Alternation in the Mandarin disposal constructions: quantifying their evolutionary dynamics across twelve centuries

1 Introduction

There seems to be a wide consensus that the basic word order of Mandarin Chinese is SVO order (e.g., Light 1979; Sun and Givon 1985) but there also exists a widely used syntactic configuration of SOV word order, whereby the object is preceded by the markers ba or jiang and is followed by the predicate verb (Sun and Givon 1985; Xing 1994: 201). This syntactic structure, which can be represented as “subject/NP1 + ba/jiang + object/NP2 + verb (+ X)” as in (1),^[1] is commonly labeled “disposal form” or “disposal construction”. The term was introduced by Wang (1947) and it has generally been accepted in Chinese linguistic circles because the marked object in the construction tends to code a patient which undergoes an explicit change of state (Chappell 2013; Peyraube and Wiebusch 2020; Wang 1947).

It is commonly accepted that the disposal markers ba and jiang can be traced back to the first verb (= V₁, meaning ‘take/hold’) in a serial verb construction, as exemplified for ba in (2):

A crucial process in the development of these disposal constructions is the grammaticalization from V₁ to object marker, resulting in examples such as (1a) and (1b), where ba and jiang can no longer be interpreted as verbs (see Chappell and Peyraube 2011: 787–790; see also Chao 1968; Hopper and Traugott 2003: 28; Li and Thompson 1981: 466–479; Sun 2015: 429–442; Wang 1944).^[2] Given that the ba- and the jiang-construction has been claimed to have undergone the same process of grammaticalization and have thus acquired the same disposal function over time (Cao and Long 2005; Wu 2003; among others), they can be considered interchangeable alternants of the Mandarin disposal construction.

The view that the two disposal constructions encode the same function, however, has been challenged in a number of diachronic and synchronic studies. Xing (1994), for instance, finds a register difference between ba and jiang: in the 16th century, ba was more likely to be used in informal texts while jiang was preferred in formal texts, and this difference has resulted in the disappearance^[3] of jiang in written texts of modern times. In another diachronic study, Zhu (2016) investigates the distribution of the ba-construction and the jiang-construction in typical works from the dynasties of the early modern Chinese period (from the Tang dynasty to the Qing dynasty, i.e., 7th–19th century). He finds that in the initial stages, i.e., from the Tang dynasty to the Yuan dynasty (7th–14th century), the use of the jiang-construction was much more frequent than the ba-construction, but from the Ming dynasty onwards (15th–17th century), the ba-construction began to overtake the jiang-construction. Like Xing, she attributes this change of frequency to register: it was in colloquial texts that the ba-construction gained ground at the expense of the jiang-construction. Based on a synchronic investigation of ancient vernacular novels, Lu (2006) finds that the difference between ba and jiang mainly lies in the following three aspects: first, the ba-construction is usually used to describe the vulgarity of people;^[4] second, the ba-construction has a more prominent descriptive function, that is, the use of it can bring a more vivid picture of the situation described; third, the object of the ba-construction can be very long while that of the jiang-construction is relatively shorter. Jing-Schmidt and Tao (2009), then, suggest that ba and jiang share the basic meaning of disposal, but contrast in terms of subjectivity and emotionality. Specifically, the ba-construction prototypically signifies subjectivity and emotionality while the jiang-construction prototypically signifies objectivity and precision (Jing-Schmidt and Tao 2009: 36). Wu (2013) finds that while the ba-construction and the jiang-construction in modern Chinese are largely interchangeable, they differ in that the jiang-construction has a strong color of Literary Chinese, viz., an antique flavor, while the ba-construction has a strong color of colloquialism and is usually used in colloquial texts.

As can be seen, these earlier studies largely ascribe the distinction between ba and jiang to pragmatic differences, with register as an important motivation (in addition to such factors as emotionality and subjectivity). In so doing, the studies subscribe to Goldberg’s (1995) Principle of No Synonymy, whereby two constructions that are formally different are also different semantically and/or pragmatically. Furthermore, rather than being based on wide-ranging corpus data, these studies are more limited in scope in that they only consider data from typical works of particular periods – whether from a diachronic or synchronic angle. The diachronic studies (Xing 1994; Zhu 2016), furthermore, only quantify the distribution differences of the two disposal constructions. In the present – diachronic – investigation, we advocate a single meaning approach, where the ba- and jiang-constructions are seen as essentially semantically/pragmatically equivalent; that is, semantic/pragmatic differences between the ba- and jiang-alternants may well exist (e.g., the register differences discussed above), but they have no truth-conditional value (see also Section 2.2). In line with the view that grammatical variation and change are multifaceted and non-deterministic (e.g., Bernaisch et al. 2014; Bresnan 2007; Theijssen et al. 2013), this study probes into the probabilistic nature of the ba versus jiang alternation, i.e., it examines the language-internal and language-external factors probabilistically constraining the alternation, as well as the change in these factors.

In sum, the present study adopts a corpus-based variationist approach to address the following three questions:

1. What are the factors that constrain the alternation between the ba-construction and the jiang-construction?

2. How did these factors change diachronically?

3. What are the motivations underlying this change of factors?

This paper is structured as follows. Section 2 describes data retrieval and annotation, as well as the circumscription of variable contexts. In Section 3, we introduce the regression model. Section 4 reports and discusses the results of mixed-effects regression analysis. Section 5 summarizes the findings of this study.

2 Methods

3 Regression analysis

We applied mixed-effects binary logistic regression using the lme4 package (Bates et al. 2015) in R. In our mixed-effects model, the binary response variable is one of two alternants of the Mandarin disposal construction, which is coded as ba-cx, or jiang-cx. As regards independent variables, mixed-effects modeling takes into account not only fixed effects, which are repeatable, but also random effects, which are non-repeatable and group-specific (Speelman et al. 2018: 1).

To find out the best-fitted model, we adopted backward stepwise selection: fitting a model with all predictors and possible interactions, and then conducting a stepwise selection (Levshina 2015: 266). Our maximal model included NP2 animacy, NP2 definiteness, NP2 length, adjunct semantics, verb type, parallelism, genre, and period as fixed effects as well as verb and corpus file as random effects. To explore the diachronic evolutionary dynamics, interactions between period and all the other fixed-effects explanatory variables were incorporated. In addition, interactions between genre and all the other language-internal fixed-effects variables were included to explore their interaction. For the stepwise selection, we used the anova() function in R to compare the log-likelihoods, AIC (Akaike Information Criterion), and BIC (Bayesian Information Criterion) in two adjacent models, which are goodness-of-fit measures for comparison of models with different number of variables (Levshina 2015: 194).

The fitted final model includes a random intercept for verb, and one for corpus file as random effects, three two-way interaction terms, NP2 animacy, log-transformed NP2 length, parallelism, and verb type as fixed effects. The final model is as follows:

variant ∼ (1 | verb) + (1 | corpus_file) + period * (AdjunctSem + genre) + genre * NP2_Animacy + log (NP2_length) + parallelism + verb_type

With the concordance index C and the Nagelkerke pseudo-R ² as important goodness-of-fit statistics of logit models, this final model can predict the variant/outcome well (p < 0.001, pseudo-R ² = 0.995, C = 0.965). According to Hosmer and Lemeshow (2000: 162), C ≥ 0.9 shows “outstanding discrimination”. Hence, the concordance index C value of the final model indicates that the model can discriminate very well between the ba-construction and the jiang-construction. The Nagelkerke pseudo-R ² ranges from 0 (no predictive power) to 1 (perfect prediction). The value of pseudo-R ² in our model (0.995) indicates good prediction of the model. The classification accuracy of the model is 89.53 %, which is much higher than the baseline of 60.10 % when the most frequent variant (the ba-construction) is selected. To check if there exists multicollinearity between explanatory factors, we computed the condition index k (11.41) and the VIFs for each factor (1.03–4.64), which are both below alarming thresholds (Baayen 2008: 182; Levshina 2015: 272). For model validation, our dataset was randomly divided 100 times into a training set (70 % of the data) and a test set (the remaining 30 %). Afterwards, the model was fitted to each training set and its predictions on the corresponding test set were computed. The results show that the prediction accuracy of the 100 models ranges from 75.67 to 78.31 % and the mean accuracy is 77.11 %, which suggests a good model fit.

4 Results and discussion

In exploring the factors probabilistically constraining the choice between ba and jiang, we first calculated the importance of each factor in the model by measuring the decrease in goodness-of-fit when leaving the factor out of the model. This was done with the Anova() function in the car package in R. The result is visualized in Figure 1.^[11] We can see that the language-external factor “period” is the most important factor in the model, followed, at a distance, by the language-internal variables, and interaction terms. The importance of the language-external factors (especially “period”) can thus be said to reflect the significant role of socio-cultural developments in the three periods examined, but most outspokenly in period 3. More specifically, the high value of the factor “period” in Figure 1 arguably correlates with the fact that the Vernacular Language Movement in China boosted the use of the ba-construction in formal and informal texts alike in period 3.^[12] This finding dovetails with the tenet of cognitive sociolinguistics that socio-cultural forces (co-)define language variation (e.g., Kristiansen and Geeraerts 2013).

Figure 1:

Variable importance (decrease in model goodness-of-fit if factor removed) of all the factors found significant in predicting the ba and jiang variants.

Table 3 reports the fixed-effects structure of the final model. The column labeled “estimate” shows the estimates of coefficients on a logit-scale. Positive estimates indicate a preference for the ba-construction (the predicted outcome) while negative estimates suggest a preference for the jiang-construction.

The table shows that “intercept” is significant in predicting the variants. The intercept coefficient (−10.297) is the estimated log odds of the predicted outcome when all quantitative/numerical predictor variables are equal to zero and categorical variables are at their reference levels. Exponentiating this coefficient yields its simple odds value, which is smaller than 0.0001.^[13] This means that the odds of the ba-construction versus the jiang-construction are nearly zero in period 2 and in formal contexts with activity verbs, concrete NP2s, adverbial adjuncts, no parallel disposal constructions in adjacent contexts, and with the length of NP2 being one character.^[14] In other words, when all categorical variables are at their reference level and the only quantitative variable “log(NP2_length)” is zero, the probability of the jiang-construction is nearly 100 %; that is, language users choose the jiang-construction in all likelihood under such circumstances.

Next, let us first look at variables which only show main effects in the ba versus jiang alternation.

The variable “verb type” is only involved in main effects. Table 3 shows that all three levels can significantly predict the ba or jiang variant, compared with the reference level “Activity”. They differ in that the category/level “ExOcc” (verbs of existence and occurrence) (see (11) and (12)) significantly boosts the odds of the jiang-construction compared with the reference category “Activity” verbs, while verbs of mental activity (13) significantly increase the choice of the ba-construction. The verb type “Other” (14), which includes all the types of verbs apart from verbs of activity, existence and occurrence, and mental activity, also boosts the ba-construction, compared with the reference category “Activity” verbs.

Figure 2 demonstrates the ratio of the ba construction compared to the jiang construction by verb type across the three historical periods. We can observe that the probability of the ba-construction steadily increases in all four verb types, signaling its overall expansion over time; in other words, the jiang-construction has increasingly lost ground to the ba-construction. Compared to the ba versus jiang proportions in “Activity” verbs – the reference level in the regression analysis – in the three periods combined, we can see that the share of the ba-construction in “Mental activity” verbs and in “Other” verbs is higher (than that of jiang-constructions) while that of “ExOcc” verbs is lower (compared to the “Activity” verbs taking up an average of 55.5 % of ba-constructions, the “Mental activity” and “Other” verbs take up 66.9 and 63.1 %, respectively, whereas the “ExOcs” verbs only take up 39 %). In the output of the regression model, these results correspond with “Mental activity” and “Other” verbs boosting the odds of the ba-construction compared to the reference level and the “ExOcc” verbs boosting the odds of the jiang-construction.

Figure 2:

The proportion of the ba and jiang constructions by verb types across three historical periods.

It is noteworthy that in the latest period (period 3), 91.3 % of mental activity verbs choose the ba-construction; it is especially this class of verbs that has contributed to the expansion of the ba-construction and its taking over of the jiang-construction. Recall that ba was initially a lexical verb with the meaning of “holding”; when it was used in serial verb constructions, the second verb usually designated physical manipulation, as in (15).

In this example, the NP following ba is the object of both ba and the matrix verb 摇 yao ‘shake’. Specifically, in order to shake the tea tree, Yangshan (the subject) holds the tree first and then shakes it. It is from contexts of this type that the ba-construction has developed into a disposal construction (cf. Cao and Yu 2000; Cao and Long 2005); that is, the lexical meaning of ba has bleached and has become a disposal marker. Thus, in the initial stages of the development of the ba-construction, its matrix verb mainly denoted physical manipulation. After that, other types of verbs (most prominently among them, mental activity verbs) increasingly entered the matrix slot of the ba-construction, which can be seen as host-class expansion in Himmelmann’s (2004) terms.

On another note, the increase of mental activity verbs in the ba-construction in period 3 is consistent with the subjectivity/emotionality and metaphoricality of the ba-construction found in synchronous studies (e.g., Jing-Schmidt 2005; Shen 2002). That is, Jing-Schmidt (2005) and Shen (2002) argue that the ba-construction in modern Chinese is endowed with the notion of subjectivity.^[15] Mental activity verbs such as ji ‘worry’ in (13) denote subjective feeling/emotion or mental activity of human beings. Hence, the finding that mental activity verbs have been used more frequently in the ba-construction chimes with the subjectivity/emotionality of the ba-construction found in previous studies.

The estimate of log(NP2_length) is −0.267, which means that the chances of the ba-construction decrease by 0.267 with every additional unit of log(NP2_length). In other words, the longer the NP2 is, the less likely the ba-construction will be chosen in comparison with the reference level – the jiang-construction. Figure 3 illustrates the distribution of NP2s with different length in the two variants. As we mentioned in Section 2.3, to reduce data skewing and the effect of outliers, NP2 length is capped at 10 characters. Thus, as Figure 3 shows, the NP2 complement ranges for both variants between 1 and 10 characters. However, the density for the two variants, shown by the two lines, is different. Overall, in both variants, the density of NP2s with a length of one to six characters is much higher. Specifically, NP2s of one character are much more densely used in the ba-construction, while NP2s with a length of two to six characters are more densely used in the jiang-construction than in the ba-construction. For NP2s of seven to ten characters, the two density lines almost overlap, which suggests that there is no significant difference between distributions of NP2s with such length in the two variants. Thus, overall, the longer the NP2 is, the less likely language users choose the ba-construction compared with the jiang-construction. This finding is not in line with assertions made in some previous studies (e.g., Lu 2006), where the object of the ba-construction is claimed to be longer than that of the jiang-construction. As noted in Section 1, these previous studies lack quantitative statistical analysis and therefore need further corroboration.

Figure 3:

The distribution of NP2s with different length in the two disposal constructions.

The final variable which only demonstrates main effects in the alternation is “parallelism”. Table 3 shows that the level “yes”, viz., having parallel disposal constructions in the adjacent context, significantly predicts the use of the ba-construction compared with the reference level, i.e., without parallel disposal constructions in the adjacent contexts. It can be argued that parallelism reflects the effect of syntactic priming (also sometimes called structural priming, syntactic persistence, or plan reuse). Syntactic priming is the facilitation of processing that occurs when a syntactic structure is repeated across consecutive sentences (Bock 1986; Kempen 1977; Levelt 1989). This facilitation manifests itself both in language comprehension and in language production. During production, it has been documented that syntactic priming is reflected in an increased tendency to choose, for instance, the same grammatical voice (active vs. passive), the same relativizer (that vs. which), the same type of dative construction (double-object vs. prepositional dative), or the same type of syntactic structure (the ba-construction vs. SVO sentences) in consecutive sentences (e.g., Bock 1986; Bock and Loebell 1990; Fang and Liu 2021; Hinrichs et al. 2015). The present study echoes these previous studies in that language users increasingly choose the same type of syntactic structure, viz., the ba disposal construction, when there is already a disposal construction (a ba-construction or a jiang-construction) in the adjacent context.

Next, let us look at the interaction terms. Figure 4 visualizes the effect of the interaction term between “period” and “adjunct semantics”. From this figure, we can see that overall, the different levels of “adjunct semantics” in each period do not make much difference in constraining the alternation as the three points representing the three levels of the variable “AdjunctSem” are close to each other in each period. Across periods, however, there is great difference: obviously, period 3 greatly boosts the use of the ba-construction while in period 1 and period 2, the probability of the ba-construction is much lower. This is in accord with the finding in Figure 1 that “period” is the most important predictor of the alternation. Concretely, in period 1, the adjunct category “complement”, as in example (16), slightly boosts the odds of the ba-construction, compared to the reference level “adverbial”. In period 2, however, the category “complement” changes to significantly predict the jiang-construction, as can be seen from the main effect of the adjunct category “complement” in Table 3. In period 3, as the three points representing the three levels almost overlap, they show little difference in constraining the alternation, namely, all three levels prefer the ba-construction compared to period 2 (the reference level of “period”).

Figure 4:

Interaction plot: predicted log odds for the ba-construction variant by the factor “adjunct semantics” across periods.

The interaction effect between “genre” and “period” is visualized in Figure 5. It shows that “informal genre” in period 1 can significantly predict the use of the jiang-construction compared with “formal genre” (the reference level), and the effect size of this prediction is very large, with the estimate being –17.563, as can be seen from Table 3. This may be ascribed to the frequency effect of the jiang-construction in this period. As we mentioned above, in the initial stage of the co-existence of the two disposal markers, jiang was still much more frequently used than ba. Figure 6 presents the distribution of the two disposal constructions across the two genres in period 1. The frequency count of the jiang-construction in the informal genre is nearly twice that of the ba-construction. A Chi-squared test shows that there exists a significant difference between the frequency counts of the two constructions in the two genres (χ ² = 10.271, p = 0.0014).

Figure 5:

Interaction plot: predicted log odds for the ba-construction variant by the factor “genre” across periods.

Figure 6:

The distribution of the ba and jiang disposal constructions across the two genres in period 1.

When it comes to period 2, we can see from the plot that the two points representing the two levels of genre almost overlap, which indicates that they do not condition the alternation in a different way. Note as well that both genres prefer the jiang-construction as their log odds are negative. When it comes to the modern period, i.e., period 3, both genres now boost the odds of the ba-construction, and “informal genre” shows a stronger preference for the ba-construction compared to “formal genre”. Thus, we can see a clear change in the preference of informal texts over time, that is, from a preference for the jiang-construction in period 1 to a preference for the ba-construction in period 3. Hence, overall, in period 3, the ba-construction is preferred to its counterpart, the jiang-construction, both in informal and formal texts, thus signaling the expansion of the use of the ba-construction over time. This observation echoes Xing’s (1994) finding that the jiang disposal marker has largely been replaced by ba as a disposal marker in modern written texts.

We submit that the growth of the ba-construction can be attributed to two general trends (one language-internal and the other language-external). First, ba’s expansion can be seen as resulting from the emergence of a new grammaticalized layer (Hopper 1991) of the disposal construction, in addition to the older jiang-construction (Xing 1994). Jiang became a grammaticalized disposal marker much earlier than ba (e.g., Lü 1955; Wang 1957). Ba did not replace jiang immediately when it developed into a disposal marker, and the two have co-existed until modern times. Nonetheless, the newcomer has been encroaching on the territory of the old disposal marker jiang. Diao (1993) notes that in the initial stage of the co-existence of the two disposal markers, the use of jiang was still much more frequent while from the Yuan dynasty onwards, ba began to overtake jiang. Second, the overall preference for the ba-construction in modern times can be related to the Vernacular Language Movement in Chinese, which began in late 1910s. The goal of the movement was to replace Literary Chinese (文言 wenyan) with spoken language (白话 baihua) in nearly all written works. As a result, by the early 1920s, vernacular in written Chinese gained nationwide acceptance as a reputable style in prose, poetry, and fiction and was the officially designated style for school textbooks. As noted by previous studies (cf. Jing-Schmidt and Tao 2009; Wu 2013), while the jiang-construction has a strong color of Literary Chinese, viz., an antique flavor, the ba-construction has a strong color of colloquialism. Thus, the popularity of vernacular from the 1920s onwards has substantially contributed to jiang losing ground to ba.

With respect to the interaction between “NP2 animacy” and “genre”, we can see from Figure 7 that compared to “formal genre” (the reference level of “genre”), there is little change in informal genre in the position of “animate” relative to that of “concrete” (the reference level of “NP2_Animacy”). What does change substantially is the position of “abstract”. Compared to concrete entities, abstract NP2s (17) in formal genre boost the odds of the ba-construction while in informal genre, they boost the jiang-construction.

Figure 7:

Interaction plot: predicted log odds for the ba-construction variant by the factor “NP2 animacy” across genres.

As regards random effects, it is noticeable from Table 4 that the by-subject adjustment “corpus_file” (N = 1,059, Variance: 511.08, Std Dev: 22.604) accounts for more variability than the by-item adjustment “verb” (N = 1,569, Variance: 1.05, Std Dev: 1.024). It should be noted that in the CCL corpus, the name of the corpus file of a particular concordance is actually the title of the book the concordance has been taken from. Hence, the number of corpus files equals that of the books. The number of books to a large extent approximates the number of authors as the majority of the books investigated in the present study were written by different authors. It has been documented that author differences play a role in language change not only in Indo-European languages such as English (e.g., De Smet 2020; Petré and Anthonissen 2020), but also in Mandarin Chinese (Li et al. 2023). Thus, it is not surprising to observe the random effects of “corpus_file” in the present study. Previous studies have also found that verb differences play a part in the diachronic evolution of Chinese syntactic variation (cf. Li et al. 2023). This effect is echoed in the present study.

5 Conclusion

While the disposal construction in Mandarin has been studied intensively, the evolution of the variation between the ba and jiang constructions has rarely been investigated in a multifactorially controlled research design. In this paper, we gauged the differential impact and interplay of a substantial set of language-internal and language-external factors. The results indicate two general trends. First of all, as the new layer of the grammaticalization of the Mandarin disposal construction, the ba-construction increased its odds over time at the expense of its counterpart, the jiang-construction, with a strong boost in period 3. Second, over time, the effect size of the significant preference for the jiang-construction in informal genres decreased from period 1 to period 2, and this preference disappeared in modern times (period 3), with both informal and formal genres converging to favor the use of the ba-construction. Ba’s dominance not only results from its emergence as a new grammaticalized layer of the disposal construction, it also needs to be explained from a language-external perspective, and specifically as the impact of the Vernacular Language Movement in China. The variable importance scale (Figure 1), with “period” as the most important variable, clearly reflects the impact of this movement in period 3.

In addition to the general trends, regression modeling quantifies the evolutionary dynamics of the alternation by revealing that there are both stable linguistic constraints (i.e., parallelism/syntactic priming, verb type, NP2 animacy, and NP2 length) and fluid constraints (i.e., adjunct semantics, and genre). Our findings concerning the stable linguistic constraints are the following: the factor “parallelism” (i.e., syntactic priming) predicts the use of the ba-construction; verbs of existence and occurrence prefer the jiang-construction, while mental activity verbs and “Other” verbs prefer the ba-construction; abstract NP2s in the informal genre significantly boost the use of the jiang-construction whereas in the formal genre, they significantly boost the odds of the ba-construction; long NP2s/objects significantly boost the jiang-construction. With regard to fluid constraints, our regression analysis shows that there are fluctuations in both their effect sizes and effect directions as a function of real time. These probabilistic fluctuations showcase the potential of multifactorial quantitative methods with a variationist/probabilistic twist, for these methods contribute to unveiling subtle grammar changes that went undetected in previous non-multifactorial approaches to syntactic alternation.

This empirical study has at least the following three theoretical contributions. First, while previous studies have mainly focused on syntactic alternations involving word order change (e.g., the ba-construction vs. SVO sentences in Chinese, and the theme-recipient alternation in Chinese; see Fang and Liu 2021; Li et al. 2023), the present study has been concerned with a new type of alternation that does not involve word order variation: the alternation between two grammatical markers, viz., object/disposal markers ba and jiang. In doing so, this study has not only advanced our knowledge of the two disposal constructions and their evolution, but has also shed new light on diachronic syntactic alternation in general, and on grammatical marker alternation in Mandarin Chinese in particular. Second, the results of our study indicate that the ba-construction has been winning out over time in its competition with the jiang-construction. Thus, it would be natural for the latter to be eventually preempted by the former in ordinary discourse.^[16] In recent literature, statistical preemption has given rise to an interesting discussion, namely whether it is compatible with a “sameness of meaning” view (the point of departure of the present study; see Section 1) or with the “No Synonymy view”. On the one hand, one could argue that statistical preemption presupposes sameness of meaning (speakers end up consistently selecting one alternative rather than the other because the two alternatives are similar in meaning). On the other hand, advocates of the No Synonymy view hold that it is precisely “no synonymy” that preemption is compatible with: as Leclercq and Morin (2023: 4) point out: “Indeed, while the principle of no synonymy posits that no two constructions have the exact same function, statistical preemption ensures that this be the case by blocking the use of an alternative (or new) form when a function is already associated with a specific construction”. Third, this study makes a methodological contribution to the empirical testing of hypotheses. It is the first one, to our knowledge, to take a corpus-based long-term perspective on the variation, from the earliest period whereby the ba and jiang constructions coexisted to the present, and to gauge the effects under multifactorial control, including an elaborate random structure. It thus contributes to a data-driven corpus analysis of complex historical linguistic issues.

Still, the results have to be approached with due caution. Because of the lack of true diachronic oral data for the periods prior to modern Chinese, we had to classify drama, family letters, and the scripts of story-telling into the informal genre as proxy. Moreover, we have ignored regional differences and other sociolinguistic factors, as this kind of metadata is absent for historical texts.