8 Evidence from Other Research

Having discussed evidence adduced primarily in my lab, it is timely at this stage to review contemporary literature to see whether its array of findings can be accounted for by the graded salience hypothesis (Giora, 1997b, 1999b; chapter 2 in this volume).

Recall that according to the interactionist, direct access view (cf. chapter 3 in this volume), processing language involves a single mechanism that is sensitive to both linguistic and nonlinguistic information and allows for contextual information to interact with lexical processes very early on. Such processes result in tapping the appropriate interpretation directly, with no recourse to contextually incompatible information. Thus, when context is specific and supportive, literal and nonliteral language involve equivalent processes.

Proponents of this view argue against the context-resistant approach, dubbed “the standard pragmatic model,” according to which interpreting an utterance figuratively is a post-access procedure, upon which the literal interpretation is activated initially, regardless of context; moreover, because it is inappropriate, the interpretation is adjusted to contextual information. On this view, only literal language is understood directly.

Alternatively, the graded salience hypothesis assumes that comprehension involves two distinct mechanisms that do not interact at the initial comprehension stage: a contextual, central system mechanism, which may have rapid predictive effects, and a lexical accessing mechanism (among others), which is modular— automatic and impervious to context effects. Lexical access is salience sensitive: salient (familiar, frequent) meanings are accessed automatically upon encounter of the linguistic stimulus, regardless of context; less salient meanings lag behind. Although context may avail the appropriate meaning immediately via the predictive and inferential mechanism, it does not block salient meanings even when they are inappropriate. Understanding unfamiliar uses of familiar instances may be rapid due to a strongly predictive context, but it would not be selective: salient, though contextually incompatible meanings would emerge nonetheless on account of their salience (see chapters 2–3 in this volume).

Like the direct access view, the graded salience hypothesis dispenses with the literal/figurative divide regarding initial processes. It assumes, instead, the salient-nonsalient continuum. It thus predicts that alternative interpretations of similar salience would be activated in parallel (for instance, the literal and figurative meanings of familiar metaphors), regardless of context. Utterances diverging in salience (such as the ironic versus literal interpretations of unfamiliar ironies and the literal and idiomatic interpretations of familiar idioms) would involve different processes. Initially, nonsalient, novel uses of familiar language would involve salient, albeit contextually incompatible meaning. Contextual information will be operative, too, either preceding, coinciding, or following the output of lexical processing. But it will affect the latter only postlexically. When these parallel operations (of the contextual and the lexical mechanisms) result in conflicting outputs (that are anticipated upon encounter of novel uses of language), consequent revisitation processes will follow.

The following facts cannot be accounted for by the traditional view: they show that context facilitates initial comprehension of figurative language. For instance, some studies attest that a rich context neutralizes the difference between comprehension of literal and nonliteral language. Contexts longer than three sentences induced similar reading times for (nonsalient) metaphoric and (relatively more salient/accessible) literal interpretations (Inhoff, Lima, & Carroll, 1984; Ortony, Schallert, Reynolds, & Antos, 1978; Shinjo & Myers, 1978). Similarly, Gerrig and Healy (1983), who manipulated metaphor and context ordering, showed that metaphors followed by a context phrase took longer to read than when preceded by a context. (The ordering manipulation, however, did not affect comprehension of literal sentences). Kemper (1981), who investigated comprehension of proverbs, also found that the length of the paragraph affected proverb interpretation: the longer the paragraph, the faster it took to interpret it figuratively. Turner and Katz (1997), Katz and Ferretti (2000, 2001), and Katz, Ferretti, and Taylor (1999) showed that familiar proverbs did not take longer to read in figuratively than in literally biasing contexts (but see Katz and Ferretti's, 2000 experiment 2 for counterexamples). Similarly, Pickering and Frisson (2001a) found no difficulty with unfamiliar metonymies (read Needham) compared to familiar ones (read Dickens) when preceded by a predictive context.

Apparently, these findings can be accounted for by the interactionist/direct access approach. According to the direct access view, elaborated and supportive contextual information interacts with lexical processes very early on and preselects the contextually appropriate meaning. Hence, the facilitation effects and equal reading times.

These findings can also be accounted for by the graded salience hypothesis. For instance, equal reading times for similarly familiar utterances is anticipated (Turner & Katz, 1997; Giora & Fein, 1999c). But even equal reading times for more and less familiar utterances need not necessarily disconfirm the graded salience hypothesis because the graded salience hypothesis allows for facilitative contextual effects. To disprove the graded salience hypothesis, salient (but contextually incompatible meanings) should be shown to be blocked upon encounter. Given that salient meanings should be activated irrespective of contextual information, the graded salience hypothesis thus anticipates processing difficulties upon encountering a less salient meaning that should involve accessing its salient, albeit inappropriate, meaning as well. For example, it predicts processing difficulties upon encountering troops while reading Regardless of the danger, the troops marched on in the metaphor-inducing context (speaking of ruthless kids, see Ortony et al., 1978). Encountering troops should involve processing difficulties such as found for unbalanced ambiguous words embedded in contexts biasing their interpretation toward their less salient (subordinate) meaning (Rayner, Pacht, & Duffy, 1994; cf. chapter 3 in this volume). Indeed, findings by Janus and Bever (1985), who measured reading times at the locus of the figurative information (troops) rather than at the end of the sentence, exhibited longer reading times for the nonsalient (figurative) than for the salient (literal) word.

Janus and Bever (1985) tested the hypothesis that the equal reading times found for literal and metaphoric utterances in Ortony et al.’s (1978) study were not affected by the prior rich context, but by the extra processing at the end of sentences (attested to by Abrams & Bever, 1969, and Just & Carpenter, 1980). This extra processing masked the difference in processing times for comprehension of literal versus metaphorical sentences and fostered an illusion of equivalence. Criticizing Ortony et al.’s methodology, but using their materials, Janus and Bever measured reading times at the end of target (vehicle) phrases rather than at the end of target sentences. Their findings show that even when the context is rich, novel metaphors present some difficulty to comprehenders (see also Brisard, Frisson, & Sandra, 2001; and see Gibbs & Gerrig, 1989, on the possibility that equal reading times of whole sentences may mask underlying processes). That rich and supportive context does not block salient though incompatible meanings has also been shown by Peleg, Giora, and Fein (2001, in press; cf. chapter 3 in this volume).

Findings regarding comprehension of idiomatic language do not corroborate the traditional view, either. They show that nonliteral meanings are activated first. For instance, in a conversational context, idioms took longer to be understood literally than figuratively (Gibbs, 1980). In Hillert and Swinney (2001), idioms were not processed literally first. In Van de Voort and Vonk (1995), familiar idioms, whose more salient meaning is idiomatic, were automatically processed idiomatically. Variant idioms, however, took longer to process than nonmodified idioms, while the opposite was true for their literal controls. A similar tendency was found in Ortony et al. (1978). Needham (1992), too, disconfirmed the hypothesis that literal meaning is activated during comprehension of idiomatic utterances. He presented participants with three target sentences preceded by a context that had a title. The targets were an idiom, a literal (anaphor) target, or a control phrase. The test word for all the three cases was identical an e idiom and to the literal (anaphor) target. Participants were told to d appeared previously in the text, but it was related only to the literal interpretation of th decide as quickly as possible whether the test word had occurred in the passage:

The test word for all three cases was pan, presented at *

Although participants’ response time in the three conditions did not differ significantly across subjects, and only marginally across items, there was a significant effect of condition on error rates across subjects and materials. The error rate for the literal (anaphor) condition was significantly lower than for either the idiom or the control condition, suggesting that the literal meaning was computed only for the literal target.

Most of these findings are accountable by the direct access hypothesis (Gibbs, 1984), which does not require that the literal interpretation of the figurative utterance be computed. They can also be accounted for by the graded salience hypothesis, which predicts that salient meanings, such as the conventional (figurative) meaning of idioms would be processed initially, regardless of context.

Let us consider other findings, which are problematic for the interactionist, direct access view. They suggest that the contextually incompatible literal meaning of figurative language is activated and at times even triggers a sequential process. Such findings can be accounted for by the graded salience hypothesis, however, as well as by the traditional view. For instance, in Turner and Katz (1997) and Katz and Ferretti (2000, 2001), unfamiliar proverbs took longer to read in figuratively than in literally biasing contexts. In Honeck, Welge, and Temple (1998), participants took longer to judge whether a less familiar proverb was an appropriate restatement of the paragraph topic than its literal counterpart. Looking into familiar and unfamiliar producer-for-product metonymies with literal expressions, Pickering and Frisson (2001a) found that, whereas familiar metonymies (read Dickens) incurred no difficulty, expressions that can be interpreted as unfamiliar metonymies (read Needham) did. (Notwithstanding, introducing Needham as an author in prior context results in no difficulty for unfamiliar metonymies.)

Another counterexample to the direct access model is found in Gibbs (1990), who showed that understanding figurative referring expressions took longer to understand than literal referring expressions retrieving the same antecedent. In his study, participants read short narratives ending with either a figurative (metaphoric or metonymic) or a literal referring expression (see following example 2). Participants were fastest at reinstating the antecedent of the literal description. They were also faster at reinstating the antecedent of the metaphoric expression than the antecedent of the metonymic expression. These findings are consistent with the predictions of the traditional view, which posits the priority of literal meanings. They are also accountable by the graded salience hypothesis, which posits the superiority of salient meanings (which here coincide with literal meanings). While the intended meaning of the literal referring expressions is salient (conventional), the intended figurative meanings of both the metaphoric and the metonymic referring expressions is somewhat novel. The difference in reading times, then, might have been induced by salience difference (although this has not been controlled):

Onishi and Murphy (1993) attempted to explain Gibbs's (1990) findings in terms of information structure. Using different materials, they first replicated Gibbs's findings concerning metaphoric versus literal referring expressions. However, they also manipulated information ordering by placing the highly informative message in predicate, noninitial position (as in “He's such a creampuff that he didn ‘t even show up” said Tracey; “He's such a loser that he didn't even show up,” said Tracey) rather than in topic, initial position as did Gibbs (as in The creampuff didn't even show up). When the information was presented noninitially, equal reading times were obtained for literal and metaphoric targets. Onishi and Murphy (1993: 770) concluded that the longer reading times found for metaphoric than for literal referring expressions in Gibbs's study “may be a property not of metaphor alone, but of any reference in which the referring expression's meaning is not highly related to the properties of the referent.” Recall, however, that placing targets in sentence noninitial position has been shown to mask lexical processes and neutralize differences between salient and less salient language (Giora, Peleg, & Fein, 2001; Peleg et al., 2001, in press; chapter 3 in this volume).

On the face of it, the work of Blasko and Connine (1993) could also be problematic for the direct access view, showing that tne salient, literal meaning of metaphors was always activated initially. In the case of less familiar metaphors, it also became available before the metaphoric meaning. (This was not true for less familiar but apt metaphors, however; see also Gentner & Wolff, 1997). In all, however, their findings cannot forcefully argue against the direct access view, since their contexts are very dull and cannot induce context effects (for a more thorough critique of Blasko & Connine, see chapter 5 in this volume; Gibbs, 2001; Glucksberg, 2001). This, however, has been remedied by Brisard et al. (2001), who showed that novel metaphors embedded in a rich and supportive context are interpreted literally first.

These findings tie up with findings by Blank (1988), Gregory and Mergler (1990), and Pexman, Ferretti, and Katz (2000). These studies examined processing strategies of conventional versus novel metaphors, and their findings suggest that, as predicted by the graded salience hypothesis, comprehension of unfamiliar metaphors involves more complex inferential processes. In contrast, conventional metaphors are understood as fast as literals.

More compelling are findings by Williams (1992), who showed that both the literal and the metaphoric meanings of conventional, metaphor-based polysemies were accessed initially regardless of context; these meanings remained active even after a long delay (for an extensive review, see chapter 5 in this volume). These findings, showing that the salient (literal) meaning of the metaphoric word was active for a while, allude to the possibility that the literal meaning of the sentence as a whole was activated and retained. Such findings cannot be accounted for by the direct access view (Gibbs, 2001).

McGlone, Glucksberg, and Cacciari (1994) also examined figurative language which varied in what I would term “degree” of salience. They provided evidence for the parallel process of figurative language and showed that, in understanding idioms such as spill the beans, but particularly in understanding less conventional (variant) idioms such as didn't spill α single bean, the literal meanings of the words themselves and the idiomatic meanings are simultaneously apprehended. McGlone et al. suggest that the hypothesis that two kinds of meanings are simultaneously apprehended also accounts for the findings in Peterson, Burgess, Dell, and Eberhard (1989).

Peterson et al. found that in literally biasing contexts (kick the ball), lexical decisions to concrete words were faster than to abstract words. In contrast, in idiomatically biasing contexts (kick the bucket), there was no difference between concrete and abstract targets. Moreover, lexical decisions to both concrete and abstract targets were faster than the decisions to abstract targets in the literally biasing contexts. McGlone et al. suggest that in the idiomatic context, both abstract and concrete noun targets are primed, because an idiomatic phrase such as kick the bucket has both a concrete (‘bucket’) and an abstract (‘die’) meaning.

Keysar and Bly's (1995, 1999) findings are consistent with the hypothesis that it is the more salient (whether literal or figurative) meaning that is activated first. They showed that transparency of idioms is a consequence of familiarity. In their research, participants became acquainted with the original and the contrived meanings (which were sometimes the opposite of the original meanings) of unfamiliar idioms. Results showed that it was the recently learned (and hence the more salient) meaning that was perceived as more transparent, regardless of whether participants learned the original meaning or its opposite.¹Close

Another counterexample to the direct access view is Gernsbacher, Keysar, Robertson, and Werner (2001). They show that understanding metaphors such as My uncle's surgeon is a butcher involves activating the literal meaning of the vehicle. Such a metaphor facilitated the reading times of the literal meaning of the vehicle Butchers use knives to a greater extent than did the literal prime My father's brother is a butcher.²Close

Recent research into ironic language also suggests that, contrary to the direct access view, irony comprehension involves more complex processes than comprehending the same utterances intended literally. A reanalysis of Gibbs's (1986b) findings (see Giora, 1995), as well as findings by Dews and Winner (1997, 1999), Giora and Fein (1999a), Giora, Fein, and Schwartz (1998), Pexman et al. (2000), and Schwoebel, Dews, Winner, and Srinivas (2000), reveal that less and nonsalient ironic interpretations take longer to process than salient, literal ones. Dews and Winner (1995, 1997, 1999) and Giora et al. (1998) show that irony comprehension involves processing the salient, literal meaning as well (for a similar view, see also Bredin, 1997). Moreover, findings in Giora et al. (1998) and in Giora and Fein (1999a) attest that the salient literal meaning of irony is induced initially by both familiar and unfamiliar ironies. However in the case of familiar ironies, this meaning is activated alongside the similarly salient ironic meaning, while less familiar ironies give rise to the compatible (ironic) interpretation only after a delay (Giora & Fein, 1999a; chapter 4 in this volume).

Further evidence comes from Colston and Gibbs (2002) and Gibbs (1998a). Colston and Gibbs presented participants with utterances such as This one's really sharp embedded at the end of stories that biased their meaning either toward their metaphoric (3a) or toward their ironic meaning (3b):

Participants took longer to read the ironic than the metaphoric utterances. According to Colston and Gibbs, such findings suggest that metaphor and irony involve different processes (see also Winner, 1988). However, an alternative explanation seems no less plausible. Because many of the metaphoric utterances involve conventional metaphors while the ironic interpretations do not, Colston and Gibbs's findings might demonstrate that salient meanings are faster to process than are less salient ones. These findings cannot be accounted for by either the direct access view, according to which, in a supportive context, salient and less salient meanings should be equally accessible (cf. Ortony et al., 1978), or the traditional view, which posits the priority of literal meanings. They are accountable by the graded salience hypothesis, which posits the superiority of salient over less salient meanings, regardless of literality or context.

In all, the evidence adduced here is only partly consistent with the traditional view, and only partly consistent with the direct access view. However, it is almost entirely consistent with the graded salience hypothesis. It demonstrates that salient meanings are not inhibited by context but are always activated upon encounter. The only finding that is not predicted by the graded salience hypothesis concerns the relative ease of processing of somewhat novel but apt metaphors compared to novel but less apt metaphors (but see the discussion in chapter 5 in this volume and in Giora, 2002b).

Gerrig (1989) provides an opportunity to examine the applicability of the graded salience hypothesis to literal language comprehension. His data concern (what is termed here) “degrees” of salience of literal language. He looked into the processes invoked by conventional versus less conventional language used innovatively. He proposed that processing conventional (salient) meanings involves sense-selection operating via a direct look up in the mental lexicon, while processing novel (nonsalient) meanings involves both sense selection and sense creation which operate in parallel, the latter depending on contextual processes. Gerrig's findings are consistent with the graded salience hypothesis (which similarly assumes two distinct mechanisms that operate simultaneously). They may be taken to show that even a heavily weighted context biased in favor of nonsalient meanings does not block salient meanings (see also Binder & Rayner, 1998, 1999; Giora, Peleg, & Fein, 2001; Peleg et al. 2001, in press; Rayner et al., 1994).

Gerrig (1989) contrasted the traditional sequential process model (labeled error-recovery) with a parallel process model of sense creation. According to the parallel process model, sense creation (4b) should not be sensitive to the time it takes to process the conventional meaning (4a) of the expression in question. Innovative interpretations of any conventional expression should involve the same processing strategies. In contrast, the sequential process model, he suggests, predicts that novel interpretations should be sensitive to the processing time of their conventional uses. If a certain conventionally intended utterance is faster to understand than another conventionally intended utterance, derivation of their respective novel meaning should exhibit the same difference.

As illustration, consider the following discourses cited in Gerrig (1989), repeated here for convenience, which give rise to both conventional and innovative interpretations of the same utterances, differing in degree of conventionality:

According to the sequential processing model, contends Gerrig, if foot race is understood faster than horse race in their conventional use (4a), it should be also understood more swiftly in their innovative use (4b). In contrast, according to the parallel process model, the reading times of both compounds used innovatively should be roughly the same, irrespective of which took longer to read in its conventional use.

The graded salience and retention/suppression hypotheses (see chapter 2 in this volume) have different predictions, however. Although a strong context is allowed to be predictive of nonsalient, contextually compatible meanings (cf. Giora, Peleg, & Fein, 2001; Peleg et al., 2001; chapters 2–3 in this volume), it further predicts that difficulty in suppressing disruptive salient meanings increases relative to increase in salience (see chapter 2). Therefore, it predicts longer processing times for high than for low salience expressions used innovatively in Gerrig's experiments.³Close

Results indeed show that sentences containing salient and innovative interpretations of low conventionality expressions exhibited no reading time difference when used conventionally or innovatively. In contrast, sentences biased toward the nonsalient, innovative interpretations of high conventionality expressions took longer to read than when biased toward their salient interpretation (for similar findings, see Coulson & Kutas, 1998; cf. chapter 6 in this volume). Such findings can be accounted for by the suppression hypothesis (cf. chapters 2 and 5, showing that highly salient meanings are more difficult to suppress than less salient ones).

Gerrig interpreted these results as disconfirming the sequential model and supporting “an elaborated” parallel process model:

[According to] an elaborated concurrent processing model, when readers understandpreempting innovations, they are dividing their resources between the processes of examining a conventional reading and creating an innovative meaning. The greater the demands of the conventional readings, the fewer resources remain for constructing the innovative meaning. If a conventional meaning is highly available, resources are divided between two processes—sense selection and sense creation—for a lengthy period of time. (1989: 199)

In spite of some different assumptions, both processing models assume that the salient (conventional) meaning should be activated on encounter, regardless of context. Findings in Gerrig (1989), then, support the view that salience plays a major role in language comprehension: salient meanings cannot be obstructed by contextual information.

Additional support for the graded salience hypothesis comes from Gibbs (1982). His findings show that the compositional, less conventional (question) meaning (‘Are you unable to be friendly?’) of utterances such as Can't you be friendly? took longer to comprehend than their conventional (request) meaning (‘Please be friendly to other people’), regardless of contextual bias. As predicted by the graded salience hypothesis, readers’ decisions as to whether a word string was or was not an English sentence was much faster when it instantiated the salient, conventional interpretation than when it instantiated the compositional, less conventional, less salient interpretation, regardless of whether it followed a context biasing the utterance toward the conventional (5b) or toward the less conventional (5a) interpretation:

In sum, the plethora of findings regarding comprehension of literal and nonliteral language is best explained by the graded salience hypothesis. The vast array of apparently conflicting findings is consistent with the view that salient meanings are always accessed automatically, regardless of either literality or contextual information. This view renders the metaphoric/literal distinction inapplicable to initial processes. As shown earlier, figurative and literal utterances involve different processes when they diverge in salience, with nonsalient meanings, whether literal or figurative, taking longer to process (e.g., less salient or novel metaphors vs. their literal interpretation, literal uses of salient idioms vs. their idiomatic use, innovative use of highly salient literal language vs. its conventional use, nonsalient ironies vs. their literal interpretation). Indeed, although both familiar and unfamiliar utterances activate their salient meaning initially, nonsalient language mostly involves an additional process of redressing the incompatible meaning. Figurative and literal utterances involve similar processes when they converge in salience (e.g., less familiar idioms and their literal interpretation, less familiar literal language vs. its innovative use, familiar ironies and their literal interpretation, familiar metaphors and their literal use). They activate their salient, literal and nonliteral, meanings in parallel, regardless of context. Thus, the graded salience hypothesis enables the reconciliation of views that have until now been in disagreement.