INCIDENTAL VOCABULARY LEARNING AND L2 VIEWING

The potential benefits of audiovisual input for vocabulary uptake have received increasing attention in recent years. This shift from a focus on written input is timely, as recent surveys have revealed that people now tend to spend more time watching television than reading books. In a recent survey on media consumption by Roy Morgan Research (2015), for example, data collected from 11 countries across the Asia-Pacific region showed that people spend an average of 8.2 hours to 29.5 hours a week watching television. In her survey that investigated Flemish EFL learners’ exposure to English language media, Peters (Reference Peters2018) found that more than 40% of these learners reported regular watching of English TV with or without subtitles. In addition, these learners, aged 16 and 19, reported seeking only limited exposure to written input such as books and magazines. This echoes Lindgren and Muñoz (Reference Lindgren and Muñoz2013), who found that young EFL learners relied more on subtitled movies than books for foreign-language exposure.

The potential of authentic audiovisual materials, such as movies and TV shows, for vocabulary acquisition has been pointed out through lexical analyses of large samples of such materials (Webb & Rodgers, Reference Webb and Rodgers2009a). Watching materials that are thematically related may lead to a good chance of encountering the same mid- and low-frequency words multiple times (Rodgers & Webb, Reference Rodgers and Webb2011). Clearly, there are merits in watching TV for vocabulary learning, which has prompted some researchers to advocate extensive TV viewing to bolster incidental L2 vocabulary learning (Webb, Reference Webb, Nunan and Richards2015), analogous to earlier proposals for extensive reading programs.

If authentic audiovisual input, including TV shows, is a good resource for vocabulary acquisition, then the next question is what can be done to make optimal use of this resource. A substantial body of research has gauged the benefits of captions, or on-screen text in the same language as the audio, in comparison with uncaptioned input. Two main benefits of captioned video have been reported. Firstly, most studies have shown better content comprehension for caption groups compared to no-caption groups (e.g., Gass et al., Reference Gass, Winke, Isbell and Ahn2019; Montero Perez et al., Reference Montero Perez, Van Den Noortgate and Desmet2013; see Winke et al., Reference Winke, Gass and Sydorenko2013 for a meta-analysis). Secondly, they have also almost consistently furnished evidence in favor of captioned viewing for vocabulary learning (e.g., Cintrón-Valentin et al., Reference Cintrón-Valentín, García-Amaya and Ellis2019; Markham, Reference Markham1999; Neuman & Koskinen, Reference Neuman and Koskinen1992; Winke et al., Reference Winke, Gass and Sydorenko2010; for a meta-analysis, see Montero Perez et al., Reference Montero Perez, Van Den Noortgate and Desmet2013). One explanation for the positive effect of captions on vocabulary acquisition is that the orthographic representation of the words helps learners to recognize word boundaries, which is more challenging when only aural input is available (Bird & Williams, Reference Bird and Williams2002; Winke et al., Reference Winke, Gass and Sydorenko2010). Surprisingly, whether the same benefit for vocabulary acquisition extends to the learning of MWEs has hardly been investigated. Thanks to a recent study by Puimège and Peters (Reference Puimège and Peters2019), we know incidental uptake of MWEs from uncaptioned audiovisual material is possible. The question, then, is whether adding captions harnesses its potential. For example, MWEs often contain function words (e.g., articles, prepositions) and these tend to be phonologically reduced in natural speech. Seeing these words written in the captions that accompany speech may help learners notice them in MWEs, distinguish them from the content words within MWEs, and thus aid uptake of the precise lexical composition (or form) of MWEs. The chances of learners attending to MWEs and their components in captions may be further improved through modifications to the captions that make MWEs stand out. This possibility is what we turn to in the following section.

MAKING MULTIWORD EXPRESSIONS STAND OUT

Incidental learning is usually defined as learning that takes place as a by-product of activities where learners focus primarily on content (Schmitt, Reference Schmitt2010). However, because it is hard to determine to what extent learners spontaneously turn their attention to language items or patterns as study objects during content-oriented activities, Hulstijn (Reference Hulstijn and Robinson2001) suggests that a critical operational feature that distinguishes intentional learning from incidental learning conditions is learners’ expectation of a language test. When input is processed for content, the precise wording of that content does not usually attract much attention. Yet, attention is vital for learning (Schmidt, Reference Schmidt and Robinson2001). To direct learners’ attention to certain language features or elements, researchers have proposed to use typographic enhancement, such as the use of underlining and bolding, to make these features or elements stand out (e.g., Sharwood-Smith, Reference Sharwood-Smith1993). A few studies have investigated whether this kind of modification of reading texts improves the chances of MWEs being noticed and remembered, and the results have been positive (Bishop, Reference Bishop and Schmitt2004; Boers et al., Reference Boers, Demecheleer, He, Deconinck, Stengers and Eyckmans2017; Choi, Reference Choi2018; Sonbul & Schmitt, Reference Sonbul and Schmitt2013; Toomer & Elgort, Reference Toomer and Elgort2019). In addition, a study by Szudarski and Carter (Reference Szudarski and Carter2016), which we will return to in the following text, showed that a combination of typographic enhancement and repeated encounters with the items led to better collocational knowledge compared to repetition alone.

Given that typographic enhancement promotes MWE learning in the context of written texts, it is conceivable that it is beneficial also in the case of captioned videos. However, there has been limited research on this, and the available research (see below) has examined uptake of single words rather than MWEs. Another modification to captions that has attracted some interest from researchers is to limit the captions to key words instead of including all the words present in the audio-recording. Keyword captioning is another means to make selected words stand out, because they are visually foregrounded relative to what is not included in the captions. Montero Perez et al. (Reference Montero Perez, Peters and Desmet2015, Reference Montero Perez, Peters and Desmet2018) included a comparison of vocabulary uptake from full captions and from keyword captions, and found an advantage for the latter, albeit mostly in posttests asking participants if they recognized the lexical items as ones they had encountered in the videos. Posttests about the meaning of the items yielded much poorer results overall, which is not surprising because captions (in L2) represent lexical forms, not the meanings of these forms. The target lexical items in Montero Perez et al. included both single words and MWEs, but no separate analyses were reported for the two categories. Teng (Reference Teng2019) also included a comparison of full captions and keyword captions, and his target items were all MWEs (verb–noun collocations). Rather surprisingly, the full captions were found to generate better learning gains than keyword captions.

To the best of our knowledge, there have only been two studies thus far that have examined the effect of typographically enhanced segments of full captions. In Montero Perez et al. (Reference Montero Perez, Peters, Clarebout and Desmet2014), L2 learners of French watched a video under one of four conditions: without captions, with regular captions, with keyword captions, and with regular captions in which the key items were highlighted. All three caption groups outperformed the no-caption group in a test that asked them if they recognized the items. They also did slightly better on a multiple-choice meaning-recognition test, but in this case only the keyword caption and enhanced caption groups outperformed the no-caption group at a statistically significant level. There were no significant differences between the test scores of the three caption groups. Montero Perez et al. (Reference Montero Perez, Peters, Clarebout and Desmet2014) also included tests about the content of the videos. Scores on these tests revealed no significant differences among the participant groups. It is worth mentioning that the content-related questions did not concern text portions featuring the target vocabulary. The second study that examined the effect of typographically enhanced captions is Cintrón-Valentin et al. (Reference Cintrón-Valentín, García-Amaya and Ellis2019). Animated video clips were integrated into an L2 Spanish course in three versions: uncaptioned, captioned with target vocabulary highlighted, and captioned with target grammar patterns highlighted. The captions with typographically enhanced vocabulary items generated significantly better vocabulary uptake than the other two viewing conditions, and even the caption condition where grammar patterns instead of vocabulary items were enhanced led to better vocabulary uptake than the condition without any captions. This, again, demonstrates that captions are to some extent beneficial for vocabulary uptake even without special steps to direct viewers’ attention to lexical items.

The studies reviewed in the preceding text furnish no evidence of a trade-off effect whereby typographic enhancement benefits uptake of what is enhanced but detracts attention from other elements. However, evidence of such an effect did emerge from a study by Choi (Reference Choi2018) on MWE uptake from reading. In this study, the students who read a text in which target collocations were enhanced recalled more of these target collocations compared to their peers who read the unenhanced version. However, the latter outperformed the former by 48% in the recall of content words that had been left unenhanced in both text versions. This suggests that typographically enhanced portions of a text can attract attention at the expense of the other portions of the text. This is reminiscent of the findings of some studies on typographic enhancement to assist grammar learning, which found that the use of enhancement promoted the uptake of grammar patterns, but impaired recollection of text content (e.g., Lee, Reference Lee2007). It seems plausible that this trade-off may also occur with captioned video, perhaps, especially because this type of input requires a greater distribution of attentional resources. In short, the role of typographic enhancement in fostering MWE uptake as well as content uptake from audiovisual input warrants further investigation.

THE ROLE OF REPETITION

Repetition, or frequency of occurrence, has been established to be beneficial for the learning of unknown words (see Uchihara et al., Reference Uchihara, Webb and Yanagisawa2019, for a meta-analysis). Provided there are enough encounters with the same words, learning can occur (Horst et al., Reference Horst, Cobb and Meara1998; Rott, Reference Rott1999; Webb, Reference Webb2007). While there is a substantial number of reading studies on the effects of repetition on the acquisition of single words, research investigating the effects of repetition on incidental MWE acquisition is far more limited.

An example of such a study is Pellicer-Sánchez (Reference Pellicer-Sánchez2017), who found that encountering the same collocations in a text eight times led to better posttest scores than encountering these collocations “only” four times, although the difference fell short of statistical significance. It is worth mentioning that the collocations in Pellicer-Sánchez (Reference Pellicer-Sánchez2017) were made up of adjective–pseudonoun combinations. The positive role of repetition for the learning of real collocations was shown in a study by Durrant and Schmitt (Reference Durrant and Schmitt2010). Unlike Pellicer-Sánchez’s study, repetition in Durrant and Schmitt’s study was not operationalized as repeated encounters of target items within a text. Instead, the participants were exposed to adjective–noun pairs under one of the following three conditions: (a) single exposure to the target collocations embedded in a sentence, (b) verbatim repetition, that is repeated exposure to the same sentences containing the target collocations, or (c) varied repetition, that is repeated exposure to different sentences containing the target collocations. Both repetition conditions yielded superior levels of recall compared to the single exposure. Further, verbatim repetition was found to lead to higher gains than varied repetition. In essence, Durrant and Schmitt’s study provides insight into how repetition or frequency of occurrence could be operationalized when using authentic unmodified materials, and suggests that exposure to target items in the same context rather than different contexts is useful for MWE learning (at least at an early stage of learning MWEs). This informs our design feature, in that repetition is operationalized as exposure to target MWEs twice, through repeated viewing of the same input video. This approach was also taken by Winke et al. (Reference Winke, Gass and Sydorenko2010), who found that watching a video twice with captions led to better vocabulary learning than watching it twice without captions.

Evidence of the role of repetition has been found in the context of bimodal input as well. In Webb et al. (Reference Webb, Newton and Chang2013), participants read short stories while listening to an audio-recording under four treatment conditions that varied in the number of encounters (in various contexts) with the target collocations. Depending on the text version, the collocations occurred up to 15 times over the course of approximately 30 minutes of reading. As expected, repetition led to more collocational knowledge.

The role of repetition in engendering MWE knowledge has thus far only been investigated in the context of written and bimodal input. To our knowledge, no study to date has investigated the role of repeated viewing in MWE learning. At least three viewing studies, however, have found a positive relationship between frequency of occurrence and the learning of single words. As part of a longitudinal study, Rodgers (Reference Rodgers2013) compared the effects of captioned and uncaptioned viewing on Japanese EFL learners’ incidental vocabulary learning. The learners encountered new words repeatedly through watching multiple episodes of a TV program. A positive correlation emerged between the number of encounters and learning, but, surprisingly, there was no significant difference between the relative vocabulary gains for the caption and no-caption conditions. A study that did find an effect for captions as well as frequency was conducted by Peters et al. (Reference Peters, Heynen and Puimège2016), but in this study participants watched videos with either captions or subtitles (L1 captions). The frequency of occurrence of the words and learners’ prior vocabulary knowledge were found to be positively associated with the learning gains. A study by Peters and Webb (Reference Peters and Webb2018) also found positive associations between incidental vocabulary learning and both the number of encounters with the target words and the learners’ prior vocabulary knowledge. The likelihood that more proficient learners and learners with comparatively good vocabulary knowledge pick up new words from audiovisual materials faster than less proficient learners has been demonstrated in other recent studies (e.g., Montero Perez, Reference Montero Perez2020; Pujadas & Muñoz, Reference Pujadas and Muñoz2019).

The studies reviewed here suggest that repeated viewing will improve the chances of incidental MWE acquisition from audiovisual input. Repeated viewing has also been reported to benefit text comprehension, as listening a second time enhances content comprehension (Lund, Reference Lund1991; Nguyễn, Reference Nguyễn2017; Sakai, Reference Sakai2009). This study seeks further evidence that repeated viewing benefits both MWE learning and comprehension.

PARTICIPANTS

One-hundred and twenty-two Malaysian L2 English learners took part in this study. The participants were college students (i.e., pre-university) working toward their diploma in various disciplines such as hospitality management and information technology systems. They ranged in age between 17 and 23 (M = 18.7; SD = .09). The participants had either Malay (n = 10) or a Chinese dialect (n = 116) as their L1. Their mean vocabulary size test (VST) score (Nation & Beglar, Reference Nation and Beglar2007) was 66 out of 140 test items (SD = 21), and the range of scores indicated they had (receptive) knowledge of between 4,000 and 9,000 most frequent word families in English. The participants came from six intact classes, and these intact classes were randomly assigned to one of the six experimental conditions, differing in number of viewings (i.e., one or two viewings) and caption condition (i.e., no captions, normal captions, or enhanced captions). There was no significant difference between the six groups (see Table 1) in their scores on the VST (Kruskal-Wallis χ2 (5) = 4.57; p = .47).

TABLE 1. Number of participants and the mean VST score (out of 140) under each condition

Notes: 1 = one viewing; 2 = two viewings.

Abbreviation: VST: vocabulary size test.

The participants were also given the Vocabulary Levels Test (Schmitt et al., Reference Schmitt, Schmitt and Clapham2001). As shown in Table 2, all six conditions had mean scores that indicate a mastery of the 2,000-word level.

TABLE 2. The mean VLT score for each condition

Abbreviation: VLT: vocabulary size test.

MATERIALS

Target Items

Exposing learners to authentic material entails that they will encounter MWEs of diverse kinds. Previous studies on MWE learning focused on certain parts of speech, such as adjective + noun collocations or verb + noun collocations, but this necessitated modifying texts to include sufficient numbers of these particular MWE types. If the aim is to explore ways of improving the chances of acquisition from authentic audiovisual materials, then it seems more ecologically valid to keep those materials intact. Moreover, there are indications that the risk of interitem interference is increased when learners are presented with sets of MWEs that are syntactically similar (Boers et al., Reference Boers, Demecheleer, Coxhead and Webb2014). In any case, the phraseology of a language naturally comprises diverse kinds of MWEs and there seems to be little justification for prioritizing certain syntactic categories over others. It is, nonetheless, acknowledged that not all MWEs are “equal” in the way that they are processed (e.g., Columbus, Reference Columbus and Wood2010), or in the challenges they pose to L2 learners (e.g., Boers, Reference Boers and Webb2020). However, in the present study this interitem variability applied across all six conditions due to the between-participant design of the experiment where all groups were exposed to the same text.

The video was first screened for MWEs likely to be as yet unfamiliar to the students in this study. To ascertain that these expressions were relatively well established in (American) English, the Corpus of Contemporary American English (COCA; Davies, Reference Davies2008) was consulted. All the potential target MWEs² yielded at least 100 hits in this corpus and/or were found to have a mutual information (MI) score (i.e., a measure of collocational strength) of at least 3.0.

After compiling a list of potential target MWEs from the video, four teachers of the learners in question were consulted on whether they believed the learners had knowledge of these items. Guided by this information, 20 MWEs were chosen as target items to be included in the tests. Each MWE only occurred once in the video. Two items attracted more than 30% correct responses at the pretest stage. This could be because these items were highly guessable, owing to the first-letter cues. As such, these items were not included as the target items. Table 3 lists the 18 final target items and their frequencies and MI scores in COCA (at the time the study was designed).

TABLE 3. Target items with corpus frequency and MI score

Abbreviations: MI: mutual information; MWE: multiword expression; COCA: Corpus of Contemporary American English.

For the normal captions condition, all the target items appeared without any enhancements (Figure 1). In the enhanced captions condition, the target items appeared bolded and underlined (Figure 2).

FIGURE 1. An example of a still from the video (Fresh off the Boat) with normal captions.

FIGURE 2. An example of a still from the video (Fresh off the Boat) with enhanced captions.

PROCEDURE

The data were collected in four sessions. In the first session, the participants were briefed on the research (without disclosing its specific purpose) before they signed the consent form. In the second session, the participants took the form-recall pretest. This was immediately followed by the VLT. Two weeks later, they watched the video under their respective conditions. Prior to watching the video, the participants were informed that they would be asked some comprehension questions. The comprehension test was administered immediately after the students had finished watching the video (once or twice). Then followed the immediate MWE recall test. Two weeks later, the delayed recall posttest was administered.

SCORING AND ANALYSIS

The responses in the MWE recall tests were scored by two raters. A lenient scoring system was adopted (e.g., see Webb and Kagimoto, Reference Webb and Kagimoto2009, Reference Webb and Kagimoto2011), in which minor mistakes were marked as correct and received a full score (e.g., a word used in singular that should have been plural, or vice versa; wrongly spelled words; wrong part of speech). While the scoring procedure was relatively straightforward for a target item in which only one content word needed to be completed, it was not the case for target items in which two or three content words of the MWEs needed to be supplied. For these items, partial credit (i.e., 0.5) was awarded for two-gap responses in which one of the words supplied by the learner contained a minor mistake, while the other contained a major mistake. A major mistake is defined as spelling mistakes that constitute a new word, affect the pronounceability of the word or generate a form that does not resemble the target item.

For an item that required participants to fill in three gaps, a two-over-three rule was imposed. This means that partial credit was given for responses in which two out of three gaps were correct or contained a minor mistake. Put differently, partial scores were only awarded for majorly inaccurate responses with two and three gaps (see Supplementary Materials for more details on the scoring procedure along with examples). The items with two and three gaps constituted 10 out of the 18 items. Because the interrater reliability was high (.97, .95, and .97 for the pretest, immediate posttest, and delayed posttest, respectively), the average score between the two raters was used. Each rater independently awarded 0 or .5 or 1 to a response. As the average score between two raters was taken, this resulted in five possible scores: 0, 0.25, 0.5, 0.75, or 1. Table 4 shows the descriptive statistics for all the MWE tests. As the pretest data were nonnormally distributed with unequal variance, the data were first transformed using Tukey Ladder of Powers before running a one-way nonparametric ANOVA (Kruskal–Wallis test). The Kruskal–Wallis test revealed no significant differences between groups in their scores in the pretest (Kruskal–Wallis χ2 (5) = 4.58, p = .47).

TABLE 4. Descriptive statistics for the form recall tests

Note: The maximum score on all tests was 18.

Because the pretest and the immediate posttest were not exactly the same format, the analysis focused on the two posttests, with performance on the pretest at the item level included as one of the fixed effects in the statistical modeling. Two clmm analyses for the immediate and delayed data respectively were carried out in R (R Core Team, 2018). The clmm function in the ordinal package (Christensen, Reference Christensen2019) was used. Apart from pretest performance, the fixed effects in all analyses included caption condition, number of viewings, and VST score. VST score was included as a fixed effect because the literature has shown that learners with a larger vocabulary tend to understand reading and listening texts better than learners with a smaller vocabulary, and this gives them an advantage when it comes to acquiring new lexical items from those texts (Elgort & Warren, Reference Elgort and Warren2014; Noreillie et al., Reference Noreillie, Kestemont, Heylen, Desmet and Peters2018; Schmitt et al., Reference Schmitt, Jiang and Grabe2011; Stæhr, Reference Stæhr2009). What is not certain is whether this comparative advantage might be mitigated by input conditions such as the availability of captions. VST score was centered prior to analysis. There was no issue of multicollinearity between pretest score and VST score. The five possible pretest scores at the item level (i.e., 0, 0.25, 0.5, 0.75, and 1) were multiplied by 4 (i.e., producing 0, 1, 2, 3, and 4) before being fitted into the models. An increase in one unit of pretest score would then relate to going up one score. The model comparison procedure started with the most complex model that included all the fixed effects, as well as the two-way interactions and three-way interactions between the factors. Each term was then incrementally removed. The significance level of fixed-effect predictors for all the models was assessed through model comparison using likelihood ratio tests (i.e., comparing a full model and a reduced model). The comparison returned likelihood ratio statistics with a chi-square distribution. This is reported in the form of likelihood ratio test (LRT), i.e., (LRT χ2 (n1) = n2, p < n3), where n1 = degrees of freedom, n2 = likelihood ratio statistic and n3 = p-value. To correct for multiple testing, Bonferroni adjustment of the alpha level was applied. This means that the adjusted level of significance is p < .025 rather than p < .05. To locate the differences between the treatment groups (e.g., between the three caption conditions), multiple comparison was carried out using the emmeans function in the emmeans package (Lenth, Reference Lenth2018).

As for the comprehension test about the content of the video, because the outcome variable was binomial, the data were analyzed by means of a generalized linear mixed model with the glmer function of the lme4 package (Bates et al., Reference Bates, Maechler, Bolker and Walker2015). The fixed effects included caption condition, number of viewings, and VST score. Model comparisons were carried out in the same way as the clmm analyses.

IMMEDIATE MWE POSTTEST

The results of the clmm analysis revealed that number of viewings (χ2 (1) = 21.9, p < .001), caption condition, (χ2 (2) = 52.5, p < .001), VST score (χ2 (1) = 66.7, p < .001), and pretest score (χ2 (1) = 92.9, p < .001) significantly predicted participants’ score on the immediate posttest. No significant interactions between the factors were found. Table 5 shows the output of the best-fit model, including the odds ratios. The odds are defined as the probability of an event occurring divided by the probability of it not occurring (Field et al., Reference Field, Miles and Field2012). As shown in the table, the odds of obtaining a score that is one level higher (e.g., 0 to .25) in the immediate posttest for participants who watched the video twice were 2.69 times the odds of those who viewed the video once. As for the effect of caption condition, multiple comparisons using the emmeans function in the emmeans package (Lenth, Reference Lenth2018) with Bonferroni adjustment applied (i.e., p < .025) revealed that the enhanced captions conditions were more likely to get a score that is one level higher compared to the uncaptioned conditions (p < .0001). Additionally, the normal captions conditions were more likely to get a score that is one level higher in the immediate posttest compared to the uncaptioned conditions (p < .0001). The difference between the enhanced and normal captions, however, failed to reach significance (p = .04) As to the effect of VST score, the odds of obtaining a score that is one level higher became greater as participants’ VST score increased. Further, a one unit-increase in the pretest score (e.g., 0.25 to 0.5) also increased the odds of obtaining a score that is one level higher in the immediate posttest.

TABLE 5. Output of best-fit model predicting a higher score in the immediate posttest

Notes: *Intercept levels: caption condition = uncaptioned; number of viewings = once.

Abbreviations: OR: odds ratio; VST: vocabulary size test.

DELAYED MWE POSTTEST

The analysis revealed that VST score significantly predicted participants’ score on the delayed posttest (χ2 (1) = 43.3, p < .001). As shown in Table 6, the odds of obtaining a better score became higher as participants’ VST score increased. The pretest score (χ2 (1) = 47.2, p < .001) and caption condition (χ2 (2) = 13.0, p < .01) were also found to be significant. A significant interaction between pretest score and caption condition was also found (χ2 (2) = 13.04, p < .01). To illustrate this interaction, the predicted probabilities of receiving the five possible delayed posttest scores on an item for all three caption conditions and all pretest scores were generated. The plot in Figure 3 shows the predicted probabilities for participants with the mean VST score (i.e., 69.69 out of 140). As there were far fewer occurrences of participants receiving a delayed posttest score of 0.25 and 0.75 than the other three scores, they were omitted from the plot. As can be seen from the plot, for participants with a pretest score of 0, the probability of receiving the same score in the delayed posttest was the lowest for the enhanced captions condition, followed by the normal captions and the uncaptioned conditions. Turning to the likelihood of receiving a full score in the delayed posttest, for participants who received 0 in the pretest, the enhanced captions had the highest likelihood of receiving a full score in the delayed posttest, followed by the normal captions and the uncaptioned condition. However, as the pretest scores became higher (e.g., 0.75 and 1), the predicted probability of getting a full score was higher for the normal captions and uncaptioned conditions, compared to the enhanced captions condition. This suggests that the effects of typographic enhancement could be stronger for participants with little or no knowledge of the target items. Put differently, participants with relatively good knowledge of the target items may not have needed typographic enhancement to pick up unknown MWEs.

TABLE 6. Output of best-fit model predicting a higher score in the delayed posttest

Notes: Intercept levels: caption condition = uncaptioned; number of viewings = once.

Abbreviations: OR: odds ratio; VLT: vocabulary size test.

FIGURE 3. Predicted probability of receiving a 0, 0.5, and 1 in the delayed posttest.

CONTENT COMPREHENSION TEST

Table 7 shows the descriptive statistics for the comprehension test. The results of the glmer analysis showed that the number of viewings had a significant influence (χ2 (1) = 8.87, p < .01). That is, the odds of obtaining a correct answer on the content comprehension test were higher for those who viewed the video twice compared to those who viewed it once. Further, VST score was also a significant predictor (χ2 (1) = 25.9, p < .001), suggesting that a higher VST score led to a higher likelihood of getting an item correct. There was also a significant main effect of caption condition (χ2 (2) = 8.64, p = .01). Participants in the normal captions condition were significantly more likely to get an item correct compared to their peers in the uncaptioned condition, as the odds of a correct response in the former were 1.77 the odds in the latter (Table 8). The difference between the enhanced and normal captions conditions did not reach significance (z = 1.25, p < .20). No interactions were found significant.

TABLE 7. Descriptive statistics for the comprehension test

Note: The maximum possible score was 22.

TABLE 8. Output of best-fit model predicting a correct response in the comprehension test

* Intercept levels: caption condition = uncaptioned; number of viewings = once.

Abbreviations: OR: odds ratio; VST: vocabulary size test.