Two decades have passed since the set of proposals commonly spoken of as the fundamental difference hypothesis (FDH; Bley-Vroman, Reference Bley-Vroman, Rutherford and Sharwood Smith1988, Reference Bley-Vroman, Gass and Schachter1989, Reference Bley-Vroman1990) were made.Footnote 1 Those proposals address two theoretical areas of SLA:
(1) a. the explanatory burden of SLA
b. the nature of the SLA system and of second language (L2) knowledge.
The term FDH has a broad and a narrow meaning. Narrowly, and originally, it refers just to a set of related alternative proposals for (1b). However, because it is proposed that SLA is different in fundamental ways, both in the explanandum (1a) and in the explanans (1b), the term can appropriately be used broadly to comprise both.
Much empirical research over the past two decades has taken place with either implicit or explicit reference to the FDH. Because both the explanandum and the explanans are formulated within the generative theory of Universal Grammar (UG), the abundant research that addresses the place of UG in SLA is broadly relevant. The present article addresses these same issues again, from a point of view two decades later. The focus of this article, as that of the original work, is chiefly conceptual and programmatic. It does not attempt a review of the specific research findings, except as they impinge directly on the conceptual issues.Footnote 2
First, the logical problem (1a), focusing on two properties of foreign language learning—namely, (un)reliability and (non)convergence—will be considered. The original arguments still hold in this case. Second, the problem of explanation (1b) is tackled; developments in learning theory, linguistics, psychology, mathematics, and philosophy require that the FDH be reconceived in radical ways.
THE LOGICAL PROBLEM OF FOREIGN LANGUAGE LEARNING
The explanatory burden of the theory of native language development is to account for how acquisition of the language can succeed, given the data to which the learning device is exposed. It is against this background that all serious native language acquisition theories are constructed, independent of whether such theories are based in mainstream generative grammar or some other framework. Debates about the adequacy of theories are always in terms of whether they will work reliably and consistently.
Bley-Vroman (Reference Bley-Vroman1990) gave ten characteristics with respect to which native language development contrasts with foreign language learning. Two of these are crucial for the logical structure of the theory. Using the terminology of Pullum and Scholz (Reference Pullum and Scholz2002), native language development is characterized by reliability and convergence. These are informally defined by Pullum and Scholz as in (2):Footnote 3
(2) Reliability: Children always succeed at language learning.
Convergence: Children end up with systems that are so similar to those of others in the same speech community. (p. 10)
In the terminology of the original “Logical Problem of Foreign Language Learning,” these correspond approximately to success-failure and uniformity-variability.Footnote 4
Reliability
There are, of course, various ways of modeling the concept of reliability in formal learning theory. One can require that the grammar learned is (eventually, in the limit) identical to the grammar that generated the input. One can require that the grammar learned (eventually) generate the same set of sentences that the target grammar generates. As in more recent approaches to mathematical learning theory, one can require that the grammar learned be probably approximately correct.Footnote 5 However, regardless of how implemented and regardless of the particular operationalization of success, native language learning theories aim to learn the target; they strive to be reliable.
Reliability includes that a theory be robust across possible input sequences, even in the presence of a certain amount of noise; for example, one would not consider a learning theory reliable that was dependent on encountering the right exemplars in the right order, or that would exhibit great fragility in the face of an occasional misleading or incorrect exemplar, or that required that unusual examples occur frequently.Footnote 6
Some kind of reliability (even as an idealization) is uniformly adopted, often implicitly, as a standard by those working in (native) language acquisition theory, natural language processing, and mathematical learning theory. The very term learnability reflects this standard: We want to show that a language is in fact learnable, given a theory of learning—that is, that the theory works reliably. No theorist of native language development has ever advocated a theory on the basis of its unreliability, or because it is less likely to learn the target than a competing theory, or because it can be seduced into error, or because its success is dependent on careful structuring of the input.
Ironically, what makes a theory a good theory of native language learning—that it works reliably—makes it a bad theory of foreign language learning. This is because of the first basic property of foreign language learning, the unreliability property defined in (3):
(3) Foreign language learning is not reliable.
Convergence
For our purposes, the second crucial property of native language acquisition is that it is convergent, as defined in (2); see Pullum and Scholz (Reference Pullum and Scholz2002). In Chomsky’s (Reference Chomsky1965) seminal formulation, there is a “striking uniformity of the resulting grammars” (p. 11).
Convergence is a relationship among the acquired grammars of learners exposed to the same target language. Reliability, in contrast, is a relationship between the target grammar and the acquired grammar. Reliability and convergence are obviously related. Reliability leads to convergence, but the converse is not true. An unreliable acquisition system can nonetheless lead to convergent grammars; for example, a foreign language acquisition theory might hypothetically (and counterfactually) predict that learners of English invariably would arrive at a grammar in which relative clauses had resumptive pronouns, or in which the encoded specificity rather than definiteness, or in which subject-auxiliary inversion occurred only with pronoun subjects. In these cases, the target is not being reliably learned, but there is a grammar—a nontarget grammar to be sure—on which the learning system converges. Because theories of native language development assume reliability as a standard, convergence follows a fortiori.
The fact, of course, is that in foreign language learning, learners exposed to the same target language may arrive at very different interlanguage grammars. The theory must permit everything from the so-called near-native cases like Julie (Ioup, Boustagui, Tigi, & Moselle, Reference Ioup, Boustagui, Tigi and Moselle1994) to cases in which the acquired grammar is dramatically different from that of the input, like Schmidt’s (Reference Schmidt, Wolfson and Judd1983) Wes or Schumann’s (Reference Schumann1978) Alberto. This leads to the second essential property of foreign language learning, the nonconvergence property defined in (4):
(4) Foreign language learning is not convergent.
THE (NARROW) LINGUISTIC THEORY OF FOREIGN LANGUAGE LEARNING
Addressing the logical problem of foreign language learning requires, in short, postulating an acquisition system that does not work reliably and does not converge. It must permit, as it were, the full range from Wes to Julie. At this level, the theory is not predictive or deterministic; that is, it cannot say for sure what grammar will be arrived at given the input. Rather, the theory is permissive. Let us use the term the (narrow) linguistic theory of foreign language learning to refer to a theory that relates the target language input to a set (not necessarily finite) of interlanguage grammars. This set will, no doubt, have some sort of internal structure, with some grammars having greater likelihood than others, and this likelihood structure may be related to the extent of exposure (modeled perhaps as the size of the input set). Finally, the structure of the native language grammar will influence the range of possibilities, especially their likelihood and the relationship of likelihood to amount of exposure. The (narrow) linguistic theory of foreign language learning is only part of the full picture of foreign language learning. Precisely because it is neither reliable nor convergent and hence cannot be fully predictive of outcome, it leaves a natural place for factors that go beyond the narrowly linguistic. In this way, it has the potential to fit in well with the rich research programs underway in other areas of L2 studies.
What would such a linguistic theory of foreign language learning look like? We are now moving from the explanandum (1a) to the explanans (1b). It would not look much like government binding/principles and parameters (GB/PP) theory.Footnote 7 The consequence of unreliability and nonconvergence for foreign language learning models is both simple and severe. Because virtually all work in generative native language acquisition theory takes reliability and convergence as desiderata for the theory, any foreign language learning theory that takes native language acquisition theory as a basis is at a serious disadvantage.Footnote 8 Researchers who take a broadly generative approach to language are in an especially difficult position. The thrust of generative acquisition theory, especially during the 1990s, and even continuing today, has been to develop a model of language that would specifically permit acquisition to take place in a reliable and convergent manner. UG as embodied in GB/PP models in particular, which form the background against which the FDH was originally formulated, incorporate extensive systems of built-in knowledge, thus minimizing what would need to be acquired and radically constraining the range of hypotheses that the learning device would need to consider. Parameters (i.e., switches) need to be set, admittedly, but the number of these is generally supposed to be manageable, and—most important—the data needed to set them are widely supposed to be readily available. Such a system is highly reliable and different learners exposed to the same target will converge on systems of striking uniformity, as native language acquisition theory appropriately requires.
THE FUNDAMENTAL DIFFERENCE HYPOTHESIS
What would an adequate theory look like then? The early FDH suggested, following ideas current in generative grammar at the time, that there are two conceptual components to an acquisition theory: (a) a definition of possible language (i.e., a hypothesis space) and (b) a learning procedure. GB/PP theory provides these components for native language development.
Bley-Vroman (Reference Bley-Vroman1990) proposed that the native language, rather than UG itself, shapes the initial hypothesis space and is also the chief source of initial expectations about the likely character of the target language: The learning system will expect something of the same general character as the native language. Indeed, the initial hypothesis may be that the target resembles the native language quite closely. Parameter setting as the learning process is rejected by the FDH. Instead, learning processes are utilized separately from their integration into a specialized language faculty. Elimination of UG and parameter setting should dispose of reliability and convergence—a welcome result. Additionally, transfer is predicted to occur, but in a much looser and more variable way than the one predicted by theories that propose that the initial state of the learner is exactly the native language grammar with parameters waiting to be reset, as in the full transfer/full access hypothesis (Schwartz & Sprouse, Reference Schwartz, Sprouse, Hoekstra and Schwartz1994, Reference Schwartz and Sprouse1996). Acquisition is expected to be piecemeal and dependent on what can be noticed in the input. This permits the FDH to articulate with the noticing hypothesis (Schmidt, Reference Schmidt1990), as part of a much more encompassing framework, with indirect ties (chiefly through noticing) to cognitive science more generally and even to language teaching.
One of the most direct statements of the FDH is Chomsky’s (Reference Chomksy1997) observation about the nature of foreign language learning:
Like other kinds of growth, language acquisition happens easily at a certain age, but not later. There comes a time when the system doesn’t work anymore. There are individual differences […] but for most people, after adolescence, it becomes very hard. The system is just not working for some reason, so, you have to teach the language as something strange. (p. 128)
Chomsky’s “just not working” and “something strange” (p. 128) in fact captures quite directly the two central tenets of the FDH as originally formulated. “Just not working anymore” (p. 128) corresponds to the FDH’s proposal that the domain-specific cognitive system of UG no longer guides acquisition. Treating “language as something strange” (p. 128) is one way of formulating the FDH’s suggestion that language is being handled not in this domain-specific faculty of mind, to which language is not strange, but rather in a domain-general central system, to which language is strange.
THE SHIFTING CONTEXT OF THE FUNDAMENTAL DIFFERENCE HYPOTHESIS
Whereas the essential elements of the logical problem (i.e., the unreliability property and the nonconvergence property) are unaffected by developments in linguistics and related fields over the past years, the same is not true of certain other specific proposals of the original FDH.
As a preliminary matter, however, one more property of foreign languages must be introduced explicitly. This property was implicit in earlier formulations of the FDH; it is the language property defined in (5).
(5) Foreign languages are languages.
Language is used here in a very broad sense. Foreign languages, like native languages, map between human conceptual-intentional structures (i.e., meaning) and a physical expression realized using the speech apparatus. Elements can be combined with other elements to make larger units, and this can go on recursively. Like native languages, foreign languages are processed in the human brain in real time. They are used to communicate, for private thinking, and for all functions of native languages. The FDH has never claimed that foreign languages are not languages in this sense.
Under the FDH, to say that foreign languages are languages is not to say that foreign languages are governed (totally or partially) by a GB/PP-type UG; for example, one could find evidence that subjacency (i.e., a principle that disallows movement across more than one bounding node) does not reliably constrain movement in a particular foreign language grammar but that grammar still would map between meaning and sound. In fact, all of the highly productive UG-in-SLA work over the past decades has operated within this assumption. This includes research that concludes that particular components of UG do not work in foreign languages in the same way or that the interface with forms is affected by maturation (Hawkins & Chan, Reference Hawkins and Chan1997; Lardiere, Reference Lardiere1998a, Reference Lardiere1998b; Prévost & White, Reference Prévost and White2000).
The important consequence of the language property is that features of a foreign language that relate to its being a language need not (necessarily) invoke UG as an explanation. To take a trivial example, you do not need UG to explain the fact that the tongue may be used to make foreign language sounds, or, you do not need UG to explain the fact that foreign languages may distinguish present from past time: The human conceptual system does make such distinctions, and one end of the sound-meaning map is that conceptual system.
The original context of work on the place of UG in SLA, and the context in which most research has taken place over the past two decades, was GB/PP theory (Chomsky, Reference Chomsky1981a, Reference Chomsky, Hornstein and Lightfoot1981b, Reference Chomsky1982, Reference Chomsky1986a, Reference Chomsky1986b). Along with mainstream GB/PP, the FDH assumed innate structural principles of language and processes of language acquisition that are elaborate and sui generis; these were clearly distinct from non-domain-specific principles of structural organization and of learning (general problem solving in the terminology current at the time). Four developments in the science of language require a reconceptualization of UG-based L2 research.
First, an elaborate and highly structured UG has given way to a minimal UG. The thrust of research from the extended standard theory through GB/PP was that UG was rich and highly articulated. There were subtheories with their own special properties (bounding, binding, control, case theory, theta theory, etc.); UG provided specific parametric options and ways to set abstract parameters based on concrete experience. Jackendoff and Pinker (2005) were not exaggerating much when they characterized the UG of GB/PP as “baroque” (p. 219). I will use the term rich UG for this trend. The very richness and abstractness of the necessary theoretical constructs and the difficulty of seeing how they could be learned from input gave strong motivation for positing an innate UG. Furthermore, the principles and processes of rich UG seemed impossible to derive in any obvious way from principles and processes outside the language faculty. They were “special principles specific to this domain” (Chomsky, Reference Chomsky1980, p. 44). For SLA research, the baroque richness, abstractness, and especially the domain-specificity of rich UG yielded testable predictions about the role of UG in SLA. The constructs of rich UG gave clear empirical content to the FDH, and because of the highly articulated structure of the overall theory, with different subcomponents, principles, and parameters, these theoretical constructs also made it possible to formulate hypotheses wherein some particular part(s) of a rich UG would fail to function normally in foreign language learning (Beck, Reference Beck1998; Clahsen & Hong, Reference Clahsen and Hong1995; Eubank, Reference Eubank1993/1994, Reference Eubank, Hoekstra and Schwartz1994, Reference Eubank1996; Hawkins & Chan, Reference Hawkins and Chan1997; Tsimpli & Roussou, Reference Tsimpli and Roussou1991). An elaborate and articulated innate rich UG also underlies a full transfer/full access view (Schwartz & Sprouse, Reference Schwartz, Sprouse, Hoekstra and Schwartz1994, Reference Schwartz and Sprouse1996).
Over the past ten years, and especially in the last few years, the rich UG concept of the language faculty has been replaced with a vision in which the specifically linguistic innate endowment (narrowly conceived) is very sparse. This is particularly clear in the Minimalist Program (Chomsky, Reference Chomsky1995, Reference Chomsky2000, Reference Chomsky and Belletti2004, Reference Chomsky2005). Jackendoff (Reference Jackendoff2005) succinctly summarized the Minimalist goal: “to eliminate the richness of Universal Grammar as conceived of in Principles and Parameters theory” (p. 191). If the Minimalist Program turns out to be headed in the right direction, there is no rich UG, and a theory of UG can be developed in which all one needs to say is that language maps between sound and meaning, that elements from a lexicon can be combined, that units can combine recursively to form larger units, and that there is some sort of displacement property that accounts for movement. All of what was previously accounted for by the elaborate domain-specific (sub)systems of rich UG can be accounted for by the interaction of this recursive system with meaning, on the one hand, and linearized physical expression, on the other.
The consequences of Minimalism for the FDH—indeed for any research program that takes UG access in foreign language learning to be an empirical issue—are serious. What would it mean to say that UG is not available in foreign language learning? The essence of UG under Minimalism is really just the language property itself. However, there is no version of the FDH, or indeed of any major foreign language learning theory, that would deny the language property. It does not even need to be stipulated (as it does in linguistic theory). In the FDH, it follows directly from the hypothesis that the native language provides information about the notion of possible language, “by making the very conservative assumption that the foreign language is not an utterly different sort of thing from the native language” (Bley-Vroman, Reference Bley-Vroman, Gass and Schachter1989, pp. 51–52). As Dekydtspotter and Hathorn (Reference Dekydtspotter and Hathorn2005), citing Hale (Reference Hale1996), noted, “given the level of generality of principles of Universal Grammar in Minimalist understanding, all grammatical principles are instantiated in all languages. The distinction [between the first language (L1) and UG] becomes vitiated” (p. 317).
The empirical content of full transfer/full access virtually disappears under strong Minimalism, because full access reduces to taking the target language to be a human language (rather than, say, a system that does not involve human cognition and recursion and displacement, but one that perhaps functions by mapping, animal-like, directly from a finite number of vocalizations to entities in the world or one that consists of things like analog danger warnings and mating calls). So far, the discussion has concentrated on Minimalism. However, similar considerations arise in other trends in contemporary linguistics. The family of approaches that one might call constructionist is also minimal in its own way: Simpler syntax is probably the best worked-out representative (Culicover & Jackendoff, Reference Culicover and Jackendoff2005; Jackendoff, Reference Jackendoff2002; Pinker & Jackendoff, Reference Pinker and Jackendoff2005). Jackendoff (Reference Jackendoff2005) indeed spoke of alternative minimalist visions. Construction grammar and item-based learning (Fillmore & Kay, Reference Fillmore and Kay1999; Fillmore, Kay, & O’Connor, Reference Fillmore, Kay and O’Connor1988; Goldberg, Reference Goldberg1995, Reference Goldberg2006; Tomasello, Reference Tomasello1998, Reference Tomasello2003) have much in common with simpler syntax.Footnote 9 In general, such approaches require syntactic structures that are much less abstract, hence more directly learnable on the basis of experience. To consider the implications of this family of approaches would take us too far afield, and our focus here will remain on mainstream generative grammar as represented by the Minimalist Program. Still, it is worth emphasizing that an elaborate system of sui generis innate linguistic knowledge—rich UG—on which much UG-in-SLA research has been built, is coming under question from both ends of the theoretical spectrum. As rich UG falls away, the FDH, and indeed all UG-based L2 research, must rethink its basic assumptions.Footnote 10
Second, the typological theory of principles and parameters, at least as originally imagined in linguistic theory and adopted in L2 research, has shown itself to be untenable. The original conception assumed a small number of parameters, with languagewide settings, that could unify apparently disparate superficial phenomena in a language. A single abstract (meta)parameter setting could derive a cluster of apparently disparate phenomena. Furthermore, the parameter values could be acquired by exposure to readily available input—triggers. The possible values, triggers, and deductive consequences are a part of rich UG. A full discussion of the demise of rich-UG parameter theory is beyond the scope of this article. See Newmeyer (Reference Newmeyer2004, Reference Newmeyer2005) and the references cited there for an extensive discussion of the reasons why this approach to typology cannot work and has been generally abandoned: “Metaparameters do not exist” (Pica, Reference Pica2001, cited in Newmeyer, Reference Newmeyer2004, p. 209). Additionally, in 20 years of SLA research, not a single study has convincingly demonstrated the sort of triggering and clustering that might have been expected. Quite the opposite: Studies have uniformly showed that aspects of a parameter (i.e., the individual phenomena) are learned separately; there is little evidence of strong deductivity. The work on the verb-movement parameter provides a particularly clear example (Ayoun, Reference Ayoun2002; White, Reference White1990/1991, Reference White1991, Reference White1992). In the Minimalist Program, it may turn out that abstract switches on lexical items may survive as a descendant of parameter theory, but their status seems quite different. Certainly, in keeping with Minimalist philosophy, these abstract switches must derive from the very nature of language as giving expression to the conceptual-intentional system. Definitely, they will not have the same consequences for L2 research as rich-UG metaparameters. This is a second reason to rethink the foundations of UG-based SLA research.
Third, there is now increasing evidence that many facts of language originally thought to be formally grammatical and attributed to the workings of rich UG may actually be the consequences of extragrammatical factors. The principle of subjacency, for example, has been widely investigated as a touchstone for the availability of UG in foreign language learning. However, there is now very good reason to believe that subjacency should be thought of as having its roots in some combination of processing and semantics-discourse structure (Hawkins, Reference Hawkins1999, Reference Hawkins2004; Kehler, Reference Kehler2002; Kluender, Reference Kluender, Culicover and McNally1998; Kluender & Kutas, Reference Kluender and Kutas1993; Levine & Sag, Reference Levine and Sag2003). Hawkins’s work is particularly persuasive in showing how a wide range of grammatical facts can be understood in this way.Footnote 11 It is, of course, possible that the grammar of a particular language may reflect, through grammaticization, such functional motivations. However, this does not mean that rich UG is required; nor can one definitely conclude, on the basis of foreign language learners’ sensitivity to such functional factors, that they have access to rich UG.
Fourth and finally, it is no longer possible to make a clear distinction between processes (especially data-processing acquisition processes) used by the language faculty and those used more generally. The FDH depended on this distinction. Current linguistic thinking explores the possibility that many of the processes of the language faculty (broadly conceived) are not exclusive to language. This position is most strikingly evident in Chomsky’s current work: “The modular view of learning of course does not entail that the component elements of the module are unique to it” (2005, p. 5), and “we need no longer assume that the means of generating structured expressions are highly articulated and specific to language. We can seriously entertain the possibility that they might be reducible to language-independent principles” (p. 9). Note the striking contrast to “special principles specific to this domain” (Chomsky, Reference Chomsky1980, p. 44) posited by rich UG. Chomsky cited as a concrete example the work of Gambell and Yang (Reference Gambell and Yang2003, published in Gambell & Yang, Reference Gambell, Yang, Moulton and Wolf2004), who showed that general principles of data analysis will give reasonable results in identifying word boundaries. One can clearly go farther than this. Work that shows the potential of general data analysis processes is not restricted to that couched in generative terminology. Indeed, such research includes—in fact is dominated by—that of scholars who are decidedly skeptical of the generative enterprise. There are now significant advances in machine learning and mathematical linguistics that point to the potential of general data-processing methods for language acquisition (see Lappin & Shieber, Reference Lappin and Shieber2007, for a review).Footnote 12 To take but one example, Clark and Eyraud (Reference Clark and Eyraud2006, Reference Clark and Eyraud2007) showed how structure-dependent auxiliary fronting in yes/no questions can be learned by distributional learning.
Work along these lines makes simplifying assumptions; for example, Gambell and Yang (Reference Gambell and Yang2003, as cited by Chomsky, Reference Chomsky2005) adopted the assumption that words are associated with a single stress; this in effect permits the system to say “there’s a word here somewhere; where are the edges?.” Clark and Eyraud (Reference Clark and Eyraud2006, Reference Clark and Eyraud2007) assumed that the system to be learned has the mathematical property of substitutability, which is essential to show the learnability of structure dependence. At this stage of the research, such simplifying—and almost certainly false—assumptions should not be disturbing. What is important is the nature of the assumptions themselves: They are linguistic. There is an interaction between features that relate to the nature of human language (e.g., stress-word associations or substitutability) and principles of data analysis surely also used in other domains. In other words, language-specificity of processing resides in the nature of the data rather than in the inherent language-specificity of the procedure.
If the data-processing acquisition processes of the language module are not unique to language, L2 research cannot use the distinction between domain-specific and domain-general processes to articulate the UG-access question or to formulate the FDH. To say that the processes of the language faculty are not unique to the language faculty is not to say that a language module does not exist. The language faculty is a system of processes, and the fact that the processes used by the faculty may also be shared by other systems does not show that the module, qua system, does not exist. The faculty of language is a “subcomponent of (mostly) the brain” (Chomsky, Reference Chomsky2004, p. 104), but “its elements might be recruited from or used for other functions” (p. 124). So, the question of whether foreign language is handled by the language faculty does not reduce to the question of whether the processes used by foreign languages are the same as those used in native languages.
To illustrate this point, an analogy may be helpful. Special-purpose computer programs—a word processor, a spreadsheet, or a concordancer—may well recruit the same routines (perhaps provided by the operating system); for example, on my computer, almost all application programs use the same text-entry and text-editing systems as well as the same file-management systems. The generalized regular expression processor of Unix is recruited in a wide range of programs, with minimal adaptations. The fact that two programs may use the same routines does not make them the same program. The user can use different programs to perform a similar function, and these different programs may use the same routines in performing those functions. In a set of corpus files, I may want to find a linguistic pattern and list the lines in which it occurs. For this, I may use a command-line pattern matcher (such as grep in Unix), or I could write a small special-purpose script (using a language like perl or python), or I can use a specialized text-processing program with a multifile search capability, or a special corpus-linguistics concordancer. These different systems all use essentially the same basic method of pattern matching; these pattern matching methods in the different systems are not sui generis. Nonetheless, the systems are different systems: The command-line grep is not the same as the concordancer. Additionally—this is important—using one program-system may be more efficient than another, depending on a more encompassing goal. A system is more than the sum of its parts: The manner in which the parts are integrated matters.
To return to the FDH: In the context of the original FDH, it was possible to say that child language development used a rich UG and that foreign language learning relied, instead, on the L1. However, rich UG may not exist. Additionally, if all UG principles—minimal as they are—are instantiated in all languages, then the distinction between access to UG and access to UG through L1 is without content.
In the original context of the FDH, it was also possible to say that child language development used domain-specific data-processing and learning procedures and that foreign language learning relied, instead, on domain-general learning and data processing. However, it is possible that the elements used by the language faculty are “recruited from or used for other functions” (Chomsky, Reference Chomsky2004, p. 124). If there is no distinction between sui generis language processes and other processes, then that distinction cannot be used for the FDH.
TOWARD A NEW FORMULATION
Here is the dilemma: Without rich UG and language-specific processes, we are left without a framework within which to formulate an account of the properties of unreliability and nonconvergence. Essentially, the property of language, if taken seriously and combined with contemporary views of the nature of language, deprives us of an explanation for nonconvergence and unreliability. The next subsections propose a way forward: The leading idea is that elements of the language faculty itself provide the seeds of nonconvergence and unreliability.
Patches as a Bridge
Consider the consequences of this quote from Wodehouse’s (1946) Joy in the Morning reproduced in (6):
(6) “I balanced a thoughtful lump of sugar on the teaspoon.” (chapter 5)
and these (also from Wodehouse) from Uneasy Money (1917) and Jeeves in the Offing (1960) in (7) and (8), respectively:
(7) “He was now smoking an uneasy cigarette and waiting for the blow to fall.” (chapter 9)
(8) “His eyes widened and an astonished piece of toast fell from his grasp.” (chapter 2)
The sugar is, of course, not thoughtful; nor is the cigarette uneasy; nor is the piece of toast astonished. Whereas Wodehouse’s examples are particularly striking, the device is actually not that unusual, rather it is the well-known transferred epithet in traditional rhetoric: There is the “restless night.” It intrigued linguists early in the development of generative grammar—and was first noted by Hall (Reference Hall1973). It is of interest because it is a novel way of associating grammatical structure with conceptual structure that does not inevitably follow from general principles; Hall spoke of the generalization as being “on the semantic plane” (p. 94). Although it is no doubt rooted in the mechanisms of metaphor (Warren, 1983) and one can understand it readily, it can become conventionalized—certainly, it became conventionalized for Wodehouse—and one can learn it.
Of course, there is nothing really new in the observation that there are idiosyncratic structures and mappings from syntax to conceptual structure in a language. The what is X doing Y (WXDY) construction, as in What is that fly doing in my soup? for instance (Fillmore & Kay, Reference Fillmore and Kay1999), is a notable example that has been an inspiration to construction grammar. Knowledge of WXDY is part of the grammar of (many, although probably not all) native speakers of English.
Morgan (Reference Morgan1972), in a study of verb agreement in English, noted that speakers of English sometimes boggle at examples like (9) or (10).
(9) ??Are/Is (either) John or his parents here?
(10) ??Are/Is (either) John’s parents or his wife here?
Morgan observed:
The speaker learns a relatively simple principle of agreement which somehow fails to extend to complex cases. This sometimes leads to patching the principle by adding subsidiary principles, and [verb agreement] VA ends up as a (possibly hierarchical) set of principles. The details of the patching vary considerably from speaker to speaker, and some speakers apparently fail to patch at all for certain cases. (pp. 284–285)
Morgan suggested that patching happens when, for some reason, the general mechanisms of grammar learning fail. Learning fails because a single general principle cannot be derived from the available data and extended to all of the relevant cases. In cases for which the usual general language acquisition processes cannot apply, the learner is driven to develop a patch—to use Morgan’s terminology—for the resulting hole caused by acquisition failure. Patches have a couple of characteristics of particular interest to those who are concerned with SLA. First, there is considerable variability among speakers (i.e., diversity in Morgan’s terminology). This contrasts with the complete unanimity that one expects to find in UG-driven development—what we are calling convergence. Second, there is often uncertainty within a speaker: Morgan said some speakers “just don’t know” (p. 284).
Emonds (Reference Emonds, Brame, Contreras and Newmeyer1986) noted the considerable interspeaker variability (and uncertainty) in the mastery of prestige rules of prescriptive grammar. These, too, are not derived by the normal acquisition system. In this case, Emonds noted, the variability is exploited socially as a device for “arbitrarily signaling membership” in the “business-government-legal-religious-academic—the high culture ‘community’ (i.e. class)” and for “re-enforcing exclusion from it” (pp. 123–124). The TOEFL exploits the variability and uncertainty in the foreign language knowledge to effect similar inclusions and exclusions.
McCawley (Reference McCawley1988) extended the notion of patch to cases like he can’t seem to do it right (where can’t clearly does not have scope over seem), the usual men were drunk (men are not usual), and similar cases. These are reminiscent of the Wodehousean transferred epithet (as McCawley himself noted). McCawley suggested that
a patch is a way of employing the syntactic means provided by the language so as to allow one to say things of the same sort that can normally be expressed by a given construction in some class of cases that the normal syntactic means of the language do not provide for. (p. 746)
One might understand “normal syntactic means” (p. 746) as the means developed in the operation of normal acquisition, to use McCawley’s terminology.
The importance of such phenomena has become apparent to contemporary mainstream generative grammar. Lasnik and Sobin (Reference Lasnik and Sobin2000) developed a theory of what they called viruses. (They explored the properties of the rules for whom.) Viruses—the concept of which seems very similar to patches—are superficial and parasitic on existing structures: These are shallow principles. Sobin’s earlier (1997) study of viruses, published in Linguistic Inquiry, included tables of grammaticality judgments that could have been taken right out of an article published in Studies in Second Language Acquisition. The grammaticality judgments presented in Sobin’s tables show the same sort of nonconvergence and uncertainty that we expect to find in SLA research.
The crucial point in all of these cases is this: Although can’t seem, the usual men, what’s that fly doing in my soup, for you and I, and the person whom I am told is responsible are not part of the core system that arises by the interaction of the principles of UG and input during development, there is no reason to believe that they are not language and that the structures cannot interact with the language module. Lasnik and Sobin (Reference Lasnik and Sobin2000) put it very clearly: “Viruses comprise a subtheory of the Minimalist Program, distinct from the core system, though interactive with it” (p. 369). Although viruses are not part of the grammar proper, they are linguistic. With foreign language learning in mind, consider this observation:
Grammatical viruses mimic the original phenomenon, but without complete success, since grammatical viruses are not of the same descriptive power as the devices of the grammar proper. Viruses lack generality and the possibility of normal interactivity with the core grammatical system. Such mimicry via grammatical viruses leaves the sort of telltale discrepant behavior which we have discussed here. (p. 370)
How much can patches do? Lasnik and Sobin (Reference Lasnik and Sobin2000) aimed to constrain patches tightly: “Virus Theory claims that such manipulations are not manipulations of the true, abstract system, but are quite superficial surface manipulations of limited features of the objects that the system produces” (pp. 367–368). Nonetheless, the system clearly has access at least to syntactic category information and to concepts like subject and theta-assigning verb. Moreover, the system can contain statements such as check ACC (on a particular element), which shows that abstract features and at least the basic ideas of a morphological mechanism are available. Lasnik and Sobin are concerned to limit the extent to which linguistic-theory-internal concepts can be referred to in a patch. (It is clear that there are dangers to Fodorian modularity if the patch system, which seems to be affected by conscious awareness and may be amenable to explicit instruction, can manipulate objects deep within the core.) In this vein, Lasnik and Sobin proposed, for example, that “subject status may be determined here by considering the theta role or meaning of elements 1-2 (that is, whom) and 3, something that a speaker must be cognizant of to understand the sentence” (p. 359). Note that the elements that can be used in viruses are those that the speaker can be cognizant of (although it is a bit difficult to see that the learner can be cognizant of the abstract feature-checking mechanism). To the extent that foreign language learners are creating patches, cognizance puts restrictions on the abstractness of the rules that can be proposed. (For foreign language learning, the connection to noticing theory [Schmidt, Reference Schmidt1990] is obvious.)
However, the tight constraints that Lasnik and Sobin (Reference Lasnik and Sobin2000) proposed may not be inherent limitations of the patching system. It may simply be that the core acquisition system of native language development is robust enough so that holes are tiny as well as few and far between. Who knows what patching might be capable of? What if the core acquisition system simply does not work (to paraphrase Chomsky, Reference Chomksy1997) and the holes are many and large?Footnote 13
Data Processing Outside the Language System
Can “is just not working” (Chomsky, Reference Chomksy1997, p. 128) be characterized within contemporary conceptualizations of the relationship of domain-general and domain-specific systems? Suppose, following Chomsky (e.g., Reference Chomsky2005) and much current work in mathematical learning theory and natural language processing, that the elements recruited by the language system are not specific to language. Suppose the language system itself is not working properly. This does not mean that these processes, which are not restricted to language, are not available. The whole system is greater than the sum of its parts. To return to the computer program analogy, a specialized concordance program—of the type that might be used in a corpus linguistics class—is basically just a packaging of otherwise available multipurpose routines into an efficient and easy-to-use special-purpose package. If you do not have such a specialized program, you can still do the homework, by pulling in individual components, cobbling something together. However, it will be less efficient than an integrated system and probably more prone to error. (Additionally, different students in the class are more likely to get different results.)
Shallow Processing
In addition to patches or viruses and the use of data-processing techniques apart from their integration into the language faculty, recent discoveries about language processing give us another piece of the puzzle. There are two parts to this language processing picture: The first involves native language and the second involves foreign language.
First, there is increasing reason to believe that (native) language processing employs both what one might call shallow mechanisms and deep mechanisms. Townsend and Bever (Reference Townsend and Bever2001)—who accept mainstream Minimalism—proposed that “while grammatically defined representations appear to be computed during language behavior, the grammatical rules that define them may not be used” (p. 45). Processing often uses relatively superficial patterns that reflect the statistical structure of language.
Second, for L2 processing, Clahsen and colleagues have reviewed existing research and reported neurophysiological work using measures such as event-related potentials (see Clahsen & Felser, Reference Clahsen and Felser2006b, and the references cited there). The evidence thus far is not completely conclusive. However, Clahsen and Felser tentatively proposed the shallow structure hypothesis: “For nonnative language processing, we suggest that the syntactic representations adult L2 learners compute for comprehension are shallower and less detailed than those of native speakers” (p. 32). Of course, if Townsend and Bever are right, then native speakers also can and do compute shallow representations, perhaps using them in parallel with grammatically defined representations. If this is true, then nonnative speakers in computing shallow representations are employing mechanisms of the language faculty, but they rely on the shallow route exclusively or primarily, perhaps because grammatically defined representations are unavailable or not fully definite (i.e., patched).Footnote 14
There is an increasing consensus among processing researchers that some kind of dual-mechanism model will have a place in accounting for differences between native language acquisition and adult foreign language learning. Ullman’s (e.g., Reference Ullman2001, Reference Ullman and Sanz2005, Reference Ullman2006) application of the declarative-procedural distinction seems particularly promising. Under Ullman’s (Reference Ullman2001) model, “linguistic forms whose grammatical computation depends upon procedural memory in L1 are posited to be largely dependent upon declarative/lexical memory in L2” (p. 105). “L2 learners should tend to rely heavily on declarative memory, even for functions that depend on the procedural system in L1” (Ullman, Reference Ullman2006, p. 99); that is, in L2 there is a “shift from procedural to declarative memory” (Ullman, Reference Ullman2001, p. 110) of grammatical processing. It is quite possible to relate this idea to the concept of patches. One might conceptualize patches as words, which would thus be handled by declarative-lexical memory in L2. Indeed, Ullman (Reference Ullman2006) suggested that L2 learners may rely largely on “stored schemas or constructions” (p. 99), referring to Fillmore et al. (1998) and L2 productivity may involve “the use of ‘rules’ learned in declarative memory” (Ullman, Reference Ullman2006, p. 99); thus a patch could be thought of as such a “rule”—although not as a rule (i.e., without quotes). Paradis (e.g., Reference Paradis and Ellis1994, Reference Paradis2004) has made a rather similar proposal, but he also focused on the distinction between implicit and explicit knowledge—these are related to level of conscious awareness.
Although there are differences among the proposals now being made in the psycholinguistics and neurolinguistics of L2 processing, there seems to be a general compatibility between them and the proposals made here in the context of the FDH. Still, the correspondence between the family of dual-processing proposals and the FDH is not completely straightforward. Much depends on working out the knotty relationship between “getting good at something” (through practice), the status of different memory and processing systems, or both, and the problem of conscious awareness.
SUMMARY AND CONCLUSION
The proposals outlined here for recasting UG-in-SLA research in light of the evolving linguistic-theoretic context share a common thread. In characterizing the fundamental nature of foreign languages, these proposals make use of possibilities that are inherent in the language module but that, for native languages, may not be utilized in the same way or to the same extent. Patches or viruses, for example, are peripheral in the case of native languages: They are a minor factor—more like a wrinkle in the system—but they are there as a potentiality, awaiting more extensive exploitation. In foreign languages, they may carry great weight. The data-processing elements recruited by the native language acquisition system are not specific to language; therefore, these elements can be used even if the language acquisition system is no longer working properly, although they may not work as well for language apart from their integration in the system. The shallow representations that foreign language employs in online processing are not absent in native language processing, but, in foreign language processing, they are relied on more heavily (or even exclusively).
The research program that would arise under this conception would have three thrusts. First, we need a much greater understanding of the nature and limits of patches and prescriptive rules, both in native languages and in foreign languages. Second, because we now understand that the data-processing elements of the language module are not sui generis, we can begin to integrate discoveries from machine learning, mathematical linguistics, probabilistic reasoning, and the like into foreign language acquisition theory. Crucial to the newly recast FDH is the speculation that such processes can work even outside a (hypothetically impaired) dedicated language acquisition system but will be less reliable when not integrated into the specialized system. Third, psycholinguistic research into online processing is required to explore further the interplay of rough-and-ready shallow processing and full grammar-driven processing.
To conclude, let us return to Chomsky’s (Reference Chomksy1997) statement of the FDH. The key notions were “just not working” and “something strange” (p. 128). We can change “just not working” (p. 128) to “just not working in the same way”: The core system is not working (or not working well), so the patching system is taking the burden. To the foreign language learning system, language is neither more nor less strange than the funny order of can’t and seem in he can’t seem to get it right. The language faculty has means of dealing with strangeness and uncertainty, and these are heavily employed in foreign language learning.
