Syntactic complexity, which involves the range and sophistication of syntactic structures, is considered a key dimension of academic writing development and quality (Ortega, 2003; Lu, 2011). A quantitative metric that has garnered heightened attention in syntactic complexity research is DD. DD offers an index for evaluating the density or dispersion of grammatical connections throughout a text. Research suggests that dependency distance minimization (DDM) reflects a universal cognitive pressure for efficient human information processing and linguistic production (Futrell et al., 2015). English writers have been found to prefer syntactic structures with shorter dependencies to reduce integration difficulty and yield more locally coherent sentences (Temperley, 2007). However, cross-linguistic differences have also been observed, with head-final languages like Japanese, Korean, and Turkish showing greater distances attributable to word order variation (Futrell et al., 2015). Chinese and English show different dynamic valency of words and syntactic dependency structures (Lu et al., 2018). Liu et al. (2009) also found that Chinese shows quite different features in dependency relations, with its dependencies tending to be governor-final and mean dependency distance (MDD) being much higher than languages like English, German, and Japanese. While research has examined DDs in native language writing, fewer studies have investigated DDs in L2 academic writing.

Research consistently shows a strong tendency for compact, local syntactic structures in academic writing by L1 writers, which is argued to reflect pressures for efficient linguistic processing and production (Liu et al., 2017). An early study by Temperley (2007) analyzed DDs in the Wall Street Journal portion of the Penn Treebank. It is found that writers favor structures with shorter dependencies, which is evidenced by their preference for short left-branching constituents. Temperley argued that writers optimize and minimize dependency lengths to yield more incrementally interpretable sentences to facilitate comprehension. Futrell et al. (2015) also concluded that DDM is a universal characteristic across human languages, suggesting that variation in language can be explained by the general properties of human information processing. The authors argue that minimizing DDs enhances the efficiency of parsing and producing natural language, reducing integration costs and enabling more efficient packing of information into sentences. Lu and Liu (2020) also discovered a tendency of DDM within noun phrases, potentially due to limitations in human working memory capacity.

However, cross-linguistic differences have also been observed, attributable to syntactic variations across languages. Futrell et al. (2015) found that head-final languages like Japanese, Turkish, and Korean show much less DDM than head-initial languages like English, Italian, and Indonesian. Temperley and Gildea also found great differences across languages in DD. Their study confirms that DDM serves as an important factor in language structure and cognition, which is evidenced by the fact that writers and speakers tend to prefer structures that reduce dependency length when a language allows for different orderings of constituents (Temperley and Gildea, 2018). Nonetheless, Futrell et al. (2015) argue that DDM remains a universal quantitative property, as overall DDs were substantially shorter than random baselines (benchmarks created by randomly reorganizing the head word and its dependents in dependency trees, without following any specific linguistic word order rules), across all 37 diverse languages in their study. They contend that despite the structural variations among languages influencing their DDs, there is a universal aim in all languages to minimize DDs for the sake of efficiency, within the bounds of their structural limitations. The following example demonstrates the impact of syntactic variations on DD, with the specifics of the dependency relations delineated in Table 1.

The two sentences above convey the same concept and have identical structures, except for the positioning of the word “out.” However, their total and mean DDs significantly differ, at 14 vs. 20 and 1.55 vs. 2.22, respectively. The varying DDs require distinct cognitive effort and working memory capacity. Example 1b where “out” is moved to the end demands more cognitive effort and working memory. For lower cognitive effort and working memory load, Example 1a, which is free from particle movement and exemplifies DDM, is preferred.

In sum, research shows syntax is optimized for brevity both within and across languages to aid production and comprehension, but differences exist due to language-specific conventions.

While numerous studies have examined L1 DDs, few studies have investigated DDs in L2 academic writing. Ouyang et al. (2022) investigated the writing proficiency of beginner, intermediate, and advanced learners by DD measures. They discovered that the MDD overall is significantly effective at distinguishing between each pair of consecutive proficiency levels. Hao et al. (2022) verified the application of DD and its probability distribution as syntactic indicators of English as interlanguage from the perspective of language typology. They found that the MDDs of L2 learners with different backgrounds of native language gradually approach that of the target language with the improvement of their L2 proficiency. Li and Yan (2021) similarly found that MDD can serve as an effective indicator to measure the syntactic complexity of Japanese EFL learners’ interlanguage. Similarly, Hao et al. (2023) found in their study that dependency parameters have universal applicability in reflecting interlanguage proficiency.

To date, very few studies have directly compared the DD profiles of L2 writers from different L1 backgrounds composing in the L2. Gao and He (2023) examined the MDD of Ph.D. dissertation abstracts written by L1 (native English) and L2 (English as a foreign language) academic writers across language backgrounds and disciplines, finding that MDD successfully distinguishes between academic texts from various linguistic backgrounds and disciplines. They argued that the authors’ efforts to make comprehension easier for readers result in the shorter MDD observed in physics and chemistry abstracts.

Overall, research on L2 DDs remains limited, with very few studies comparing profiles of different L1 groups in natural academic writing tasks. In particular, few studies examined whether L2 DDs and dependency directions are influenced by L2 writers’ native language. This represents a significant gap, as investigating cross-linguistic differences can elucidate the role of L1 transfer versus universality in L2 syntactic development, with key theoretical and pedagogical implications (Ortega, 2003).

In examining the impact of native language transfer in L2 acquisition, scholars have extensively investigated how syntactic features influence L2 language production. Whong-Barr and Schwartz (2002) reveal that in the L2 acquisition process, children’s mastery of English dative constructions is significantly shaped by their native linguistic backgrounds, underscoring the influence of L1 syntactic frameworks and prevalent overgeneralization patterns on their learning trajectory. Chan (2004) demonstrates evidence of syntactic transfer from Chinese to English among Hong Kong Chinese ESL learners, revealing that learners often think in Chinese before writing in English, leading to interlanguage structures that closely resemble or mirror the syntactic patterns of their first language, particularly in complex target structures and among learners of lower proficiency levels. Recent advancements in second language acquisition research have introduced theories such as the Interpretability Hypothesis by Tsimpli and Dimitrakopoulou (2007), which argues that learners can acquire L2 features interpretable across syntax and other cognitive systems like semantics or pragmatics, regardless of their presence in L1; the Interface Hypothesis by Sorace and Filiaci (2006), highlighting the particular difficulties learners face with language elements that integrate syntax with semantics or discourse; and the Feature Reassembly Hypothesis by Lardiere (2009), emphasizing the primary challenge of reconfiguring L1 features to conform to the target language’s system, often leading to substantial learning challenges, especially where the languages’ feature systems notably diverge.

Though these new theories have been much discussed in recent years, language transfer theory remains relevant due to its powerful explanatory capabilities, able to account for many phenomena in L2 acquisition. This study aims to explore syntactic transfer through the lens of DG. We contend that the dependency patterns of an individual’s native language can influence those in their L2 writings. For instance, consider Chinese, which is predominantly a head-final language, and English, primarily a head-initial language. The MDD of Chinese stands at 3.662, markedly higher than English’s MDD of 2.543. We hypothesize that this significantly larger MDD in Chinese will lead to extended MDDs in L2 English writings by Chinese learners, as a consequence of language transfer. This study will verify our hypothesis.

In the present study, we intend to answer the following three questions:

Is there a significant difference in DD between English L2 academic writings and native academic writings?

The data for the present study was sourced from Scopus, selected for its status as the most extensive academic database and its established reliability as a data collection source across numerous studies (Crosthwaite et al., 2022, 2023; Zakaria and Aryadoust, 2023). Articles were chosen from the disciplines of Arts and Humanities and Social Sciences, limiting the selection to the “article” type while excluding “book review” and “book chapter” types, with an additional restriction for language to English only. In exporting the articles, we included the complete metadata for each article. Altogether over 2.65 million articles were extracted and stored in CSV files, with each metadata item allocated to a distinct column. Articles originating from various countries were segregated into separate CSV files, amounting to 178 files in total. The methodology for cleaning and processing the raw data is outlined in the following section.

We cleaned and processed the raw data using the following procedures. First, we wrote an R script to extract the “abstract” and “affiliation” columns. Second, we extracted the country names from the affiliation column, removed rows in the abstract column where no abstract was available using an R script, and manually checked the rows in the affiliation column where the country names were not available. Following the cleaning process, we obtained more than 2.22 million abstracts, totaling over 408.9 million tokens. Next, we calculated the dependency distance and direction for each abstract in each CSV file based on Eq. 2. This calculation was performed in Python, utilizing the Stanford CoreNLP package version 4.5.5. This package was selected due to its strong performance and established reliability as an NLP tool, as evidenced by various studies (Manning et al., 2014; Blšták and Rozinajová, 2022; Hashemi-Namin et al., 2023; He and Ang, 2023).

To categorize the dataset into L2 and native writings, we implemented a two-phase approach. In the first phase, we classified the abstracts according to the authors’ country of origin, identifying writings from the United Kingdom, United States, Australia, Canada, South Africa, New Zealand, Ireland, Bermuda, Jamaica, Trinidad and Tobago, Guyana, Barbados, and the Bahamas as native. In contrast, abstracts originating from any other country were designated as L2 writings. However, using affiliations to distinguish L2 from native writings is not entirely reliable, as L2 writers may study or work in countries where English is the primary language. To enhance the accuracy of classifying L2 and native writings, we introduced a second step involving a Python script that utilizes the nationalize.io API. This web-based service predicts nationalities from names using a vast database of names linked to their corresponding countries. We then compared the nationality predictions from nationalize.io with the initial phase’s results. In cases of discrepancies between the two sets of results, which were infrequent, we performed manual verification by consulting online sources to ascertain the authors’ nationalities. Through this dual-step approach, we significantly improved the precision of our classification between L2 and native writings. Despite this thorough double-check, there might still be a small number of cases where the classification was not accurate. However, given the vast size of our dataset, totaling over 2.22 million entries, these few discrepancies are unlikely to significantly impact our overall findings. Besides, other factors might affect the quality of L2 writing. The experiences of L2 writers, such as studying abroad or having their work edited by native speakers, could contribute to the subtleties of their writing. Nevertheless, we maintain that these factors do not substantially alter the fundamental linguistic characteristics of L2 writings. While there may be exceptional instances where they do, these are not expected to cause major deviations in our overall findings. Future research could consider incorporating these variables into their study designs to further enrich and complement our findings.

The table below presents the overall descriptive statistics of the data used in this study.

The overall descriptive statistics show that L2 writings have longer MDDs and higher percentages of governor-final dependencies (DDI) than native writings for both datasets. As the size of the datasets reached beyond the limit of Shapiro–Wilk and Student’s t-test, both of which require a sample size below 5,000, we used the Anderson-Darling normality test and Kolmogorov–Smirnov test to compare the native writings and L2 writings. The results show that native writings and L2 writings are significantly different in both MDD and DDI for both datasets (p < 0.0001), indicating that L2 writings have significantly longer MDDs and higher percentages of governor-final dependencies than native writings.

Based on the descriptive statistics, we can offer a positive answer to our research questions 1 and 2 regarding whether there is a significant difference in dependency distance and direction between English L1 and L2 academic writings. There is a significant difference in the MDD and DDI of the two groups. Yet, it is not sure whether the significant difference is influenced by native language transfer. In the next section, we will discuss it based on more detailed results.

In our datasets, the sample size of some countries is very small (for example, Gambia and Guinea). Thus, we only selected those countries with a sample size of 500 and above. Figures 2, 3 report the dependency distances of the selected countries in the Arts and Humanities group and the Social Sciences group, respectively. (Detailed reports of the MDD and DDI are available upon request).

In Table 3, MDD(sd) is the standard deviation of MDD, DDI is the mean ratio of dependency relations where governors are preceded by their dependents, in other words, DDI is mean governor-final ratio, and DDI(sd) is the standard deviation of DDI. The results in both sub-datasets show that the MDDs of both English native and L2 academic writings are much longer than that of the MDD of English (2.543) according to Liu (2008). Liu’s calculation of the MDD of English is based on news texts. Our study examines the MDD of academic texts. News texts and academic texts are different genres showing different linguistic features. Their differences in MDD show that genre is a factor that affects MDD, which is partially in line with the findings of Wang and Liu (2017). According to their study, genre affects dependency distance and direction significantly, but the effect is very small. They hold that “dependency distance is primarily determined by universal cognitive factors rather than genre-specific stylistic factors” (Wang and Liu, 2017: 135). Yet, in our study, we find that English native academic writings have an MDD of 2.9, much larger than the MDD of English news texts, which is 2.543 (Liu, 2008). In Wang and Liu’s (2017) study, the ratio of dependency relations where governors precede their dependents is between 46 and 51%, while in our study, this ratio is around 33% as the ratio of governors following dependents is around 68%. This again shows that the genre of English academic writings has a much higher ratio of governor-final dependencies. Such a finding indicates that genre has a significant influence on MDD, at least in terms of the genre of academic writing. Yet, as we only examined one genre, it is not safe to claim that genre has a large effect on its influence over DD, which needs further investigation with samples from different genres.

We proceeded to make Mann–Whitney U tests between native writings and L2 writings with different language backgrounds for the two sub-datasets. Mann Whitney U test is chosen because the native group has a very large sample size and is not in a normal distribution. As the results of Mann–Whitney U tests include many pairs of comparison, which takes up much space, they are not reported here and are available upon request. The results reveal that English native academic writings are significantly different in both MDD and DDI from English L2 academic writings of different language backgrounds for both sub-disciplines, as the p values are all below the significance level (0.05), with large effect sizes (R > =0.8) for most pairs. This finding further confirms the result reported previously in Table 3 where English native academic writings are found to be significantly different from English L2 academic writings on the whole.

As the Mann–Whitney U test examines whether two samples come from the same population, but does not reveal the correlation between variables, we did correlation analyses to find whether the differences in MDD and DDI are related to the nature of the samples, that is, native academic writings or L2 academic writings, to explore whether the language backgrounds of English L2 academic writings affect their dependency distances and dependency directions. We made a binomial logistic regression in R by the basic function glm with the two levels of the abstract type, Native vs. L2 as the response variable and MDD and DDI as predictor variables. Besides, we did another two analyses, linear regression analysis in R by the basic function lm and correlation analysis in R by the basic function cor.test. The three analyses are made for mutual corroboration.

The results of the binomial logistic regression in Table 4 show a significant correlation between article type and MDD and DDI for both sub-datasets, with p values below the significance level (0.05). Since the article type is a binary categorical variable, native vs. L2, the strong correlation indicates that whether the article type is native or L2 has a significant influence over MDD and DDI.

The results of linear regression in Table 5 and correlation analyses in Table 6 both confirm the correlation between article type and MDD and DDI, with p values much lower than the significance level (0.05). As the three correlation-related analyses all show a significant influence of article type on MDD and DDI, it could be claimed that on the whole there is an effect of native language transfer on the MDD and DDI of English L2 academic writing. Currently, no study has been found to examine the MDD and DDI of all languages in the world due to various reasons, though some studies have investigated the DDs of some languages, such as Chen and Gerdes (2020), Jing and Liu (2017), Futrell et al. (2015) and Liu (2008). However, examining studies such as Liu (2008) reveals a trend: the greater the MDD in the background language of English L2 academic writings, the longer the MDD tends to be in English L2 academic writings themselves. For example, in the data of Social Sciences, L2 writings with Chinese as their background language have a much higher DD than the English native ones, with their MDD being 2.9896 vs. 2.8950, compared to the MDD of original Chinese and English, which is 3.662 vs. 2.543 (Liu, 2008). It is the same, for instance, with Hungarian, German, and Spanish, 2.9500 vs. 2.8950 and 3.446 vs. 2.543, 2.9464 vs. 2.8950 and 3.353 vs. 2.543, 2.9400 vs. 2.8950 and 2.665 vs. 2.543.

To further confirm or examine the effect of native language transfer on the MDD and DDI of English L2 academic writing, we categorized the language backgrounds into several groups based on the classification of language families by Katzner (2002) and Brown and Ogilvie (2009). Then, we calculated the MDD and DDI of the English L2 academic writings of different language background groups and made Mann–Whitney U test.

Tables 7, 8 report the MDD and DDI of the samples grouped by language family, which are accompanied by Figures 4, 5 for better visualization. The results show that English native academic writings are significantly different in MDD and DDI from L2 academic writings with different language family backgrounds, which can be confirmed by the results of Mann–Whitney U test reported in Table 9, as the p values of the comparison of most pairs between English native academic writings and L2 academic writings are below significance level (0.05). One exception is the pair of Indo-European_Native vs. Pidgin-Creole in Arts & Humanities (p = 0.0503), showing no significant difference. This insignificance arises probably because pidgins and creoles are hybrid languages formed by the blending of different languages. For example, a pidgin language is one with vocabulary “of English, French, Spanish, or Portuguese origin” (Katzner, 2002: 32). As a result, it may share syntactical and lexical features with English, which in turn can influence the pidgin speakers’ English L2 academic writing. On the whole, a significant difference exists in MDD and DDI between English native academic writings and English L2 academic writings.

Similar results in Figures 6, 7 are found between English native academic writings and English L2 academic writings grouped by language family group. (Detailed reports of the MDD and DDI are not presented here and are available upon request as they take much space). A significant difference arises between English native and L2 academic writings, which is confirmed by the significant p values, which is below 0.05. (Detailed reports of the Mann–Whitney U test are available upon request to save space here).

We also grouped the English L2 academic writings according to whether English is regarded as an official language of the countries where these L2 writings are from. As Table 10 shows, English L2 academic writings with a background of non-English official languages have a longer MDD and higher ratio of DDI, with a significant difference (p < 0.05) from the English native academic writings as Table 11 shows. For those L2 writings with English as the official language of their countries, no significant difference (p = 0.5798) from English native academic writings is found for samples from Social Sciences, though a significant difference is found for those from the Arts and Humanities.