Since the early days of SLA research it has become increasingly clear that concepts such as the complexity of L2 forms and the difficulty of learning them are complex and difficult to pin down. In this short article I will try to present a taxonomy of sources of difficulty, stress how they interact, and illustrate some of these interactions and the methodological issues involved in studying them with examples from the empirical studies in this issue. Finally, I will make some suggestions for future research. It should be clear from the outset that my comments only concern complexity as a learning problem, not complexity as a desirable component of proficiency (see, e.g., various contributions in Housen, Kuiken, and Vedder, Reference Housen, Kuiken and Vedder2012). At the same time, I do not want to restrict difficulty to narrowly defined “structural complexity” (Pallotti, Reference Pallotti2015), that is, complexity completely inherent in the language regardless of the learning process, because even restricting the discussion to the learning and not the proficiency side of the spectrum, we can distinguish not only linguistic but also contextual and subjective difficulty.
Linguistic difficulty is due to the characteristics of the L2 target: the meanings to be expressed, the forms used for expressing them, and most crucially, the form-meaning mappings (cf. DeKeyser, Reference DeKeyser2005). Forms can be difficult because they lack phonological salience (e.g., nonsyllabic morphemes), because of morphophonological alternations, or because they are hard to pronounce. Meanings can be difficult because they are both novel (not represented in L1) and abstract, and hence hard to infer from the input. The semantics of articles and of grammatical aspect marking on the verb are notorious examples of this. Classifiers in languages like Chinese, Japanese, or Korean are somewhat less abstract but are still novel for many learners; conversely, person, number, and tense marking on the verb are far from obvious for speakers of these languages. Form-meaning mapping, however, is the essence of language learning, given that language is a set of tacit conventions about how to link forms and meanings; this mapping in itself can be hard because the relationship between form and meaning can be far from transparent, let alone salient. It is particularly difficult in inflectional languages when person, number, and tense together yield one morph so that there is no single morpheme corresponding to any of the semantic elements that are expressed by that morph. Homonymy of morphs as well as allomorphy help to obscure the form-meaning relationship (cf. Goldschneider and DeKeyser, Reference Goldschneider and DeKeyser2001). The English third-person –s combines just about all the elements of difficulty here: it is a little chunk of sound that only pops up when three elements of semantics come together, and at the same time it shows allomorphy (/s/, /z/, /əz/) and homonymy with genitive –s and plural –s. Add to that the lack of phonological salience (as its most frequent allomorphs are nonsyllabic), and it is not surprising that this seemingly “simple form” remains problematic for many long-term residents in the L2 English community.
As the previous paragraph already suggests, L1-L2 differences can mitigate or exacerbate linguistic difficulty in the sense that it is usually easier to drop a distinction that is present in the native language than the other way around. At the same time, however, when L1 and L2 are in a superset-subset relationship, dropping a possibility that exists in L1 but is absent in L2 can also be problematic, as various studies on the acquisition of adverb placement in English by speakers of French have shown (e.g., White, Reference White1991). Moreover, this subset-superset relationship is often characterized by extreme lack of salience: learners seem to take the view that “absence of evidence is not evidence of absence” and continue to use structures allowed by L1 (in this case Subject-Verb-Adverb-Object order) but not by L2. The reasoning does not work in reverse, however: it is not necessarily the case that positive evidence for a superset is salient. Where the subject in English and Spanish is concerned, Spanish offers a possibility that English does not, that is dropping the overt subject, but as this extra possibility happens to consist of the absence of something, it is also nonsalient, which in part explains why English learners of Spanish keep oversupplying pronominal subjects till fairly advanced levels (e.g., DeKeyser, Reference DeKeyser2010) and keep misunderstanding object pronouns in initial position as if they were subjects (e.g., VanPatten, Reference VanPatten, Gass and Mackey2012).
Interestingly, while lack of salience can compound subset-superset relationships (as is illustrated in Della Putta, this issue), sometimes a high degree of salience makes up for what would otherwise be difficult to acquire, for example in the case of preposition stranding instead of pied piping (Bardovi-Harlig, Reference Bardovi-Harlig1987) or in the case of VF-V1 structures (Tagarelli, Ruiz, Moreno Vega, and Rebuschat, this issue). In these last two cases the advantage of salience prevails over the disadvantage of syntactic complexity.
Footnote 1
The subset learning problem and the lack of salience in general call for explicit teaching in the form of negative evidence and enhanced input, respectively, which brings us to the second aspect of difficulty: contextual difficulty. The ease or difficulty of form-meaning mapping, everything else being the same, depends to a large extent on the context in which input is given: completely unmodified input, enhanced input, interactionally modified input, structured input, grammar explanation, and grammar explanation followed by systematic practice constitute a spectrum from the most to the least challenging for form-meaning mapping.
Clearly, picking up rather opaque form-meaning relationships from completely natural spoken input is a formidable task for a learner who is still in the process of acquiring the basic grammar of the language. Enhanced written input (through, e.g., color, font, or underlining) may help in some cases (Lee, Reference Lee2007; White, Reference White, Doughty and Williams1998), but overall the record is rather mixed (e.g., Della Putta, this issue; see esp. the meta-analysis by Lee and Huang, Reference Lee and Huang2008). This is probably because only the form can be enhanced, not the form-meaning mapping. Input enhancement is probably most useful, therefore, for mere form-form relationships such as morphophonological alternations like –s/-z/-iz for the plural of the noun or the third-person singular of the verb in English, or gender agreement of adjectives with nouns and articles in the Romance languages.
In spoken language, interactional modification is the equivalent of input enhancement. Long’s (Reference Long1983a, Reference Longb) seminal work on negotiation of meaning and modified input has made it clear that interactional modification rather than simplification can help learners to understand input and learn new form-meaning mappings at the same time; both negotiation of meaning and interactional modification focus the learner’s attention on the form while its meaning is clear from the context (and perhaps a contrasting form-meaning relationship is also present in the learner’s mind). Learning from modified input is most likely for the most concrete meanings, where the mapping is more salient, as between lexemes and their referents. Where the meaning is more abstract, as for a grammatical morpheme, negotiated interaction is probably more likely to be successful when the learner is already aware of the form-meaning mapping but has not fully grasped or proceduralized it yet.
For new elements of morphosyntax, however, even deftly modified input is unlikely to be enough, and some explicit (metalinguistic) teaching of rules, or at least systematic presentation of examples from which the learner can explicitly induce them, is most likely required. Structured input in the sense of processing instruction (e.g., VanPatten, Reference VanPatten, Gass and Mackey2012) has repeatedly been shown to work well for certain structures like for the understanding of the previously mentioned Object-Verb-Subject (OVS) structures in Spanish, where the learner has a strong tendency to interpret the initial pronoun as subject, because of the combined novelties of both the preverbal clitic and the lack of overt subject. Structured input consists of a set of sentences where the learner is forced to be conscious of (i.e., aware of and tuned to) the form-meaning mapping in order to comprehend the sentence correctly and choose the corresponding picture or translation out of a set of distractors that the learner would opt for if the form-meaning mapping were not correctly applied. Even this form of task-essential input practice, however, may not work for all structures and all learners. Prieto Botana (Reference Prieto Botana2013), for example, shows that while task-essentialness may be sufficient for OVS, explicit presentation before the input practice contributed substantially to the outcome for the ser/estar distinction (Spanish copulas), which has a more complex (multifaceted) semantic component (cf. also DeKeyser & Prieto Botana, Reference DeKeyser and Prieto Botana2015).
For many elements of grammar, then, no form of input, no matter how cleverly spoon-fed to the learner, will suffice (at least not in the amount of time available in instructed second language learning, and in many cases not even after years of exposure for untutored adult immigrants). Some form of explicit teaching and systematic communicative practice will be necessary in those cases; which form it should take largely depends on the characteristics of the learner.
This brings us to the third aspect of difficulty. Different elements of the language are more or less difficult, and different learning contexts can reduce the difficulty of learning them with different degrees of success, but the picture is not complete without considering subjective difficulty, that is, the degree of difficulty experienced, for a given structure in a given context, by different learners. (This subjective difficulty is actually the sense in which Robinson [Reference Robinson and Robinson2001] used the term difficulty as opposed to structural complexity.) Among the factors that determine difficulty for an individual learner, I already mentioned the fact that difficulty in learning a given structure depends in part on what corresponds to it in the native language. This is rather obvious to any language learner or teacher, but even that became a point of contention in the 1970s and 1980s. It is also true, by contrast, that even to this day, it is not always easy to predict exactly what shape L1 to L2 transfer will take, in particular at what level of abstraction, for example, how helpful it is for the L2 learner of a language with gender agreement to have gender agreement in L1 if that gender agreement takes a very different form (e.g., Dussias, Valdés Kroff, Guzzardo Tamargo, and Gerfen, Reference Dussias, Valdés Kroff, Guzzardo Tamargo and Gerfen2013; Foucart and Frenck-Mestre, Reference Foucart and Frenck-Mestre2011), and that things get even more complicated in the case of L3 or, more generally, Lx learning, where transfer may come from L1, L2, or, more generally, Lx-1.
In a more truly individual sense, one has to take into account proficiency level, aptitudes, and experience. Less is known about the exact role of these three variables; while the role of L1 has been a very prominent topic in the literature since the 1950s, with only a short and relative hiatus around 1980, and aptitude has seen a resurgence of interest in the last couple of decades, the role of proficiency level and of learning experience (in terms of number of languages learned, how, and for how long) are represented mainly in rather small bodies of research such as, respectively, the processability theory literature and the literature on L3 learning. Where individual structures are concerned, readiness to take the next step in a sequence (e.g., Pienemann, Reference Pienemann2005; see also the earlier literature on developmental sequences for negation, question formation, and relative clauses) is a more useful predictor of difficulty of acquisition than overall proficiency. By contrast, learning experience is a predictor of ease of learning in general, but more specifically of developing fluency. Many years of experience communicating in nonnative languages dramatically increase strategic competence (Canale and Swain, Reference Canale and Swain1980), so that highly experienced learners in that sense may be able to have fairly fluent conversations and write flawless short texts in an Lx in which they have very limited linguistic competence.
The role of aptitudes, however, is predictive of both broad gains (e.g., DeKeyser and Koeth, Reference DeKeyser, Koeth and Hinkel2011) and ease of learning for specific structures (e.g., Robinson Reference Robinson and Robinson2002). I am using aptitudes rather than aptitude here because it has become very clear that even when using the word aptitude in the narrow cognitive sense, the term is really shorthand for a constellation of different aptitudes. Even in the literature on older “aptitude tests” such as the MLAT and the PLAB, it was already clear that a number of different, separable aptitudes are measured by such tests, and that has become even clearer with newer tests such as the CANAL-F (Grigorenko, Sternberg, and Ehrman, Reference Grigorenko, Sternberg and Ehrman2000), the LLAMA (Meara, Reference Meara2005; see esp. Granena, Reference Granena, Granena and Long2013), and clearest of all in the Hi-LAB (Linck et al., Reference Linck, Hughes, Campbell, Silbert, Tare, Jackson and Doughty2013), which incorporates various tests that are widely used outside of SLA into the battery used to predict success at foreign language learning, including, first and foremost, various aspects of working memory.
While it may be rather trivial by now to assert that “foreign language learning aptitude” is indeed predictive of foreign language learning, it is still far from clear which specific aptitudes play an important role at which stage of learning for what structures under what conditions. Even the picture for working memory by itself is fairly complicated (see the reviews in Juffs and Harrington, Reference Juffs and Harrington2011; Williams, Reference Williams, Gass and Mackey2012; and especially the meta-analysis in Linck, Osthus, Koeth, and Bunting, Reference Linck, Osthus, Koeth and Bunting2014).
What we need, then, is research that does not ask about the beneficial effect of aptitudes or instructional contexts or the difficulty of particular structures, but that looks at the interaction of two or more of these variables. Many paradoxical findings in the SLA literature stem from the lack of interaction research, and this special issue is meant to address that problem head-on. Where research on difficulty is concerned, there has been a strong tendency to ignore the concepts of contextual and subjective difficulty, that is, lack of relativization of difficulty. At the same time, where the objective difficulty of form-meaning mapping is concerned, there is a hidden relativization in the literature: one structure is called difficult and the other easy, because one is more difficult than the other, but they could both be easier or harder than average (certainly for specific populations, but even generally speaking). It is no surprise, then, that we find divergent points of view in the literature, some hypothesizing that explicit teaching, for instance, is most useful for difficult structures (e.g., Hulstijn and de Graaff, Reference Hulstijn and de Graaff1994), others finding that it is most useful for easy structures (e.g., DeKeyser, Reference DeKeyser1995), and a meta-analysis concluding that difficulty does not make a difference in the sense that explicit teaching always helps (Spada and Tomita, Reference Spada and Tomita2010). I still believe that in the absence of further specification of contextual and individual variables, explicit teaching of grammar is most useful for structures of moderate difficulty, that is, not so easy that the form-meaning mappings are easily learned inductively from input alone, and not so difficult that even with explicit teaching most students cannot grasp them (DeKeyser, Reference DeKeyser, Doughty and Long2003). When we do take contextual and learner variables into account, however, “things are not so straightforward,” as Rodríguez Silva and Roehr-Brackin (this issue) put it, and we should be able to make more specific recommendations based on the interaction of two or three variables.
These authors look at the interaction of one linguistic variable (difficulty in the linguistic sense) and one cognitive variable (implicit vs. explicit learning). Two more articles in this issue also investigate linguistic and cognitive variables: Yalcin and Spada (passive vs. past progressive × grammatical sensitivity and memory) and Godfroid (different vowel changes in strong verbs × implicit/explicit knowledge). Two others look at the interaction of treatment with linguistic or cognitive variables: Della Putta (+/- enhancement × subset/superset) and Cerezo, Caras, and Leow (deductive instruction vs. guided induction × hypothesis-testing, rule formation, and activation of prior knowledge; the cognitive variables are measured post hoc in this case). Finally, only Tagarelli et al. look at the interaction of cognitive variables (working memory and procedural learning capacity) with a linguistic one (complexity) and a treatment variable (incidental vs. instructed). Each one of these studies finds interactions. In what follows I will only select a few of the findings that are particularly relevant against the background sketched in the preceding text, discuss their interpretations, and provide some pointers for future research.
Let us start with the only experiment on the interaction of three independent variables in this issue: incidental/instructed × (two) aptitudes × complexity in Tagarelli et al. The authors actually also made a distinction between implicit and explicit knowledge as outcome variables, so in that sense they have a four-way interaction. Not surprisingly, the resulting pattern is very complex, but it does suggest that the instructed learners relied more on explicit knowledge and the incidental learners more on implicit knowledge. This interpretation is consistent with the fact that instructed learners do better in those instances where they say they are relying on rules and the implicit learners do better when they are relying on intuition. As the analysis of different item types suggests, however, even incidental learners rely on working memory for the explicit learning of salient rules. Perhaps learners in the incidental and instructed conditions both tend to rely primarily on one type of aptitude for one type of learning but are capable of “switching” when the type of item makes that an advantage.
Compared to Tagarelli et al., even Robinson’s (Reference Robinson1997) design and findings may seem simple. The interacting variables in that study were implicit/incidental/rule-search/instructed × aptitude/memory × complexity. The learning of easy structures was predicted by grammatical sensitivity but not by memory and the learning of complex structures by memory but not by grammatical sensitivity, for instance, and this pattern was only obtained for one of the four instructional conditions (“rule search,” i.e., inductive explicit learning).
The findings of these studies, however complex and different, do not seem incompatible with Li’s hypothesis that “other things being equal, language analytic ability is implicated in implicit conditions in the learning of easy, transparent structures that are within one’s processing capacity, and in explicit conditions in the learning of hard, opaque structures where the internalization of available metalinguistic information sets heavy processing demands on internal cognitive resources” (Li, Reference Li2013, p. 648). “Implicit” here means implicit teaching, not implicit learning, which explains the effect of analytic aptitude. This hypothesis nicely summarizes a pattern that would be perfectly in sync with the common finding in aptitude-treatment interaction research that the more the treatment puts the information-processing burden on the learner, the more important the role of aptitudes becomes, but it also suggests there are limits: having to induce the rules is hard, and opaque structures are hard enough, but combining the two sources of difficulty may cause a floor effect and make aptitude useless, at least aptitude for explicit learning. By contrast, when the rules are fairly simple and they are taught explicitly, aptitude may not play much of a role either, because the explicit teaching has a leveling effect (Erlam, Reference Erlam2005). Where the rules and the treatment are both of moderate difficulty (learning to use English articles correctly through metalinguistic feedback), the effect of aptitude is strong (Sheen, Reference Sheen2007).
If there is anything the studies cited here can teach us unambiguously, it is that complexity is a complex matter, because of its multiple components and their interactions. At the same time, in spite of the obvious theoretical and practical importance of these issues, research along these lines has been very scant, even less common than aptitude-treatment interaction research in SLA, which of course studies the cognitive and treatment variables but leaves the complexity variable out (for a recent overview, see Vatz, Tare, Jackson, and Doughty, Reference Vatz, Tare, Jackson, Doughty, Granena and Long2013).
We clearly need more research that looks at three-way interactions if we are to have any hope of finding a replicable, stable pattern. However, we have to be careful not to complicate research designs too much, because the more variables are involved, the harder it is to achieve enough statistical power, and the less likely it is that all the interactions in a pattern of findings are replicated exactly, regardless of statistical issues. Therefore, my own preference would be not to throw the implicit/explicit knowledge distinction into the mix. Godfroid (this issue) and Tagarelli et al. (this issue) make a better case for implicit learning/knowledge than many, but I am still skeptical, for several reasons. In Godfroid’s case, the learners had years of experience with explicit learning, including with the target forms and similar ones. In the case of Tagarelli et al., the nature of the semi-artificial language made structures more salient and explicit learning more likely. In both studies the treatment was short, which also privileges explicit learning. Moreover, even reasonably pure measures of implicit knowledge are hard to come by, and neither reaction times in speeded grammaticality judgments nor elicited imitation are valid measures of implicit knowledge (Suzuki and DeKeyser, Reference Suzuki and DeKeyser2015). As implicit learning is very limited in a classroom context, its measurement is very difficult, and the interactions we are looking at are very complex already, we may do well to concentrate on explicit learning with these designs.
Not only do we need to look at more interactions; we also need to be more specific about the variable that is interacting. On the linguistic side, maybe salience is the issue, for instance, or maybe it is only one of its components, like phonological salience or the transparency of form-meaning mapping. On the cognitive side, when we think working memory is the issue, it may really be something more narrow, like inhibitory control or attention-shifting capacity. On the treatment side, is it the explicitness of presentation that is at issue, for instance, or more specifically the systematic form-focused practice that goes with it, or the way the two are linked? More often than not we tend to use all-purpose evergreens as independent variables instead of more narrow ones where there would be better theoretical justification for hypothesizing interactions. After all, the fact that an aptitude predicts is not particularly interesting theoretically, and the fact that a treatment works better is not likely to be useful if it cannot be generalized for lack of understanding of what exactly in the treatment made the difference and what aptitude it drew on. Both for theoretical and practical reasons, then, understanding interactions between linguistic, contextual, and cognitive (or more broadly subjective, individual) variables is of central importance if we even want to make SLA research scientifically respectable and practically useful at the same time (DeKeyser, Reference DeKeyser2012).
In the meantime, there are some findings that have practical implications already. The rather straightforward finding that lack of salience is often an obstacle to learning implies that an important aim of any kind of treatment should be to make nonsalient structures more salient. In that sense the finding that nonsalient morphemes can be particularly nonsalient in classroom discourse (Collins, Trofimovich, White, Cardoso, and Horst, Reference Collins, Trofimovich, White, Cardoso and Horst2009) is particularly worrisome. However, given the complexity of interactions we are faced with, and even assuming they replicate well in future research, it is unrealistic to expect teachers to take them all into account, either at the level of the curriculum or in minute-to-minute interaction. In this regard there is an important role for computer-assisted instruction: if we are confident that we understand the three-way interactions between linguistic, contextual, and individual variables, technology can provide curricular alternatives for individual students’ learning of specific structures. A more detailed student model that is dynamic in nature, that is, that follows the students’ progress in understanding, applying, and automatizing the target structure, and adapts in real time would take us even further (cf. e.g., Nakic, Granic, and Glavinic, Reference Nakic, Granic and Glavinic2015, for examples of intelligent tutoring using student models in various subject areas; and Koedinger, Booth, and Klahr, Reference Koedinger, Booth and Klahr2013 for a taxonomy of factors to consider in instructional design beyond the ones discussed here). Cerezo, Caras, and Leow (this issue) show that guided induction in computer-delivered instruction can be very useful in drawing attention to form and keeping the learner’s attention on it in subsequent practice for a particularly problematic construction. More work along these lines, on a larger scale, and with more input from theoretically motivated interaction research and from work in student modeling holds great promise for language teaching in the twenty-first century.
You have
Access
Since the early days of SLA research it has become increasingly clear that concepts such as the complexity of L2 forms and the difficulty of learning them are complex and difficult to pin down. In this short article I will try to present a taxonomy of sources of difficulty, stress how they interact, and illustrate some of these interactions and the methodological issues involved in studying them with examples from the empirical studies in this issue. Finally, I will make some suggestions for future research. It should be clear from the outset that my comments only concern complexity as a learning problem, not complexity as a desirable component of proficiency (see, e.g., various contributions in Housen, Kuiken, and Vedder, Reference Housen, Kuiken and Vedder2012). At the same time, I do not want to restrict difficulty to narrowly defined “structural complexity” (Pallotti, Reference Pallotti2015), that is, complexity completely inherent in the language regardless of the learning process, because even restricting the discussion to the learning and not the proficiency side of the spectrum, we can distinguish not only linguistic but also contextual and subjective difficulty.
Linguistic difficulty is due to the characteristics of the L2 target: the meanings to be expressed, the forms used for expressing them, and most crucially, the form-meaning mappings (cf. DeKeyser, Reference DeKeyser2005). Forms can be difficult because they lack phonological salience (e.g., nonsyllabic morphemes), because of morphophonological alternations, or because they are hard to pronounce. Meanings can be difficult because they are both novel (not represented in L1) and abstract, and hence hard to infer from the input. The semantics of articles and of grammatical aspect marking on the verb are notorious examples of this. Classifiers in languages like Chinese, Japanese, or Korean are somewhat less abstract but are still novel for many learners; conversely, person, number, and tense marking on the verb are far from obvious for speakers of these languages. Form-meaning mapping, however, is the essence of language learning, given that language is a set of tacit conventions about how to link forms and meanings; this mapping in itself can be hard because the relationship between form and meaning can be far from transparent, let alone salient. It is particularly difficult in inflectional languages when person, number, and tense together yield one morph so that there is no single morpheme corresponding to any of the semantic elements that are expressed by that morph. Homonymy of morphs as well as allomorphy help to obscure the form-meaning relationship (cf. Goldschneider and DeKeyser, Reference Goldschneider and DeKeyser2001). The English third-person –s combines just about all the elements of difficulty here: it is a little chunk of sound that only pops up when three elements of semantics come together, and at the same time it shows allomorphy (/s/, /z/, /əz/) and homonymy with genitive –s and plural –s. Add to that the lack of phonological salience (as its most frequent allomorphs are nonsyllabic), and it is not surprising that this seemingly “simple form” remains problematic for many long-term residents in the L2 English community.
As the previous paragraph already suggests, L1-L2 differences can mitigate or exacerbate linguistic difficulty in the sense that it is usually easier to drop a distinction that is present in the native language than the other way around. At the same time, however, when L1 and L2 are in a superset-subset relationship, dropping a possibility that exists in L1 but is absent in L2 can also be problematic, as various studies on the acquisition of adverb placement in English by speakers of French have shown (e.g., White, Reference White1991). Moreover, this subset-superset relationship is often characterized by extreme lack of salience: learners seem to take the view that “absence of evidence is not evidence of absence” and continue to use structures allowed by L1 (in this case Subject-Verb-Adverb-Object order) but not by L2. The reasoning does not work in reverse, however: it is not necessarily the case that positive evidence for a superset is salient. Where the subject in English and Spanish is concerned, Spanish offers a possibility that English does not, that is dropping the overt subject, but as this extra possibility happens to consist of the absence of something, it is also nonsalient, which in part explains why English learners of Spanish keep oversupplying pronominal subjects till fairly advanced levels (e.g., DeKeyser, Reference DeKeyser2010) and keep misunderstanding object pronouns in initial position as if they were subjects (e.g., VanPatten, Reference VanPatten, Gass and Mackey2012).
Interestingly, while lack of salience can compound subset-superset relationships (as is illustrated in Della Putta, this issue), sometimes a high degree of salience makes up for what would otherwise be difficult to acquire, for example in the case of preposition stranding instead of pied piping (Bardovi-Harlig, Reference Bardovi-Harlig1987) or in the case of VF-V1 structures (Tagarelli, Ruiz, Moreno Vega, and Rebuschat, this issue). In these last two cases the advantage of salience prevails over the disadvantage of syntactic complexity. Footnote 1
The subset learning problem and the lack of salience in general call for explicit teaching in the form of negative evidence and enhanced input, respectively, which brings us to the second aspect of difficulty: contextual difficulty. The ease or difficulty of form-meaning mapping, everything else being the same, depends to a large extent on the context in which input is given: completely unmodified input, enhanced input, interactionally modified input, structured input, grammar explanation, and grammar explanation followed by systematic practice constitute a spectrum from the most to the least challenging for form-meaning mapping.
Clearly, picking up rather opaque form-meaning relationships from completely natural spoken input is a formidable task for a learner who is still in the process of acquiring the basic grammar of the language. Enhanced written input (through, e.g., color, font, or underlining) may help in some cases (Lee, Reference Lee2007; White, Reference White, Doughty and Williams1998), but overall the record is rather mixed (e.g., Della Putta, this issue; see esp. the meta-analysis by Lee and Huang, Reference Lee and Huang2008). This is probably because only the form can be enhanced, not the form-meaning mapping. Input enhancement is probably most useful, therefore, for mere form-form relationships such as morphophonological alternations like –s/-z/-iz for the plural of the noun or the third-person singular of the verb in English, or gender agreement of adjectives with nouns and articles in the Romance languages.
In spoken language, interactional modification is the equivalent of input enhancement. Long’s (Reference Long1983a, Reference Longb) seminal work on negotiation of meaning and modified input has made it clear that interactional modification rather than simplification can help learners to understand input and learn new form-meaning mappings at the same time; both negotiation of meaning and interactional modification focus the learner’s attention on the form while its meaning is clear from the context (and perhaps a contrasting form-meaning relationship is also present in the learner’s mind). Learning from modified input is most likely for the most concrete meanings, where the mapping is more salient, as between lexemes and their referents. Where the meaning is more abstract, as for a grammatical morpheme, negotiated interaction is probably more likely to be successful when the learner is already aware of the form-meaning mapping but has not fully grasped or proceduralized it yet.
For new elements of morphosyntax, however, even deftly modified input is unlikely to be enough, and some explicit (metalinguistic) teaching of rules, or at least systematic presentation of examples from which the learner can explicitly induce them, is most likely required. Structured input in the sense of processing instruction (e.g., VanPatten, Reference VanPatten, Gass and Mackey2012) has repeatedly been shown to work well for certain structures like for the understanding of the previously mentioned Object-Verb-Subject (OVS) structures in Spanish, where the learner has a strong tendency to interpret the initial pronoun as subject, because of the combined novelties of both the preverbal clitic and the lack of overt subject. Structured input consists of a set of sentences where the learner is forced to be conscious of (i.e., aware of and tuned to) the form-meaning mapping in order to comprehend the sentence correctly and choose the corresponding picture or translation out of a set of distractors that the learner would opt for if the form-meaning mapping were not correctly applied. Even this form of task-essential input practice, however, may not work for all structures and all learners. Prieto Botana (Reference Prieto Botana2013), for example, shows that while task-essentialness may be sufficient for OVS, explicit presentation before the input practice contributed substantially to the outcome for the ser/estar distinction (Spanish copulas), which has a more complex (multifaceted) semantic component (cf. also DeKeyser & Prieto Botana, Reference DeKeyser and Prieto Botana2015).
For many elements of grammar, then, no form of input, no matter how cleverly spoon-fed to the learner, will suffice (at least not in the amount of time available in instructed second language learning, and in many cases not even after years of exposure for untutored adult immigrants). Some form of explicit teaching and systematic communicative practice will be necessary in those cases; which form it should take largely depends on the characteristics of the learner.
This brings us to the third aspect of difficulty. Different elements of the language are more or less difficult, and different learning contexts can reduce the difficulty of learning them with different degrees of success, but the picture is not complete without considering subjective difficulty, that is, the degree of difficulty experienced, for a given structure in a given context, by different learners. (This subjective difficulty is actually the sense in which Robinson [Reference Robinson and Robinson2001] used the term difficulty as opposed to structural complexity.) Among the factors that determine difficulty for an individual learner, I already mentioned the fact that difficulty in learning a given structure depends in part on what corresponds to it in the native language. This is rather obvious to any language learner or teacher, but even that became a point of contention in the 1970s and 1980s. It is also true, by contrast, that even to this day, it is not always easy to predict exactly what shape L1 to L2 transfer will take, in particular at what level of abstraction, for example, how helpful it is for the L2 learner of a language with gender agreement to have gender agreement in L1 if that gender agreement takes a very different form (e.g., Dussias, Valdés Kroff, Guzzardo Tamargo, and Gerfen, Reference Dussias, Valdés Kroff, Guzzardo Tamargo and Gerfen2013; Foucart and Frenck-Mestre, Reference Foucart and Frenck-Mestre2011), and that things get even more complicated in the case of L3 or, more generally, Lx learning, where transfer may come from L1, L2, or, more generally, Lx-1.
In a more truly individual sense, one has to take into account proficiency level, aptitudes, and experience. Less is known about the exact role of these three variables; while the role of L1 has been a very prominent topic in the literature since the 1950s, with only a short and relative hiatus around 1980, and aptitude has seen a resurgence of interest in the last couple of decades, the role of proficiency level and of learning experience (in terms of number of languages learned, how, and for how long) are represented mainly in rather small bodies of research such as, respectively, the processability theory literature and the literature on L3 learning. Where individual structures are concerned, readiness to take the next step in a sequence (e.g., Pienemann, Reference Pienemann2005; see also the earlier literature on developmental sequences for negation, question formation, and relative clauses) is a more useful predictor of difficulty of acquisition than overall proficiency. By contrast, learning experience is a predictor of ease of learning in general, but more specifically of developing fluency. Many years of experience communicating in nonnative languages dramatically increase strategic competence (Canale and Swain, Reference Canale and Swain1980), so that highly experienced learners in that sense may be able to have fairly fluent conversations and write flawless short texts in an Lx in which they have very limited linguistic competence.
The role of aptitudes, however, is predictive of both broad gains (e.g., DeKeyser and Koeth, Reference DeKeyser, Koeth and Hinkel2011) and ease of learning for specific structures (e.g., Robinson Reference Robinson and Robinson2002). I am using aptitudes rather than aptitude here because it has become very clear that even when using the word aptitude in the narrow cognitive sense, the term is really shorthand for a constellation of different aptitudes. Even in the literature on older “aptitude tests” such as the MLAT and the PLAB, it was already clear that a number of different, separable aptitudes are measured by such tests, and that has become even clearer with newer tests such as the CANAL-F (Grigorenko, Sternberg, and Ehrman, Reference Grigorenko, Sternberg and Ehrman2000), the LLAMA (Meara, Reference Meara2005; see esp. Granena, Reference Granena, Granena and Long2013), and clearest of all in the Hi-LAB (Linck et al., Reference Linck, Hughes, Campbell, Silbert, Tare, Jackson and Doughty2013), which incorporates various tests that are widely used outside of SLA into the battery used to predict success at foreign language learning, including, first and foremost, various aspects of working memory.
While it may be rather trivial by now to assert that “foreign language learning aptitude” is indeed predictive of foreign language learning, it is still far from clear which specific aptitudes play an important role at which stage of learning for what structures under what conditions. Even the picture for working memory by itself is fairly complicated (see the reviews in Juffs and Harrington, Reference Juffs and Harrington2011; Williams, Reference Williams, Gass and Mackey2012; and especially the meta-analysis in Linck, Osthus, Koeth, and Bunting, Reference Linck, Osthus, Koeth and Bunting2014).
What we need, then, is research that does not ask about the beneficial effect of aptitudes or instructional contexts or the difficulty of particular structures, but that looks at the interaction of two or more of these variables. Many paradoxical findings in the SLA literature stem from the lack of interaction research, and this special issue is meant to address that problem head-on. Where research on difficulty is concerned, there has been a strong tendency to ignore the concepts of contextual and subjective difficulty, that is, lack of relativization of difficulty. At the same time, where the objective difficulty of form-meaning mapping is concerned, there is a hidden relativization in the literature: one structure is called difficult and the other easy, because one is more difficult than the other, but they could both be easier or harder than average (certainly for specific populations, but even generally speaking). It is no surprise, then, that we find divergent points of view in the literature, some hypothesizing that explicit teaching, for instance, is most useful for difficult structures (e.g., Hulstijn and de Graaff, Reference Hulstijn and de Graaff1994), others finding that it is most useful for easy structures (e.g., DeKeyser, Reference DeKeyser1995), and a meta-analysis concluding that difficulty does not make a difference in the sense that explicit teaching always helps (Spada and Tomita, Reference Spada and Tomita2010). I still believe that in the absence of further specification of contextual and individual variables, explicit teaching of grammar is most useful for structures of moderate difficulty, that is, not so easy that the form-meaning mappings are easily learned inductively from input alone, and not so difficult that even with explicit teaching most students cannot grasp them (DeKeyser, Reference DeKeyser, Doughty and Long2003). When we do take contextual and learner variables into account, however, “things are not so straightforward,” as Rodríguez Silva and Roehr-Brackin (this issue) put it, and we should be able to make more specific recommendations based on the interaction of two or three variables.
These authors look at the interaction of one linguistic variable (difficulty in the linguistic sense) and one cognitive variable (implicit vs. explicit learning). Two more articles in this issue also investigate linguistic and cognitive variables: Yalcin and Spada (passive vs. past progressive × grammatical sensitivity and memory) and Godfroid (different vowel changes in strong verbs × implicit/explicit knowledge). Two others look at the interaction of treatment with linguistic or cognitive variables: Della Putta (+/- enhancement × subset/superset) and Cerezo, Caras, and Leow (deductive instruction vs. guided induction × hypothesis-testing, rule formation, and activation of prior knowledge; the cognitive variables are measured post hoc in this case). Finally, only Tagarelli et al. look at the interaction of cognitive variables (working memory and procedural learning capacity) with a linguistic one (complexity) and a treatment variable (incidental vs. instructed). Each one of these studies finds interactions. In what follows I will only select a few of the findings that are particularly relevant against the background sketched in the preceding text, discuss their interpretations, and provide some pointers for future research.
Let us start with the only experiment on the interaction of three independent variables in this issue: incidental/instructed × (two) aptitudes × complexity in Tagarelli et al. The authors actually also made a distinction between implicit and explicit knowledge as outcome variables, so in that sense they have a four-way interaction. Not surprisingly, the resulting pattern is very complex, but it does suggest that the instructed learners relied more on explicit knowledge and the incidental learners more on implicit knowledge. This interpretation is consistent with the fact that instructed learners do better in those instances where they say they are relying on rules and the implicit learners do better when they are relying on intuition. As the analysis of different item types suggests, however, even incidental learners rely on working memory for the explicit learning of salient rules. Perhaps learners in the incidental and instructed conditions both tend to rely primarily on one type of aptitude for one type of learning but are capable of “switching” when the type of item makes that an advantage.
Compared to Tagarelli et al., even Robinson’s (Reference Robinson1997) design and findings may seem simple. The interacting variables in that study were implicit/incidental/rule-search/instructed × aptitude/memory × complexity. The learning of easy structures was predicted by grammatical sensitivity but not by memory and the learning of complex structures by memory but not by grammatical sensitivity, for instance, and this pattern was only obtained for one of the four instructional conditions (“rule search,” i.e., inductive explicit learning).
The findings of these studies, however complex and different, do not seem incompatible with Li’s hypothesis that “other things being equal, language analytic ability is implicated in implicit conditions in the learning of easy, transparent structures that are within one’s processing capacity, and in explicit conditions in the learning of hard, opaque structures where the internalization of available metalinguistic information sets heavy processing demands on internal cognitive resources” (Li, Reference Li2013, p. 648). “Implicit” here means implicit teaching, not implicit learning, which explains the effect of analytic aptitude. This hypothesis nicely summarizes a pattern that would be perfectly in sync with the common finding in aptitude-treatment interaction research that the more the treatment puts the information-processing burden on the learner, the more important the role of aptitudes becomes, but it also suggests there are limits: having to induce the rules is hard, and opaque structures are hard enough, but combining the two sources of difficulty may cause a floor effect and make aptitude useless, at least aptitude for explicit learning. By contrast, when the rules are fairly simple and they are taught explicitly, aptitude may not play much of a role either, because the explicit teaching has a leveling effect (Erlam, Reference Erlam2005). Where the rules and the treatment are both of moderate difficulty (learning to use English articles correctly through metalinguistic feedback), the effect of aptitude is strong (Sheen, Reference Sheen2007).
If there is anything the studies cited here can teach us unambiguously, it is that complexity is a complex matter, because of its multiple components and their interactions. At the same time, in spite of the obvious theoretical and practical importance of these issues, research along these lines has been very scant, even less common than aptitude-treatment interaction research in SLA, which of course studies the cognitive and treatment variables but leaves the complexity variable out (for a recent overview, see Vatz, Tare, Jackson, and Doughty, Reference Vatz, Tare, Jackson, Doughty, Granena and Long2013).
We clearly need more research that looks at three-way interactions if we are to have any hope of finding a replicable, stable pattern. However, we have to be careful not to complicate research designs too much, because the more variables are involved, the harder it is to achieve enough statistical power, and the less likely it is that all the interactions in a pattern of findings are replicated exactly, regardless of statistical issues. Therefore, my own preference would be not to throw the implicit/explicit knowledge distinction into the mix. Godfroid (this issue) and Tagarelli et al. (this issue) make a better case for implicit learning/knowledge than many, but I am still skeptical, for several reasons. In Godfroid’s case, the learners had years of experience with explicit learning, including with the target forms and similar ones. In the case of Tagarelli et al., the nature of the semi-artificial language made structures more salient and explicit learning more likely. In both studies the treatment was short, which also privileges explicit learning. Moreover, even reasonably pure measures of implicit knowledge are hard to come by, and neither reaction times in speeded grammaticality judgments nor elicited imitation are valid measures of implicit knowledge (Suzuki and DeKeyser, Reference Suzuki and DeKeyser2015). As implicit learning is very limited in a classroom context, its measurement is very difficult, and the interactions we are looking at are very complex already, we may do well to concentrate on explicit learning with these designs.
Not only do we need to look at more interactions; we also need to be more specific about the variable that is interacting. On the linguistic side, maybe salience is the issue, for instance, or maybe it is only one of its components, like phonological salience or the transparency of form-meaning mapping. On the cognitive side, when we think working memory is the issue, it may really be something more narrow, like inhibitory control or attention-shifting capacity. On the treatment side, is it the explicitness of presentation that is at issue, for instance, or more specifically the systematic form-focused practice that goes with it, or the way the two are linked? More often than not we tend to use all-purpose evergreens as independent variables instead of more narrow ones where there would be better theoretical justification for hypothesizing interactions. After all, the fact that an aptitude predicts is not particularly interesting theoretically, and the fact that a treatment works better is not likely to be useful if it cannot be generalized for lack of understanding of what exactly in the treatment made the difference and what aptitude it drew on. Both for theoretical and practical reasons, then, understanding interactions between linguistic, contextual, and cognitive (or more broadly subjective, individual) variables is of central importance if we even want to make SLA research scientifically respectable and practically useful at the same time (DeKeyser, Reference DeKeyser2012).
In the meantime, there are some findings that have practical implications already. The rather straightforward finding that lack of salience is often an obstacle to learning implies that an important aim of any kind of treatment should be to make nonsalient structures more salient. In that sense the finding that nonsalient morphemes can be particularly nonsalient in classroom discourse (Collins, Trofimovich, White, Cardoso, and Horst, Reference Collins, Trofimovich, White, Cardoso and Horst2009) is particularly worrisome. However, given the complexity of interactions we are faced with, and even assuming they replicate well in future research, it is unrealistic to expect teachers to take them all into account, either at the level of the curriculum or in minute-to-minute interaction. In this regard there is an important role for computer-assisted instruction: if we are confident that we understand the three-way interactions between linguistic, contextual, and individual variables, technology can provide curricular alternatives for individual students’ learning of specific structures. A more detailed student model that is dynamic in nature, that is, that follows the students’ progress in understanding, applying, and automatizing the target structure, and adapts in real time would take us even further (cf. e.g., Nakic, Granic, and Glavinic, Reference Nakic, Granic and Glavinic2015, for examples of intelligent tutoring using student models in various subject areas; and Koedinger, Booth, and Klahr, Reference Koedinger, Booth and Klahr2013 for a taxonomy of factors to consider in instructional design beyond the ones discussed here). Cerezo, Caras, and Leow (this issue) show that guided induction in computer-delivered instruction can be very useful in drawing attention to form and keeping the learner’s attention on it in subsequent practice for a particularly problematic construction. More work along these lines, on a larger scale, and with more input from theoretically motivated interaction research and from work in student modeling holds great promise for language teaching in the twenty-first century.