The effects of glosses on English L2 incidental vocabulary learning through reading: a meta-analysis

1 Introduction

Vocabulary learning continues to be a central focus within the field of second language (L2) learning. Although explicit instruction often results in rapid improvements in learners' ability to recall lexical meanings in testing contexts, incidental vocabulary learning (IVL), broadly defined as the learning of lexical items that occurs when vocabulary learning is not the learners' explicitly assigned primary goal during meaning-focused activities such as reading or listening (Laufer and Hulstijn, 2001), is widely acknowledged as a fundamental mechanism for developing sustainable and functional vocabulary knowledge. In the present study, incidental vocabulary learning is operationalized in gloss-supported reading contexts where reading comprehension remains the primary assigned task goal, even though glossing may direct learners' attention to unfamiliar lexical items. We therefore acknowledge that some glossing treatments may involve a semi-incidental degree of lexical attention and are treated here as operationally incidental in a task-goal-based sense rather than as strictly unintentional learning. A growing body of empirical research has underscored the significance of IVL in second language learning (Webb et al., 2023). Nonetheless, studies have also revealed that this process tends to be relatively inefficient, yielding modest learning gains and often proving inadequate for the mastery of more sophisticated lexical knowledge (Schmitt, 2010). One common way of supporting incidental vocabulary learning is through glossing, which can make unfamiliar lexical items more noticeable while preserving the meaning-focused nature of reading. Glossing has therefore been widely examined as a form of support in reading-based vocabulary learning research. Among various enhancement techniques, lexical glossing (brief glosses accompanying target words) has been shown to effectively direct learners' attention to unfamiliar vocabulary while simultaneously supporting comprehension (Öztürk and Yorganci, 2018; Lofgren, 2022; Fakher Ajabshir et al., 2025; Jung, 2025). Within the domain of incidental vocabulary learning, numerous studies have explored diverse glossing techniques designed to enhance learners' comprehension and retention of lexical items. These glosses vary in their presentation format (e.g., in-text, marginal, or bottom-of-page; e.g., O'Donnell, 2012), linguistic type (e.g., L1, L2, or bilingual glosses; e.g., Kim et al., 2024), and modality (textual, pictorial, or multimodal combinations incorporating text, imagery, audio, or video; e.g., Rassaei, 2018; Ramezanali et al., 2021; Çakmak and Erçetin, 2018; Mirzaei et al., 2023; Shaojie et al., 2022; Zhang and Zou, 2022; Hoggan, 2024). Additionally, researchers have examined interactional and temporal dimensions, including hyperlinked vs. static glosses (e.g., Chen and Yen, 2013), as well as design-based contrasts, involving single vs. multiple gloss options and varying glossing frequencies (e.g., high- vs. low-density glossing conditions; e.g., Taylor, 2014; Zhang and Ma, 2024).

Overall, there is a general consensus within second language learning research that glossing exerts a beneficial impact on incidental vocabulary learning, serving as a form of input enhancement that draws learners' attention to target lexical items and facilitates comprehension (Smith, 1993; Mohsen et al., 2024). Nevertheless, scholarly debate persists concerning the precise scope and nature of these benefits. Points of contention include how different glossing contrasts may be associated with different learning advantages, and the extent to which they promote deeper lexical knowledge (e.g., Wang, 2022; Alcaraz Mármol, 2021). Some previous studies (e.g., Li et al., 2022) have discussed modality-related glossing effects with reference to theoretical perspectives such as Dual Coding Theory (Paivio, 1990) and Cognitive Load Theory (Sweller, 2005). In the present study, however, these perspectives are treated only as supplementary explanatory perspectives rather than as the primary interpretive framework, which is provided by ILH+. These ongoing debates underscore the need for systematic synthesis to elucidate the conditions under which glosses most effectively facilitate L2 incidental vocabulary learning. The Involvement Load Hypothesis (Laufer and Hulstijn, 2001) was initially formulated to explain the factors that modulate the effectiveness of incidental vocabulary learning. According to this framework, the depth of lexical processing and learning outcomes is determined by the degree of task-induced involvement, which is operationalized through three components: need, referring to the motivational requirement to learn a word; search, referring to the attempt to find its meaning or form; and evaluation, referring to the comparison or use of the word in relation to other lexical options or contextual meanings. Each component can vary in both presence and intensity. Although the hypothesis has provided valuable explanatory insights, it has been criticized for its limited predictive precision regarding vocabulary learning outcomes (Yanagisawa and Webb, 2021a). To refine the model, (Yanagisawa and Webb 2021b) proposed the Involvement Load Hypothesis Plus (ILH+), which integrates additional predictive variables such as input modality, Lexical Unit Type, test length, and input frequency. Drawing on this extended framework, the present study adopts a meta-analytic methodology to examine how moderator variables associated with glossing are related to incidental vocabulary learning outcomes across studies. In the present study, the Involvement Load Hypothesis Plus (ILH+) serves as the primary interpretive framework for organizing the moderator analyses and discussion. The present meta-analysis does not directly test the core components of ILH+; rather, it uses the framework to interpret variation across studies in a more systematic way. Other theoretical perspectives, including the Input Enhancement Hypothesis, the Noticing Hypothesis, Dual Coding Theory, and Cognitive Load Theory, are referred to only in a limited supplementary way where they help contextualize particular findings, and are not treated as parallel explanatory frameworks for the meta-analysis.

Against this background, most previous research syntheses on glossing have focused on relatively specific dimensions of gloss design, such as different forms of glossing in reading, textual glosses, gloss language, or multimodal glossing, rather than offering an integrated comparison of multiple glossing contrasts within a common reading-based analytic frame. For example, (Yanagisawa et al. 2020) examined how different forms of glossing contribute to L2 vocabulary learning from reading, (Zhang and Ma 2024) synthesized the effects of textual glosses on L2 vocabulary learning from reading, (Kim et al. 2024) focused on the relative effects of L1 and L2 glosses, and (Ramezanali et al. 2021) examined multimodal glossing. These studies have made important contributions, but each addresses only part of the broader glossing landscape. The present meta-analysis differs from previous syntheses in two main respects. First, it focuses specifically on English L2 incidental vocabulary learning in reading contexts and compares several recurrent operational glossing contrasts identified in the primary literature within a single quantitative synthesis. Second, beyond estimating the overall effect of glossing, the study examines a broader set of potential moderators relevant to variation in glossing effectiveness, including input frequency, treatment duration, test length, vocabulary knowledge type, lexical unit type, educational level, and input modality within reading-based environments.

Another moderator relevant to variation in glossing effects is the type of vocabulary outcome assessed. Vocabulary knowledge is commonly conceptualized as comprising two interrelated dimensions: breadth and depth. Because glossing studies differ in the outcome dimensions they assess, effect sizes may vary not only by gloss design but also by the aspect of vocabulary knowledge measured. Breadth typically refers to learners' knowledge of form–meaning connections (Schmitt, 2014, p. 915), which can be further operationalized through subcomponents such as form recognition, form recall, meaning recognition, and meaning recall (Schmitt, 2010). Depth of vocabulary knowledge, by contrast, encompasses learners' understanding of the semantic, morphological, and collocational relationships among words, including derivational and collocational knowledge. While acknowledging this multidimensional framework, the present meta-analysis focuses exclusively on breadth-related outcomes, as comparable data on depth measures are limited. Despite the growing interest in the role of glosses in incidental vocabulary learning, few studies have systematically examined how different dimensions of vocabulary knowledge interact with glossing effects. Existing evidence suggests that even within a single glossing condition, learning gains may differ across vocabulary dimensions. For example, (Hu et al. 2014) reported that L1 glosses yielded stronger retention on delayed posttests, whereas (Zhu 2023) found contrasting results, highlighting the need to interpret glossing effects with reference to the specific aspect of vocabulary knowledge assessed.

Input frequency is commonly defined as the number of times a target lexical item reoccurs within the input. Empirical research has consistently shown a positive relationship between exposure frequency and the efficacy of incidental vocabulary learning (Yanagisawa and Webb, 2021b). Nevertheless, in studies employing glossed input, frequency manipulations have typically been limited to between one and eight encounters. The precise mechanisms through which frequency interacts with glossing effects, however, remain insufficiently understood and warrant further empirical investigation.

Test length refers to the number of target lexical items included in a vocabulary assessment. As noted by (Yanagisawa and Webb 2021b), when fewer target words are presented, learners are able to devote more attentional resources to each item, thereby increasing the likelihood of successful incidental learning. In the context of glossed input, similar reasoning has been proposed: the facilitative effects of glosses tend to emerge more clearly when the lexical load is relatively small. Nevertheless, given the considerable variation in the number of target words across existing studies, it remains essential to investigate the moderating influence of test length in order to empirically substantiate this assumption.

Beyond the aforementioned factors, lexical unit type under investigation may also influence the outcomes of incidental vocabulary learning. Studies on glossed input have examined both single-word items and multiword expressions such as collocations and idioms. Owing to their greater formal and semantic complexity, multiword units are generally more demanding to process and acquire than isolated lexical items (Schmitt, 2010), thereby elevating overall task difficulty. Nevertheless, the extent to which lexical complexity interacts with glossing to affect learning outcomes remains insufficiently understood. In addition, treatment duration and learners' educational level were included because they may plausibly condition glossing effects. Longer treatments may provide greater cumulative exposure and more opportunities for consolidation (Ebersbach et al., 2022), whereas educational level may be associated with differences in reading strategies, metalinguistic awareness, and learners' ability to make effective use of glossing support.

Taken together, the existing literature suggests that glossing is generally facilitative, but several issues remain insufficiently resolved. Previous reviews and meta-analyses have tended to concentrate on particular dimensions of gloss design rather than comparing multiple recurrent glossing contrasts within a single synthesis. More specifically, earlier syntheses have more commonly examined gloss presence, gloss form, gloss language, or modality-related contrasts, whereas moderators such as vocabulary knowledge type, test length, treatment duration, educational level, and lexical unit type have been less consistently incorporated into gloss-focused meta-analytic models. As a result, existing syntheses provide valuable evidence for the overall usefulness of glosses, but still offer only a partial account of the conditions under which glossing effects vary across studies. This creates a clear need for a broader synthesis that compares multiple glossing contrasts while examining a wider range of moderators within a common reading-based framework.

As a quantitative research methodology, meta-analysis offers an objective framework for synthesizing empirical findings by transforming outcomes from individual studies into a standardized metric of effect size. This approach not only enables the quantitative integration of disparate datasets and the estimation of overall intervention effects but also facilitates moderator analyses that identify variables influencing these effects while controlling for potential confounds. Meta-analytic procedures are therefore particularly well-suited to evaluating the efficacy of glosses in incidental vocabulary learning and to delineating the contextual conditions under which their benefits are maximized. Although previous meta-analyses in the domain of incidental vocabulary learning (e.g., Yanagisawa and Webb, 2021a,b; Webb et al., 2023) have confirmed the generally facilitative role of glosses, the range of moderators examined has been relatively narrow, leaving several potentially influential factors insufficiently explored. Against this backdrop, the present study adopts a meta-analytic approach to address three research questions:

• 1) What is the overall effect of glossing on L2 learners' incidental vocabulary learning?
• 2) To what extent do different operationally defined glossing contrasts differentially facilitate L2 incidental vocabulary learning?
• 3) Which potential moderator variables influence the effectiveness of glosses in promoting L2 incidental vocabulary learning?

2 Methodology

2.1 Retrieving primary sources

The identification of studies for this meta-analysis was conducted through a multi-step search procedure designed to enhance transparency and replicability. First, electronic searches were carried out in MLA International Bibliography, Linguistics and Language Behavior Abstracts (LLBA), Education Resources Information Center (ERIC), Web of Science, and ProQuest for studies published between January 1993 and December 2024. Search strings were adapted slightly to the indexing conventions of each database, but the core search string was as follows: (“gloss^*” OR annotation^* OR “marginal gloss^*” OR “multimedia gloss^*” OR “electronic gloss^*” OR “hypertext gloss^*” OR “e-gloss^*”) AND (“incidental vocabulary learning” OR “incidental vocabulary acquisition” OR “vocabulary learning” OR “vocabulary acquisition” OR “vocabulary retention” OR “lexical learning” OR “lexical retention”) AND (reading OR “reading comprehension”) AND (“second language” OR “foreign language” OR L2 OR EFL OR ESL).

Second, a manual search was conducted in the following journals: Language Learning, Applied Linguistics, Language Teaching Research, The Modern Language Journal, Studies in Second Language Acquisition, TESOL Quarterly, and Reading in a Foreign Language. These journals were selected because they are major publication outlets in second language acquisition, vocabulary studies, reading, and technology-assisted language learning, in which studies on glossing, technology-assisted reading, and L2 vocabulary learning are commonly published (Bahari and Gholami, 2023). These journals constituted the full set of journals manually searched in this study.

Third, citation chasing was used to identify additional eligible studies. Backward citation searching involved screening the reference lists of six previous research syntheses on glossing: (Yanagisawa et al. 2020), (Ramezanali et al. 2021), (Kim et al. 2024), (Mohsen et al. 2024), (Zhang and Ma 2024), and (Yun 2011). In addition, one relevant book chapter, (Jung 2025), was screened during the study identification phase. Forward citation searching was conducted through Google Scholar to identify more recent studies citing these seven sources. Google Scholar was also used as a supplementary search tool to locate recently published studies that might not yet have been fully captured in the database searches. After removing duplicate records, a total of 1,032 studies were identified at the end of this stage.

2.2 Inclusion and exclusion criteria and study selection procedure

As illustrated in Figure 1, the study selection process was conducted in two stages: title-and-abstract screening followed by full-text eligibility assessment. Two authors of the present study independently screened the records retrieved from the database, manual, and citation-chasing searches. At the title-and-abstract stage, studies were excluded as clearly irrelevant when they were obviously unrelated to glossing, incidental vocabulary learning, reading-based tasks, English as a second or foreign language, or empirical intervention research. After duplicate removal, 1,032 records were identified. Following title-and-abstract screening, 980 records were excluded for not meeting the research focus, leaving 52 records for full-text eligibility assessment. Two authors independently conducted both the title-and-abstract screening and the full-text screening. Inter-rater reliability was calculated prior to consensus discussion using Bennett, Alpert, and Goldstein's S index. The resulting coefficients were S = 0.957 for the title-and-abstract screening stage and S = 0.846 for the full-text screening stage. Any disagreements at either stage were discussed and resolved through consensus. After all eligibility criteria had been applied, 26 studies were included in the meta-analysis.

The full texts were then examined against nine inclusion criteria. Specifically, eligible studies had to: (1) be published between 1993, following the introduction of the Input Enhancement Hypothesis, and December 2024; (2) examine English as a second or foreign language; (3) investigate reading-based learning contexts; (4) maintain reading comprehension as the primary task goal; (5) operationalize vocabulary learning as incidental in the task-goal-based sense adopted in the present study, such that reading comprehension remained the primary assigned task and deliberate vocabulary study was not explicitly required before, during, or after reading; (6) examine the effects of glosses or glossing-related contrasts on vocabulary learning outcomes; (7) employ an experimental, quasi-experimental, randomized controlled, or repeated-measures between-group design; (8) include a control or comparison group; and (9) report sufficient statistical information, including means, standard deviations, and sample sizes, to enable effect-size calculation.

Furthermore, studies were required to investigate reading contexts in which comprehension served as the primary task goal, with vocabulary learning occurring as a by-product of text processing rather than as an explicitly assigned instructional objective. We recognize, however, that the construct status of gloss-supported learning within incidental vocabulary learning research is not straightforward, because glosses may increase the salience of target lexical items and encourage a degree of lexical attention during meaning-focused reading. In the present review, “incidental” was therefore operationalized primarily in terms of task goal rather than the complete absence of lexical attention.

More specifically, studies were retained when participants were primarily instructed to understand the text and vocabulary learning was not presented as the explicit main purpose of the activity, even when glosses, highlighting, repetition, or other input-enhancement features directed attention to unfamiliar words. By contrast, studies were excluded if they included explicit instructions to study, memorize, rehearse, or deliberately learn target vocabulary before reading, during reading, or in post-reading instructional activities. Post-reading vocabulary tests or delayed posttests used to measure learning outcomes were not treated as evidence of intentional learning, provided that the instructional phase itself did not require deliberate vocabulary study.

Borderline cases were handled by applying the same task-goal criterion consistently across studies. If a study appeared to encourage partly intentional attention to lexical items because of gloss salience or repeated exposure, it was retained only when reading comprehension remained the primary assigned goal and deliberate vocabulary learning was not explicitly required. Such cases are therefore treated in the present review as operationally incidental, while acknowledging that some may be more appropriately characterized as semi-incidental rather than strictly unintentional. Because reporting practices varied across primary studies, the present review does not assume complete uniformity in participants' awareness or the total absence of lexical attention. When study descriptions were ambiguous, eligibility decisions were made conservatively on the basis of whether reading comprehension could still be reasonably identified as the primary assigned task and deliberate vocabulary study was not explicitly required. Accordingly, the term incidental vocabulary learning in this meta-analysis should be interpreted as an operational, task-goal-based category rather than as a claim that all included learning conditions were entirely free from intentional attention to vocabulary. After all eligibility criteria had been applied, 26 studies were included in the meta-analysis.

2.3 Quality assessment

Following the screening process, the methodological quality of the included studies was evaluated using the QualSyst appraisal tool (Kmet et al., 2004). This instrument comprises 14 criteria (see Table 1), each rated on a three-point scale: yes = 2, partly = 1, and no = 0. Criteria deemed not applicable to a particular study design were coded as NA and excluded from the final calculation. For inter-rater reliability, only item ratings that were applicable and independently scored by both reviewers were included in the calculation; item pairs involving NA were excluded. Higher total scores reflect stronger methodological quality. Based on the percentage of the total attainable score, studies were categorized as high quality (≥75%), medium quality (55–75%), or low quality (< 55%). Two independent researchers conducted the quality assessment. Inter-rater reliability for this stage was calculated prior to consensus discussion using Bennett, Alpert, and Goldstein's S index. The resulting coefficient was S = 0.93, indicating a high level of agreement between the two reviewers. Any discrepancies in scoring were discussed and resolved through consensus. Based on the QualSyst appraisal, 22 studies were rated as high quality, 4 as medium quality, and none as low quality. Because the distribution across quality categories was limited and uneven, study quality was not examined as a moderator in the main analyses. However, to assess the robustness of the overall findings, we conducted a sensitivity analysis excluding the four medium-quality studies. The restricted analysis yielded a substantively similar overall pattern (d = 0.945, 95% CI [0.777, 1.113], I² = 89.6%), suggesting that the main finding was not driven by the lower-rated studies in the evidence base.

Table 1. Assessment of study quality via the QualSyst tool.

2.4 Data coding

The studies meeting the inclusion criteria were analyzed using the coding framework presented in Table 2. This framework was initially developed with reference to prior meta-analyses on glossing (e.g., Yun, 2011; Yanagisawa et al., 2020). The coding was conducted independently by two authors of the present study. Inter-rater reliability was initially estimated using Cohen's kappa, which indicated a high level of coding agreement (κ = 0.92). Any discrepancies between the coders were discussed and resolved through mutual agreement.

Table 2. Coding scheme.

As presented in Table 2, the coding framework consisted of three categories of variables, representing different analytical levels. At the study level, variables captured general bibliographic information, including the author(s), year of publication, and publication type. At the treatment level, variables described characteristics of the instructional intervention. Specifically, the eight variables examined as potential moderators in the meta-analysis were input modality (paper-based and computer-based), lexical unit type (single words and phrases), treatment duration, input frequency, test length, vocabulary knowledge type (meaning recognition, meaning recall, form recognition, and form recall), educational level (university and school-level learners), and glossing subgroups. For coding purposes, the five glossing subgroups were operationally defined as follows: bilingual glosses, used here as a broad label for gloss-language contrasts, referred to glosses that differed in the language used to present lexical meaning (e.g., L1 or L2); positional glosses referred to glosses that differed in their spatial location relative to the target item (e.g., in-text, marginal, or bottom-of-page glosses); interactional glosses referred to glosses that differed in the degree or mode of learner interaction required to access them (e.g., hyperlinked/clickable vs. static glosses); modality-based glosses referred to glosses that differed in the modality of presentation (e.g., textual, pictorial, audio, video, or multimodal glosses); and experimental glosses referred to other design-based glossing contrasts that did not fall neatly into the preceding categories, such as single vs. multiple gloss options or differences in glossing frequency or density. These subgroups were not intended to constitute a single hierarchical taxonomy of glossing categories. Rather, they represent recurrent operational contrasts in the primary literature that were coded because they provided comparable between-condition glossing contrasts for quantitative synthesis. In other words, the basis of comparability in the present meta-analysis was not that all categories existed at the same conceptual level, but that each represented an empirically examined contrastive glossing condition in reading-based studies. For this reason, the subgroup analysis should be interpreted as a comparison of operationally defined glossing contrasts rather than as a test of a single unified taxonomy of glosses. Accordingly, differences across these subgroups are reported as descriptive subgroup patterns and should not be overinterpreted as evidence that the categories are conceptually equivalent. Audiovisual input was classified under the computer-based modality because it is typically delivered through digital media or computer-mediated interfaces. Finally, at the outcome level, variables captured statistical information required for effect size computation, including test type, sample size, mean, and standard deviation for both experimental and control groups.

2.5 Study characteristics

A total of 26 studies involving 3,232 participants satisfied the inclusion criteria. Sample sizes varied considerably across studies, ranging from 45 to 237 participants. The participants represented two educational levels: middle school learners (n = 3) and university students (n = 23). This distribution indicates that the evidence base for educational-level comparisons was notably imbalanced. Across the included studies, a variety of annotation-based interventions were examined, including experimental glossing (n = 7), positional glossing (n = 2), modality-based glossing (n = 7), interactional glossing (n = 2), and bilingual glossing (n = 8). Intervention duration ranged from less than 1 to 7 weeks, with most studies employing short-term treatments (n = 9). Regarding methodological design, 15 studies adopted randomized controlled trials (RCT), seven used quasi-experimental designs, and four implemented other forms of between-group experimental designs. Outcome measures were predominantly researcher-developed tests (RDT; n = 25), with only five studies utilizing externally developed vocabulary assessments. A comprehensive summary of study characteristics is presented in Table 3.

Table 3. Characteristics of the included studies.

2.6 Data analysis

In the present study, the meta-analysis was conducted using Comprehensive Meta-Analysis (CMA) software, version 2.2. The results showed that all Q statistics were significant and the I² values exceeded 75%, indicating a substantial level of heterogeneity across the included studies. Consequently, a random-effects model was applied to estimate the overall effects. The pooled effect size was therefore interpreted as an average effect across diverse study conditions rather than as a uniform estimate expected to apply equally across all contexts. Considering that the relationship between glossing and incidental vocabulary learning may be influenced by various moderator variables, further moderator analyses were performed. Specifically, categorical moderators were examined through subgroup analyses, whereas continuous moderators were first tested using meta-regression. To ensure robustness and interpretability, these continuous moderators were subsequently categorized and re-examined via subgroup analyses.

To evaluate potential publication bias, we employed multiple complementary diagnostics, including visual inspection of the funnel plot, Rosenthal's fail-safe N (Rosenthal, 1979), and the Duval and Tweedie trim-and-fill method. As illustrated in Figure 2, the funnel plot exhibited a certain degree of asymmetry in the distribution of effect sizes around the central line, which represents the fixed-effect weighted mean (d). The fail-safe N, calculated using Rosenthal's method based on 123 independent samples with computable standard errors, was 31,832, far exceeding the empirical threshold of 5k + 10 = 1,000. Under Rosenthal's criterion, this result suggests that a large number of unpublished studies would be required to nullify the overall significant effect, although this inference should also be interpreted cautiously in light of the heterogeneity in the dataset. Furthermore, the simplified Duval–Tweedie trim-and-fill analysis, applied to the same subsample, indicated no missing studies on the right side of the funnel plot but identified 36 missing studies on the left side. After adjustment, the estimated overall effect size changed only slightly from d = 0.503 to d = 0.508. Although this pattern does not suggest that publication bias alone fully explains the observed overall effect, the result should be interpreted cautiously. Publication bias diagnostics are informative rather than definitive, particularly when funnel plot asymmetry is present and the dataset is characterized by substantial heterogeneity and methodological diversity (see Zhang et al., 2021, pp. 13–15).

3 Results

In interpreting the empirical outcomes, this study aligns with the evaluative framework articulated by (Plonsky and Oswald 2014), adopting d values of 0.400, 0.700, and 1.000 as conventional benchmarks that demarcate small, medium, and large magnitudes of effect, respectively, in between-group designs situated within the domain of second language learning research.

3.1 The overall effect size of glosses on incidental vocabulary learning outcomes

As shown in Table 4, the pooled overall effect of glossing on incidental vocabulary learning was statistically significant and approached a large magnitude (d = 0.935). The heterogeneity test yielded a significant result (Q = 1,128.51, df = 122, p < 0.001, I² = 89.2% > 75%), indicating substantial heterogeneity among the included studies. This suggests that the variability in the observed effects may be attributable to the influence of moderator variables, thereby warranting subsequent moderator analyses to further explore potential sources of heterogeneity. Accordingly, the overall effect size should be interpreted as a broad mean tendency across highly diverse instructional conditions, rather than as evidence of a uniformly strong effect in all glossing contexts. The very high I² value also indicates that the pooled estimate summarizes substantial cross-study variability in study design, learner populations, glossing conditions, and outcome measures, and should not be interpreted as a single stable effect that would be expected to apply similarly across all contexts. A sensitivity analysis excluding the four medium-quality studies yielded a substantively similar overall effect (d = 0.945, 95% CI [0.777, 1.113], I² = 89.6%), indicating that the overall result was robust to the exclusion of the lower-rated studies.

Table 4. Analysis of overall effect size.

According to the data summarized in Table 5, the between-group comparison revealed significant differences in effect-size estimates across the five operationally defined glossing subgroups (Q_between = 38.93, p < 0.001). These subgroup estimates varied in magnitude. Specifically, positional glossing was associated with a small effect-size estimate (d = 0.375), and modality-based glossing was associated with an estimate approaching the medium range (d = 0.638). By contrast, the experimental (d = 1.354), interactional (d = 1.494), and bilingual subgroups were associated with relatively larger subgroup estimates. Because these subgroup categories represent different dimensions of gloss design rather than a single hierarchical taxonomy, the results should therefore be interpreted as comparative subgroup patterns rather than as evidence of a uniform ranking across conceptually equivalent gloss categories.

Table 5. Effect-size estimates across operationally defined glossing subgroups.

3.2 Moderator analysis

As shown in Table 6, the moderator analysis indicates that input frequency, test length, treatment duration, and vocabulary knowledge type had statistically significant effects (p < 0.001) on glossing-related learning outcomes. A marginally significant pattern was also observed for educational level; however, this result should be interpreted with considerable caution because the educational-level comparison was based on a markedly imbalanced evidence base, with only three school-level studies and twenty-three university-level studies. By contrast, input modality and lexical unit type did not exert statistically significant effects. Furthermore, the meta-regression results presented in Table 7 reveal that among the continuous variables, input frequency and treatment duration were positively correlated with the effectiveness of the intervention, whereas test length exhibited a negative correlation with intervention efficacy.

Table 6. Results of moderator analyses.

Table 7. Meta-regression results for continuous moderators.

As shown in Table 8, the subgroup analyses revealed variation in effect-size estimates across moderator categories. These subgroup estimates are reported primarily as descriptive patterns. Unless supported by a formal test of subgroup differences, they should not be interpreted as conclusive evidence of clear between-subgroup differences. Regarding input modality, the estimate for computer-based reading (d = 1.000) was numerically larger than that for paper-based reading (d = 0.858). However, because input modality did not emerge as a statistically significant moderator, this pattern should be interpreted cautiously as a descriptive difference rather than as evidence of a reliable between-subgroup advantage. For lexical unit type, the estimate for single-word learning (d = 0.956) was numerically larger than that for phrase learning (d = 0.651); however, this subgroup difference should be treated as descriptive rather than conclusive. With respect to vocabulary knowledge type, subgroup estimates varied across outcome categories. Form recall (d = 2.022) and meaning recognition (d = 1.039) showed relatively larger estimates, whereas meaning recall (d = 0.777) and form recognition (d = 0.800) showed comparatively smaller estimates. In terms of educational level, the estimate for university students (d = 0.969) was numerically larger than that for middle school students (d = 0.656), although this descriptive difference should not be taken as evidence of a robust educational-level advantage. For treatment duration, the 5–7 week subgroup (d = 1.566) showed a larger estimate than the shorter-duration subgroups ( ≤ 1 week, d = 1.004; 2–4 weeks, d = 0.791), but this pattern is reported descriptively rather than as definitive evidence of superiority. Regarding input frequency, the 3–8 times subgroup (d = 1.309) showed a larger estimate than the 1–2 times subgroup (d = 0.942). This difference is described here as a subgroup pattern rather than as a conclusive between-subgroup effect. Finally, for test length, the 1–20 item subgroup (d = 1.269) showed a larger estimate than the 21–40 item subgroup (d = 0.622) and the 41–60 item subgroup (d = 0.706). These differences should be interpreted as descriptive subgroup variation.

Table 8. Results of subgroup analyses for moderator variables.

4 Discussion

4.1 The overall effect of glossing on incidental vocabulary learning

The results indicate that glossing is associated with a substantial and statistically significant facilitative effect on incidental vocabulary learning. However, given the high degree of heterogeneity across the included studies, this overall effect is best interpreted as a broad average across diverse instructional conditions rather than as a uniformly robust effect expected in all contexts. This caution is especially important given the substantial diversity across the included studies in learner populations, glossing conditions, research designs, and outcome measures. This overall pattern is broadly consistent with attention-based perspectives such as the Input Enhancement Hypothesis and the Noticing Hypothesis, in the sense that glosses may make target lexical items more noticeable during meaning-focused reading. In the present study, however, these perspectives serve only as general background accounts, whereas ILH+ provides the primary interpretive framework for organizing moderator-related variation across studies.

4.2 The effects of different glossing contrasts on incidental vocabulary learning

The subgroup analysis revealed variation in effect-size estimates across the operational glossing contrasts included in the present meta-analysis. It should be emphasized, however, that these categories were not intended to constitute a single hierarchical taxonomy of glossing categories. Rather, they represent recurrent operational contrasts in gloss design found in the primary literature and were compared here as operationally defined subgroups because each captured an empirically examined glossing contrast in reading-based studies.

Within this analytic framework, interactional, experimental, and bilingual contrasts were associated with relatively larger subgroup estimates, whereas modality-based and positional contrasts were associated with relatively smaller estimates. These differences should be interpreted cautiously as comparative subgroup patterns rather than as evidence of a strict ranking across conceptually equivalent glossing categories. One possible interpretation is that some glossing contrasts may be more likely than others to provide different kinds of processing support or learner involvement during reading. However, because the present meta-analysis cannot directly verify the specific cognitive processes engaged in the primary studies, these interpretations remain theoretical and provisional. The findings are therefore best read as descriptive cross-study variation across operational glossing contrasts, interpreted with due caution regarding the uneven conceptual status of the subgroup categories.

4.3 The effect of moderating variables on the overall effect size

4.3.1 Input modality

Input modality did not emerge as a statistically significant moderator in the present meta-analysis. Although the computer-based subgroup showed a numerically larger estimate than the paper-based subgroup, this descriptive difference should not be interpreted as evidence of a reliable modality advantage. One plausible interpretation is that the broad distinction between paper-based and computer-based input may itself be too coarse to capture meaningful variation across glossing conditions. Computer-mediated environments can differ substantially in interactivity, multimedia richness, pacing, learner control, and the way glosses are accessed, and these differences are not well-represented by a simple paper-vs.-computer contrast. The present null result should therefore not be taken as evidence that modality is necessarily unimportant; rather, it may indicate that this moderator, as operationalized in the present meta-analysis, obscures potentially more informative distinctions within computer-mediated conditions. Accordingly, the current finding is best interpreted as a limitation of the available coding contrast rather than as evidence for a clear modality-based advantage in either direction.

4.3.2 Lexical unit type and vocabulary knowledge type

The subgroup results for lexical unit type and vocabulary knowledge type suggest that glossing outcomes may vary according to the type of target item and the type of vocabulary knowledge assessed. For lexical unit type, the estimate for single words was numerically larger than that for phrases. However, because lexical unit type did not emerge as a statistically significant moderator, this difference should be interpreted cautiously as a descriptive subgroup pattern rather than as evidence of a reliable between-subgroup advantage. One possible interpretation is that phrase-level targets may present greater processing demands than single-word items, although the present meta-analysis cannot directly verify this possibility.

Variation was also observed across vocabulary knowledge types, with form recall and meaning recognition showing relatively larger subgroup estimates than meaning recall and form recognition. These differences may reflect variation in the kinds of processing demands associated with different assessment formats. Such patterns can be viewed as broadly consistent with perspectives such as the Depth of Processing framework and Cognitive Load Theory (Craik and Lockhart, 1972; Sweller, 2011), but the present findings should not be taken as direct confirmation of those mechanisms. Overall, the results are best understood as descriptive outcome-related differences that warrant cautious interpretation.

4.3.3 Input frequency and treatment duration

The significant effects of input frequency and treatment duration suggest that glossing outcomes may vary depending on how frequently learners encounter target items and how long the treatment is sustained. In the present study, these findings are interpreted cautiously as moderator patterns rather than as direct confirmation of specific cognitive mechanisms. Within the ILH+ framework, one possible interpretation is that more frequent exposure and longer treatment duration may provide learners with greater opportunities to attend to, retrieve, and consolidate target lexical items over time (Teng, 2019; Kaivanpanah et al., 2022; Middleton et al., 2019). However, because the present meta-analysis cannot directly verify the actual cognitive processes engaged in the primary studies, this interpretation should remain provisional. Overall, the results are best understood as a descriptive pattern broadly consistent with the view that repeated and sustained encounters may support stronger vocabulary learning outcomes.

4.3.4 Test length

The meta-regression analysis revealed a significant negative association between effect size and test length, which is broadly consistent with the predictions of ILH+ (Yanagisawa and Webb, 2021b). One possible interpretation is that when a larger number of target items is included in assessment tasks, learners may have fewer opportunities to allocate attention and processing effort to each individual item. In this sense, longer tests may be associated with smaller learning gains at the item level.

This pattern can also be viewed as broadly consistent with perspectives such as Cognitive Load Theory and the Noticing Hypothesis (Sweller, 2011; Schmidt, 1990), in that larger target sets may make it more difficult for learners to process each item deeply during meaning-focused reading. However, the present meta-analysis does not directly test these underlying mechanisms. The result should therefore be interpreted as a moderator pattern consistent with reduced per-item processing opportunity, rather than as direct evidence of a specific cognitive process.

4.3.5 Educational level

The meta-regression analysis suggested a possible association between educational level and the effect of glossing. However, this pattern should be interpreted with considerable caution because the educational-level comparison was based on a markedly imbalanced evidence base, with most included studies involving university learners and only a small number involving school-level learners. Although the subgroup estimate for university learners was numerically larger than that for school-level learners, this difference should be treated as descriptive rather than as conclusive evidence of a robust educational-level moderator.

One possible interpretation is that differences in learner experience or study context may shape how glossing support is used during reading, although the present evidence base is far too uneven to support strong conclusions on this point. Overall, the present findings do not permit a firm conclusion about educational level as a moderator; rather, they suggest at most a tentative pattern that requires further evidence from a more balanced study base.

5 Conclusion

This meta-analysis systematically examined the effectiveness of glossing in second language (L2) incidental vocabulary learning and quantitatively integrated findings across varied input and learning conditions. Specifically, glossing was associated with a significant facilitative effect on learners' incidental vocabulary learning, although this pooled estimate should be interpreted as a broad average across heterogeneous study conditions rather than as a uniformly robust general effect expected to apply similarly across all glossing contexts. Variation was also observed across the operational glossing contrasts synthesized in the study, with interactional and experimental contrasts associated with relatively larger subgroup estimates and positional and modality-based contrasts associated with relatively smaller estimates. These subgroup patterns should be interpreted cautiously, however, because the glossing categories captured different dimensions of gloss design rather than a single hierarchical taxonomy. The study used the Involvement Load Hypothesis Plus (ILH+; Yanagisawa and Webb, 2021b) as its primary interpretive framework for organizing moderator-related variation across studies, without directly testing the framework's core mechanisms. Pedagogically, the findings should be treated as cautious, evidence-informed tendencies rather than instructional prescriptions. In some contexts, more interactive or functionally supportive glossing conditions, repeated and sustained exposure to target items, and relatively lower target load may be associated with stronger vocabulary learning outcomes. However, these possible implications must be interpreted in light of the substantial heterogeneity of the dataset, the conceptual unevenness of some glossing subgroups, and the limited evidence base for certain comparisons. Despite its comprehensive scope, the study acknowledges certain limitations: cognitive and affective factors, identified by (Yanagisawa and Webb 2021b) and (Li and Wang 2020) as influential in incidental vocabulary learning, were underrepresented in the primary data and thus not included as moderator variables. In addition, the glossing categories synthesized in this meta-analysis captured different dimensions of gloss design rather than a single hierarchical taxonomy, which should be borne in mind when interpreting subgroup comparisons. A further limitation concerns the construct status of gloss-supported learning within incidental vocabulary learning research. The findings should therefore be interpreted as applying to gloss-supported, meaning-focused reading conditions that were operationalized as incidental, rather than to uniformly unintentional vocabulary learning in the strictest possible sense. A further methodological limitation is that the original item-level coding records were no longer available for retrospective recalculation of alternative interrater reliability indices such as the S index. Future research should further examine how cognitive, affective, and methodological factors shape the relationship between glossing and vocabulary learning outcomes, thereby enriching theoretical understanding of the cognitive–affective dynamics underlying L2 lexical learning.

This review was prospectively registered in PROSPERO (International Prospective Register of Systematic Reviews) under the registration number CRD420251237727.

Generative AI statement

The author(s) declared that generative AI was not used in the creation of this manuscript.

Any alternative text (alt text) provided alongside figures in this article has been generated by Frontiers with the support of artificial intelligence and reasonable efforts have been made to ensure accuracy, including review by the authors wherever possible. If you identify any issues, please contact us.

Publisher's note

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.

Footnotes

Edited by: Irina Elgort, Victoria University of Wellington, New Zealand

Reviewed by: Hassan Banaruee, University of Education Weingarten, Germany

Yuya Arai, Chiba University, Japan