Native users of a language often use formulaic language (i.e., recurrent sequences of words) in spoken and written communication. This includes the use of collocations, which are juxtapositions of two or more words that are used together frequently, i.e., phrases, such as “black coffee” or “a quick shower” (Hill, 2000). These phrases are considered to be distinct from other word conjunctions that include compound words, such as “football” and “sunflower,” or hyphenated compounds, such as “machine-made,” where the conjunction of words produces a new or distinctive meaning. Crucially, the use of formulaic expressions, such as collocations, is believed to be a hallmark of linguistic proficiency and therefore, potentially important for the development of linguistic competence by L2 language learners (Cowie, 1998; Wray, 2000; Granger and Bestgen, 2014). This raises the question of whether L2 language learners use knowledge of formulaic language to benefit their production and comprehension. Bearing this issue in mind, with the present experiment, we used measures of eye movements to investigate the L2 processing of collocations during reading comprehension.

The experiment was motivated by the growing evidence that knowledge of collocations can facilitate processes of word recognition in reading. This includes evidence from research that has used variants of the lexical decision task to reveal a recognition advantage for collocations over non-collocative phrases (e.g., Ellis et al., 2009; Durrant and Doherty, 2010; Wolter and Yamashita, 2014). In one such experiment by Durrant and Doherty, participants viewed a series of two-word displays in which a prime word was swiftly followed by a target that was either a real word or a pseudoword. The participant’s task was to indicate whether the target was a real word as quickly and accurately as possible by pressing one of two response keys. The principal finding was that participants were quicker to make lexical decisions when the prime and target word combined to form a collocation (e.g., parish church) rather than a novel phrase (feature church). Such findings, therefore, suggest that readers can use their knowledge of collocations to speed up the process of word recognition.

Other researchers have used eye movement measures to determine whether a similar recognition advantage is observed in normal reading. This approach is based on the assumption that reader’s eye movements are sensitive to how easily linguistic information can be processed so that their gaze dwells for longer on words that are more difficult to recognize (see Rayner, 1998, 2009). There is substantial evidence for this close yoking between eye movements and the process of visual word recognition, including from studies showing that less time is spent looking at words that have a higher frequency of usage in text and so likely to be more familiar to readers (e.g., Inhoff and Rayner, 1986; Rayner et al., 1996). Collocations by their very nature are encountered more often than other phrases and so might be expected to produce similar frequency effects during reading. Consistent with this, an experiment by Vilkaite (2016) showed that collocations, such as “provide information,” are read more quickly compared to non-collocations, such as “compare information.” Similarly, an experiment by Siyanova-Chanturia et al. (2011) showed that binomial phrases, which are a conjunction of words in a specified order (e.g., “bride and groom”), are read more quickly when presented in this order compared to when the word order is reversed (e.g., groom and bride; see Yu et al., 2016, for similar effects for Chinese idioms).

Other research has additionally shown that readers’ eye movements are sensitive to the frequency of collocation usage. Information about a collocation’s frequency of usage can be computed in several ways, i.e., in terms of phrasal frequency, which provide a count of how frequently combinations of words appear together as a phrase (Gries and Ellis, 2015), or in terms of mutual information (i.e., MI, Hunston, 2002), which is a conditionalized count (i.e., a ratio) of how often these words appear together in a phrase rather than separately. Some researchers have also used transitional probabilities, which assess how likely it is that one word will follow another in text, to measure word co-occurrences (McDonald and Shillcock, 2003a,b). An experiment by Sonbul (2015) examined participants’ eye movements when reading sentences containing synonymous phrases that were classified as strong collocations (e.g., fatal mistake), weaker collocations (e.g., awful mistake), or non-collocative phrases (e.g., extreme mistake) based on their relative phrasal frequency and MI. The findings showed that the strong collocations were read fast by L1 readers, that the weaker collocations were read more slowly, and that non-collocations were read most slowly. Such findings suggest that readers’ eye movements are highly sensitive to a collocation’s frequency of written usage.

With the present experiment, we used eye movement measures to investigate whether L2 readers are also sensitive to collocation usage. As might be expected, L2 readers typically have less knowledge of collocations (e.g., Siyanova and Schmitt, 2008) and make more errors using them in both comprehension and production (e.g., Farghal and Obiedat, 1995; Weinert, 1995; Wray, 2002; Nesselhauf, 2003). However, there is evidence to suggest that L2 readers nevertheless develop sensitivity to collocation usage. This includes evidence from phrase judgment tasks showing that L2 readers are faster to recognize collocations when compared to novel phrases with a similar meaning (Wolter and Gyllstad, 2013; see also Ellis et al., 2008). Moreover, eye movement studies with L2 readers show that collocations, such as “provide information,” are read more quickly than similar non-collocative phrases (Vilkaitë and Schmitt, 2017). Similarly, while Sonbul (2015) found that L1 participants could categorize collocations and non-collocations more accurately than L2 readers in an offline rating task, these two groups of readers showed similar sensitivity to the frequency of collocation usage in a subsequent eye movement experiment. The indication, therefore, is that L2 readers can rapidly acquire knowledge about phrasal usage and use this to read more efficiently. We investigated this issue further in the present research using methods from a recent study by Li et al. (2021).

In this recent experiment, L1 participants read either strong collocations, such as “black coffee,” or weaker collocations, such as “bitter coffee,” in sentence contexts where the phrases were either highly predictable or unpredictable from the previous sentence context. This allowed Li et al. (2021) to investigate the relative contribution of frequency of usage and contextual predictability to the processing of collocations. The key finding was that strong collocations were read faster than weak collocations, consistent with Sonbul (2015) and providing further evidence that readers are sensitive to collocation frequency. The findings also showed that phrases were read more quickly in predictable when compared to neutral sentence contexts, consistent with substantial other evidence showing that readers use contextual knowledge to facilitate word recognition (for a review, see Staub, 2015). Crucially, however, this influence of prior sentence context was similar for the strong and weak collocations, suggesting that knowledge about collocation frequency can be used independently of context. Given these findings, our aim was to determine if L2 readers show similar sensitivity to collocation frequency and contextual predictability.

Accurately answering the comprehension questions that followed sentence displays was generally high (M = 90%) and was not significantly different across experimental conditions (ps > 0.1), indicating that the L2 readers could understand the sentences well. Before analyzing the data, short fixations (< 80 ms) were pooled with adjacent fixations, and then fixations less than 80 ms or more than 1,000 ms were removed following a standard procedure. In addition, fixations greater than 2.5 SD from each participant’s condition mean were deleted as outliers (accounting for < 3% of data).

The remaining data were analyzed using linear mixed-effects models (LMEMs, Baayen et al., 2008) using the R statistical programming environment (R Core Team, 2019) and the lme4 statistical package (version 1.1-26; Bates et al., 2012). For all analyses, participants and stimuli were treated as crossed random effects, and context and collocation strength were treated as fixed factors. The “contr.sdif” function in the MASS package was used to implement contrasts that compared the different levels of the fixed factors (Venables and Ripley, 2002). A full random structure model was used when possible (Barr et al., 2013). However, if the full model did not converge, its random structure was trimmed until it did converge successfully (starting with removing correlations between factors, then interactions). Significant interactions were investigated further using the “emmeans” package (Lenth et al., 2018). For all analyses, effects, where t/z > 1.96, were considered to be statistically significant (see, e.g., Baayen, 2008).

Eye movement measures are reported for the collocation. These included first-pass measures for the collocation (i.e., its initial processing before a saccade was made to the right of the collocation or a regression was made to its left), which were informative about factors affecting its initial processing. We examined first-pass reading time (FPRT, summed first-pass fixations within a region) and regressions-out (RO, probability of a regressive eye movement from a region) as measures of first-pass processing, and regression-path duration (RPD, summed fixations starting from the first fixation within a region, and following a regression, prior to a fixation to its right; see Liversedge et al., 1998), total reading time (TRT, summed fixations within a region), and regressions-in (RI, probability of a regressive eye movement back into a region), as measures that are informative about the later processing of the collocation. Note that skipping rates (i.e., the probability of not fixating the collocation during first-pass reading) were very low, even for collocation nouns. Specifically, for the collocation nouns, in predictable contexts we observed skipping rates of 8% (SD = 1%) for strong collocations and 6% (SD = 1%) for weak collocations, and for neutral contexts, we observed noun skipping rates of 4% (SD = 1%) for strong collocations and 5% (SD = 1%) for weak collocations.

The means of the eye movement measures are reported in Table 1 and the corresponding statistical results are reported in Table 2. Main effects of context predictability were obtained in all measures, with the exception of RO. These were due to faster reading times and fewer regressions back to collocations in predictable than neutral contexts. In addition, the main effects of collocation strength were obtained for all measures with the exception of RO. These were due to faster reading times and fewer regressions back for the strong relative to the weak collocations. Finally, a significant interaction between context and collocation strength was observed in TRTs (as shown in Figure 2). This reflected a larger predictability effect (i.e., shorter reading times in predictable compared to neutral contexts) for weak relative to strong collocations. Ideally, we would compare the interaction’s effect size against the minimal detectable effect size we report in the Method. However, calculating effect size (and using this to derive statistical power) is notoriously non-trivial for complex LMEMs with more than one fixed effect and two random effects (see, e.g., Westfall et al., 2014; Brysbaert and Stevens, 2018; Kumle et al., 2021). This was therefore not possible to compute for the interaction effect, although we note that, having computed the minimal detectable effect size, our design is relatively well-powered to detect small- to medium-sized effects.

The present experiment was conducted to shed light on the use of knowledge about collocation frequency and contextual predictability in L2 reading. The Chinese L2 English readers who took part had relatively long reading times and high rates of regression (collocation reading times averaged 680 ms or more, and collocation regression rates averaged 25% or more), consistent with their status as intermediate L2 readers. Despite this, the eye movement behavior of the L2 readers showed sensitivity to both the frequency of usage of collocations and their predictability from the prior sentence context.

We investigated the effects of collocation frequency by examining eye movements for collocations that were categorized as either strong or weak, depending on their frequency of written usage (as defined using phrasal frequency and MI). The results showed that less time was spent gazing at strong as compared to weak collocations; while, there were larger regression effects for the weak collocations. These effects were obtained in both measures of first-pass and later processing. The presence of effects in measures sensitive to the first-pass processing of the collocation suggests that collocation strength influenced how readily these phrases could be recognized by the L2 readers, with faster recognition for the stronger, i.e., higher frequency and collocations. This aspect of our findings is in line with other evidence from L1 and L2 readers showing that collocation frequency can influence the first-pass processing of a phrase (e.g., Sonbul, 2015; Li et al., 2021). Such findings are important as they imply that frequency of exposure to a specific collocation influences how information about this phrase is encoded in the mental lexicon and used subsequently to facilitate recognition of these expressions. The finding that the processing advantage for strong over weak collocations was also observed in later measures of processing (i.e., regressions back to the collocation) suggests that this processing advantage persists until later during processing and may affect how readily the phrases can be integrated as part of the sentence meaning. This contrasts with the findings from Sonbul (2015), who observed the effects of collocation frequency only in first-pass reading. However, similar effects were reported for L1 readers by Li et al. (2021) using the same stimuli, and so this late effect is unlikely to be unique to L2 readers. We note, however, that we did not test this formally, by directly comparing L2 data from the current experiment with L1 data from this previous research, as these two groups may differ in other respects.

We additionally investigated the effects of contextual predictability on phrase processing by placing the collocations in either a predictive or neutral sentence context. The results showed that readers spent less time at the collocations when they were predicted by the context and with increased regression rates when the collocations were unpredicted. This pattern of findings is consistent with a wealth of other research showing that readers make use of their knowledge of the prior linguistic context to guide the processing of new linguistic information (for a review, see Staub, 2015). Moreover, our findings are in line with the findings from other studies of L2 readers showing that eye movements are affected by both the frequency of usage and contextual predictability of words (e.g., Whitford and Titone, 2017; Mor and Prior, 2021). The predictability effects we observed were emerged in measures of first-pass processing and were also observed in later measures of processing. It, therefore, appears that L2 readers can make rapid use of their knowledge of the prior context to guide the recognition of collocative phrases.

Crucially, the present experimental design also allowed us to assess the conjoint effects of collocation strength and contextual predictability on the performance of the L2 readers. These variables were shown to interactively influence the processing of the collocations. This interaction effect was emerged in relatively late eye movement measure, in TRTs for the collocation. This effect (see Figure 2) was due to larger predictability effects (i.e., longer reading times in neutral vs. predictive contexts) for the weak when compared to strong collocations. As this effect was emerged relatively late in processing, it seems likely that it reflects the integration of the collocation with the sentence context rather than an influence on word recognition processes. The pattern of effects suggests that, in the absence of a constraining context, L2 readers experienced greater difficulty when attempting to integrate the weaker collocations with the sentence. This interaction between predictability and collocation strength stands in contrast with evidence showing that these two factors independently influence L1 readers’ eye movements (Li et al., 2021). Again, while this would be interesting to test formally, by comparing the performance of the two groups, we chose not to do this because of likely uncontrolled group differences. We nevertheless consider that these findings might be suggestive of L2 readers who had difficulty in integrating words with the context in the absence of strong contextual or lexical cues.

Taken together, the present findings add to the growing evidence that knowledge about collocations can influence the eye movements of L1 and L2 readers, and therefore that knowledge about formulaic language has an important influence on reading. Other evidence for such effects comes from studies of idioms (e.g., kick the bucket, Underwood et al., 2004; Conklin and Schmitt, 2008), binomial phrases (e.g., salt and pepper, Siyanova-Chanturia et al., 2011) and collocations (e.g., Sonbul, 2015; Vilkaite, 2016; Vilkaitë and Schmitt, 2017; Li et al., 2021). Among these studies, Sonbul also showed that L1 and L2 readers are sensitive to the frequency of usage of collocations, providing evidence that eye movement behavior is sensitive to the co-occurrence of words as a phrase. Such findings are important to our understanding of what information is accumulated through reading experience. It also challenges a key assumption made by current models of reading (e.g., the E-Z Reader model; Reichle et al., 1998, 2003) which assumes that the language processor employs information relating to the frequency of individual words but not phrases. Research with formulaic language, and collocations, in particular, reveal that L1 and L2 readers are highly sensitive to the frequency with which words occur together in phrases. The present findings additionally show that the integration of phrases with context is a product of the reader’s knowledge of collocations, such that better-known collocations (which will have a higher frequency of usage), are more readily integrated in the absence of contextual constraint. Moreover, the fact that such effects are detectable in eye movements reveals that such knowledge is rapidly brought to bear on the reading process by L1 and L2 readers. Further work is nevertheless needed to understand mechanisms underlying the acquisition and usage of this knowledge and the extent to which differences are observed as a function of reading skills. This will include understanding how similarity (and dissimilarity) between L1 and L2 collocations might affect processing, including for example whether an equivalence between L1 and L2 collocations might contribute to frequency effects.

A stimulus list, eye movement datafiles and analysis scripts are published on Figshare with the following DOI: 10.25392/leicester.data.17693798. Files are available from to publication from the following link: https://figshare.com/s/80b83e946d1854f46aee.

The studies involving human participants were reviewed and approved by Zhejiang University. The patients/participants provided their written informed consent to participate in this study.

HL, XW, and KP designed the experiment. HL collected the data. HL and KW analyzed the data. HL and KP wrote the manuscript. XW gave critical comments. All authors contributed to the article and approved the submitted version.

The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.