Priming of semantic classifications by novel subliminal prime words

Abstract

Four experiments demonstrate category congruency priming by subliminal prime words that were never seen as targets in a valence-classification task (Experiments 1, 2, and 4) and a gender-classification task (Experiment 3). In Experiment 1, overlap in terms of word fragments of one or more letters between primes and targets of different valences was larger than between primes and targets of the same valence. In Experiments 2 and 3, the sets of prime words and target words were completely disjoint in terms of used letters. In Experiment 4, pictures served as targets. The observed subliminal priming effects for novel primes cannot be driven by partial analysis of primes at the word-fragment level; they suggest instead that primes were processed semantically as whole words contingent upon prime duration.

Introduction

To what extent are subliminal stimuli processed semantically? One line of research addressing this question has employed binary categorization tasks, in which visible target stimuli are preceded by masked primes and are to be classified in one of two semantically opposite categories (Greenwald, Klinger, & Schuh, 1995). The primes are either in the same category as targets (congruent) or not (incongruent). A so-called category congruency effect is said to occur if participants respond faster or more accurately (or both) on congruent than on incongruent trials.

For example, when words are to be classified as positive or negative, target words (e.g., sunshine) are often evaluated faster and more accurately when they are preceded by evaluatively congruent prime words (e.g., priest) than by evaluatively incongruent prime words (e.g., hate; Klauer & Musch, 2003). Category congruency priming of this kind differs from so-called semantic priming (Neely, 1991) in that both congruent and incongruent prime-target pairs are categorically related: Congruent primes and targets share the same category membership (e.g., the category of positive words), incongruent primes and targets are members in semantically opposite categories (e.g., the target is a member of the set of positive words, whereas the prime is a member of the set of negative words). As pointed out by Lambert et al. (2003), a more appropriate structural analogy than the semantic-priming paradigm is Jacoby’s (1991) process-dissociation paradigm. For congruent prime-target pairs, responding on the basis of the target as well as responding on the basis of the prime lead to the correct answer. In the language of the process-dissociation method, responding in a controlled way to the target word leads to the same response as responding in the absence of control to the task-irrelevant prime word. For incongruent prime-target pairs, responding in a controlled way to the target and responding to the task-irrelevant prime in the absence of control lead to contradictory responses. Unintentional influences of the prime word are reflected in differences in performance for targets in congruent relative to incongruent pairings.

In this line of research, evidence for reproducible subliminal priming has accumulated. For example, Draine and Greenwald (1998), using a valence-classification task, found replicable category congruency effects for masked prime words selected from the same set as the target words. One way to explain the observed subliminal category congruency effects is to assume that primes are processed semantically. For example, Dehaene et al. (1998) argued that participants unconsciously apply the task set for target processing to the prime words. Congruent prime words thereby acquire the power to evoke the same response as target words, leading to facilitation, whereas incongruent prime words bias the opposite response, leading to interference.

Subsequent results suggested alternative explanations of these findings. Abrams and Greenwald (2000) as well as Damian (2001) found that primes were only effective if they had repeatedly been responded to as visible targets. In other words, there was little priming by new stimuli that had not appeared as targets in the classification task before. Abrams and Greenwald (2000) demonstrated that even fragments consisting of a few letters of prior targets sufficed to elicit category congruency effects; it was not necessary to present the entire word as prime. In a similar vein, Kouider and Dupoux (2004) showed that pseudowords formed by transposing the letters of previous targets can engender similar category congruency effects as the targets themselves when used as primes. Finally, Greenwald and Abrams (2002) found that even single consonants (repeated in a letter string, e.g., LLLLL) from prior targets (tulip) can have this effect. The explanations that were proposed for these findings differ in several respects, as explained next, but they agree that masked primes are not processed as deeply as targets, typically only at the level of word fragments of one or more letters, undermining the hope to demonstrate semantic processing of subliminal primes by means of category congruency priming.

For example, according to an account by evolving automaticity, as words are repeatedly classified, an association between the word and the appropriate response (Damian, 2001) or between the word and a more abstract response-related representation such as its response category (Abrams, Klinger, & Greenwald, 2002) is formed, curtailing the need for semantic processing of the word. The associated response-related representation is also activated, although perhaps only weakly, when a target later appears as prime. This biases the response to the current target, accounting for category congruency priming. Note that the account by evolving automaticity must postulate that associations are also formed between word fragments and response-related representations to be able to explain the above-reviewed findings.

Another explanation assumes that targets’ mental representations are strongly activated in the course of repeated classifications. In this highly activated state, a representation including semantic information about the target’s category membership can already be triggered by partial visual information such as by a masked fragment of the whole word. The triggered semantic information then interacts with the category information that is gathered from the current target in an evidence-accumulation response-selection process (Abrams and Greenwald, 2000, Broadbent and Gathercole, 1990), leading to category congruency effects in the absence of a proper semantic analysis of prime words.

In a related account, Kouider and Dupoux (2004) postulated that target identity is sometimes reconstructed from letters or fragments of a word. In their view, masked priming effects are actually effects of reconstructed, fully conscious primes.

Finally, according to the action-trigger approach by Kunde, Kiesel, and Hoffmann (2003), participants intentionally specify action triggers in elaborating the task instructions and as a consequence of practice in the task. These action triggers are templates against which target stimuli are matched perceptually. In the case of a match, the response belonging to the matching action trigger is released. Primes have the power to bias responses if they match one of the action triggers perceptually.

All of these accounts require only relatively low-level perceptual analysis of primes to explain category congruency priming. Novel primes should therefore be ineffective as primes if and when they do not share the critical visual features (letters, word fragments) that, as the case may be, are directly associated with response codes, trigger preactivated semantic representations, are used to reconstruct an intact word, or match action triggers. Recently, Naccache and Dehaene, 2001, Reynvoet et al., 2004, Greenwald et al., 2003 did however report category congruency effects for novel primes in a number-classification task. For example, Naccache and Dehaene’s (2001) participants practised quantity classifications (smaller or larger than five) with just four target numbers: 1, 4, 6, and 9. Nevertheless, unpractised numbers, 2, 3, 7, and 8 engendered significant category congruency effects in masked priming.

Can category congruency priming by novel primes be explained in terms of only partial analyses of primes? According to the above, it needs to be claimed that the critical visual features driving the effects for old primes are shared by novel primes; a claim that is difficult to refute unless it is specified a priori what features precisely can be used (Broadbent, 1987). The account by Kunde et al. (2003) involves a semantic element that allows for even more flexibility: action triggers can include templates for stimuli that were never shown. For example, if participants know that they will be shown one-digit numbers and that their task is to discriminate numbers smaller than five from those larger than five, they are likely to set up triggers that perceptually match all numbers from 1 to 4 for one response and all numbers from 6 to 9 for the other response even before any number stimuli are shown. Since action triggers are responsible for category congruency effects, this means that priming by novel primes is expected as soon as the task context makes it likely that action triggers for such primes are formed. Note that semantic analysis of subliminal primes needs not be assumed.

Nevertheless, as acknowledged by Kunde et al. (2003), this latter argument cannot explain category congruency effects for novel primes from task categories with many perceptually dissimilar members such as the set of all positive (or negative) words or the set of all small (or large) objects. And indeed, Abrams and Greenwald, 2000, Greenwald and Abrams, 2002, using valence classifications, and Damian (2001), using size discriminations, obtained little evidence for category congruency effects by novel prime words drawn from such potentially large stimulus pools.

The question addressed by the present research is whether there are category congruency effects engendered by masked novel word primes. As just reviewed and further elaborated in the General Discussion, the question is theoretically important. There are three previous failed attempts to find the effect. Two of these were reported by Damian (2001; Experiments 2 and 3), the third by Abrams and Greenwald (2000, Exp. 3). All three experiments had relatively small test power; there were 16 participants in each of Damian’s experiments and 12 participants in the group that saw novel prime words in Abrams and Greenwald’s experiment. The probability of detecting a priming effect of medium effect size ($d=d_3\sqrt{2}=0.50$; Cohen, 1988, Chapter 2.3) at the 5% level of significance with 16 (12) participants is however only 25% (21%), and for priming effects of smaller size, it is of course even smaller. Thus, the evidence against the existence of category congruency priming by masked novel prime words is not very compelling to date.

Related studies using number stimuli and quantity-classification tasks have produced mixed results. As already mentioned, a number of studies found category congruency priming by masked novel number stimuli (Greenwald et al., 2003, Naccache and Dehaene, 2001, Reynvoet et al., 2004) although such findings were criticized by Kunde et al. (2003) on the grounds that action triggers might be formed for the number stimuli as a consequence of elaborating the task instructions, explaining priming without semantic analysis of masked novel primes. Three experiments by Kunde et al. (2003; Experiments 2, 3, and 4) aimed at making it unlikely that action-trigger templates would be formed for novel number stimuli. In their Experiment 2, category congruency priming was absent for novel primes, but with 12 participants it was unlikely that a small category congruency effect would have been detected according to the above power analysis. In their Experiment 3, a reversed category congruency effect was obtained for novel primes that was however attributed to a non-semantic confounding in visual features between primes and targets. Finally, in their Experiment 4 (n = 24), primes were either digits or number words; notation format was varied between participants. Category congruency effects emerged for novel primes given notation matches (i.e., when both prime and target were digits, or both were number words), but was absent for novel and previously classified primes given a notation mismatch between prime and target. However, as recently shown by Van Opstal, Reynvoet, and Verguts (2005) when a slightly different mask was used, category congruency priming appeared even for primes and targets that differ in notation independently of prime novelty (but see Kunde, Kiesel, & Hoffmann, 2005). Finally, Reynvoet, Gevers, and Caessens (2005) recently reported small priming effects by novel number and letter primes in three experiments that they argued could not be explained by alternative explanations of the kind just reviewed and that were instead attributed to semantic processing of novel prime numbers and letters.

Whereas the studies using number and letter stimuli bear only indirectly on the question of the present paper, a number of findings suggest more directly that there might be a small amount of category congruency priming by masked novel prime words. Dehaene et al. (2004) using behavioral and neuroimaging techniques recently reported evidence for subliminal activation of visual word-form information above the letter and word-fragment level, although it remained an open question in this research whether the activated representations contain semantic information. Naccache et al. (2005) furthermore showed that subliminal presentation of novel emotional words that were to be classified as threatening or non-threatening modulated the activity of the amygdala at a long latency, larger than 800 ms, providing evidence for semantic access to emotional valence of these words, in three epileptic patients. Forster, Mohan, and Hector (2003) and Forster (2004; Experiment 4) found category congruency effects by masked novel prime words in an animal-categorization task in two experiments that did not however include a test of prime visibility. A recent study by Greenwald and Abrams (2002), based on first names and a gender-classification task, reported category congruency priming by novel subliminal prime words, but no systematic effort was made to control for confoundings between primes and targets at the level of word fragments of one or more letters. Moreover, we recently proposed a way to disentangle a semantic component from a response-related one in category congruency priming (Klauer, Musch, & Eder, in press). The semantic component of the observed masked category congruency effects was smaller than the response-related one, but unlike the response-related component, it was independent of the amount of practice that primes had previously received as targets. This led us to assume that a residual semantic congruency effect might give rise to masked category congruency priming even by novel word primes.

Three of the present four experiments use the valence-classification task, rendering it unlikely that action triggers (Kunde et al., 2003) can be set up a priori to encompass arbitrary novel primes from the large and diverse task categories of positive and negative words. Nevertheless, it is difficult and perhaps impossible to rule out conclusively that relatively low-level perceptual features of the particular used primes and targets are confounded with their category memberships. The difficulty arises from the fact that the class of potentially relevant features is very large. Most of the above approaches have however suggested that word fragments consisting of one or more letters are the critical features. In the first three experiments, we therefore control for confoundings at the letter and word-fragment level. This means that we can rule out alternative explanations of possible category congruency effects in terms of partial analysis of prime words at the word-fragment level. In Experiment 3, the effects of prime novelty and prime duration are additionally explored. In a fourth experiment, we try to rule out orthographical overlap of any kind or level between primes and targets by using pictures rather than words as targets.