Readers use visual input to retrieve word representations from memory and derive meaning from the text. A critical question is whether word recognition during reading happens by directly accessing meaning from written forms, or whether this process requires activation of speech sounds (i.e., phonology). Although there is robust evidence for the role of phonology in a range of word recognition tasks (see Leinenger, 2014), there is mixed evidence for the strength of the effect in silent reading, particularly at the earliest stages of the process (i.e. phonological preview benefit – faster processing of a target word following parafoveal perception of its homophone compared to an orthographic control). A recent Bayesian meta-analysis of condition means across studies, concluded that the phonological preview benefit effect likely exists but is small (Vasilev et al., 2019), under the assumption that variation in the effect is due to it being unreliable, rather than due to systematic differences that have yet to be rigorously disentangled. In the current study, we investigated stimulus-level factors (i.e., sentence constraint and preview lexicality) and participant-level factors (i.e., spelling and phonological decoding abilities) that may impact the magnitude of the phonological preview benefit to better understand the circumstances under which readers engage in phonological recoding in order to identify words during silent reading.

Phonological codes do appear to be activated during visual word recognition, but many of the tasks used to study this phenomenon have such a long delay from when the word is first perceived to when a response is produced that the time course is difficult to pinpoint. For example, when making decisions about the meaning of a word or sentence, people are slower and less accurate to reject semantically inappropriate words if they sound like a semantically appropriate word (i.e., are a homophone of the correct word) compared to when they are equally orthographically similar to the target but do not sound like it (i.e., an orthographic control; Coltheart et al., 1988, Coltheart et al., 1994, Lesch and Pollatsek, 1998, Treiman et al., 1983, Van Orden, 1987, Van Orden et al., 1988, Jared and Seidenberg, 1991). However, an important question is whether phonological activation actually aids visual word recognition or if it is an epiphenomenal byproduct of the process; eye movements provide a real-time measure of the moment-to-moment processes involved in word recognition during reading (Rayner, 1998, 2009) and the time course of these effects can be telling.

Many studies that have used eye tracking to measure in-the-moment processing of phonology employ a proofreading task (i.e., explicitly looking for typos; Daneman et al., 1995, Daneman and Stainton, 1991, Jared et al., 1999) or “homophone foil” paradigm (i.e., only reading for comprehension even though typos are present; Daneman and Reingold, 1993, Daneman and Reingold, 2000, Feng et al., 2001, Folk, 1999, Jared and O’Donnell, 2017, Rayner et al., 1998). In general, these studies find less of a disruption to reading (e.g., shorter fixation durations, fewer regressions) for homophones compared to orthographic controls. These patterns have been taken to suggest that phonological codes are used in initial word recognition as well as in implicit repair processes, even if the task does not explicitly involve a proofreading decision. However, as a consequence of these paradigms, the readers in these studies directly encounter (i.e., fixate on) semantically incorrect words (i.e., homophones of the correct word, or orthographic controls) and are therefore generally consciously aware of the typos. Therefore, it is difficult to infer too much from these studies about natural reading because the fact that readers frequently encountered semantically incorrect words could have introduced strategic processes that made them read differently than they would under normal circumstances.

To circumvent the issues related to readers fixating on semantically incorrect words, and also to investigate an even earlier stage of the word recognition process (i.e., parafoveal processing), other researchers have turned to the gaze-contingent boundary paradigm (Rayner, 1975). In this paradigm, an invisible boundary is triggered when a person’s eyes cross it, causing the word on the opposite side of the boundary to change from an experimentally manipulated parafoveal preview to a semantically correct foveal target. Because the earliest stages of visual word recognition during normal reading can occur before a word is even directly fixated (i.e., during parafoveal preview), this paradigm can be used to investigate what types of information, and how much of it, readers extract at the earliest moments of the reading process. According to the E-Z Reader model (Rayner et al., 1998), once the currently fixated word is recognized but the eyes have not yet moved forward, attention shifts to the upcoming word and certain features of that word begin to be processed while eye movements are being programmed towards it. During this in-between time, the reader gets a head-start on word recognition of the upcoming word (see Schotter et al., 2012; Schotter, 2018), which lies outside of the fovea in lower-acuity parafoveal vision. This head-start is referred to as a parafoveal preview. Because the parafoveal preview is the earliest possible bottom-up information that can be extracted from a word, it is an ideal place to investigate how early phonological codes come online.

The increased speed with which a word is processed when a parafoveal preview was available compared to when it was denied is the preview benefit and has been estimated to be a reduction in gaze durations of about 20–50 ms (Rayner et al., 2003; for a Meta-analysis see Vasilev & Angele, 2017). Recently, differences in fixation durations between an unrelated and related preview condition have sometimes been conceptualized as a preview cost for unrelated previews (Hutzler et al., 2013, 2019; Kliegl et al., 2013; Marx et al., 2015; Vasilev & Angele, 2017), but we maintain the terminology used when the phenomena were first investigated. Many studies have found benefits for word length and orthography (see Schotter et al., 2012), and even semantics (see Schotter, 2018; Andrews & Veldre, 2019). Preview benefits related to phonological processing have also been investigated, but with mixed results in terms of whether it is observed and robust.

A phonological preview benefit was originally reported by Pollatsek et al. (1992), who compared reading times on a target (i.e., “cent” in the sentence “The generous man gave every cent to charity.”) when the preview was a homophone (i.e., “sent”) or an orthographically matched word (i.e., “rent;” they also included identical and unrelated preview conditions). They reported an average phonological preview benefit of 20 ms in first fixation duration (duration in milliseconds of the first fixation on the target word; 275 ms for the homophone preview and 295 ms for the orthographic control preview), and 14 ms in gaze duration (the duration of the sum of all fixations on the word prior to leaving). This reduction in reading time for targets with homophone previews compared to orthographically matched previews presumably suggests that parafoveally perceived words can be decoded into their phonological forms, which facilitate word recognition above and beyond the information activated from the orthographic forms.

A number of studies have followed up on the phonological preview benefit (see Leininger, 2014 for a review), and a Bayesian meta-analysis (Vasilev, et al., 2019), which included a number of unpublished studies, aimed to estimate the size of the effect. Among the studies included in this meta-analysis, some replicated the original phonological preview benefit effect (Miellet and Sparrow, 2004, Blythe et al., 2018, Jouravlev and Jared, 2016; Leininger, 2018, Experiment 4), some replicated it after segmenting the data by reading ability or age (Chace et al., 2005, Tiffin-Richards and Schroeder, 2015), and some failed to replicate it altogether (Choi & Gordon, 2014; Leininger, 2018, Experiment 3). Ultimately, Vasilev et al. (2019) concluded that there is a high probability that the phonological preview benefit effect exists (>92 %) but that the size of the effect was relatively small (i.e., 4.5 ms in gaze duration compared to the 14 ms effect originally reported by Pollatsek et al., 1992). It is not surprising that a phonological effect is small in magnitude, particularly in alphabetic languages, because the typical design of these studies involves comparisons between a homophone and an orthographic control, which contains as much orthographic overlap with the target word, but only minimally less phonological overlap. With respect to the example stimulus provided by Pollatsek et al. (1992), the orthographic control “rent” shares 75 % of its phonology with “cent,” which is only slightly less than the 100 % overlap in phonology between “sent” and “cent.” The close correspondence between orthography and phonology in alphabetic languages poses some difficulty in isolating effects that are due purely to phonological versus orthographic processing. In English, most homophones have a substantial amount of orthographic overlap. However, there is some variability in the extent of homophone orthographic similarity (e.g., steel and steal are more orthographically similar than shoot and chute), and this variability in orthographic overlap has been shown to impact the magnitude of preview effects (Milledge et al., 2022, Pollatsek et al., 1992). Therefore, in this study, we were careful to ensure that the homophone and orthographic control preview were closely matched with respect to orthographic similarity to the target (see Method sections) so that any effect of the preview condition could be interpreted as evidence of phonological processing/overlap above and beyond orthographic processing/overlap.

Related to the potentially small effect size for the phonological preview benefit, Vasilev et al. (2019) highlighted that many studies on the phonological preview benefit are underpowered and note that, “while the effect very likely reflects a true difference in the population, only high-precision experiments may be able to reliably capture it.” They provided a heatmap figure of a power analysis that suggests that, assuming an effect size of 4.5 ms in gaze duration, a study with 50 items per condition would require about 90 participants to have a power level of.80. Because there are only a limited number of homophones in the English language that are length-matched (a criterion for boundary paradigm studies to not produce a very salient visual manipulation), 50 target items is about the limit of what is possible in a study of this nature, so the only way to increase power is to collect data from an increasingly large number of subjects. The only study with that many subjects is an unpublished study by Drieghe et al. (2016), which was included in the meta-analysis but was not reported in detail so there is little information on the characteristics of the stimuli or the participant sample. Therefore, we ran a high-powered test of the phonological preview benefit and also investigated potential factors that may qualify the magnitude of the effect (see below).

The meta-analysis highlights the inconsistencies in the phonological preview benefit literature (e.g., varying effect sizes and some null results), and suggests that the variation in the effect across the individual studies is due to the effect itself being small and unreliable. However, the included studies varied in the types of homophones, sentence contexts, participant populations, and the languages used. Recognizing a word in a sentence is a multifaceted process that involves the integration of perceptual information, existing word knowledge, and the preceding sentence context, to solve the complex problem of extracting meaning from printed symbols. Therefore, it is possible that the variation in the effect size reported in these studies is due to systematic differences in the characteristics of the experimental design or participant sample. As a first step toward understanding variability in the phonological preview benefit, the current study focuses on lexical and sentential characteristics of the language stimuli that may impact the magnitude of the phonological preview benefit, and measures individual differences in the quality of the readers’ lexical representations, which may be a driving factor in the extent to which phonological representations are used during reading.

Sentence context generates expectations that preactivate information about upcoming words and boost the efficiency of word recognition for predictable compared to unpredictable words (see Staub, 2015). Predictability is determined using a cloze task (Taylor, 1953), in which participants are asked to produce the word that they most strongly expect to come next in the sentence. The cloze probability is then calculated as the probability of a particular word being produced across all participants. Eye tracking studies have demonstrated that the higher the cloze probability of a word in a given sentence context, the more likely a reader is to skip that word (Balota et al., 1985, Drieghe et al., 2004, White et al., 2005) and the shorter their fixation durations are (Inhoff, 1984, Rayner and Well, 1996). Furthermore, the semantic expectations generated by the sentence context may also feed down to pre-activate sublexical representations of orthography, phonology, or both. In fact, there is clear evidence that sentence-generated expectations do feed down to orthography; semantically incorrect words or pseudowords that are orthographically similar to expected words produce larger skipping rates than orthographically dissimilar words (Balota et al., 1985, Veldre and Andrews, 2017).

When it comes to previous phonological preview benefit studies, sentence constraint has generally not been carefully controlled or manipulated; some studies do not even report the cloze probabilities of their target words. Therefore, variability in sentence constraint across studies could interact with parafoveal processing of phonological codes, leading to variability in the effects across studies. For example, Pollatsek et al.’s (1992) original phonological preview benefit study that reported the largest effect size did not report the predictability of the target words or include a list of the sentence frames. Only two studies on typical English-speaking college populations reported the cloze predictability of their target words (<.25,.39, and.65 in Chace et al., 2005; Leinenger, 2019, Experiment 3; and Leinenger, 2019, Experiment 4, respectively). These studies reported phonological preview benefit effects in gaze duration of 9, 7, and 8 ms, respectively.

While the similarity in effect size between Leinenger’s (2019) Experiment 3 and Experiment 4, despite the difference in cloze probabilities, might suggest that constraint does not have a large impact, those two studies also differed in terms of whether the previews were real words (Experiment 3) or pseudowords (Experiment 4), and this confounding difference limits any strong conclusion about the effect of sentence constraint. The lexicality of the previews in previous studies could also potentially impact the effect size of the phonological preview benefit. In Vasilev et al.’s (2019) meta-analysis of phonological preview benefit effects there was variability across studies with respect to whether the previews were real-word homophones or pseudoword homophones and it is unclear whether word familiarity (i.e., lexicality) has a sizable impact on the magnitude of the effect. It has been argued that any evidence for phonological preview benefit from pseudoword previews would provide the strongest test of the early generation of phonological codes because there is no lexical representation of the stimulus from which phonological codes can feed down once lexical access has been achieved (Leinenger, 2019). However, reading behaviors such as fixation durations are strongly influenced by word familiarity (e.g., word frequency and lexicality; see Rayner, 2009). Furthermore, the frequency of the preview word relative to the target word is an important variable that determines whether the preview effect is positive (i.e., a standard preview benefit in which the unrelated preview leads to longer fixation durations than the identical preview) or negative (i.e., a reversed preview benefit in which the unrelated preview leads to shorter fixation durations than the identical preview; Milligan et al., 2022, Schotter and Fennell, 2019, Schotter and Leinenger, 2016, Schotter et al., 2018, Schotter et al., 2019). This suggests that words that are more familiar lead to earlier activation of linguistic information and allow readers to get further into processing based on parafoveal information (Schotter, 2018; Schotter & Leinenger, 2016). Therefore, the extent of phonological processing prior to fixating the target may have qualitatively different effects when a word is registered as familiar parafoveally (which is unlikely for pseudowords), versus when its semantics are fully recognized, versus when only low-level features are extracted.

There is also likely variability in the participants’ lexical quality both within and between the experiments in the Vasilev et al.’s (2019) meta-analysis. The lexical quality hypothesis proposes that a primary difference between skilled and unskilled readers lies in the quality of their lexical representations (i.e., the degree to which they consist of unified orthographic, phonological, and semantic codes (Perfetti & Hart, 2002, p. 190). Although lexical quality has been argued to drive efficiency in a wide range of reading processes (Andrews, Veldre, & Clarke, 2020), including parafoveal preview (Veldre & Andrews, 2015, 2016, 2017), little is known about how lexical quality relates to phonological preview benefit. Only one of the studies included in Vasilev et al.’s (2019) meta-analysis examined the effects of lexical quality (measured as general reading ability) on the phonological preview benefit (Chace et al., 2005). That study found that the effect was present only for ‘good readers’ which lends evidence to the possibility that lexical quality influences the use of phonological codes in silent reading. However, in a non-boundary eye tracking study (i.e., in which readers directly fixated on incorrect words) that investigated reader skill in the same way as Chace et al. (2005), the authors found that poor readers showed larger effects of phonology (i.e., shorter reading times on homophones than orthographic controls; Jared, Levy, & Rayner, 1999).

Some of the inconsistencies between Chace et al. (2005) and Jared et al. (1999) could be related to the differences in task (i.e., directly fixating on semantically incorrect words or not), but they could also be related to the fact that these earlier studies investigating individual differences effects on phonological activation assessed reading comprehension ability (i.e., the Nelson-Denny reading test; Brown, Fishco, & Hanna, 1993) rather than lexical quality per se. Reading comprehension has since been argued to be a less precise measure of lexical quality and a worse predictor of individual differences in reading compared to spelling ability (Andrews et al., 2020). Readers with high lexical quality, particularly reflected in spelling ability, show overall better reading abilities (Andrews et al., 2020), rely less on sentence context to recognize words (Andrews & Bond, 2009; Hersch & Andrews, 2011), and obtain more information parafoveally (Slattery & Yates, 2017; Veldre & Andrews, 2015). Therefore, the inconsistencies between Chace et al. (2005) and Jared et al. (1999) may be partially due to the fact that they were indirectly tapping into effects of lexical quality via the reading comprehension measures. However, this measure itself may be inconsistent at reflecting the latent individual differences variable (i.e., lexical quality) that is most relevant for the phonological contribution to early semantic activation. For example, individuals with higher spelling ability specifically may be more capable of rapidly activating phonological information from the visual orthographic information that could feed forward to semantics in a redundant fashion.

It is also unclear whether the effects of lexical quality (i.e., spelling ability) interact with lexical or context-level variables when it comes to phonological processing. In two of the three experiments that investigated reader skill in Jared et al. (1999), the sentences were constructed to be low constraint and the mean target word cloze probability was <3 % whereas in Chace et al. (2005) the sentences were low to moderate constraint (i.e., <25 % cloze probability). Subsequent work by Jared and O’Donnell (2017) that focused only on good readers reported an effect of phonology, which was larger when the semantically correct word was very high frequency. This study used low constraint sentences (i.e., 9.1 % close for the very high frequency target words and 3.6 % for low frequency target words1), which could mean that, in general, phonology is more likely to be activated when the semantically correct word is more expected, either because of its frequency of occurrence in the language or because of the preceding sentence context. Because the other phonological preview benefit studies did not control for or investigate reading or spelling ability, it is possible that the average phonological preview benefit effects that have been reported are obscuring a more complicated picture in which better readers have larger effects and worse readers have smaller or null effects.

The goal of this study was to provide a more complete picture of the nuances that impact phonological parafoveal preview benefit in silent reading by taking into account sentence context, preview lexicality, and individual differences in lexical quality. We conducted two high-powered eye tracking experiments using the boundary paradigm (Rayner, 1975; see Fig. 1) in which a homophone or orthographic control parafoveal preview changed to a semantically correct word when the eyes crossed an invisible boundary.

Because lexicality is informative about the stage of lexical processing at which phonological preview benefit occurs, we conducted separate experiments for word previews (Experiment 1; see examples 1a-1b) and pseudoword previews (Experiment 2; see examples 2a-2b). We were careful to match the target words in the two experiments on average word length and frequency, which are known to have strong influences on reading behavior (Rayner, 1998, 2009), in order to rule out these as potential confounds when inferring the effects of preview lexicality. For every target word we created two sentence contexts so that in one version the target word was highly predictable (1a & 2a) and in the other version it was unpredictable (1b & 2b), and these manipulations were closely matched across experiments, as well.

(1a) The boy bought his crush a single red (rows/rods) rose for Valentine's Day.

(1b) The thoughtful man bought his wife a beautiful (rows/rods) rose for her birthday.

(2a) The king and queen live in the large (cassul, casmol) castle in the countryside.

(2b) They went on a trip to see the ancient (cassul, casmol) castle that is in ruins.

We also collected assessments of the participants’ language abilities to investigate individual differences in language skill within a sample for each experiment. Additionally, we used these measures to ensure that the samples for the two experiments were comparable on language skill in order to facilitate comparisons of effects based on lexicality between the two experiments. We assessed spelling ability, phonological decoding ability, and semantic knowledge (i.e., vocabulary and reading comprehension).

We focused on three dependent eye tracking measures that are associated with the earliest stages of word recognition (i.e., skipping rates, single fixation duration, and gaze duration). First we describe our hypotheses that apply to both experiments, and we discuss our hypotheses for the differences based on lexicality in the introduction to Experiment 2.

We expected to observe a phonological preview benefit – higher skipping rates and shorter single fixation durations and gaze durations in the (pseudo)homophone conditions compared to the control conditions – in both studies. Although Vasilev et al.’s (2019) meta-analysis suggests we may not find such an effect if it is indeed quite small, we designed our experiments to be adequately powered so we expected the preview benefit effects to be significant.

With respect to our manipulation of sentence constraint, we hypothesized the phonological preview benefit effects to be larger in the high constraint conditions than in low constraint conditions. This two-way interaction between sentence constraint and preview condition would suggest that sentence context generates predictions about phonological forms, which would lead to a benefit for the preview that matches the preactivated phonological representation (i.e., the homophone). However, if expectations generated by context are limited to only semantics, or feed down only to orthography, we would not expect differences in the phonological preview benefit based on sentence constraint and therefore there should be no two-way interaction between sentence constraint and preview condition.

With respect to individual differences in lexical quality, we hypothesized that individuals with high lexical quality would show larger phonological preview benefits, particularly in high constraint sentences, which would manifest as a three-way interaction between lexical quality, sentence constraint, and preview condition. This would suggest that high lexical quality generates stronger predictions about the lower-level features of a word and readers with high lexical quality would be able to use the preactivation of these features to more efficiently extract phonological form from the bottom-up input in the parafovea. However, we were unsure which measure of lexical quality – spelling ability or phonological decoding ability – would be the stronger driver of this interaction. If the three-way interaction involves spelling ability, it would suggest that the phonological preview benefit results from more efficient processing of the orthographic forms, which would in turn allow earlier activation of phonological forms. Alternatively, if the significant three-way interaction involves phonological decoding ability, it would suggest that the phonological preview benefit results from better connections between orthographic and phonological representations in memory, which would speed efficiency only once the orthographic form has been recognized.