including propositional connectives such as or and and. As we will show, when viewed in the context of other pragmatic phenomena, the conditional and the data it generates appear exceptional. This ultimately leads us to present a different view on the pragmatic enrichment of conditionals and ultimately to a rejection of Geis and Zwicky’s original and influential proposal. In what follows, we present a brief review of the findings on pragmatic enrichments linked to utterances having logical terms other than if-then, such as the existential quantifier, Some, the connective for disjunction, or, and the conjunction, and. The first two are known to prompt what are known as scalar inferences since they result from a narrowing of a relatively weak expression. For example, Some often leads to the exclusion of All (i.e. the semantics of Some can be glossed as Some and perhaps all and it is narrowed to Some and not all); the notion of scales arises because All is a quantifier comparable to Some and entails it. The case of and is similar in that it is assumed to be readily interpreted, e.g., as and then (Mary bought an iPad and spent the evening surfing). However, it comes with no assumption of scales. Scalars and conjunctions are often viewed as two different species of pragmatic enrichment. For example, according to Levinson, scalars are examples of Heuristic 1 (What isn’t said isn’t) and a conjunctive enrichment is an example of Heuristic 2 (What is simply described is stereotypically exemplified). Conditional perfection is considered an example of Heuristic 2 as well. What investigations about scalars and conjunctions tell us Many in the linguistic-pragmatic literature (e.g., see Levinson, 2000) assume that scalar terms come readily with enrichments so that Some means by default Some but not all or that or becomes automatically exclusive (or but not both). It follows that in order to get at semantic readings (which would be some and perhaps all and or and perhaps both, respectively), one would need to cancel default readings. In other words, it is often assumed that enriched readings are primary and that semantic readings require an effortful second step. However, experimental research has largely debunked that idea (see Noveck & Reboul, 2008 for a summary). It now seems quite clear that semantic readings are often sufficient for making on-line interpretations with utterances containing scalars and it is readings with enrichments that come with extra costs, even if they are relatively light. There is little evidence that utterances containing scalar expressions are automatically enriched to be read as some but not all or as or but not both. This can be seen developmentally, in terms of reading times, as well as in EEG experiments. We quickly describe examples of such findings. Developmentally speaking, enriched readings occur more frequently with age, indicating that younger participants are more likely than older ones to rely on non-enriched (semantic) readings. This is supported by findings showing that children are less likely than adults to reject as false or inadequate underinformative sentences such as Some boxes contain a token (Certaines boites contiennent un jeton) when presented with four boxes each containing a token. The developmental trajectory (from non-enriched to enriched) is present consistently and across multiple languages (French, Greek, Italian, English & German, to name a few) and even as types of underinformative sentences vary (Noveck, 2001; Papafragou & Musolino, 2004; Guasti et al., 2005; Pouscoulous et al., 2007). One finds this effect in sentence processing too. Using a self-paced task, Breheny et al. (2006) compared the reading times of disjunctive phrases across two kinds of contexts, Lower-bound contexts (where the non-enriched reading of the scalar is more appropriate) and Upper-bound contexts (where the enriched meaning is). For example, compare (1) and (2) below (a slash bar indicates where readers are required to hit a key to advance a text): (1) Lower-bound context John heard that / the textbook for Geophysics / was very advanced. / Nobody understood it properly./ He heard that / if he wanted to pass the course / he should read / the class notes or the summary. (2) Upper-bound context John was taking a university course / and working at the same time. / For the exams / he had to study / from short and comprehensive sources./ Depending on the course, / he decided to read / the class notes or the summary. The disjunctive phrase in (1) takes significantly less time to process than the same phrase in (2). This is, by the way, consistent with findings from a sentence evaluation study from Bott and Noveck (2004) with existential quantifiers. They found that True responses to underinformative sentences such as Some goats are mammals were as fast as responses to a set of true and false control items (e.g., Some mammals are goats and All goats are insects) while False responses to underinformative sentences were consistently linked with significantly longer evaluation times. Further evidence of the ready use of semantic readings comes from a recent EEG study from Nieuwland et al. (2010). While exploiting individual differences first reported in Noveck (2001) and in an EEG study from Noveck and Posada (2003), Nieuwland et al. showed through EEG recordings that the moment that underinformativeness becomes apparent is generally not noticeable to most participants. Consider sentence (3) from Experiment 2 of their study: (3) Some people have lungs that require good care. The word “lungs” which is arguably underinformative at the point at which it is expressed (since all people have lungs) apparently prompts an EEG profile that is indistinguishable from the word “pets” in (4) when it is expressed, even though the latter does not carry have the potential to appear similarly underinformativeeness: (4) Some people have pets that require good care. It is important to note that Experiment 1 of the same paper -- with the same paradigm -- had shown that EEG profiles appear much different when (a) participants are broken down into two groups based on their “Communication Score” in Baron-Cohen’s Autism Quotient (a higher score indicates that an individual is higher on the Autism Spectrum), plus when; (b) a comma is inserted prompting a sort of “clausal wrap-up.” For example, (3) and (4) were presented as (3’) and (4’): (3’) Some people have lungs, which require good care. (4’) Some people have pets, which require good care. With these sentences, those lower on the scale (who are thus considered more “pragmatically skilled”) produced EEG profiles indicating a greater sense of surprise (an N400) to the underinformative item in a sentence such as (3’) when compared to (4’) than those who are higher on the scale and whose N400’s are less remarkable. Thus, it confirms that there are apparent individual differences in the way underinformative sentences are processed. The moral upshot here is that a subgroup proportion of participants can be encouraged to enrich a sentence on-line through linguistic cues. Nevertheless, it remains hard to argue that the existential quantifier in utterances is generally enriched by default. The account of scalars presented here indicates that pragmatic enrichments are highly contextual and that, very often, unenriched treatments are good enough. The account also extends to conjunctions. Noveck et al. (2010) showed that younger children are more likely than adults to respond affirmatively to a question such as “Guillaume ate dinner at a friend’s and took a cat into his arms?” after reading a short vignette that just described the opposite sequence of events. Again, the minimal semantic reading here (two conjoined events occurred) is often sufficient in on-line interpretations while the enriched reading (two conjoined events occurred and in their order of mention) is linked with extra effort. Such data go against default accounts (Levinson assumes that conjunctions are “buttressed” by conventional enrichments so that and is readily understood as and then or as and therefore). In this case, too, pragmatic enrichments appear to require some amount of effort. Three insights from the experimental Summary concerning pragmatic literaturescalars and conjunctions One lesson learned from the scalar literature is that pragmatic enrichment serves as a way to gain some amount of information. When a participant narrows the meaning of an underinformative utterance she is ultimately doing some sort of narrowing. This much is not controversial. Perhaps a hotter topic of discussion is the claim, one largely supported by data, that suchis the sort of refinement thats are is typically associated with extra cognitive effort (Bott & Noveck, 2004; Chevallier et al, 2008). The second lesson from the scalar literature is that enrichments do not occur generally and by default. That is, a literal reading of, say, some, can suffice for reaping adequate cognitive effects; enriched readings, which can occur when contextual clues and an addressee’s predisposition encourage them, are not automatic. A scalar inference is not a genie that needs to be put back into its bottle; it is an enrichment that comes out under favorable circumstances and and, as we indicated, at a (perhaps slight) cost. One of the main findings from the literature on scalars is that extra work linked to pragmatic enrichments is due to their production and not to their cancellation (again, see Noveck & Reboul, 2008). Finally, and relatedly, one ought to find individual differences – both within populations and across age groups. Experiments in the scalar literature have repeatedly shown that one can find groups of participants who typically enrich a sentence in characteristic ways as well as groups that do not (Noveck & Posada, 2003; Bott & Noveck, 2004; ). This does not account for all the data for thereThere are often participants who provide inconsistent responses. However, within-subject analyses still show that, even among them, pragmatic enrichments require are linked with extra effort (see Bott and Noveck, 2004; Exp. 3). The division, often reported in the experimental pragmatic literature, most likely depends on task demands and abilities (e.g., see . Multiple experiments across various paradigms largely confirm this result ; Degen, Reeder, Carbary & Tanenhaus, 2009). For instance, the scalar literature has popularized the notion that children, as a group, are less “pragmatic” than adults (Noveck, 2001; Papafragou & Musolino, 2003). This appears to hold for a range of pragmatic phenomena, including for conjunctions. Distinctions between children and adults, often reported in the experimental pragmatic literature, also depend on task demands and abilities (e.g., see . The data from Nieuwland et al. can be viewed through this prism as well. Pragmatically inclined individuals are more likely than the less inclined to take advantage of linguistic cues in order to enrich out-of-the-blue sentences. Applying lessons from the scalar inference literature to conditional processing Assuming that invited inferences as described by Geis & Zwicky are for real, one would expect them to share the main signature of other pragmatic enrichments described above: Their appearance should co-occur with extra effort (with age or further processing time) where faster treatments or younger ages ought to be linked with a semantic reading and slower treatments and increasing age ought to be linked with a greater likelihood of invited inferences. One should then find that these extra efforts are employed in ways that reveal individual differences. It should also be safe to assume that invited inferences do not arise by default. However, when one turns to the conditional reasoning literature and the expectations concerning the way invited inferences are proposed to arise, one finds descriptions as well as behavior that are diametrically opposed to what was just described. To simplify our discussion we will focus on examples that concern the fallacy of the Affirmation of the Consequent. We do this mostly to avoid the negations in Denying of Antecedent cases. Negations are universally recognized as the source of added complexity to both sentence processing and to linguistic analysis. Note the following four features that are recurrent in the conditional reasoning literature and that do not match up with characteristics in the scalar literature. First, while growing up and extra effort are linked with pragmatic enrichments with respect to scalars, advancing age and further effort are linked to fewer “pragmatic” interpretations with conditionals (Taplin et al, 1974; Barrouillet et al, 2000). Research on conditionals generally shows that when extra cognitive resources are required they are linked with discouraging the endorsement of fallacies, such as AC arguments. Second, reaction time measures also show that endorsements of, say, AC fallacies are fast while rejections (indicating logically valid reasoning) are time-consuming. For example, Marcus and Rips (1979, Experiment 2) presented the four sorts of conditional syllogism -- Modus Ponens (MP), Affirmation of the Consequent, Denial of the Antecedent, and Modus Tollens tachistoscopically and their participants’ reaction times showed that, while those participants who endorsed Modus Ponens and AC were statistically comparable (1813 and 1907 msec, respectively), only a third of the responses were endorsements of AC (in the meantime, MP prompted 100% endorsements). Those trials that rejected AC arguments were more common and took noticeably longer (2437 msec). The authors did not report statistical tests, so this point is based on eyeballing the presented data. For other data that point in the same direction, the reader is directed to Van der Henst et al. (2006), to be described later. Third, assumptions behind invited inferences do not ring true. Unlike present views on pragmatic enrichment, invited inferences are characterized in one of two ways. One is to say that they arise with the conditional. This leaves the impression that the participant who accepts the AC conclusion as true must have produced the invited inference with the arrival of the conditional premise and that she then treats the minor premise in a way similar to Modus Ponens . Others refer to the way pragmatic inferences “interfere” with conditional processing (e.g. Daniel, 2006). This characterization also implies that invited inferences readily arise and that semantic readings (and logically valid judgements) result from a cancellation. Such descriptions sound much like a default which is itself a questionable assumption. Fourth, individual differences with scalars concern distinctions between those who are more or less skilled pragmatic hearers. Dissimilarly, performance with conditionals leads to different types of individual differences having to dothat concern with memory abilities and IQ (De Neys et al., 2005; Evans et al., 2008, 2009; Evans, Handley, Neilens, & Over, 2007). Indications of pragmatic interpretations increasing with age are not to be found. Overall, the data are not consistent with the notion that extra work effort is linked with accepting AC arguments (through the inclusion of invited inferences), but rather with rejecting them. In other words, the data show that extra work is linked with maintaining a refined interpretation (i.e. the only way to make an inference appear at all is with the expectedone that accepts only the conditional’s antecedent as a minor premise). So, when viewed through the prism of pragmatic enrichment behavior described earlier, explanations of conditional processing based on invited inferences seem paradoxical. Reconciling two pragmatic phenomena: An alternative account of conditional enrichment There are two ways to resolve this paradox. One route is to say that scalars and conditionals represent two species of pragmatic enrichment that behave in diametrically opposing ways. This is what both Horn (1992) and Levinson (2000) propose by arguing for two sorts of heuristics. From this point of view, conditional perfection is akin to conjunctive enrichments (see Footnote 2). Claims from this sort of strategy have been punctured already by the fact that conjunctions and scalars pattern together even though they are representative of the two different species and, as we just summarized, conditionals are distinguishable from both. Another (more parsimonious) route is to suppose that there are overarching pragmatic principles that can account for both phenomena. If one takes the second approach, then we need to do more explaining. It is the second approach we take. We assume that pragmatic processes are common across all utterances and that one of the main signatures of narrowed readings is evidence of longer processing and age-related changes when compared to less vigilant readings. The question then is what makes scalars on the one hand and conditionals prompt what appears to be different phenomena? WWhile assuming that the pragmatic processes in both cases are similar, we thus assume that it is the semantics of each that differs. Below, we sketch out how interpreting scalars is fundamentally different from interpreting conditionals even before pragmatic principles come into play. An utterance with a scalar term does not come with a built-in, on-line interpreter that forcefully enriches the meaning of, say, some through narrowing. That much has been established by the experimental literature. Whether one looks at developmental data, on-line data or EEG measures, one arrives at the same conclusion -- a minimal interpretation can often be good enough. When an utterance includes Some there is no, context-free and specific, expectation that its interpreted meaninginterpretation will go beyond its coded semantic meaning. Contrarily, it appears that athe conditional statement does come with a minimal expectation, which is that its antecedent be imminently available. come with a built-in minimal expectation about what should follow. An utterance introduced with a conditional of the form if p then q raises expectations – and specifically about the antecedent. When one hears If you mow the lawn then I’ll give you five dollars, the focus is implicitly on mowing the lawn. The questions raised are: Will you, the addressee, mow the lawn? Is there a doubt you will? There is a focus on the antecedent that, at least initially, leaves unconsidered anything else. This expectation is pragmatic potential in itself and is eventually at the root of further pragmatic processing (see Allott & Uchida, 2009, for a similar argument). In any case, the arrival of a minor premise that potentially sets up a fallacy (such as the minor premise in an AC argument) violates such an expectation. A hearer is expecting more about the antecedent and is, instead, served something else. It is our view that the minor premise of the fallacies is not the immediate source of an online pragmatic modification nor is the conditional itself. Rather, the presentation of a fallacious minor premise represents a violation, which speaks to the semantics of the conditional and the expectations it raises. The claim that conditionals come with built-in expectations concerning the antecedent has found empirical support in our lab. In an EEG study using arbitrary letters as materials, Bonnefond and Vvan der Henst (2009) found unique P3b activity linked to the Minor Premise of a conditional argument when it is presented in its traditional position (If P then Q; P) as opposed to when the minor premise was presented first (P; If P then Q). This indicates that participants’ expectations are satisfied with the arrival of the antecedent as a Minor Premise. The arrival of a minor premise that potentially sets up a fallacy (such as the minor premise in an AC argument) violates such an expectation. A hearer is expecting more about the antecedent and is, instead, served something else. It is our view that neither the conditional nor the minor premise of the fallacies is the immediate source of an online pragmatic modification (such as a Geis & Zwickian invited inference). Rather, the presentation of a minor premise that points to a fallacy is akin to an outright violation, which sets in motion a process that is different from an enrichment. That is why the presentation of a minor premise that leads to a fallacy ought to appear unexpected (Bonnefond & Vvan der Henst, 2009; Bonnefond et al., in preparation). For example, Bonnefond et al. (in preparation) show – in another EEG studythrough reaction times and EEG -- that a minor premise indicating an AC inference in a standard conditional syllogism prompts a reaction indicating surprise even among participants who ultimately accept AC arguments. This simply should not occur from the the point of view that AC arguments are accepted due to invited inferences that transform conditionals into biconditionals. This speaks to the semantics of the conditional and the expectations its appearance raises. The claim that extra time (further processing) yields greater proportions of rejection of an AC argument has also been demonstrated by having participants devote more time to reading conditional arguments. showed participants MP and AC arguments either one word at a time or a sentence at a time. Those who were required to spend more time reading the arguments (the one-word-at-a-time group) were significantly more likely to reject AC arguments. This shows that individual differences concerning memory and intelligence can take a back seat to subtle variations of time and processing, the sort of variables that are intrinsic to pragmatic processing. We interpret these findings to mean that extra time in processing the arguments is linked with developing a narrowed interpretation of what one can expect following a conditional. RHow participants reacting to an unexpected minor premise should not be confused with supposing the supposition that hearers enrich conditional statements on-line with invited inferences. It is our view that the fallacies, such as those brought on by AC arguments, actually reveal how participants react after being faced with an unexpected minor premise that violates one’s expectations. Once something goes awry (because one did not get a minor premise that was expected), one can react either by being generous or by being strict in one’s interpretations. A generous reaction can ultimately lead to finding ways to endorse the argument (dashed expectations notwithstanding) and a strict reaction would be to recognize the flaw and to not refuse to budge. We assume that everyone detects that an expectation was not met in an AC argument and that, between the two ways available to react, the latter sort of reaction (the outright rejection of the conditional argument) is more effortful because it requires maintained vigilanceattention and ultimately a rejection (which is usually time consuming). The ultimate acceptance of an AC argument implies that its minor premise allowed the participant to make further inferences once looking past a noticeable flaw. These claims are not obvious so to make these a participant’s two potential reactions more salient, we describe a non-conditional situation that entails a dashed expectation and two reactions to it. Imagine arriving at a Mediterranean hotel on a hot day and being informedtold (at the front desk) that your room has air conditioning; when you get to your room, all you find is a fan. One can react in two ways. One way is to be generous and reconcile the contradictory information (e.g., by supposing the speaker meant the fan). T he other way is toor one can note the contradiction and conclude that the speaker had no right to assert that there was air conditioning in the first place, i.e. the fan does not follow from the utterance about air-conditioning; thus, one can reject outright the front-desk’s utterance. Either way, one has to first register disappointment in not having had what was expected and either way one is doing work that is post-hoc (in the face of unexpected information). Similarly, a conditional sentence raises expectations about itsthe antecedent. When it does not arrive at the moment expected (as a minor premise), one has a choice. One can accept the disappointment and make do with it (assume that the speaker intends the listener to make profitable use of it) or else effortfully maintain a (logically valid) stance and, stubbornly refuseing to accept anything that follows from what is essentially a false claimthe disappointment. . Arguably,Arguably, the second approach requires more mental energy to carry outmaintain. To summarize, we propose that there are further efforts to be made in conditionals – that do resemble the kind of pragmatic enriching described above with scalars -- but they arise with respect to expectations about the antecedent of the conditional. We also propose that the pragmatic-like costs of effort are twofold. One is linked to maintaining the conditional’s built-in expectation, viz. participants seek useful information that allows the conditional to go through (the antecedent that prompts modus ponens). The data traditionally associated with conditional reasoning bear this out. As people get older, they increasingly are more likely to realize that only a satisfied antecedent is informational. The other place for costs arises when faced with fallacious arguments, such as AC. The costs vary as a function of the way people react in the face of disappointment (when the expected antecedent does not occur). There are those who reject the argument in light of a disappointing minor premise and there are those who make do with it by seeking ways to make further inferences (regardless of their validity). One can see that those who reject an AC argument never need consider invited inferences and, according to our account, those who accept an AC argument are not inserting invited inferences on-line. Unlike the case with scalars (where a minimal meaning is good enough), the minimal acceptable interpretation that engages conditionals is one that is joined by its antecedent in a separate line. When this does not come to pass at the moment expected, all participants note some measure of surprise and disappointment. Those who do accept AC arguments nonetheless arguably adjust the argument in a generous way once they get past their disappointment. In fact, they then might consider that the conditional Those who reject AC arguments refuse to consider ways to make the argument pay off. Either way, such a view highlights that invited inferences --– the on-line and ubiquitous pragmatic enrichments as characterized by Geis and Zwicky as on-line and ubiquitous -- are simply never (or are very rarely) in play. References PAGE 6