Critical contingency competition in L2 clause positioning acquisition: the case of concessive clause by Chinese EFL learners

3.1.1 Corpora

To achieve comparability, we retrieved research data from the L1 and L2 argumentative essays. Specifically, native data were extracted from BAWE (the section which includes native argumentative essays from the Arts and Humanities disciplinary group; Nesi et al. 2008), LOCNESS (the section which includes argumentative essays; Granger 1998), and NESSIE (Xu 2012). Apart from a small section of the native corpora, which consists of British A-level argumentative essays (60,209 words out of the 1,882,981 word tokens), all other essays are of university level. Learner data were from university-level argumentative essays in WECCL (Wen et al. 2009) and TECCL (Xu 2016). The general corpora profile is listed in Table 4.

3.1.2 Data extraction

The data points we extracted are complete concessions composed of one matrix and one concessive adverbial clause each. A special operation was implemented on run-on and embedded concessive sentences.

Run-on sentences, also known as fused sentences, occur when two complete sentences are squashed together without a coordinating conjunction or proper punctuation (e.g., a period or semicolon). They exist in large quantities in Chinese learner corpora and less frequently in the native corpora of this research.

For run-on sentences with exceptionally long spans, the necessary segmentations were made at the boundary of complete concessive subordinations. In Example (12), only the bold part was extracted for the present dataset because it is both a part of a run-on sentence (missing coma before and) and a complete concessive sentence.

Concessive sentences embedded in larger complex sentences were extracted using the complete least principle. Complete principle refers to the extraction of complete concessive subordinations. Complex sentences embedded in larger complex sentences as complements, subjects, objects, coordinations, and adverbials were usually incorporated in the research datasets if they were complete sentences (bold part of Example 13). However, when a complex sentence acts as a componential part of a relative clause, it was not extracted due to its incompleteness as a sentence (bold part of Example 14).

According to the least principle, the words which introduce the embedded concessive sentences were not extracted. In other words, only concessive subordinations with a matrix and concessive adverbial clause constructions were extracted. Words such as that, and, but, where, and when were not included in the sentence boundary studied in the present research. Therefore, based on the complete least principle, the bold part of Example (13) without the introducing word so was extracted.

All the extracted data points are listed in Table 5. The concessive subordinations are distributed unevenly in the research corpora. Specifically, though-led subordinate clauses are few in number. To achieve balance, we then tagged the concessive clause positions of all though-led clauses and 1,000 randomly selected although-led clauses from the native and learner corpora, respectively. It was found that concessive clauses had strong tendencies to be preposed in both datasets (see Table 6). To explore the distinctions between preposed and postposed concessive clauses, we extracted research data based on the minimum count of the sub-clause positions – 339 native postposed clauses. Finally, we extracted 1,356 concessive subordinations, 339 native postposed clauses, and 339 randomly selected native preposed, learner preposed, and learner postposed data points.

The cue distribution of the 1,356 extracted datapoints is demonstrated in Table 7.

4.1.1 Effective cues on native speakers’ part

We first performed the drop1() function in R language to measure the effectiveness of the nine cues listed in Section 2 (see Table 8). P values and asterisks demonstrate cue effectiveness. Three, two, and one asterisks mark significance values (p values) smaller than 0.001, between 0.001 and 0.01, and between 0.01 and 0.05, respectively. The p values in Table 8 indicate that six cues (spc.clau.rel, sub.main.ratio, ini.adv, sub.clau, subordinator and bridging) are effective (deranking being moderately effective) in tuning the concessive clause positions on the native speakers’ part, but two cues (embeddedness and hedging) are ineffective.

We then fed the seven effective cues to a binary logistic regression model via the function glm(), which both checked the drop1() test results and yielded the tendency of the effective cues in placing the concessive adverbial clauses (preposed or postposed). Table 9 reports the results of the logistic regression model.

Coefficient values in the second column of Table 9 show the tendency of independent variable levels (e.g., subordinatorthough) to co-occur with the non-referent level of the dependent variable (here, the preposed position of the concessive adverbial clause). The positive/negative values show the attraction/repulsion of the independent variable levels to/from the preposed positions. For example, the coefficient value of subordinator though is negative (−0.9303), indicating that compared to although, though has fewer chances of leading the preposed concessive clauses than the postposed ones. It must be noted that the absolute coefficient values only demonstrate comparative power between certain independent variable levels when the other independent variables are controlled at their referent levels; thus, they do not reveal the absolute power of the variables compared to other independent variables.

On the whole, the polarity values of the coefficients suggest that the preposed positions would most frequently co-occur with the subordinator although, anaphoric terms in adverbial clauses, indirect concessive clausal relationships, shorter and simpler adverbial clauses, non-initial adverbials, and complete (non-deranked) adverbial clauses.

Then, we compared the test results to similar multifactorial studies on concessive clause positioning – Wiechmann and Kerz (2013) and Kang and Xu (2021). The former studied 2,000 although- and whereas-led concessions from the BNC, while the latter examined 1,738 although-led concessions from the BAWE. These two studies have investigated five and eight of the nine cues examined in the present study (see Table 10).

We found that six cues are comparatively stable across studies: subordinator, bridging, clausal relationships, sub-clause length, smaller clauses in concessive sub-clauses, and initial adverbials. These cues have identical co-occurrence tendencies with the concessive clause positions based on various datasets.

Conversely, three cues perform unstably across studies: hedging terms in the adverbial clauses, deranking status of the adverbial clauses, and embeddedness of the whole concessive sentence. For different datasets, these three cues vary either in their effectiveness or their tendency to co-occur with concessive clause positions.

4.1.2 Effective cue ranking on native speakers’ part

We then measured the power ranking of the seven effective cues (six stable and one unstable) via random forest modelling. Specifically, we performed the function randomForest (), with parameters of 1,500 tress and Mty = 3. The Auc value was 0.749, suggesting a satisfactory performance of the model (Figure 1).

Figure 1:

Cue importance in concessive clause positioning – native data.

The cues form roughly into three clusters – the Strong Group with spc.clau.rel assuming the utmost importance, the Mediocre Group with ini.adv, sub.clau, subordinator and sub.main.ratio being moderately important, and the Weak Group with deranking and bridging being very weak. The five stable and effective cues roughly echo Kang and Xu’s (2021) finding that native speakers focus more on meaning expressing than on easy processing in placing although-led clauses.

The superior power of inter-clausal relationships in tuning concessive clause positions implies native speakers’ priority on clause positioning based on clausal relationships. That is, they most likely tend to prepose adverse conditions to strengthen the main clauses and add restrictive clauses to soften them.

The Mediocre Group cues imply possible principles that people observe. The rankings of sub.clau and sub.main.ratio follow inter-clausal relationships, indicating a secondary focus on observing the end-weight principle (Quirk et al. 1972: 766) by native speakers – placing heavier clauses with sub-clauses and more words in the sentence final parts. The possible reasons for the end-weight principle are that (1) people have more time and space to plan for the end-positioned constituent of a construction when it is more complex, longer, or fuller (Arnold et al. 2000); (2) English is a right-branching language, which means it is easier to parse complex sub-clauses when they follow the main clauses than when they precede them (Hawkins 1994, 2005]; Quirk et al. 1985: 1037–1039). Cue ini.adv ranks third, suggesting native speakers’ tendency to observe the horror aequi principle, which refers to language users’ reluctance to employ linguistic components that are similar in form or meaning in nearby positions (McCawley 1988; Rohdenburg 2003). Preposed adverbial clauses and sentence-initial adverbials consisting of words and phrases bear similar functions, namely connecting sentences, setting the background or tone of the propositions, or intensifying the writer’s intentions (Biber et al. 1999; Quirk et al. 1985; Talmy 2008; Virtanen 2004). According to the principle of horror aequi, adverbial clauses are usually postposed in complex sentences with initial adverbials to avoid repetition. Cues subordinator reveals native speakers’ sensitivity to morphological-lexical changes.

The Weak Group consists of the stable cue bridging and the unstable cue deranking. The former suggests the native speakers’ tendency to maintain coherence by referring to the preceding clauses, although the tendency is quite weak. The latter implies that native speakers have a weak tendency to prepose deranked concessive clauses.

4.2.1 Effective cues and their ranking on learners’ part

On the learners’ dataset, we also performed drop1 (), logistic regression, and random forest models, yielding cue effectiveness in Table 11, cues’ tendency of co-occurring with concessive clause positions in Table 12, and the cue power ranking in terms of their importance values in Figure 2.

Figure 2:

Cue importance in concessive clause positioning – learner data.

Tables 11 and 12 demonstrate that, compared to native speakers, learners react to fewer cues (six vs. seven), some of which perform distinctly from those on the native dataset. Specifically, only four cues (spc.clau.rel, ini.adv, subordinator and sub.clau) tune the concessive clauses in the same direction as native speakers; one cue (sub.main.ratio) tunes the concessive clauses opposite to native speakers, and one cue (embeddedness) that is ineffective on the native dataset takes effect on the learner dataset.

In addition, the effective cues on the learner dataset are less predictive than those on the native dataset (0.71 vs. 0.8), which runs contrary to Wiechmann and Kerz (2014) (0.84 vs. 0.76). It must be noted that the two studies are based on different datasets with distinct subordinators and learner native language backgrounds, which might contribute to different model predictions.

Random forest modelling (with the parameters of 1,500 tress and Mty being 3, and an Auc value of 0.706) generated cue importance in terms of mean decrease accuracy values (see Figure 2). Generally, the six effective cues are not distinguished as sharply as those on the native dataset. Nevertheless, similar to native dataset cues, they fall roughly into three categories – the Strong Group consisting of spc.clau.rel and ini.adv, the Mediocre Group consisting of sub.main.ratio, subordinator and embeddedness, and the Weak Group consisting of sub.clau. It is worth noting that in addition to the tendency distinction on tuning concessive clause positions (by sub.main.ratio) and effectiveness (embeddedness), cue ranking also changes across datasets. That is, ini.adv ascends from the Mediocre Group in the native dataset to the Strong Group in the learner dataset, while sub.clau descends from the Mediocre Group in the native dataset to the Weak Group in the learner dataset (see Tables 13 and 14).

Although learners are sensitive to six cues, only one fewer than the native speakers, the effective cues vary greatly from those on the native speakers’ part. Table 14 shows that only two cues, spc.clau.rel and subordinator, are consistent in both the tuning tendency and power groups, and the remaining four cues perform distinctively in the native and learners’ datasets. This corresponds to previous studies’ finding that learners react to fewer cues than native speakers (Bates and MacWhinney 1987; Ellis 2006), lay different focus on cues that are effective on native speakers’ datasets (Wiechmann and Kerz 2014), and may create their own features (learner creativity) in the language acquisition process (Deshors et al. 2018), which may contribute to the complex language acquisition system.

As has been reviewed in Section 1, the complex language acquisition system relates to contingency, salience of the cues, as well as L1 impact in terms of preemption and inhibition (Ellis 2006; Larsen-Freeman 1997, 2002], 2012], 2019]), and we will thoroughly explore the contribution of contingency, salience, and L1 transfer to the contrasted cue performances between native speakers and learners.

4.2.2 Contingency effects

Contingency, or the reliability of form-meaning mapping, has long been argued to be the driving force in language learning (Ellis 2006; Ellis et al. 2016; Shanks 1995). Specifically, it is the co-occurrence possibility of the examined outcome and cue minus the possibility of the outcome in the context without the examined cue, the ∆P value (see Table 15 and the formula).

Studies on contingency effects on language learning have mostly focused on lexical acquisition, such as the acquisition order of VAC (verb-argument constructions, Ellis et al. 2016), its competition with availability, and formulaicity in lexical acquisition (Murakami and Ellis 2022). So far, little attention has been directed to the contingency of syntactic alternation and cues. In this research, we calculated the contingency values of the seven effective cues and the concessive clause positioning on native speakers’ part.

Contingency statistics were performed on a naturally distributed dataset. The native dataset on which we based the logistic regression and random forest modelling was extracted on a clause positioning basis. That is, we incorporated equal quantities of datapoints with preposed and postposed concessive clauses to explore cues that tune concessive clause positions. However, contingency must be calculated in a natural language use environment; therefore, we adjusted the former dataset to a naturally distributed one. Among all the native although-led concessions (1,397), we annotated 1,000 data points, and the annotation rate was 0.7158. Based on this rate, we annotated 381 though-led concessions among all 532 native though-led concessions (381 = 532 × 0.7158). In total, we annotated 1,000 although-led and 381 though-led concessions on a natural distribution basis. Table 16 lists the co-occurrence counts of the cues and the concessive adverbial clause positions. The ∆P values of the preposed position and cue conditions are listed in the last column. One exception is the ∆P value of meta-discursive causal relationship with the postposed position because spc.clau.rel includes three levels. Compared to the reference level of the direct causal relationship, the indirect causal relationship prefers the preposed position, while the meta-discursive causal relationship prefers the postposed position. The calculation of the contingency value of the indirect clausal relationship and the preposed position is listed as an example of ∆P calculation.

The ranking of contingency values is incorporated in Table 13 for comparison between the ranking groups by native speakers and learners. We observed categorization by contingency values. Cues with the two biggest values (spc.clau.rel and ini.adv) correspond exactly to the learners’ cue ranking; cues with the mediocre values (sub.clau and subordinator), though not corresponding to the learners’ cue ranking, tune the learners’ concessive clause positioning to the same tendency as that of the native speakers; cues with the lowest values (bridging, sub.main.ratio and deranking) either change in their effectiveness (bridging and deranking) or in their tuning tendency (sub.main.ratio). This categorization implies that, in our case, contingency assumes an important role in tuning learners’ concessive clauses.

4.2.3 L1 transfer

L1 transfer hinders L2 acquisition when L1 and L2 distinctions exist in concepts or language structures (Ellis 2006; Ellis et al. 2016; Jarvis 2000; Jarvis and Pavlenko 2008; Larsen-Freeman 2019; MacWhinney 2008). In our research, L1 transfer is related to five cues: spc.clau.rel, sub.clau, sub.main.ratio, bridging, and deranking.

Distinction in performing concessive softening functions between Chinese and English prevents Chinese L2 learners from using meta-discursive concessions to restrict the main clause propositions. In Chinese, concessions are mostly realized by complex sentences led by 虽然 (sui1ran2, meaning although or though), 虽(sui1, meaning although or though) or 虽然……但是(sui1ran2…dan4shi4, functioning as although or though construction) (Li 1992: 222–223; Xing 2001: 47, 467). Clauses led by 虽然 (sui1ran2, meaning although or though) typically precede the matrixes (Li and Thompson 1981: 632–635; Wu 2015: 241). The clauses that convey a softening or weakening effect are often introduced by 但是 (dan4shi4, meaning but), as noted by Xing (2001]: 293), or 只是 (zhi3shi4, meaning just that), as mentioned by Li (1992]: 209), and these usually follow the main clauses. The infrequent application of 虽然(sui1ran2, meaning although or though) in softening functions may prevent Chinese L2 learners from constructing meta-discursive concessions. It turned out that learners not only produced meta-discursive concessions infrequently (less than 1/3 those of the than native speakers, 54 vs. 174), but they also responded inactively in placing the concessive clauses to the sentence end positions in meta-discursive concessions.

Distinction in sentence construction patterns imposes tolls on observing the end-weight principle and producing deranked clauses by Chinese learners. Specifically, this distinction leads to different cue performances of sub.clau, sub.main.ratio, and deranking. While English native speakers observe the end-weight principle, the Chinese language does not have that feature. In fact, Chinese features run-on (flowing) sentences, which do not show a particular weight distribution (Shen 2012; Wang 2019). In addition, Chinese modifiers (such as attributives and adverbials) all have positions before the modified components (such as noun components and predicates) (Li and Thompson 1981: 104, 320–322), which would most likely change the end-weight-related cue performance – sub.clau and sub.main.ratio. This was verified by our research results that the cue sub.clau became weaker in dragging the heavy clauses to the sentence end position, while sub.main.ratio performed opposite to that in the native dataset. Chinese lacks strict morphological alterations, as discussed by Lv (1979), Ma (1983 [1898]), Li (1992]: 17), and Lu (2003]: 6). Specifically, it does not possess reflexive transformations, as noted by Karlgren (2017 [1949]: 59). This typological distinction can pose challenges for Chinese students. They may find it difficult to employ full predicates in the “verb+ participle (in reflexive form)” structure and even more so when navigating the less common deranked clauses compared to complete ones. In our research, deranked clauses produced by learners are only 18, compared to the 78 produced by native speakers (see Table 7).

Distinctions in cohesion patterns hinder Chinese learners from bridging close clauses. English cohesion is usually maintained via theme-rheme progression in a flow fashion. However, Chinese anaphoric cohesion frequently covers longer spans than between neighboring clauses, which is referred to as the sprung type of cohesion in Yang and Wang (2017). This contradiction in cohesion maintenance may potentially contribute to the ineffectiveness of the preceding clause referring to anaphoric items in tuning concessive clauses. Our research findings run contrary to two similar studies by Rezaee and Golparvar (2017) and Wiechmann and Kerz (2014), which reported the effectiveness of cue bridging in tuning concessive clause positions by L2 learners. The contradictions may be caused by different learner L1 backgrounds as they examined L2 learners with L1 backgrounds of Indo-European languages (Persian and German), while our research investigated learners with a Chinese language background, which belongs to the Sino-Tibetan language family.

4.2.4 Salience

Perceptually salient cues prime the related outcomes more strongly than non-salient ones (Ellis 2006; Larsen-Freeman 2019), such as temporal adverbials, which are more effective cues in constructing a past event construal than verbs in past tense forms. In our research, salient cues are categorized into two types: morphologically salient cues and syntactically salient cues. The former feature easy perception in their morphological forms, such as ini.adv and subordinator, while the latter can be identified following certain syntactic rules. For example, cues sub.clau, bridging and embeddedness could be perceived by word items such as that, which, he, or anaphoric content words.

4.2.5 Competition among contingency, L1 transfer, and salience

Cue power consistency in terms of tendency and ranking consistency across datasets reveals how close learners’ concessive clause positioning is to that by native speakers. It has been reported to be related to contingency values, eroded by L1 transfer, and promoted by cue salience (Ellis 2006, 2008]; Larsen-Freeman 2019), which is demonstrated by the formula below.

We listed the contingency values, salience status, and L1 transfer status in Table 17 based on the analysis in the previous sections (Sections 4.2.2–4.2.4) for a thorough observation of their specific relationships. It must be noted that for clarity, contingency of spc.clau.rel is set as the mean value (0.663) of its two possible conditions: indirect (0.509) and meta-discursive (0.816).

Table 17:

Competition among contingency, salience, and L1 transfer.

1. TC = tendency consistency; RC = ranking consistency.

It can be seen that contingency values, compared to Salience and L1 transfer, play a baseline role in cue power consistency across datasets. Cues with high contingency values tend to exert similar effects on learner and native outcomes and vice versa. A critical line classifies cues into two groups: those in or not in power consistency with the native dataset. Above this line, L1 negative effects do not lead learners to produce distinct outcomes from native speakers. Below this line, L1 negative effects usually overshadow contingency effects and cue salience, thus driving learners to perform distinctively from native speakers.

The two cues spc.clau.rel and ini.adv with the two largest contingency values are consistent across datasets both in tendency and power ranking. Although L1 transfer exists on spc.clau.rel, it only affects learners in producing meta-discursive concessions and thus tuning concessive clause positions. Learners are highly sensitive to preposing indirect concessive clauses to the sentence’s initial position, which is identical in the case of native speakers. Cue ini.adv, free from L1 transfer and with morphological salience, reveals that horror aequi principle underlies both learners and native speakers in concessive clause positioning.

The two cues sub.clau and sub.main.ratio are both subjected to the L1 transfer of insensitivity to the end-weight principle; however, cue sub.clau is consistent, while sub.main.ratio is not. In addition to the larger contingency values of sub.clau (0.192 vs. 0.074), the salient grammatical markers of sub.clau might also provide clues for learners to follow when placing concessive clauses appropriately.

The cue subordinator’s comparatively large contingency value and salience status assist learners in performing similarly to native speakers on this cue.

The four inconsistent cues (ineffective or performing opposite to those on the native dataset) have smaller contingency values, and three of them are also subjected to L1 transfer. This explains, to some extent, the findings by which learners react to fewer cues than native speakers in language production (Bates and MacWhinney 1987; Ellis 2006; Wiechmann and Kerz 2014).

One interesting finding in our research is that the cue embeddedness, not valued by native speakers, influences learners’ concessive clause positioning. The tuning tendency of this cue indicates that Chinese EFL learners would postpose concessive clauses when the concessive complex sentence is embedded in a larger sentence (see Example 15).

This would usually arouse ambiguity as to what the matrix of the postposed concessive clause is – the immediately preceding clause or the still larger one that incorporates the preceding clause – and is thus often avoided by native speakers (Kang and Xu 2021; Quirk et al. 1985, although this cue is not effective on the native dataset in this research). What force then drives learners to place the embedded concessive clauses at the sentence end position at the cost of arousing ambiguity? Embedded complex sentences are usually complements such as findings, statements, or propositions. Considering learners’ language proficiency, it may be speculated that they are eager to put forward the cores of the complex concessions – the matrixes – instead of the concessive conditions, to guarantee the communication of the propositional meaning. This unique feature of learners is not caused by L1 transfer and is discussed as learner creativity in Deshors et al. (2018).

Overall, in our case of concessive clause positioning, we concluded a critical contingency competition pattern among factors of contingency, salience, and L1 transfer. Specifically, (1) contingency exerts baseline effects on learners’ language production; (2) consistent concessive clause positioning by learners and native speakers is boosted by salience but hindered by L1 transfer; (3) a critical contingency value marks the border of cue consistency between learners and the native speakers in terms of tuning tendency and power ranking. Cues with contingency values below the critical value are quite vulnerable to L1 transfer in language production; (4) learners may also have unique performance free from the L1 transfer influence, such as postposing concessive clauses when the concessions are embedded in larger complex sentences.