Citation competence is an important rhetorical ability to position a writer’s research in relation to previous work and to build connections between them ([25]; [35]). High citation competence enables a writer not only to integrate others’ words or findings but also to present the purpose of the research convincingly and to show findings persuasively ([17]; [26]). Large-scale diachronic evidence confirms the centrality of this ability: citation in published research has grown steadily denser and more varied over the past five decades ([19]). For second language writers, integrating sources appropriately remains a well-documented challenge at both undergraduate and postgraduate levels ([10]; [21]; [37]). Because citation competence is realised in concrete citation features, it can be evaluated by analysing those features in research writing. With reference to appraisal theory ([27]), [6] focused on the interpersonal dimension of citations and proposed three sub-dimensions—writer’s stance, textual integration of the source, and the nature of the source. [16] applied and revised Coffin’s framework to compare citation competence across disciplines. These frameworks, however, are based on the interpersonal metafunction alone.

[39] established a more comprehensive framework drawing on the ideational, interpersonal, and textual metafunctions. In Xu’s framework, the ideational dimension concerns citing “what” (the content of citations); the interpersonal dimension concerns citing “whom” (the function of citations and the writer’s relationship with cited researchers); and the textual dimension concerns “how to cite” (the type and placement of citations). Compared with Coffin’s interpersonal-only model, Xu’s framework offers a more complete evaluation of citation competence and has yielded useful findings when applied to applied-linguistics theses ([40]). The present study, following [22], is based on and further extends Xu’s framework.

It should be noted that, in the Chinese research context, citation has also been studied from a bibliometric and informetric perspective, in which the citation is treated as a unit for evaluating research rather than as a feature of writing competence. This strand includes the evaluation of the innovativeness of academic papers through their citing intention ([36]), the improvement of journal-evaluation methods from the perspective of single-paper citation ([38]), and citation-based topic clustering ([13]). Although valuable, these studies analyse citations for the purpose of research evaluation; the present study, by contrast, analyses citation as a realisation of a writer’s competence, and therefore adopts the writing-oriented framework of [40].

Following [22], this study expands Xu’s framework by adding the writer’s stance to the interpersonal dimension, since stance shows the writer’s attitude towards the cited text and is closely related to the relationship the writer builds with cited researchers. The expanded framework thus comprises four features—content, function, writer’s stance, and type—distributed across the three dimensions. The function of citations is analysed using the six categories proposed by [8], who also studied citations in the research writing of Chinese English-major students: supporting, defining, application, topic-relevance, agreement, and disagreement. The writer’s stance is analysed using [17] widely adopted classification: acceptance, disagreement, positive, neutral, tentative, and critical. Apart from acceptance and disagreement, the other four are realised covertly, with the writer making the cited author responsible for the attitude. The reporting verbs that introduce integral citations are a central resource for this covert positioning, and their use has been shown to vary systematically across disciplines and genres ([4]).

Citation features differ across the sections of a research text because the sections perform different rhetorical functions ([12]; [34]). The Introduction establishes the territory and significance of a study and points out a niche, whereas the Literature Review builds a network of previous research, presents theoretical background, and positions the writer’s own study ([20]; [26]; [31]). It is reasonable to expect that the densities and the distribution of citation features in these two sections differ. Although scholars have investigated citation features in single sections, fewer studies have made a detailed comparison between the Introduction and the Literature Review in the writing of Chinese students, which this study undertakes.

Following [39] division, the content of a citation falls into six categories: research topics (citing previous topics or concrete details), concepts/definitions/terms, theories/models (theories, models, frameworks, or constructs), methodology (research designs, instruments, or procedures), results/findings, and viewpoints/explanations (attitudes, evaluations, or explanations of concepts shown by previous authors).

The function of citations ([8]) comprises supporting (presenting cited standpoints or findings to support the writer’s claim), defining, application (applying a cited construct, model or theory in one’s own analysis), topic-relevance (making comparisons with previous studies), agreement, and disagreement. The writer’s stance ([17]) comprises acceptance and disagreement (explicitly presenting the cited information as true or false) and the four covert stances: positive, neutral, tentative, and critical, in which the writer makes the cited author responsible for the attitude.

The type of a citation ([32]; [34]) is either integral or non-integral. Integral citations are divided into verb-controlling (the cited author acts as the agent controlling a verb) and naming (the cited author functions as a noun phrase). Non-integral citations are divided into source (attribution of a proposition to a cited author), identification (clarification of an agent within the sentence), reference (introducing other texts to the reader, often with “e.g.”), and origin (indicating the originator of a concept or product) (Table 2).

Table 2.Definitions and examples of the four citation features in the three dimensions.

Examples are authentic instances from the present corpora.

The corpus consists of two sub-corpora. The Chinese Student Corpus (CSC) comprises 30 English-medium theses written by fourth-year Chinese English-major undergraduates, who conducted small-scale empirical studies in English language teaching and applied linguistics. The Professional Writer Corpus (PWC) comprises 15 articles published in leading international journals in applied linguistics, EAP, and second language writing (e.g., Journal of English for Academic Purposes, Journal of Second Language Writing), including classic citation research (e.g., [28]; [26]) and recent empirical studies. The student theses were sampled exhaustively from a single institutional cohort: all available BA theses defended by English majors at one Chinese university in 2024 and 2025 were included, with one thesis being replaced by another from the same cohort owing to an unrecoverable file-conversion fault. The professional articles were purposively sampled against three criteria: (a) publication in peer-reviewed international journals in applied linguistics, EAP, or second language writing; (b) topical comparability with the theses (academic writing, English language teaching, and second language learning); and (c) availability of clean full text for annotation; topics in both corpora therefore fall within applied linguistics broadly defined. After cleaning (removal of titles, headers, footers, tables, figures, appendices, and reference lists), the CSC contained 802,026 words and the PWC 122,968 words. Individual texts averaged 26,734 words (SD = 5,645) in the CSC and 8198 words (SD = 2071) in the PWC. The Introduction and Literature Review of each text were identified from section headings; the cross-sectional analysis (RQ2) was conducted on the student corpus, whose theses have clearly demarcated Introduction and Literature Review chapters (Table 3).

Table 3.The corpus and its sub-corpora.

Density = citations per 1,000 words of the cleaned corpus.

A combination of automatic and manual procedures was used so that the annotation could be applied consistently to a large corpus while preserving the interpretive judgement that the framework requires. Three complementary kinds of tool were employed to this end. First, purpose-built text-processing scripts cleaned the texts, segmented each text into rhetorical sections from its headings, retrieved every in-text citation together with its citing sentence, and assigned a provisional code for each of the four features. Second, AntConc was used as a concordancer to inspect reporting structures and the immediate context of citations during manual verification. Third, the qualitative-analysis software NVivo was used to manage the coding scheme and to store, retrieve, and revise the codes for the four features. Frequencies were normalised as densities per 1,000 words; the individual text was the unit of analysis, the density of each feature in each text being computed as its raw frequency divided by the text’s cleaned word count and multiplied by 1000; group comparisons were computed with independent-samples (Welch) t-tests on these per-text densities (PWC n = 15; CSC n = 30), so that degrees of freedom reflect texts rather than citations, and Cohen’s d was calculated for each comparison as a measure of effect size (the magnitude of the difference), since statistical significance alone does not indicate how large a difference is. Because 24 feature-level tests were conducted, the Benjamini–Hochberg false-discovery-rate procedure ([1]) was applied across them at q < 0.05; all feature-level differences reported as significant below survive this correction unless explicitly noted, and given their number these comparisons are best read as exploratory. In addition, the reporting verbs introducing integral citations were extracted and classified by their evaluative loading—factive, non-factive, or counter-factive—following [18], in order to examine the lexical realisation of the writers’ stance. Each reporting verb was identified within the citing clause of an integral citation, lemmatised, and assigned its loading from a closed lemma list adapted from [18] (factive, e.g., find, show, demonstrate, identify; non-factive, e.g., suggest, propose, argue, believe; counter-factive, e.g., fail, overlook); classification was thus based on the verb lemma alone rather than on its local discourse context. For the cross-sectional comparison (RQ2), the association between section (Introduction versus Literature Review) and the distribution of each feature in the student corpus was additionally tested with chi-square tests on the raw frequencies. Welch degrees of freedom are rounded to integers for reporting, and 95% confidence intervals (CIs) of the between-group differences are provided for the principal significant comparisons.

The annotation followed a two-stage, software-assisted-then-manual workflow designed to maximise both consistency and validity. In the first (automatic) stage, the scripts assigned every citation a provisional value for each feature: citation type and integration were determined from surface patterns (author placement, quotation marks, single vs. multiple sources), while content, function, and writer’s stance were assigned from explicit lexico-grammatical cues—reporting verbs, evaluative lexis, and contextual markers—operationalised from the definitions in Table 2. In the second (manual) stage, the researcher checked and, where necessary, corrected every provisional code against the framework definitions, using AntConc concordances to resolve context-dependent cases; the manually verified codes are those reported below. Explicit counting and disambiguation rules were applied: a parenthetical string citing several sources in support of a single proposition (author–year references separated by semicolons) was treated as one citation and coded once, its type recorded as reference; distinct propositions within one sentence, each supported by its own citation, were counted and coded separately; and where more than one content, function, or stance category was plausible, the category realised by the main clause of the citing sentence was coded (a primary-function rule, cf. [28]), residual ambiguities being resolved by adjudication against the category definitions during the manual stage.

The scheme’s reliability was established in three ways, following [22]. A pilot annotation of three texts from each sub-corpus was first carried out to refine the operational definitions and adjudicate borderline cases. The accuracy of the automatic stage was then assessed on a stratified random sample of citations by comparing the software-assigned codes with the researcher’s independent manual coding; agreement was 78% for content, 88% for function, 88% for writer’s stance, and 88% for type. These figures are comparable to the 91.7% inter-coder agreement reported by [22]; the lower value for content reflects the difficulty of the fine-grained semantic judgment of what is cited, and all content codes were therefore manually verified. Third, because agreement between the automatic and the manual codes indexes the accuracy of the automatic stage rather than coder reliability, inter-coder reliability was additionally assessed on a random sample of 100 citations, re-coded independently and blind to the first-round codes (three detection false positives were identified and excluded, indicating a detection precision of about 97%). Raw agreement was 69.1% for type, 77.3% for stance, 66.0% for function, and 35.1% for content, and Cohen’s κ was 0.59, 0.53, 0.23, and 0.21 respectively. Because κ is unstable when category distributions are highly skewed ([11]), [14] AC1 was computed alongside it, yielding 0.64, 0.74, 0.62, and 0.24. The contrast between κ (0.23) and AC1 (0.62) for function indicates that its low κ chiefly reflects a prevalence effect—the supporting category accounts for roughly nine tenths of all function codes—rather than pervasive disagreement, with residual divergences concentrated at the supporting–defining boundary; for content, by contrast, all indices were low, with disagreements concentrated at the research topics–concepts boundary. Importantly, the two coding rounds converged on the core between-group contrasts: in the re-coded sample the non-integral share was 64.6% (PWC) versus 30.3% (CSC) under the first-round codes and 64.5% versus 33.3% under the blind re-coding, and the direction of the neutral-stance difference was likewise preserved. When the six content categories were collapsed into three broader classes (conceptual, methodological, and results-oriented), agreement rose to 79.4% (κ = 0.42); the content findings are therefore robust at this coarser level, while the six-way content results are best read as exploratory. All disagreements were resolved by consensus adjudication. The low-frequency categories—disagreement (function and stance) and origin (type)—were retained because they help distinguish the two groups of writers.

The overall citation density of the professional writers (9.79 per 1000 words) was more than twice that of the student writers (4.39), and the difference was significant (t(18) = 9.05, p < 0.001). This indicates that the professional writers engaged with the literature far more frequently. The difference in mean per-text density was 5.27 per 1000 words (95% CI [4.05, 6.49]). A sensitivity check confirmed that the inclusion of one response article in the PWC did not drive this result: its density (7.94 per 1000 words) lies close to the PWC mean, and excluding it changes the mean per-text density only marginally (from 9.61 to 9.73 per 1000 words). The subsections that follow examine the ideational, interpersonal, and textual dimensions in turn.

Table 4 reports the density of each content category in the two corpora, with Welch tests and effect sizes.

Table 4.Content of citations (PWC vs. CSC).

Density per 1,000 words: Mean (SD); Welch t-test; d = Cohen’s d (small ≥ 0.2, medium ≥ 0.5, large ≥ 0.8). Significant differences (p < 0.05) are shown in the p column.

In the ideational dimension (Table 4), the professional writers cited research topics (6.14 (1.68) vs. 2.28 (0.69) per 1000 words; t(16) = 8.51, p < 0.001; 95% CI of the difference [2.90, 4.82]) and results/findings (t(17) = 4.27, p = 0.001) significantly more often than the student writers; differences for theories/models (t(20) = 2.06, p = 0.053) and methodology (t(17) = 1.95, p = 0.069) were not statistically significant. This indicates that, beyond a shared core of topic-oriented citation, the professional writers engaged more fully with the substance of prior work—above all, its empirical findings—rather than merely naming the topics it addressed.

(1) Academic vocabulary, encompassing words like “approach”, “infer”, and “notion”, is commonly used across various subject disciplines in higher education (Coxhead, 2020). [PWC, research topics]

Analysis: The professional writer cites a foundational source to establish the research territory; such topic-level citation is the shared core of both corpora but is used far more densely by professionals.

(2) [18] conducted a large-scale study with 645 university students in China and found significant differences between arts and science students. [PWC, methodology]

Analysis: Here, the citation reports the design and scale of a prior study, illustrating the professionals’ fuller engagement with how earlier research was carried out—a content type on which the two groups did not differ significantly, though the professionals still used it more.

(3) Ji (2011) proposed a comprehensive taxonomy that categorizes feedback by source, form, and focus. [CSC, theories/models]

Analysis: When the student writers moved beyond topics, they most often cited a theory or model to frame the construct under study, as here; theories/models were one of the few content types they used relatively frequently.

(4) However, in a previous study, it was found that immediate correction of oral errors was more desirable than correction after speaking ([25]), which is different from this study. [CSC, results/findings]

Analysis: Although this student cites a prior empirical finding, results/findings content remained comparatively rare in the student corpus (0.28 vs. 0.98 per 1000 words), where topic-level citation predominates.

In the interpersonal dimension, two features—the function of citations and the writer’s stance—were examined. Table 5 reports the results for the function of citations.

Table 5.Function of citations (PWC vs. CSC).

Density per 1000 words: Mean (SD); Welch t-test; d = Cohen’s d (small ≥ 0.2, medium ≥ 0.5, large ≥ 0.8). Significant differences (p < 0.05) are shown in the p column.

For function (Table 5), supporting was by far the most common function in both groups, but the professional writers used it far more densely (8.47 (1.83) vs. 3.95 (1.08); t(19) = 8.86, p < .001; 95% CI [3.45, 5.59]). More tellingly, the professionals fulfilled the higher-order functions of application (t(14) = 3.41, p = 0.004) and topic-relevance (t(16) = 4.17, p = 0.001) significantly more often, while the difference for agreement was not statistically significant (t(15) = 2.02, p = 0.063). This finding closely parallels [22], who likewise found application and agreement to separate the more competent writers: writers with greater competence are better able to apply previous models in their own research, to connect studies to one another, and to express direct agreement with cited claims, rather than merely presenting them.

(5) Following Egbert and Biber (2023), we considered Cohen’s d values exceeding 0.2 to be the threshold for a meaningful effect. [PWC, application]

Analysis: The professional writer applies a cited criterion directly in the study’s own analysis—one of the functions that most clearly separates the two groups (t = 3.41, p = 0.004).

(6) Consequently, this paper uses the systematic evaluation method proposed by Cheng (2011) as the theoretical basis for evaluating textbooks. [CSC, application]

Analysis: When students did apply a source, they typically adopted a single framework in its entirety, as here, rather than integrating and adapting several—a narrower realisation of the same function.

(7) In keeping with the development of intercultural communicative competence, [5] broadened Byram’s model by introducing symbolic understanding as a necessary addition. [CSC, topic-relevance]

Analysis: This topic-relevance citation connects two studies to each other; the students used such connective moves far less often than the professionals (0.07 vs. 0.33 per 1000 words), reflecting a more list-like presentation of sources.

Table 6 presents the distribution of the writer’s stance across the two corpora.

Table 6.Writer’s stance (PWC vs. CSC).

Density per 1000 words: Mean (SD); Welch t-test; d = Cohen’s d (small ≥ 0.2, medium ≥ 0.5, large ≥ 0.8). Significant differences (p < 0.05) are shown in the p column.

For writer’s stance (Table 6), the covert neutral stance dominated both corpora, again at a much higher density in the professional writing (7.13 (2.06) vs. 3.20 (1.00); t(17) = 6.98, p < 0.001; 95% CI [2.74, 5.12]). The professional writers also took a positive stance (t(23) = 3.99, p = 0.001) and a tentative stance (t(15) = 3.02, p = 0.009) significantly more often. The explicit stances of acceptance and disagreement were rare in both groups, and although the professionals adopted a critical stance more frequently, the difference was not significant. Overall, the student writers refrained from signaling evaluation, presenting cited material largely without committing to a position.

(8) Research targeting first-year university students in Hong Kong SAR has shown that these students find academic vocabulary to be challenging and overwhelming (Chung & Newton, 2024). [PWC, positive]

Analysis: The factive reporting verb “shown” aligns the writer with the cited claim, presenting it as established knowledge; this covert positive stance was the second most common in both corpora and significantly denser in professional writing.

(9) [10] asserts that a host of variables, including trust, can frame learners’ use of technology. [PWC, critical]

Analysis: “Asserts” covertly signals a critical stance, making the cited author responsible for a contestable claim; such evaluative positioning was more available to the professionals than to the students.

(10) Graham and Harris (1989) proposed that self-efficacy essentially boils down to students’ assessments of their own capabilities in successfully executing a given writing task. [CSC, tentative]

Analysis: “Proposed” conveys a tentative stance, but instances like this were the exception; the students overwhelmingly defaulted to a neutral stance rather than taking graded positions on their sources.

(11) When it comes to English literature education, [10] asserts that teachers must craft activities that align with key competencies, particularly cultural awareness. [CSC, critical]

Analysis: This rare student instance of a critical stance shows that the linguistic resource is available to the student writers, yet it was seldom deployed—the students largely refrained from evaluating the sources they cited.

Table 7 reports the density of each citation type in the two corpora.

Table 7.Types of citations (PWC vs. CSC).

Density per 1000 words: Mean (SD); Welch t-test; d = Cohen’s d (small ≥ 0.2, medium ≥ 0.5, large ≥ 0.8). Significant differences (p < 0.05) are shown in the p column.

In the textual dimension (Table 7), the two groups differed most clearly in their use of non-integral citation. The professional writers used the source type (4.12 (1.63) vs. 1.37 (0.78); t(17) = 6.20, p < 0.001; 95% CI [1.81, 3.69]) and the reference type (t(15) = 5.70, p < 0.001) significantly more often, emphasizing the cited content and introducing readers to multiple related texts. The student writers, by contrast, used verb-controlling citation—in which the cited author is the grammatical subject—more than the professionals (1.02 (0.53) vs. 0.67 (0.44); t(34) = −2.31, p = 0.027, uncorrected; this difference does not survive the false-discovery-rate correction), foregrounding the authority of the cited researcher rather than the proposition. This pattern echoes the finding that more competent writers favor non-integral citation to assert their own voice ([24]; [33]).

(12) Cultural tools objectify culturally developed knowledge as they are interpreted and used in interaction ([37]). [PWC, source]

Analysis: The non-integral source citation places the author outside the sentence, foregrounding the proposition rather than the researcher—the concept-oriented style used by the professionals far more than the students (4.12 vs. 1.37 per 1000 words).

(13) Unlike technical vocabulary, academic vocabulary is typically not the primary focus of instruction in subject-area courses (Dang, 2022; Nation, 2022; Skoufaki & Petrić, 2021). [PWC, reference]

Analysis: The reference type bundles several sources to introduce a body of work to the reader; the professionals used it roughly nine times as densely as the students, reflecting a broader command of the literature.

(14) Entwistle (2000) and Ramsden (2003) developed the concept and theory of deep learning and surface learning. [CSC, verb-controlling (integral)]

Analysis: Here, the cited authors are the grammatical subjects that control the verb, giving them prominence over the proposition—the author-prominent integral style that the students used significantly more than the professionals.

Because integral citation is realised through reporting verbs, a closer analysis of these verbs refines the interpersonal findings (Table 8).

Table 8.Reporting verbs introducing integral citations (PWC vs. CSC).

Evaluative loading follows [18]: factive verbs present the cited claim as true; non-factive verbs attribute it without commitment; counter-factive verbs portray it as flawed.

The two groups used reporting verbs at a broadly similar rate (14.31 vs. 12.90 tokens per 10,000 words), but they used them very differently. The professional writers favoured factive, research-oriented verbs—such as find, identify, report, and show—which present the cited claim as an established finding that the writer accepts; 52.8% of their reporting verbs were factive. The student writers instead relied on non-factive, discourse-oriented verbs—above all propose, point out, believe, and argue—which attribute a view to the cited author without committing the writer to its truth; only 34.2% of their verbs were factive, while non-factive verbs accounted for 64.6%. Even though the larger student corpus contained more verb types in absolute terms (50 vs. 35), the students concentrated more heavily on a few high-frequency verbs (their top five verbs covered 46.5% of all tokens, vs. 42.6% for the professionals), and three of those five—propose, point out, believe—are precisely the non-committal attribution verbs that realise a covert neutral stance. Counter-factive verbs (e.g., fail, overlook) were marginal in both groups (about 1%), confirming that neither group frequently contested its sources. The reporting-verb data thus provide the lexical mechanism behind the stance results in Table 6: the students’ narrow, non-factive verb repertoire is the concrete means by which they maintain distance from evaluative commitment.

The student writers’ citations were heavily concentrated in the Literature Review, consistent with its rhetorical function of building a network of prior research. Within-section percentages reveal how the function of each section shaped the distribution of citation features. Table 9 shows the distribution of content categories in the two sections.

Table 9.Content of citations in the introduction and literature review (student corpus, %).

Within-section percentages.

In the ideational dimension (Table 9), research topics accounted for a larger share of citations in the Introduction (66.28%) than in the Literature Review (53.95%), whereas concepts/definitions/terms (2.68% vs. 8.09%) and theories/models (10.34% vs. 15.04%) made up a larger share in the Literature Review. The overall association between section and content distribution was significant (χ²(5) = 22.59, p < .001). This is consistent with the Introduction’s role in establishing the topic and the Literature Review’s role in engaging with concepts and theoretical frameworks.

Table 10 shows the corresponding distribution of citation types in the two sections.

Table 10.Type of citations in the introduction and literature review (student corpus, %).

Within-section percentages.

In the textual dimension (Table 10), the contrast was sharpest. The non-integral source type accounted for nearly half of the Introduction citations (48.28%) but only 25.27% of the Literature Review citations, whereas integral naming citations rose from 25.67% in the Introduction to 39.18% in the Literature Review, and verb-controlling citations increased from 17.24% to 27.66%. The association between section and citation type was likewise significant (χ²(5) = 67.78, p < 0.001, N = 2821). In other words, the student writers foregrounded cited authors (integral citation) precisely where they reviewed and organised the literature, while using more compact, content-focused source citations to establish the research territory in the Introduction. This mirrors [22] finding that integral, verb-controlling citation is more frequent in literature reviews.

Examples (15)-(17) illustrate this section-sensitive shift in the student writing. In the Introduction, sources are typically realized as compact, non-integral source citations that establish the research territory (15), whereas in the Literature Review the same writers foreground cited authors through integral naming (16) and verb-controlling (17) citations as they organize and engage with the literature.

(15) Existing studies have pointed out that deep learning is the basic path to developing students’ core competency (Cui, 2019). [CSC, Introduction—non-integral source]

Analysis: In the Introduction the source is placed in parentheses to establish the research territory economically; source citation accounted for almost half (48%) of the student writers’ Introduction citations.

(16) Bloom (1956) divides cognitive goals into six dimensions in his book Taxonomy of Educational Objectives: knowledge, comprehension, application, analysis, synthesis, and evaluation. [CSC, Literature Review—integral naming]

Analysis: In the Literature Review, the same writers bring the cited author into the sentence as a naming element, giving the source greater prominence; naming rose from about 26% of Introduction citations to 39% in the Literature Review.

(17) In the realm of practical research, Long (2018) merges the Activity-based Approach with the technique of continuation writing. [CSC, Literature Review—integral verb-controlling]

Analysis: This verb-controlling citation, in which the author is the acting subject, likewise concentrates in the Literature Review, consistent with that section’s function of organizing and attributing prior work.

For function and stance, the distributions were more stable across the two sections: supporting and the covert neutral stance dominated both, indicating that—regardless of section—the student writers tended to present sources to display knowledge rather than to evaluate or apply them.