Metadiscourse and Voice Construction in Discussion Sections in BA Theses by Chinese University Students Majoring in English

Research on Voice in Academic Writing

Voice is an elusive and multifaceted construct, which has been conceptualized in different ways, ranging from being one’s personal and individual features in writing to showing one’s identity within certain disciplinary and social circumstances (Tardy, 2016; Yoon, 2017). This study concurs with Matsuda’s (2001) definition of voice as “the amalgamative effect of the use of discursive and non-discursive features that language users choose, deliberately or otherwise, from socially available yet ever-changing repertoires” (p. 40). More specifically, voice in this study is viewed as the writer’s textual presence. It should be noted that voice, in line with Matsuda’s (2001) definition, is usually perceived holistically from discursive and nondiscursive sources, intentionally or involuntarily. Besides, voice is both individual and social, meaning that the writer always constructs the unique self while conforming to the conventions and practice within a specific academic community. In addition, voice is also viewed as the textual effect on the reader, which is generated from the interaction between the writer, the reader, and the text (Matsuda & Tardy, 2007). Matsuda and Tardy (2007) argued that voice is jointly constructed by writers and readers because ultimately it is the readers who sense the effect of voice. Hence, voice research can benefit from an inclusion of the reader’s perspective.

While voice is widely accepted as indispensable in academic writing, recent empirical attention has been attached to issues as to what discursive devices contribute to voice and how to measure voice strength. As mentioned above, voice does not seem to rise from the use of any single textual feature; instead, Tardy and Matsuda (2009) found in their study that editorial board members of journals constructed a writer’s identity based on a wide range of features such as strange use of jargon or tentative style. Researchers have often subscribed to Hyland’s (2008) text-oriented voice model to examine voice elements, which is further introduced in the next section. Based on Hyland’s (2008) categories of metadiscourse, Zhao (2013) developed an analytic rubric for measuring voice in argumentative writing. Through factor analysis and also qualitative methods, such as think-aloud and interviews, the rubric was validated to capture three dimensions of voice: “(1) the presence and clarity of ideas in the content; (2) the manner of the presentation of ideas; and (3) the writer and reader presence” (Zhao, 2013, p. 201).

The analytic voice rubric established by Zhao (2013) has inspired ensuing studies into voice and its relationship with text quality. In her analysis of 200 Test of English as a Foreign Language (TOEFL) essays, Zhao (2017) found that, while the three dimensions of voice were correlated with holistic essay scores, the presence and clarity of ideas mostly predicted essay quality. This proposition was endorsed by Stock and Eik-Nes (2016) who cautioned that overemphasis on linguistic features may blind us from noticing “content-related features that might be more relevant in the construction of voice” (p. 89). Yoon (2017) drew on this rubric and reported, in 219 argumentative essays written by Greek L1 EFL students, weak to moderate correlations between textual voice elements and holistic voice strength, but weak correlation between holistic voice strength and essay quality. These findings collectively suggest that a strong voice should not be simplistically assumed to lead to high quality of essays.

While researching voice in essay writing by means of rigorous measurement has achieved a breakthrough, research on voice in graduation theses has mostly adopted holistic measures. For instance, Morton and Storch (2019) investigated voice development in two sets of comparable texts written by three PhD students in the first year and toward the end of their candidature. They adopted the reader’s perspective and invited five PhD supervisors to evaluate the writer’s voice conveyed in the selected texts based on a rubric adapted from Tardy (2012). Morton and Storch’s (2019) study innovatively elicited qualitative findings by asking the supervisors to mark with a highlighting pen any fragments of the text that gave them a sense of the writer’s voice and to attend a follow-up interview. This study concluded that the supervisors’ impressions of voice varied across their language backgrounds, disciplinary specialties, personal histories, and preferences. The less conclusive results compared with those obtained through measuring voice in essays may be because thesis writing entails disciplinary knowledge and methodological expertise besides manipulating language to express ideas. In other words, academic conventions such as the use of citation, knowledge display, and knowledge creation (Geng & Wharton, 2016) may complicate the measurement of voice in a thesis, not to mention that voice is multifaceted in its own right.

Voice as Interaction Through Metadiscourse

While voice can be expressed by various discursive means, Hyland (2008) proposed a text-oriented approach to capturing voice. This approach stresses the essence of interaction in voice construction, taking voice as the way writers “position themselves and their work in relation to other members of their group” (Hyland, 2008, p. 6). Hyland (2005) established an interpersonal model of metadiscourse, which distinguishes between interactive and interactional metadiscourse. Interactive metadiscourse concerns linguistic resources for organizing the text and for guiding the reader to comprehend the text, whereas interactional metadiscourse concerns the ways in which writers position their stance and attitudes, and involve readers into the text.

Of the two types of metadiscourse, interactional metadiscourse has received more research attention, particularly since Hyland (2008) explicitly proposed an interactional model in which voice is realized by two subtypes of interactional metadiscourse: stance and engagement. Stance is writer-oriented and refers to “the ways writers present themselves and convey their judgements, opinions, and commitments” (Hyland, 2008, p. 7) by means of four types of linguistic resources: hedges, boosters, attitude markers, and self-mention. Engagement tends to be reader-oriented, which functions to acknowledge the presence of the readers and align them toward the writer’s arguments or opinions by means of five textual elements: reader mention, directives, questions, knowledge reference, and personal asides. Hyland’s (2008) interactional model of voice has provided a solid theoretical foundation for the research of voice in academic writing. In their study, comparing the results and discussion chapters of master’s theses written in English and in Spanish, respectively, by English L1 and Spanish L1 students, Lee and Casal (2014) reported that the English texts contained significantly higher occurrences of metadiscourse, specifically, hedges, boosters, and attitude markers; the Spanish texts consisted of more self-mentions and engagement markers, which was attributed to the importance attached to in-groupness and politeness strategies in Spanish culture. Burneikaitė (2008) also identified significantly more frequencies of emphatic markers (a notion related to boosters) in master’s theses in linguistics written by English L1 writers than in those by Lithuanian-speaking EAL writers.

Arguably, interactive metadiscourse also reflects the writer’s voice. Interactive metadiscourse includes transitions, frame markers, endophoric markers, evidentials, and code glosses (Hyland, 2005; presented later). These resources can serve as signposts, guiding the readers through the text by anticipating their needs and expectations. Thompson (2012) rightly pointed out that metadiscourse can be used to augment textual coherence, which can “help cement the voice of authority” (p. 125). This insight found empirical support in Morton and Storch’s (2019) study where a PhD supervisor believed that coherence shows the writer’s control over their material and is “thus an important criterion in assessing authorial voice” (p. 19). Hyland and Tse (2004) also explicitly postulated that “all metadiscourse is interpersonal in that it takes account of the reader’s knowledge, textual experiences, and processing needs” (p. 161). Li and Wharton (2012) stressed that writers with a broad repertoire of metadiscourse can construct an intentional stance and voice in text. Therefore, it is fair to state that interactive metadiscourse, with its inherent interpersonal nature, builds connections between the writer, the reader, and the text, and also conveys the writer’s voice. Compared with interactional metadiscourse, interactive metadiscourse has been under-explored (Cao & Hu, 2014), let alone its association with voice. Lee and Casal’s (2014) study reported that the English thesis extracts, compared with the Spanish extracts, consisted of significantly more instances of transitions, endophoric markers, and evidentials, which assist readers’ textual navigation and are typical in English rhetorical conventions. In Burneikaitė’s (2008) study, the Lithuanian students employed significantly more text-connectives (a notion related to transitions). This result was attributed to “an (over-)excessive emphasis on connectives in academic writing instruction at secondary or undergraduate level” (Burneikaitė, 2008, p. 43).

The view that metadiscourse can function as resources for constructing voice has also been articulated in other studies besides Hyland (2008). For instance, Helms-Park and Stapleton (2003) developed a Voice Intensity Rating Scale for measuring voice in argumentative writing. This scale contains four components, of which assertiveness can be established through the use of hedges and boosters, self-identification through self-mentions, and reiteration of central point arguably through code glosses. The last component, authorial presence and autonomy of thought, is evaluated primarily based on readers’ impression of the written text (Zhao & Llosa, 2008). Such impression can potentially derive from readers’ sense of being involved into the writer’s stance or attitudes and being guided through the discourse, which may be created by metadiscourse (Hyland, 2008). This being said, the roles of metadiscourse markers in conveying voice may not be on a par. Ädel (2006) pointed out that “the interpersonal function is, in fact, fundamental to the overall category” (p. 17). It is likely that interactional metadiscourse markers (e.g., hedges, boosters, and attitude marker) better demystify voice as they naturally reflect the writer’s stance and attitudes.

In brief, important issues regarding metadiscourse in relation to voice construction and voice improvement in BA thesis writing have remained unaddressed. Many EAL undergraduate students are inexperienced in both empirical research and English academic writing whose rhetoric conventions can be different from the L1 writing (Hirvela & Belcher, 2001). Efforts in this light can generate important implications for supporting undergraduates in their thesis writing, preparing them for further academic pursuits, and enhancing their interpersonal skills in written communication. This study was guided by the following research questions:

Corpus Materials and Informants

The corpus materials included 35 sets of discussion sections taken from the first and the final versions of 35 BA theses written by English majors studying in a university in Southern China. English majors in this university have to complete a graduation thesis of about 6,000 to 8,000 words at the end of their final year. Students are required to, within a period of about 6 months, submit at least two drafts to their supervisors and revise the drafts according to the supervisor’s feedback before they are approved to submit the final version. Therefore, students’ engaging in process-oriented writing is mandated to improve the quality of the theses. The 35 sets of discussion sections were all taken from theses in applied linguistics with the consent of their writers. The decision to focus on applied linguistics was motivated by the generally shared argument that “academic writing varies systematically across disciplines” (Li & Wharton, 2012, p. 346). While the discussion section is often a necessary component in evidence-based research in applied linguistics, in the fields of literary or translation research, considerable structural variations can exist and there may not be a distinct discussion section in the theses, as was found in many theses archived in China National Knowledge Infrastructure (CNKI), which is the largest database of academic works in China.

We kept only the discussion section in each thesis, with other parts removed. As in some theses results/findings and discussion were blended into one section, we retained only the scripts related to discussion according to the function of a discussion section, that is, discuss the meaning and significance of the results/findings, as specified in Bitchener (2010). While this may affect text coherence, we considered it applicable in the current context because, as we found, the undergraduate writers basically followed the results-discussion pattern. These texts formed a corpus of 57,372 words, with 25,869 words in the first-draft subcorpus and 31,503 words in the final-draft subcorpus.

To triangulate the corpus analysis of metadiscourse in relation to voice construction, we also incorporated the reader’s perspective. Two native Chinese-speaking faculty members in the university where the corpus materials were collected were invited as informants, who were given the pseudonyms Jane and Mary. They both held a PhD degree, obtained overseas, and were experienced in publishing research articles in prestigious international journals such as those indexed in the Social Sciences Citation Index (SSCI). They were supervisors of the English majors, but not the supervisors of the theses analyzed in this study. They were invited to first review 10 of the 35 pairs of the discussion sections, rate the voice strength of the sample texts based on a rubric used in Morton and Storch (2019), and then to attend a follow-up interview.

Instruments

The instruments used in this study consist of three parts. The first is the UAM CorpusTool Version 3.3 (O’Donnell, 2019), which is a freeware annotation tool. This tool was used to facilitate the coding of metadiscourse in the discussion sections included in the corpus and it could return with the frequencies of the coded types of metadiscourse.

The second instrument is a rubric adapted from Morton and Storch (2019) to elicit the reviewers’ impressions of the sample discussion extracts. The rubric evaluated the extracts in terms of central focus, style, language, the writer’s credibility, and authorial presence. Voice strength examined in this study was based on the reviewers’ ratings on a 5-point scale in response to the statement “The writer’s authorial presence in this paper is” anchored, respectively, from weak/not noticeable (score = 1) to strong/very noticeable (score = 5).

An interview protocol was used to elicit the reviewers’ comments on their impression of voice construction and voice development in the English majors’ theses and factors contributing to their impression. This protocol is shown in Appendix B.

Procedures

This study was conducted in two phases. In the first phase, the discussion sections in the corpus were analyzed and coded using UAM CorpusTool (O’Donnell, 2019). The data were coded by the second author and, to enhance the reliability of the coding, five pairs of the texts (14.3% of the data) were randomly selected and independently coded by the first author. Intercoder reliability calculated for each subtype of interactive and interactional metadiscourse ranged between 90% and 93%, which were “in the 90% range” specified by Miles and Huberman (1994, p. 64). Upon completing the coding, normalized frequencies per 1,000 words of the two main types of metadiscourse and their subtypes, that is, (frequency of each type/number of words in the text) * 1,000, were calculated by the UAM CorpusTool and obtained for subsequent statistical analyses.

In the second phase, the two informants were invited to review 10 sample pairs of texts, each being labeled with a number plus “a” or “b” (e.g., Text 1a, Text 1b). They were informed that the labels of “a” and “b” in the file name were randomly used without suggesting the sequence of a pair of texts. The reviewers were asked to highlight any words, expressions, or other textual chunks that gave them a sense of the writer’s voice (see Morton & Storch, 2019). Following this, the metadiscourse elements highlighted by the reviewers were identified and categorized according to Hyland’s (2005) framework, and their frequencies in the first and final versions of each pair, respectively, were then calculated. The reasons for presenting only 10 pairs of texts to the reviewers were twofold. First, language teachers nowadays are already “steeped in work and responsibility” (Hobbs & Kubanyiova, 2008, p. 503), which was the case for the two reviewers at the time this study was conducted, so it was unpractical to request them to review 35 pairs of texts. In addition, as the purpose of the second phase was to complement the corpus-based analysis with qualitative findings from the reader’s perspective, the 10 pairs of texts were considered suitable for fulfilling this purpose. The reviewers were also asked to rate each text’s voice strength based on the rubric adapted from Morton and Storch (2019). They were then interviewed individually for about 20 to 30 min. The interviews were recorded with their consent and transcribed verbatim.

Analytic Framework of Metadiscourse

The corpus was analyzed based on Hyland’s (2005) interpersonal model of metadiscourse, which distinguishes between interactive metadiscourse and interactional metadiscourse. Interactive metadiscourse includes five subtypes: transitions, frame markers, endophoric markers, code glosses, and evidentials. Transitions (therefore) function to signal logical connections within the text and thus enhance cohesion. Frame markers are used to organize texts in terms of sequences (second), topics (with regard to), discourse stages (thus far), and discourse goals (aim to). Endophoric markers are used to refer to other parts of the text (as mentioned above). Code glosses function to exemplify or rephrase propositional meanings (in other words). Evidentials are used to introduce information from other sources into the text through citations (according to).

While interactive metadiscourse is used to maneuver textual elements to meet readers’ expectations of the unfolding text, which enhances textual coherence and thereby reflects writers’ presence, interactional metadiscourse is used to directly address the readers. Five subtypes of interactional metadiscourse were examined in this study: hedges, boosters, attitude markers, self-mentions, and engagement markers. Hedges (perhaps) are the devices used to mitigate assertiveness and open dialogic space, whereas boosters (demonstrate) are used to emphasize writers’ certainty and to close down alternatives. Attitude markers are used to convey writers’ attitudes toward propositions, which include attitude verbs (prefer), adverbs (interestingly), and adjectives (important). Self-mentions refer to the use of first-person pronouns and possessive structures to indicate author presence in the text (I). Engagement markers are devices used to explicitly address the reader (you) or pull them into the discourse (note; see Appendix A for the metadiscourse markers examined in this study).

Data Analysis

Data analysis includes several steps. Based on the framework presented above, tokens of interactive and interactional metadiscourse and their subtypes were identified and annotated. Several practices were adopted. First, the identification of self-mentions was based on Hyland’s (2005) framework, also considered a broad approach (Alotaibi, 2018; Hyland, 2017), which sidesteps a distinction made by Ädel (2006) between references to the world of discourse and those to the real world, which represents a narrow approach. While differentiating self-mentions in the two “worlds” (Ädel, 2006) surely allows a focused lens on the text per se and on “the current writer and reader qua writer and reader” (p. 183), in practice, self-mentions referencing to the world of discourse and real world are not always clear-cut (Alotaibi, 2018) and may depend on genres (Alotaibi, 2018) and disciplines (Lee & Casal, 2014). Therefore, to avoid the risk of “eliminating much of what makes metadiscourse a powerful analytic tool” (Hyland, 2017, p. 27), Hyland’s (2005) broad approach was adopted in this study.

Following Cao and Hu’s (2014) analysis, intra-sentential transitions (e.g., because, since) were not counted as transitions because they basically fulfill syntactic functions instead of acting as “metadiscoursal logical markers” (p. 19). Besides, the often-used pair of phrases, on the one hand and on the other hand, were coded as frame markers signaling spatial sequences when they were used to introduce sequences, but coded as transitions when they were used to show logical connections (Cao & Hu, 2014). Modal verbs like could and would were counted as hedges only when they were used to qualify degrees of certainty, but were excluded when used “in a subjunctive mood or past tense that had nothing to do with hedging” (Zhao, 2013, p. 205). In addition, a distinction was made between self-mention we and engagement marker inclusive we, the former referring to authorship, whereas the latter functioning to capture readers’ attention or involve readers in the discourse (Hyland, 2005).

To address the first research question, we performed dependent t tests to compare between the first and final versions the normalized frequencies of interactive metadiscourse and interactional metadiscourse, respectively. The assumptions that sample means and the differences between sample means were normally distributed were met (Field, 2009). However, the data of the subtypes of each metadiscourse violated the assumption of normal distribution. Hence, Wilcoxon signed-rank test, which is the nonparametric equivalent of dependent t test (Field, 2009), was performed to test the differences in each subtype of metadiscourse between the first and the final drafts.

To answer the second research question, the metadiscourse elements highlighted by the reviewers were categorized according to Hyland’s (2005) framework and their frequencies in the first and final versions of each pair, respectively, were then calculated. As the purpose here was to complement the corpus-based analytic findings from the reader’s perspective, inferential statistics was not processed.

The third research question was addressed by drawing on several types of data. The voice strength of each text was obtained by averaging the two reviewers’ ratings. We then examined the voice strength of each text and the frequencies of metadiscourse markers highlighted by the reviewers as reflecting the writer’s presence for the purpose of spotting association or compatibility, if any, between the two types of information. In addition, qualitative content analysis (Bryman, 2012) was performed on the interview transcripts. This involved repeatedly reading and searching in the reviewers’ comments for themes regarding voice improvement in the sample texts and possible underlying reasons. The themes reported in this article concern the roles of metadiscourse and other aspects that were perceived by the reviewers to account for differences in voice between the first and final versions of the texts. In addition, a theme regarding supervisors’ role in guiding students’ voice construction also inductively emerged from the interview data.

Use of Metadiscourse in the First and Final Versions of the Texts

For the dependent t tests and Wilcoxon test conducted in this study, effect size (ES) was obtained, referencing Field (2009), by calculating r in t test (r = t 2 / t 2 + d f ) and r in Wilcoxon test (r = Z / n ). The cutoff values of r, suggesting a small, medium, and large effect are 0.1, 0.3, and 0.5, respectively (Field, 2009).

The results of dependent t tests showed that both interactive and interactional metadiscourse occurred more frequently in the final-draft subcorpus (M = 19.39, SD = 7.02; M = 19.40, SD = 7.22, respectively) than in the first-draft subcorpus (M = 16.61, SD = 8.06; M = 13.24, SD = 7.44, respectively). These differences were statistically significant (t = −2.83, p = .008, r = .44, medium ES, for interactive metadiscourse; t = −5.77, p < .001, r = .70, large ES, for interactional metadiscourse).

Detailed differences in the subtypes of interactive and interactional metadiscourse between the two subcorpora were further examined by running a series of Wilcoxon signed-rank tests. Table 1 shows the descriptive statistics and the test results.

OPEN IN VIEWER

Table 1.

Descriptive Statistics and Wilcoxon Signed-Rank Test Results.

Subtype	First version		Final version		z	p	r
	M	SD	M	SD
Interactive
Transitions	7.52	4.88	7.81	3.85	−0.14	.89	.02
Frame markers	1.96	2.51	1.86	1.91	−0.74	.46	.09
Endophoric markers	1.10	1.27	1.30	1.16	−0.97	.33	.12
Evidentials	4.05	3.53	5.09	3.39	−2.22*	.03	.27
Code glosses	2.20	2.55	3.23	2.64	−1.75	.08	.21
Interactional
Hedges	8.46	5.46	13.00	5.53	−4.42**	.00	.53
Boosters	1.80	2.10	2.96	2.96	−2.37*	.02	.28
Attitude markers	1.09	1.32	1.17	1.32	−0.54	.59	.06
Self-mentions	0.13	0.42	0.14	0.84	−0.67	.50	.08
Engagement markers	1.76	2.85	2.19	1.94	−1.79	.07	.21

*

p < .05. **p < .01.

As seen in Table 1, the final-version subcorpus contained significantly more evidentials (z = −2.22, p = .03, r = .27), more hedges (z = −4.42, p < .001, r = .53), and more boosters (z = −2.37, p = .02, r = .28) than the first-version subcorpus. Except for the three subtypes, no significant differences in other subtypes were identified between the two subcorpora. The increased use of evidentials, hedges, and boosters in the two versions are illustrated in the following three pairs of examples, with these markers in the second version shown in bold:

As seen in Examples 1a and 1b, the writer in the final version employed the evidentials in the form of citation to align his or her result with that of previous studies, which was not found in the first draft. The use of the evidential to represent an idea from another source in the final version can convey the writer’s voice through his or her alluding to the disciplinary literature.

In contrast to Example 2a, Example 2b contains the hedge may, which indicated the writer’s awareness of calculating the weight given to his or her assertion (Hyland, 2005). The current finding adds evidence to Hyland’s (2008) finding that hedges are “the most frequent feature of writer perspective” (p. 12). Zhao’s (2013) factor analysis result also showed that the use of hedges is one unique dimension of voice.

In Examples 3a and 3b, the use of the booster found in the final version indicated the writer’s commitment to the proposition, through which the readers are encouraged to take on board his or her opinion.

Metadiscourse Perceived to Convey Voice

The metadiscourse elements highlighted by the reviewers as giving them a sense of voice were categorized and calculated. The reviewers’ ratings of each text’s voice strength were averaged. Table 2 shows the frequencies of the highlighted metadiscourse markers in each text and its corresponding voice strength.

OPEN IN VIEWER

Table 2.

Frequencies of Metadiscourse Markers and Voice Strength of the Sample Texts.

Sample text	A1	A2	A3	A4	A5	B1	B2	B3	B4	B5	Column total	Voice strength
Text1_first	4	0	3	10	0	2	1	2	0	0	22	1.5
Text1_final	6	1	9	12	1	6	5	1	0	0	41	3.0
Text2_first	3	0	11	6	2	11	6	2	0	2	43	4.0
Text2_final	9	0	6	8	1	5	3	1	0	3	36	3.25
Text3_first	4	0	1	3	1	6	2	3	0	1	21	4.5
Text3_final	2	4	3	4	1	6	3	1	0	1	25	3.25
Text4_first	11	3	5	3	0	6	1	0	0	1	30	3.75
Text4_final	14	0	3	3	2	13	1	1	0	0	37	4.25
Text5_first	0	0	3	2	1	2	2	1	0	0	11	3.0
Text5_final	3	1	4	1	2	5	5	0	0	1	22	4.25
Text6_first	0	0	0	0	1	3	1	0	0	0	5	1.75
Text6_final	8	1	2	2	1	8	5	0	0	1	28	3.25
Text7_first	0	0	1	0	2	2	2	0	0	0	7	2.0
Text7_final	2	0	1	0	1	2	4	1	0	0	11	3.0
Text8_first	2	0	0	1	0	1	0	0	1	3	8	3.5
Text8_final	1	0	1	2	0	2	0	0	0	7	13	2.0
Text9_first	2	2	1	1	3	0	2	0	0	0	11	2.5
Text9_final	1	3	1	2	3	2	1	1	0	7	21	2.5
Text10_first	5	1	1	3	0	1	3	3	4	0	21	3.5
Text10_final	6	1	2	2	0	2	1	3	1	0	18	3.5
Row total	83	17	58	65	22	85	48	20	6	27

Note. A1 = transitions; A2 = frame markers; A3 = endophoric markers; A4 = evidentials; A5 = code glosses; B1 = hedges; B2 = boosters; B3 = attitude markers; B4 = self-mentions; B5 = engagement markers.

The first result found in Table 2 is that the metadiscourse markers giving the reviewers a sense of voice included all subtypes. Of these subtypes, hedges, transitions, and evidentials were the top three in terms of frequencies (i.e., 85, 83, and 65, respectively). Coincidently, as reported above, two of these three subtypes, that is, hedges and evidentials, were found to occur significantly more frequently in the final-version subcorpus than in the first-version subcorpus.

Table 2 also shows that metadiscourse markers exhibited more instances in the final version of eight pairs of the texts and showed fewer instances in the final version of only two pairs (i.e., Text 2 and Text 10). This result was also largely consistent with the Wilcoxon signed-rank text results reported previously, which together may imply a rough tendency toward an increased use of metadiscourse in the final version. In other words, the increased use of metadiscourse played a part in capturing the reviewers’ attention when they interpreted the writer’s voice in the text.

It can also be observed in Table 2 that the pairs of Texts 6, 1, and 5 ranked the top three in the increased use of metadiscourse (i.e., 28 − 5 = 23, 41 − 22 = 19, and 22 − 11 = 11, respectively), and these three pairs happened to rank top three in voice strength growth (3.25 − 1.75 = 1.50, 3 − 1.5 = 1.5, and 4.25 − 3 = 1.25, respectively). Similar changes in the same direction could be observed in the pairs of Texts 2, 4, and 7. Although these anecdotal findings certainly could not amount to any conclusive evidence, it may be inferred that authorial voice perceived by the reviewers may, at least partially, be attributed to the use of metadiscourse. The following two pairs of examples illustrate the metadiscourse markers not found in the first drafts of Text 4 and Text 1 but were highlighted by the reviewers in their final versions.

The proposition in Example 4b was more sufficiently elaborated than in Example 4a, and the writer added two hedges probably, could when offering his or her interpretation. The use of these hedging devices indicated the writer’s recognition of other alternative opinions and cautions to allow space for negotiation (Hyland, 2005). Notably, this uncertainty made in staking a claim did not seem to erase the writer’s voice but instead enabled the writer to enhance his or her presence in the text by inviting other voices in the proposition.

Similarly, in Examples 5a and 5b, the original sentence was revised into two sentences in the final version, and the transition in addition to and hedge may were added. The use of the transition here indicated the writer’s awareness of readers’ expectations and his or her efforts in guiding the readers through the text. Hence, this transition could function to improve the information flow and thereby textual coherence, which resulted in a clearer presence of the writer as the text organizer.

Voice Construction in the First and Final Versions of the Sample Texts

During the interviews, both Jane and Mary commented that overall a certain extent of improvement in voice construction could be perceived when comparing each pair of the texts, but emphasized that much variations existed across different pairs. This was corroborated by the figures in Table 2, where six pairs registered increased voice strength (Texts 1, 3, 4, 5, 6, and 7), two no such increase (Texts 9 and 10), and two a drop in voice strength (Texts 2 and 8). Based on the last column in Table 2, the voice strength of the first and final versions of the 10 texts could be averaged, respectively, and the resulted values were 3 and 3.225, which indicated a minor increase in voice strength.

When asked to consider the role of metadiscourse in constructing voice, both Jane and Mary acknowledged that metadiscourse plays important roles. For instance, “signaling words” could “bring the writer’s voice to the readers” (Jane); words like to summarize could “draw the readers’ attention, that is, I am going to summarize my above-mentioned opinion” (Mary). Hedging devices such as may, possibly could “soften the writer’s argument” and when combined with verbs, “the writer’s voice could become clearer” (Jane).

Besides metadiscourse, three content-related aspects were considered by the reviewers as factors accounting for differences in voice construction between the two versions of the sample texts, or lack thereof. The first aspect is clearer connections between the present study and previous studies presented in the final version. When commenting on the pair of Text 1, Jane noted,

I think in the final draft the writer at least mentioned many consistencies between his/her findings and those in the literature, which was not mentioned in the first draft. The first draft reads like he/she was reporting what others had done, as in the Literature Review section, which showed unclear connections between [others’ studies] and his/her study.

Jane’s opinion was exactly shared by Mary, who also commented on the final version of Text 1, “He or she also built connections with previous studies, saying what previous studies have found, and my research confirmed their findings, etc.”

The second aspect contributing to voice are the explicit statements of the significance of the present study found in the final version. Jane emphasized, “How to highlight the significance? How to highlight the findings? What are the significant points? If the writer could make revisions in these lights, his or her voice could be projected.” Mary also addressed, “In term of authorial voice, I think he or she highlighted his or her findings, so [the voice] was stronger.”

The reviewers also mentioned that more elaborations or explanations on the writer’s propositions found in the final version contributed to a stronger voice. For instance, when comparing the pair of Text 5, Jane noted that the writer’s explanations on his or her propositions in the final version were much clearer than in the first draft. She added, “If the authorial voice was stronger, it was mainly due to some revisions, such as giving clearer explanations.” Coincidently, Mary also mentioned that the final draft of Text 5 was “overall written in a clearer way” and “the writer’s explanations on the present findings were clearer.”

Notably, near the end of the interview, both reviewers acknowledged the important role that supervisors could play in guiding students to convey voice in their BA thesis writing, as Jane noted,

I think if the teacher consciously does this thing and tells students that “in your final version, I hope you can express your own voice as a researcher,” better effect may be achieved. Otherwise, if [voice construction] was carried out unconsciously in their writing, maybe there is no much effect.

However, based on the sample texts they reviewed, they doubted that voice was not stressed in the supervision process, as Mary mentioned,

When I was highlighting [the textual elements], my feeling was that voice was not the concern of the student writers’ supervisors.

It can be seen that with their professional training and experience in academic writing, both Mary and Jane regarded voice construction important, but for students to successfully construct their voice in thesis writing, explicit guidance from their supervisors is indispensable.