Language Development for English-Medium Instruction: A Longitudinal Perspective on the Use of Cohesive Devices by Chinese English Majors in Argumentative Writing

Abstract

In alignment with a strong increase of fully or partially English-taught programmes worldwide, intense research interest has been drawn to English as a medium of instruction (EMI) in higher education. In the meanwhile, much research has been done concerning cohesion and coherence in second language writing, which reveals that the appropriate use of cohesive devices will ultimately lead to text coherence. This study adopts a longitudinal perspective to examine the use of cohesive devices by Chinese English majors in argumentative writing in the EMI context. Thirty-one second-year Chinese English majors participated in this study. They were given three writing prompts at the initial, middle, and final time-points of one academic year, respectively. Altogether, 93 argumentative writings were collected, and 30 cohesion indices were selected and investigated at local, global and text levels via two automatic computational tools: Coh-Metrix and TAACO. The major findings of the study indicate (1) Chinese English majors were capable of a variety of cohesive devices in their writing. They tended to use local cohesive devices, especially connectives, more than global and text cohesive devices; (2) Most of the cohesion indices demonstrated growth. Of the 30 selected cohesion indices, 13 presented significant development over one academic year, and among them, 3 were at local level, 10 at global level and 3 at text level. Informed of the possible learning trajectories of cohesive devices, teachers may be guided to focus on their teaching in some specific areas of cohesion, and students may be more aware of what contributes to their writing performance and what to emphasize when writing in English.

1. Introduction

Currently, English is widely used as a global language [1]. In fact, there is no overestimation in saying that English proficiency has become a most coveted form of cultural capital in Chinese society [2] (p. 564). Better education and English competence are increasingly conceived by emerging Chinese middle-class families as essential to their children securing a better future [3]. To guarantee a competitive edge, large numbers of Chinese families decide to invest heavily in their children’s learning of English, and many of them choose to send their children abroad, in particular to anglophone countries where education is delivered through the medium of English [4,5]. As a matter of fact, with the furthering of Englishization (the use of English in educational contexts where local languages were previously used), little can be done “to halt the express train of English Medium Instruction (EMI)” [6] (p. 300). In Hong Kong, starting from 2010, with the implementation of a “fine-tuning” compulsory language policy, some Chinese-medium schools have started to adopt EMI fully [7]. In the following year, the Chinese Ministry of Education issued a directive in which universities were urged to offer 5–10 percent of undergraduate courses in English (or other foreign languages) in three years [8]. Because of the perceived importance of English proficiency, the current Chinese national college English curriculum requires that college English instruction take up as much as 10 % of the total credit hours for undergraduate studies. At the national level, English is seen as having a crucial part to play in China’s ambitious development agenda to strengthen its innovation capacity, access cutting-edge knowledge available in English, enhance its competitive edge in international cultural and economic activity, and fully integrate into the world system [2].

As far as English learning is concerned, EMI has been shown to positively impact the development of language skills, namely listening, speaking, reading, and writing, regardless of the learners’ language proficiency [9]. Huang et al. [10] carried out an empirical study by way of a mixed-methods needs analysis to investigate the sustainability and success of English for nursing communication courses, affording a detailed picture of the various speech functions that realize engagement and how they are used at various stages of nurse-patient interactions. However, in the EMI context, students often find and report their linguistic limitations concerning vocabulary, writing, and speaking. For instance, in writing, learners often find particularly difficult the organization of essays and the use of appropriate academic style [11], as they are not used to writing extended texts in English. With the data collected from 356 students studying at eight English-medium secondary schools in Hong Kong, Pun and Jin [12] investigated the structural relationships between L2 students’ English proficiency, their use of language in the science classroom, their self-perceived difficulty in using English in the science classroom, their science learning self-concept, and their science achievement. The path analysis revealed strong connections between the variables, suggesting that L2 students’ self-perceptions and their English competence play important roles in their acquisition of scientific knowledge. Although L2 students’ English proficiency was viewed as a strongest predictor of science achievement, negative perceptions of EMI caused by lack of English competence also facilitate students’ academic outcome in science, to some extent. Besides, it was found that a mixed-language instruction appeared to be more beneficial than pure English-medium instruction when teaching science subjects. Following from this, the present study is intended to examine the sustainability and success in the learning and using of cohesive devices by Chinese English majors in argumentative writing in the EMI context, from a longitudinal perspective.

To the best of our knowledge, previous studies on cohesion, or rather the use of cohesive devices, are mostly cross-sectional, where the use of cohesive devices is examined across different grades or proficiency levels at a given point of time [13,14,15]. As proposed by Crossley et al. [16], longitudinal methods of data collection can help control for writer-specific variables (i.e., demographic differences), which is not possible in cross-sectional ones. Besides, previous studies are mostly focused on the analysis of some explicit types of cohesive devices such as logical connectors, which only touch upon surface cohesion [17,18,19]. In those studies, data are collected by means of manually marking and counting the number of cohesive devices in each writing, which involves a lot of time and effort. Moreover, deeper-level linguistic features are nearly impossible for researchers to recognize in this way. Therefore, it is of great necessity to probe deeper and draw a more complete picture by means of automatic tools like Coh-Metrix and TAACO, which make up for the deficiency of previous studies as well as making most of the work automatic.

The significance of this study is two-fold. On one hand, unlike most studies stopping at only manually analysing explicit cohesive devices, the use of advanced computational tools affords us a chance to dig deeper into the use of cohesive devices at three different levels, namely local, global and text levels in the EMI context. On the other hand, the longitudinal perspective adopted in this study renders it possible to track the changes in the use of cohesive devices by the same group of subjects. It is hoped that this one-academic-year longitudinal study may shed light on both EFL teaching and learning with regard to the use of cohesive devices.

1.1. Classification of Cohesive Devices

Halliday and Hasan [20] distinguished and described the domains of cohesive devices into grammatical ones and lexical ones as reference (e.g., reference by means of function in the speech situation through the category of person), substitution (e.g., one, ones, the same), ellipsis (e.g., ellipsis within the nominal group), conjunction (e.g., and, nor, or, furthermore, likewise, that is), and lexical cohesive devices (e.g., same word, synonym or near-synonym, superordinate and general word). Based on Halliday and Hasan’s taxonomy, Huang [21] concluded cohesive devices into grammatical device, lexical cohesion, logical connector (i.e., cohesive devices indicating logical meanings) and pragmatic and semantic implication. Likewise, Zhu et al. [22] made a distinction between grammatical cohesive devices and lexical ones. Grammatical cohesion is further divided into reference, ellipsis, substitution, and conjunction; lexical cohesion is further divided into repetition, synonymy/antonymy, hyponymy/superordinate, meaning, collocation. In the same vein, Zhang and Liu [23] further grouped cohesive devices into three parts: lexical grammar, pronunciation (pronunciation pattern) and intonation (tone group). Lexical grammar can be further divided into transitivity, tone, information structure, thematic structure, non-structural cohesion. Among them, the non-structural cohesion has five sub-categories: reference, substitution, ellipsis, conjunction, and lexical cohesion.

Nonetheless, when a larger number of longer writings are called for, such a method where people identify and mark cohesive devices takes too much time and effort. Therefore, it is of great necessity to develop an automatic tool, which can afford speed, flexibility, and reliability [18].

1.2. The Computational Tools of Coh-Metrix and TAACO

The computational tools help break through the limitation of analysing a text only at the surface level in the past [17]. Coh-Metrix was developed by McNamara et al. [24], which can make a comprehensive text analysis with the help of WorldNet (i.e., including semantic and syntactic features of nouns, verbs, adjectives, and other content words in the English language), MRC Psycholinguistic Database (i.e., storing human ratings of thousands of words on familiarity, imagery, concreteness, and meaningfulness), CELEX Lexical Database (i.e., estimating how frequently English words are used), etc. It is capable of quantifying surface as well as deep structure of a text by means of statistical algorithms such as Latent Semantic Analysis (LSA) which considers semantic overlap between explicit words and words that are implicitly similar or related in meaning [19]. Under the guidance of similar principles, TAACO, which was designed later by Crossley et al. [25], also has these advantages.

Generally speaking, Coh-Metrix focuses on local and text cohesion while TAACO reports on a greater number and variety of local (i.e., devices related to sentence level cohesion), global (i.e., devices related to cohesion between larger chunks of texts), and overall text (i.e., devices related to cohesion across an entire text) cohesive devices [26,27]. Moreover, the two computational tools are complementary. TAACO, as opposed to Coh-Metrix, provides a greater breadth of global cohesion indices. TAACO also offers synonym overlap indices and part of speech (POS) tagged cohesion indices. By contrast, Coh-Metrix uses Latent Semantic Analysis (LSA) to demonstrate semantic overlap. In other words, Coh-Metrix and TAACO calculate distinct indices while overlaps also exist. To be specific, Coh-Metrix measures semantic similarity, syntactic cohesion, causal cohesion, spatial cohesion, and temporal cohesion, whereas TAACO measures synonym overlap, lexical overlap, and giveness. Both tools measure connectives, lexical overlap, and lexical diversity. With the help of Coh-Metrix and TAACO, cohesive devices can be examined in a fairly comprehensive way.

1.3. Previous Studies in Relation to the Use of Cohesive Devices

Studies on the use of cohesive devices have been conducted among both L1 and L2 writers. For L1 writers, lexical cohesive devices are most frequently used. For example, Witte and Faigley [28] selected 10 out of 90 essays written by freshmen from University of Texas. Following Halliday and Hasan’s cohesion theory, they found the largest number of cohesive ties, about two-thirds of the total ties, fall into the general category of lexical cohesion. For L2 writers, though with different L1, similar findings are concluded using Halliday and Hasan’s taxonomy of cohesive devices and their framework for analysis, that is, lexical cohesive devices are most common. For example, Zhang [29] investigated the use of cohesive features in 107 writings by Chinese undergraduates and found that lexical devices were the most frequently used, followed by conjunctions and reference devices. Kim and Na [30] examined 57 writings produced by Korean EFL university students and found that the students used lexical devices most frequently, followed by reference and conjunction devices. Abdul Rahman [31] examined college-level Arabic L1 users’ command of cohesive devices and found that the students used certain types of cohesive devices most often, e.g., repetition (belongs to lexical cohesion), reference, and connectives.

Meanwhile, a large number of studies have investigated the changes in the use of cohesive devices in essay writings, though most research concerning the development of cohesive devices has been conducted for L1 writers than L2 writers, resulting in a paucity of available information about cohesion development in L2 learners [32]. For L1 writers, when they are young, they tend to use explicit cohesive devices that are local in nature to link sections of text together, e.g., referential pronouns and connectives [33]. However, around the 8th grade, developing writers begin to use fewer explicit cohesive devices to organize text [34]. This trend continues into high school and beyond. For adolescent and adult writers, the use of such explicit local cohesive devices is generally decreasing [35]. In other words, native writers experience a movement from explicit to implicit cohesion. Nonetheless, agreement has not been reached as regards the changes in the use of cohesive devices by L2 writers. Some researchers found a similar trend, that is, the use of explicit cohesive devices decreases as L2 writers become more skilled. For instance, Crossley and McNamara [14] examined 200 essays by Hong Kong high school students for the Hong Kong Advanced Level Examination (HKALE). Results showed that L2 writers categorized as highly proficient produce essays that are less apparently cohesive. Other researchers found that more proficient L2 writers used a greater number of connectives and thus produced more cohesive text. Yang and Sun [15] compared the argumentative writings of second and fourth-year Chinese undergraduates and found that more advanced learners used a greater number of cohesive devices and used them more accurately.

One point which is worthy of special note is that the use of cohesive devices develops separately from other writing skills. Berninger et al. [36] found children’s ability to connect sentences in paragraphs, which is largely by means of cohesive devices, was unrelated to their ability to write well-structured sentences or to spell well. They also found no relation among skills of lexical sophistication, sentence complexity, or cohesion within writing produced by the same individuals. Thus, there is a need to look into cohesive devices specifically. Besides, in order to help understand what characterizes L2 writing development and how it might be most effectively investigated and fostered, Manchón [37] recommended longitudinal designs, which can capture what is developed and how development is observed. Considering the possible interference of intergroup differences in cross-sectional studies, it is of great necessity to depict a fuller picture of the changes in the use of cohesive devices by the same group of individuals over a period of time. Therefore, the present study is intended to address the following two questions (1) What is the status quo of Chinese English majors’ use of cohesive devices in argumentative writing in the EMI context? (2) What are the changes in the use of cohesive devices by Chinese English majors in argumentative writing over one academic year?

2. Methodology

2.1. Setting and Participants

The study was positioned in a 1-year-long EMI Comprehensive English course, which was given to second-year English majors in a renowned comprehensive university in Jiangsu province, mainland China. The participants for this study were 31 Chinese sophomores majoring in English, aged between 20 to 22, and all with about 10 years of formal English learning experience. They came from an intact class in the School of Foreign Languages, but they were still kind of varied in terms English proficiency. The participants who come from cities or eastern parts of mainland China are comparatively more competent in English while those coming from rural areas or the inland provinces are relatively weaker in English. As a rule, for English majors in mainland China, the vast majority of courses, including both practical and theoretical ones, are given in English where students are in fact immersed in the English environment, at least in class.

2.2. Materials

For English majors in mainland China, they are supposed to take the TEM (Test for English Majors)-4 in their sophomore year, which is composed of listening, reading, and writing. The genres of writing in TEM-4 are either argumentative or expository. Over the academic year, the participants had a lot of writing practices in the two genres in the course of Comprehensive English, of which the argumentative wring was selected as materials for this study. To be specific, the materials used in this study consist of 93 argumentative writings by 31 Chinese sophomores majoring in English. The collection lasted one academic year, during which subjects were assigned three independent argumentative tasks of no fewer than 300 words within 50 min. The writings were collected at the beginning, the midpoint and the end of the academic year. Topics of these argumentative writings were determined based on daily life to ensure every participant could express his or her viewpoints (see

Table 1. Writing Prompts.

Initial Writing	Some people believe that the best way of learning about life is by listening to the advice of family and friends. Other people believe that the best way of learning about life is through personal experience. Which do you think is preferable? Use specific examples to support your preference. You should write at least 300 words.
Middle Writing	Some people claim that staying in a place all one’s life is good for one’s growth, while others disagree with the view, and they prefer changing the place. Discuss both views and then give your own opinion. Use specific reasons and details to support your answer. You should write at least 300 words.
Final Finding	Some people think that universities should provide students with the knowledge and skills needed in the workplace. Others think that the true function of a university should be to give access to knowledge for its own sake, regardless of whether the course is useful for an employee. Write an essay under the title “The Essence of University Education”. Give reasons for your answer and include any relevant examples from your knowledge or experience. Write about 300 words.

).

2.3. Cohesion Indices

Numerous cohesion indices are listed in Coh-Metrix and TAACO, among which overlaps exist. Moreover, both tools report some measures that are redundant with one another. Therefore, only the indices that measure unique cohesion features are selected under the premise of covering all three levels. To report on cohesion in argumentative writing, 30 cohesion indices from Coh-Metrix and TAACO were selected in this study. A sketch of all selected indices is shown in

Table 2. Cohesion features, categorization, tool, and examples.

Feature	Cohesion Types	Tool	Description	Example
Lexical overlap	Local and global	Coh-Metrix 3.0/TAACO 2.0.4	Overlap between nouns, arguments, stems and content and function words	The sun was high in the sky. The day was sunny.
Semantic similarity	Local, global and text	Coh-Metrix	LSA cosine values to measure similarity between text segments	The dog was hungry. So was the cat.
Connectives	Local	Coh-Metrix/TAACO	A number of theoretical and rhetorical lists of connectives	First, it was dark outside. Second, that man was not trustworthy.
Synonymy overlap	Local and global	TAACO	Overlap of synonyms across sentences and paragraphs	He kept a dog in his house. Every day he went home as early as possible.
Lexical diversity	Text	Coh-Metrix	Word repetition across a text	The little kid chased after the little dog.
Causal cohesion	Text	Coh-Metrix	Use of causal verbs and particles	He shouted because he was angry.
Temporal cohesion	Text	Coh-Metrix	Repetition of tense and aspect	The boy had breakfast and went to school.
Giveness	Text	TAACO	Ratio of pronouns to nouns; incidence of demonstratives	The man was satisfied he had that.

.

2.4. Data Analysis

As to the first research question concerning the status quo of Chinese English majors’ use of cohesive devices, descriptive statistics was done in terms of local cohesion, global cohesion, and text cohesion to get the means and standard deviation at the beginning of the academic year. As to the second research question regarding the changes in the use of cohesive devices by Chinese English majors, repeated measures ANOVAs were conducted to examine if significant differences in cohesive devices existed between initial, middle, and final compositions.

3. Results

3.1. The Use of Cohesive Devices by Chinese English Majors in Argumentative Writing

To investigate the use of cohesive devices by Chinese English majors in their argumentative writing, descriptive analysis of SPSS 16.0 was employed to get the means and standard deviation of 30 selected cohesion indices at the beginning of the academic year. The results are shown in

Table 3. Descriptive statistics of cohesion indices at the beginning of the academic year.

Index	Type	Tool	M	SD
Noun overlap, adjacent sentences, binary, mean	local	Coh-Metrix	0.28	0.13
Argument overlap, adjacent sentences, binary, mean	local	Coh-Metrix	0.57	0.16
Stem overlap, adjacent sentences, binary, mean	local	Coh-Metrix	0.35	0.15
Content word overlap, adjacent sentences, proportional, mean	local	Coh-Metrix	0.09	0.03
LSA overlap, adjacent sentences, mean	local	Coh-Metrix	0.12	0.04
All connectives incidence	local	Coh-Metrix	118.03	17.29
Causal connectives incidence	local	Coh-Metrix	32.98	11.03
Logical connectives incidence	local	Coh-Metrix	48.43	10.76
Adversative and contrastive connectives incidence	local	Coh-Metrix	17.37	11.97
Temporal connectives incidence	local	Coh-Metrix	19.00	8.21
All_positive	local	TAACO	0.08	0.01
All_negative	local	TAACO	0.02	0.01
Syn_overlap_sent_noun (adjacent sentence overlap noun synonyms)	local	TAACO	0.80	1.61
Syn_overlap_sent_verb (adjacent sentence overlap verb synonyms)	local	TAACO	0.85	2.02
LSA overlap, adjacent paragraphs, mean	global	Coh-Metrix	0.34	0.08
Adjacent_overlap_cw_para (adjacent paragraph overlap content lemmas)	global	TAACO	0.16	0.06
Adjacent_overlap_fw_para (adjacent paragraph overlap function lemmas)	global	TAACO	0.34	0.08
Adjacent_overlap_noun_para (adjacent paragraph overlap noun lemmas)	global	TAACO	0.20	0.08
Adjacent_overlap_verb_para (adjacent paragraph overlap verb lemmas)	global	TAACO	0.14	0.07
Adjacent_overlap_pronoun_para (adjacent paragraph overlap pronoun lemmas)	global	TAACO	0.28	0.14
Syn_overlap_para_noun (adjacent paragraph overlap noun synonyms)	global	TAACO	4.22	1.53
Syn_overlap_para_verb (adjacent paragraph overlap verb synonyms)	global	TAACO	4.29	3.47
LSA given/new, sentences, mean	text	Coh-Metrix	0.26	0.02
Lexical diversity, MTLD, all words	text	Coh-Metrix	79.08	13.76
Ratio of causal particles to causal verbs	text	Coh-Metrix	0.62	0.45
Temporal cohesion, tense and aspect repetition, mean	text	Coh-Metrix	0.82	0.08
Pronoun density	text	TAACO	0.05	0.14
Pronoun noun ratio	text	TAACO	0.24	0.09
Repeated content lemmas	text	TAACO	0.26	0.04
Repeated content and pronoun lemmas	text	TAACO	0.30	0.04

.

As can be seen in Table 3, all the three types of cohesive indices, namely, local, global, and text, are used. Among the 14 local cohesion indices, 4 belong to lexical overlap; 1 belongs to semantic similarity (between sentences); 7 belong to connectives; 2 belong to synonym overlap between sentences. Among the 8 global cohesion indices, 1 belongs to semantic similarity (between paragraphs); 5 belong to lexical overlap between paragraphs; 2 belong to synonym overlap between paragraphs. Among the 8 text cohesion indices, 4 are calculated by Coh-Metrix, and 4 others are giveness according to TAACO. It can be safely concluded that Chinese English majors are capable of using cohesive devices not only at the local level, but also at the global and text levels. Such a general finding confirms Liu’s [38] corpus-based study. Besides, the noticeably high standard deviations in connectives and lexical diversity suggest huge differences in vocabulary among students.

3.2. The Changes in the Use of Cohesive Devices by Chinese English Majors in Argumentative Writing over the Year

To probe into the changes in the use of cohesive devices by Chinese English majors in argumentative writing over one academic year, descriptive data of means and standard deviations of three time-points were first elicited and compared. Then repeated measures ANOVAs were conducted to obtain the overall growth patterns and ascertain possible significant differences in cohesive devices between initial and middle compositions, initial and final compositions, and between middle and final compositions. The results of repeated measures ANOVAs are displayed in

Table 4. Comparison of cohesion indices used by Chinese English majors in writing over the year.

Index	Initial		Middle		Final		F (2,60)	Result
	M	SD	M	SD	M	SD
Noun overlap, adjacent sentences, binary, mean	0.28	0.13	0.30	0.13	0.35	0.17	2.01	/
Argument overlap, adjacent sentences, binary, mean	0.57	0.16	0.50	0.15	0.51	0.15	2.66	/
Stem overlap, adjacent sentences, binary, mean	0.35	0.15	0.39	0.14	0.43	0.16	2.04	/
Content word overlap, adjacent sentences, proportional, mean	0.09	0.03	0.10	0.03	0.13	0.04	8.53 *	Final > Middle Final > Initial
LSA overlap, adjacent sentences, mean	0.12	0.04	0.17	0.05	0.20	0.05	27.92 *	Middle > Initial Final > Initial
All connectives incidence	118.03	17.29	117.21	14.12	125.90	18.07	1.90	/
Causal connectives incidence	32.98	11.03	33.39	9.41	31.09	10.37	1.02	/
Logical connectives incidence	48.43	10.76	50.47	11.41	53.31	11.64	1.64	/
Adversative and contrastive connectives incidence	17.37	11.97	17.92	7.42	23.63	8.27	4.65 *	Final > Initial Final > Middle
Temporal connectives incidence	19.00	8.21	14.97	5.75	17.26	7.42	2.25	/
All_positive	0.08	0.01	0.07	0.01	0.08	0.02	1.34	/
All_negative	0.02	0.01	0.01	0.01	0.03	0.01	2.26	/
Syn_overlap_sent_noun (adjacent sentence overlap noun synonyms)	0.80	1.61	0.61	0.39	0.70	0.33	0.34	/
Syn_overlap_sent_verb (adjacent sentence overlap verb synonyms)	0.85	2.02	0.57	0.55	0.66	0.33	1.07	/
LSA overlap, adjacent paragraphs, mean	0.34	0.08	0.35	0.10	0.48	0.10	23.97 *	Final > Middle Final > Initial
Adjacent_overlap_cw_para (adjacent paragraph overlap content lemmas)	0.16	0.06	0.19	0.05	0.20	0.04	6.07 *	Final > Initial Middle > Initial
Adjacent_overlap_fw_para (adjacent paragraph overlap function lemmas)	0.34	0.08	0.44	0.07	0.49	0.77	25.13 *	Middle > Initial Final > Middle Final > Initial
Adjacent_overlap_noun_para (adjacent paragraph overlap noun lemmas)	0.20	0.08	0.21	0.05	0.19	0.07	0.78	/
Adjacent_overlap_verb_para (adjacent paragraph overlap verb lemmas)	0.14	0.07	0.20	0.07	0.21	0.07	11.84 *	Middle > Initial Final > Initial
Adjacent_overlap_pronoun_para (adjacent paragraph overlap pronoun lemmas)	0.28	0.14	0.45	0.16	0.48	0.16	15.55 *	Middle > Initial Final > Initial
Syn_overlap_para_noun (adjacent paragraph overlap noun synonyms)	4.22	1.53	4.29	2.42	5.94	2.62	6.62 *	Final > Middle Final > Initial
Syn_overlap_para_verb (adjacent paragraph overlap verb synonyms)	4.29	3.47	6.31	3.32	7.20	3.32	7.49 *	Middle > Initial Final > Initial
LSA given/ new, sentences, mean	0.26	0.02	0.30	0.03	0.32	0.03	48.22 *	Middle > Initial Final > Middle Final > Initial
Lexical diversity, MTLD, all words	79.08	13.76	82.75	16.56	84.29	17.71	1.42	/
Ratio of causal particles to causal verbs	0.62	0.45	0.51	0.41	0.53	0.38	1.90	/
Temporal cohesion, tense and aspect repetition, mean	0.82	0.08	0.82	0.06	0.83	0.07	0.08	/
Pronoun density	0.05	0.01	0.05	0.02	0.05	0.01	0.93	/
Pronoun noun ratio	0.24	0.09	0.24	0.08	0.20	0.07	2.03	/
Repeated content lemmas	0.26	0.04	0.26	0.04	0.29	0.05	6.86 *	Final > Initial Final > Middle
Repeated content and pronoun lemmas	0.30	0.04	0.31	0.05	0.33	0.05	4.69 *	Final > Initial

* p < 0.05.

.

According to Table 4, 18 out of the 30 cohesion indices demonstrate consistent growth patterns while the other 12 do not show ups or downs over one academic year. Among the 18 indices, 6 are at local level, 7 are at global level, 5 are at text level. Moreover, the results of pairwise comparisons show that 13 of the 18 cohesion indices displaying growth demonstrate significant differences.

Of the 13 cohesion indices exhibiting significant changes over the year, three are local cohesion indices. To be specific, the score of content word overlap between adjacent sentences (F (2, 60) = 8.53, p < 0.05) is significantly higher in final compositions than in initial (MD = 0.04) and middle (MD = 0.03) compositions. There is no significant difference between initial and middle compositions (MD = 0.01). The score of LSA overlap between adjacent sentences (F (2, 60) = 27.92, p < 0.05) is significantly lower in initial compositions than in middle (MD = 0.05) and final (MD = 0.08) compositions. There is no significant difference between middle and final compositions (MD = 0.03). The score of adversative and contrastive connectives incidence (F (2, 60) = 4.65, p < 0.05) is significantly higher in final compositions that in initial (MD = 6.26) and middle ones (MD = 5.71). There is no significant difference between initial and middle compositions (MD = 0.55).

Seven are global cohesion indices. The score of LSA overlap between adjacent paragraphs (F (2, 60) = 23.97, p < 0.05) is significantly higher in final compositions than in initial (MD = 0.14) and middle (MD = 0.13) ones. There is no significant difference between initial and middle compositions (MD = 0.01). The score of adjacent paragraph overlap content lemmas (F (2, 60) = 6.07, p < 0.05) is significantly lower in initial compositions than in middle (MD = 0.03) and final (MD = 0.04) ones. There is no significant difference between middle and final compositions (MD = 0.01). The score of adjacent paragraph overlap function lemmas (F (2, 60) =25.13, p < 0.05) is significantly higher in final compositions than in initial (MD = 0.43) and middle (MD = 0.33) ones. It is also significantly higher in middle compositions than in initial ones (MD = 0.10). The score of adjacent paragraph overlap verb lemmas (F (2, 60) = 11.84, p < 0.05) is significantly lower in initial compositions than in middle (MD = 0.06) and final (MD = 0.07) ones. There is no significant difference between middle and final writings (MD = 0.01). The score of adjacent paragraph overlap pronoun lemmas (F (2, 60) =15.55, p < 0.05) is significantly lower in initial writings than in middle (MD = 0.17) and final (MD = 0.20) ones. There is no significant difference between middle and final writings (MD = 0.03). The score of adjacent paragraph overlap noun synonyms (F (2, 60) = 6.62, p < 0.05) is significantly higher in final writings than in initial (MD = 1.72) and middle (MD = 1.65) ones. There is no significant difference between initial and middle writings (MD = 0.07). The score of adjacent paragraph overlap verb synonyms (F (2, 60) = 7.49, p < 0.05) is significantly lower in initial writings than in middle (MD = 2.02) and final (MD = 2.91) ones. There is no significant difference between middle and final writings (MD = 0.89).

Three are text cohesion indices. The score of LSA given/new (F (2, 60) = 48.22, p < 0.05) is significantly higher in final compositions than in initial (MD = 0.06) and middle (MD = 0.02) ones. It is also significantly higher in middle compositions than in initial ones (MD = 0.04). The score of repeated content lemmas (F (2, 60) = 6.86, p < 0.05) is significantly higher in final compositions than in initial (MD = 1.72) and middle (MD = 1.65) ones. There is no significant difference between initial and middle compositions (MD = 0.07). The score of repeated content and pronoun lemmas (F (2, 60) = 4.69, p < 0.05) is significantly higher in final compositions than in initial ones (MD = 0.03). There is no significant difference between initial and middle compositions (MD = 0.01) or between middle and final compositions (MD = 0.02).

In all 13 indices, final compositions demonstrate significantly higher levels of cohesion indices than initial compositions. Final compositions demonstrate significantly higher levels than middle compositions and middle compositions demonstrate higher level that initial ones in 7 indices. In Crossley et al.’s one-term research [32], no significant differences were found between middle and final essays in terms of the production of cohesive devices. Thus, the importance of this longer study, which lasted one academic year, is demonstrated.

Generally speaking, over half of the cohesion indices, namely, 18 out of 30, are on the increase and the 13 indices displaying significant differences over the three time-points cover several subcategories of local, global and text cohesion indices. It can thus be inferred that the compositions by Chinese English majors demonstrate greater local, global and text cohesion over one academic year. This finding supports the study by Crossley et al. [32], where upper-level L2 learners performed better in final compositions during one semester. Furthermore, over half of the 13 indices are at paragraph level, demonstrating outstanding growth especially in global cohesion indices.

Among the 13 indices showing significant differences, the strongest growth is LSA given/new based on the effect size, indicating a large increase in text cohesion. LSA given/new is a unique measure in Coh-Metrix so it has been rarely used in most previous studies under the guidance of Halliday and Hasan’s Taxonomy. According to McNamara, “text constituents can be classified into three categories: given, partially given (based on various types of inferential availability), or not given (i.e., new)”. Two other LSA-based indices, LSA overlap between adjacent sentences and LSA overlap between adjacent paragraphs, have the second and fourth strongest growth. All of them measure semantic similarity which are implicit in nature. Evidence can also be found in Ji and Guo’s [39] study. They contrasted the cohesion devices used by Chinese and American writers in published papers from SCI source journals and found no significant difference between the three indices mentioned above. Therefore, it can be inferred that as Chinese writers’ proficiency increases, they are getting closer to native writers in achieving cohesion with semantic coreferentiality.

The third strongest variable is adjacent paragraph overlap function lemmas, which is global in nature. Function word generally refers to conjunction, preposition, pronoun and article. As is mentioned by Li [40], logical connectors demonstrate semantic relations between two or more sentences as well as between different parts in a text, and conjunction can thus be seen as most influential in terms of adjacent paragraphs among function lemmas. At local level, connectives are investigated as well, in more detailed classifications, though. Only adversative and contrastive connectives show significant development. It confirms Yang and Sun’s finding in a cross-sectional study that seniors employed significantly larger quantity of adversatives to illustrate arguments than sophomores did. It also illustrates that English majors are capable of functions of adversative connectives such as interpretation, extension, etc. other than opposition only [39]. On the whole, significant growth can be found in not only adversative and contrastive connectives between sentences but also in logical connectors between paragraphs. It’s noteworthy that most connectives, even all connectives incidence, demonstrate no significant increase. Crossley et al. [32] got the same finding that most connective indices did not show significant growth patterns, indicating that the L2 writers did not make obvious progress regarding explicit cohesive devices. Moreover, since the overuse of connectives does no good to coherence [17], English majors are likely to have recognized this problem, thus appropriately dealing with connectives.

In contrast to function word, significant growth can be seen in content word overlap at not only local level but also global and text level. Evidence comes from four indices: content word overlap between adjacent sentences, adjacent paragraph overlap content lemmas, repeated content lemmas and repeated content and pronoun lemmas. It is partly consistent with Crossley et al.’s [32] finding that writers increasingly repeated content words across a text (i.e., text cohesion). This study also found gains at the local and global levels, indicating that English majors are becoming increasingly skilled in repeating content words at the sentence, paragraph, and text level to achieve cohesion.

Four other significant indices are all at global level: adjacent paragraph overlap verb lemmas, adjacent paragraph overlap pronoun lemmas, adjacent paragraph overlap noun synonyms and adjacent paragraph overlap verb synonyms. It can be inferred that English majors improve their use of repetitions and synonyms between paragraphs, which is at a more implicit level. These findings are kind of different from those of Crossley et al. [32]. They found that L2 writers increasingly used noun synonyms between paragraphs which is consistent with this study; while the strongest growth was an increase in noun overlap between paragraphs which does not demonstrate significant trend in this study. Considering that Crossley et al. investigated descriptive writings and this study examined argumentative writing, it may be the genre that makes the difference, i.e., the former genre calls for repetitive nouns more often to describe vividly while the latter genre focuses more on arguing reasonably. This explanation can also be justified by the insignificant growth in noun overlap between adjacent sentences. In addition, in contrast to two out of three adjacent paragraph overlap synonyms which are included in 13 significant indices, no adjacent sentence overlap synonyms (i.e., adjacent sentence overlap noun synonyms, adjacent sentence overlap verb synonyms) show significant differences across three time points. It indicates that English majors make more remarkable progress in using global cohesive devices than in local ones.

4. Discussion and Implications

As regards the use of cohesive devices by Chinese English majors, among local cohesion indices, connectives are most frequently used, followed by synonym overlap. This result is not quite consistent with most previous studies, where reference is of highest frequency. This finding can be accounted for from two perspectives. Firstly, due to the adoption of different frameworks by researchers, varied results come out. Li [41], Gu [42], Guo [43] all found reference, one of the grammatical cohesive devices under the guidance of Halliday and Hasan’s taxonomy, to be most often employed; while in this study connectives, computed by means of computational tools, were found to be most often used. Secondly, the possible cause exists in the difference of subjects. Li [41] examined non-English majors, Gu [42] looked into senior high school students, Guo [43] investigated junior high school students, while this study focused on English majors who are “more proficient and competent in using English” [44]. In Yang and Sun [15], the fourth-year college students employed a significantly larger quantity of conjunctions whereas the second-year college students employed a remarkably larger number of references. Their findings that subjects of higher proficiency use more connectives and fewer references provide an effective explanation for this study. Since a majority of previous studies are based on Halliday and Hasan’s cohesion theory which is local in nature, not many touch upon global and text cohesive devices. At global level, the score of LSA between adjacent paragraphs is higher than LSA between adjacent sentences, indicating that a better performance is made at paragraph level with regard to LSA. It’s also worth mentioning that the score of function word overlap between adjacent paragraphs is over twice as that of content word. It can be inferred that English majors tend to use function words to achieve cohesion in their writing. These findings are consistent with those in Liu’s study [38]. As regards text cohesion indices, the ratio of LSA given information to new information and the scores of all the four indices related to giveness are relatively low. These indices are similar but complementary since giveness indices approximate the proportion of given information to new information by examining pronoun density, pronoun to noun ratios, and repeated content lemmas [45]. The finding here is consistent with what was concluded in Li [46] that proficient learners are more likely to use new words and information. However, excessive new information weakens the connections built by lexical co-occurrence between sentences as well as paragraphs, thus lowering text cohesion and coherence. It can be seen here that English majors are not skilful in balancing between given information and new information.

Concerning the changes in the use of cohesive devices by Chinese English majors, over half of all the selected 30 indices, namely, 17 out of 30, show either nonlinear trend or insignificant differences among the initial, middle, and final writings. Apart from most connective indices mentioned before, indices related to situation models, at text level, also deserve investigation. Crossley et al. [32] faced the same situation and they attributed it to genre and subjects. It may be the case that such links are unnecessary to argumentative writings or that L2 writers may not have yet developed the skills to create such links in their writing. It is also noticeable that, at local level, argument overlap between adjacent sentences and stem overlap between adjacent sentences do not show significant differences over one academic year, which can be explained in two ways. For one thing, Chinese is an isolating language while English is a fusional language [47]. Derivation is one of the three most productive English word formations, adding prefix or suffix to root. And both stem and argument can function as root, which is quite different from Chinese. For another, Chinese writers, as nonnative language users, are limited in terms of vocabulary depth and vocabulary breadth. Therefore, it is quite difficult for Chinese writers to master stem overlap and argument overlap as well as native writers [39].

The implications of this study are two-fold. On the one hand, the findings yielded in this study could help inform teachers of the possible learning trajectories of the use of cohesive devices in student writing in the EMI context. It may guide and improve their instruction in the specific areas of cohesion. For example, there is a need for teachers to avoid only focusing on the word and sentence levels but to go beyond to the paragraph and whole text level, assisting students in achieving unity as well as cohesion. On the other hand, in order to raise students’ awareness of cohesion, students should be encouraged to take full advantage of the EMI context, which can afford learners the most authentic situation and materials. For instance, apart from the EMI in classroom, texts written by native speakers can be used as another form of EMI to practice students’ critical and analytical comprehension, both in class and outside of class. After features of a reading material are learned and internalized, the students need to produce a writing modelled on the material. In this manner, the students are expected to make the best of EMI mode and become more sensitive to varied types of cohesive devices and more capable of creating writing of high cohesion.

5. Conclusions

Regarding the use of cohesive devices, this study demonstrates that Chinese English majors are capable of applying various cohesive devices in argumentative writing. They tend to use local cohesive devices, especially connectives, more than global and text cohesive devices. In reference to the changes in the use of cohesive devices over the year, a majority of the cohesion indices investigated exhibit growth and among them, thirteen present significant changes from the initial to the final writing over the year. At sentence level, local overlap, semantic similarity between sentences and connectives show significant differences over time. Most of the cohesion indices on the increase are found at paragraph level, indicating that the greatest increase over one year lies in global cohesion, including semantic similarity between paragraphs, lexical overlap between paragraphs, and synonym overlap between paragraphs. At text level for LSA given/new, giveness demonstrates significant growth trend over the year.

When interpreting the findings of this study, several limitations should be acknowledged. First of all, the investigated corpus in this study is composed of 93 compositions, which is not a sizeable sample. Future studies would benefit from a larger sample size. Moreover, this study only looks into argumentative writing. Comparing the present results with those yielded in previous ones, we find differences exist between different genres. For example, the strongest growth is found in LSA given/new in argumentative writing of this study while the greatest increase is found in noun overlap between paragraphs in descriptive writing of a previous study. Follow-up studies may examine other genres and make comparisons in order to conclude possible genre effects on the use of cohesive devices. In addition, although one academic year is not short for a longitudinal study, it is still desirable that future studies conduct similar research over a longer period of time. Finally, this study only focuses on one linguistic feature, cohesive devices. As Crossley et al. [32] put it, while cohesion is an important element of writing, it does not operate in isolation. Therefore, future studies may consider how cohesion interacts with other linguistic elements such as syntactic complexity and lexical sophistication in explaining growth and predicting writing performance.