The Onset and Time Course of Semantic Priming during Rapid Recognition of Visual Words

Renske S Hoedemaker; Peter C Gordon

doi:10.1037/xhp0000377

. Author manuscript; available in PMC: 2018 May 1.

Published in final edited form as: J Exp Psychol Hum Percept Perform. 2017 Feb 23;43(5):881–902. doi: 10.1037/xhp0000377

The Onset and Time Course of Semantic Priming during Rapid Recognition of Visual Words

Renske S Hoedemaker ^1,², Peter C Gordon ¹

PMCID: PMC5403553 NIHMSID: NIHMS837774 PMID: 28230394

Abstract

In two experiments, we assessed the effects of response latency and task-induced goals on the onset and time course of semantic priming during rapid processing of visual words as revealed by ocular response tasks. In Experiment 1 (Ocular Lexical Decision Task), participants performed a lexical decision task using eye-movement responses on a sequence of four words. In Experiment 2, the same words were encoded for an episodic recognition memory task that did not require a meta-linguistic judgment. For both tasks, survival analyses showed that the earliest-observable effect (Divergence Point or DP) of semantic priming on target-word reading times occurred at approximately 260 ms, and ex-Gaussian distribution fits revealed that the magnitude of the priming effect increased as a function of response time. Together, these distributional effects of semantic priming suggest that the influence of the prime increases when target processing is more effortful. This effect does not require that the task include a metalinguistic judgment; manipulation of the task goals across experiments affected the overall response speed but not the location of the DP or the overall distributional pattern of the priming effect. These results are more readily explained as the result of a retrospective rather than a prospective priming mechanism and are consistent with compound-cue models of semantic priming.

Keywords: semantic priming, ocular response tasks, lexical decision, response-time distributions, eye tracking during word reading

Semantic priming refers to facilitation in the processing of a word when it is preceded by a related word. Meyer and Schvaneveldt (1971) first demonstrated this effect for response times in the lexical decision task (LDT), in which participants make speeded judgments categorizing letter strings as words or nonwords. Since then, semantic priming has become a staple phenomenon in the study of cognition (McNamara, 2005; Neely, 1991). Its effects are robust as measured in a variety of isolated word recognition tasks including LDT, word naming, and semantic categorization (Neely, 1977; 1991; Hutchison et al., 2013; De Wit & Kinoshita, 2014; 2015a). Patterns of priming have played a fundamental role in the development of models of language processing and memory (Anderson, 1983; Collins & Loftus, 1975; McRae, De Sa, & Seidenberg, 1997; Rumelhart & McClelland, 1986; Ratcliff & McKoon, 1988; Masson, 1995; McNamara, 1992; Plaut & Booth, 2000), and semantic priming effects are often used as a tool for assessing other cognitive and psychological phenomena (McNamara, 2005).

The nature of the mechanisms by which semantic relations affect word recognition times continues to be a topic of investigation. A substantial body of the research on semantic priming has focused on distinguishing automatic and strategic priming effects (e.g., Neely, 1977; 1991; Hutchison, 2003), in which automatic processes are defined as fast and unaffected by intention or awareness, and strategic processes as slower, intentional, and consciously controlled (Posner & Snyder, 1975). In addition, substantial efforts have been made to distinguish prospective priming processes, those that begin before the target word in a prime-target pair is encountered (Balota, Yap, Cortese & Watson, 2008; Yap, Balota & Tan, 2013; Neely, 1977), from retrospective priming processes that only begin after a target word has been encountered (Neely & Keefe, 1989). As discussed by Thomas, Neely and O’Connor (2012), the prospective-retrospective distinction is neither parallel nor orthogonal to the automatic-strategic distinction. Prospective priming may result from the automatic pre-activation of a prime word’s related targets (Balota et al., 2008; Neely, 1977; Hutchison, Neely & Johnson, 2001) or from strategic processes that generate expectations about the target based on the semantic properties of the prime (e.g. Becker, 1980; Neely, Keefe & Ross, 1989). Similarly, retrospective priming processes may occur because the prime-target relation contributes to the development of a compound cue that can facilitate retrieval of the target word (Ratcliff & McKoon, 1988), or because prime-target relations are strategically recruited to facilitate target resolution. For example, during lexical decision the presence of a prime-target relationship can be used as a strategic cue that the target is a real word, a strategy referred to as retrospective semantic matching (Forster, 1981; Neely, 1977; Neely et al., 1989; Stanovich & West, 1983) or post-lexical coherence checking (De Groot, 1984).

Distributional effects of semantic priming

Analyses of the effect across the full RT distribution have generated new insights regarding the time course and underlying cognitive mechanisms of semantic priming. Ex-Gaussian distributions are typically a good fit for distributions of RTs in both isolated word recognition tasks and eye movements during sentence reading (Balota et al., 2008; Balota & Yap, 2011; Staub & Benatar, 2013, Staub, White, Drieghe, Hollway, & Rayner, 2010, Staub, 2011; White & Staub, 2012). The ex-Gaussian distribution is a convolution of a Gaussian and an exponential distribution described by three parameters (Ratcliff, 1979). The mean and standard deviation of the Gaussian portion of the distribution are described by μ and σ respectively, and τ represents mean and standard deviation of the exponential component of the distribution. Changes in μ reflect distributional shifts that maintain the general shape of the distribution, changes in σ represent changes in RT variability of the Gaussian component of the distribution, and changes in τ represent changes in the exponential component of the distribution, which reflects the amount of skew. Although the resulting distributions are shaped somewhat differently, an increase in the σ and/or τ parameters reflects an increase of the magnitude of the effect across the slow tail of the distribution. The ex-Gaussian distribution is not based on a theory of response time, so the mapping of distributional parameters to cognitive processes must be supported with additional theoretical and empirical evidence (Balota & Yap, 2011; Matzke & Wagenmakers, 2009). Nonetheless, ex-Gaussian parameter estimates allow us to capture effects of experimental manipulations across the RT distribution.

Recent literature on the effects of semantic priming on response time distributions has reported several distinct patterns. In some cases, semantic priming has been found to affect only estimates of μ, indicating a shift of the RT distribution between the related- and unrelated-prime conditions. Balota and colleagues (2008) found semantic priming resulted in a distributional shift (affecting only μ) for both speeded pronunciation (at both short and long SOAs) and lexical decision tasks (LDT, at relatively long SOAs; see also Yap et al., 2013). This distributional shift was interpreted to reflect a prospective priming mechanism. According to the prospective priming account, the prime pre-activates its related targets which results in a processing head start for related targets compared to unrelated targets. This head start mechanism would yield a semantic priming effect that is constant across the RT distribution. A pattern of semantic priming resulting solely from a distributional shift was also observed by De Wit and Kinoshita (2014; 2015a) for responses in a semantic categorization task. Similarly, the authors ascribe the distributional shift to a processing head start for related targets. According to this account, the semantic categorization decision is based on a process of evidence accumulation (see Norris & Kinoshita, 2008). The task-relevant (i.e., category-diagnostic) features of related primes overlap with those of the target, allowing the accumulation of relevant evidence about the category membership of the target to begin earlier on related compared to unrelated trials. Importantly, this account was proposed specifically for the semantic categorization task; the authors observed a different distributional pattern of priming during LDT and propose a different, task-specific mechanism in each context. In summary, patterns of semantic priming reflected solely in a distributional shift (estimates of the μ parameter) are generally considered to reflect a prospective priming process in the form of a metaphorical processing head start, although accounts vary on the precise mechanism by which this head start is established.

Nonetheless, in the majority of cases, semantic priming has been observed to affect not only μ, but both μ and either σ or τ, with the resulting distribution reflecting semantic priming effect for both fast and slow responses and the magnitude of the effect increasing as a function of response time. Balota and colleagues (2008) found that LDTs with a short SOA (250 ms) showed a priming effect on μ and σ, and when targets were visually degraded, semantic priming affected μ and τ (Balota et al., 2008; Yap et al., 2013, both experiments used an 800 ms SOA). De Wit and Kinoshita (2015a) also observed a combination of μ- and τ-based priming for a lexical decision task with visually intact targets at a short SOA using a high proportion of related trials. Furthermore, the distributional pattern of priming can be affected by individual differences in vocabulary knowledge. Yap, Tse and Balota (2009) found that semantic priming affected both μ and τ for low-vocabulary individuals on low frequency targets, whereas high vocabulary individuals showed priming in the form of a distributional shift only (μ effect) regardless of target frequency (see also Hutchison, Heap, Neely, & Thomas, 2014). This pattern combining a distributional shift and increasing effects in the slow tail of the distribution has been argued to reflect a mixture of prospective and retrospective influences of the prime. Balota and colleagues (2008) proposed that the μ+σ and μ+τ effects of semantic priming reflect a ‘race’ between bottom-up processing of the target (aided by the prospective influence of the related prime) and retrospective utilization of the prime to facilitate target processing. The authors suggest that the increased magnitude of the priming effect on slow responses (effects on the σ or τ parameter) is considered to reflect a greater reliance on related prime information for targets that are more difficult to process, for example because they are visually degraded (Balota et al., 2008; Yap et al., 2012) or in the case of low frequency targets for low-vocabulary individuals (Yap et al., 2009). Importantly, this mechanism does not necessarily reflect a conscious reliance on the prime, as a similar pattern of μ+τ based priming was observed when primes were highly masked and thus unavailable for conscious processing (Balota et al., 2008, but see De Wit & Kinoshita, 2015b).

De Wit and Kinoshita (2015a; 2015b) have proposed a somewhat different account for the μ+τ based pattern based on Ratcliff and McKoon’s (1988) compound-cue model. According to their account, slow responses allow more time for the prime to affect responses through the LDT-specific mechanism of retrospective semantic matching. While the target is being processed, information about the prime-target relationship is added to the developing compound cue, the collection of cues used to make the lexical decision. More direct evidence that semantic priming depends on a retrospective process is provided by Thomas et al. (2012), who showed that the finding of greater priming for visually degraded targets in the slow tail of the distribution was observed only for targets with a strong backward association, meaning that the prime was a strong associate of the target but not vice versa (e.g., small – shrink), and symmetrically associated pairs (e.g., east – west), but not for pairs that shared only strong forward associative connections (e.g., keg – beer, see also Hutchison, Heap, Neely, & Thomas, 2014). These findings support the notion that the increase in priming across the RT distribution depends on an active process of retrospective recruitment of information about the prime or the target-to-prime relationship in service of target word recognition, a process that cannot begin until after the target word is encountered. Although De Wit and Kinoshita (2015a,b) have proposed that such retrospective use of prime information takes place specifically in service of the word-nonword discrimination component of the LDT, similar data patterns have been observed in speeded naming tasks, which do not require a word-nonword discrimination, when targets were visually degraded (Balota et al., 2008; Thomas et al., 2012). In sum, while in some cases analyses of semantic priming effects have shown only a distributional shift, a substantial body of research shows a distributional shift plus changes in skew, thereby suggesting that semantic priming is a combination of prospective and retrospective influences.

Task goals and response speed

While ex-Gaussian distribution fits provide information about the development of semantic priming as a function of target processing time, interpretation of this information must take task factors into account. First, responses in most isolated word recognition tasks are relatively slow. For both manual LD and speeded naming tasks, response times average around 600 ms (Balota et al., 2007; McNamara, 2005; Balota & Chumbley, 1984), whereas reading times for the same words presented in a sentence take only half that time or even less (Inhoff, 1984; Rayner, 1998; Morris, 1994). This is an important consideration especially when interpreting effects of priming as function of response time. A second, related concern is that manual key presses, and arguably even vocal naming, are not familiar ways of responding to recognition of written words. The response of moving the eyes from one word to the next during reading is highly practiced for skilled readers, so that there is a tight link between word recognition and saccade execution (Engbert, Nuthman, Richter & Kliegl, 2005; Gordon, Plummer & Choi, 2013; Rayner & Pollatsek, 1989; Reichle, Rayner & Pollatsek, 2003; Reilly & Radach, 2006). In contrast, isolated word recognition tasks require participants to use a far less practiced response mode together with response mappings that have little connection to natural reading.

Ocular LDTs (Hoedemaker and Gordon, 2014a) eliminate the unusual manual response mappings typically used in LDTs with a more natural eye-movement response that leads to much faster RTs. In Hoedemaker and Gordon’s study, participants read a sequence of three letter strings and were instructed to move their eyes from one letter string to the next if the letter string was a word, but to keep their eyes still and press a button if the letter string was a nonword. Semantic relatedness of the first (prime) and second (target) words of the sequence in each trial was manipulated. Mean target processing times in the ocular LDT were shorter than those typically observed in manual LDT. Average lexical decision time for target words not preceded by a related prime was 384 ms, compared to 595 ms for the same words in the English Lexicon Project (Balota et al., 2007, average speeded naming time for the same words in the ELP was 599 ms). Nonetheless, ocular response times were found to be highly sensitive to lexico-semantic word characteristics. First, across all words in the study the word frequency effect was marginally stronger for ocular lexical decisions (R² = .27) than for the ELP’s manual lexical decisions (R² = .18). Second, gaze durations in the ocular LDT showed a statistically robust 23 ms priming effect. The magnitude of this effect was smaller in absolute terms than the 34 ms effect observed for the same prime-target pairs in the manual LDT experiment from which the stimuli were adapted (Lupker & Pexman, 2010, Experiment 4), but when considered as a proportion of the response time in the unrelated condition the ocular effect was slightly larger than the effect observed with manual response times (6.0% of the 384 ms baseline response time in the ocular task as compared to 5.6% of the 606 ms baseline in Lupker and Pexman’s manual task). In sum, the ocular LDT has an advantage over manual LDT because its response mapping (a forward saccade to indicate successful recognition of a word) closely resembles normal reading while maintaining the task goals of an LDT and showing robust sensitivity to lexico-semantic word characteristics. As a result, this task allows us to assess a portion of the word recognition response time distribution that is not accessible to manual response tasks.¹

Mean target processing times in the Hoedemaker and Gordon’s (2014a) ocular LDT showed a robust semantic priming effect. However, contrary to the results of prior literature using manual response tasks, Hoedemaker and Gordon (2014a) did not observe a priming effect on μ. Instead, the semantic priming effect was concentrated in estimates of τ, indicating that the prime effect primarily affected slower responses, with the influence of the prime increasing gradually as a function of response time. As discussed, previous observations of priming effects on μ (indicating a distributional shift) have been interpreted to reflect a metaphorical head start or prospective priming mechanism. However, the faster response times afforded by the ocular response mode allowed a portion of the LDs to be completed before the time needed to show significant priming, showing that the observation of a distributional shift depends in part on the response time floor that is dictated by the measure that is used. Furthermore, ex-Gaussian analyses provide information about the development of semantic priming over time, but they do not provide an estimate of the earliest moment at which the prime has an effect on behavior. Survival analysis of fixation durations during reading complements ex-Gaussian distribution fits by providing information about the earliest time point at which effects such as priming may be detected (Reingold, Reichle, Glaholt & Sheridan, 2012). With this method, survival curves are computed for each 1 ms bin over a time window by determining the proportion of fixations that are slower than the time of the bin (i.e., fixations that ‘survive’ as they have not yet been terminated by a saccade). The earliest point at which there is a discernable difference between the curves for two different conditions is known as the Divergence Point (DP). Using a combination of ex-Gaussian and survival analyses, Reingold, Sheridan and colleagues have found that a variety of factors have relatively fast-acting effects on first-fixation duration during normal reading (DPs ranging from 139 ms to 145 ms) including word frequency (Reingold et al., 2012), predictability (Sheridan & Reingold, 2012a) and lexical ambiguity (Sheridan & Reingold, 2012b).

Current study

The current study follows Hoedemaker & Gordon (2014a) in using fast ocular responses to determine the minimum duration at which semantic priming affects behavioral responses to word recognition (using survival analysis) and to assess how the influence of the prime varies as a function of response time (using both survival and ex-Gaussian analyses) and task-based goals. Experiment 1 uses the ocular LDT (Hoedemaker & Gordon, 2014a) in order to establish the divergence point and distributional pattern of semantic priming when the metalinguistic judgment of lexical status is indicated by movement of the eyes. Experiment 2 tests participants with the same set of words as Experiment 1 but replaces the LDT with an episodic recognition task in which each word in a trial set must be encoded in preparation for an episodic recognition probe immediately following the set. The episodic recognition task allows us to access an even earlier portion of the response time distribution than is available using ocular LDTs, and analyses of the divergence point and distributional pattern of priming across tasks allows us to distinguish between the effects of task goals and response speed. If the distributional pattern of effects observed in manual LDTs (priming reflected in both μ and σ or τ) reflects a retrospective matching procedure applied specifically in service of the word-nonword discrimination task (Balota et al., 2008; De Wit & Kinoshita, 2015a b), we should not observe this effect on single-word reading times when the task only requires encoding for subsequent episodic recognition. In an attempt to further distinguish between prospective and retrospective priming mechanisms, we adapted Thomas et al.’s (2012) approach and varied the degree of forward and backward association strength. As discussed above, prospective priming mechanisms are hypothesized to rely on forward (prime-to-target) associative connections, whereas retrospective priming mechanisms make use of backward (target-to-prime) associations. Therefore, an effect of either forward or backward association strength would provide further evidence for a prospective or retrospective priming mechanism respectively.

Experiment 1

Experiment 1 uses the ocular LDT to assess the onset and distribution of semantic priming in a lexical decision task. As in Hoedemaker and Gordon (2014a), participants were presented with sets of letter strings and on each string made a lexical decision by moving their eyes to the next string in the set to indicate ‘word’ and by keeping their eyes still and pressing a button to indicate ‘nonword’. The task and stimulus presentation in the current experiment were identical to those used by Hoedemaker and Gordon, except that the current experiment presented sets of four words instead of three. The current experiment had three goals: to establish the extent to which this new stimulus set shows the distributional pattern of priming during ocular LDT observed by Hoedemaker and Gordon, to determine the earliest point at which an effect of priming may be observed on fixation durations in the ocular LDT, and to assess how the strength of the forward and backward prime-target connections affects the magnitude of the priming effect. The results of Hoedemaker and Gordon lead to the prediction that the magnitude of the priming effect will increase across the slow tail of the distribution, supporting the notion that priming depends on a mechanism that relies on the availability of target information. Hoedemaker and Gordon did not observe μ-based priming, indicating that a portion of LDs was completed faster than the time needed for priming to affect response time. Based on these earlier findings, the divergence point of the semantic priming was predicted to occur after the start of the distribution, allowing a portion of responses to occur before the divergence point. Together, these results would provide evidence for a priming mechanism that depends mostly on information about the target. Thomas et al.’s (2012) found that priming effects observed in the degree of distributional skew depend crucially on the presence of target-to-prime (backwards) associative connections, suggesting the effect depends on a retrospective priming process. Consistent with this hypothesis, we expect greater priming for items with stronger backward associative connections.

Method

Participants

A total of 33 undergraduate students from the University of North Carolina at Chapel Hill participated in the experiment for course credit. All participants were native speakers of English with normal or corrected-to-normal vision. One participant was excluded from all analyses because of high skipping rates (over 40% of trials included at least one skip), leaving a total of 32 subjects in the analysis.

Materials and design

The stimulus words were presented in sets of four, with the experimental primes and targets appearing in the first and second position of each set. For the purpose of all three experiments in the current study, a total of 160 associatively related prime-target pairs were selected from the Semantic Priming Project’s (SPP) lexical decision database (Hutchison et al., 2013) on the basis of showing strong associative priming in the SPP study (mean standardized priming effect: z = .42, sd = .20, range: .20 – 1.10). In addition, the prime-target pairs were selected to represent a range of forward (.01 – .83) and backward (.00 – .82) association strengths based on the Nelson, McEvoy, and Schreiber (1998) association database. The word frequency of primes and targets is reported as the log10 of the number of occurrences per 51 million (SUBTLEXus; Brysbaert & New, 2009) and averaged 3.02 (range 1.30 – 4.87), mean length was 5.96 letters (range 4 – 14 letters) and mean orthographic neighborhood size was 4.14 (range 0 – 28 neighbors). All experimental pairs are provided in the appendix. To keep the duration of the lexical decision task in the current experiment under 45 minutes while accommodating the use of filler trials required for the LDT design, a total of 120 experimental prime-target pairs were randomly selected from the full set. For the selected pairs the mean standardized priming effect was z = .42 (range: .20 – .97), primes and targets had a mean word frequency of 3.04 (range 1.3 – 4.87), mean length was 6.08 letters (range 4 – 14) and mean orthographic neighborhood size was 4.18 (range 0 – 28). To create the stimulus lists, each prime was re-paired with a different prime’s related target in order to create a set of unrelated prime-target pairs. The related and unrelated pairs were divided into two lists that were shown to difference participants. Each list contained every target, half preceded by a related prime and half preceded by an unrelated prime, and no words were repeated within a list.

A word or a pronounceable nonword was added in the post-target position for each prime-target pair, so that half of the related and half of the unrelated pairs were followed by a nonword. For those pairs followed by a word in the third position, a word or nonword was added in the fourth and last position, distributed equally across related and unrelated trials. Filler trials were added so that on 20% of the trials (60 trials per list) a nonword appeared in the first (prime) position, and on 40% of the trials (120 trials per list) a filler word appeared in the first position followed by a nonword in the second (target position). As a result, there was always a .5 probability of a nonword appearing in the second, third or fourth position given that the previous position contained a word. The filler words were selected from the Nelson et al. (1998) association database and nonwords were selected from the English Lexicon Project (Balota et al., 2007). The experimental and filler words were equivalent in mean frequency, t(446) = −.77, p = .44 and the experimental words, fillers and nonwords were equivalent in word length, F(2,715) = 1.54, p = .22, and orthographic neighborhood size, F(2,715) = 1.61, p = .20.

Procedure

Eye movements were recorded in a dimly-lit room from the participant’s dominant eye using an SR EyeLink 1000. Eye dominance was determined using the Miles or ‘hole-in-the-hand’ test (Miles, 1929; Roth, Lora, & Heilman, 2002). Chin and forehead rests were used to minimize head movements. The stimuli appeared on a 22″ Samsung LCD monitor at a viewing distance of 57 cm with a 120 Hz refresh rate and a 1680x1050 display resolution. Use of a 20-point monospace font rendered each letter about 11 pixels wide; one degree of visual angle spanned approximately 3 characters. A 9-point calibration procedure preceded each experimental session. After initial calibration, each experimental session started with 10 warm-up trials. These warm-up trials did not contain any of the words used in the experimental list and were excluded from all analyses. Following the warm-up trials, all experimental trials were presented in random order in a single block. Each experimental session lasted approximately 30 minutes.

Operation of the gaze-contingent display is depicted in Figure 1. The start of each trial was marked by a fixation point on the left side of the screen. Once this point was fixated, the next screen appeared containing four masks made up of hash marks. The first mask appeared six blank character spaces to the right of the fixation point, and the subsequent masks were separated by two blank character spaces. Gaze-contingent invisible boundaries were placed in between each mask. The gaze contingencies were set to prevent parafoveal processing and rereading of the non-fixated words. Each word was unmasked only when the eyes entered its region on the screen from left to right. Once the eyes left the word across the right boundary (thus simultaneously entering the next region and unmasking the next word), the mask reappeared and the word was no longer visible regardless of whether the participant made any regressive eye movements. Participants were instructed to read the four words silently, and for each letter string decide whether it was a word or a nonword. Each time they decided a letter string was a word, they were to indicate this by moving their eyes as quickly as possible to the next letter string in the set (or to the final hash mark in the case of the fourth word). They were instructed to keep their eyes still and use a speeded key press on a hand-held console each time they decided the string was a nonword. This key press ended the trial. In the case of a correct decision on the final word, the words “Correct! Please press the button to proceed to the next trial” appeared in response to fixating the final fixation point. The word ‘INCORRECT!’ was presented after an incorrect eye movement (i.e. making a forward saccade to the next letter string in cases where the currently fixated string was a nonword) or an incorrect button press (i.e., pressing the button while fixating a real word). Any incorrect response ended the trial regardless of which word position had been reached at that moment.

Presentation of stimuli in the ocular LDT (Experiment 1) on the gaze-contingent display. For each letter string in the set, participants were instructed to move their eyes to the next string to indicate ‘word’ and to hold their eyes still and press a button to indicate ‘nonword.’ The dashed vertical lines represent the invisible boundaries used to trigger the gaze-contingent display changes. Gaze contingencies were set up to prevent rereading of previously-seen words and preview of upcoming words.

Analysis of eye movements

Fixations shorter than 80 ms and within 1 degree of a longer, immediately subsequent fixation were merged with the longer fixation by an automatic procedure in the EyeLink software. Trials on which either the prime or the target was skipped (6.4% of critical trials) were removed from the analyses, as were trials on which a boundary was inadvertently triggered due to a blink (.7% of critical trials) or because the eye fixated on or very near the boundary instead of on the word (4.9% of critical trials). Trials on which the participant regressed from the target back to the (then masked) prime rather than progressing to the post-target word were also removed (.2% of trials). The excluded trials were distributed equally across the related and unrelated conditions, with an average of 51 usable critical trials remaining in each condition. When brief delays in the display change caused a word to be unmasked slightly after the onset of the first fixation on a word (13.9% of words), the timestamp of the fixation onset was adjusted to reflect the onset of the word display, excluding any time the participant was fixating the mask rather than the word; the adjustments averaged 7 ms (range 1 – 71 ms). Finally, a reading time cutoff was determined at three standard deviations above the mean for each position in the triplet in each relatedness condition. Gaze durations longer than the relevant cutoff were removed, affecting 1.8% of all words on critical trials, equally distributed across the related (1.7%) and unrelated (1.8%) trials.

Gaze duration (GZD) is the sum of all first-pass fixation durations on a word. In the context of the ocular LD task, gaze duration was interpreted as a measure of both encoding and lexical decision time. Results for two other widely-used measures of first-pass reading are reported for completeness. Single-fixation duration (SFD) is the fixation duration for those words that received only one first-pass fixation. First-fixation duration (FFD) is the duration of the first fixation on a word regardless of the total number of fixations on that word. However, GZD is our primary measure of interest for several reasons. First, GZD is widely used as a measure of lexical encoding in eye-tracking studies of reading (Inhoff, 1984; Morris, 1994; Rayner, 1998) and although there are arguments for using SFD or FFD instead of GZD, both measures also have substantial drawbacks. Although SFD may be the most straightforward eye movement measure of word encoding, limiting our analyses to trials in which the target received a single fixation would result in a large amount of data loss (the target received more than one fixation on over 30% of all critical trials). Previous studies examining distributional effects of lexical characteristics on eye movements sometimes focus on FFD instead of GZD, but this measure can be considered less stable than GZD as the decision to refixate a word may be influenced by non-lexical factors such as the initial landing position within the word or other oculomotor targeting errors (Rayner, 1998). Finally, because participants were instructed to indicate a ‘word’ decision by moving their eyes to the next letter string, GZD in this context is operationally equivalent to button press RT during manual LDT.

Survival curves were computed for gaze duration and first-fixation duration on the target word in the related and unrelated conditions. For each 1 ms time bin within a 1 ms – 1040 ms window, the proportion of reading times that was longer than the time bin was considered the proportion of ‘surviving’ fixations. Survival curves were computed separately for each participant, and the averaged curves are presented in Figure 2 (top row). The Confidence Interval DPA procedure of Reingold and Sheridan (2014), which uses bootstrap resampling of the data (Efron & Tibshirani, 1994), was used to determine the earliest time bin at which the proportion survival differed as a function of relatedness condition; see also Reingold & Sheridan (2014) for additional DPA bootstrapping procedures and Inhoff & Radach (2014) for an alternative DPA procedure. On each of 10,000 bootstrap iterations, the collection of data for each condition within each participant was randomly resampled with replacement and the individual participant survival curves were computed and averaged across participants. For each bootstrap iteration, the divergence point was determined as the first bin in a run of five consecutive bins on which the proportion survival in the unrelated condition was at least 1.5% greater than in the related condition (following the criteria recommended by Reingold and Sheridan, 2014). Subsequently, the 10,000 divergence point estimates were rank ordered and the median of all divergence points was used as the divergence point estimate for the sample. The 250^th and 9750^th observed divergence point values were taken as the lower and upper bound of the 95% confidence interval.

Survival plots for gaze duration and first-fixation duration in Experiments 1 and 2. The dashed line presents the divergence point (DP). The grey boxes represent the 95% confidence interval of the divergence point.

Ex-Gaussian parameter estimates for target word reading times were obtained separately for each participant in each relatedness condition using the QMPE v2.18 program (Cousineau, Brown, & Heathcote, 2004) for quantile maximum probability estimation. Quantile estimates were calculated by ranking RTs for each participant in each condition from fastest to slowest, and dividing them into 11 equally spaced bins. Ten observed quantile estimates were then generated by taking the average of the slowest trial in one bin and the fastest trial in the next bin. This approach to distributional analysis has the advantage that data from all participants is represented equally across the ten quantiles, thereby effectively standardizing the effect across the distribution. As such these analyses avoid the problems typically associated with of interpreting effects across individuals, groups or items with different baseline response latencies (Faust, Balota, Spieler, Ferraro, 1999; Hutchison, 2003).

Results

Mean accuracy across subjects on all (critical and filler) trials was 98% for words (range: 84% – 100%) and 85% for nonwords (range: 69% – 99%). These levels of accuracy are similar to those found in the English Lexicon Project for these particular words (words: 97%, nonwords: 86%). For gaze duration and first-fixation duration the average number of correct trials per participant available for analysis after trimming was 49 in each condition. The average number of available single-fixation trials was 33 in the unrelated and 35 in the related condition. Table 1 presents word reading times across the prime, target and third-word position and relatedness conditions. Mean gaze duration was 440 ms (sd = 62 ms) for unrelated words across all three positions (i.e., target words (second position) in the related-prime condition were excluded). Average ocular LD times were significantly faster than manual LDs (M = 635 ms, sd = 68), t(299) = 56.97, p < .001 and speeded naming times (M = 623 ms, sd = 46), t(299) = 52.7, p < .001, for the same words in the ELP. Mean gaze durations on individual words were correlated with manual response times in the ELP, r = .59, R² = .34 p < .001 (N = 300), but the effect of SUBTLEX frequency (Brysbaert & New, 2009) on response times was greater for ocular LDs, r = −.60, R² = .36, p < .001 (N = 300) than manual LDs as observed in the ELP database, r = −.49, R² = .24 (N = 300), Fisher’s z = 2.61, p < .01 (N = 300). The average manual response time to nonwords (correct responses only, trimmed to exclude RTs more than three standard deviations above the mean for that word position) was 710 ms (sd = 137).

Table 1.

Word reading times in Experiment 1. All times in are in milliseconds. Asterisks indicate a significant effect.

Measure	Condition	Word Position
Measure	Condition	Prime	Target	Post-Target
Word GZD (sd)	Unrelated pairs	451 (82)	425 (80)	417 (71)
	Related pairs	459 (89)	398 (76)	431 (82)
	Mean	455 (85)	411 (79)	424 (77)
	*Priming*		27*	−14*

Word FFD (sd)	Unrelated pairs	250 (59)	349 (67)	330 (64)
	Related pairs	254 (68)	333 (66)	332 (69)
	Mean	252 (63)	341 (66)	330 (66)
	*Priming*		16*	−2

Word SFD (sd)	Unrelated pairs	388 (83)	397 (81)	377 (69)
	Related pairs	386 (86)	372 (73)	385 (74)
	Mean	387 (84)	384 (77)	381 (71)
	*Priming*		25*	−8

Open in a new tab

Semantic priming

Target word reading times showed a significant effect of relatedness, such that ocular LDs for targets were faster in the related compared to the unrelated condition across all eye movement measures, GZD: t₁(31) = 4.65, p < .001, t₂(119) = 6.73, p < .001, SFD: t₁(31) = 4.04, p < .001, t₂(119) = 4.90, p < .001; FFD: t₁(31) = 3.44, p < .01, t₂(119) = 4.08, p < .001. Consistent with the priming effect on target word RTs, accuracy rates were higher for related (M = 98%) compared to the unrelated (M = 96%) target words, t₁(31) = 2.16, p < .05, t₂(119) = 2.56, p < .05. Regression analyses within individual participants showed that priming on the target was greater for prime-target pairs with higher forward association values for SFD t(31) = −2.78, p < .01 and FFD, t(31) = −3.63, p < .01, and the effect was marginal for gaze duration, t(31) = −1.84, p = .08. Priming did not vary as a function of backward association strength, ts < 1.

Gaze durations on the post-target word showed a reversed effect of prime-target relatedness, such that post-target reading times were longer after a related compared to an unrelated prime-target pair, t₁(31) = −2.44, p < .05, t₂(59) = −2.28, p < .05, though this effect was not significant for single-fixation duration, t₁(31) = −1.39, p = .17, t₂(59) = −.75, p = .45, or first-fixation duration, t₁(31) = −.32, p = .75, t₂(59) = −.50, p = .62. Response accuracy on the post-target word also showed a reversed effect of relatedness, such that accuracy was lower following a related (M = 96%) compared to an unrelated (M = 97%) prime-target pair, an effect that reached significance by subjects, t₁(31) = −2.25, p < .05, but not by items, t₂(59) = 1.64, p = .12.

Survival analysis

The divergence point for gaze durations in the current experiment was estimated to occur at 260 ms (95% CI: 250 – 277 ms). This analysis indicates that the earliest point which a semantic priming effect could be detected was 260 ms. For first-fixation duration, the divergence point was also estimated at 260 ms (95% CI: 153 – 312 ms). For each individual participant, we computed the proportion of target word reading times (across related and unrelated conditions) that were faster than the divergence point of the sample. Averaged across participants, .05 (SD = .11) of gaze durations were faster than the GZD divergence point, and .22 (SD = .16) of first-fixation durations were faster than the FFD divergence point.²

Ex-Gaussian distribution fits and quantile analyses

The quantile estimates for gaze duration and first-fixation duration are plotted in Figure 3. There was a main effect of relatedness on gaze duration, F(1,31) = 19.4, p < .001, single-fixation duration, F(1,29) = 12.69, p < .001, and first-fixation duration, F(1,31) = 8.0, p < .01. For gaze duration and first-fixation duration, there was also a significant interaction between quantile and relatedness, indicating that the effect of relatedness increased across the slow tail of the distribution, GZD: F(1,31) = 5.6, p <.05, FFD: F(1,31) = 8.3, p <.01. The effect was not significant for single-fixation duration F(1,29) = 2.66, p = .11. In order to ensure that the interaction did not depend entirely on effects occurring only in the slowest tenth quantile (which showed a great deal of variability), the ANOVA was repeated including only the first nine quantiles. The results remained the same regardless of tenth-quantile inclusion.

Quantile plot for mean gaze durations (panel A) and first-fixation durations (panel B) on the target word in the ocular lexical decision task (Experiment 1) when the target was preceded by a related or an unrelated prime. Quantile estimates were calculated by ranking RTs for each participant in each condition from fastest to slowest, and dividing them into 11 equally spaced bins. Ten observed quantile estimates were then generated by taking the average of the slowest trial in one bin and the fastest trial in the next bin. Quantiles are arranged from fastest to slowest on the x-axis. Error bars show the standard error of the quantile value across subjects and the dashed lines represent predicted quantile values based on mean parameters of the estimated ex-Gaussian distribution.

Table 2 shows the average ex-Gaussian parameter estimates for gaze duration and first-fixation duration generated by the QMPE program (single-fixation duration did not yield a large enough number of observations to allow for ex-Gaussian distribution fits). The parameter estimates were used as dependent variables in a paired-samples t-test. There was no effect of relatedness on μ, GZD: t(31) = 1.66, p = .11, FFD: t(31) = −1.37, p = .18, or σ, GZD: t(31) = −.58, p = .57, FFD: t(31) = −.05, p = .96. Estimates of τ showed a significant effect of relatedness on first-fixation duration, t(31) = 2.54, p < .02, indicating an increase in priming across the slow tail of the distribution. This effect did not reach significance for gaze duration, t(31) = 1.62, p = .12. However, the effect was significant when two subjects with the worst ex-Gaussian model fits (computed as the average discrepancy between the observed and estimated quantile estimates across the first nine quantiles, excluding the often-noisy 10^th quantile) were excluded from the model³, t(29) = 2.54, p < .05.

Table 2.

Ex-Gaussian parameter estimates in Experiment 1. All times are in milliseconds. Asterisks indicate a significant effect.

Measure	Condition	Parameter
Measure	Condition	Mu	Sigma	Tau
GZD (sd)	Unrelated	324 (70)	27 (30)	104 (55)
	Related	313 (65)	30 (31)	87 (48)
	*Priming*	11	−3	17

FFD (sd)	Unrelated	267 (57)	54 (34)	82 (64)
	Related	275 (60)	54 (34)	58 (54)
	*Priming*	−8	0	24*

Open in a new tab

We also assessed the effect of prime word reading time on the distribution of target word reading times. To do this, we ranked each participant’s prime word gaze durations within relatedness conditions and divided these into 5 equally sized bins. We computed each participants’ average target word reading time for each prime-reading time bin in each condition⁴, and used a 5 (bins 1–5) by 2 (related vs unrelated) ANOVA to assess the effect of prime word reading time on the magnitude of the target word priming effect. As can be observed in Figure 4, besides the main effect of relatedness that was already established, GZD: F(1,31) = 23.0, p < .001, FFD: F(1,31) = 9.7, p < .01, target word reading times were positively correlated with prime word gaze duration, reflecting a within-trial “rhythm” effect, GZD, F(1,31) = 56.1, p < .001, FFD: F(1,31) = 22.3, p < .001. However, the magnitude of the priming effect on target word reading times did not vary as a function of the amount of time spent on the prime, Fs < 1. Supplementary analyses showed these results were not affected by the time stamp corrections applied to account for the occasional brief delays in the gaze contingent display change.

Mean target word gaze durations as a function of the gaze duration on the prime word on the same trial in Experiment 1. Trials were divided into 5 bins for each participant in each condition based on prime word gaze duration and average target word gaze duration in each bin and condition was plotted. Error bars represent the standard error of the bin mean across subjects.

Discussion

Consistent with Hoedemaker and Gordon (2014a) we observed a robust semantic priming effect in the ocular LDT, even though average word reading times were much shorter than those typically observed using manual-response or speeded-naming tasks. In addition, the word frequency effect was stronger for ocular LDs in the current experiment than for manual LDs in the ELP (Balota et al. 2008), confirming that this measure is sensitive to lexical properties. Divergence point analysis revealed that the earliest detectable influence of the prime on target word reading times, whether measured for gaze duration or first-fixation duration, occurred around 260 ms after the target word was fixated. The timing of this DP is roughly consistent with the time scale of semantic priming as observed using EEG. Typically, the ERPs associated with target word processing in LDT are shown to begin diverging between 200 and 250 ms with the effect peaking in the N400 region (e.g., Bentin, McCarthy & Wood, 1985; Holcomb, 1988) although in some cases the effect is not reported to emerge until 300 ms after the onset of the target (Holcomb & Neville, 1990; Anderson & Holcomb, 1995; Rossell, Price & Nobre, 2003).

A possible concern about these data is that fast ocular responses include a high proportion of guesses, which would explain why the priming effect was attenuated in the fast tail of the distribution. The accuracy data showed a bias to respond ‘word’ (higher error rates for nonwords than words), as might be expected if participants adopted a liberal criterion for moving their eyes to the next letter string. However, this word bias was similar in magnitude to that found for manual LDs for the same words in the ELP (Balota et al., 2007), suggesting that response bias contributes equally to ocular and manual LD responses. Importantly, there was no relationship between individuals’ nonword error rates and the extent to which the effect of relatedness within participants was expressed in effects of priming on either μ (GZD: r = .26, p = .15, FFD: r = .11, p = .56) or τ (GZD: r = −.07, p = .69, FFD: r = −.03, p= .86). Therefore, participants who adopted a more conservative strategy did not show greater μ-based priming than participants who adopted a more liberal strategy. In addition, incorrect eye movement responses (moving one’s eyes from a nonword onto to the next letter string) were slower on average (M = 548 ms, sd = 151 ms) than correct eye movement responses (i.e., moving one’s eyes from a word onto the next letter string). This suggests that incorrect classifications of nonwords as words were not due to fast guesses or an inability to suppress fast, involuntary forward saccades. A final indication that reading times in the ocular LDT reflect LDT-related processes (rather than guesses) is the fact that we observed a reverse priming effect on the post-target (slower reading times following a related prime-target pair). This result replicates the findings in Hoedemaker & Gordon (2014a) and suggests participants adopted a stricter decision criterion following words that were easier to process by virtue of having been preceded by a related prime. Such criterion adjustments have also been observed in manual LDTs in the form of first-order sequential effects (Perea and Carreiras, 2003) and frequency blocking effects (Glanzer & Ehrenreich, 1979; Lupker, Brown & Colombo, 1997).

The 11 ms effect of priming on estimates of μ did not reach statistical significance and therefore the results of the ex-Gaussian analysis did not provide clear evidence that semantic priming caused a distributional shift, an outcome that is consistent with the results of Hoedemaker and Gordon (2014a) where the priming effect on μ was 5 ms and not statistically significant. However, across participants only 5% of gaze durations were faster than the 260 ms DP. This suggests that semantic priming affected the majority of ocular LD responses, including many of those in the fast tail of the distribution. As mentioned, additional analyses of the ocular LDT data presented in Hoedemaker and Gordon showed a DP of 270 and on average 14% of responses were faster than the DP. Thus, while this experiment and Experiment 1 of Hoedemaker and Gordon showed similar DPs, gaze durations in the earlier experiment were slightly faster than in the current data, resulting in a clear attenuation of the priming effect in the fast tail of the distribution. Taken together, these results suggest that the DP of the semantic priming effect is quite stable across experiments and measures (e.g., gaze duration and first-fixation duration), but that the observed distributional pattern based on ex-Gaussian distribution fits varies depending on the response time floor of a particular task. Response times are faster for ocular as compared to manual LDTs and the greater speed of responding is accompanied by attenuation or elimination of the priming effect in the fast tail the ocular RT distribution. Comparison of the current and previous ocular LDT experiment shows that even small differences in baseline response times (in this case likely due to average word frequency being lower in the current experiment than in Experiment 1 of Hoedemaker and Gordon) can also affect interpretations of the distributional pattern of priming.

The priming effect on estimates of τ reached significance once two participants with poor model fit were removed from the data. The notion that the magnitude of the priming effect increased as a function of response time is also supported by the quantile analysis, which shows a significant increase of the effect across quantiles. As discussed, there is no a priori mapping between distributional parameters and cognitive processes. However, the analysis of priming as a function of prime-word reading times (as opposed to target-word reading times) provides important information about the possible mechanisms driving the skew-based priming effect. Consistent with Hoedemaker and Gordon (2014a) participants adopted a within-trial rhythm such that prime and target word reading times within a trial were positively correlated. As shown in Figure 4, target reading times increased across the slower prime reading time bins, a pattern that is consistent with the previous finding. However, the magnitude of the priming effect did not increase as a function of prime reading time bin, which indicates that the magnitude of the priming effect increases with target response latency as a function of target-specific processing effort rather than trial-general processing. When target word reading time is elevated due to factors that are not specifically related to target word difficulty (such as general within-trial rhythm as also measured on prime reading time), there is no corresponding increase in priming. However, when target word reading time is elevated due to target-specific factors, we observe a greater influence of the prime. These results support Balota et al.’s (2008) account of skew-based priming effects as reflecting greater utilization of the prime when target processing is more effortful, showing that the account applies also in cases where processing effort is related to properties of the target itself (word frequency and length being likely candidates to affect processing difficulty) rather than visual degradation.

Thomas et al. (2012) proposed that active recruitment of prime information in service of target word recognition on more difficult trials depends crucially on the availability of target-to-prime backward associative connections. That relationship between the magnitude of priming and the strength of backward association between prime and target was not observed in this experiment. Instead, the data provide some evidence that the magnitude of priming increased with increases in the strength of the forward association between prime and target but that the magnitude of priming did not vary with the strength of backward associative connections. This pattern is not consistent with a mechanism in which retrospective priming is characterized as dependent on backward associations that are not available until the target word had been at least partially recognized. However, it does not rule out alternative memory-search models of priming, such as compound-cues models, that do not depend specifically on an active search for target-to-prime relations. The implications of this hypothesis will be explored in more detail in the general discussion.

Experiment 2

The goal of Experiment 2 was to assess the extent to which the pattern of priming found in the ocular LDT depends on LDT-specific processes and to explore the distribution of the priming effect in a task with an even lower response time floor. Participants read sets of four words on the same gaze-contingent display used in Experiment 1, and on each trial indicated whether a subsequently presented episodic recognition-memory probe word had been among that trial’s four words. As such, the goal of the reader was to encode primes and targets (as well as fillers) for the purpose of a relatively easy recognition-memory task. We refer to this task as the episodic-recognition task, as the only explicit response required from the participant is a yes/no recognition memory judgment to the probe word following each trial. However, the main measure of interest, word reading time on the primes and target, reflects word encoding time. Word encoding times in the episodic-recognition task do not reflect the meta-linguistic judgment required by an LD, and forward saccades reflect an implicit decision that the word has been sufficiently encoded rather than an explicit task-based response. Whereas the ocular LDT might have encouraged backwards checking for semantic relations (prime-target relatedness provided a reliable indication that the target was a word), such strategies are not applicable in the episodic recognition task. The episodic-recognition task was adapted from Brysbaert (1995) who applied it to investigate sequential effects on the encoding of Arabic numerals and from Hoedemaker and Gordon (2014b) who used it to study encoding of words. Using a similar paradigm, Deacon, Hewitt, Yang and Nagata (2000) obtained a significant N400 effect of semantic priming both when the prime was masked and when it was not masked; this suggests that the task is well-suited for the investigation of semantic priming effects on the encoding of words. If priming in the ocular LDT is driven entirely by the meta-linguistic decision-making processes required by the LD task, we would not expect to see an effect of priming on target word reading times in the recognition task. Alternatively, if the magnitude of the priming effect primarily depends on response time, the faster word reading times afforded by the current task will shift the distribution relative to the DP of the priming effect, resulting in a larger proportion of responses that is faster than the DP and thus not affected by priming. The observation of a τ-based priming effect would indicate that the influence of the prime increases for slower responses even when the decisions regarding when and where to move the eyes are driven by processes of general word recognition rather than a meta-linguistic judgment.