A cross-modal analysis of lexical sophistication: EFL and ESL learners in written and spoken production

2.1 Lexical sophistication

Lexical sophistication is operationalized as a metric that measures “the proportion of relatively unusual or advanced words in the learner’s text” (Read 2000: 203). Traditional indices assessing lexical sophistication include the ratio of high-grade words (e.g., those introduced at grade 9 or later) to the total word count (Linnarud 1986) and the proportion of low-frequency words relative to all words (Laufer and Nation 1995). With the advancements in computational programs and text analysis techniques, recent research has introduced finer-grained lexical sophistication indices, capturing the multidimensional aspects of L2 lexical knowledge. In this section, we discuss 14 widely adopted lexical sophistication indices, which fall into four dimensions (Kyle and Crossley 2015).

Word frequency and range. Word frequency has been recognized as a key indicator of L2 production abilities. It is widely accepted that high-frequency words are typically learned earlier and more easily than low-frequency words (Ellis 2002). The underlying assumption of this frequency effect is that repeated exposure to a word strengthens the association between its form and meaning, allowing for more rapid and efficient access and retrieval during production (Ellis 2006; Gollan et al. 2008). Previous studies have commonly assessed word frequency by either counting the number of words falling within specific frequency bands determined by their ranking in a reference corpus (e.g., Laufer and Nation 1995; Morris and Cobb 2004), or by tallying the frequency of a word’s appearance in a reference corpus (e.g., Kyle and Crossley 2015).

However, the word frequency index has a limited application since it often falls short of covering specific types of words that remain uncaptured in a broad term of frequency. For instance, some technical words may be highly frequent in a narrow set of contexts but appear less frequently in general language use (Kyle et al. 2018). To address this issue, the concept of word range (known variably as dispersion, entropy, and contextual diversity) has been introduced. Word range counts the number of documents in a reference corpus in which a target word occurs, with a lower range value indicating a word that appears in fewer contexts and is considered more specialized and sophisticated (Kyle et al. 2018). Previous research on L2 lexical sophistication has measured this index by counting the number of documents in which a word appears, both in spoken contexts (e.g., Crossley et al. 2013; Kyle and Crossley 2015) and written contexts (e.g., Kim et al. 2018).

Academic language. Sophisticated lexical items also include words and phrases commonly encountered in academic contexts (Kyle et al. 2018). Two notable academic word lists serve as key indices of lexical sophistication: the Academic Word Lists (AWL; Coxhead 2000) and the Academic Formulas List (AFL; Simpson–Vlach and Ellis 2010). The AWL includes academic words identified in written academic corpora, while the AFL contains lists of formulaic sequences documented in both written and spoken academic corpora. Previous studies computed frequency scores of academic language based on these lists, revealing significant associations with L2 speaking proficiency (Kyle and Crossley 2015) and writing proficiency (Kim et al. 2018).

Psycholinguistic word information. The psycholinguistic properties of words are associated with words’ learnability (Salsbury et al. 2011). Specific indices within this dimension include word concreteness, imageability, meaningfulness, familiarity, and age of acquisition (Coltheart 1981; Kuperman et al. 2012). Concreteness refers to the extent to which a word is concrete or abstract. Words evoking more tangible and perceptible mental representations (e.g., orange, doctor) are deemed more concrete and, consequently, considered less sophisticated compared to abstract words (e.g., impossible, truth). Imageability reflects the ease with which a mental image of a word can be created. Meaningfulness measures the degree to which a word is associated with other words. Familiarity assesses how commonly a word is experienced (e.g., paper vs. encephalon), with this index strongly correlated with word frequency (Kim et al. 2018). Age of acquisition refers to the age at which a particular word is learned.

Previous studies have utilized native speakers’ judgment ratings from the Medical Research Council Psycholinguistic Database (Coltheart 1981) to evaluate words’ concreteness, imageability, meaningfulness, and familiarity (e.g., Crossley and McNamara 2012; Kim et al. 2018). In addition, native speaker judgment scores reported by Kuperman et al. (2012) have been employed to determine the age of acquisition of words (Kim et al. 2018). These psycholinguistic word indices have been shown to account for variance in the quality of written texts (Crossley and McNamara 2012; Kim et al. 2018) and reflect the longitudinal development of L2 vocabulary in spoken languages (Crossley et al. 2016; Salsbury et al. 2011).

N-gram frequency, range, and association strength. N-grams, referring to sequences of multiword units (Hyland 2008), encompass bigrams (e.g., each other, thank you) and trigrams (e.g., out of the, a lot of). These n-grams serve as valuable indicators of phraseological complexity in learner production, revealing how learners combine different lexical items to create a variety of word combinations (Paquot 2019). Several studies have adopted n-gram features as constructs of lexical sophistication, which is a subset of phraseological sophistication (e.g., Kim et al. 2018; Kyle and Crossley 2015).

Indices related to n-grams include frequency, range, and association strength. Similar to the assessment of word frequency and range, previous research computed n-gram frequency and range by tallying the occurrences of n-grams and determining the number of documents in which they appear in a native-speaker reference corpus (Kim et al. 2018; Kyle and Crossley 2015). These indices have been identified as strong indicators of L2 proficiency in both written (e.g., Kim et al. 2018) and spoken production (e.g., Kyle and Crossley 2015).

N-gram association strength measures the likelihood of a sequence of words occurring together. A widely used index for this purpose is Delta P, a well-established measure for evaluating the association strength of word sequences. Delta P is calculated as the difference between the probability of word A occurring with word B and the probability of word A occurring with any word other than word B (Kyle et al. 2018). Given its consideration of directionality between words, Delta P effectively addresses issues related to word order effects. Texts exhibiting higher proficiency tended to include bigrams and trigrams with stronger associations (Kim et al. 2018; Kyle et al. 2018; Paquot 2019). Consequently, higher Delta P scores are interpreted as indicative of a higher degree of lexical sophistication.

2.2 Effect of modality on L2 production

It is well established that L2 learners show varying degrees of linguistic complexity in writing versus speaking (e.g., Biber et al. 2016; Ellis and Yuan 2005; Hwang et al. 2020; Kim and Hwang 2022; Kormos 2014; Vasylets et al. 2017). However, the majority of studies have focused on lexical diversity and/or syntactic complexity, with few exploring lexical sophistication. For example, Kormos (2014) examined various linguistic complexity measures in the written and spoken texts produced by secondary-school foreign language learners with upper-intermediate proficiency. Her findings indicated higher lexical accuracy, greater lexical variation, and more complexity in noun phrases in written output compared to spoken output. Similarly, Ellis and Yuan (2005) found that adult Chinese EFL learners demonstrated increased vocabulary accuracy in writing than in speaking.

The impact of production modality, as observed in previous research, can be captured in terms of different psycholinguistic mechanisms underlying the two production modes (Kormos 2014). In general, writing is considered a cognitively less demanding process than speaking. For instance, Vasylets et al. (2017: 397) characterized writing as “offline, permanent, and visible” while describing speaking as “online and evanescent”. According to their descriptions, writing involves a cyclical process in which a writer can move back and forth between message planning, linguistic encoding, and writing processes (Halliday 2002; Kellogg 1996). In contrast, speaking is a linear, time-constrained process that places significant processing demands on the speaker, particularly for L2 speakers who are more constrained than native speakers in the automatic formulation and articulation of speech (Kormos 2014).

The distinct nature of the writing and speaking processes suggests that modality can directly influence L2 learners’ vocabulary production. However, there is a gap in research specifically investigating the effect of modality on L2 lexical sophistication. Instead, studies have examined L2 lexical sophistication separately in the domains of written and spoken production. For example, Kim et al. (2018) analyzed college students’ written essays using various lexical sophistication indices, as mentioned in the previous section. Their results showed that several indices, including word frequency and range, age of acquisition, and n-gram properties (frequency, range, and association), significantly contributed to the variance in L2 writing proficiency. In another study, Eguchi and Kyle (2020) explored lexical sophistication in oral interviews of Japanese-speaking learners, demonstrating that indices such as word concreteness, word frequency, orthographic/phonological neighbors, word accessibility, and bigram association significantly explained the variance in the learners’ proficiency scores.

While each of these studies identified different lexical sophistication indices as significant components of L2 proficiency, it is impossible to make direct comparisons of these outcomes due to differences in data and analysis methods. This limitation underscores the need for further research that examines both written and spoken production using consistent methodologies. Moreover, it remains unclear whether the modality effect interacts with other factors such as L2 learning contexts. In what follows, we outline differences between EFL and ESL contexts and consider how these differences may interact with modality to influence L2 lexical sophistication.

2.3 Differences between EFL and ESL contexts

The defining characteristics of EFL and ESL contexts are best captured by Kachru’s (1985) landmark descriptions of the three major English use contexts: the inner circle (L1 varieties, e.g., Australia, the United States, United Kingdom), the outer circle (ESL varieties, e.g., India, Malaysia, Pakistan, Philippines, Singapore), and the expanding circle (EFL varieties, e.g., China, Greece, Indonesia, Japan, Korea). As members of the outer circle, speakers in the ESL context use English for daily communication (Tarnopolsky 2000), either as the official language or as an additional language of the government. As a result, ESL learners encounter a variety of linguistic and sociolinguistic features of English in diverse, realistic situations beyond classroom settings. In contrast, the use of English in the EFL context is confined to formal instructional settings, offering limited opportunities for students to use the language for communicative purposes in real-world situations (Tarnopolsky 2000).

These differences imply a significant divergence in both the quantity and the type of L2 input provided to ESL and EFL students. Given that ESL students are exposed to English in various social settings, they tend to have ample opportunities to receive a multiplicity of written and spoken linguistic input on a daily basis (Barrot and Gabinete 2021). Conversely, students in the EFL context encounter English predominantly within a classroom setting, with limited exposure to real-world communicative situations (Barrot and Gabinete 2021; Tarnopolsky 2000). Moreover, textbooks serve as the primary source of input in the EFL context, resulting in students mostly receiving written input, with few experiences of spoken language through in-class practices (Saito 2017).

Based on the distinctive features of these two English learning contexts, several studies have compared vocabulary usage between EFL and ESL students, with a primary focus on examining written production. The findings of these studies have been varied. For example, Barrot and Gabinete (2021) observed that ESL learners tend to use a greater number of words compared to their EFL counterparts. Similarly, in their analyses of thesis abstracts written by EFL learners, ESL learners, and English native speakers, Nasseri and Thompson (2021) found that ESL learners and native speakers exhibited a higher number of lexical words and a wider variety of vocabulary compared to EFL learners. In contrast, Zhang and Kang (2022) showed that Chinese EFL students displayed a higher degree of lexical diversity in written production compared to Hong Kong ESL learners.

Regarding lexical sophistication, to the best of our knowledge, only one study has conducted comparative analyses between EFL and ESL learners. Choemue and Bram (2021) examined scientific journal articles written by EFL authors from Indonesia and Thailand and ESL authors from the Philippines and Malaysia. Their assessment of lexical sophistication included calculations based on the number of advanced word tokens as defined by Laufer and Nation (1995), along with measuring lexical density and lexical diversity. Their results indicated no significant differences between the two learner groups across all three domains.

While these studies have contributed to our understanding of diverse aspects of lexical complexity influenced by learning contexts, several questions remain unanswered. Although Choemue and Bram (2021) found little difference between EFL and ESL learners, this finding may not necessarily apply to non-scientific contexts of production. Moreover, their focus on lexical sophistication primarily centered on word frequency, such as the use of advanced words. It thus remains unclear whether differences exist between EFL and ESL learners in other aspects of lexical sophistication. Crucially, most of the previous investigations have focused on written production, leaving open the question of how learning contexts interact with production modality in affecting L2 lexical sophistication.

Examining the influence of L2 learning contexts provides a valuable avenue for testing various theoretical models that establish a link between language experience and L2 vocabulary usage. One noteworthy model in this context is the usage-based approach to L2 acquisition (Bybee 2008; Ellis 2006). This approach emphasizes the role of a learner’s language usage experience as a driving force behind language acquisition. It posits that the memories of prior utterances within an individual’s language use profile, combined with frequency-based statistical learning mechanisms, serve as the primary sources of their linguistic knowledge (Ellis 2006). Of particular importance in the context of usage-based vocabulary learning is the role of word exposure frequency. According to this approach, words that are experienced frequently become firmly entrenched and automatized in a learner’s mental lexicon, enabling more efficient processing, storage, and retrieval of lexical information (Gollan et al. 2008; Verspoor and Schmitt 2013).

In line with this perspective, we hypothesize that L2 learners will exhibit varying degrees of lexical sophistication contingent upon their language experience. Specifically, within the usage-based approach, when a language learner is exposed to a lexical item with greater frequency, their knowledge of the word, including its form and meaning, becomes more deeply entrenched, facilitating the acquisition and retrieval of the lexical information (Ellis 2002, 2006; Ellis and Wulff 2015). From this standpoint, one might anticipate that ESL learners, who encounter English in natural settings, exhibit higher degrees of lexical sophistication compared to EFL learners, who predominantly learn English vocabulary through classroom instruction. Despite this prediction, however, empirical evidence supporting the impact of learning contexts on L2 lexical sophistication, particularly within the theoretical framework of the usage-based approach, remains scarce.

4.1 Corpus selection

The dataset utilized in this study comprised 400 written essays (200 from EFL learners and 200 from ESL learners) and 400 speech samples (200 from EFL learners and 200 from ESL learners) extracted from the International Corpus Network of Asian Learners of English (ICNALE; Ishikawa 2014). According to the corpus developer’s descriptions, the written and spoken samples were independently collected from college students in various regions of Asia, including six EFL areas, such as China, Indonesia, Japan, South Korea, Taiwan, and Thailand, and four ESL regions, such as Hong Kong, Pakistan, the Philippines, and Singapore.

Although different individuals participated in the written and spoken tasks, the written and spoken samples in ICNALE were closely matched in terms of the genre and topics (Ishikawa 2013, 2014). In highly controlled production tasks, students provided written essays of 200–300 words within a time limit of 40 min and spoken monologues within a time limit of 60 s. Each task presented students with two topics: part-time jobs for college students and banning smoking at restaurants. Each topic was accompanied by a question prompting students to provide their opinions (e.g., Do you agree or disagree with the following statements? Use reasons and specific details to support your opinion. “It is important for college students to have a part-time job”). In the writing task, each student produced one essay per topic, while in the speaking task, two speech monologues were generated per topic. To mitigate potential practice effects associated with repetition, we selected the first round of speech samples from each student for analysis. To maintain consistency in the topic, our data analysis focused on written and speech samples produced under the topic of part-time jobs for college students.

Furthermore, we controlled for participants’ English proficiency in our sample data. The ICNALE categorizes each student’s proficiency level in alignment with the Common European Framework of Reference (CEFR), classifying learners into four levels: A2 (Waystage), B1_1 (Threshold lower), B1_2 (Threshold upper), and B2 (Vantage or Higher). In our analysis, we specifically focused on data from the students in the B1_2 proficiency level, as this category covered a sufficient number of samples across both EFL and ESL groups. Table 1 provides a summary of information regarding the samples in each group.

4.2 Lexical sophistication indices

The present study measured the following 14 lexical sophistication indices using TAALES (Kyle et al. 2018) version 2.2. The list of indices is presented in Table 2.

Word frequency and range. The word frequency and range scores were computed using either the BNC written corpus or the BNC spoken corpus. It is noteworthy that content words and function words exhibit distinct features, thus requiring separate analyses (Durrant et al. 2019). However, in our analysis, we found that word frequency and range scores for both content and function words were significantly correlated with those of all words. Consequently, we reported outcomes based on all words (see Appendix B for separate analyses for content and function words). TAALES calculates the average score of each index in a text by dividing the sum of the frequency/range scores for all words by the total number of words in the text (Kim et al. 2018; Kyle and Crossley 2015). To meet the assumption of normality, we applied logarithm transformation to the frequency and range scores (e.g., Eguchi and Kyle 2020).

Academic language. Between the two measures of academic language lists provided by TAALES, namely the AWL (Coxhead 2000) and the AFL (Simpson−Vlach and Ellis 2010), we specifically referred to the AFL list, which included formulaic sequences used in both written and spoken contexts. In our analysis, we utilized the normed scores, calculated by dividing the number of tokens of academic sequences by the total number of word tokens (Kyle and Crossley 2015).

Psycholinguistic word information. TAALES provides scores for word concreteness, imageability, meaningfulness, and familiarity, which are based on native speakers’ judgment ratings derived from the Medical Research Council Psycholinguistic Database (Coltheart 1981). Additionally, the tool offers scores for the age of acquisition based on native speaker judgment, as reported by Kuperman et al. (2012). In our analysis, we normalized the scores from each psycholinguistic index using log transformation.

N-gram frequency, range, and association strength. Using TAALES, we computed the frequency and range scores of bigrams and trigrams with reference to two subcorpora in the Corpus of Contemporary American English (COCA; Davies 2009). Since TAALES employed either the written academic or spoken section as a reference corpus, we conducted separate analyses using each reference corpus for the written and spoken data. The n-gram frequency and range scores were log-transformed for normality of the data. Regarding n-gram association strength, we relied on the Delta P scores derived from the COCA.

4.3 Data analyses

Prior to data analysis, we conducted correlation analyses to examine collinearity among indices within each category. The outcomes of these correlation analyses can be found in Appendix C. In case where we identified indices showing a strong correlation (r ≥ 0.7), we opted to retain one of them (e.g., Kyle et al. 2018). To examine the modulating effects of modality and language learning contexts on each lexical sophistication index, we constructed linear mixed-effects models using the lme4 package (Bates et al. 2015) in R version 4.2.1 (R Core Team 2022). P values were calculated using the lmerTest package (Kuznetsova et al. 2017). Each model included two fixed effects of Modality (spoken, written) and Context (EFL, ESL), alongside a random effect of participant. The fixed factors were centered around the means and deviation coded by assigning –0.5 to spoken and EFL conditions and 0.5 to written and ESL conditions. When the model revealed a significant interaction between Modality and Context, we conducted post-hoc analyses using the emmeans function in R (Lenth 2018), investigating the effect of Modality in each context group. The results of these post-hoc tests are presented with the corrected alpha level implemented via Tukey adjustments. To determine the effect size for each comparison, we calculated Cohen’s d using the eff_size function in R.

5.1 Word frequency and range

Due to strong correlations between the word frequency and range scores obtained from the written and spoken reference corpora (see Appendix C), we chose to present only the results of word frequency derived from the written reference corpus as a representative measure. The model outcomes are presented in Table 3, with a visual representation provided in Figure 1. The model revealed a significant effect of Modality, indicating higher scores in speaking than in writing. Additionally, there was a significant effect of Context, with higher scores in the EFL than the ESL group. However, no interaction was observed between the two factors, suggesting that the frequency score difference in spoken and written production remained consistent between the EFL and ESL groups. These results suggest that while both groups produced more frequent words in speaking than in writing, the EFL group generally produced higher-frequency words compared to the ESL group.

Figure 1:

Results of word frequency (log-transformed raw scores).

5.2 Academic language

As shown in Table 4, the model demonstrated a main effect of Context, indicating that the ESL group produced more academic phrases than the EFL group. This outcome suggests an enhanced ability of the ESL group to produce formulaic sequences frequently appearing in academic contexts. While there was no main effect of Modality, an interaction was found between Modality and Context. Post-hoc tests using emmeans revealed that this interaction was driven by the two groups’ distinct production patterns depending on the production mode. Specifically, the EFL group produced a greater number of academic language in written than in spoken production (b = −0.013, SE = 0.004, t = −2.960, p = 0.017, Cohen’s d = −0.296). In contrast, the ESL group exhibited the opposite pattern, producing a greater number of academic language in spoken than in written production (b = 0.014, SE = 0.004, t = 3.241, p = 0.007, Cohen’s d = 0.324). As visible in Figure 2, the academic phrases produced by the ESL group in spoken production outnumbered not only those in writing but also those produced by the EFL group in spoken and written production. These findings highlight the strong impact of L2 learning experience on the production of formulaic sequences used in academic contexts.

Figure 2:

Result of academic language (normalized raw scores).

5.3 Psycholinguistic word information

We excluded two indices: (a) Imageability due to its strong correlation with Concreteness and Meaningfulness and (b) Concreteness, which was strongly associated with Meaningfulness (see Appendix C). The model outcomes are presented in Table 5, followed by graphical illustrations in Figure 3. We first focus on Meaningfulness and Familiarity, which exhibited similar results, and subsequently report the result for Age of acquisition. The models for the first two indices yielded significant effects of Context and a significant interaction with Modality, indicating a differential effect of modality on the EFL and ESL groups. In light of the interaction, we conducted separate by-group comparisons, breaking down these interactions for each of the two indices.

Figure 3:

Result of psycholinguistic information (log-transformed raw scores).

For Meaningfulness, the EFL group had a significantly lower score in written than in spoken production (b = 0.009, SE = 0.003, p = 0.043, Cohen’s d = 0.264), while modality did not affect the ESL group’s production (b = −0.003, SE = 0.003, p = 0.802, Cohen’s d = −0.106).

For Familiarity, we found significantly lower scores in written than in spoken production for both the EFL (b = 0.007, SE = 0.001, p < 0.001, Cohen’s d = 1.082) and the ESL group (b = 0.004, SE = 0.001, p < 0.001, Cohen’s d = 0.700). However, as indicated by the effect sizes and the graph in Figure 3, the score difference between written and spoken production was greater for the EFL than the ESL group, suggesting a more pronounced impact of modality on the EFL group.

Turning to the results of Age of acquisition, the model revealed a main effect of Context, indicating significantly higher scores for the ESL than the EFL group, and a main effect of Modality, driven by significantly higher scores in written than in spoken production. Unlike the previous two indices, no significant interaction emerged between the two factors, indicating similar patterns across the two production modes for both groups. However, as shown in Figure 3, the ESL group’s score in spoken production was even higher than the EFL group’s score in written production. These results suggest the ESL group’s increased ability to produce later acquired words.

5.4 N-gram information

We observed a strong correlation between range and frequency in both bigrams and trigrams (see Appendix C). Given that frequency is a more commonly used measure, we excluded n-gram range from our analyses. Additionally, we identified strong correlations between Delta P scores derived from the written reference corpus and those from the spoken reference corpus for both bigrams and trigrams. Therefore, we only used bigram and trigram association strength derived from the written reference corpus.

Table 6 presents the model outputs for bigram information, followed by graphical illustrations of the mean scores in Figure 4. Our analyses revealed different outcomes depending on the type of reference corpus used. First, when analyzing bigram frequency using the written reference corpus, we observed a main effect of Modality and a main effect of Context, without their significant interaction. On the other hand, when analyzing bigram frequency using the spoken reference corpus, we found a main effect of Modality, a main effect of Context, and their significant interaction. Post-hoc tests using emmeans showed a significantly lower score in written than in spoken production for both the EFL group (b = 0.223, SE = 0.036, p < 0.001, Cohen’s d = 0.616) and the ESL group (b = 0.116, SE = 0.036, p = 0.008, Cohen’s d = 0.321), yet the effect size was greater for the EFL group.

Figure 4:

Result of bigram frequency (log-transformed raw scores) for written reference corpus (left) and for spoken reference corpus (middle) and bigram association (right).

When analyzing the bigram association using the written reference corpus, we observed a main effect of Modality, with higher scores in written than in spoken production. There was also a main effect of Context, with higher scores in the ESL than the EFL group, and a significant interaction between the two factors. Post-hoc tests revealed a significantly higher score in written than in spoken production for the EFL group (b = −0.172, SE = 0.030, p < 0.001) but not for the ESL group (b = −0.016, SE = 0.030, p = 0.953). These results indicate that the EFL group produced more strongly associated bigrams in written than in spoken production, whereas the ESL group produced strongly associated bigrams equally in both written and spoken production.

Turning to the results of the trigram information, Table 7 presents the model outcomes, and Figure 5 displays visual representations. The analysis using the written reference corpus revealed a main effect of Modality, a main effect of Context, and their significant interaction. Post-hoc tests indicated that the EFL group had a significantly higher frequency score in written production than in spoken production (b = −0.336, SE = 0.053, p < 0.001, Cohen’s d = −0.641), whereas the ESL group had similar scores in both modes (b = −0.107, SE = 0.053, p = 0.174, Cohen’s d = −0.204). When using the spoken reference corpus, the model only showed a main effect of Context, with the EFL group producing more frequent trigrams than the ESL group. These results suggest that while the EFL learners tended to produce more frequent trigrams in written than spoken production, the ESL learners did not show this difference.

Figure 5:

Result of trigram frequency (log-transformed raw scores) for written reference corpus (left) and for spoken reference corpus (middle) and bigram association (right).

When analyzing the trigram association using the written reference corpus, we found a main effect of Modality, indicating higher scores in written than in spoken production, and a main effect of Context, with higher scores in the ESL group than the EFL group. However, there was no significant interaction between the two factors. These results suggest that while both groups produced more strongly associated trigrams in written than spoken production, the ESL group exhibited a higher propensity to produce more strongly associated trigrams than the EFL group.

6.1 RQ1: Effect of modality

Building on previous findings highlighting the influence of modality on L2 production, our study explored whether this modality effect extends to L2 lexical sophistication. The results of our analyses revealed a robust effect of modality, without its interaction with the group factor, in three indices, including word frequency, age of acquisition, and trigram association. Learners in our study demonstrated a tendency to produce less frequent words, later acquired words, and more strongly associated trigrams in writing compared to speaking. These outcomes align with previous findings indicating that writing enables learners to produce more complex and accurate linguistic units compared to speaking (e.g., Ellis and Yuan 2005; Hwang et al. 2020; Kim and Hwang 2022; Kormos 2014; Vasylets et al. 2017). Our findings go further in unveiling the intimate link between production mode and one’s ability to retrieve infrequent and advanced words.

In keeping with the line of research documenting the effect of modality in L2 production, we invoke the different psycholinguistic processes underlying writing and speaking as the major explanation for the modality effect emerging in our sample. As reviewed earlier, the recursive nature of writing confers more cognitive advantages on language users compared to the linear, time-constrained process of speaking (Halliday 2002; Kellogg 1996). From this vantage point, the cognitive burden imposed on our L2 learners may have been less demanding in writing than in speaking, allowing them to produce more sophisticated words in the writing task. In contrast, due to the high cognitive demands during speaking, the learners may have relied more heavily on high-frequency items in the speaking task, which require relatively fewer memory resources for lexical retrieval, leading to reduced lexical sophistication in the speech data.

6.2 RQ2: Interaction of modality and context

Another crucial question we asked was whether the effect of modality would be influenced by L2 learning contexts. For six indices, including academic language, meaningfulness, familiarity, bigram frequency (when analyzed with the spoken reference corpus), bigram association, and trigram frequency (when analyzed with the written reference corpus), we found significant interactions between modality and L2 learning contexts.

We observed similar patterns in terms of meaningfulness, familiarity, bigram frequency, and bigram association, which showed a greater score gap between writing and speaking for the EFL compared to the ESL group. Specifically, the ESL group produced words with similar degrees of meaningfulness and bigrams with similar degrees of association in both writing and speaking, while the EFL group produced significantly less meaningful words and more strongly associated bigrams in writing than in speaking. Additionally, although both groups produced less familiar words and less frequent bigrams in writing than in speaking, this tendency was more pronounced for the EFL group. The EFL learners’ greater lexical sophistication in writing may be attributed to their increased experience with written language. As previously discussed, EFL learners have limited exposure to L2 input, mostly from textbooks, and have limited access to spoken input (Saito 2017). This imbalanced and restricted L2 experience may have resulted in fewer opportunities for EFL learners to encounter and use a variety of words in spoken language, making it more challenging to produce sophisticated words in speaking.

However, we also found an interesting pattern where the EFL group produced more frequent bigrams and trigrams in written production than in spoken production when analyzed with the written reference corpus. This pattern appears inconsistent with the group’s increased sophistication in written production for various indices, as more frequent n-grams suggest reduced sophistication. We speculate that these outcomes may be due to the EFL group’s greater exposure to written input, which includes a variety of n-grams, both frequent and infrequent. In contrast, this group may have employed less frequent n-grams in speech due to their limited experience with spoken language, which restricts them to a small number of infrequent n-grams. Further research is needed to validate this speculation by examining the specific usage of n-grams among EFL learners.

On the other hand, we observed a different pattern of interaction regarding academic language. In this case, the ESL group tended to use more academic phrases in speaking compared to writing, even outnumbering the EFL group’s usage of academic phrases in both written and spoken production. This outcome might be attributed to the ESL group’s language experience, reflecting the influence of their exposure to a significant amount of written and auditory input on a daily basis. It is widely assumed that ESL learners have substantial exposure to both written and spoken language in their everyday lives (Barrot and Gabinete 2021; Tarnopolsky 2000). In some cases, ESL learners may have greater exposure to spoken language, extensively used in various social contexts, compared to written language, which is primarily employed in classroom settings. The ESL group in our study demonstrated a higher use of academic language in spoken production, presumably because they had encountered academic language in diverse social contexts and had numerous opportunities to use it in oral communication.

6.3 Implications of the findings

The findings of the current study, demonstrating the robust effect of modality and its interaction with L2 learning experience in lexical sophistication, contribute novel evidence to the existing literature in the field of L2 lexical production. The previously attested effect of modality has primarily focused on lexical complexity measures such as the number of word tokens and types (e.g., Kormos 2014; Vasylets et al. 2017). While these indices provide useful information with respect to the breadth of L2 vocabulary knowledge (i.e., how many different words one can produce), more in-depth, finer-grained analyses of L2 lexical knowledge may require additional information, such as lexical sophistication, which informs learners’ frequency-based usage of words (Crossley et al. 2016). Our findings thus expand and refine previous work by demonstrating that the degree of L2 lexical sophistication is closely aligned with the usage-based factors related to learners’ language experience, such as modality and L2 learning context. Future research exploring the relationship between L2 learners’ lexical sophistication and the amount of input they receive will advance our understanding of the role of language input in L2 lexical development.

The present study provides empirical evidence supporting the usage-based approach to L2 acquisition (Bybee 2008; Ellis 2006). As outlined earlier, this approach places significant emphasis on the role of L2 experiences in the development of lexical sophistication. Based on this approach, we predicted that the ESL group, benefiting from frequent exposure to words in natural settings, would exhibit a higher level of lexical sophistication compared to the EFL learners, who primarily acquired vocabulary in instructional settings. Consistent with this prediction, our findings furnish evidence of the effect of learning contexts in several lexical sophistication indices. It appears that the ESL learners, through their repeated use of words in diverse contexts, enhanced their ability to process and retrieve words, leading to the use of sophisticated vocabulary during both writing and speaking tasks. In contrast, the EFL learners may have encountered challenges in efficiently entrenching vocabulary knowledge, potentially impeding their ability to produce sophisticated words. Notably, the EFL learners’ restricted experience with spoken language may have contributed to their reduced degree of lexical sophistication in spoken production compared to written production. While these results align with the usage-based approach, further research is required to generalize these findings by exploring the effect of L2 experiences on different aspects of lexical knowledge among learners with diverse L2 learning histories and backgrounds.

6.4 Limitations and suggestions for further work

We acknowledge some limitations that need to be addressed in future research. First, although we controlled for several factors influencing L2 vocabulary use, such as L2 proficiency, genre, and topics, the written and speech samples we analyzed did not come from a homogeneous group, raising the possibility that individual differences might have affected the results. Future studies should address this concern by analyzing written and spoken texts produced by the same learner group to provide a clearer confirmation of the effect of modality.

An additional concern is raised regarding the potential influence of the learners’ L1. As a reviewer pointed out, the learners within the EFL and ESL groups originated from different regions. Given this diversity, we conducted additional analyses where we included the learners’ region as a random factor in our models. However, the model outcomes consistently showed similar results across all the analyzed indices, suggesting that the learners’ region did not significantly affect the observed effects of modality and learning contexts. Nevertheless, it is important to note that this approach does not consider the specific influence of the learners’ individual L1, which can significantly influence L2 production (e.g., Lu and Ai 2015). To address this concern, future studies should incorporate information about the learners’ specific L1 and examine how this factor interacts with modality and learning contexts in shaping L2 lexical sophistication.

In addition, there is a concern regarding the notable difference in the number of words between the written and spoken data. The unbalanced data size reflects the inherently more challenging nature of speaking compared to writing. As pointed out by a reviewer, this disparity could potentially impact the validity of comparisons between these two corpora and introduce biases into the analysis. To address this concern, it is crucial to explore methods that can either balance the data or account for these differences. For instance, future research could consider providing task instructions that limit the number of words in written production, thereby aligning it more closely with the word counts in spoken production.

Another limitation is related to our selection of the ESL and EFL groups. Despite the variable nature of L2 learning experiences, our sample populations were confined to learners in EFL and ESL contexts. While we identified a robust interaction of L2 learning contexts and modality within this specific classification, these learner groups offer a limited representation of L2 learners. Therefore, future work should capture individual variation in language learning experiences by considering individuals’ L2 learning profiles such as the amount of input they receive in diverse contexts.

Furthermore, the exclusive focus on a single topic (i.e., part-time job) in our dataset may have restricted the learners’ specific vocabulary, potentially affecting the outcomes. To enhance the flexibility in vocabulary selection, future studies should involve learner data derived from diverse prompts.

In addition, we recognize a limitation in our study related to the proficiency levels of our participants. In the ICNALE data, students were categorized into different proficiency groups primarily based on their scores in a vocabulary size test. As this test assesses participants’ vocabulary sizes rather than their overall proficiency, students in the same proficiency level may exhibit similar lexical abilities, possibly obscuring variations in lexical sophistication among them. To address this issue, further research should recruit participants with their proficiency fully controlled based on standardized test scores.