Lexical access and frequency sensitivity: frequency saturation and open/closed class equivalence.

Whether closed-class words use the same lexical access route(s) as open-class words has been intensely debated recently. Differences in frequency sensitivity have been suggested as one manifestation of separable access routes. However, all the lexical decision studies have been limited by floor effects at the higher frequency ranges which could mask theoretically important differences in the behavior of the two classes. We studied lexical decisions to highand very-high frequency words of both classes using stimulus masking and speeded responses, in order to minimize floor effects, to try to reveal potential differences between the behavior of the two classes, and to contrast theories of lexical access. We did not find evidence to support the view that closed-class words have a different or special access route;. Neither word class showed any appreciable frequency effect for Kui;era-Francis frequencies of 400lmillion or greater, on either reaction time or error analyses. We did find open-class words to have somewhat faster responses than comparable closed-class words, but this may contradict some explanations of the reported word class effect (Bradley et al., 1980). Moreover, our data also show what may be word-specific influences on lexical decision times-effects which may be impossible to factor out of the word class effect in English. In order to accommodate the frequency insensitivity that we found, logogenbased models of lexical access have to be amended to include a floor on threshold settings. Resonance models (Gordon, 1983), already predict this frequency *This study was supported by NINCDS Teacher-Investigator Award l-K07 NS00721 to the first author and by NINCDS grant NS-14099 to The Johns Hopkins University. The authors thank Dr. Steven G. Lapointe, Barbara Mroz, Joan Sobkov, and Kate Van Dyke for help in creating the stimuli, Andrea Berger, Russ Carstens, and Joan Sobkov for testing subjects, Clara Marin and Antoinette Reed for help with the analysis, and Renee Gordon for editorial refinements. Dr. David Caplan and an anonymous reviewer helped enormously with a number of truly constructive criticisms. Reprint requests should be sent to Barry Gordon, Cognitive Neurology, Meyer 2-222, Johns Hopkins Hospital, 600 N. Wolfe Street, Baltimore, Maryland 21205, U.S.A. OOlO-0277/85/$6.80 0 Elsevier Sequoia/Printed in The Netherlands 96 B. Gordon and A. Caramazza insensitivity. It should be possible to distinguish between these two accounts by comparing masked and routine lexical decisions, but the unexpected word-specific effects prevented us from doing so. In languages such as English, two groups of words have been distinguished by a number of criteria. Differences in set size and membership privileges have inspired their nomenclature: The “closed” class is a small set, of about 150-250 words, which only rarely gets new members. The “open” class, in contrast, contains hundreds of thousands of words, with a constantly expanding membership. These classes show a number of other, perhaps more fundamental, differences. They play markedly different grammatical roles: the closed classcontains the free grammatical morphemes (articles, prepositions, and the like); the open class contains the nouns, verbs, and adjectives. The two classes also have different acquisition patterns (Bloom, 1970), phonological structure (Kean, 1977), different roles in word formation and in normal speech error patterns (Garrett, 1980), and contrasting impairments in aphasic speech (Caramazza & Berndt, 1985). Because the two classes have such distinct properties in so many theoretically important aspects of language use, the hypothesis has been entertained that the two classes of words are served by completely different access routes. We will term this the “dual route” hypothesis. A form of this hypothesis has been proposed by Bradley and her colleagues (Bradley, 1978; Bradley & Garrett, 1983; Bradley, Garrett, & Zurif, 1980). They have argued for two separate lexical access systems: One, a general purpose access mechanism serving both openand closed-class words; the other, a special purpose access mechanism serving only the closed-class words. We will term this view of the organization of the lexical access system the “common plus special route” hypothesis. Bradley et al. have made some fairly specific claims about how processing differs in the two systems. Their claims have been based on data from a very influential series of lexical decision (and other) tasks with normal and aphasic subjects. One operational marker that they have identified for these different processing modes is the word frequency effect. It is well established that lexical decision time varies with frequency, at least for open-class words (e.g., Gordon, 1983); generally, higher frequency words have faster decision times than lower frequency words. However, Bradley et al. reported that closedclass words did not show this effect in normal individuals, although matched open-class words did. Paradoxically, moreover, in agrammatic Broca’s aphasics, closed-class words did show a frequency effect, comparable in magnitude to that of the open-class words. Since normal closed-class access is generally Frequency sensitivity 97 assumed to be impaired in agrammatic Broca’s aphasics, Bradley et al. reasoned that their performance was due to the spared abilities of another lexical access system, one which could clearly make accurate, frequency-sensitive decisions to both openand closed-class words. Since normal subjects were presumably making use of an intact special-purpose access system, and did not show frequency effects for closed-class words, the special purpose system could not be frequency sensitive. Bradley et al. interpreted these and other data as support for the common plus specific route hypothesis, and as evidence that the two routes had to differ in frequency sensitivity. Bradley and her colleagues also studied nonword interference effects, using nonwords whose initial segments were closedor open-class words (Bradley, 1978; Bradley et al., 1980). Their normal subjects did not show any interference with nonword decisions when the initial segment was a closed-class word, but their agrammatics did. To account for this finding, Bradley et al (1980) were led to the further conclusion that access via the special closedclass mechanism must “inhibit the consequences” (p. 283) of the general-purpose access mechanism. They left the nature and site of this inhibition indeterminate. We will therefore have to speculate about possible observable consequences of this inhibition, but we can feel fairly secure about two points. First, Bradley et al. must have felt that the lexical decision task could index lexical access through either system, since they accepted its results in normals (where they were presuming closed-class performance was mediated by the special-purpose system) and in agrammatics (where they presumed that only the general-purpose route was preserved). Second, in normals, decisions about very high frequency closed-class words are no faster than those for lower frequency open-class words, as both Bradley (1978) and Gordon and Caramazza (1982) showed. But Bradley’s hypothetical generaland specialpurpose processors are both intact in normal subjects. Being frequency sensitive, the general-purpose processor should have been able to process these very high frequency closed-class words faster than the special-purpose processor could, and presumably make a faster decision about them than for lowerfrequency words. Since decisions about these words are not faster, we must conclude that the special-purpose route is identifying closed-class words (perhaps not individually, but as a class) and then blocking processing in some fashion through the general-purpose route. This is, of course, the same conclusion Bradley et al. came to from their nonword interference results. It should be noted, however, that we may be focusing too narrowly on frequency sensitivity (or lack of it) as a means of distinguishing between possible lexical access mechanisms. One reason for de-emphasizing frequency as a distinctive characteristic of the different processing systems is that the data on frequency dependence that Bradley et al. adduced to support dual-ac98 B. Gordon and A. Caramazza cess models have been called into question. Lexical decision studies in English (Gordon & Caramazza, 1982, 1983), French (Sequi, Mehler, Frauenfelder & Morton, 1982)) Dutch (Kolk & van Grunsven, 1981)) and German (Friederici & Heeschen, 1983; Heeschen, Friederici & Drews, 1984) with normal subjects have been interpreted as not showing any appreciable differences in the frequency sensitivity of openand closed-class words, for normal subjects. Gordon and Caramazza (1983) also failed to find frequency related differences in closed-class lexical decisions between agrammatic Broca’s and other aphasic patients. Therefore, frequency may not be the critical variable distinguishing the two hypothetical access systems. Operationally, the two systems might be distinguished by either their dependence upon lexical access codes, or by their lexical access speeds. For example, in -a reading task, access to the lexical representations of high frequency open-class words might take place directly from the visual representation (cf. McCusker, Hillinger, & Bias, 1981), while access to closed-class word representations might depend upon phonological recoding first. This would actually be an example of a difference in both codes and speed, since it is entirely possible that phonological recoding might take longer than direct access (McCusker et al., 1981). It is also possible to postulate a pure speed advantage for one class over the other. There may be reasons why one access route should be faster and more efficient than the other. Perhaps one route has to be optimized to meet the demands of on-line speech processing. Or perhaps one route could benefit because it need only identify a small number of items, and therefore it can operate with either higher efficiency or greater speed, or both. We have deliberately not identified which access system we think should have which property, because we wish to emphasize more general points and experimental opportunities: While there could well be a number of differences postulated between the two access systems, many of them might be expressed as differences in speed. Even coding differences might well show up as differences in the speed of stages of coding/access. This being the case, then, a number of hypothetical differences between openand closed-class processing routes may be testable by two modifications of the straightforward lexical decision task: masked stimulus presentation, and a response deadline (Pachella, Smith, & Stanovich, 1978). These changes might simply exaggerate normal processing differences. Presenting the stimulus string under tachistostopic, masked, stimulus-limited conditions, for instance, should limit the orthographic information available to make the decision, and so penalize whichever access route is less optimal or whichever is slower. Imposing a response deadline should have a similar effect. It is also possible that these different experimental demands might force the subject to change his/her Frequency sensitivity 99 lexical decision strategy from what it would normally be. For example, a subject struggling to meet a response deadline might be tempted to monitor the special, rapidly-working closed-class access system Bradley et al. proposed, thus giving those words an advantage. Clearly, we cannot make strong a priori arguments about what the outcomes from these tasks should be, since we cannot reason from a wellspecified base of proposed openand closed-class differences. Even so, we think it is clear that even a negative result from these types of experimental manipulations could be valuable for setting limits on how empirically distinguishable the different explanations can be, in the absence of any other guides. In addition, masking and probed response experiments with these high frequency words will address both a practical problem in earlier experiments, and what may be a significant theoretical point about the origin of the frequency effect. Both the Gordon and Caramazza (1982) and the Kolk and van Grunsven (1981) experiments found a flattening or absence of frequency sensitivity for words with frequencies at or above about 400/million (log frequency of about 2.6). Therefore, whether or not there were true differences between openand closed-class words in this range could not be tested fairly. A masking and/or a speeded response method could test if this limit was due to peripheral rather than central factors. If an upper limit on perceptual appreciation speed had been the limiting factor under normal viewing conditions, limiting the available visual information by masking could eliminate this bottleneck. Similar reasoning would apply for the speeded respo ,! se situation, if motor response speed had been the floor factor. If neither manipulation succeeds in removing the floor on reaction times for very high frequency words, then this floor most likely represents a true central saturation of the word frequency effect. This finding would be relevant for theories of the lexical access process itself. The major class of such theories we will consider here are those with parallel access’ such as Morton’s (1970, 1982) logogen model and Gordon’s (1981, 1983, in press) resonance model. In the version of Morton’s model2 developed by Coltheart and his colleagues (Coltheart, Davelaar, Jonasson, & Besner, 1977), lexical access occurs through logogens, counting devices for the perceptual evidence for a word (or other unit). A logogen is incremented whenever there is an input ‘Sequential access theories such as Forster’s (1978) arc the other major class. However, these theories have not been developed sufficiently to be directly applied to our experimental tasks and concerns. *We should note that Morton (1982) himself does not contend that logogens are responsible for the frequency effect. 100 B. Gordon and A. Caramazza which satisfies one of its defining attributes. The logogen signals its activation to subsequent stages only when the accumulated count exceeds some critical threshold value, which is different for each logogen. Coltheart et al. postulated that highand low-frequency word logogens accumulate information at the same rate, but that the frequency effect arises from variations in threshold settings. The logogens for the higher frequency words have lower thresholds, and so register the presence of their words more rapidly (with less evidence). However, Morton (1970) has also suggested that the thresholds are under some degree of subject control. Specifically, he argued that subjects might lower all their thresholds to make the system more sensitive, if perceptual factors such as stimulus degradation slowed the rate of information accumulation. If we were to find a flat frequency response, this would imply that the relevant logogens were all at the same, low, threshold setting. If we were to additionally find a flat function for very high frequency words under clear presentation conditions, that remains flat with stimulus degradation, this would imply that there is a lower limit on the threshold settings. Morton (1970) anticipated this possibility, but to our knowledge it has not been tested. In Gordon’s (1981, 1983, in press) resonance model, the rate and degree of activation of internal word units is proportional to their frequency. The subject sets a threshold for responding “yes” to this activation in a lexical decision task. Together, these factors imply that a saturation of the frequency effect would occur. But this awaits empirical verification. To summarize, the primary motivation for this study is that, if the notion of separate openand closed-class access routes has any substance, we expect to see differences emerge between the classes under either tachistoscopic, masked presentation, or with speeded responses. Furthermore, finding a saturation of the frequency effect, particularly one independent of masking or response speeding, would be of additional importance for the logogen and resonance accounts of lexical access mechanisms. Experiment 1: Tachistoscopic, masked presentation