14.1 Introduction
The question to what extent the linguistic features found in English as a Lingua Franca (ELF) interactions are recurring – or even systematic – and to what extent they are the result of accommodational processes remains at the heart of ELF research. This is, at least in part, due to the fact that ELF encounters are typically transient interactions between speakers of various first languages, at varying stages of proficiency in English, and from different linguacultural backgrounds. ELF interactions can occur between speakers who meet repeatedly and over an extended time period, but often they are only temporary and take place only once in a specific situation and within a specific constellation of people (so-called ‘Transient International Groups’, Pitzl Reference Pitzl2018). This leads to a complex situation of language contact, in which speakers contribute their Individual Multilingual Repertoires (IMRs) to a common Multilingual Resource Pool (MRP), from which they can choose during the ongoing encounter (Pitzl Reference Pitzl2016: 295–9; Reference Pitzl2018: 33). In consequence, Pitzl (Reference Pitzl2018: 36–7) identifies two major methodological challenges for the study of ELF: (1) what are suitable models to capture the fleeting nature of ELF encounters, and (2) how can the fact that grammar is constantly being negotiated by language users be considered in research? Indeed, the concept of (non‑)canonicity is particularly relevant in ELF, since this is the context in which the notion of the ‘elusive target’ is central: ELF users may or may not subscribe to a (supra-)regional or external standard; they may or may not have received formal education in English; they may or may not have spent significant time abroad; etc. Hence, it is difficult to determine what a ‘canonical’ structure in a specific ELF context might look like.
In this chapter, we therefore draw on insights from the theory of emergent grammar to investigate (potentially) non-canonical language usage in ELF face-to-face interactions. This means that we regard conversational interaction as consisting not ‘of sentences generated by rules, but of the linear on-line assembly of familiar fragments’ and grammar as ‘emergent and epiphenomenal to the ongoing creation of new combinations of forms in interactive encounters’ (Hopper Reference Hopper, Auer and Pfänder2011: 26). As we show in the following, this usage- and experience-based view of grammar acknowledges the somewhat idiosyncratic nature of ELF encounters while explaining shifting feature frequencies in ELF. To that end, we conduct a qualitative and quantitative analysis of minus-plurals in Asian and European ELF to investigate how much speaker L1s and their corresponding language families as well as other extra- and intra-linguistic factors play a role in the decision-making of ELF speakers. The term ‘minus-plurals’ describes examples such as We had two beer-Ø instead of We had two beers and has been preferred over other terms as it represents an ideologically neutral term (see Rüdiger Reference Rüdiger2019: 48, and Section 14.3 for more details). Furthermore, we provide a discussion of what statistics can and cannot contribute to an analysis of ELF encounters, which, so far, have been predominantly analysed using qualitative approaches. The two main research questions we address in this chapter are the following:
(1) Are minus-plural constructions in Asian and European ELF encounters the result of typological pressure, that is, are they selected because of their dominance in the linguistic ecology of the speech situation (cf. Ansaldo Reference Ansaldo2009) and therefore the result of ‘cooperative restructuring’ (Thompson Reference Thompson2017: 209)?
(2) Which other extra-linguistic and intra-linguistic context factors have a significant influence (if any) on the use of minus-plurals?
While the first research question primarily scrutinises L1 transfer as an explanation for the use of minus-plurals, the second research question focuses on the age and gender of the interactants, the animacy of the head noun, and the presence of a quantifier as potential factors.
We use two comparable ELF corpora to analyse the data from a qualitative and quantitative perspective since, hitherto, few studies have combined a close reading of select examples with statistical methods in ELF contexts. The goal of the quantitative study is to identify potentially impactful variables for occurrences of minus-plurals. Ultimately, we seek to improve our understanding of how linguistic features in ELF are employed and to what extent they can be attributed to (socio-)linguistic factors, cooperative restructuring, or a combination of these factors. Furthermore, we address the potential of explaining feature emergence in ELF as based on regionally identifiable but also diffuse cores in which certain languages and language families dominate, leading to more frequent occurrences of specific contact features.
In Section 14.2, we establish the theoretical framework of communicative dynamism and cooperative restructuring in ELF and introduce and explain minus-plurals as the feature in focus of our study. We then provide detailed information on ACE and VOICE, our dataset, the predictor variables, and the applied statistical method in Section 14.3. In Section 14.4, we present and discuss the results of our study. Finally, in Section 14.5, we conclude the chapter and provide an outlook.
14.2 Grammar in ELF: Syntactic Borderlands?
14.2.1 Emergent Grammar in ELF
ELF interactions come in manifold forms, which makes it hard – or maybe even impossible – to define what ‘ELF grammar’ actually looks like. In consequence, recent research has moved away from regarding ELF as a stable code and rather conceptualises it as ‘a series of more or less demanding communicative situations where speakers come with whatever their language skills to tackle the communicative tasks at hand’ (Ranta Reference Ranta, Jenkins, Baker and Dewey2017: 247). According to this view, ELF grammar is in a state of fast change, strongly depending on the linguistic backgrounds and proficiency levels of its speakers, who are engaging in specific interactional encounters, which can be temporary and clearly task-based (such as asking for directions) but might also be more stable and general (such as roommates communicating through a lingua franca) (see Mauranen Reference Mauranen2018: 107–8). Of course, this does not imply that morphosyntactic aspects cannot be investigated in ELF contexts; however, it highlights the necessity of suitable methods and an awareness of the transient nature of ELF when working in the field of ELF grammar. As Pitzl highlights, interactants in ELF situations bring their own individual multilingual resource pools, consisting of ‘all the linguistic resources … [they] have at their disposal’ (Reference Pitzl2016: 298), but they also have access to the shared multilingual resource pool which emerges during the encounter itself and is constantly modified and adapted (Reference Pitzl2018: 34–6; see also Hülmbauer Reference Hülmbauer, Mauranen and Ranta2009: 325). In other words, ELF speakers tend to ‘bridge lingua-cultural boundaries, resulting in code-mixing and foreign language use, and the use of forms or form-function relationships which are non-codified’ (Osimk-Teasdale & Dorn Reference Osimk-Teasdale and Dorn2016: 373).
Thus, rather than merely listing features which seem to be diverging from native speaker ‘standard English’ – whatever this term is supposed to mean – recent research in ELF primarily focuses on the strategies interactants use to establish mutual understanding, that is, on how they ‘effectively employ language in communication by drawing on a set of resources and … strategically use them in interaction’ (Vettorel Reference Vettorel2019: 184). This includes, for instance, accommodation processes in general but also very specific strategies like restructuring or reformulating on the side of the speaker and repair initiations, such as requests for clarification, on the side of the hearer(s) (e.g., Wong Reference Wong2000: 247; Cogo Reference Cogo, Mauranen and Ranta2009; Vettorel Reference Vettorel2019). However, there has been considerably less research on the question why a certain strategy is preferably selected in specific interactional contexts, and, in particular, larger quantitative investigations of this topic are still missing (but see Neumaier Reference Neumaier2023a, Reference Neumaier, Wilson and Westphal2023b, who looks at the interface of conversational patterns and grammar in Southeast Asian interactions).
Although World Englishes research typically focuses on long-term situations of language contact, the question of how grammatical conventions come into being through language use is central to its research agenda, and the methodological approaches it has developed might be fruitfully applied to ELF contexts. When describing the emergence of creoles, Mufwene lists several aspects which seem to influence whether a feature is likely to be selected from a common pool of multilingual repertoires, including its ‘statistical frequency, semantic transparency, regularity, salience, [and] social status of the model speakers’ (Reference Mufwene2008: 19). Mufwene’s concept of the linguistic feature pool strongly resembles the idea of a shared multilingual resource pool advocated in current ELF research (as discussed previously). In fact, as Thompson puts it, ‘in view of the highly diverse interaction environments that embed ELF, we might expect that not only the extent but the capacity for exploitation is also greater – the virtual language, or resource pool, is bigger – than it is in non-lingua-franca contexts’ (Reference Thompson2017: 214). That this is the case has already been illustrated in previous studies on minus-features in ELF conversations, for instance, on copula usage by (South)-East Asian speakers (see Leuckert & Neumaier Reference Leuckert and Neumaier2016).
Accepting that ELF grammar is inherently adaptive and in a process of constant restructuring means, of course, that any investigation of syntactic patterns must be able to take this fluid nature into account. How, indeed, can one analyse something which is influenced not only by various linguistic inputs but also by the context-specific contingencies of the very specific communicative situation, where the focus is – in many cases – not predominantly on grammatical correctness but rather on intelligibility? The concept of ‘emergent grammar’ (Hopper Reference Hopper, Auer and Pfänder2011) as well as the related idea of ‘probabilistic grammar’ put forward by Szmrecsanyi et al. (Reference Szmrecsanyi, Grafmiller, Heller and Röthlisberger2016) have been proposed as ways to account for this complexity and will also form the underlying framework for the present study. These conceptualisations of grammar as something dynamic go together well with the framework developed by Pham et al. (Reference Pham, Leuckert, Dreschler, Götz, Günther, Kircili, Lange, Mycock, Neumaier and 329Rüdiger2024), as they also regard non-canonical syntactic constructions as logical products of language usage and acknowledge their theoretical potential to develop into unmarked constructions in specific contexts. Emergent grammar has been described as ‘an alternative to the standard lexical-item-and-rules model of linguistic description’ (Hopper Reference Hopper, Auer and Pfänder2011: 23). In communication, speakers resort to their individual linguistic feature pool which consists of both individual items but also formulaic constructions and fragments, and the appliance of grammatical rules is subject to ‘different degrees of adaptation to meet syntactic constraints and the requirements of context’ (Widdowson Reference Widdowson1989: 135). Importantly, this does not imply that grammar cannot be investigated systematically; on the contrary, it is regarded as essential to establish a basis of analysis. The concept of emergent grammar acknowledges, however, that
when the study of language is directed toward spoken conversational interactions, the relevant results of traditional linguistics are soon exhausted. It is understood that categories don’t exist in advance of the communicative setting. Instead, they are constantly being elaborated in and by communication itself. They are unfinished and indeterminate. … Emergent Grammar focuses on the boundaries of categories rather than their prototypes, exploring the leading edges and the territory around them as they move.
Although Hopper’s statement does not refer to interactions between non-native speakers per se, we suggest that the concept of emergent grammar is well suited to investigate the ‘syntactic borderlands’ researchers enter when dealing with ELF conversations.
Probabilistic grammar models embrace this usage-based approach but also ‘incorporate statistical regularities derived from experience [of the speakers, and] … associate these quantitative patterns not (only) with surface forms or lexical items … but with abstract features or constraints’ (Grafmiller et al. Reference Grafmiller, Szmrecsanyi, Röthlisberger and Heller2018: 3). This gives rise to three predictions:
(a) The influence of certain cognitive factors on quantitative syntactic variation in across [sic] different (sub)varieties of a given language should be relatively stable in terms of the directions of those factors’ influence. (b) Subtle variation in the types and frequencies of constructions will lead to gradient, yet detectable differences in the strength of different factors’ influence on speakers’ syntactic choices. (c) This variation in the use of specific constructions may be driven by stylistic preferences among registers or speakers, by situational forces such as language/dialect contact, by cognitive pressures related to language processing, or by normal dialectal drift.
To date, probabilistic approaches have typically been concerned with grammatical patterns in second-language varieties of English. Heller et al. (Reference Heller, Bernaisch and Th. Gries2017), for instance, compare the use of the genitive in Asian Englishes with British English. Recently, however, researchers have started to acknowledge the enormous potential of probabilistic models for the study of ELF (Deshors Reference Deshors2020), revealing that ‘ELF is not only a discourse-driven and fuzzy phenomenon’ but exhibits underlying – and statistically measurable – systematicity while at the same time ‘ELF users not only react passively to ongoing processes but also contribute to create patterns within the core grammar of English’ (Laitinen Reference Laitinen2020: 440).
14.2.2 Minus-Plurals in ELF
In this chapter, we suggest using a quantitative approach to investigate some of the intra- and extra-linguistic factors influencing the use of a specific non-canonical syntactic construction, the lack of overt morphological plural marking in nouns, in ELF contexts. Examples (1) and (2) illustrate the phenomenon.
there are so many ethnic (.) group
handle five beers but i can- cannot handle three beers and two coffee
Thus, while in (1) the head of the noun phrase (NP), group, is clearly semantically marked for plural by the preceding quantifier many as well as the copula, it is not morphologically marked as such. Similarly, in (2) the head noun coffee also remains morphologically unmarked, even though it is modified by a numeral quantifier indicating plurality – and even though the speaker explicitly marks plural in the preceding NPs (five beers, three beers). In the description of the selected feature, we rely on Rüdiger’s (Reference Rüdiger2019: 48) classification, illustrated in Table 14.1. The main reason for using the term ‘minus-plurals’ instead of a more common one is its relative neutrality, since other labels, such as ‘omission’, ‘lack’, or ‘underuse’, are, to a degree, associated with more prescriptive ideas (consciously or subconsciously). Our decision to focus on minus-plurals is directly connected to our research questions: minus-plurals are a well-known feature of many varieties of English (see, for instance, eWAVE feature 57) and may result from language contact or represent simplification tendencies. In research on Asian ELF, minus-plurals have been reported regularly (e.g., Thompson Reference Thompson2017; Ji Reference Ji2016; Kirkpatrick Reference Kirkpatrick2010).
Table 14.1Long description
The table is divided into 4 columns with the labels Superordinate term, Subordinate terms, Other labels, and example. The data is filled in a row, where the subordinate term has two subdivisions: plus and minus. The data from left to right is filled as follows:For the term non-canonical use:
The relevant data for plus plural are superfluous items, overuse, and two childrens.
The relevant data for the minus plural are omission, lack, underuse, and two coffee.
14.3 Data and Method
Following the theoretical contextualisation of ELF and minus-plurals in the previous section, we now introduce the dataset, consisting of segments of ACE and VOICE, as well as our methodology.
14.3.1 ACE and VOICE
In order to trace minus-plurals in spoken ELF conversations from different contexts, we analysed the files containing non-institutionalised, spontaneous language in the Asian Corpus of English (ACE) and the Vienna-Oxford International Corpus of English (VOICE). ACE and VOICE each represent 1‑million-word corpora of naturally occurring face-to-face interactions of ELF speakers, with VOICE predominantly featuring European and ACE featuring Asian speakers from countries that are part of the Association of Southeast Asian Nations (ASEAN)Footnote 1 as well as China, Japan, and South Korea. The corpora can be accessed freely via http://corpus.ied.edu.hk/ace/ and https://voice.acdh.oeaw.ac.at/, respectively. ACE has been modelled to match the compilation procedure and structure of VOICE, with both corpora containing conversations in educational contexts (ca. 25%), leisurely contexts (ca. 10%), professional business (ca. 20%), professional organisation (ca. 35%), and professional research/science (ca. 10%). The dataset we used is summarised in Table 14.2, which also gives the figures for the tokens of overtly marked plurals and minus-plurals (see Section 14.3.2 for further details on token extraction and annotation). To reach roughly equal word counts, we extracted conversations from two interactional contexts in ACE (leisure and education) and one interactional context in VOICE (leisure).
Table 14.2Long description
The table is divided into 3 columns with the labels Asian Corpus of English or A C E, and Vienna-Oxford International Corpus of English or V O I C E. The data is filled in 4 rows from left to right as follows:
The relevant data for word count is c a. 76000 words and c a. 72000 words.
The relevant data for the number of speakers is 36 and 63.
The relevant data for overtly marked plurals is 1029 and 917.
The relevant data for minus-plural is 256 c a. 20% of all cases and 31 c a. 3% of all cases.
It is important to note that we did not normalise frequencies since we focus on the relative proportions of minus-plurals to overtly marked plurals, which means that frequency normalisation would not provide further insights.
ACE and VOICE are compiled with the goal of accessing natural ELF conversations, meaning that L1 speakers of English and speakers at all levels of proficiency are generally included. Consequently, it is important to stress that neither ACE nor VOICE are corpora of learner English. Just like Osimk-Teasdale and Dorn point out in their paper on Part-of-Speech tagging in VOICE, the ‘goal [is] not to tag “errors” or to register degrees of systematicity or conformity in reference to established norms of conventional usage’ (Reference Osimk-Teasdale and Dorn2016: 374) but to find out to what extent variation in plural marking in ELF contexts can be explained by language contact and by relevant interactional strategies.
14.3.2 Token Retrieval and Annotation
After deciding on a dataset for analysis, we manually identified and tagged all relevant constructions by reading through the corpus files and tagging each regular plural noun that is morphologically marked in Standard English, that is, with the inflectional {-s}-ending. In order to be able to make sound statements about the relative frequency of minus-plurals, we tagged both overtly marked plural forms as well as minus-plurals. Focusing on variation this way allows investigating ‘alternate ways of saying “the same” thing’ (Labov Reference Labov1972: 188). We annotated all relevant tokens yielded by the process described above for the predictor variables listed in Table 14.3.
Table 14.3Long description
The table is divided into 2 columns with the labels predictor variable and levels. The data is filled in 7 rows from left to right as follows:Sociolinguistic variables:
The data for the age group is young, mid, old, and n or a.
The data for gender is male, female, and n or a.
The data for Speaker’s L 1 is language code according to I S O 639-3.
The data for language family is given in Table 14.4. Intra-linguistic variables:
The data for quantifiers in N P is quantifier or no quantifier.
The data for animacy of the head noun in N P is animate or inanimate.
The data for corpus is A C E or V O I C E.
We included four variables that can be described as sociolinguistic in nature. Age is potentially relevant because of apparent-time change (e.g., Chambers Reference Chambers, Chambers, Trudgill and Schilling-Estes2004) that may be visible in the data, that is, there might be differences between generations due to younger speakers adopting new features or features disappearing or changing over time.Footnote 2 Gender, in turn, is relevant because female speakers are known as innovators in language development, which means that minus-plurals could possibly be more frequent in their language use. Corpus-linguistic methods have been identified as highly useful to investigate gender in World Englishes contexts as they contribute to ‘obtain[ing] a sociolinguistically and empirically valid picture’ (Bernaisch Reference Bernaisch and Bernaisch2021: 5).
Both speaker L1 and language family are crucial in that they allow us to discuss the potential impact of language contact on occurrences of minus-plurals. Terassa, for instance, found that ‘speakers of HKE [Hong Kong English] might omit the plural suffix because Cantonese does not inflectionally mark its nouns for plural’ (Reference Terassa2017: 139), although she identified low rates of minus-plurals in Singapore English (SgE) and Indian English, whose substratum languages do have inflectional plural marking. Language families may also provide insight, since language areas can have significant effects in statistical modelling (see, for instance, Bentz Reference Bentz2018: 121).
We would like to point out that we used the language codes according to the ISO 639-3, which is organised by the Summer Institute of Linguistics (SIL). The SIL, which also publishes the Ethnologue, has been criticised for its missionary activities and at times questionable language naming practices, but the ISO language codes offer a tool to quickly refer to languages with clearly identifiable codes. The language families, the languages, and the language codes relevant for our data are listed in Table 14.4. We decided to use sub-groups for the Indo-European languages, since we otherwise would have had to exclude language family from the statistical analysis due to an extreme skew towards Indo-European (IE) in the dataset.
Table 14.4Long description
The table is divided into 2 columns with the labels Language families, 17 branches, and Languages and language codes I S O 639-3, 31 languages. The data is filled in several rows from left to right as follows:
The data for the branch Afro-Asiatic or A f r is Maltese, m l t.
The data for the branch Albanian, A greater than I E,) is Albanian, a l b.
The data for the branch Lolo-Burmese, B is Burmese, b u r.
The data for the branch Dravidian, D is Tamil, t a m.
The data for the branch Finno-Ugric, F is Finnish, f i n.
The data for the branch Germanic G greater than I E, is Danish, d a n, Dutch or d u t, English or e n g, German, g e r, Norwegian, n o, Swedish, s w e.
The data for Indo-Aryan, I n greater than I E, is Hindi, h I n.
The data for Indo-Iranian, I r greater than I E, is Iranian, i r a.
The data for Japonic, J is Japanese, j a p.
The data for Korean, K, is Korean k o r.
The data for Malayo-Polynesian, M is Cebuano c e b, Filipino f i l, Indonesian Malay i n d), Tagalog t g l.
The data for Romance R greater than I E is Catalan c a t, Italian i t a, Spanish s p a.
The data for Sinitic S is Cantonese or Yue y u e, Mandarin c m n, Putonghua p u t.
The data for Slavic S l a greater than I E is Czech c z e, Polish p o l, Serbian s r p.
The data for Tai T a is Thai t h a.
The data for Turkic T is Kyrgyz k I r.
The data for Vietic V is Vietnamese v i e.
In addition to the sociolinguistic variables, we considered the presence or absence of a quantifier in the NP and the animacy of the head noun in the NP as potential intra-linguistic predictors. The role of quantifiers in SgE in relation to plural inflection, for instance, ‘has been examined in a number of accounts … with contradictory findings’ (Terassa Reference Terassa2017: 119). We decided not to include usage frequency of the nouns as a variable (as Terassa Reference Terassa2017 does in her study on plural inflection in Asian Englishes), since our dataset was relatively small while still allowing for inferential modelling. Animacy has been included since ‘human nouns are more likely to have plural marking than non-human (especially inanimate) nouns’ (Haspelmath Reference Haspelmath, Dryer and Haspelmath2013 in WALS) across all languages and animacy may therefore be a relevant predictor.
The final predictor in our model was corpus, that is, if the token occurs in ACE or VOICE. This predictor was included to find out if this essential distinction causes any significant splits in the model and since the two corpora, while compiled with comparability in mind, are different in some fundamental ways. First, they have been compiled in different timeframes – VOICE data were recorded between 2001 and 2007, while data collection for ACE started in 2009, with the corpus being released in 2014. This means that the corpora do not represent language use in their respective regions at identical periods of time. Second, and more importantly, the dominant language ecologies differ in the corpora and, by extension, in the regions they were created in. These differences are accounted for to an extent by the language families and speaker L1s, but the corpora are better suited to represent Asia and Europe at large (although ELF is by nature multilingual and there are no clear-cut boundaries). Third, and again in spite of the intended comparability, the two corpora are two different datasets created by different teams and in different situations.
14.3.3 Statistical Approach: RePrInDT
We subjected the variables described in the previous section to a statistical analysis using conditional inference trees. Tree-based methods, such as conditional inference trees and random forests, have gained in popularity in recent years and their increasing use is symptomatic of what has been called the ‘quantitative turn’ in linguistics (see Kortmann Reference Kortmann2021). Such methods ‘function by repeatedly splitting data sets up in two parts such that the split leads to the best increase in terms of classification accuracy or in terms of some other statistical criterion’ (Gries Reference Gries2020: 618; see also Bernaisch et al. Reference Bernaisch, Th. Gries and Mukherjee2014; Lohmann Reference Lohmann2013; Tagliamonte & Baayen Reference Tagliamonte and Harald Baayen2012). Gries (Reference Gries2020) addresses the widespread conception that tree-based methods are ‘easy to interpret’ even though they offer various pitfalls. Key issues in this context are that ‘there can be patterns in data that make trees underperform considerably when it comes to accuracy, variable importance/parsimony, and effects interpretation’ (Gries Reference Gries2020: 644). An advancement of the conditional inference trees available, for instance, as part of R’s partykit package (Hothorn & Zeileis Reference Hothorn and Zeileis2015), is Weihs and Buschfeld’s (Reference Weihs and Buschfeld2021) repeated undersampling in PrInDT (RePrInDT), which is itself a further developed version of PrInDT. RePrInDT improves conditional inference trees in R by undersampling the larger class in the dependent variable (in our case: 15% of overtly marked plurals) in order to achieve better prediction rates in the trees (cf. Weihs & Buschfeld Reference Weihs and Buschfeld2021: 5). This function is particularly useful in such cases as ours, in which one class is (much) bigger than the other and, as a result, prediction would always favour the larger class.
The script runs numerous trees and provides information on all results as well as the three trees with the highest accuracy. Furthermore, a big advantage is the in-built calculation and plotting of the balanced accuracy, that is, the accuracy for both the larger and smaller class in the predictor variable. We decided to work with RePrInDT since, just like in Weihs and Buschfeld’s (Reference Weihs and Buschfeld2021) case study on zero and realised subject pronouns in SgE, there is a strong imbalance in our dataset due to the much higher frequencies of overtly marked plural forms. It is important to note that, at an earlier stage, we had included speaker as a variable in the tree but decided to discard it since it did not appear as an important predictor variable in any of the best-performing trees.
14.4 Results: Minus-Plurals in ACE and VOICE
In this section, we first present the results of the statistical analysis based on RePrInDT before moving on to a qualitative analysis highlighting usage contexts of minus-plurals in ACE and VOICE.
The best acceptable conditional inference tree yielded by RePrInDT is built on 579 observations and has eight terminal nodes, as can be seen in Figure 14.1. Speaker L1 shows to have the strongest effect and is responsible for the first split in the tree. From this first node, the tree splits into two further branches, both of which select corpus as the strongest predictor, thus separating ACE from VOICE speakers. If further splits occur, speaker L1, language family, or age play a role. Somewhat surprisingly, the purely intra-linguistic factors animacy and quantifier in the NP do not appear as important predictors in the model.
Figure 14.1 Best acceptable conditional inference tree
Figure 14.1Long description
Tree diagram with multiple nodes representing a decision-making process based on the variable speaker.L 1 with a significance level of p less than 0.001 at the top node, node 1. The tree branches into two main paths from node 1, evaluating different conditions related to the speaker's language background. As the tree progresses downward, each node is labeled with conditions, sample sizes n, and p-values, which provide insight into the significance of each decision point. Node 4, with n equals 56, node 5, with n equals 160, node 7, with n equals 21, node 8, with n equals 14, node 11, with n equals 69, node 12, with n equals 57, node 14, with n equals 21, and node 15, with n equals 181, are terminal nodes. These nodes are followed by bar charts that split into two segments: one representing plus and the other representing minus. The vertical bars display proportions, with values ranging from 0.0 to 1.0 on the y-axis.
A closer look at the left branches of the tree reveals that node 2, corpus, splits the group into Southeast Asian and European speakers of ELF.Footnote 3 For the ACE group, minus-plurals are most strongly predicted if the speakers’ L1 is Burmese, English, Thai, or Vietnamese (node 5). This finding is partly expected, as the majority of these languages does not code plural via suffixes but rather relies on numeral classifiers or reduplication (cf. Iwasaki & Ingkaphirom Reference Iwasaki and Ingkaphirom2005, on Thai) instead – according to WALS (feature 55A), numeral classifiers are obligatory for plural marking in Thai, Vietnamese, and Burmese. English seems to be the odd one out in this group, but it should be kept in mind that several Southeast Asian speakers indicated English as their L1; hence, it is likely that English in ACE refers to a Southeast Asian variety rather than Standard American or British English. Overall, our finding thus corroborates the hypothesis that the use of minus-plurals in English might be influenced by transfer from L1 structures (cf. also Deterding Reference Deterding2007; Wee & Ansaldo Reference Wee, Ansaldo and Lim2004: 64).
Node 4 of the tree also indicates relatively high rates of minus-plurals for L1 speakers of Indonesian Malay and Cantonese/Yue. Again, this is unsurprising from a typological perspective, as plurality is not coded through suffixation in these languages (WALS feature 33): while Indonesian Malay uses reduplication for plural marking, plural is not morphologically marked in Cantonese/ Yue. As before, the high frequency of minus-plural constructions thus suggests potential transfer from speakers’ L1s. Previous research on Chinese ELF already pointed at possible L1 transfer in more formal settings of ACE (Ji Reference Ji2016), and our study provides evidence that this also seems to hold for informal speech situations. However, as mentioned above, we could not find evidence for a strong effect of an existing quantifier in the NP, which has been suggested to trigger overt inflectional plural marking in similar linguistic contexts, such as SgE (Alsagoff & Ho Reference Alsagoff, Lick, Foley, Kandiah, Zhiming, Gupta, Alsagoff, Lick, Wee, Talib and Bokhorst-Heng1998: 144).
For VOICE speakers, we find very low rates of minus-plural for speakers with English as their L1 (node 8). The relatively high rates for L1 speakers of Albanian, Italian, or Korean (node 7) are somewhat unexpected, however, and require further analysis. Again, typological transfer from the L1 might play a role: although Korean has an inflectional plural marker, for instance, plural is typically only overtly marked for emphasis and not if number is inferable from the context. In both Italian and Albanian, plural marking often involves inflectional vowel changes in the stem rather than the addition of inflectional morphemes.
On the right side of the tree in Figure 14.1, it is again corpus which is responsible for a first split in the data (node 9). For VOICE speakers, language family then splits the tree into a branch with hardly any predicted minus-plurals – this is the group of Maltese (Afro-Asiatic) and Finnish (Finno-Ugric) speakers as well as speakers of Romance and Slavic languages (node 11). Some minus-plurals are predicted if Germanic and Turkic languages are involved (node 12). For the group of ACE speakers, age turned out to be a relevant predictor: minus-plurals are predicted much more often if speakers are younger, that is, not older than 30 (node 14). Still, as this group is very small and only involves 21 participants, this finding should not be overinterpreted.
Overall, the tree has an acceptable balanced accuracy of 0.75, with a slightly better predictive power for plural marking (0.80) than for minus-plural (0.71) constructions. Two variables did not seem to have a central effect, and both of them are intra-linguistic: the existence of a quantifier in the NP and the animacy of the noun. Our findings are confirmed when looking at the second-best tree yielded by RePrInDT. As Figure 14.2 shows, the tree exhibits the same splits – speaker L1 is at node 1, followed by corpus. Language family and age of the speakers also have an effect: for VOICE, we find more minus-plurals if the speakers have an L1 which is Albanian, Italian, or Korean; for ACE, Burmese, English, Cantonese/Yue, Thai, and Vietnamese seem to trigger more minus-plurals.
Figure 14.2 Second-best acceptable conditional inference tree
Figure 14.2Long description
The tree diagram is composed of ovals connected by lines, representing a decision-making process based on probability values. The diagram is structured horizontally with nodes displaying descriptions and associated probability values. At the top, node 1 evaluates the speaker. L 1 with p equals 0.001, which splits into two branches. The left branch leads to node 2 labeled corpus with p equals 0.001, which further splits into node 3 evaluating speaker L 1 with p equals 0.01. Node 3 then leads to two results: node 4 representing English and node 5 representing Albanian, Italian, and Korean. Node 2 also branches into node 6, evaluating language-family with p equals 0.035, and node 8 representing M, each with their respective sub-nodes.
On the right side of node 1 is node 9, evaluating corpus with p less than 0.001. This node splits into node 10, evaluating language-family with p equals 0.048, and node 13, evaluating age with p less than 0.001, both with further splits. Each node is followed by a bar chart showing the distribution of plus and minus, with black and gray bars representing the proportion of results at each terminal node. The diagram provides a clear hierarchical structure of decisions based on probability values and statistical significance.
This pattern is repeated when we re-run the model. Taken together, the three best trees yielded by RePrInDT all show language family, speaker L1, corpus, and age as best predictors, while the effect of the language-internal variables animacy and quantifier in the NP seems to be small. The balanced accuracy of the three best trees is 0.75, and thus acceptable though not excellent. When the ensemble of all 1,001 trees is considered, this balanced accuracy decreases slightly to 0.73. The pattern identified in the three best trees is confirmed in the overall ensemble. The language-internal predictors animacy and quantifier in the NP play a slightly greater role when all trees yielded by the model are taken into account.Footnote 4
Overall, our model predicts well. The histogram in Figure 14.3 shows the balanced accuracies of all trees from undersampling. The bold line represents the median of all balanced accuracies, which lies at 0.72. In total, 500 trees yielded by our model come with a greater balanced accuracy than this median value.
Figure 14.3 Balanced accuracies of all 1,001 trees
Although statistical modelling did not find animacy and quantifier in the NP to constitute major factors for predicting minus-plurals, a qualitative analysis still finds highly interesting patterns related to them. Generally, minus-plurals are found to occur primarily with inanimate head nouns, as in (3) or (4).
our our contract are yearly (.) contract are yearly
NPs with inanimate head nouns account for the vast majority (75%) of all minus-plural constructions. Animate head nouns tend to be marked for plural in both the Southeast Asian and the European dataset, even though minus-plurals also occur in this context,Footnote 6 as shown by (5).
but er many many tourist especially in autumn and er summer yes and in winter mhm not so much because it’s cooler and er there’s high water
Furthermore, 68% of all minus-plurals can be identified in NPs without quantifiers, as in (6), taken from ACE.
[first name13] is a person who used to work in er university of the Philippines er and er we er we are friend online
One third (32%) of all minus-plurals involve quantification of some kind, for instance via a numeral determiner, as in (7).
yeah there are two supervisor
Hence, although quantifiers seem to lead to overt inflectional plural marking in most cases, which supports Alsagoff and Ho’s (Reference Alsagoff, Lick, Foley, Kandiah, Zhiming, Gupta, Alsagoff, Lick, Wee, Talib and Bokhorst-Heng1998: 144) observation for SgE, there is still a considerable amount of utterances in which quantification of the NP is combined with the use of a minus-plural construction. One possible explanation for this tendency might be the wish to avoid redundancy, which is often described as a typical feature of ELF interactions (e.g., Cogo & Dewey Reference Cogo and Dewey2012: 89).Footnote 7 That is, speakers might not see the need for additional inflectional plural in cases where plurality is already indicated through a quantifying expression.
14.5 Discussion and Conclusion
In this chapter we analysed minus-plurals in the ELF corpora ACE and VOICE. Our main goal was to identify the frequencies of minus-plurals in conversational contexts. If frequency differences between the two corpora could be found, we further wanted to investigate whether these might best be explained by focusing on typological influence from the speakers’ respective L1s, or whether they could be due to other processes, such as grammatical restructuring in ELF interactions. By closely examining typological structures involved in ELF encounters and including language families and individual L1s, we intended to acknowledge that ‘multilingualism rather than English is to be understood as the overarching framework within which ELF communication takes place’ (Jenkins Reference Jenkins2015: 67).
The statistical analysis revealed that speaker L1 as well as language family seem to have an effect on plural marking in ACE and VOICE, which means that, at least from a statistical perspective, transfer of L1 structures is a possible explanation for occurrences of minus-plurals in ELF encounters. However, minus-plurals do not constitute a preferred choice in ACE and VOICE in general, since they are (still) strongly linked to specific L1s. It is important to emphasise again that there are notable frequency differences of minus-plurals between ACE and VOICE. Thus, while there is no ground for speaking of minus-plurals as a ‘preferred choice’ in VOICE, they might be on their way to becoming unmarked or even ‘canonical’ in Asian ELF – at least based on a frequency-based definition of canonicity. In cases where minus-plurals occur, they confirm that the ‘multilingual repertoires of ELF users, who are by definition at least bilingual, are an integral part of communication moves in general, and often used in close combination with communication strategies in the co-construction of meaning, and become part of a shared Multilingual Resource Pool’ (Vettorel Reference Vettorel2019: 203). Many Asian contact languages in ACE employ little to no affixation to mark number, and L1 emerged as a highly predictive factor across all statistical models in our study.
Another important aspect to consider is the role of ‘emergent grammar’ that we outlined in Section 14.2. Hopper finds that utterances in spontaneous spoken language ‘do not conform to fixed prior grammatical forms. Instead, they conform to norms, which means they more or less conform to recognizable patterns’ (Reference Hopper, Auer and Pfänder2011: 42). Two aspects related to this idea are relevant to our study. First, speakers in ACE are more likely to encounter minus-plurals in Asian Englishes than speakers in VOICE are in European varieties of English and, as we have pointed out for various languages in the previous section, many languages that are part of their multilingual resource pools do not overtly mark number. Thus, minus-plurals (or the general absence of morphological number marking) are indeed at least one of the norms ASEAN speakers encounter – in addition to the norms associated with the varieties that are closer to Standard English (such as Standard Singapore English) and formal classroom English. Second, the use of minus-plurals does not in any significant way hinder effective communication. If speakers are used to succeeding with the linguistic tools available to them, there is no need to resort to other strategies. Indeed, repeated usage of a linguistic feature or construction has important consequences both in the short and the long term (see Bybee Reference Bybee2006 on frequency effects). Thus, minus-plural marking, which might subconsciously be considered as effective or more effective than overt plural marking, may become one of the norms that Asian ELF speakers base utterances on in subsequent interactions. In this context, it is important to acknowledge that many ELF speakers employ strategies of second-language acquisition. These are challenging to operationalise as part of statistical modelling, but ELF is a variety used by bi- or multilinguals who may consciously or subconsciously resort to second-language acquisition strategies in order to achieve their communicative goals.
In methodological terms, we hope to have shown that sophisticated statistics can meaningfully be applied to ELF data. However, we also believe that caution is required when interpreting the statistical analysis. ELF encounters are multifaceted and influenced by a range of factors, such as the other languages that are part of the larger as well as the conversation-specific language ecology, the speakers’ individual sociolinguistic profiles and their linguistic backgrounds, processes of linguistic accommodation, etc. In addition, the linguistic feature at hand and its properties, be they phonetic/phonological, lexical, or morphosyntactic, might present certain constraints that make it more or less dynamic and, hence, challenging to analyse. It is hardly possible to operationalise all of these factors as part of a statistical procedure; instead, a mixed-method approach involving both qualitative and quantitative analyses seems most promising. In the present chapter, we addressed this problem by a close reading of the corpus, enriching the statistical analysis with insights from language typology, and scrutinising selected examples.
Finally, we share the belief that World Englishes and ELF research are not only distantly related but, instead, crucially share their interest in multilingual contact scenarios. Specifically, evolving modes of communication in the digital realm (e.g., Leuckert Reference Leuckert and Jansen2020) but also in other contexts, such as international trade and tourism, lead to ever-increasing ELF encounters that frequently involve a range of varieties of English. The concept of ‘probabilistic indigenisation’ (Szmrecsanyi et al. Reference Szmrecsanyi, Grafmiller, Heller and Röthlisberger2016) has been proposed for L2 varieties of English. At least based on our findings, this concept also proves helpful in improving our understanding of ELF encounters by framing them as dynamic and flexible but not as random. This understanding of ELF also ties in with how canonicity is viewed in ELF contexts, with some factors playing bigger and others smaller roles in the process of feature selection. In order to advance our understanding of ELF grammar as probabilistic grammar, future studies involving additional datasets and additional linguistic features will be required.
Open access