This study deals with the analysis of syntactic complexity in professional academic writing and is based on a corpus of so-called 'hard' and 'soft' papers published in leading international journals. We aim at describing the main complexity features of academic discourse and testing the hypothesis that there is considerable disciplinary variation in linguistic complexity. We conclude that, first, clausal complexity strategies are more prevalent in the 'hard' sciences, while phrasal-complexity features dominate in the 'soft' ones. Second, the data reveal a continuum across subdisciplines within the broad categories of 'soft' and 'hard' genres with respect to the adoption of complexity strategies.
The phenomenon of complexity has been extensively approached in corpus linguistics over the
recent years. Specifically, the complexity of writing has been studied in terms of the comparison of
L2 and L1 writing [e.g. 1], correlations between text complexity, language proficiency and task types
[e.g. 2], and the development of text complexity after intensive instruction [e.g. 3]. However,
complexity in professional academic writing has been relatively under-researched to date despite the
potential pedagogical implications of such studies. In this respect, we contend that following the
linguistic conventions of a particular discipline plays a crucial role in identifying the writers as
experts in their own discourse communities [
The analysis of linguistic complexity in professional academic writing has been conducted on a 775,000-word corpus of research papers in four ‘soft’ arts and social sciences (business studies, linguistics, history and political science), and four ‘hard’ life and physical sciences (mathematics, engineering, chemistry and physics) which were published in leading peer-review journals indexed in Scopus Quartile 1, in 2016 and 2017. Once collected, the texts were manually cleared from tables, formulas, graphs, charts, metadata and reference lists for further analysis. The size and details of the corpus are given in Table 1.
Journals 16 18 13 17 64 10 10 10 11 41 97,947 95,852 98,430 99,003 391,232 95,350 95,603 99,303 93,366 383,622
Chem Physics Letters B (PL)
Reviews in Physics (RP)
Compositio Matematica (CM) The Journal of Differential Geometry (JDG)
Automatica (Auto) Materials Characterisation (MC)
Contemporary European History (CEH) The Journal of Modern History (JMH)
Political Analysis (PA) World Politics (WP)
In this study we undertake both the quantitative analysis of measures automatically generated by the complexity analyser and the qualitative scrutiny of a number of syntactic patterns associated with syntactic complexity. Firstly, to accomplish the quantitative analysis, the corpus texts were processed using Lu’s L2 Syntactic Complexity Analyser (hereafter L2SCA). L2SCA provided the 14 indices given in Table 2 along with their descriptions, as in Lu [6: 43]. Such indices were categorised into: (i) metrics of structural complexity: indices reporting the length of units (sentences, T-units, clauses2), measured by counting the number of words; (ii) metrics of syntactic complexity: indices reflecting syntactic depth and dependency, that is, those based on coordination and subordination ratios as well as on clausal/T-unit embedding within other superordinate units; and (iii) metrics of categorial complexity: indices expressing the pervasiveness of nominal and verbal categories in the text.
At the second stage of the analysis, we carried out the qualitative analysis of the clausal and the
phrasal complexity features, based on the taxonomy in Staples et al. [
The automated complexity indices are given in Table 3.
In an attempt to determine the relative weights of the complexity indices, a binomial linear regression analysis was applied to the data, implemented via the function ‘glm’ (‘stats’ package, R Core Team 2020). We operationalised a (backward-steps) reduction of the number of indices that led to the model in (1), with only the indices VPT (Verb phrases per T-unit), DCS (Dependent clause ratio), TS (T-unit/sentence ratio) and CPT (Coordinate phrases per T-unit). Both the C(oncordance) 0.918 and Nagelkerke R2 0.653 discrimination indices indicate that the model is very good at explaining the variation. (1)
Definitive glm model (‘***’: 0,001, ‘*’: 0,05)
Estimate Std, Error z value Pr(>|z|) (Intercept) -25,9115 4,0116 -6,459 1,05e-10 *** vpt 3,4756 1,0531 3,300 0,000966 *** dcs 10,6276 5,0567 2,102 0,035580 * ts 10,1416 2,6373 3,845 0,000120 *** cpt 3,8392 0,7312 5,250 1,52e-07 ***
The interpretation of the findings revealed by the statistical analysis of the complexity indices per broad discipline, that is, hard and soft sciences, is as follows. The reduction of the indices led to a model with only 4 indices evincing different dimensions of linguistic complexity:
(i) syntactic complexity mirrored by pervasive coordination, as reflected by the index CPT, which calculates the ratio of coordinated phrases per T-unit
(ii) syntactic complexity determined by subordination within clausal units, as evinced by the index DCC, which expresses the amount of subordinate dependent clauses in matrix clauses, and in sentences, which has been corroborated by the statistical significance of the index TS, a telling indicator of the ratio of T-units per sentence
(iii) categorial complexity associated with the frequency of, specifically, verbal constituents in Tunits, here captured by the index VPT.
Random Forests have demonstrated, on the one hand, that, out of the indices that proved to be very
strong in the model, those measures evincing complexity triggered by coordination (CPT) and by the
profusion of verbal categories (VPT), contribute to the variation of hard versus soft science to a
greater extent than DCC and TS. On the other hand, the probability of higher values in the four
complexity indices increases in academic writings categorised as soft science. In other words, greater
ratios of coordination, subordination and the ‘verby’ status of texts can be taken as proxies for the
categorisation of a research paper within the domain of social sciences and humanities. These results
are in line with Biber et al, [10: 29] when they claim that “complexity is not a single unified construct,
and it is therefore not reasonable to suppose that any single measure will adequately represent this
construct”. However, some remarks are in order here as regards the interpretation of our findings in
light of the conclusions drawn by Biber and colleagues. In their multidimensional analysis of
academic writing versus other more informal genres, Biber et al, [
Subsequently, a more qualitative analysis of the frequencies of the features associated with clausal and phrasal complexity was carried out. The results of the such an analysis are shown in Figure 2, which provides the normalised frequencies (per 100,000 words) of the features.
All the differences in the use of the complexity features in hard and in soft sciences were found to be statistically significant at the level of 1%, except that of verb+that-clauses, which was significant at the 5% level. As can be seen in Figure 2, adverbial clauses were found to be more common in the corpus of the hard-science papers. A closer look at the types of adverbial clauses extensively employed in life and physical sciences revealed that the most frequently used one is the conditional clause, which accounts for almost a third of all adverbial clauses. This type of adverbial clauses is typically used in the comments for various calculations, formulas and theorems (see example 1). As regards the two features evincing complementation strategies, wh-clauses prevail in the soft research papers, whereas that-clauses are more frequent in the hard disciplines. Finally, the data demonstrates that, overall, phrasal complexity features, particularly, adjectival and prepositional phrases prevail in the soft-science texts, while nominal categories are more frequent in the hard sciences, particularly in chemistry, where they are used in long names of chemical entities and processes (see example 2). (1) The next lemma expresses the important fact that if qC > 0 and if the excess measured relative to C is much smaller than the excess measured relative to pairs of planes with higher-dimensional axes… (JDG-2017-3). (2) In addition, methyliminodiacetic acid (MIDA)-protected boronate esters were well tolerated (Chem-2016-4)
This study has tackled the analysis of linguistic complexity in professional academic writing in English. The analysis of automated indices of complexity in a corpus of research articles published in leading journals in hard (mathematics, chemistry, physics, engineering) and soft (linguistics, history, business, political science) science papers led to the following conclusions. Soft sciences demonstrate a significantly larger number of features associated with syntactic complexity, subordination and coordination ratios than the hard-science genre. The data have also revealed that the clausalcomplexity indices, in particular, the occurrence of sentence-final adverbial clauses, are significantly more frequent in the corpus of the hard-science papers. Phrasal complexity, measured here by the amount of adjectival and prepositional phrases, proved to prevail in the soft-science category, whereas the hard-science texts exhibited greater ratios of nominal categories.
An in-depth description of linguistic complexity in professional academic texts, along the lines of
analyses of objectively depicted indices, can benefit the teaching of EAP/ESP writing in terms of
guiding the production of discipline-specific language-learning materials that will address the needs
of learners of different sciences in a more effective way. From the perspective of Data Driven
Learning (DDL) approaches [