Syntactic network, constructional relations, and L1 and L2 representations

Syntactic constructions are interrelated as an associative network, and each of them is situated in a particular position determined by its relationship to other relevant constructions (Diessel, Reference Diessel2019, p. 200). On the one hand, constructions that belong to the same semantic category are categorized as one constructional family (Goldberg & Jackendoff, Reference Goldberg and Jackendoff2004); on the other hand, “wherever there is a difference in form in a language, there is a difference in meaning” (Clark, Reference Clark and MacWhinney1987, p. 1). Members of a syntactic family usually exhibit a great deal of idiosyncrasy.

In Mandarin Chinese, for instance, a variety of syntactic constructions are available to express a causative event such as “The older sister kicked the door broken,” including the Ba- (1a), Bei- (1b), topic-comment constructions (1c), and the basic SVO structure (1d). However, their syntactic features entail different pragmatic priorities due to the intrinsic relationship between form and function (Hoffmann & Trousdale, Reference Hoffmann, Trousdale, Hoffmann and Trousdale2013). Specifically, they use different word orders (OV in 1a-c vs. canonical VO in 1d) and function words (把 bǎ and 被 bèi in 1a and 1b, respectively) to encode grammatical relations (Tai, Reference Tai and Haiman1985). As a result, compared to 1d, 1a-c have a contrastive/emphatic interpretation (Sun & Givón, Reference Sun and Givón1985) and highlight the causative result, as their VP takes up the sentence-final position (Xu, Reference Xu2004). They further differ in the relative salience of referents in the expression of the causative event, owing to the presence or absence of the causer (1a and 1b vs. 1c) and the placement of the causer and causee (1a vs. 1b) (Shi, Reference Shi2019).

It is worth noting that while native speakers can recognize this kind of dynamic relationship between the constructions in a constructional family, L2 learners are less aware of their subtle differences, particularly when there are discrepancies between L1 and L2. Consequently, native speakers’ usage is usually characterized by asymmetric statistical distributions (Ellis, Reference Ellis, Pütz and Sicola2010) and partial productivities (Goldberg, Reference Goldberg2016), whereas L2 learners are generally less sensitive to competing alternatives during comprehension (Robenalt & Goldberg, Reference Robenalt and Goldberg2016) and tend to produce them interchangeably without any discrimination (Teixeira, Reference Teixeira2020).

L1 and L2 acquisition of Mandarin Ba-construction and its SVO counterpart

The Mandarin Ba-construction and its SVO counterpart belong to the same syntactic family and both express the same semantic causation. The different usage patterns between native speakers and L2 learners have been repeatedly observed. Native speakers tend to show syntactic preference in specific contexts, whereas most L2 learners seem to use them interchangeably.

In over thirty years of investigations, L1 studies have discovered that L1 children are particularly receptive to the Ba-construction (Erbaugh, Reference Erbaugh1982). They start to produce Ba sentences as early as two years old and reach adult-like capacity by the age of five (Tsung & Gong, Reference Tsung and Gong2020). Their use of the Ba-construction is regulated by its semantic and syntactic constraints (Deng, Mai, & Yip, Reference Deng, Mai and Yip2017; Li, Reference Li1993), and it is highly sensitive to pragmatic factors (Chang & Zheng, Reference Chang and Zheng2017; Yang & Xiao, Reference Yang and Xiao2008).

More recently, research has started to pay increasingly more attention to the interaction between the Ba-construction and its competing alternatives, particularly with regard to its SVO counterpart. Zhang (Reference Zhang2020), for instance, examined under what conditions SVO sentences in the classical version of Diary of An Envoy to Britain were converted into Ba sentences in the vernacular version of the same book. The results, which supported observations made by several researchers in the field, showed that the Ba-construction was commonly employed when (i) the communicative intention focused on the resultant state of the predicate, for example, the focus of the predicate of sentence (2) below was on the complement 完 wán “finish” (Sun, Reference Sun, Wang and Sun2015); (ii) the VP was structurally complex and semantically bounded; for instance, Liu (Reference Liu1997) stated that there must be some element (e.g., 完 wán “finish” in (2)) other than the basic verb (e.g., 看 kàn “read” in (2)) in the predicate of the Ba-construction to indicate the terminal point or resultative state of an event; (iii) subjectivity was expressed (Jing-Schmidt & Tao, Reference Jing-Schmidt and Tao2009), such as in sentence (2), where the imperative sentence was used for a command or request, which inherently involved the speaker’s attitude.

Most L2 studies reveal the divergent use of the Ba-construction by L2 learners. Since most world languages do not have corresponding constructions, the Ba-construction presents a seemingly universal challenge to L2 learners. Unlike native speakers, L2 learners are less sensitive to the constraints of the Ba-construction, which results in either erroneous Ba productions (Liu & Wang, Reference Liu and Wang2003; Wen, Reference Wen2012) or non-target-like usage patterns, either avoiding the Ba-construction (Du, Reference Du and Tao2016; Yu, Reference Yu2000) or overusing it (Xu, Reference Xu2012; Zhang, Reference Zhang2010). Importantly, L2 non-target-like use of the Ba-construction persists even to advanced levels. For instance, Du (Reference Du, Everson and Shen2010) used a picture description task to collect written productions by L2 learners at three proficiency levels and found that all L2 learners, regardless of their proficiency, tended to substitute the Ba-construction with its SVO counterpart.

L2 learners’ non-target-like usage patterns indicate the general difficulty of L2 acquisition of the Ba-construction, especially when the L1 does not feature an equivalent structure. Such difficulty has conventionally been attributed to its complex linguistic characteristics (See Wen, Reference Wen2012 for a summary), but this fails to explain the differences between L1 and L2 acquisition, considering that linguistic complexity is something both native speakers and L2 learners have to deal with.

A complementary view, which takes learners’ cognitive aspects into account, has recently started to emerge. Hou (Reference Hou2021) proposed that, for L2 learners whose L1 does not have the equivalent of the Ba-construction, the acquisition of the Ba-construction is subject to a blocking effect from competing constructions. Utilizing a design that embedded an overshadowing treatment in a construction learning paradigm, this study asked beginning-level L2 learners without the equivalent of the Ba-construction in their native language to learn the Ba-construction and its SVO counterpart simultaneously. It was found that although they were able to quickly notice the semantic overlap between the Ba-construction and its SVO counterpart, L2 learners exhibited a reduced learning of the Ba-construction due to its unfamiliarity. They not only tended to replace the Ba-construction with its SVO counterpart but also made more mistakes when producing Ba sentences. The present study intends to further explore the blocking effect in L2 acquisition of the Ba-construction and examines whether or not the blocking effect can be identified in later stages of acquisition and how it manifests at the level of syntactic representations.

Blocking and L2 morphological acquisition

Ellis and colleagues found the acquisition of L2 morphological inflections to be particularly susceptible to blocking. They conducted a series of investigations to provide evidence of how L2 acquisition of verbal inflections was blocked by lexical adverbs and to demonstrate how L2 morphological acquisition was filtered through the lens of the native language (Cintrón-Valentín & Ellis, Reference Cintrón-Valentín, Ellis, Blumenthal-Dramé, Hanulíková and Kortmann2017; Ellis et al., Reference Ellis, Hafeez, Martin, Chen, Boland and Sagarra2014; Ellis & Sagarra, Reference Ellis and Sagarra2010a, Reference Ellis and Sagarra2010b, Reference Ellis and Sagarra2011; Sagarra & Ellis, Reference Sagarra and Ellis2013).

For example, Ellis & Sagarra (Reference Ellis and Sagarra2010a, experiment 2) asked Chinese and English native speakers to first learn congruent Latin adverb-verb combinations and their temporal reference (e.g., hodie cogito “today I think”). They then instructed them to judge whether an adverb-verb combination referred to the past, present, or future, including incongruent ones (e.g., hodie cogitavi “today I thought”). The results showed that Chinese native speakers overwhelmingly relied on adverbs, while English native speakers used both cues in a more balanced fashion. The results indicated that morphological blocking was dependent on the bias shaped by L1. Due to the absence of verb morphology in their native language, Chinese native speakers were less able than English speakers to acquire (hence rely on) Latin verbal inflection.

Ellis and Sagarra (Reference Ellis and Sagarra2011) replicated the investigation and extended it to four L1 groups. They found that successful acquisition of Latin verbal inflections correlated with the L1. Spanish learners were the most reliant on verbs, followed by Russian learners, English learners, and Chinese learners (who were the least reliant on verbs). The results thus suggested that L2 learners’ sensitivity to Latin tense morphology depends on the degree to which the L1 makes use of morphological inflections and the similarity between Latin and L1 verbal inflection.

Morphological blocking was also found at later stages of acquisition. Ellis and Sagarra (Reference Ellis and Sagarra2010b) employed an eye-tracking technique to compare cue reliance between English and Spanish monolinguals, and beginning and intermediate L2 Spanish learners of English. Participants were asked to read sentences containing both lexical and morphological cues for temporal reference. The experimental sentences had a tense conflict between verbal and adverbial cues. By gauging participants’ gaze times, the study found that (i) cue reliance was language-dependent, with English and Spanish monolinguals having longer gaze times on adverbs and verbs, respectively; (ii) L1 processing bias blocked the acquisition of L2 morphological cues, with beginning L2 Spanish learners spending significantly more time on adverbs and proving insensitive to tense incongruencies; and (iii) blocking was also detected in L2 learners who already had eight semesters of Spanish learning experience, and although intermediate L2 Spanish learners increasingly attended to verbal cues, they continued to rely primarily on adverbs, diverging from Spanish monolinguals.

Blocking and L2 syntactic acquisition

A limited number of studies examined blocking effects at the syntactic level. Hou (Reference Hou2021) exposed thirty L2 Chinese learners, whose native language did not have the equivalent of the Ba-construction, to the Ba-construction (3a) and its SVO counterpart (3b) via videos. When they were tested for comprehension and production of the target constructions, it was found that the comprehension of both constructions was equally accurate, but the production of the Ba-construction was vulnerable to interference from its SVO counterpart, resulting in avoidance of use and higher error rate.

Since L2 morphological blocking was found both at the initial and later stages of acquisition, it is natural to inquire whether L2 syntactic blocking can also be found at later stages of acquisition. To our knowledge, no studies have pursued this issue to date. It is therefore important to decide what the most appropriate method would be.

For morphological blocking, lexical and morphological cues often appear together in the same sentence. This makes investigation of the overt attentional bias possible (e.g., Sagarra & Ellis, Reference Sagarra and Ellis2013, with eye-tracking techniques). However, competing syntactic constructions rarely, if ever, appear in the same context (Goldberg & Jackendoff, Reference Goldberg and Jackendoff2004). Attentional bias as a consequence of syntactic blocking is mainly covert, which is more appropriately inferred from analyzing learners’ usage patterns during comprehension and production tasks.

More importantly, different usage patterns of the Ba-construction and its SVO counterpart between native speakers and L2 learners suggest their different representational relations. Specifically, the Ba-construction should be represented taxonomically from its SVO counterpart in native speakers’ mind, allowing them to show a preference toward the Ba-construction in certain contexts in order to fulfill certain pragmatic functions (e.g., expressing subjectivity). On the contrary, L2 learners, whose L1 lacks an equivalent for the Ba-construction, seem to represent the two target constructions less distinctly. As a consequence, they are less sensitive to the pragmatic functions specific to the Ba-construction and tend to use the two constructions interchangeably without any discrimination.

Participants

In total, 35 Mandarin native speakers and 61 L2-Mandarin learners took part in the experiment. Five L2 participants were excluded due to an incomplete vocabulary test used to assess proficiency (n = 1), technical difficulties (n = 1), or low accuracy (n = 3). Table 1 summarizes the descriptive data of their demographic information. The two groups were comparable in terms of age (t = -.492, p = .624), but there was a significant gender difference (χ2(1, N = 91) = 5.232, p = .022).

Table 1. Participant information

L2 participants were international students from a university in China, who came from Central Asia (Kazakhstan, Kyrgyzstan, Tajikistan, and Uzbekistan), Russia, and Sudan. They were all multilingual except for one Russian participant who was bilingual in Russian and Chinese. Besides Mandarin, participants from Central Asia and Russia were also fluent in Russian and a Turkic language (e.g., Kazakh, Uzbek) and/or Dungan, and participants from Sudan were all Arabic native speakers with advanced proficiency in English. They were further divided into two subgroups based on whether they spoke the Dungan language. Accordingly, L2 Dungan (L2DG) and Non-Dungan (L2NDG) groups were formed, each consisting of 28 participants. The two L2 groups did not differ significantly with respect to age (t = −.822, p = .415) or gender (χ²(1, n = 56) = .664, p = .415).

Irrespective of their L1 background, all L2 learners could be considered as having an intermediate Mandarin proficiency level and had been abundantly exposed to the Ba-construction. They had at least passed Level 3 of the HSK (a standardized Chinese language proficiency test), and the majority had actually passed Level 4, which is supposed to be the counterpart of the B2 Level of the CEFR. Their average length of residency in China was 2.2 years (SD = 0.9), meaning they had been exposed to and practiced the Ba-construction for at least 1.5 years by the time of the experiment, as the Ba-construction was introduced by the end of the first semester.

The proficiency level of the two L2 groups was comparable based on the results of a vocabulary test and participants’ self-evaluations. The vocabulary test was adopted from Zhang (Reference Zhang2017), and the full score was 99. The self-evaluation was a 7-point scale survey, with 1 meaning very poor and 7 meaning excellent (adapted from Bernolet, Hartsuiker, & Pickering, Reference Bernolet, Hartsuiker and Pickering2013). The descriptive data are shown in Table 2. No statistical differences were found in the results of the vocabulary test (t _vocab = –1.628, p = .109) and self-evaluations (t _overall = –1.263, p = .212; t _grammar = –1.584, p = .119; t _listening = –1.195, p = .237; t _speaking = −.473, p = .638).

Table 2. Chinese proficiency evaluation results

Note: SD in parentheses.

Materials, procedure, and coding

The critical trials consisted of 24 target pictures depicting common resultative causative events and 24 pairs of prime sentences. To create the target pictures, 24 Ba sentences were first constructed. Three native speakers with zero linguistics training were asked to perform a structural transformation task to ensure that every sentence could be successfully converted to their SVO counterpart. Then, 24 pictures were created based on those sentences, going through two rounds of ratings and revisions to make sure they accurately depicted the meaning of the sentences. Each VP was printed below its corresponding picture in both Chinese characters and Pinyin (see Figure 1).

Figure 1. The Presentation of the Picture in the Experiment; the Vp used in this Example Means ‘emptied’.

We expected that native speakers would be more likely to use the Ba-construction for the target pictures because of the salience of the causative result as well as the inevitable involvement of the subjectivity in picture description. To verify this, 25 native speakers were recruited to complete a picture description task without syntactic priming. They produced 321 Ba sentences, 196 SVO sentences, and an additional 83 sentences with other constructions. The proportions between the Ba and SVO production were 62.1% vs. 37.9%, confirming a bias toward the use of the Ba-construction (χ² (1, n = 517) = 29.741, p < .001).

The prime sentences corresponded to two prime conditions (see Table 3). Naturalness was evaluated by 10 native speakers on a 5-point scale. All the sentences were judged highly natural (m = 4.9, SD = 0.3). Since L2 participants were involved, lexical complexity was also controlled, with over 90% of the words within the 1 to 2,500 frequency band (Xiao, Rayson, & McEnery, Reference Xiao, Rayson and McEnery2009). The remaining words were also supposed to be familiar to our L2 participants based on feedback from their instructors.

Table 3. Sample prime sentences

An additional 24 pairs of prime sentences and their paired pictures were created as filler items. Four L2 error-prone syntactic constructions were used, namely the manner complement (4a), comparative (4b), preverbal locative (4c), and attributive clause constructions (4d) (see Table 4 for samples and paired pictures).

Table 4. Filler item examples

Each trial of syntactic priming began with a headphone icon displayed for 1 second, after which the audio prime sentence started to play. Participants were instructed to repeat out loud the sentence they had just heard before completing the picture description task (Figure 2). The events represented in the prime sentence and the paired picture were semantically unrelated to avoid content word repetition.

Figure 2. Flowchart of the Procedure of the Syntactic Priming Task.

In addition, (backward) digit recall tasks were constructed to further disguise the research purpose by increasing the variety of task types and the number of filler items. All the materials were then compiled as PowerPoint slideshows. The repeated sequences of digit recall, filler syntactic priming, backward digit recall, and critical syntactic priming gave 24 blocks in total (see Figure 3). One potential drawback to using PowerPoint was that the slides could only be presented in a fixed order. In order to avoid discernible patterns caused by the fixed sequence of trials, twelve lists of critical items were created, in which (1) the order of the critical pictures was randomized; (2) the order of the prime sentences was randomized; (3) each picture was paired with a different prime sentence in different slideshows; and (4) the first prime sentence in the first block was the Ba-construction in half of the slideshows, and its SVO counterpart in the other half to avoid the primacy effect.

Figure 3. Order of the Tasks in Each Block.

Participants were randomly assigned to one of the lists. They were briefed on how each task would work and given 2 blocks of practice trials. Three breaks were placed after blocks 6, 12, and 18. It took about 25 and 40 minutes for L1 and L2 participants, respectively, to complete all 24 blocks. L2 participants also completed a vocabulary size test, which took them another 30 minutes on average.

The coding was as follows: the recall of prime sentences was coded as “Y” or “N” depending on whether or not participants correctly repeated the prime sentence. The responses for the picture description were coded as “Ba,” “SVO,” “Other,” (i.e., non-target construction), and “X” (i.e., erroneous production). Twenty-five percent of the participants’ responses were independently coded by a second coder, with strong intercoder agreement in coding of the sentence recall (Cohen’s kappa κ = .826, p < .001) as well as the picture description (κ = .899, p < .001) (McHugh, Reference McHugh2012). Disagreement was resolved through discussion.

Data analysis

The 91 participants produced 2,184 sentences in the critical trials. Trials in which participants failed to correctly recall the prime sentence or in which responses for the picture description were coded as “Other” and “X” were eliminated. This screening process led to a loss of 79 and 203 sentences, respectively. The analysis was based on the remaining 1,902 data points.

Analyses were conducted in R (Version 4.0.2), employing generalized linear mixed models (GLMM). Models were built with the glmer function from the lme4 package (Version 1.1.21). The procedure followed the steps outlined in (Gries, Reference Gries2021). After initial screening, a full model was built: the fixed effects structure (FES) consisted of Group (L1, L2DG, and L2NDG), Prime Type (Ba and SVO) and their interactions as independent variables, and the First (block construction) Type (Ba and SVO) as a covariate. First Type was effect coded, and Group and Prime Type were dummy coded, with the L1_Group and the SVO_Prime as the reference levels. As to the random effects structure (RES), a maximal RES was constructed with random intercepts for subjects and items, by-subject random slopes for Prime Type, and by-item random slopes for Prime Type, Group, First Type, and L1.

A series of variable selection and model comparisons was then implemented to determine the best RES, followed by the best FES. The best RES was decided based on the results of principal component analysis. Following the principle of marginality, I simplified the RES by removing what seemed to do least. Each time a random effect component with the smallest variance was removed, a model comparison using likely ratio tests was performed to see if the removal of the component negatively impacted the model fitting. This process resulted in removing by-subject random slopes of Prime Type and removing by-item random slopes of L1, First Type, and Group were removed sequentially. The best FES was decided by computing likelihood ratio tests with a significance threshold set to .05. The result indicated that no predictors should be removed. Last, model diagnostics were performed for the model with the best FES and RES. The influential Subject 1 and Item 11 were removed as they were the influential data points. Neither collinearity nor overdispersion was identified in the final model.

Results

Descriptive data are visualized in Figure 4. Participants in all three groups overall used the Ba-construction more than the SVO construction. However, the Ba preference in the L1 and L2DG groups was across the prime conditions, and the change of the prime construction did not drastically alter the proportions of Ba and SVO production. By contrast, the prime condition had an obvious impact on the L2NDG group’s production, as the use of the Ba-construction was at chance level in the SVO prime condition (i.e., 52.5%), the proportion increasing to 84.2% in the Ba prime condition.

Figure 4. Responses Under Different Prime Conditions Across the Three Groups.

The results of the GLMM shown in Table 5 confirm the above-mentioned observations with statistical significance. The intercept (b = 1.004) is represented as the log-odds of the Ba production of reference group (i.e., L1) under the reference prime condition (i.e., the SVO condition), averaged over the levels of First_Type. Under the same SVO prime condition, the L2DG group produced significantly more Ba sentences (b = 2.208, p < .001), while the L2NDG group significantly fewer Ba sentences (b = −.787, p = .015), compared to the L1 group. For the L1 group and L2DG, changing the prime to the Ba condition did not significantly increase the production of the Ba-construction (L1: b = .119, p = .519; L2DG = .757, p = .109). However, for the L2NDG group, the Ba prime led to significantly more Ba production (b = 2.303, p < .001). In addition, a primacy effect was found. If the Ba-construction was used in the first block, participants tended to produce more Ba sentences (b = .381, p = .005).

Table 5. Mixed model parameters

Note: a. The final model formula is Production ∼ Group * Prime_Type + First_Type + (1 | Participants) + (1 | Item); b. Marginal r² = 0.281, Conditional R² = 0.510.