Transfer five ways: applications of multiple distinctive collexeme analysis to the dative alternation in Mandarin Chinese

1.1 Previous applications of MDCA

As mentioned above, MDCA is one member in the family of collostructional analyses first developed by Stefanowitsch and Gries (Gries and Stefanowitsch 2004a, 2004b; Stefanowitsch and Gries 2003, 2005). All types of collostructional analyses determine the strength of association between two kinds of constructions, most typically a lexical and a syntactic construction (although the methods can, and have been, applied also to determine lexeme-lexeme, morpheme-lexeme, lexeme-variety, and other association strengths). (M)DCA, in particular, “looks at how much constructions (usually words/lemmas) (dis)like to occur in a slot of a usually more schematic construction as opposed to alternative, functionally similar constructions” (Gries 2023: 352). In other words, (M)DCA is used to zoom in on a variable slot in a limited number of functionally very similar constructions and typically asks the questions whether the word types that occur in a variable slot shared by the alternatives show any bias towards one constructional variant and what that might mean for how the constructions differ in terms of their functions (usually, their semantics).

DCA contrasts two constructional variants (or alternations) with each other, and it has been applied extensively. MDCA is a conceptual extension of DCA, allowing the comparison of three or more functionally similar constructions (e.g., Gilquin 2006, 2010, 2013). Compared to the number of studies that have applied DCA, the number of studies using MDCA is much smaller, but MDCA has arguably been applied to a wider range of phenomena, which we take to be indicative of its usefulness. Most studies using MDCA focused on differences between a set of near-synonymous syntactic or argument structure constructions (Béchet 2015; Gilquin 2006, 2010, 2013; Schönefeld 2013a, 2015; Su and Chen 2019). For example, Gilquin (2006) uses MDCA to uncover the differences in meaning between 10 English periphrastic causative constructions with regard to the verbs that are distinctively associated with the non-finite verb slot of each construction. Less typically, MDCA has also been applied to the comparison of near-synonymous words (Desagulier 2014; Rajeg 2014, 2019, 2020) and morphemes (Van Goethem and Norde 2020).

As the above examples illustrate, the most typical application of MDCA is to compare three or more constructions and identify word types in a variable slot in these constructions that aid in identifying systematic semantic and/or functional differences between these multiple constructions. However, the ‘multiple’ part of MDCA has also been reconceptualized such that rather than looking at word types in multiple constructions, researchers have examined word types or constructions in multiple registers or in multiple time periods. For instance, Hilpert (2006) applies MDCA to the diachronic study of the infinitive complements of the English auxiliary shall. This method, which Hilpert (2008) refers to as ‘diachronic distinctive collexeme analysis’, has been widely used to track changes in the preferred collocates of individual constructions across multiple time periods (Fonteyn and Hartmann 2016; Hilpert 2006, 2008, 2012a, 2012b; Rens 2017; Zhou 2023). These shifting collocational preferences can be interpreted as changes in the meaning/function of the construction (Hilpert 2006, 2008). Yet another application of MDCA is to identify the register-specific behavior of a given construction (Schönefeld 2013b; Zhou 2023). For example, Schönefeld (2013b) uses MDCA to identify what types of co-occurring elements are preferred by the English go un-V-en construction in the four registers (academic prose, newspaper texts, fiction and conversation) in the British National Corpus. Less typically, MDCA has also been applied to determine the lexeme-morpheme (Wulff et al. 2009) and lexeme-variety (Mukherjee and Gries 2009) association strengths.

As mentioned above, existing applications of (M)DCA reveals that most studies targeted Indo-European languages (one of those most targeted being English) – only few studies exist that examine non-Indo-European languages such as Chinese (Su and Chen 2019) and Indonesian (Rajeg 2014, 2019, 2020). The present paper, however, is concerned with an alternation phenomenon in a language that has been studied collostructionally much less: dative alternation in Chinese, to which we turn now.

1.2 Dative alternation in Mandarin Chinese

Argument structure alternations refer to the phenomenon in which pairs/sets of formally distinct argument structure constructions (called the variants of the alternation) share substantial aspects of their semantics such that they can accommodate a common set of lexical items to fulfill similar functions (Perek 2015: 145–146). The dative alternation is a classic example of argument structure alternation in the world’s languages. In English, specifically, the double-object construction (the ditransitive) (Stefan gave Shengyu the cookie) alternates with the prepositional to-dative construction (Stefan gave the cookie to Shengyu).

While extensively studied in English and other Germanic/Western languages (e.g., Ambridge et al. 2012; Bresnan and Ford 2010; Bresnan et al. 2007; Colleman 2009; Colleman and Bernolet 2012; de Marneffe et al. 2012; De Vaere 2020; De Vaere et al. 2021; Gries 2003; Gries and Stefanowitsch 2004a; Gropen et al. 1989; Reali 2017; Theijssen 2012; Valdeson 2021), comparatively little is known about the status of dative alternation in Mandarin Chinese. According to previous studies, there are at least five variants of the dative alternation in Mandarin Chinese that share substantial aspects of their semantics (Chang 2014; Lu 2007; Ma 2013; Peyraube and Lü 2019):

However, previous usage-based constructionist studies on dative alternation in Mandarin Chinese either included only a subset of the five constructions or did not distinguish among them at all. For example, Hsiao and Mahastuti (2020) carried out a distinctive collexeme analysis of constructions a, b, and c (they grouped constructions b and c as expressing the same construction, in contrast to construction a) and found that they differ in the number and completion of transfer events. However, according to the Principle of No Synonymy (Goldberg 1995), formally distinct constructions should also differ functionally. In other words, variants of an alternation are seen as related yet separate constructions. By contrast, Peyraube and Lü (2019) proposed that construction b and construction c differ in the types of verbs that can occur in them and the semantic roles of the recipient-like argument (recipient, source, or beneficiary) involved in the construction. Thus, it might be problematic for Hsiao and Mahastuti (2020) to group variants b and c as expressing the same construction. Zhang and Xu (2019) conducted a multifactorial corpus analysis of constructions a, b, c, and e – grouping them into patterns a and e versus b and c – and found with a binary logistic regression that the definiteness of theme, the concreteness of theme, and the animacy of recipient are strong predictors of speakers’ choice of construction. Recently, Zhang and Xu (2023) examined the factors determining speakers’ choice between constructions a and c with a mixed-effects logistic regression analysis and reported that it was significantly affected by the animacy, pronominality, and definiteness of the recipient, the accessibility and concreteness of the theme, and the length difference between the theme and the recipient. Only Ma (2013) included instances of all five constructions, yet performed only a collostructional analysis (CA) on the combined data rather than running a MDCA.

In summary, it is yet unclear what distinguishes the five variants. The studies referenced above have shown that intra-lingual factors (the length, animacy, pronominality, and definiteness of the theme and the recipient, or the sense of the verb) impact speakers’ choices. However, there are also several relevant extra-lingual factors that have been shown to matter in the dative alternation, such as the language variety (e.g., Bernaisch et al. 2014; Bresnan and Ford 2010; Röthlisberger 2018; Röthlisberger et al. 2017), mode/register (e.g., Buysse 2012; Theijssen 2008; Wang 2013), as well as gender and age of the speaker (e.g., Jenset et al. 2018; Lorente Sánchez 2018; Röthlisberger 2021; Theijssen 2012). Syntactic priming has also been proven to play a role (e.g., Bresnan and Ford 2010; de Marneffe et al. 2012; Gries 2005). While a comprehensive study of all these factors in the dative alternation in Mandarin Chinese is out of scope for the present study, this is the first study to identify verb-specific preferences of all five constructions.

2.1 The corpus data

Our data come from the ToRCH2014 (Texts Of Recent CHinese) corpus (Xu 2017), a balanced corpus of written Mandarin Chinese, which consists of 983,651 tokenized words and 68,453 word types in 657 files. For the most part, the corpus was built following the sampling scheme of the Brown corpus (Kučera and Francis 1967), covering texts (mostly published in 2014) from 15 text categories, but there are also two differences: First, Brown’s category N (western and adventure fiction) was replaced with a category of martial arts fiction, following the Lancaster Corpus of Mandarin Chinese (LCMC, McEnery et al. 2003). Second, the sampling frame of 500 text segments of approximately 2,000 words was not strictly replicated for text files in categories A–E. There are more text files with varying length belonging to those categories in ToRCH2014 than in the Brown corpus, leading to a total of 657 files. The genres covered in ToRCH2014 are shown in Table 1.

For spoken Chinese corpus data, we decided to use the CallFriend-Mainland Mandarin corpus (Canavan and Zipperlen 1996), which consists of 43 files transcribed from casual Mandarin telephone conversations between friends, totaling 273,327 word tokens and 12,705 word types. Since only seven of the transcripts were segmented, we removed all the whitespaces between all words and then tokenized all files using the Stanford Word Segmenter^[2] (Tseng et al. 2005) with the Chinese Penn Treebank word segmentation standard. Since the texts in ToRCH2014 was originally segmented using a different segmenter, the ToRCH2014 corpus was resegmented after all white spaces were removed, following the same procedure to maintain consistency. After word segmentation was done, all the spoken and written texts were tagged using the Stanford Log-linear Part-Of-Speech Tagger^[3] (Toutanova et al. 2003).

We adopted a verb-centered approach in the retrieval of all five constructions. A total of 354 verb candidates were identified by combining the verb lists of Mandarin ditransitive verbs studied in previous research (Hsiao and Mahastuti 2020; Liu 2006; Ma 2013; Yao and Liu 2010). The searches and retrieval are based on word form/PoS-tag combinations (c.f., Stefanowitsch 2019). Given the wide-spread cross-PoS polysemy in Mandarin Chinese, making the retrieval PoS sensitive seemed more sensible than basing the retrieval purely on word forms. For example, the word 还 can be a verb (还 huán ‘return’) or an adverb (还 hái ‘still’), and is tagged as VV or AD accordingly (Xia 2000). There are altogether 1939 instances of 还_AD and 5 instances of 还_VV, so searching by word form/PoS-tag combinations can help us filter out those irrelevant uses.^[4] Since 给 gěi ‘to’ can be attached to some ditransitive verbs in variant b and variant d, the query term is formed as follows: Verb_VA|Verb_VC|Verb_VE|Verb_VV|Verb给_VA|Verb给_VC|Verb给_VE|Verb给_VV. Specifically, the search term for 还 huán ‘return’ is 还_VA|还_VC|还_VE|还_VV|还给_VA|还给_VC|还给_VE|还给_VV. All the sentences in which the candidate verb/PoS-tag combinations occur were extracted using AntConc (Anthony 2020). Retrieving all instances of these verbs yielded 26,523 attestations in the ToRCH2014 corpus and 9,825 attestations in the CallFriend corpus for a total of 36,348 attestations.

Since (i) this initial sample is too large to comb through manually, (ii) the written and the spoken corpus are of dramatically different sizes, and (iii) a substantive number of verbs occur in either corpus only rarely (the frequencies of the verbs display the expected Zipfian distribution), we developed a sampling procedure that would ensure reasonable coverage but also minimum frequencies of occurrence that allow for the MDCAs to be applied meaningfully. We decided to set a frequency threshold such that only verb types occurring at least 5 times in each of the written and the spoken corpus would be included in the final sample, which resulted in 76 verb types (with a combined token frequency of 28,434, consisting of 18,743 in ToRCH2014 and 9,691 in the CallFriend corpus). That in turn means that these 76 verb types account for over three quarters of all the tokens of the entire data sample initially retrieved.

In order to minimize possible effects of speaker (approximated by the corpus file) and sample maximally diversely across the two corpora, the following sampling procedure was implemented: the complete data set was sorted by verb, then by file (with the files in random order), then by occurrence in the file (with the occurrences in random order). The first author then checked each attestation in a given file to identify a true hit (that is, an instantiation of either of the five constructions) until a true hit was found. Once a true hit was found in a file, Author 1 moved on to the next file. This process was continued until the end of the concordance was reached. A total of 1,015 instances from 50 different verb types were sampled this way. The frequency distributions of each constructional variant are as follows, ordered from the most frequent variant e to the least frequent variant a with their frequency in parentheses: construction e (584), construction c (232), construction d (123), construction b (44), and construction a (32).

2.2 Multiple distinctive collexeme analysis

Distinctive collexeme analysis (DCA, Gries and Stefanowitsch 2004a) identifies lexical items that show a strong preference for one member of a group of two or more seemingly synonymous constructions as opposed to the other(s). This method has been frequently applied in the context of syntactic alternations, such as the dative alternation. To calculate the association strength between a particular lexical item l and a member of two semantically/functionally similar constructions cx₁ and cx₂, the following frequencies are needed: the frequency of l in cx₁, the frequency of l in cx₂, the frequency of lexical items that are not l in cx₁, and the frequency of lexical items that are not l in cx₂, as shown in Table 2 (adapted from Stefanowitsch 2014).

This contingency table can then be submitted to a contingency test that determines how much the observed distribution differs from the expected one. Typical association measures (AMs) that are used to express this difference are the Fisher-Yates exact test (p_FYE), G², or a Mutual Information (MI) score. Once this procedure has been repeated for all members of L (l _1…n ) occurring in cx₁ or cx₂ in a given corpus, lexical items are ranked by their association strength. Lexical items that occur (significantly) more frequently than expected in one of the constructions are called (significantly) distinctive/differential collexemes of that construction (Stefanowitsch 2013, 2014, 2020).

As we discussed earlier, conceptually speaking, multiple distinctive collexeme analysis (MDCA) is a rather straightforward extension of a DCA to cases where more than two constructions are compared to one another, but mathematically speaking, the departure from the simple binary contrast requires a different statistical approach. In the following, we briefly outline how MDCAs were originally computed (exact binomial tests), and how they can alternatively be computed using multiple distinctive log odds ratios, Pearson residuals, or Kullback-Leibler divergence scores.

4.1 Linguistic findings/implications

The results from the multiple distinctive collexeme analysis help us identify the similarities and differences between the five constructional variants, as summarized in Table 3.

Table 3 shows that while there is significant overlap in semantic potential, the variants differ in terms of potential semantic scope, with some being much narrower than others. Variant e is the most frequent (in our data anyway), attracts the widest range and highest number of collexemes (i.e., verbs) and thus is the most inclusive semantically, and can express transfer bidirectionally – if we had to suggest which variant to teach first to second language learners of Chinese, our recommendation would be construction e. Construction c, the next most frequent variant, seems more focused on short-distance transfer events, and shares with construction e that it can be used to express transfer in either direction. As the reviewer pointed out, the ease of comprehension should also be considered when deciding which variant to teach first to L2 Chinese learners. The constructional variant in Chinese that has a parallel construction in the learner’s L1 (structurally or/and semantically) is conceptually more accessible and easier for the learner to acquire (Teo 2023). For example, construction c (the ditransitive) in Chinese (我给了玛丽一本书 Wǒ gěi-le mǎlì yī běn shū ‘I gave Mary a book’) and the ditransitive construction in English (I gave Mary a book) are parallel constructions structurally in the sense that both constructions have the same syntactic ordering (i.e., SBJ_agent V OBJ_recipient OBJ_theme) and also semantically because both are more focused on short-distance/direct transfer events (c.f., Gries and Stefanowitsch 2004a). From that perspective, construction c would be the easiest construction for L1-English L2-Chinese learners to understand and therefore should be taught first. The remaining constructions are only used to express transfer unidirectionally, with construction d being the one with the widest semantic range of transfer events of the three. Variant b is similar to c in its restriction to short-distance transfer scenarios, while variant a is specifically associated with long-distance transfer events.

The overall similarity of variants b and c poses an interesting question by re-raising the question of whether they should actually be considered the same construction or not, which in turn affords us the opportunity to highlight possible follow-up analyses our methods offer. In general, one would argue that two constructs instantiate two separate constructions if the constructs differ on the formal or the functional side (or of course both). In Section 1.2 above, the difference between the two examples is the presence of 给 gěi ‘to’ versus 了 le (glossed as an aspectual marker) and qualitative considerations do support the notion that b and c exhibit less than the high degree of similarity one would expect to see for conflation. Variants b and c differ along aspectual lines such that the former usually describes a completed action while the latter can be used to describe both a completed and a future action. Also, the presence of 给 gěi ‘to’ limits the direction of the transfer events that variant b can express, i.e., it makes b unidirectional. Some ‘acquire’ verbs, such as 收 shōu ‘receive’, can occur with variant c to denote the movement of the theme from the source to the agent (see [8]), but cannot occur with variant b because of 给 gěi’s presence. Another ‘acquire’ verb, 买 mǎi ‘buy’, denotes opposite directions when occurring with variant b (from agent to recipient) and variant c (from source to agent).

In addition to these qualitative considerations, we can approach the issue of the similarity of b and c using the quantitative data we have: If variants b and c were really just one construction, the distributional hypothesis underlying all of collostructions and nearly all of quantitative corpus linguistics would lead to the expectations that

1. their collexemes would behave very similarly, and that

2. the correlation between the distinctive collexeme strengths for b and the distinctive collexeme strengths for c would be high.

However, our results suggest both do not seem to be the case, as can be seen when one, first, plots the distinctive collexeme strengths of b and c against each other and when one, second, checks the correlation between their collexeme strengths, which is not strong at all (Spearman’s rho = 0.026 with a bootstrapped 95 % confidence interval of [−0.21, 0.273]). In fact, the correlations and an exploratory cluster analysis on them suggest that b is more similar to a than to anything else, which corroborates Liu’s (2006) observation. Thus, if we take the distributional hypothesis seriously, as one has to once one has committed to any kind of collocational/collostructional analysis, it does not seem as if b and c should be grouped (as opposed to Hsiao and Mahastuti [2020] and Zhang and Xu [2019]).

4.2 Methodological findings/implications

This study is one of a small number of MDCAs, small at least compared to collexeme and distinctive collexeme analyses. It is one of an even smaller number of such studies following recent discussion by using more than one association measure value (like p_FYE or an exact binomial test), namely a frequency value together with an association measure that is much less correlated with, or tainted by, co-occurrence frequency (Gries 2019, 2022b, 2023, 2024), and it is among the first to do so with a directional association measure while also considering the dispersion of each verb over the constructions. In bringing together this variety of innovations, this study, on top of the linguistic findings, was also a methodological proof of concept. It is, however, also essential to consider which MDCA approach is most useful under what circumstances. In general, we feel that the following conclusions and recommendations are warranted:

1. Of the two measures that are highly correlated with co-occurrence frequency – the legacy exact binomial tests and the Pearson residuals – there is no good reason to use the former anymore. The results both measures produce are extremely highly correlated but the former runs into the -Inf problem of really small probabilities much more often than the latter, and taking steps to address that problem can sometimes amount to be computationally very resource-intensive and slow; computing all Pearson residuals for one analysis, by contrast, usually takes no time at all, which in turn allows for bootstrapping approaches etc. for confidence intervals of collexeme attraction values. However, downsides even for the Pearson residuals approach are the high correlation with co-occurrence frequency and the bidirectionality, which means its application should maybe be restricted to exploratory approaches.

2. The two measures that are more independent of co-occurrence frequency – the multiple log odds ratio and the contributions to KLD – are both fast to compute; the main criterion differentiating between them is the ability of the latter to distinguish verb-to-construction and construction-to-verb associations. While we have focused only on one direction here, depending on one’s goals, this can be a very attractive option, especially when coupled with confidence intervals (Olguín Martínez and Gries to appear).

That means, for any kind of study that has theoretical and/or explanatory goals (rather than just exploratory), the contributions to KLD approach is probably the best-suited one. It combines speed (for bootstrapping confidence intervals) with ‘expressiveness’ (capturing association largely independently of co-occurrence frequency) and directionality – it’s hard to see why this should not be the method of choice for future studies.