The Impact of Lexical Bundle Length on L2 Oral Proficiency

Abstract

Lexical bundles (LBs) are crucial in L2 oral proficiency, yet their complexity in terms of length is under-researched. This study therefore examines the relationship between longer and shorter LBs and oral proficiency among 150 L2 learners of varying proficiency levels at a UK university. Through the analysis of oral presentation data (scores ranging from intermediate to advanced) and employing a combined text-internal and text-external approach (two- to five-word bundles), this study advances an innovative text-internal LB refinement procedure, thus isolating the unique contribution of LB length. Robust regression, dominance analysis, and random forest statistical techniques reveal the predictive power of bigram mutual information (MI) and longer three-to-five-word sequences on higher proficiency scores. Our results show that learners using higher MI score bigrams tend to perform better in their presentations, with a strong positive impact on scores (b = 14.38, 95% CI [8.01, 20.76], t = 4.42; dominance weight = 58.63%). Additionally, the use of longer three-to-five-word phrases also contributes to better performance, though to a lesser extent (dominance weight = 18.80%). These findings highlight the pedagogical potential of a nuanced approach to the strategic deployment of LBs, particularly bigram MI, to foster oral proficiency. Suggestions for future LB proficiency research are discussed in relation to L2 speech production models.

1. Introduction

In recent years, there has been a growing interest in studying multiword sequences (MWSs) in second language (L2) oral proficiency. Foundational research by Pawley and Syder (1983) emphasized the critical role of lexical phrases, or “chunks”, in achieving native-like fluency. They argued that a large part of native speakers’ fluency stems from their use of these prefabricated chunks, which reduce the cognitive load during speech production by allowing speakers to retrieve whole phrases from memory rather than constructing sentences word-by-word. This process, known as chunking, enables more fluent and efficient language use and is crucial for effective communication.

Building on these insights, subsequent research has underscored the significance of MWSs in L2 learning and speaking proficiency (Schmidt 1992; Nation 2013). Lexical bundles (LBs), defined as frequent, contiguous MWSs, have gained particular attention for their role in speech fluency (McGuire and Larson-Hall 2021; Tavakoli and Uchihara 2020) and oral proficiency (Garner and Crossley 2018; Kyle and Crossley 2015; Zhang et al. 2021). The findings from these studies suggest that incorporating MWS-focused activities in language teaching can significantly enhance students’ speaking skills. Despite this, previous studies have predominantly focused on shorter LBs (typically two- and three-word sequences), thus potentially limiting our understanding of learners’ phraseological knowledge and its impact on oral proficiency. Exploring this gap by examining both longer (e.g., as you can see) and shorter (e.g., there are) LBs is critical, as longer LBs could offer enhanced proficiency and show processing efficiencies that are essential for fluent speech production. Responding to calls from recent literature (e.g., Hougham et al. 2024; Tavakoli and Uchihara 2020), our study addresses this gap by investigating the predictive power of both shorter and longer LB usage on oral presentation scores among learners with varying proficiency levels. We aim to isolate the unique contributions of longer LBs using an innovative text-internal and text-external approach, thus providing a more comprehensive understanding of how LB length influences L2 oral proficiency. Addressing this gap is crucial for advancing our theoretical understanding of phraseological knowledge in L2 learners and for developing practical teaching strategies that leverage the benefits of various types of LBs to enhance oral proficiency.

2. Literature Review

2.1. Multiword Sequences and Lexical Bundles

Multiword sequences include various combinations of words that commonly co-occur, such as idioms (e.g., hit the nail on the head), collocations (e.g., heavy rain), phrasal verbs (e.g., look up), proverbs (e.g., make hay while the sun shines) and LBs (e.g., in terms of) (Garner and Crossley 2018). These sequences can be identified through either a phraseological approach or a frequency-based approach (Granger and Paquot 2008; Nesselhauf 2004). The phraseological approach categorizes MWSs using linguistic criteria (Cowie 1981), which are often based on the intuitions of first language (L1) speakers. Previous studies employing this approach (e.g., Boers et al. 2006) frequently rely on human raters to assess the formulaic nature of MWSs, but this can result in low inter-rater reliability. For example, Boers et al. (2006) reported a reliability coefficient of less than 0.60, which is significantly lower than the median inter-rater reliability of 0.92 reported in SLA research by Plonsky and Derrick (2016).

On the other hand, the frequency-based approach, developed by Sinclair (1991) and Nesselhauf (2004), uses corpus-based automated extraction techniques to identify frequent word combinations based on quantitative criteria such as frequency and range. These are often termed lexical bundles or n-grams and can include both structurally complete (e.g., at the end of the day) and incomplete sequences (e.g., in the middle of), regardless of their idiomaticity or structural status (Biber et al. 1999). The objective nature of the frequency-based approach has made it popular in learner corpus research (Paquot and Granger 2012). For instance, Ebeling and Hasselgård (2015) emphasized the benefits of this approach, such as accessing a large quantity of data for quantitative analysis.

In our study, we use the term LB to refer to sequences identified using automatic extraction software based on criteria like minimum frequency, range, and mutual information (MI) scores, which measure the strength of association between word pairs. Our aim is to distinguish LBs from other MWSs, even though there may be some conceptual overlap (e.g., high MI score bigrams like global warming can also be categorized as collocations).

In learner corpus research, LB production can be analyzed using either text-internal or text-external approaches. The text-internal approach (e.g., Biber and Gray 2013) focuses on data within learner corpora. In contrast, the text-external approach (e.g., Tavakoli and Uchihara 2020) examines LBs in terms of criteria like the frequency of co-occurrence in external reference corpora consisting of L1 speaker language. Both approaches have their limitations such as the arbitrary frequency cutoffs in the text-internal approach (Myles and Cordier 2017) and the challenges in measuring longer bundles with text-external analysis tools (e.g., some software programs such as the Tool for the Automatic Analysis of Lexical Sophistication: TAALES (Kyle et al. 2018) can only analyze bigrams and trigrams). Additionally, text-external techniques may not ensure that frequently occurring word combinations in external corpora are psycholinguistically relevant to the learners being studied (Ellis et al. 2009; Myles and Cordier 2017). Our study employs both approaches to assess the contributions of shorter and longer LBs with respect to L2 oral proficiency.

2.2. From Theory to Empirical Findings

Pawley and Syder (1983) lay foundational insights into the importance of multiword sequences in language proficiency through their exploration of lexical phrases. They propose that a speaker’s command of fixed and semifixed phrases, or “chunks” of language, is crucial for achieving native-like fluency. Such lexical phrases ease the cognitive load during speech production and serve as essential building blocks for fluent and idiomatic language use. This early perspective is significant, as it suggests that the repertoire of lexical phrases is a key factor distinguishing more proficient speakers from less proficient ones, thereby enabling more efficient processing and production of language. Building upon the understanding of lexical phrases, Levelt’s (1989) speech production model provides further theoretical justification for the link between LBs and proficiency. The model describes three stages in speech production: conceptualization, formulation, and articulation. During the conceptualization stage, speakers plan the content of their speech. In the formulation stage, they encode lexical and grammatical items in the mental lexicon, activating appropriate lemmas and constructing syntactic structures. Finally, the articulation stage involves implementing the phonetic plan, resulting in speech production. Originally designed for L1 speakers, Kormos (2006) adapted this model to address the specific challenges of L2 speakers, such as a smaller and less structured mental lexicon with fewer formulaic expressions. The model suggests that speakers with a larger repertoire of MWSs can retrieve these sequences as easily as single words during the formulation stage, which reduces cognitive load and provides a processing advantage. This efficiency allows L2 learners to allocate cognitive resources to other aspects of speech production, such as lexical and grammatical accuracy or complexity (Kormos 2006; Skehan 2014). Utilizing MWSs can boost oral fluency, particularly during lexical selection at the formulation stage (Kormos 2006; Levelt 1992). In contrast, speakers with a limited MWS repertoire may struggle, as they expend more cognitive resources when retrieving individual lexical items during the formulation stage. Alongside these frameworks, usage-based theory, as articulated by Ellis (2002) and supported by Bybee (2006), further complements our understanding of how language proficiency develops from the experience and frequency of usage. Ellis (2002) introduces the notion that linguistic competence is significantly influenced by the accumulation of exposure to MWSs, which reinforces their retrieval and production. This perspective aligns with the proceduralization of language use, where proficiency emerges from the frequent and familiar use of language structures. Bybee suggests that repeated exposure to MWSs not only facilitates their entrenchment in the learner’s mental lexicon but also highlights the importance of frequency and familiarity for L2 learning. Together, these theoretical perspectives—Pawley and Syder’s lexical phrases theory, Levelt’s speech production model, and the usage-based theories proposed by Ellis and Bybee—form a multifaceted view of the mechanisms underlying L2 oral proficiency. The current study aims to bridge these theoretical insights with empirical evidence, exploring how the length, frequency, and association strength of LBs contribute to oral proficiency. We seek to illuminate the relationship between LB use and language proficiency, grounded in a broad spectrum of linguistic theory.

To date, most learner corpus-based studies have focused on LBs used in writing rather than speaking (e.g., Appel and Wood 2016; Siyanova-Chanturia and Spina 2020; Staples et al. 2013). This emphasis can be attributed to two main reasons. First, natural speech within L2 classes/tests is often not recorded, whereas L2 students regularly submit written work, thus making written corpora more readily available for analysis. Second, spoken corpora require more time to analyze, as there are additional steps in processing. Specifically, the speech first needs to be transcribed into a text form, and it may need to be divided into speech units for ease of analysis, thus making the analysis of spoken language more labor-intensive and time-consuming than written texts. Given these constraints, most research has naturally gravitated towards written corpora. However, the findings from studies focusing on written language—highlighting the complexity of the MWS–L2 proficiency link and trends regarding the use and proficiency-related evolution of LBs—underscore a rich area for exploration within spoken language as well. Specifically, the research exploring written LBs has revealed general trends, including the following: (a) the quantity of LBs produced by learners decreases as proficiency increases (Appel and Wood 2016; Staples et al. 2013) or with time spent in an English-speaking country (Groom 2009). Less proficient learners overuse bundle tokens and under-use bundle types, which is a trend resembling the use of “phraseological teddy bears” (i.e., overusing high-frequency phrases with which one feels comfortable, as demonstrated in Hasselgård 2019). Research has shown that proficient learners have a firmer and more creative command of lower-frequency bundles whose constituent words are nonassociated (Siyanova-Chanturia and Spina 2020); and (b) less proficient learners rely more on LBs copied from writing prompts or source texts due to their relatively limited lexical repertoires (e.g., Appel and Wood 2016; Staples et al. 2013). Overall, these studies suggest that much remains to be understood about the relationship between LB use and general writing proficiency. While most existing learner corpus-based studies have focused on the use of LBs in written contexts, the insights they provide into the relationship between MWS–L2 proficiency potentially apply to the spoken domain as well. The understanding that learners’ reliance on LBs decreases with proficiency or extended exposure to an English-speaking environment, and that proficiency dictates a learner’s command over lower-frequency bundles, offers a compelling framework to consider for spoken data. The current study seeks to bridge this gap by exploring how these patterns manifest in speech proficiency, which has been less frequently examined. There is potential that the dynamics of LB use in writing, as detailed in these studies, could find parallels in oral production, thereby providing richer insights into speech proficiency and its intricacies.

Several recent studies have examined the link between LB use and general speaking proficiency from a text-external perspective, focusing on the extent to which L2 learners use L1 target-like two- and three-word (bi- and trigram) measures in terms of quantitative indices (frequency, proportion, and association) (e.g., Garner and Crossley 2018; Kyle and Crossley 2015; Zhang et al. 2021). The findings from these studies differ from the typical trend of writing-centered studies (which proposed that less-proficient L2 writers (over)use a greater number of high-frequency bundles). Kyle and Crossley (2015) found significant correlations between (human-rated) oral proficiency scores and several n-gram scores, of which the strongest predictor of speaking proficiency came from high-frequency trigrams, thus suggesting that more skillful L2 speakers use a larger number of highly frequent trigrams. Garner and Crossley (2018) found that beginning-level L2 learners showed the greatest increase in oral production of high-frequency bigrams over the course of their four-month longitudinal study. Zhang et al. (2021) reported that several n-gram measures (e.g., bigram proportion and association: MI and t scores) significantly correlated with (human-rated) oral proficiency scores on story retelling and monologic tasks. Such studies highlight the important role of proficiency in the development of LB use but suggest that further research is required to bring clarity to this research area.

Relatively few studies (e.g., Biber and Gray 2013; De Cock 2004) have examined LB use in spoken corpora from a text-internal perspective. De Cock (2004) examined two- to six-word bundle use among advanced EFL learners compared to L1 speakers. She found that learners’ preferred bundles were less interactional and included relatively few vagueness markers (e.g., or something, kind of) compared to L1 speakers. Biber and Gray’s (2013) study of spoken and written responses to the TOEFL iBT showed a slightly more complex pattern than other n-gram studies. They reported that intermediate-level participants produced a greater number of bundles (four-word units) than their lower- and higher-proficiency groups, thus suggesting a general developmental progression in which lower-level participants use a smaller number of bundles, intermediate-level participants overuse a larger number of bundles, and high-scoring participants show greater control and creativity in using the bundles they have acquired (p. 37). To summarize, these studies suggest a complex perspective and a need for further research on how language use varies among individuals from diverse backgrounds.

While such studies have offered insights into the patterns of LB use in spoken corpora, a broader context emerges when we consider the significant relationship between MWS use and speech fluency, and it is clear across various teaching contexts. Studies that show significant and positive relationships between MWS use and speech fluency (including L2 proficiency in the broader sense) come from a range of different teaching contexts (e.g., Boers et al. 2006; Hougham et al. 2024; McGuire and Larson-Hall 2017, 2021; Stengers et al. 2011; Suzuki et al. 2022; Tavakoli 2011; Tavakoli and Uchihara 2020; Uchihara et al. 2021; Wood 2009, 2010). Boers et al. (2006) and Stengers et al. (2011) found strong links between the number of MWSs used (in story retelling tasks) and (perceived) oral ability scores in a Belgian EFL context. Wood (2009) examined the effect of MWS-focused teaching on MWS use and oral fluency in a case study (N = 1) in a Canadian ESL context. Wood found that MWS-focused instruction can lead to increased MWS use and increased spoken fluency over a short period (six weeks). Wood (2010) also found similar results with a slightly larger sample size (N = 11) in a similar context over a longer period (six months). Tavakoli (2011) compared the pausing patterns of L1 versus L2 speakers’ performance in a UK university context. She found that L2 learners rarely paused in the middle of multi-word units, thus providing further corroborating evidence that lexical chunks facilitate fluency. Similarly, Uchihara et al. (2021) found that speakers who provided more low-frequency MWS (collocational type) responses to a word association task (Lex30) spoke more rapidly with fewer silent pauses. McGuire and Larson-Hall (2017) replicated Wood’s (2009) study in an American ESL study abroad context. They reported a moderately strong relationship between all participants’ MWS use and fluency measures. Tavakoli and Uchihara’s (2020) study, reporting the link between two- and three-word LBs and one objective measure from each aspect of utterance fluency (speed, breakdown, and repair) across assessed proficiency levels in a UK university context, represents the first systematic study of its kind. Tavakoli and Uchihara reported that greater LB use (a larger proportion of frequent LBs and more frequent LBs) was positively and significantly related to higher speaking ability scores and with some fluency aspects (faster articulation rate and fewer pauses within clauses). Suzuki et al.’s (2022) task repetition intervention study examined the use of single words and trigrams on speed, breakdown, and repair fluency aspects. They found that the recycling of more complex MWSs through task repetitions seemed to facilitate proceduralization (i.e., more efficient retrieval of MWSs), but they also found that such reuse had both positive and negative influences on midclause pauses specifically, as well as fewer but longer pauses within clauses, which may show that learner encoding systems were in the process of restructuring. Hougham et al. (2024) examined the relationship between both shorter (bi- and trigram) and longer (four–to-five-word) LBs and three dimensions of fluency (speed, breakdown, and repair) using both text-internal and text-external techniques. They found that using longer LBs, specifically four-to-five-word sequences of high collocational quality (those with high MI scores), significantly enhanced speech fluency by reducing the frequency of pauses and repairs. Moreover, they uncovered a correlation between frequent combinations of two words and a faster rate of speech, whereas complex combinations (those with high mutual information scores) were found to slow down speech.

2.3. Gaps and Unexplored Areas in Previous Research

While the studies reviewed here support the hypothesis of a positive relationship between MWS use and proficiency, they might be limited in at least six important ways. Foremost, studies focusing on relatively short (two- and/or three-word) sequences might not fully capture learners’ actual phraseological knowledge and how it relates to oral fluency units. This is exemplified by multiple studies (e.g., Garner and Crossley 2018; Kyle and Crossley 2015; Kyle et al. 2018; McGuire and Larson-Hall 2021; Suzuki et al. 2022; Tavakoli and Uchihara 2020; Zhang et al. 2021). For example, using longer LBs might be more beneficial for improving aspects of oral fluency and increasing high-stakes assessment scores. Emphasizing the potential importance of longer LBs, Tremblay et al. (2011) demonstrated that longer (four- and five-word) LBs offer online processing advantages over non-LBs in receptive tasks. Hougham et al. (2024) also showed that longer LBs of high collocational quality enhanced various aspects of fluency. Despite this understanding, the effects of longer linguistic units on achieving fluency in speech production have not been fully explored. Given what we know from Pawley and Syder’s insights into the significance of lexical phrases, the comprehensive framework provided by Levelt’s speech production model, as well as the principles of usage-based theory by researchers such as Bybee and Ellis, it is reasonable to hypothesize that employing longer LBs can lead to enhanced processing efficiency. Second, many previous studies have had methodological or contextual limitations, such as measuring MWSs subjectively using a criteria checklist and L1 speaker intuition (e.g., McGuire and Larson-Hall 2017; Wood 2009). Third, most previous research is restricted to investigating the MWS proficiency link with a learner-external approach (i.e., examining learners’ use of selected sequences that are thought to be formulaic in L1 speaker English and identified in advance as formulaic or quantifying a text’s formulaicity by checking the frequency of all of its constituent word sequences against an external reference corpus) (e.g., Garner and Crossley 2018; Tavakoli and Uchihara 2020; Zhang et al. 2021). Only one study by Hougham et al. (2024) has systematically attempted to employ both text-external and text-internal methods to analyze learner-produced LBs in relation to aspects of oral fluency, but their study did not examine oral proficiency scores. Fourth, some previous studies have suffered from small sample sizes (e.g., N = 19 in McGuire and Larson-Hall 2017; N = 1 in Wood 2009; N = 11 in Wood 2010). Fifth, the LBs examined in studies based on learner corpora frequently have varying lengths, typically ranging from two to six words. Additionally, the criteria for extracting these LBs, such as frequency and dispersion, differ significantly across different studies. Therefore, it is important to view the aforementioned findings and general patterns as hypotheses that require further testing using alternative corpus data within diverse contexts across various proficiency levels.

3. The Current Study

Informed by prior theoretical insights (e.g., Levelt 1989; Pawley and Syder 1983) and responding to the call for a more comprehensive approach to MWSs in EFL research (Hougham et al. 2024; Tavakoli and Uchihara 2020), our study explores the under-investigated area of longer LB usage, hypothesizing that these can offer significant processing advantages for L2 speakers, which is a notion previously suggested but not empirically tested across a range of proficiency levels. Our research question is designed to directly address the identified need for more comprehensive analyses of LB usage in relation to proficiency scores. This is done by building on and extending the work of Tavakoli and Uchihara (2020), as well as Hougham et al. (2024). We investigate the relationship between the use of shorter (bi- and trigrams) and longer (three-to-five-word) LBs and oral proficiency scores. By comparing and contrasting findings across different learner populations and proficiency scores, the current study seeks to contribute to a more detailed understanding of how LBs function in L2 speech production models and inform future LB proficiency research directions. By examining a broader dataset (150 L2 learners at varying proficiency levels from a UK university’s presessional course) and using an innovative text-internal LB refinement procedure that identifies more structurally complete and useful LBs, the current study allows us to explore the relationship between LB usage and oral proficiency, aiming to provide a more comprehensive understanding of the relationship. Additionally, the current study seeks to explore the extent to which both shorter and longer LBs can predict speaking proficiency scores. By employing robust regression, dominance analysis, and random forest techniques, we not only aim to validate previous findings but also uncover new patterns, potentially leading to more effective pedagogical strategies for enhancing oral proficiency in diverse L2 learning contexts.

The current study addresses the following research question: To what extent do longer lexical bundles (three-to-five-word units) predict oral proficiency scores compared to shorter lexical bundles (bi- and trigrams) in L2 learners?

Based on the findings in Tavakoli and Uchihara (2020) and the theoretical frameworks of speech production (e.g., Kormos 2006), the current study hypothesizes that longer LBs will significantly contribute to higher oral proficiency scores, potentially more so than shorter LBs, due to their ability to enhance processing efficiency and fluency.

4. Materials and Methods

4.1. Participants

The participants were 150 language learners taking a presessional course at a UK University. They were from 20 different L1 backgrounds (outlined in

Table 1. Participants according to L1 backgrounds (N = 150).

Nationality	n	Nationality	n	Nationality	n
Chinese	101	Indonesian	2	Taiwanese	1
Saudi Arabian	9	Bangladeshi	1	Chilean	1
Turkish	9	German	1	Egyptian	1
Japanese	7	Bahraini	1	Georgian	1
Thai	6	Kuwaiti	1	Pakistani	1
Indian	2	Ghanian	1	Russian	1
Colombian	2	Italian	1

), with most of the participants being L1 Chinese (n = 101), L1 Saudi Arabian (n = 9), or L1 Turkish (n = 9). Participants’ raw presentation scores (described in the next section) were first converted into IELTS bands ranging from 6.5 to 7.5. Next, they were categorized into three groups depending on these bands. From a larger pool, we randomly selected 50 individuals for each level, ensuring balanced representation across the three bands. There were 50 participants at IELTS level 6.5, 50 at IELTS level 7, and 50 at IELTS level 7.5 (see

Table 2. Participants’ IELTS-rated speaking proficiency levels.

IELTS speaking level	6.5	7	7.5	Total
n	50	50	50	150

).

4.2. Oral Presentation Tasks and Proficiency Scores

All participants completed oral presentations comprising a 7 min presentation (monologue) in small groups with the aid of PowerPoint slides as part of the presessional course. It is important to note that these presentations were not conducted specifically for the purposes of the current study but were completed as a requirement of the presessional course at Queen Mary University of London. Their outputs were video recorded online and assessed by teachers using grading descriptors (in equal measure): presentation content, presentation structure, seminar leadership, language fluency, and language accuracy (see the full descriptors in Appendix A). The oral presentation tasks were designed to cover a wide range of subtopics under the overarching theme of globalization, which had been the focal point of the participants’ 5-week academic English course. While globalization served as the central theme, participants were encouraged to explore this broad topic through various lenses, ranging from economic and legal perspectives (e.g., how globalization affects development and international trade law) to more specialized areas (e.g., acoustic telemetry in fisheries). This diverse range of subtopics allowed participants to delve into areas aligned with their academic interests and expertise. The students were from a wide range of disciplines: humanities, law, science, technology, engineering, and mathematics majors. From the video recordings, we took 3 min speech samples starting at the 30 s timestamp. As the 3 min sections of speech analyzed came from the “presentation” part of their seminar, we are specifically analyzing monologic oral presentations rather than freer dialogic speech. These samples were transcribed using Sonix.ai web-based software, and the transcriptions were checked for accuracy by a research assistant and double-checked by the first researcher. We used the transcripts in the lexical analyses (described in detail below), and we used the raw scores and the banded IELTS scores as the measures of oral proficiency in the current study.

4.3. Measuring Lexical Bundles: A Two-Pronged Approach

For the current study, we adopt a frequency-based approach using both text-internal and text-external techniques to isolate the unique contribution of shorter versus longer LBs.

4.3.1. Text-External Lexical Bundle Analysis and Measures

Following previous studies (Garner and Crossley 2018; Hougham et al. 2024; Tavakoli and Uchihara 2020), we used three n-gram indices (proportion, frequency, and association) to objectively measure the use of shorter LBs, specifically two- and three-word contiguous sequences (i.e., bi- and trigram tokens) in our learner corpus. TAALES version 2.0 was used to calculate three kinds of n-gram scores, producing six score indices (two proportion, two frequency, and two association indices). As our external reference corpus, we chose the spoken subsection of the Corpus of Contemporary American English (COCA Davies 2009), which comprises 79 million words from transcriptions of a wide range of TV and radio programs. Our choice of this spoken corpora was in alignment with research findings showing a gap in L2 learners’ spoken and written vocabulary sizes (Uchihara and Harada 2018) and differences in lexical profiles between spoken and written modes (Dang et al. 2017). We maintained consistency between the modality in which L2 words were elicited and the modality of the reference corpus based on the practices used in previous studies (e.g., Uchihara and Clenton 2020; Uchihara et al. 2021).

In our current study, we use proportion score indices to measure the occurrence of bi- and trigrams in our learner speech sample data. These bi- and trigrams are also among the 30,000 most frequent ones in the external reference corpus (COCA). Higher proportion scores show that participants in the sample produced a higher percentage of high-frequency, target-like bi- and trigrams. Higher frequency scores show that participants in our sample produced a larger number of high-frequency target-like bi- and trigrams. Logarithmic bi- and trigram scores, instead of raw frequency scores, were used to control for Zipfian effects common in word frequency lists (Kyle and Crossley 2015; Tavakoli and Uchihara 2020). Association score indices measure the association strength between individual words within bigrams and trigrams. Of the five association measures available in TAALES, the one association measure we used was mutual information score.1 MI score measures the strength of association between two words. MI scores show the strength of word associations, with higher scores suggesting stronger associations. However, MI also focuses on word pairs that are not commonly found together (Schmitt 2010, p. 130). Before n-gram analysis using TAALES, all the transcripts were cleaned by correcting any misspellings and mispronunciations and removing any markings of filled pausing (i.e., ums, uhs, etc.). The resulting transcripts ranged between 216 and 436 words (M = 310.78, SD = 47.32).

4.3.2. Text-Internal Lexical Bundle Identification and Refinement Procedures

We adopted a text-internal approach to isolate and measure the unique contribution of longer LBs to proficiency. As a first step, we conducted frequency analyses using AntConc (Anthony 2022) to generate lists of the most frequently used four-word LBs in the learner corpus. The frequency and dispersion thresholds used to identify lexical bundles vary from study to study. Figures used for “frequency cut offs are somewhat arbitrary” (Hyland 2008, p. 8) depending on both the size and specificity of the corpus. For relatively small spoken corpora like the one in this study, a raw cutoff frequency has often been used, ranging from two to ten occurrences (e.g., Altenberg 1998; Biber and Barbieri 2007; De Cock 1998). Given the small size of the spoken corpus in the current study (46,617 words), for four-word combinations to qualify as lexical bundles, we used a cutoff point of three or more occurrences in at least three texts, following Biber and Barbieri (2007). These minimum figures help to ensure that the identified bundles are not idiosyncrasies confined to occurrences produced by an individual speaker. This resulted in 447 instances of four-word LBs that met these criteria. To deal with the issues of overlap and structural incompleteness among these instances, we used a refinement procedure developed by previous researchers (Wood and Appel 2014) aiming to identify more structurally complete and useful LBs. We split each four-word sequence (e.g., I will give you) into two constituent three-word clusters (e.g., I will give, will give you). The frequency of the two three-word clusters in the corpora were identified and compared. If the frequency of one three-word cluster was at least double the frequency of the other, the more frequent cluster was classified as the root structure and the fourth word was considered as a word that commonly occurred with that structure and was put in parentheses. For example, I will give occurred over two times (freq = 41) than will give you (freq = 20) in the current data set. Therefore, the final resultant structure is in this case: I will give (you). Another example from the current data set is the sequence the first part is. Since the first part (freq = 41) occurred over two times more frequently than first part is (freq = 17), the final resultant structure is the first part (is). The refinement process produced a list of 119 multiword structures, primarily comprising core three-word phrases and four-word structures, with a smaller number of longer five-word structures included.

Table 3. Numbers and examples of structures identified at different lengths.

Lexical bundle length	three-word units	four-word units	five-word units	Total
Number of units	55	58	6	119
Examples	I’m going to the first part I will give	I would like to at the end of at the same time	if you have any questions the presentation will be brief in a socially responsible way

Note: Contracted forms (e.g., I’m) were counted as a single word; for example, I’m going to was considered to be a trigram rather than a quadgram.

shows the number of three-word (55), four-word (58) and five-word structures (6) identified. The examination of the extensive list provided in Appendix B shows that many of the resulting three-word structures are self-contained units in terms of semantics or structure. For instance, with the development of forms a complete unit. This observation suggests that the refinement procedure successfully pinpointed additional core structures.

It is crucial to differentiate the newly identified structures from traditional LBs (and other multiword structures) when describing them, as this refinement procedure goes beyond the usual LB approach. Recall that the traditional LB approach strictly identifies multiword sequences within two parameters: frequency and range. For this reason, when making modifications or refinements to traditional LB methods, previous researchers (e.g., Simpson-Vlach and Ellis 2010; Wood and Appel 2014) have used different terminology to refer to such refined or modified LBs. Simpson-Vlach and Ellis (2010), for instance, started with an LB approach and then applied additional criteria (e.g., human ratings of formula teaching worth combined with MI score as a measure of collocation strength), thus aiming to identify more useful multiword units for teaching purposes. Simpson-Vlach and Ellis (2010) adopted the general term “formulaic language” to describe the word combinations identified in their study. Another example is Wood and Appel (2014), who developed the LB refinement procedure used in the current study. Wood and Appel adopted the general term “multiword constructions” to refer to the refined LBs identified in their study. “Formulaic language” and “multiword constructions” have also been used as umbrella terms in other studies such as Liu (2012). Because of the fuzzy nature of boundaries between many types of semifixed multiword combinations, and because each study’s identification and refinement methods are different, there is no consensus in the literature as to which terms apply in all cases. Although it is challenging to pin down a consistent definition of MWSs across all studies, most researchers agree that it is important to distinguish between different types of MWSs where possible and to report clearly how MWSs are identified in each study to facilitate comparisons of research findings across studies. In the current study, to avoid terminological confusion, we use the term “refined LBs” to refer to the refined LB list (see Appendix B) produced by the above-described refinement procedure.

4.4. Scoring the Use of Text-Internal Bundles

It is important to quantify usage of the list of refined LBs so that we can run various quantitative analyses (e.g., multiple regression) and compare the relationships between different types of LBs (unrefined text-external vs. refined text-internal; shorter vs. longer) with oral presentation scores. To do so, we awarded one point for each identified refined LB used by each participant. We tallied up the total number of points for each participant, arriving at a three-, four-, and five-word usage score for each participant. As for text-internal MI scores, we extracted MI scores for sequences of various lengths using the Collocate 2.0 software program (Barlow 2015). MI scores for three-to-five-word sequences have been used in several studies (e.g., Ellis et al. 2008; Simpson-Vlach and Ellis 2010), as they appear to offer a reliable indication of phrasal coherence. For each refined bundle used by each participant, we awarded the corresponding MI score. We then tallied all MI scores and gave each participant a total MI score.

4.5. Statistical Analyses

We analyzed six text-external n-gram (two- and three-word) measures and two text-internal refined n-gram (three-, four-, and five-word) measures. To examine our research question, that the use of LBs of varying length can predict oral presentation raw scores, we selected regression analysis. Initially, we examined the assumptions underlying regression models. The presentation raw scores variable yielded a Shapiro–Wilk p-value of 0.04, thus showing a significant deviation from a normal distribution. To identify influential outliers within the data, Cook’s distance was employed with a conservative threshold of 4/n, thus facilitating the assessment of the impact of individual data points on the regression model. Several data points emerged as influential outliers, surpassing the Cook’s distance threshold. Their presence implies a potential influence on the estimated regression coefficients (see Supplementary Materials for detailed results of these checks). Subsequently, guided by Larson-Hall (2015, p. 264), we conducted a robust regression using MM estimation with the “rlm” function in the MASS package in R (R Development Core Team 2019). We chose the “rlm” function because it aptly accommodates data sets with non-normal distributions and outliers.

In an effort to make the results more robust, we decided it was relevant to conduct additional analyses on the relative importance of each predictor variable in explaining the variance in the model. We pursued this inquiry through a dominance analysis (DA) using the “calc.relimp” function from the “relaimpo” package in R (Grömping 2006). DA can effectively address correlations among predictor variables and can help in better understanding the unique contribution of each PV to the criterion variable in multiple regression analysis as opposed to relying solely on possibly misleading standardized beta coefficients (Mizumoto 2022). DA facilitates comprehension by computing dominance weights for each predictor, which show the mean impact of a variable on the predictability of all potential subsets of predictors, consequently presenting a thorough picture of the influence of each predictor on the outcome.

To achieve a more precise estimation of the importance of each variable in the multiple regression model, it is important to conduct dominance analysis in combination with random forests analysis (Mizumoto 2022). The random forests approach is a nonparametric machine learning model, meaning it can offer more precise outcomes when multiple regression assumptions are violated (Liakhovitski et al. 2010). Using random forests allows researchers to acquire a nuanced perspective on variable importance. Hence, following the guidelines by Mizumoto (2022), we integrated the random forests analysis using the Boruta package in R (Kursa and Rudnicki 2010). The Boruta algorithm, specifically designed for feature ranking based on random forests, runs the random forests multiple times. It labels features (or predictors) as “confirmed”, “rejected”, or “tentative” based on their significance compared to randomized shadow features. “Confirmed” predictors are deemed significant, “rejected” ones are considered unimportant, and “tentative” labels are reserved for predictors whose importance remains uncertain. We show these using boxplots in the figures presented in the following section.

Integrating Boruta with robust regression and DA allows for a comprehensive analysis, where each method compensates for the limitations of the others. Robust regression ensures that our model is not unduly influenced by outliers, providing reliable coefficient estimates even when the data distribution is non-normal. DA helps clarify the relative importance of each predictor by showing their unique contributions to the model while addressing correlations among predictors. Meanwhile, Boruta offers a rigorous feature selection process that ranks predictors based on their importance and independently of the distributional assumptions. By using Boruta, we can validate which predictors are truly significant, providing a layer of verification to the results obtained from robust regression and DA. This combination of methods allows for triangulation, where the results from one method can support and validate the findings of the others, thus possibly enhancing the overall reliability and robustness of our conclusions. In what follows, we present the descriptive statistics first, we then report the robust regression model in combination with DA and random forests to address our research question.

5. Results

Table 4. Descriptive statistics for presentation scores and n-gram measures.

Measure	M	SD	Median	Minimum	Maximum
Presentation raw scores	62.71	5.53	63.00	53.00	78.00
Bigram log frequency	1.32	0.12	1.31	1.04	1.73
Bigram proportion	0.44	0.07	0.44	0.26	0.63
Bigram association MI	1.54	0.17	1.54	1.13	2.09
Trigram log frequency	0.68	0.13	0.67	0.41	1.07
Trigram proportion	0.12	0.04	0.12	0.04	0.26
Trigram association MI	2.53	0.25	2.54	1.91	3.21
Three-to-five-word usage	10.29	5.10	10.00	1.00	27.00
Three-to-five-word MI	132.39	73.57	117.97	19.87	399.56

Note: M = mean. SD = standard deviation. MI = mutual information.

shows the descriptive statistics for the presentation raw scores and different LB measures used in this study. The table includes the mean (M), standard deviation (SD), median, minimum, and maximum values for each measure. These statistics provide an overview of the distribution and central tendency of the presentation scores and n-gram measures, highlighting the variability and range of the data.

A boxplot was created to visualize the distribution of presentation scores across different L1 groups (Figure 1). The boxplot provides a clear representation of the central tendency and variability within each group. Although a Kruskal–Wallis test did not find statistically significant differences in the presentation scores across L1 groups (Kruskal–Wallis chi-squared = 29.825, df = 19, p-value = 0.054), the boxplot offers valuable insights into the data distribution. It shows that most L1 groups have similar median scores, but there is considerable variability within some groups. The blue dots represent individual data points, highlighting the spread of scores within each group.

Table 5. Robust regression and dominance analysis (criterion: presentation raw scores).

Predictor	B	95% CI	SE	T	Dominance Weight (%)
Intercept	39.02	16.45, 61.58	11.51	3.39
Bigram frequency	8.10	−4.17, 20.37	6.26	1.29	0.008 (3.41%)
Bigram proportion	−12.09	−39.57, 15.39	14.02	−0.86	0.006 (2.64%)
Bigram MI	14.38 *	8.01, 20.76	3.25	4.42	0.129 (58.63%)
Trigram frequency	−3.07	−13.10, 6.95	5.11	−0.60	0.003 (1.23%)
Trigram proportion	4.15	−58.81, 67.11	32.12	0.13	0.007 (3.23%)
Trigram MI	−2.05	−6.24, 2.14	2.14	−0.96	0.007 (3.28%)
Three-to-five-word usage	0.53*	0.06, 0.99	0.24	2.22	0.041 (18.80%)
Three-to-five-word MI	−0.02	−0.05, 0.01	0.02	−1.25	0.019 (8.78%)
Total					0.220 (100%)

Note: MI = mutual information. * Results are marked significant if the 95% confidence interval excludes zero.

shows the findings from the robust regression and dominance analysis with the criterion variable being presentation raw scores. The robust regression analysis unveils multiple predictors impacting the presentation raw scores. Bigram MI displayed a strong positive association (b = 14.38, 95% CI [8.01, 20.76], t = 4.42), accounting for a dominant 58.63% of the variance in presentation raw scores. This implies that learners using more bigrams with higher MI scores tend to score higher on their presentations, holding other variables constant. Examples of bigrams with high MI scores produced by high-scoring participants include carbon dioxide, Kyoto Protocol, and global warming (for more examples, see Appendix C). Three-to-five-word usage also showed a significant positive influence on the model (b = 0.53, 95% CI [0.06, 0.99], t = 2.22), contributing to 18.80% of the variance in the presentation raw scores. This suggests that individuals employing specific three-to-five-word phrases achieve higher presentation scores. Other predictors like trigram frequency, trigram proportion, and trigram MI indicated lesser influence, with each contributing less than 4% to the total dominance weight.

In Figure 2, the dominance weights are presented in descending order starting from the predictor variable with the highest weight (bigram MI) and ending with the one with the lowest weight (trigram frequency). Figure 1 helps us visualize each predictor’s relative importance, highlighting bigram MI as the most important predictor among all variables in our study.

Figure 3 shows a variable importance plot derived from random forests using the Boruta algorithm. The Boruta results confirm the importance of five attributes: Bigram MI, three-to-five-word usage, trigram proportion, three-to-five-word MI, and bigram proportion. Two attributes, trigram frequency and trigram MI, were confirmed as unimportant, while bigram frequency remained tentative. Detailed results and the R code are available in Supplementary Materials.

Overall, the Boruta analysis, along with dominance and robust regression analyses, together highlight the substantial influence of bigram MI and three-to-five-word usage on presentation scores. These consistent findings across different analytical techniques underscore the importance of specific lexical choices, especially bigram MI, in determining speaking performance, thus offering a triangulated insight into how LBs impact speaking proficiency.

6. Discussion

The overarching aim of this study was to explore the impact of LB usage on oral proficiency across a broad dataset and to employ a novel LB refinement technique for a more detailed analysis. Our research question asked to what extent LB usage of varying lengths could predict raw presentation scores. The following discussion has been structured to address this research question in relation to theoretical frameworks, previous findings, and our own hypothesis.

The robust regression and dominance analysis identified bigram MI as a significant positive and dominant predictor of raw presentation scores (b = 14.38, 95% CI [8.01, 20.76], t = 4.42; dominance weight = 58.63%). This was complemented by the finding that three-to-five-word LB usage was also found to be a significant positive and powerful predictor of presentation success (b = 0.53, 95% CI [0.06, 0.99], t = 2.22; dominance weight = 18.80%). These two LB measures’ significances were reinforced by the variable importance plot generated by the Boruta algorithm, showing consistency across different approaches. Our discussion will thus focus on these two LB measures.

Our finding that learners who effectively use text-external high MI bigrams achieve higher proficiency levels are consistent with Pawley and Syder’s (1983) lexical phrases theory. Pawley and Syder argued in their seminal work that the native-like selection of expressions, which includes collocations and idiomatic phrases captured by measures like bigram MI, is an integral aspect of speaking a language fluently. Their lexical phrases theory posits that a speaker’s proficiency is marked by the ability to produce sequences of words that native speakers recognize as familiar, suggesting that native-like proficiency is to some extent a function of memory for lexically stored sequences rather than just rules for combining words. Our findings support and extend this notion by suggesting that not just the presence of LBs, but their “rare exclusivity”, which is the main practical effect of the MI score (Gablasova et al. 2017, p. 10), differentiates more proficient speakers from their less proficient counterparts. The empirical evidence presented in this paper supports the view that language proficiency, especially in productive skills such as speaking, involves the use of language patterns that are characteristic of L1 speaker usage. Our findings suggest that as L2 proficiency increases, so does the use of high MI score bigrams, indicating a narrowing gap in collocational usage between L2 and L1 speakers. These finding highlight the importance of integrating specific types of LBs, particularly those with high MI, into language assessment and instruction, supporting the idea that proficiency is not merely a matter of lexical range but also involves the strategic use of linguistically sophisticated and exclusive lexical patterns.

Our results also align with Kormos’s (2006) extension of Levelt’s model to L2 speakers by illustrating the formulation stage’s critical role in speech production. Specifically, the increase in text-external bigram MI and longer text-internal LB usage with proficiency suggests a more efficient lexical selection process among higher proficiency learners, echoing the notion that advanced speakers can activate and employ appropriate lemmas with greater ease. This efficiency likely contributes to the freeing up of cognitive resources for other processing needs, which is essential for achieving fluency. It provides empirical support for the theory that a larger MWS repertoire enables L2 learners to enjoy a processing advantage by reducing the demands on cognitive resources, thereby facilitating more fluent and sophisticated language production. In addition, our findings that bigram MI and longer (three-to-five-word) LB usage predicts higher scores echoes the tenets of usage-based theory, specifically as it relates to language proficiency. Usage-based theory suggests that language learning is exemplified through the increased use and understanding of recurrent patterns or constructions in language, which are acquired through exposure and use (Bybee 2006; Ellis 2002). Bigram MI within the usage-based framework can be seen as a proxy for the type of patterned use that is a hallmark of language proficiency. Bigram MI measures the exclusivity and specificity of the co-occurrence between two words, reflecting how their joint appearance significantly exceeds what would be anticipated based on their independent distribution across texts. This measure not only highlights word pairs that share a unique connection but also underscores the meaningful combinations that are preferentially utilized by more proficient speakers. The high bigram MI scores in our study suggest that more proficient language users are more likely to employ exclusive and information-rich lexical sequences. In a sense, bigrams with high MI scores are those that are entrenched in the linguistic repertoire of proficient speakers, thus reflecting common usage patterns.

Our findings, highlighting the predictive power of bigram MI for proficiency levels, are also consistent with key insights from corpus-based SLA studies that have explored collocational usage differences between L1 and L2 writers. These findings, while encouraging, are for writing, while our paper considers speaking. Notably, the research in this area has consistently found that L1 users tend to produce collocations with higher MI score values compared to L2 users (Durrant and Schmitt 2009; Ellis et al. 2015; Schmitt 2012). The work of Schmitt (2012, p. 6) emphasizes this disparity, concluding that the absence of high MI collocations is a distinctive marker of non-native versus native production. Our findings also echo the findings of previous research, which compared bigram usage among L2 writers across different proficiency levels. Granger and Bestgen (2014) compared the use of collocations by intermediate and advanced L2 writers, finding that essays scoring in the advanced range of the CEFR had higher proportions of bigrams with high MI scores than essays scoring in the intermediate range. These convergent findings highlight the significance of the MI score as a measure for distinguishing between the collocational choices of L1 and L2 users, as well as between intermediate and advanced L2 users, thus highlighting its utility and growing importance in language research. The differential use of high MI score collocations between L1 and L2 users and between intermediate- and higher-proficiency L2 users highlights a key aspect of language proficiency: the ability to employ exclusive, less frequent word combinations beyond common collocations.

However, a notable divergence arose concerning our bigram MI finding and the other existing literature focusing on spoken production. Notably, Tavakoli and Uchihara (2020) found a general decrease in MI with rising proficiency, suggesting a broadening between LB combinations as learners become more proficient. Conversely, our results showed an increase in bigram MI with proficiency levels (except for a slight decrease in trigram MI). Although both studies used TAALES software to measure bigram MI through a text-external approach, the difference in findings could be due to the distinct data sets, methodologies, or proficiency level groupings employed in the two studies. While Tavakoli and Uchihara used a Kruskal–Wallis H test because of violations of the homogeneity of variance, our study applied robust regression to account for outliers, which may have contributed to the contrasting outcomes in the MI trends. While both studies corroborate the trend of increased LB usage with proficiency, the current analysis contributes a novel perspective by documenting the pattern of bigram MI increase and by bringing to light the intricate usage of longer LBs as learners progress. Such contrasting results underscore the need for further research to explore the intricate dynamics of bigram MI use and proficiency levels, hopefully enriching our understanding of effective language use by L2 speakers.

Regarding our hypothesis that the use of LBs of all lengths would positively correlate with raw oral presentation scores, the results show a mixed but insightful picture. The robust regression and dominance analysis, corroborated by the Boruta algorithm, strongly support the hypothesis in the case of bigram MI and three-to-five-word usage. Bigram MI, which emerged as the most powerful and dominant factor, accounted for a significant portion of the variance in presentation scores. This likely reflects the importance of conciseness in effective oral communication, where the use of meaningful and compact word pairs (bigrams with high mutual information) enables speakers to convey their points succinctly, thereby engaging the audience effectively. The positive influence of three-to-five-word usage also highlights the value of specificity. These slightly longer LBs, which we measured through the text-internal approach, are critical for articulating complex or specific ideas in a clear and focused manner, enhancing the speech’s informativeness without becoming overly wordy. However, our hypothesis found less support with other types of LBs. Bigram proportion, for instance, showed a negative nonsignificant association with presentation scores, and trigram-related measures (frequency and MI) had lesser influence. This suggests that, while certain LBs positively correlate with presentation performance, not all LB types show the same level of predictive power. These findings partially validate our hypothesis, thus underscoring the significant predictive power of specific types of LBs, particularly bigram MI, in relation to raw presentation scores.

In summary, while our hypothesis that all lengths of LBs would positively predict oral proficiency scores was only partially supported, the current study contributes insights into the specific lexical features (bigram MI in particular) that are most predictive of higher proficiency scores. The current study’s findings enhance the existing literature by providing a nuanced picture of how LB usage evolves at higher levels of language learning. Our findings also contribute to research methodology by showing the effectiveness of combining robust statistical techniques with machine learning algorithms like Boruta to strengthen the robustness of educational research. The convergence of evidence from different analytic techniques strengthens the reliability of the current study’s findings. Using robust regression helped to mitigate the influence of outliers, dominance analysis provided insights into the relative importance of predictors, and the Boruta algorithm offered an additional layer of confirmation about which LB attributes are truly influential. Such methodological triangulation enhances the study’s credibility, allowing for more confident conclusions regarding the predictive power of LB usage for speaking proficiency.

7. Limitations and Suggestions for Future Research

Although the results are encouraging, there is still room for future research to overcome the current study’s limitations. One notable area that was not explored in the current study is the potential influence of different presentation topics on LB usage and presentation scores. This unexplored area might shed further light on findings such as the increased use of high MI score bigrams among higher-scoring learners. For example, Gablasova et al. (2017) emphasize that the MI score favors less frequent and more specialized combinations, such as technical terms, which can vary significantly with the nature of the presentation topic. They note that “the technical nature of a specific topic can influence the strength of collocations as measured by the MI score”, often revealing hidden patterns when only generalized MI score rankings are considered across whole corpora. (Gablasova et al. 2017, p. 20). This observation could be highly pertinent to our research, indicating that the frequency and variety of high MI score word pairs we have observed might not only reflect the speakers’ language proficiency but also the specific vocabulary requirements of their chosen topics. Technical presentations, for instance, might necessitate the use of specialized vocabulary, thereby increasing the MI scores of the collocations used. Future studies could benefit from incorporating an analysis of the presentation topics, examining how the choice of topic influences the use of high MI score bigrams across proficiency levels.

At least five other potential limitations warrant consideration. First, the proficiency level of the participants was confined to intermediate- and higher-proficiency learners (IELTS bands 6.5 to 7.5), thus excluding lower-proficiency speakers and potentially limiting generalizability. Second, the study’s cross-sectional design limits the ability to trace language proficiency development over time or establish causality between variables. A longitudinal design could be useful in future studies, monitoring LB usage and proficiency over time. Third, the frequency-based approach using corpus analysis software has certain limitations, especially when dealing with spoken corpora. The current study used a minimum frequency of three and a minimum range of three to identify LBs in the learner corpus in order to keep the analysis manageable. These minimum frequency cutoffs mean that this study did not identify all multiword units in a comprehensive way. Some multiword units were used only once or twice or used idiosyncratically or in a nonstandard way by the L2 learners in our corpus. Many multiword sequences tend to blend into the linguistic context in transcripts, and many are frames or have larger fillable slots, which present real challenges for automatic extraction techniques. Such infrequent and/or semifixed units were not detected in the current study’s approach. The current study focused on bi- and trigrams using TAALES and three-to-five-word phrases using AntConc, possibly leaving out other multiword structures that may affect speaking proficiency. Future research could expand the scope of MWS analysis for more comprehensive insights. Fourth, the operationalization of LB usage through MI and frequency does not account for the qualitative aspects of contextual appropriateness in conversation. Lastly, since MI scores were primarily designed for two-word collocations, and since they do not consider the order of the words (Biber 2009; Hyland 2012), the MI scores might not reliably measure longer lexical strings. Future research should aim to address these limitations.

8. Pedagogical Implications and Suggestions

The findings support including bigram MI and longer LBs in assessing and teaching English proficiency, especially in speaking. They highlight the importance of focusing on the quality of LBs, particularly high MI bigrams, rather than just their quantity. This insight can inform instructional design to enhance learners’ speaking proficiency. Here are a few practical steps for implementing these findings in the classroom:

• Select relevant texts: Choose texts that align with students’ interests and proficiency levels.
• Identify high MI bigrams: Use Tom Cobb’s “Phrase Extractor” tool (https://www.lextutor.ca/multiwords/phrase/, accessed 3 February 2024) to extract high MI bigrams from the selected texts. These bigrams often serve as the foundation for longer LBs and expressions, thus making them a useful starting point.
• Practice and raise awareness:
  • Develop exercises: Create exercises that target high MI bigrams and longer LBs.
  • Highlight during activities: Have students notice and highlight high MI bigrams and longer LBs during reading and listening activities.
  • Assessment criteria: Include criteria related to use of high-quality LBs in speaking (and writing) grading rubrics.

By integrating and building on these steps, educators can effectively enhance students’ proficiency and awareness of high-quality LBs in English.

9. Conclusions

The current study has provided detailed insights into the patterns of LB usage that correspond to higher oral presentation scores. By extending beyond the scope of Tavakoli and Uchihara (2020), it has illuminated the intricate relationship between proficiency and both the frequency and complexity of LBs. Through extensive statistical analysis, it has established the link between linguistic sophistication and presentation scores, emphasizing the importance of both shorter and longer LBs in speaking proficiency. The current study’s findings make a persuasive case for the critical role of LB usage, particularly bigram MI, in predicting English language proficiency and presentation performance. These findings enhance our understanding of the relationship between lexical choice and speaking performance, thus offering practical insights for language assessment and teaching. The current research has highlighted bigram MI (and to some extent three-to-five-word usage) as a reliable indicator of English language scores on oral academic presentations. This is consistent with linguistic theories first put forward in the 1980s (e.g., Pawley and Syder 1983; Levelt 1989) that emphasize the efficient use of preformed chunks of language as a requirement of fluent speech and therefore a hallmark of proficiency. The implications for language teaching are considerable, suggesting that educators should include a focus on teaching strategies that improve learners’ awareness and command of high MI bigrams.