The use of metadiscourse by secondary-level Chinese learners of English in examination scripts: insights from a corpus-based study

2.1 Conceptualisation of metadiscourse

Metadiscourse has been defined and classified in various ways. Many of the early taxonomies drew on Halliday and Matthiessen’s (2013) metafunctions of language and categorised metadiscourse as either textual or interpersonal, depending on the function it performs in a text (see Crismore et al. 1993; Hyland 1998; Vande Kopple 1985). Textual metadiscourse refers to elements that help organise the discourse and create cohesion among propositional content, while interpersonal items mainly serve to convey authorial attitudes. A few scholars (e.g., Ädel 2006; Mauranen 2010) consider some interpersonal elements (e.g., attitude and stance markers) to be non-metadiscursive devices that refer to text-external rather than text-internal phenomena. Their approach to metadiscourse focuses on analyzing aspects of the text and writer-reader interaction. However, concern has grown over the intrinsic shortcomings of such a bidirectional categorisation of metadiscourse as either textual or interpersonal. Functional overlaps in academic writing make it difficult to label specific metadiscourse markers that fall into a grey area between the two categories (Ädel 2006; Hyland 2019). Consequently, Hyland (2019) proposed a revised taxonomy and classified metadiscourse into two categories of interpersonal resources—interactive and interactional—offering a means to understand “the interpersonal resources writers use to organise texts coherently and to convey their personality, credibility, reader sensitivity and relationship to the message” (p. 71). This model assesses the construction of a cohesive, consistent, and persuasive argument using metadiscourse for readers in an academic context and better facilitates the exploration of metadiscourse in discourse analysis and academic writing.

According to Hyland, interactive resources (e.g., transitions, frame markers, endophoric markers, evidentials, and code glosses) enable writers to manage the flow of information to convey their preferred interpretations explicitly. Interactional resources (e.g. hedges, boosters, attitude markers, engagement markers, and self-mentions) highlight the participants of writer-reader interactions and aim to display the writer’s persona and a tenor compatible with the practices of disciplinary communities. Metadiscourse is therefore a “recipient design filter” facilitating the reception of an intended message by offering a running commentary (Hyland 2017, p. 17). It demonstrates the writer’s awareness of their readers and the necessity to clarify, elaborate, interact with, and guide them through the use of language. Management of these resources enables the writer to convey their affective position towards the content and reader, build writer-reader rapport, and eventually construct a text that is considered persuasive or successful (Lee and Deakin 2016).

2.2 Factors affecting the use of metadiscourse in L2 writing

Many studies have focused on the differences in the use of interpersonal metadiscourse markers amongst L2 learners of English by analyzing written learner corpora. This line of investigation has assessed the effects of various learner external and internal factors and revealed that metadiscourse use varies considerably across academic levels (Ruan 2019), disciplines (Hyland 2004, 2010), L2 proficiency (Ho and Li 2018; Lee and Deakin 2016), learning contexts (Li and Wharton 2012), and lingua-cultural backgrounds (Lee and Casal 2014), as well as between students and academics (Aull and Lancaster 2014) and between English L1 and L2 student writers (Ädel 2006; Leedham and Cai 2013).

Along with these factors, use of metadiscourse also varies depending on the genre of the texts. According to Bhatia (2004), genres are largely shaped by the communicative purpose and the social context of the discourse, and these in turn determine the schematic structure of the text and influence the choice of language, register, and style, including the use of metadiscoursal resources (Hyland 2019; Swales 1990). Studies have examined the role of metadiscourse in a variety of genres and how its use is mediated by different aspects of the communicative events. For example, Yoon (2020), examining the use of interactional metadiscourse by Chinese, Japanese, and Korean learners of English in two timed argumentative essays on different topics, found that the writing topic influenced the use of metadiscourse more than proficiency level. Further, Kuteeva (2011), who examined argumentative texts produced by L2 learners of English on wikis, revealed that the writer’s audience awareness was a factor in using L2 interpersonal metadiscourse markers and suggested that improving students’ awareness of the audience could produce a more reader-oriented text. Similarly, Bogdanović and Mirović (2018) found a positive relationship between the level of audience awareness and the use of interpersonal metadiscourse in research articles written by advanced L2 learners of English.

While researchers have tended to focus on written academic genres such as undergraduate and postgraduate dissertations (Hyland 2004; Hyland and Tse 2004), university argumentative essays (Ho and Li 2018; Ruan 2019; Wu 2007), and research articles (Dahl 2004; Keshavarz and Kheirieh 2012), there has been a growing body of research into metadiscourse use in other genres, such as business letters and reports (Crismore 2004; Huang and Rose 2018; Hyland 1998), job advertisements (Fu 2012), instruction manuals (Herriman 2022), magazines (Fu and Hyland 2014), newspaper (Dafouz-Milne 2008), and workplace emails (Ho 2018), since Hyland’s (1998) seminal work. By comparing CEOs’ letters and directors’ reports, Hyland (1998) found that while both genres used textual metadiscourse, CEOs’ letters contained considerably more interpersonal devices, particularly hedges, boosters and attitude markers, than directors’ reports. Hyland attributed such a marked contrast to the different communicative purposes of the two genres, suggesting that while directors’ reports laid out an objective assessment of the companies’ performance, CEOs’ letters served to communicate achievements and promote a positive corporate image. The latter thus involved using more interactional resources to influence and build trust with readers or potential investors. The differences in metadiscourse use between letters and reports are also apparent when comparing the findings of Alyousef (2015) and Crismore (2004). In analysing management reports written by postgraduate accounting students, Alyousef (2015) noted that the students appeared to be aware of the conventions of academic report writing and made considerable use of interactive metadiscourse, especially transitions, code glosses, and frame markers, to organise and present key findings and arguments in a cohesive and coherent manner. In contrast, Crismore (2004) examined fundraising letters and found that the writers used more interpersonal metadiscourse, especially personal pronouns, hedges, and attitude markers, to persuade readers to donate to their organisations.

While studies have shown that metadiscourse is genre-specific, few have factored text genre into the analysis of metadiscourse use or compared metadiscourse use in L2 secondary school students’ writing in different genres. Among those few, Hong and Cao (2014) compared the use of interactional metadiscourse in descriptive and argumentative essays written in English by young Chinese, Polish, and Spanish students from Grades 3 to 12. They found that the argumentative essays contained a significantly greater number of hedges and self-mentions than the descriptive essays, while essay topics and task prompts also impacted the use of self-mentions. In another recent study, Qin and Uccelli (2019), who compared the use of metadiscourse markers in academic and colloquial texts written by Chinese, French, and Spanish students at different proficiency and academic levels, found that boosters, engagement markers, and self-mentions appeared more frequently in colloquial texts, while code glosses appeared more often in academic texts.

Researchers have also compared the use of metadiscourse between high- and low-rated L2 learners’ essays to uncover the relationship between metadiscourse use and writing quality. In general, these studies have found that high-rated L2 writers tended to use more and draw on a greater variety of interactional metadiscourse markers to construct their voice and engage with the readers in their writing (Ho and Li 2018; Intaraprawat and Steffensen 1995; Lee and Deakin 2016; Wu 2007). For example, Intaraprawat and Steffensen (1995) reported that, in an early investigation of metadiscourse use in L2 writing, high-rated examination scripts exhibited a greater diversity of metadiscourse markers overall and more frequent use of interactional metadiscourse markers than the low-rated scripts. They noted that stronger writers tended to display a higher level of audience awareness and a greater ability to guide and engage with their intended readers through the use of metadiscourse in their writing, which contributed to the positive assessment of their essays. More recently, Ho and Li (2018) observed that Chinese first-year undergraduate students in Hong Kong used more interactional than interactive metadiscourse markers in timed argumentative essays, with the high-rated scripts containing significantly more hedges and attitude markers, but fewer engagement markers than the low-rated scripts. Comparing the use of interpersonal metadiscourse in successful (A-graded) and less successful (B-graded) argumentative essays written by Chinese L2 university students and their L1 peers, Lee and Deakin (2016) also found that the high-rated essays contained significantly more instances of hedges than the low-rated ones; however, the instances of attitude markers, boosters, and engagement markers observed in both sets of essays did not differ significantly.

2.3 Metadiscourse in pre-tertiary L2 writing

Research on metadiscourse in L2 writing has focused on texts produced by university students, with only a few studies examining the use of metadiscourse in pre-tertiary L2 writing. Among those, Field and Yip (1992) examined argumentative essays written by secondary-level English and Cantonese speakers in Australia and Hong Kong. They found that the ESL learners employed far more connective devices that functioned as transitions (e.g. moreover, however) and frame markers (e.g. first, secondly, thirdly) compared to their L1 English-speaking counterparts. In a later study, Hyland and Milton (1997) investigated how epistemic stance was expressed in timed examination scripts written by high school leavers in Hong Kong and the UK. They observed that ESL learners in Hong Kong relied on a narrower range of stance markers and tended to express stronger commitments, than the UK students. Further, they found that the use of epistemic devices among the learners varied across proficiency levels, with writers of high-rated scripts incorporating more hedging devices in their writing to moderate their claims, while writers of low-rated scripts employed far more boosters and first-person pronouns to express their stance. L2 learners’ tendency to rely on a limited range of interactional devices and employ markers common in informal and spoken registers was noted in Hong and Cao’s (2014) analysis examining how English learners from Grades 3 to 12 used various interactional metadiscourse devices in writing to express their stance and interact with their intended audience. Specifically, most hedges were realised through a small number of epistemic modal auxiliaries (e.g., would, could) and adverbs (e.g., maybe), and the use of boosters was limited to a few intensifying adverbs (e.g., really, very) often occurring in the spoken register. In addition, most self-mentions and engagement markers used by the learners were first-person pronouns (e.g. I, my) and second-person pronouns (e.g. you, your), respectively, reflecting a high level of interpersonal involvement, albeit with a relatively conversational and informal style of writing.

4.1 Context

This study used the written texts selected from the Hong Kong Diploma of Secondary Education (HKDSE) examination. First administered in 2012, the HKDSE examination is a public examination measuring the academic competence of local school students (aged 17–18) upon completing six years of secondary school. English is one of four compulsory subjects in the HKDSE examination and comprises four papers: reading, writing, listening and integrated skills, and speaking. Writing has two parts: Part A requires students to write approximately 200 words on one question, and Part B requires students to write about 400 words on one of eight prescribed themes (sports communication, drama, songs and poems, debate, short stories, popular culture, workplace communication, and social issues). Students’ ability to write texts using appropriate tone, style, register and the salient features of different genres is assessed. In Part B, for example, they may be asked to write a letter to the editor or a report that offers suggestions or advice based on facts corresponding with the prescribed theme of social issues. The Hong Kong Examination and Assessment Authority (HKEAA) prepares the rubrics for marking the examination scripts and organising the moderation sessions. The HKDSE examination adopts standard-referenced reporting of assessment results, and examinees’ achievement falls within one of seven levels: 1 (unclassified), 2, 3, 4, 5, 5*, and 5** (highest). Level 2 is equivalent to the overall IELTS band score of 4.79–5.07, whilst Level 5** corresponds to 7.51–7.77.

4.2 Corpus construction

From a collection of 120 essays produced in the HKDSE examination (2017 and 2018), we built a corpus of 63,485 words. While the essays were written by Hong-Kong-based Chinese ESL students, the demographic and educational background of the student writers was not disclosed by the HKEAA due to confidentiality. According to the HKEAA reports from 2012 to 2018, “learning English through social issues” was the most popular theme in five of the six years, with an average of 28.56% of examinees choosing this theme. With assistance from the HKEAA, we chose 120 digital images of marked examination scripts – 60 from the 2017 cohort and 60 from the 2018 cohort – on this theme from the HKDSE English Language Paper 2 (2017 and 2018). As Figure 1 depicts, the 2017 and 2018 questions included a scenario presenting a social issue, followed by an invitation to provide suggestions and advice. In 2017, examinees were asked to write a report of suggestions/advice and in 2018, a letter of suggestions/advice.

Figure 1:

Examination prompts in 2017 and 2018.

Two trained assessors double-marked each written examination response according to a norm-referenced scale, later converted into numerical grades by the HKEAA, ranging from 0 to 42. The 120 examination scripts were divided according to three proficiency levels, as detailed in Table 1. Among them were low-, medium- and high-rated (20 in each year). The low- and high-rated examination scripts were written by the bottom and top 5% of the candidates. The medium-rated scripts averaged a score of 22.5 marks (out of 42), which was the mean score of 2017 and 2018 (HKEAA 2017, 2018).

4.3 Classification and annotation of metadiscourse

In the present study, we adopted Hyland’s (2019) classification scheme as the starting point for identifying and analysing the metadiscourse markers used in the examination texts. Among the different taxonomies (e.g., Ädel 2006; Crismore et al. 1993; Hyland 2019), Hyland’s classification was considered the most suitable because it is arguably the most comprehensive and detailed in the existing literature (Bax et al. 2019). It has also been widely used to study metadiscoursal practices in a variety of genres and communicative contexts (Herriman 2022). In Hyland’s model, metadiscourse comprises two interpersonal dimensions – interactive and interactional – with each dimension containing five categories of metadiscoursal markers. As revealed below, almost all the frame markers identified in our corpus were sequencing markers. To avoid data distortion and misrepresentation of a metadiscoursal category (Bax et al. 2019), we decided to treat the four sub-types of frame markers as individual categories for the purpose of analysis. Accordingly, interactive metadiscourse in our analytical framework included transitions, sequencing, topic shifts, label stages, announce goals, endophoric markers, evidentials, and code glosses, while interactional metadiscourse contained hedges, boosters, attitude markers, engagement markers, and self-mentions (see Table 2).

To annotate metadiscoursal markers in our corpus, two research assistants first transcribed and cross-checked the 120 digital images. Each transcript was segmented into clauses and then coded and analysed using the Child Language Data Exchange System (CHILDES) transcription conventions. Next, the Computerized Language Analysis (CLAN) programme (MacWhinney 2000) was used to tag all the items featured in the list of metadiscoursal devices provided by Hyland (2019). As Hyland (2019) stressed, metadiscourse is context-sensitive; items that perform metadiscoursally in one context may not do so in others. For example, the word ‘around’ may function as a hedge (around six years), but it can be non-metadiscursive in some contexts referring to text-external relations (around the world). Furthermore, metadiscourse features can be multi-functional (Ädel 2006; Hyland 2017). The first person pronoun ‘we’, for example, can serve as self-mention (we have received your letter) or as an engagement marker (we’re all in this together). Considering these differences, each instance of all the potential metadiscoursal items identified was manually checked in its sentential context to ensure that all the non-metadiscursive elements and extraneous items were excluded from the analysis. Issues and criteria regarding identification of metadiscourse markers were discussed and established before the coding process started. The two research assistants coded the essays independently and cross-checked the data analysis to ensure inter-rater reliability, with the percentage agreement reaching 95%. Disagreements between the two coders regarding the coding were resolved through discussions by the researchers afterwards.

4.4 Analysis

To determine the contribution of genre (letters vs. reports of advice) and grade awarded (low, medium, high) towards variations in using interactional and interactive metadiscourse markers, the frequencies of each metadiscourse marker were normalised to n per 1,000 tokens, ensuring fair cross-corpus comparison.

The inferential statistical measures used on the data followed a two-step process for regression to inferentially confirm the significance and effect size of the latent constructs underlying the observed variables (Norouzian and Plonsky 2018). Partykit package in R was used to create conditional inference trees (CITs) (Tagliamonte and Baayen 2012). CITs, like other multivariate tree-based methods, recursively partition the data into two sections to maximise prediction accuracy, making them a versatile multivariate method with easily interpretable output (Baayen et al. 2013). CITs, unlike traditional classification and regression trees (CARTs), use significance tests to establish whether a particular split is warranted (Gries 2020). This technique decreases the need for pruning (Hothorn et al. 2006), while variables with more potential splitting points are not artificially preferred (Boulesteix et al. 2015). CITs are thus used to discover correlations between the predictors and dependent variables, which is important for reducing irrelevant variables in subsequent regression analyses.

Next, the glmulti regression package for easy automated model selection was used to confirm the CITs’ findings (Calcagno and de Mazancourt 2010). Glmulti builds all possible unique models involving all input variables and their pairwise interactions. Models are fitted with standard R functions, e.g., glm. The n best models are returned allowing model selection and multi-model inference through standard R functions. To ensure the regression models met required assumptions, a Durbin-Watson test was conducted on each regression analysis to determine the potential for autocorrelation (also called serial correlation) in residuals. Each test statistic was approximately 2.0, with test statistic values in the range of 1.5–2.5 considered relatively normal while values under 1 or more than 3 were a cause for concern (Field 2009). An R Notebook containing the procedure for both the CITs and subsequent regression for the first analysis on interactive metadiscourse can be downloaded at https://tinyurl.com/bp7dhf7z.

Finally, we compared the specific wordings of metadiscourse features through log-likelihood, using the Log-Likelihood Calculator (Rayson and Garside 2000).

5.1 Interactive metadiscourse by genre and grade

Looking first at the relationships among interactive metadiscourse, genre and grade, a CIT was generated including the combined normalised frequencies of all interactive metadiscourse markers as the dependent continuous variables, and grade (low, medium, high) and genre (letters vs. reports) as independent fixed-effect variables (Figure 2).

Figure 2:

CIT with impact of genre and grade on use of interactive metadiscourse markers.

The CIT revealed a significant interaction of genre (reports vs. letters, p < 0.001) and grade on the use of interactive metadiscourse, with both low-rated reports (p < 0.001) and letters (p = 0.026) containing a higher normalised frequency of such features.

Next, to confirm the effect size of this relationship, the glmulti R package was used for regression analysis with the frequency of interactive metadiscourse markers as the dependent variable and genre and grade (and the interactions of these at each level) as independent fixed-effect variables. One model out of five potential models conducted in glmulti was within 2 IC units and was used as the final model, with an AICC value of 841.49 and an R ² value of 0.426, denoting a medium effect size (Table 4).

The model suggests that letters were significantly less likely to contain interactive metadiscourse markers, while examination scripts with high grades were also significantly less likely to contain such markers. There was also a significant interaction between letters and medium-rated examination scripts, with these being less likely to contain a high frequency of interactive metadiscourse.

5.2 Interactional metadiscourse by genre and grade

Regarding the relationships among genre, grade, and the use interactional metadiscourse markers, another CIT was generated including the combined normalised frequencies of all interactional metadiscourse markers as the dependent continuous variable, and grade (low, medium, high) and genre (letters vs. reports) as independent fixed-effect variables (Figure 3).

Figure 3:

CIT with impact of genre and grade on use of interactional metadiscourse markers.

The CIT revealed a significant interaction of genre (reports vs. letters, p < 0.001) and grade, with low-rated reports containing a higher normalised frequency of interactional metadiscourse markers (p = 0.001) and low-rated letters also containing a higher frequency of such features (p = 0.003).

To confirm the relationship revealed in the CIT, regression analysis was conducted with the frequency of interactional metadiscourse markers as the dependent continuous variable, and genre and grade as the independent fixed-effect variables. Two models out of a potential five conducted in glmulti were within 2 IC units, with one reaching 95% of evidence weight, and this one was used as the final model. The AICC value was 1,173.12, with an R ² value of 0.845, denoting a large effect size (Table 5).

The model revealed that letters were significantly more likely to contain interactional metadiscourse markers, while examination scripts with medium and high grades were significantly less likely to contain such markers.

5.3 Specific interactive metadiscourse categories by genre

We now turn to the relationship between genre, and interactive/interactional metadiscourse, to answer RQ1. Focusing first on the relationship between the use of specific interactive metadiscourse categories and genre, a CIT was generated, this time with genre as the dependent dichotomous variable and the normed frequencies of each interactive metadiscourse marker as the independent continuous variables (Figure 4). This process helped us determine which interactive metadiscourse categories should be included in the subsequent regression analysis.

Figure 4:

CIT with the impact of individual interactive metadiscoursal categories by genre.

Following correction for multiple tests, the CIT revealed that sequencing markers are a significant source of variation across genres (p < 0.001), with reports appearing to contain a higher frequency of such markers. There was also an interaction between the use of sequencing markers together with code glosses (p = 0.012). Using glmulti, we sought to confirm this relationship, this time with sequencing markers as the dependent continuous variable, and genre and code glosses as the independent variables; the code glosses were included to determine the significance of any interaction between their use and that of sequencing markers. Three models out of a potential six were within two IC units, with two models reaching 95% of evidence weight. The final preferred model had an AICC value of 621.31 with an R ² value of 0.181, for a small effect size (Table 6).

The model revealed that it was highly likely for reports to contain a high frequency of sequencing markers (p < 0.001), with no significant interactions reported for code glosses. Table 7 shows the forms of sequencing markers used across genres.

There did not appear to be significant differences in the specific wordings of sequencing markers across genres, except for ‘second’ (LL = 6.34, p < 0.050) and ‘secondly’ (LL = 13.27, p < 0.001), which was more prominent in reports (Examples 1, 2).

This use of sequencing terms, i.e., “second,” is an artifact of typical classroom instructions in report writing; students are often told to choose two to three supporting arguments and list them sequentially through numbering.

5.4 Specific interactional metadiscourse features by genre

An additional CIT focused on the relationship between the use of individual interactional metadiscourse categories and genres (Figure 5).

Figure 5:

CIT with impact of individual interactional metadiscoursal categories by genre.

Following corrections for multiple tests, the CIT indicated that the use of engagement markers varied significantly across the two genres (p < 0.001), with reports appearing to contain a lower frequency than letters. There was also an interaction between the use of engagement markers and self-mentions (p < 0.001), with students using pronouns for both (with inclusive/exclusive forms constituting the different categories).

Using glmulti to confirm this relationship with engagement markers as the dependent continuous variable and genre and self-mentions as independent variables, two models from six were within two IC units, and two within 95% of evidence weight. The final preferred model had an AICC value of 1,104.15, with an R ² value of 0.810, indicating a very large effect size (Table 8).

The model suggests reports were significantly less likely to contain engagement markers (p < 0.001), although there was also a significant interaction between the use of engagement markers and self-mentions specific to reports (p = 0.034). Table 9 shows the engagement markers used by genre.

Terms that were significantly more prominent in reports than letters include the use of ‘we’ (LL = 36.06, p < 0.0001) and ‘our’ (LL = 12.44, p < 0.001), given the difference in audience specification between the two tasks (Examples 3–6).

We

Our

Terms that were more prominent in letters include ‘you’ (LL = 2,159.19), ‘your’ (LL = 1,335.54), ‘should’ (LL = 164.14) and ‘must’ (LL = 26.52) at the p < 0.0001 level, with ‘you’ and ‘your’ representing a different audience specification, while ‘should’ and ‘must’ were used as recommendations and imperatives to the audience (Examples 7–13).

You

Your

Should

Must

5.5 Specific interactive metadiscourse categories by grade

We now turn to the relationships among interactive and interactional metadiscourse and grades to answer RQ2. Focusing first on the relationship between interactive metadiscourse categories and grade, a CIT was generated including grade as the dependent variable, and the normed frequencies of all individual interactive metadiscourse categories as independent variables (Figure 6) to determine which interactive metadiscourse categories should be included in the subsequent regression analysis.

Figure 6:

CIT using individual interactive metadiscourse categories by grade.

Following correction for multiple tests, the CIT indicated that the use of sequencing markers varied significantly across essays of different grades (p = 0.007) with low and medium-rated texts appearing to have a higher frequency of such markers. Using glmulti to confirm this relationship, this time with sequencing markers as the continuous dependent variable, one of the two models was within two IC units, so it was used as the final model. The AICC value was 627.83 with an R ² value of 0.119, denoting a small effect size (Table 10).

The model revealed that high-rated texts were significantly less likely to have sequencing markers (p = 0.001). Table 11 shows the sequencing markers used by grade.

Regarding specific terms, the use of ‘first’ (LL = 10.22, p < 0.010), ‘finally’ (LL = 8.73, p < 0.010), and ‘next’ (LL = 16.90, p < 0.0001) were more frequent in low-rated texts because these forms are a mechanical method of structuring a complete text (Examples 14–19).

First

Finally

Next

5.6 Specific interactional metadiscourse categories by grade

Next, a CIT was generated to determine the relationship between the use of interactional metadiscourse categories and the grade awarded to a text, including grade as the dependent variable and the normed frequencies of all individual interactional metadiscourse markers as independent variables (Figure 7).

Figure 7:

CIT with use of individual interactional metadiscourse categories by grade.

Following correction for multiple tests, the CIT indicated that the use of boosters, but not other interactional markers (e.g., hedges, self-mentions), varied significantly within texts of different grades (p < 0.001), with low-rated texts appearing to contain a high frequency of such markers. Using glmulti to confirm this relationship, this time with booster use as the continuous dependent variable and grade as the independent fixed-effect variable, one of two models was within two IC units, so it was used as the final model. The AICC value was 856.28 with an R ² value of 0.193, denoting a small effect size (Table 12).

The model revealed that medium- and high-rated texts were unlikely to contain many boosters (p < 0.001). Table 13 shows the booster forms used across low, medium, and high-rated texts.

The boosters ‘think’ (LL = 104.70, p < 0.0001), ‘always’ (LL = 42.40, p < 0.0001) and ‘really’ (LL = 13.33, p < 0.001) were prominent in low-rated texts. ‘Think’ was commonly used with the self-mention ‘I’ to convey the author’s stance or to assume the stance of others (Examples 20, 21), while ‘always’ was regularly used to discuss habits or imperatives (Examples 22, 23). ‘Really’ was primarily used in low-rated texts as an adjective modifier.

Think

Always

Really

On the other hand, the use of ‘in fact’ (LL = 6.37, p < 0.050) was associated with high-rated texts, both in the sentence-initial position and the post-verbal position (Examples 26, 27).

In fact

6.1 RQ1. Use of metadiscourse in different genres under timed written examination conditions

Regarding RQ1, we found that genre significantly affects the choice of metadiscourse in secondary-level Chinese learners’ writing in English. In line with previous research on the metadiscoursal practices employed in letters (Crismore 2004) and reports (Alyousef 2015), letters of advice contained fewer interactive devices than those seen in reports while containing more interactional devices. This implies that the writers were sensitive to genre expectations when composing reports; in effect, they were more concerned about guiding the readers through the text than interacting with them, emphasising the structure and logical development of their arguments (see Alyousef 2015). When writing letters, they emphasised communicating ideas and engaging with their readers (see Crismore 2004; Hyland 1998). Our analysis reveals that only sequencing and engagement markers significantly varied across the two genres, with the former identified in reports and the latter in letters. Therefore, it appears that pre-tertiary learners modify their use of metadiscourse according to the requirements of the writing prompts and the communicative contexts of individual tasks (Hong and Cao 2014).

Since the prompt for report writing required explaining and suggesting solutions to the rising number of NEETs in Hong Kong, the examinees had to explicitly signal transitions, label text stages, indicate topic shifts, and present their arguments in an orderly manner with sequencing markers and other interactive resources, and this was reflected in the high frequency of such features observed in the reports. In contrast, the letter prompt required examinees to give personal advice, so writers had to directly address the reader by incorporating engagement markers—namely, second-person pronouns (you, your) and modal verbs (should, must)—which included the reader as a discourse participant (Hyland 2019). However, concerning the use of engagement markers, the inclusive ‘we’ and ‘our’ were frequently used in reports, signalling a potential mismatch between the genre expectations for authorial engagement and intrusion within report writing, and this may have resulted in lower grades for texts where this use of metadiscourse was prominent.

6.2 RQ2. Metadiscourse markers and exam writing grades

Regarding RQ2, the frequency of metadiscourse use was a significant indicator of L2 writing quality. However, our finding that higher-rated scripts were significantly less likely to contain interactive and interactional metadiscourse markers than lower-rated scripts contradicts previous studies (cf. Intaraprawat and Steffensen 1995; Lee and Deakin 2016), which have suggested the opposite, although these studies were in different contexts.

Furthermore, our data revealed that the mechanical use of certain subtypes of metadiscourse was associated with low-rated scripts. This correlation was particularly notable in the use of sequencing markers and boosters. Specifically, the frequent occurrence of sequencing markers was found to be a negative predictor of writing quality, with high-rated scripts having fewer instances of such markers, especially those indicating chronological order (e.g., first, second, next and finally). This may be because examinees often received a higher grade if they demonstrated the ability to employ different types of cohesive devices or syntactic patterns to structure their writing in a sophisticated manner. However, those who relied heavily on sequencing markers may have been penalised for the excessive and mechanical use of such devices, which may have made transitions between sentences or paragraphs appear artificial, adversely affecting the readability, comprehensibility, and overall writing quality. Despite the importance of metadiscourse in writing, extensive use of such textual and interpersonal resources does not necessarily contribute to a higher perceived quality of writing.

Our findings also show that the varying use of boosters may have affected grading decisions, with high-rated and medium-rated scripts having a low frequency of them. This finding concurs with the patterns observed in L2 pre-tertiary writing (Hyland and Milton 1997) and some studies on tertiary writing (Crosthwaite and Jiang 2017). Contrary to expectations, hedges in our study were not a significant predictor of grades, despite the claim that the greater inclusion of hedges is likely to contribute to successful writing (Ho and Li 2018; Lee and Deakin 2016). Thus, the use of hedges may not be a major consideration of the assessors in our context, as the assessment criteria did not address the ability to mitigate claims and maintain subjectivity—a more important consideration in tertiary academic writing.