Investigating the lexico-grammatical resources of a non-native user of English: The case of can and could in email requests

1.1 Ontologies of grammar in SLA

In SLA, NNS English grammar has been typically described with reference to “deviations” from NS norms or as features of Interlanguage states on a learning trajectory towards an assumed “final state” of NS competence (cf. Han and Tarone 2014 for a recent re-appraisal). The issue of variable outcomes in grammar learning has been addressed in work on “ultimate attainment”, focusing on the assessment of proficiency following instruction (e. g. Birdsong 2004; Lardiere 2006). But whether the nature of the target final state is simply asserted (as in generativist approaches) or is ascertained using corpora (as in functionalist approaches, e. g. Ellis et al. 2014), the explicit or implicit assumption has generally been that it is monolingual NSs of a “standard variety” who provide the yardstick. A logical consequence of this enduring monolithic ontology of target grammar is that accuracy is taken as an essential component in the conceptualization and assessment of language proficiency (Hall 2014). Accuracy, i. e. the extent to which an individual’s production conforms to externally specified norms (normally NS “standard varieties”), is the main criterion for assessing grammatical attainment and, in consequence, for evaluating the relative effectiveness of different approaches to grammar instruction. Although some in SLA argue that the “A” in the “CAF” dimensions of L2 performance and proficiency (complexity, accuracy, and fluency) should also be taken to include appropriateness and acceptability (Housen et al. 2012: 4), success in learning English grammar is still generally understood in terms of the extent to which NS grammatical norms are accurately reproduced.

Cook’s (1991, 2002) notion of multi-competence represents a powerful counter-argument to the dominant position, stressing that learners/users inevitably develop and use qualitatively different kinds of competence from that of monolingual users. But as Hall et al. (2006) point out, the ontology of language assumed in multi-competence work has not been fully specified, and its characterizations of monolingual competence are essentially monolithic, leading to a view of multilinguals as “possessing two or more separate, internally consistent language systems” (p. 225). Despite pockets of non-conformity (notably Harris 1981), the dominant ontology of language in linguistics, both theoretical and applied, has until recently been one in which English (and other named languages) exist as monolithic, internally consistent community systems.

Usage-based linguistics (e. g. Tomasello 2003) is an approach that provides a radically alternative conceptualization of grammar, as overlapping sets of structural regularities determined by individual mental analyses of the frequency and distribution of form-meaning pairings in the input received during usage events. Developing a grammar involves the assembly of “more or less schematic constructions on the basis of recurring lexically specific exemplars of the construction in question” (Eskildsen 2015: 56). This developmental process has been demonstrated for several constructions in L2 English (Bardovi-Harlig 2002; Mellow 2006; Eskildsen 2008; several papers in Collins and Ellis 2009; Gries and Wulff 2009; Wulff and Gries 2011; Deshors 2014; Deshors and Gries 2014; Eskildsen 2015). Eskildsen’s (2008) study, for example, concentrates on the piecemeal, item-based development of constructions using modal can in an L1 Spanish speaker, which he traces back to specific usage events featuring the fixed expression I can write. He argues that the constructions develop as a function of the learner’s situated experience, with “no evidence of a journey along the interlanguage continuum toward structural completion and increasingly native-like conformity” (p. 352).

1.2 Community grammars and NS norms

The grammar of “standard” varieties of English is an idealized social construct, abstracted away from the behaviour of many individuals and interpenetrated with non-linguistics factors. Many scholars argue that it is in fact an ideological construct, rather than a linguistic one (e. g. Firth and Wagner 1997; Gumperz and Cook-Gumperz 2005; Davis 2013). This is a particular problem for generative linguistics, which adopts Chomsky’s (1986) concept of I-language, an individual’s mentally-represented linguistic (grammatical) resources, but in practice idealizes away from I-language to describe grammatical rules and principles on the basis of (educated) NS intuitions about “well-formedness” (cf. Hall 2013). In Cognitive Linguistics, a grammar is an inventory of “conventional linguistic units” (Langacker 1987: 62), i. e. symbolic resources which are shared (to a greater or lesser degree) by a speech community. Well-formedness is understood as the degree to which an expression conforms to linguistic convention. These cognitively-oriented approaches in mainstream linguistics assume a homogeneity of outcomes in L1 acquisition which is being increasingly questioned. Experimental evidence from a variety of constructions, over several languages, demonstrates that grammatical knowledge varies between NSs, most probably because of differences in usage experience, principally in education and literacy (Dąbrowska 2012). This implies different degrees of entrenchment of constructions across individuals, i. e. variation in the extent to which the associated neural event type is stored and accessed as a unit in memory as “an established routine that can be carried out more or less automatically once it is initiated” (Langacker 1987: 100).

To what extent can learners of L2 English be expected to come to know the “conventional linguistic units” (CLUs) of the language, and what is the nature of the knowledge developed? Deliberate learning of rules can, of course, result in declarative knowledge of aspects of the community grammar, but it is broadly accepted that such knowledge will rarely, on its own, constitute an adequate basis for effective use (cf. Bialystok 1981; VanPatten 2010). Neurobiological work on the distinct roles of procedural and declarative memory in L1 and L2 development (cf. Ullman 2014) suggests that instructed learning is less likely than “immersion-like” learning to lead to “native-like” processing. This again relates to the issue of CAF. Explicit learning in declarative memory will bolster accuracy in the CLUs used by NSs, but this is likely to be at the expense of fluency, a dimension of language use more associated with automatization (entrenchment) in procedural memory, and complexity, understood as increasing sophistication in grammar through restructuring of the implicit system. Research on the usage of English as a Lingua Franca (ELF) demonstrates that the implicit system, developing on the basis of uninstructed experience, is likely to diverge substantially from the NS norms modelled in instruction (Seidlhofer 2011, Mauranen 2012, Hall forthcoming).

ELF scholars have argued that a unique focus on NS norms as a model for learning is likely to be unhelpful for many circumstances of use (Seidlhofer 2011). Corpus work on ELF has shown that users often develop their own “natural” norms in order to communicate effectively in lingua franca contexts (Mauranen 2012:6). Such contexts, in which NS interactants are generally absent or at least in a minority, now probably account for a large proportion of global English usage events. ELF interactional data provide evidence of group norms emerging on the fly (e. g. Hynninen 2013), but the extent to which these norms may be retained by individuals across usage events, regulating the composition of their own lexico-grammatical resources, is not yet clear.

1.3 Corpus work on individual grammars

Our understanding of the nature of plurilithic Englishes is very poor, and the best prospect for detailed knowledge, namely through examination of corpus data, is as yet only very scantly fulfilled. Beyond authorship identification studies in literary criticism (e. g. Love 2002) and forensic linguistics (e. g. Coulthard 2004), there are very few corpus studies of individual NS usage. Mollin (2009) investigated the collocational preferences of Tony Blair on the basis of a corpus of transcripts of his public speaking and Barlow (2013) studied the idiolectal features of several White House press secretaries. A finding of the latter study that we will be returning to below is that while individuals tend to align their production with that of their interlocutors, they still maintain robust, idiosyncratic structural preferences which can resist patterns in the input.

There are almost no studies which focus on the lexico-grammatical resources of individual NNS English users post-instruction. The only work that we are aware of (Vetchinnikova 2015 and Vetchinnikova forthcoming) uses ELF data from NNS academic writers and a prolific blog poster to show that: (a) individual users tend to re-use multi-word units (MWUs) that have appeared in the input to which they have been exposed; and (b) individual users’ lexical patterning is more “fixed” (less variable) than at the community level. The first finding contradicts claims that in ELF contexts users show a preference for the “open choice principle” over the “idiom principle” (Sinclair 1991). The second finding, on the other hand, confirms Barlow’s (2013) observation that input is not the sole determinant of an individual’s lexico-grammatical resources. On the basis of interactional analysis in ELF contexts, Hynninen (2013) also found that NNSs maintained their “idiolectal preferences” across usage events.

1.4 An ontological framework

In the following analysis, we adopt the ontological framework for language introduced in Hall (2013) and applied in Hall et al. (2015). The relevant ontological categories are as follows.

1. I-language: the mentally represented linguistic resources of an individual (I suggesting “internal”, “individual”);

2. I-register: the I-language resources associated with use in a specific social/functional domain or genre;

3. E-language: expressed, situated utterances or texts (E suggesting “external”);

4. N-language: named linguistic resources like “(Standard) English” (N suggesting also “normed,” and “national”);

5. P-language: idealized linguistic systems independent of cognition and social use (the P stands for Platonic, as in the concept of Platonic ideal).

The categories I-, E-, and P-language were all originally introduced by Chomsky (1986). As alluded to previously and argued in greater detail in Hall (2013), the object of study in generativist linguistics is not actually I-language, but rather an idealized form of it filtered by “grammaticality judgements” which are inevitably biased by N-language conceptualizations. Hall (2013) interprets I-languages as repertoires of linguistic resources, often drawn from what are socially constructed as different N-languages, and deployed in different domains/genres of use. In bilinguals, such I-registers will be “polylingually constituted” (Jørgensen 2008) and involved in translanguaging (Otheguy et al. 2015).

1.5 Research questions and research strategy

Our object of interest is NNS I-language, studied not according to the implicitly monolithic N- or P-language principles of traditional linguistic analysis, but instead using a more psychologically plausible epistemology based on inferences from E-language, in the form of corpus data. Specifically, we wanted to address the following questions:

1. To what extent are individual NNS lexico-grammatical resources (I-language) determined by input (E-language)?

2. To what extent do they align with (E-language) genre norms?

3. What other factors of linguistic awareness and experience might explain the individual’s I-language configuration?

4. How can I-language grammars best be characterized?

Our strategy was, accordingly, to collect longitudinal usage data from an individual NNS user of English and to identify a representative lexico-grammatical feature to examine within a particular I-register. Computer-mediated communication (CMC) has obvious advantages as a source of data for such a project, principally: (a) it already exists in device histories such as email inboxes; (b) it includes both input and output; (c) it is already transcribed; and (d) much of it is genre-specific. We therefore sought to identify informants with the appropriate linguistic profile who would be prepared to share with us a large sample of their recent CMC interactions with other NNSs and NSs.

2.1 The informant and his correspondents

Our informant Antonio (a pseudonym) gave us access to the emails in his inbox for the previous 18 months and his outbox for the previous 36 months. At the time of data collection, he was working at managerial level in an industrial corporation in South Africa, where he had lived for over 25 years. His first language is Italian. He studied English at school in Italy and spent two years in the UK, returning to Italy with his English NS wife for eight years. In South Africa he learned sufficient Afrikaans for many purposes, but he reported English as his dominant language. He estimated that he used English around 60 % of the time, followed by Afrikaans at 35 %, and Italian at only 5 %. On 10-point proficiency self-assessment scales from the LEAP-Q questionnaire (Marian et al. 2007), Antonio rated himself with a mean score of: 8.75 for English; 3.5 for Afrikaans; and 8.25 for Italian. He reported having a “moderate foreign accent” in English and that he was taken as a NNS 50 % of the time. There are certainly NNS features in the corpus data, such as:

1. Subject-Auxiliary Inversion in embedded WH-clauses (e. g. Do you know where can we find the drawings?)

2. NNS prepositional complements (e. g. interested on straightening; result on the suspension)

But the samples of Antonio’s writing that we have seen use sophisticated vocabulary and syntax and appear to be communicatively effective. The fact that he worked for an Anglophone company at a very high level of responsibility confirms that Antonio is an expert user.

Antonio also provided information on a subset of his correspondents, totalling 292 individuals. In terms of L1, Afrikaans (43 %) and English (41 %) dominate. Next most frequent are NSs of Sesotho and Zulu (13 % combined), followed by NSs of several European languages (3 %). Many of his email exchanges were thus performed in an ELF context. Antonio coded the majority of his correspondents (61 %) as work colleagues, followed by service providers (22 %), friends and relatives (11 %), and work clients/suppliers/partners (7 %). These data indicate that most of the messages (89 %) were sent to and/or received from correspondents in the business domain or in personal business/service transactions. Lastly, Antonio reported knowing most of his correspondents (59 %) for less than two years, but the largest single category (37 %) was between two and five years.

2.2 The corpora

The Outbox (OB) corpus of sent mail comprises 57,291 word tokens, representing 5,661 types, in 2,179 messages sent over a period of 18 months. The Inbox (IB) corpus of mail received is larger, comprising 74,145 word tokens, representing 6,525 types, but drawn from a smaller number of messages (1,600) received over a period of 36 months (coterminous with the OB data). The messages were stripped of all headers, footers, copies of other messages, junk mail, and extensive quoted text. The OB constitutes the main corpus, with the IB as reference corpus. A second reference corpus was the EnronSent (EN) corpus (Styler 2011). This is a very large subset of a substantial collection of emails from a multinational corporation, in the public domain since 2003. It comprises almost 14 million words from over 96,000 messages. The corpus is untagged and there is no metadata on authors or timeframe. We took the OB data to be a more-or-less representative E-language reflex of Antonio’s I-language in this domain (an I-register). Similarly, we took the IB data to represent the kind of E-language to which Antonio was exposed in this domain prior to and during the period sampled, and used the EN data to identify broader international norms for the genre. Table 1 provides a summary of the corpora.

4.1 Initial learning context

Antonio’s development of the constructions will have been influenced by the instructional/usage contexts in which he encountered them during his early experiences with English. Regrettably, we do not have access to relevant data, so cannot conduct the kind of analysis that Eskildsen (2008) performed to track his informant’s development of modal can during instruction. Nevertheless, there is some research on other Italian NSs’ learning of the English modals can and could which is germane to our study. Research using cross-sectional learner corpora has reported that can is used significantly more frequently by Italian NS learners than by NSs of English, in both academic writing (Neff et al. 2003) and CMC settings (Dalziel and Helm 2008). Neff et al. suggest that their elevated numbers for can are due in part to its use in epistemic senses (e. g. “I think that this [...] can be the only way to improve our society”) where NSs would use could or might (p. 216). They attribute this to cross-linguistic influence (CLI) from the Italian modal potere, which has broader functional scope than English can. We might expect Antonio to have initially forged a strong interlingual association (cf. Hall et al. 2009) between his mental representation of the entrenched present indicative forms of potere and their perceived L2 translation equivalent can, leading to its use in constructions where NSs might use could, might, may or other modals. As Neff et al. (2003: 216) suggest, this might also be expected given that can is generally the first modal taught to learners. It may also be over-represented in teaching materials, compared with NS usage (Römer 2004).

Most accounts of lexical CLI demonstrate a weakening of associations between L1 and L2 items as exposure to the L2 increases (cf. Hall and Reyes 2009). L2 items can become functionally dissociated from their original L1 hosts as a result of learners’ participation in meaningful usage events in which the L2 forms are encountered in context. Some frequently experienced items will become stored as part of semi-fixed expressions (SFEs) as well as on their own. So, although an initial association between can and indicative potere, together with instructional experience, might make this modal more available for production than could, both modals will be experienced in the input and will gradually become stored also with associated lexical material (e. g. specific verbs) for functions which are divorced from the original meanings associated with the modals on their own. In this way, CnYV and CdYV will get stored as SFEs and directly activated for requests. But why does Antonio come to use the former construction so much more than the latter in post-instruction usage?

4.2 The pragmatics of requests

According to conventional pragmatic analysis (e. g. Leech 2014) and limited NS and NNS judgement data (Carrell and Konneker 1981), could is considered more polite than can. An informal inspection of EFL textbooks and Internet resources confirms that this is also the message that learners are likely to receive. There is a parallel alternation to CdYV/CnYV for requests in Italian, with potere in the indicative taken as less formal than the conditional, matching the way the isolated modals can and could are presented in instruction. This suggests that influence from the L1 should be positive (a case of “transfer to somewhere” in Andersen’s [1983] words), leading to straightforward initial mapping to CdYV and CnYV.

But closer analyses of usage data have revealed that can/could you V are now pragmatically unmarked alternates for requests, at least for NSs (Watts 2003; Curl and Drew 2008). Nevertheless, responding to follow-up questions after data collection, Antonio expressed the belief that CdYV was more formal/polite, and that he used it more with correspondents he did not know so well. We therefore inspected the (albeit patchy and coarse-grained) data on his relationship with correspondents to check for any correlation. ^[3] Over 75 % of both constructions were used in requests to work colleagues and service providers, who together account for over 80 % of all correspondents. The other two groups yielded numbers which were too small for meaningful interpretation; however, it is perhaps significant that CdYV was used five times more often than CnYV with friends, suggesting that the reported association of this construction with greater formality/politeness is not a factor in Antonio’s usage. This is consistent with the findings of Watts (2003) and Curl and Drew (2008) and inconsistent with the speaker’s own intuitions. ^[4]

Additionally, a correlation with length of relationship revealed that, contrary to his own intuitions, Antonio uses CnYV more with newer acquaintances than older ones, and vice-versa for CdYV. The result was significant at p < 0.001 according to a Fisher exact test. Also, presence of please was more frequent than absence in both constructions, indicating politeness across requests. Peterson et al. (2011) found that in the EN corpus, more formal language was used overall in requests vs. non-requests, so perhaps genre-specific factors eliminate any formality distinction associated with the construction alternation outside the genre. In sum, we find no evidence that Antonio’s request pragmatics can account for his preference for CnYV over CdYV.

4.3 Verb use

We turn now to an analysis of the lexical verbs used in the constructions, to explore the extent to which the OB patterns may be a function of the collocational strength of particularly frequent verbs and the overall verb frequency patterns in the input to which Antonio has been exposed and in his own output. An examination of collocational association, using the t-score measure of confidence, shows that the six most frequent verbs in the CnYV construction are particularly significantly associated with the can you bigram (t-score > 2). ^[5] For verbs in CdYV, on the other hand, only forward is significantly associated with could you. Table 2 lists the association strengths, with their t-score, by order of frequency (absolute and normalized per 100 instances to allow comparison across constructions and corpora). Of the six most frequent verb types in the OB, only two are strongly associated with the modal in the IB: assist in CnYV (t=2.7) and advise in CdYV (t=2.4). See Table 3.

The verb assist is the main exemplar for CnYV for both the OB and the IB. There are fewer types and tokens overall for CdYV, but there is still partial alignment at the high frequency end. This time advise is the most frequent verb in the IB, and it is also the second most frequent verb for Antonio in this frame. Overall, these data reveal partial coincidence of frequent verb types across the IB and the OB, but with some significant differences.

Some of the same verbs also crop up with proportionally high frequency in the EN corpus, suggesting that they are typical of the genre overall, as Table 4 shows. ^[6] But the most frequent couple of verbs across the constructions in both the OB and IB are relatively infrequent in EN: assist occurs in CnYV 12 times (normalized 0.22) and in CdYV only twice (normalized 0.07); and advise occurs 31 times (normalized 0.58) in CnYV and 11 times (normalized 0.38) in CdYV. This suggests that in Antonio’s context, there is a pattern of usage of verbs in this construction alternation which is not present in the broader genre: a set of “natural” local norms (Mauranen 2012: 6).

Overall, these analyses show: (a) considerable coincidence of frequent verbs in the constructions across corpora, but with some notable differences between OB and IB, and OB/IB and EN; and (b) an especially tight association between a handful of frequent verbs and Can you in Antonio’s production.

We now take into account the distribution of less frequent verb types used in the construction. Figure 2 plots the token frequencies of OB verbs in CnYV from highest to lowest. It will be observed that the type/token distribution at least partially approximates a Zipfian curve (Ellis 2013: 369–370), in which the highest frequency items account for a disproportionately high number of tokens: in CnYV, Antonio uses a small number of verbs a lot, and a large number of verbs only once. It has been widely documented for infants that the acquisition of constructions is typically driven by the initial dominance in production of a single, prototypical item or small set of items, typically verbs (cf. Tomasello 2003: 117–122). Evidence is emerging for similar learning trajectories in L2 (e. g. Ellis and Ferreira-Junior 2009) and is essentially what Eskildsen (2008) observed in his study of an L2 English learner’s development of can. From a usage-based perspective, learning and usage are two sides of the same coin, so patterns of comprehension and production will continually influence the configuration of a speaker’s I-language, in both L1 and L2. We may therefore expect to see item-based learning patterns recapitulated in usage patterns, such as the ones under analysis here. In line with this expectation, Vetchinnikova (forthcoming) found similar Zipfian effects with adjectives used by a NNS blogger in the it is ADJ that construction.

Figure 2:

Verb token frequencies for CnYV in the OB.

Psycholinguistic work on accommodation and structural priming (cf. Pickering and Ferreira 2008) demonstrates that speakers tend to reproduce grammatical constructions that have been receptively processed in preceding discourse, especially when they are associated with specific lexical material, typically verbs (Gries 2005). This work addresses the psychological mechanism through which input from usage events can continue to shape the dynamic language system. But usage, of course, includes a speaker’s output as well as input, and this would appear to be a major factor in Antonio’s elevated use of CnYV compared with the lower levels of usage in the IB and EN. Barlow (2013) documents similar individual divergence from community norms in his study of idiolectal variation in NS White House press secretaries. He speculates that despite the tendency to align one’s production to patterns in the input, this can be offset by “an inbuilt inertia and preference for familiar routines leading to a consistency in frequency of usage of language expressions by individual speakers over a period of time” (pp. 471–472) with the result that “the patterns of production data do not follow the patterns of comprehension data as closely as might have been thought” (p. 472). This is also consistent with Vetchinnikova’s (forthcoming) finding that individual lexical patterning in ELF appears to be more “fixed” than community norms.

Accordingly, we suggest that the main determinant of Antonio’s elevated usage of CnYV is his idiolectal preference for the exemplar assist, by far the most frequently occurring verb in CnYV for both OB and IB (see Tables 2 and 3). This preference might also be due, in part, to CLI, since the verb is cognate with Italian assistere, whereas its near-synonym help is generally translated by the non-cognate form aiutare. Assist is also a prototypical one for the construction, distilling the request function from the object of the request and shifting the latter into another construction, typically a gerund complement of a prepositional phrase headed by with. A selection of concordance lines from the OB in Figure 3 illustrate the pattern.

Figure 3:

Concordance lines illustrating the Can you assist with V-ing SFE in the OB.

Antonio’s preference for the more indirect Can you assist with V-ing over direct Can you V may also stem in part from the enhanced overall politeness/formality he adopts in this genre, as reflected in the markers of formality that emerged in the keyword analysis and his frequent use of please in both constructions. This would also account for his preference for assist, designated “a formal word” by the corpus-based Longman Dictionary of Contemporary English, over help, “the most general verb” of the synonym group. But the co-occurrence of the SFE with Sorry mate in the last concordance line in Figure 3 suggests that the construction has subsequently become dissociated from active pragmatic choices. In sum: Antonio’s original preference for more formal language in email requests (attested across the genre, as evidenced by Peterson et al.’s [2011] work on the EN data), together with the frequency of Can you assist in the email requests he received, and possible CLI from assistere, has led to a proliferation of assist in his own use of CnYV, which in turn sets up a bias for new verbs to be used in the same construction.