2.1 Introduction

This chapter presents the first of our six chapters adopting a short-text MDA-driven approach to studying discourse in learner language. The primary aim of this chapter is to interrogate the efficacy of the short-text approach to MDA, discussed in Chapter 1, through the use of a severe test (McEnery and Brezina, Reference McEnery and Brezina2022: 95–96). To do so, we undertake a micro-structural analysis of discourse at the turn level, which, in our view, constitutes a severe test as turns provide less data than would be available if we took a whole text, task or discourse unit view of the data. Moreover, if we carry out an analysis of discourse at the level of the largest micro-structure – the turn – we can gain a view of discourse at that level which we can then draw upon when undertaking a macro-structural analysis at the level of the discourse unit. When exploring the data to discern discourse functions at the turn level, we will bring a wide range of metadata into play to see to what extent the short-text MDA may give us insights into the impact of present variables on learner performance.

A secondary aim of this chapter is to provide a detailed outline of the short-text MDA procedure. For this first analysis, we provide a lot of detail regarding the dimensions themselves because, as well as exploring the dimensions observed, we also want to be clear about our method and provide to readers the data that will allow them, to a reasonable extent, to critically evaluate our findings. Subsequent short-text MDAs in this book are focused more on ‘use’ than ‘proof’, as they are the result of a similar procedure. As such, in the interests of space, we will not produce a step-by-step discussion of the steps undertaken in those later analyses, and we will trim back some of the reporting of features of the dimensions, relying on this first analysis as a template and justification for the analyses to come.

While the primary focus of this chapter is on learner discourse, it must be noted that examiners play a key role in the examination. Not only do they contribute to discourse in the corpus, they often shape it. For example, we might expect them to produce explanatory sequences when they are introducing tasks, when they tell the student that the task is changing and when they introduce a new task. The social context warrants the examiner to perform those functions, not the student. Hence, we might expect, on these grounds alone, that the discourse functions of the examiner may differ from those of the examinee at the micro-structural level. This is not the case in macro-structural analyses – there the relative contribution of the examiner and examinee may vary, but the discourse units are always co-constructed. They are not monologic, the macro-structures are a blend of the examiner and examinee discourse functions at the micro-structural level. To permit a clearer view of this, while the main focus of this chapter will be on learner language, at the end of the chapter we will briefly introduce the result of the short-text MDA of examiner turns. To contextualise the analysis to come, we begin with a discussion of existing MDA-based research of discourse in learner language.

2.2 MDA of Learner Language

For reasons outlined in Chapter 1, MDA-based studies of learner language are the closest, methodologically, to what we may achieve with short-text MDA. There are, broadly, two approaches to such studies. The first approach is a full MDA. To recap, this is where dimensions of linguistic variation are computed by subjecting the relative frequencies of numerous lexico-grammatical features across the texts of a corpus to factor analysis. This analysis uncovers a series of dimensions comprising the major patterns of linguistic co-occurrence across that corpus. These dimensions are then interpreted as continuums of functional variation based on the notion that frequent patterns of co-occurring linguistic features tend to be motivated by at least one underlying communicative function (Biber, Reference Biber1988). As discussed in Chapter 1, Biber (Reference Biber1988) used this approach on a corpus of spoken and written English and discovered six major dimensions of linguistic variation.

The second approach to MDA is not aimed at computing new dimensions of linguistic variation. Rather, it is concerned with comparing and projecting new texts or registers onto existing dimensions of linguistic variation by measuring the relative frequencies of the linguistic features used in the original analysis against those in new texts or registers. Then, based on the mean and standard deviation scores of the original analysis, factor scores of the new texts and registers are calculated (Berber-Sardinha, Reference Berber Sardinha, Berber Sardinha and Veirano Pinto2014), enabling multi-dimensional functional descriptions of the ways in which the new texts vary with respect to the registers and texts used in the original analysis.

Both approaches to MDA have been applied to corpora of written and spoken learner language. For the most part, previous research has tended to favour projecting corpora of learner language onto existing dimensions of linguistic variation, especially dimensions of spoken and written English, as a way of investigating various research questions, such as (i) how the language produced in learner interviews compares to other spoken and written English registers (e.g. Aguado-Jiménez, Pérez-Paredes and Sánchez, Reference Aguado-Jiménez, Pérez-Paredes and Sánchez2012); (ii) how spoken learner language varies according to different elicitation tasks and in different test conditions (e.g. Connor-Linton and Shohamy, Reference Connor-Linton, Shohamy, Conrad and Biber2001); (iii) how native and non-native speakers vary in their use of particular communicative functions in essay writing (e.g. Van Rooy and Terblanche, Reference Van Rooy and Terblanch2006) and during elicitation tasks in interviews (e.g. Pérez-Paredes and Sánchez-Tornel, Reference Pérez-Paredes, Sánchez-Tornel, Callies and Götz2015); (iv) how writing varies in different kinds of learners of English (e.g. Pakistani, English as a Foreign Language (EFL), English as a Second Language (ESL), English as a Native Language (ENL)) with different cultural backgrounds (e.g. Abdulaziz, Mahmood and Azher, Reference Abdulaziz, Mahmood and Azher2016); and (v) how the language produced by learners varies over time after English for Academic Purposes (EAP) instruction (e.g. Crosthwaite, Reference Crosthwaite2016).

These studies demonstrate that language use varies according to a variety of factors, including task (written or spoken). For example, Connor-Linton and Shohamy (Reference Connor-Linton, Shohamy, Conrad and Biber2001) projected the oral production interview of L1 Hebrew EFL students onto Biber’s (Reference Biber1988) dimensions of linguistic variation of spoken and written English. They examined whether the communicative functions employed by the learners varied according to different elicitation tasks: talking about oneself and two role-playing tasks (complaining about noise and requesting an extension to a deadline). They found that complaints were less narrative and much more associated with situation-dependent reference than the other tasks, as complaints typically involve talking in the present tense about things in the immediate context. Requesting or talking about the self often concerned events in the past or future. Complaints and requests were more persuasive than telling about the self. While many of the linguistic co-occurrence patterns were used by the learners for the same communicative function as the written and spoken English registers in Biber’s (1998) original analysis, Connor-Linton and Shohamy (Reference Connor-Linton, Shohamy, Conrad and Biber2001) suggest that the features associated with Dimension 5 (abstract v. non-abstract) were potentially being used differently by the learners in comparison to native speaker registers. This is in part a motivation for the study in this book, as the finding indicates the need to investigate exactly how learners use language from a bottom-up perspective by conducting a full MDA as opposed to projecting texts onto existing dimensions based on native speaker data.

However, our study is not the first to take this bottom-up approach. Although less common, full MDA has been applied to explore corpora of written (e.g. Ascención-Delaney and Collentine, Reference Asención-Delaney and Collentine2011; Friginal and Weigle, Reference Friginal and Weigle2014) and spoken (e.g. Friginal and Polat, Reference Friginal and Polat2015; Friginal et al., Reference Friginal, Lee, Polat and Roberson2017; Staples et al., Reference Staples, Laflair and Egbert2017) learner language. For example, Ascención-Delaney and Collentine (Reference Asención-Delaney and Collentine2011) investigated the major communicative functions employed in the various writing tasks of second- and third-year university-level learners of Spanish. They found three major dimensions, including narrative vs. expository, descriptive expository prose and expository prose with a stance. They then explored the overall association of the different written text types to these dimensions. For example, they found that argumentative essays were more strongly associated with ‘expository prose with a stance’, whilst narrative texts were the text type least associated with this dimension.

In another study of learner writing, Friginal and Weigle (Reference Friginal and Weigle2014) conducted a longitudinal MDA of writing from L2 students in three time periods during a semester (early in the semester, midway through and late in the semester). They identified four dimensions and then assessed the relationship between both the average assessment scores and the date when the essay was written with the average dimension score of each essay. They found patterns between the average dimension scores of the learners’ writing varied over time. For example, they found that the average score of the learners’ first piece of writing was more involved. However, over time their writing got less involved and more informational. They found that this also correlated with the students’ assessment scores. Essays from students with lower assessment scores were more involved, whilst essays that had higher assessment scores were informational.

A full MDA of spoken learner language was conducted by Friginal and Polat (Reference Friginal and Polat2015). They used the Louvain International Database of Spoken English Interlanguage, which allowed them to use data from speakers with eleven different L1 backgrounds. Each speaker in the corpus performs three tasks – a set topic discussion, a free discussion and a picture description task. The set task is drawn by the L2 speaker from three options. Friginal and Polat aimed to investigate language variation in relation to differences in both the speakers’ L1 background and the interview tasks. They identified four dimensions of linguistic variation and then compared speakers of different L1 backgrounds by averaging the factor scores of the different groups. One finding was that, overall, Swedish learners of English were considerably more involved and conversational, whereas Japanese learners were more informational. Additionally, they identified that Dutch, Swedish, Chinese and Spanish speakers were the groups most strongly associated with the communicative function ‘complex statement of opinion’. Italian was the least strongly associated with this function. Despite operationalising how speakers of different L1 backgrounds compare by averaging the factor scores of the different groups, Friginal and Polat’s comparison of the different interview tasks in relation to the dimensions of linguistic variation is far less systematic, tending to be much more subjective. This is because they have factor scores for each interview as opposed to factor scores for the interviews separated by task. Accordingly, they were not able to compute average factor scores for the individual tasks, so any association of task to function is masked. They were only able to claim that particular linguistic co-occurrence patterns were more associated with the interview tasks through close reading of a sample of interview texts. So, while this study sheds some light on the relationship between L1 background and the use of particular linguistic repertoires, it also revealed a strong need to examine the major communicative functions at a more specific level than at the whole interaction level.

Friginal and Polat’s study provides an important motivation for our use of short-text MDA to explore the Trinity Lancaster Corpus (TLC). As noted, their study suffers from a degree of data aggregation that masks an important dimension along which functions in their data may vary – task. In the TLC, tasks are clearly demarcated, allowing us to make a more systematic observation of how function varies according to task as well as other variables. By looking at this corpus of conversations between L1 examiners and L2 examinees, we can thus achieve the primary aim of this chapter by investigating the major communicative functions of learner language at a low level of aggregation – the turn level – in the corpus.

2.3 Corpus Analysis

We began our analysis by carrying out the short-text MDA procedure outlined in Chapter 1. That procedure relies, of course, on reasonably accurate part-of-speech tagging, so we undertook a short test of tagging accuracy on the data. The precision of the tagger on the learner’s turns is 0.97, which was calculated from an investigation of 100 random instances of each linguistic feature, whilst the recall rate is 0.91, which was calculated by assessing the number of false positives (incorrect tags) and false negatives (missed tags) in 100 random transcript lines. The overall F-score (a score which combines precision and recall) of the tagger on learner’s turns is 0.94.Footnote ¹ We deemed these results sufficiently accurate to proceed with our analysis, though in interpreting our results we were, of course, aware that errors in tagging could, in principle, impact our results. In particular, throughout this book we were mindful of this when undertaking the close reading of examples to interpret the dimensions that we explore.

After tagging, each turn in the training and test datasets was automatically analysed for the presence or absence of the 130 linguistic features encoded for our analysis.Footnote ² These results were recorded in a categorical data matrix, where each row represented a turn in a transcript and each column was a linguistic feature. Le Roux and Rouanet (Reference Le Roux and Rouanet2010) advise that very infrequent features (e.g. those that occur in <5 per cent of the data) either need to be pooled with other related features or they might need to be discarded because infrequent features can overly influence the results of the Multiple Correspondence Analysis, as they contribute more to the overall variance. Thus, in line with Le Roux and Rouanet (Reference Le Roux and Rouanet2010), features that occurred in fewer than 5 per cent of the turns were either pooled with a more general grammatical category, if appropriate, or removed from the data matrix altogether, leaving forty-three linguistic features (see Appendix A for a description of the feature set and pooling decisions). Time adverbs, for example, did not occur in more than 5 per cent of the turns. As a result, these were pooled with other adverbs into a ‘General Adverbs’ category. In circumstances where a more general category did not exist, such as with reflexive pronouns, these features were discarded from the feature set. While we will not return to a discussion of this procedure, note that it is used in each of the short-text MDAs presented in this book.

We identified a number of features that may be productive in distinguishing the language use of the learners, including the speaker’s L1, country of origin, grade/proficiency level and overall mark, as well as various combinations of this information. Each turn from a learner was then analysed for the presence or absence of these features, as well as the interview task the turn came from (Greeting, Listen, Conversation, Discussion, Interactive Task, Presentation) and the length of the turn (in word tokens). This metadata was added to the data matrix. The final data matrix for the whole TLC training corpus is made up of individual turns, each assessed for length (in word tokens) and the presence or absence of 191 variables (43 linguistic features and 148 pieces of metadata). This data matrix was then subjected to Multiple Correspondence Analysis in R using the ‘FactoMineR’ package (Husson et al., Reference Husson, Josse, Le and Mazet2020).

For our analysis we also needed to specify any supplementary variables; these can be qualitative and quantitative. Turn length was specified as a supplementary quantitative variable, whilst the other 147 metadata variables were qualitative. The reason for specifying turn length as a supplementary variable is owing to the possibility of text length confounding the analysis because it has not been controlled for in the analysis of the presence/absence of linguistic features. In particular, the short-text version of MDA used here does not analyse the relative frequencies of features. The relative frequencies of features are measured in standard MDA as a way to control for texts of different lengths. Measuring the relative frequencies of features, as opposed to their absolute frequency, means that texts of different lengths can be compared reliably as the frequencies of features are relative to the length of the text. Thus, by only measuring the presence or absence of features in this version of MDA, text length is not controlled for and could confound the analysis, as the more words a turn has the more likely it is to contain a variety of different linguistic features. Defining turn length as a supplementary variable enables the assessment of the degree to which turn length is correlated to the results of the analysis. Supplementary qualitative variables and turns were assigned coordinates revealing their association to the dimension patterns, enabling the assessment of the degree to which the task type and the speaker’s L1, country of origin, grade, proficiency and overall mark are associated with the communicative functions of the turns.

In the corpus, the short-text MDA revealed forty-three dimensions of linguistic variation in descending order of importance. Each category of a linguistic feature (e.g. presence of Nouns and absence of Nouns) was assigned a positive or negative coordinate on each dimension. Additionally, each category of a linguistic feature was also assigned a contribution value for each dimension. Contributions show which categories of features are the most important contributors to the dimensions. In this way, they are similar to factor loadings in factor analysis. Le Roux and Rouanet (Reference Le Roux and Rouanet2010) suggest that the categories of variables with a higher than average contribution should be interpreted, as these represent the patterns of variation with the most discriminatory power. All the contributions assigned to the categories of linguistic features for each dimension equal 100 and there are 2 categories for each of the 43 active variables (2 × 43 = 86). Therefore, the average contribution of a category is 1.16 (100/86). Categories of features with contributions above 1.16 were interpreted for their underlying function based on the notion of linguistic co-occurrence (Biber, Reference Biber1988). Unlike factor loadings, the contributions do not have polarity. Thus, to interpret the dimensions, we used the coordinates of the categories of features in combination with their contributions, which reveal the distribution of the features across the turns. In this way, features that are distributed in similar ways have coordinates close to each other, and features that are not distributed in similar ways have coordinates that are far apart from each other, that is, in opposite quadrants (Le Roux and Rouanet, Reference Le Roux and Rouanet1984, Reference Le Roux and Rouanet2010). Specifically, we consider the linguistic features with above-average contributions, but we interpret those features with positive coordinates in opposition to those with negative coordinates as a continuum of variation.

To assist interpretation, we examined the turns strongly associated with the dimensions to view the co-occurring features in context. The short-text MDA also assigned contributions and coordinates to each turn on each dimension. Similar to the categories of linguistic features, the turns with high positive and negative coordinates that most strongly contribute to the dimension were then interpreted along with the features associated with the corresponding side of the dimension for their underlying communicative function. Throughout the book, we illustrate the functions we discuss by drawing on such examples, which we will call prototypical examples. To interpret our dimensions, starting with Dimension 1, each dimension was interpreted until the dimensions were no longer readily interpretable, as is common practice in MDA. In total, for the learner turns, we interpreted six dimensions. The sixth dimension was uninterpretable, indicating the end of our analysis. The first dimension primarily represented text length and so it was excluded, leaving four meaningful dimensions that characterised our data. These dimensions account for 93 per cent of the variance in our dataset, calculated using the modified rates for short-text MDA (Benzecri, Reference Benzecri1992: 412).

2.4 Results

Each interpreted dimension is presented in the following sub-sections. Each dimension includes a table consisting of the features most strongly contributing to it and a table showing the turns most strongly associated with the linguistic co-occurrence patterns. Each example given is followed by the name of the corpus file it is taken from. Following the description of the communicative function linked to the dimension’s linguistic co-occurrence patterns, the association of the supplementary variables to the communicative function is also presented. Whilst the TLC consists of interviews from learners from over thirty-seven linguistic and twenty-six cultural backgrounds, the corpus does not always contain enough learners in a particular group to assess whether the patterns identified are predominantly a result of the learners’ linguistic/cultural backgrounds, meaning that the patterns cannot be extended to other similar learners of this linguistic/cultural background (a known issue with learner corpora, as discussed by McEnery et al., Reference McEnery, Brezina, Gablasova and Banerjee2019). For example, our corpus only includes one learner whose first language is Czech. The language produced by this learner will not only be influenced by their first language but also their grade and proficiency, among many other factors. As a result of so few learners being in this group, it would be impossible to assess if the overall association of this learner’s turns were only a result of their first language. Therefore, our assessment of variation according to the different groups of learners will be limited to those groups that have enough learners and enough turns produced by those learners to sustain our analysis. Groups were included if their turns comprised more than 5 per cent of the corpus. So, for instance, whilst there is an adequate number of learners from grade 7 that scored an overall pass, those learners do not produce enough turns in the interactive task to be assessed and compared to other groups of learners in this task, whereas they produce enough turns in the discussion task to be included and compared. As will become apparent in the next section, at the micro-structural level this still leaves us with a wide degree of variation to consider.

Although this approach factors in the absences of features, which are arguably just as important as what is present, it is hard to interpret the function of the absence of features in the context of turns as absent features are unobservable. Moreover, features which are absent and are strongly contributing to one side of the dimension tend to also be present and strongly contributing to the other side of the dimension. Thus, to avoid repetition (i.e. discussing the potential function of the absence of features on one side of the dimension and then discussing the function of their presence on the other), the absences of features are presented in this chapter, but ignored for the purpose of discussion and only the observable presences of features are interpreted.

In what follows, for each dimension explored we consider that the variation is caused by a series of variables, including task. The task variable is particularly important to consider as, should we see variation of discourse functions by task, it shows that there is an interaction between the functional analysis of utterances and the task-related functional purposes of the texts in the corpus. Given that previous studies have not been able to study this level in a bottom-up fashion well, if we are able to use the short-text MDA approach to analyse the impact of task, that would cast fresh light on a relatively neglected feature influencing learner performance.

A final note before we begin the discussion of our dimensions. Throughout the book, while a dimension is introduced with reference to its number, when discussed in the text the labels for either side of the dimension are used, rather than calling them ‘positive Dimension 1’ or ‘negative Dimension 2’ or some such. This is because we want to shift the discussion to the functional level and away from a simple discussion of rather abstract polarities and dimensions as quickly as possible. This will be done with every short-text MDA in this book. Later, when we are discussing and comparing short-text MDAs from different analyses, we also clearly note from which short-text MDA the function hails, for example, we may preface a function name with ‘learner turns’ or some other descriptive label. We do this not only because it will help readers refer back, where they wish, to the description of the function in question, but also, and importantly, because we will find that, as the book progresses, the same, or similar, functions arise in different analyses, and the short descriptive premodification of the function label thus becomes vital in distinguishing the same function relating to different analyses.

2.5 Examiner Turns

Having outlined a turn-based functional characterisation of the learner turns in the corpus, we now move on to consider examiner speech. Of course, the focus of this book is learner language, so our discussion of examiner speech will be correspondingly brief. It cannot, however, be omitted. This is because, as noted, the data in all of the corpora used in this book are generally dialogic. The functions outlined so far are not being produced by the L1 speakers at random or exclusively as a monologue. The functions are being selected with communication with an interlocutor in mind. We may assume that the choice of the functions in any turn in the L1 is produced with a variable range of intents – a response to the examiner’s speech, an intention to elicit a response from the examiner or as a move in macro-structure, such as a narrative, for example. In such a context it would be folly to try to understand learner speech without also understanding the speech of their interlocutor. This will become clearer in the next chapter. For now, we will present a summary study of examiner turns, based on a short-text MDA of them, to give a sense of how the speech of the examiners and the examinees, when viewed this way, is similar to, or differs from, each other. The analysis is summarised in Table 2.26. However, in discussing this table we will foreshadow some of the findings of the next chapter by focusing on a subset of the functions present in the speech of the examiners, which will later be shown to have a particular role to play in influencing the short-text MDA of the TLC discourse units. In the table, functions which are also present in the turn level short-text MDA of the L2 speakers are underlined.

One finding is very clear from this table. The functions of the examiner speech and those of the learner speech are quite distinct at this micro-structural level. With the exception of Narrative, there are no shared functions between the two at the level of the turn. As hypothesised at the start of this chapter, the social roles of the interlocutors in the data lead to them producing distinct repertoires of discourse functions. Those functions may interact, but they are not congruent – in the context of an exam in which they have quite distinct roles, this functional differentiation between the two speakers is as understandable as it is striking. Yet to assume that the speakers were simply distinct and to leave our analysis there would be to misconstrue the nature of discourse and to mischaracterise the interdependencies between examiner and examinee functions which, while they may appear distinct, could be mutually dependent. To begin our transition to a consideration of interaction, let us examine two dimensions, 2 and 7, from the examiner short-text MDA, which will be of importance in the next chapter. Dimension 2 splits between the descriptive and the information seeking. The following examiner turn, from the discussion task of a grade 7 exam taken by an Indian student (file 2_7_IN_11), shows an example of a turn which is associated with a Descriptive function:

The features which combine to perform the Descriptive function are apparent in this example, for instance, stative forms (sounds), contracted forms (I’ve, haven’t), demonstratives (that) and the pronoun it. The function is used by the examiner to simply outline a factual statement, in this case regarding the examiner’s viewing habits. By contrast, we find examiner turns with a function of Information Seeking on the negative side of Dimension 2. The following is one such turn from the conversation task of an Argentinian student taking a grade 6 exam (file 2_6_AR_27):

The question, seeking a statement of information from the student, orients the turn towards the Information-Seeking function, with features which combine to perform that function present including a WH-word (what), Auxiliary DO, a question mark and a second-person pronoun (you). It also tries to guide the selection of the function for the next turn by the student – given an Information-Seeking turn from the examiner, it is easy to find examples of the student being guided towards a Reveal function, even if they fail initially to produce such a turn. Consider the following exchange from a Mexican student taking a grade 6 exam (file 2_6_ME_108):

In this sequence the examiner uses an Information-Seeking function, eliciting a bare Informational response from the student. The examiner then prompts the student with another Information-Seeking response and this time, instead of using the Informational discourse function from their repertoire of turn-level discourse functions, the student produces a Reveal function instead. In the transcript the examiner then allows the student to continue with the Reveal function (the following two examiner turns are phatic). So we may, at this point, presume that examiner turns and functions, while distinct, may influence and shape a student turn, and vice versa.

What of Dimension 7? This is composed of two functions, Narrative on the positive side of the dimension and Stance Seeking on the negative side of the dimension. The Narrative function relates to turns that do not necessarily constitute a full narrative in themselves but which, in context, contribute elements to a narrative. Consider the following examiner turn, from the interactive task of a grade 7 Argentinian student:

Again, we see ample evidence of the grammatical features which bundle to produce this function. For example, a subordinator (that), third-person singular verb (wants) and a past tense verb (said). There is clearly a past orientation here, and some suggestion of a flow of events through time – this is clearly a response to a statement of some sort. When we look at the broader context, this is one of a series of such turns which together constitute a narrative. This is the discourse unit in which the turn occurs (from file 2_7_AR_16):

This makes the example slightly clearer in terms of function. The student is eliciting, across a number of turns, a narrative in response to a prompt for the task from the examiner, in this case ‘my nephew used to dress very well but now he’s totally changed his appearance I’m not sure what to think about it’. In terms of narrative, as will be explored later in the book (see Chapter 8), this is a complicating action—the genesis of the narrative. In the discourse unit just given, we have proceeded from that complicating action to the first discussion of the issue with the nephew. Following this discourse unit, the discussion proceeds through a series of interactions and issues until we arrive at a resolution, which is a, perhaps impractical, suggestion by the student that the examiner should oblige the nephew to dress better. Throughout, turns abound, which are key elements of narrative. We will not explore further here whether the Narrative function exists beyond the turn level – we have argued for one such example here, but this is certainly suggestive of the possibility that, when we analyse discourse units in Chapters 3–7, Narrative will extend beyond the micro- to the macro-structural level. What is important for now, however, is to note that a Narrative function exists at the turn level in the examiner turns.

2.6 Conclusion

This initial exploration of the data using short-text MDA at the micro-structural level (the turn) has established that the technique can be used to group turns such that, when we investigate turns, we are able to discern discourse functions that are associated with the grouped data. This indicates a link between form (the basis on which the grouping occurs) and function (the discourse functions we have assigned to the learner and examiner speech). Reflecting on our findings, we might hypothesise that the likelihood of congruence between the turn-level analysis in this chapter and the discourse unit analysis to be presented in the next chapter is low – the discourse units contain both learner and examiner speech and are co-constructed in a dialogic exchange. We saw that the set of functions used by the examiners and examinees were largely different. This, in turn, allowed us to reflect on how the role of the interlocutor, as warranted by the social context in which the interaction occurs, can influence the discourse functions they employ. Likewise, we showed, for the examinees, that task was a major context within which variation occurs. While not a principal focus of this book, we also saw how cultural background seemed to be linked to variation.

It might, at this point, be possible to argue that we can gain an adequate view of the functions of discourse by taking this micro-structural approach. However, we wish to explore, in the next two chapters, how, and whether, taking a macro-structural approach can reveal that, sitting above the micro-structural level, separate macro-structure discourse functions may be discerned. If there is a high degree of congruence between the micro- and the macro-structural analyses, then we may doubt that the discourse unit level is adding much to our understanding. If the functions at the turn level are the same as those at the discourse unit level, then while we might be able to comment on how the micro-structure level meshes with the macro-structural level to achieve this – for example, which repertoire of discourse functions at the turn level coalesce to form a specific function at the macro-level – our view of the functions of discourse would not change. We would have observed those functions at the micro-level. On the other hand, our second goal will be achieved if we find there is no, or imperfect, congruence between the functions revealed in our micro- and macro-structural analyses, a point indicated by the differences identified in examiner discourse.

Such a view would help not merely to show how the levels coalesce, it would also show how higher, macro-level functions are projected from the functions at the micro-level, potentially revealing that the two may be relatively distinct functionally, but they are closely linked structurally. If we do discern such functions, and can argue that they are independent wholly or by degree of the functions at the micro-level, we may also see whether those discourse functions in turn are sensitive to the role of interlocutor, proficiency and task. Exploring these ideas will be the focus of the following two chapters.