Initial studies on the acquisition of two Spanish copulas, ser and estar, emerged from a research agenda exploring whether second language (L2) development was driven by universal mechanisms manifested in stages rather than an accumulation of entities (Rutherford, Reference Rutherford1987). Concerning the Spanish copulas, the stages of acquisition through which Spanish L2 learners pass seem – with a few exceptions – to be largely uniform (e.g. estar + locative, ser/estar + adjective; VanPatten, Reference VanPatten, VanPatten, Dvorak and Lee1987; Ryan & Lafford, Reference Ryan and Lafford1992; Geeslin, Reference Geeslin, Leow and Sanz2000). Cheng, Lu and Giannakouros provide one of the first studies which partially test the assumption of universality, examining Mandarin Chinese speakers' copula behaviors over different levels of development. The work of Geeslin as well as Cheng, Lu and Giannakouros is valuable not so much because it provides insights into copula development but rather because this work is building a model for predicting the interaction in general terms between aspectual, semantic, and pragmatic factors and the acquisition of both grammatical and lexical L2 phenomena.

To contribute to this discussion, I focus on two issues. First, I will argue that the Cheng et al. data begs us to consider expanding the scope of this model-building enterprise even beyond the consideration of the semantics and pragmatics of copula frames (i.e. ser/estar + adjectival/adverbial complement) to a consideration of the interaction between discourse level features and learner performance. Second, I caution future researchers to carefully consider the assumptions underpinning the use of both binomial and Poisson distributions when choosing their analytical tools.

Like Geeslin (Reference Geeslin, Leow and Sanz2000), Cheng et al. found, for the most part, an increasingly important role for semantic and pragmatic factors with more advanced Mandarin Chinese L1 learners of Spanish. Copula usage at the initial stages was largely a function of the immediate syntactic context, thus the high influence of adjective class on copula choice at the initial stages of acquisition and the greater influence of formulaics such as estar infectado “to be infected”. As learners begin to process increasingly larger amounts of information, the wider, discursive context (e.g. the interplay between these local copula frames and discourse type) appears to become a consideration, at least in writing. The importance that Cheng et al. attribute to the discursive nature of the writing context on copula usage is intriguing because it underscores the close relationship between discourse type (broadly speaking, e.g. narrative, descriptive, expository) and grammatical phenomena. Cheng et al. note that the discursively more abstract register of “exploratory writing” evoked greater estar + adjective usage. Their account of this observation asks us to consider the interplay between communicative demands and learners' processing abilities, such that greater processing demands might lead to more estar + adjective use. They note that estar + adjective might be more compatible with the semantic and pragmatic goals of narratives or descriptions (which, one would think, would entail more considerations of the states of affairs of referents or background changes).

I do not view Cheng et al.'s results as providing support for the theory that greater processing demands (like those of an expository essay) necessarily lead to the use of more complex grammatical frames, such as estar + adjective. Although some theoretical perspectives predict that communicative pressures lead to more complex output (Givón, Reference Givón and Slobin1985), certain task demands must be in place for learners to intentionally produce their most complex grammatical tools to affect meaning (Robinson, Reference Robinson2005). It is not apparent that the consequences of not communicating well in these tasks were important considerations for these learners. What is more likely is that estar + copula usage appears in texts not only of a narrative nature but also those of a persuasive one.

To be sure, to explore the extent to which estar + adjective usage is a function of discourse type and not just local semantic and pragmatic variable, I sampled one thousand texts at random from the twenty-million-word Corpus del español (see Biber, Davies, Jones and Tracy-Ventura, Reference Biber, Davies, Jones and Tracy-Ventura2006) for all occurrences of ser + adjective and estar + adjective. As is customary in corpus research, the frequency of each copula frame was normed across mode (oral and written) and discourse type (e.g. editorials, news, literature, speech) for number of words per text (which total 5.43 million words for the sample), providing an insight into the relative use of each frame by mode and register independent of how much Spanish was produced within each level. A loglinear analysis indicated that there was no three-way interaction between mode, discourse and copula type (ser or estar) + adjective (χ²=0.0; df = 103; p = .99), but that there were significant two-way interactions in the native-speaker corpus (χ²=888.7; df = 51; p < .001). Of the two possible two-way interactions, one significant pattern was, perhaps, obvious, with mode and discourse interacting significantly (partial association: χ²=835.4; df = 25; p < .001); both ser + adjective and estar + adjective were more frequent in any given 1000 words in speech than in written language by a ratio of 1.8–1. What is more important to my point here is that the loglinear analysis also uncovered a significant interaction between copula frame and text type, which I interpret to be a reflection of discourse type (partial association: χ²=52.9; df = 25; p < .001). Ser + adjective was relatively frequent in almost all discourses, while estar + adjective was most frequent per 1000 words in dramas (which entail much evaluative language and monologues). Furthermore, estar + adjective was notably present in history, editorial, scientific and technical texts, suggesting that it is an important tool not only for expressing one's perspective of a referent (evaluative or in terms of a comparison, as may be the case in drama) but also for advancing a writer's argument. In other words, a consideration of native-speaker speech together with the learner patterns that Cheng et al. report suggests that grammatical phenomena occur in learner production at slightly more advanced levels not only for the immediate semantic characteristics surrounding the copula frame, but also because the communicative goals necessitate the function that the frame can serve.

The work that Geeslin and her colleagues have carried out to essentially expand our “models” of learner development to include sociolinguistic factors has been novel to say the least. However, perhaps for the sake of refining the methodology she has developed, it has also ignored a phenomenon that corpus-based linguistics has started to make increasingly salient. Bundles of lexico-grammatical structures have tight associations with types of discourses. Sometimes the associations are not obvious (e.g. while the preterit and imperfect serve narrative functions, research has shown that third-person pronouns and morphology do as well; see Biber et al., Reference Biber, Davies, Jones and Tracy-Ventura2006), and some phenomena occur at a higher rate than others although the entire set serves the discursive goals (e.g. preterit/imperfect forms predominate in narratives while third-person phenomena are less frequent although just as necessary). For example, with the data presented by Cheng et al. and with the close association between form and (discursive) function that has been uncovered as of late, the challenge for the research focused on copula frames is to provide insights into the lexico-grammatical phenomena that accompany copula usage and how they too interact with copula and adjective choice. Not only would such an expansion of the research agenda complement the insightful analyses published to date, it would also provide a level of objectivity to an enterprise that depends much – to date – on rater judgments of the semantic intent of copula frames.

Cheng et al. employ a Poisson regression analysis to study the relationship between the discursive, semantic and pragmatic variables and their effect on estar + adjective use. This approach deviates from previous regression approaches to the estar + adjective issue in an important theoretical way (see Geeslin, Reference Geeslin, Leow and Sanz2000). The Poisson regression analysis compares sampled behavior not to the binomial distribution but rather to a distribution that assumes that occurrences over some continuum (in this case, either learner level or – most likely – types of discourse) are random and considered unanticipated/unlikely. The Poisson distribution presumes that either estar + adjective frequency is stochastic or that it adheres to the law of small numbers, meaning that where it occurs in a given composition is always improbable (see Oakes, Reference Oakes1998). That is, the use of a Poisson regression implies a priori that there is no principled reason to suspect that estar + adjective is likely to occur in almost any context. There are important reasons for approaching an analysis of a linguistic data set to study copula frames using this sort of analytical tool with caution. As argued above, estar copula frame frequencies are predictable within certain discourse types.

The Poisson distribution is commonly used in computational linguistics for understanding the relationship between texts and the occurrence of lexical phenomena (technically speaking, lemmas). The assumption that a lexeme's occurrence is random can be particularly useful when, for instance, attempting to identify whether the use of a particular class of words is (un)likely to have been produced by just anyone or by a particular author (i.e. in author identification research). However, the copula frame estar + adjective seems, based on the mini corpus analysis presented in point 1, to be quite predictable within certain text types. Furthermore, small frequency rates do not imply that stochastic mechanisms are at play, especially in the case of grammatical phenomena. The corpus work of Biber et al. (Reference Biber, Davies, Jones and Tracy-Ventura2006) shows that in Spanish small frequencies of a given phenomenon often constitute enough of a critical mass to forward the rhetorical goals of the speaker/writer (if other associated phenomena are also present), and so the phenomenon's occurrence while barely “audible” or “detectable” is in no way “unlikely”.

It may be argued that the Poisson regression analysis is, nevertheless, a fair methodology because the distribution of estar + adjective frames is not normally distributed across the samples that a researcher analyzes. Caution is in order when a linguist fears that his or her sample frequencies do not fit well to the binomial distribution. First, research on (parametric) analyses of variance have shown that linear regression methodologies, for instance, are quite tolerant of compromises to their basic assumption about sample distributions (Tabachnick and Fidell, Reference Tabachnick and Fidell2001). Second, linguistic phenomena are almost always positively skewed, with sample means lying below the 50th percentile. A good deal of research in computational and corpus linguistics has shown that parametric methodologies of analysis make very fair predictions about natural language behavior, and we may be able to assume the same for learner language.

Most importantly, researchers must be very careful in deciding whether the goodness of fit of their data set will be more in line with a binomial or a Poisson distribution. The key considerations, I have argued, are: (i) whether the focus of the analysis is an element likely to occur randomly or with a low probability – as in the case of particular adjectives, in which case the Poisson method would be appropriate; (ii) whether the focus' rate of occurrence is “greater than” unlikely, albeit at low levels. Binomial approaches assume some level of causality between factors and dependent variables, whereas Poisson regression makes much less of such an assumption. As we seek to generate viable models of learner development, we must identify very carefully what is likely to occur in a causal relationship and what has a higher probability of occurring spuriously.

In conclusion, the work by Cheng et al., studying the development of ser and estar within a corpus-based methodology, contitutes a good example of how corpus-based techniques can shed light on the developmental process, and complement other techniques employed to date. I contend that we must expand our model building enterprise to consider not just semantic and pragmatic variables that affect developmental patterns but also the role of discursive features. The analytical tools with which we study these processes must be considered with regard to the relationship of the phenomenon in question to other lexico-grammatical features and to the frequency of the phenomenon.