The articles in this special issue present very interesting data on the acquisition of relative clauses (RCs) in Japanese, Korean, and Cantonese. The errors and developmental patterns are partially different from those that have been observed in European languages, and they appear to go against predictions made by the Keenan and Comrie (1977) noun phrase accessibility hierarchy (NPAH). Whenever we have data that are challenging like this, we need to ask some basic questions. In the present case, I see two big issues:

1. What exactly are the universals of relativization?
2. What do they predict for the acquisition of first languages (L1s) and second languages (L2s), respectively?

With respect to point 1, there is an important prior question: What exactly is a RC? This is a nontrivial issue in the East Asian context. Comrie (1998a, 1998b), following Matsumoto (1997), proposed that we need to distinguish between RCs and attributive clauses and that many structures analyzed hitherto as RCs in Asian languages should be reclassified as attributive clauses. RCs, in turn, can be of different types—head-external, head-internal, or headless—and the precise grammar of each type needs to be defined and contrasted with that of attributive clauses (see Kanno's article, this issue, for some relevant issues). The Keenan and Comrie (1977) article distinguished different RC strategies, defined in terms of linear ordering and case-coding, and their universals were actually formulated in terms of these strategies. Maxwell (1979) defined the strategies differently. My 1999 article (Hawkins, 1999) built on Keenan and Comrie's work and revised their proposed strategies in order to bring them into (what I argued was) better alignment with the reverse hierarchies for gaps and resumptive pronouns across languages and with the processing typology approach of Hawkins (1994, 2004).

How we resolve these questions and define RCs will determine the universal generalizations that we can formulate and the predictions we make for language acquisition. If Japanese does not have a (filler-gap type) RC but an attributive clause, then universal generalizations formulated for RCs do not apply to this language and to its acquisition. If Japanese does have RCs, then these generalizations do apply (again, see Kanno, this issue). Similarly, universals formulated in terms of strategies will depend on how the strategies are defined. I argued in Hawkins (1994, 1999) that the Keenan and Comrie (1977) data lend themselves most readily to implicational generalizations that include the following:

Defining the gap strategy in generalization (ii) is not trivial. Do we postulate an actual gap coindexed with the head of the relative, as in government binding (Chomsky, 1981) and other trace theories? Or do we coindex the head with the subcategorizer in the RC, by a direct association, as in head-driven phrase structure grammar (Pollard & Sag, 1994)? I proposed a hybrid definition of a “filler-gap domain” in Hawkins (2004, p. 175) that tried to incorporate the insights of both approaches.

Descriptive universals in any area of grammar will generally be shaped by the kind of explanation we believe to underlie them and by the details of that explanation. Yip and Matthews (this issue) argue for a processing explanation for generalizations (i)–(iii), and they propose some interacting considerations of learning and L1 transfer to account for the mismatches between these generalizations and their Cantonese-English acquisition data. I find their account compelling. There are so many universals of relativization that are quite mysterious from a purely grammatical point of view—hierarchies of subjacency effects, limitations on otherwise productive grammatical processes, reverse hierarchies for gaps versus resumptive pronouns—and I have argued (Hawkins, 1999, 2004) that processing efficiency and complexity can give a unifying and motivated account of these. There are different ways of formulating the processing complexity that might underlie the NPAH. My definition (Hawkins, 2004) of the structural domains for relativizable positions on the NPAH included all terminal and nonterminal nodes on the path from the filler to the relevant subcategorizer or gap. O'Grady, Lee, and Choo (2003) focused only on higher phrasal nodes in the tree geometry. Diessel and Tomasello (2005) have critiqued both and argued for a number of complexity factors that correlate with the NPAH and impact the order of acquisition in English and German.

Classical generative theories, by contrast, argue for the innate basis of universals, including variation-defining parameters (Chomsky, 1981). These theories have never had much to say about the NPAH and its correlated generalizations. Chomsky has now scaled back the universal generalizations for which he proposes an innate grammar, reducing the innateness of grammar to recursion only (Hauser, Chomsky, & Fitch, 2002), whereas Newmeyer (2005) argues that all universals of language variation should be handled by theories external to Universal Grammar (UG), such as processing. These proposals impact the assumption that a richly specified UG is available to L1 and L2 learners (Hoekstra & Kooij, 1988; White, 1989), they limit the potential relevance of UG as a solution for the negative evidence problem in language acquisition (Bowerman, 1988), and they increase the role of learning from positive data, something that Tomasello (2003), connectionist modelers like MacDonald (1999), and linguists like Culicover (1999) have all been arguing for independently. These converging developments greatly increase the number of descriptive universals for which a processing explanation now becomes a real possibility. Also, a processing approach enables us to recast the learnability problem and to see the positive data as less impoverished and more learnable than previously thought (Hawkins, 2004; O'Grady, 2005).

It should be clear from this brief discussion that the answer to question 1 is not straightforward. Defining the universals of relativization—even the limited set that are relevant for the NPAH—requires agreed definitions of RC and of RC strategy. Yet, important details are still unclear descriptively, and there is much debate about deeper explanations. This is not an ideal foundation on which to base universal predictions for the acquisition of these structures in L1 and L2 contexts, especially in languages whose status is at the center of the debate, and it clearly complicates the discussion and the conclusions drawn in the articles of this special issue. However, let us see what can nonetheless be said about acquisition and what cannot, and let me pull out some points from the articles that I think are important in relation to question 2 and in relation to research that needs to be done in the future.

Jakobson (1941/1968) was perhaps the first to argue that one can derive interesting predictions for acquisition (and for language loss) from universals, and he tested a number of these in the area of phonology using implicational generalizations that were available at the time. These predictions included order of acquisition (the consequent property in the implicational universal should be acquired before the antecedent), more errors of one type than another (more antecedent property errors than consequent property errors), and predictions for the directionality of substitutions in the event that one property or structure is substituted for another (with the consequent substituting for the antecedent, rather than vice versa). I summarized these predictions in Hawkins (1987), proposed a reformulation (to which I will return), and tested them on some additional data that went beyond phonology.

The basic premise of any translation from universal patterns in natively spoken languages to corresponding patterns in acquisition and interlanguages is that we expect, all things being equal, that the same principles of variation will be found in both. If, say, there are languages like Malagasy that allow relativization only on subjects, then (Jakobson, 1941/1968, would argue) it is reasonable to expect acquisition stages in prenative English that are like Malagasy with subject (SU) relatives only, or with more errors on direct objects (DO) than on SU, or with substitutions of SU relatives for object relatives. Indeed, there is much support for these acquisition stages and for this general logic.

However, we need to bear in mind the limitations of this exercise (i.e., what cannot be predicted from universals for acquisition), we need to be clear about what the predictions are, and we need to know whether there are any other factors that can interfere with the predictions.

First, we have to make sure that the relevant property or structure in the acquisition data is a genuine instance of the property or structure that figures in the universal, if we are going to apply the universal to these data. Assume, for the sake of argument, that we can resolve the RC versus attributive clause issue and that we have a set of universals defined on RCs. Are the kinds of prerelatives that we find in early acquisition, even in head-initial European languages like English (Diessel & Tomasello, 2005), really RCs? If not, the universals do not apply to them. Once nativelike structures are clearly attested that satisfy whatever definition we have of a RC, then predictions can be made for developmental stages, errors, and so forth. This obviously requires some nontrivial analysis of the acquisition or interlanguage data.

My next point is often overlooked. I argued in Hawkins (1987) that Jakobson (1941/1968) did not get the relationship between universals and acquisition quite right. The problem is this: Given an implicational universal of the form “if a language has property P, then it has property Q,” you cannot logically derive the prediction that property Q should be acquired before property P. More concretely, from the observation that relativization on a direct object (P) implies relativization on a subject (Q), you cannot derive the prediction that learners will necessarily use SU relatives first or make fewer errors relativizing on subjects than on objects. Why not? Because in addition to languages like Malagasy that have subject-only relatives (−P & Q), there are languages like English that relativize on subjects and objects (P & Q). The predictions for acquisition are derived from the set of all natively spoken languages whose properties the universal describes. Hence, the most we can predict for successive acquisition or interlanguage stages is that any stage will either be like native Malagasy or like native English, and it will not be like the unattested native languages and have DO relatives only (*P & −Q).

An “if P then Q” universal accordingly predicts that an early acquisition stage with neither property (−P & −Q) can be followed either by the acquisition of the consequent property Q alone (−P & Q) or by the acquisition of both P and Q together and that the antecedent P should not precede the acquisition of Q (*P & −Q).

Jakobson (1941/1968) assumed that an implicational universal does predict the acquisition of Q first. So do subsequent researchers, whose work I summarized in Hawkins (1987). It is also assumed in some of the articles of this special issue. It is important that we clarify this issue, because it impacts whether we can say that the predictions of the NPAH are actually correct in East Asian languages.

The NPAH predicts [for relativization in general, cf. generalization (i)] that each higher position will be relativizable either before or simultaneously with all lower positions, and it rules out relativization on a lower position prior to a higher position. More precise predictions for the different gap and resumptive pronoun strategies [in generalizations (ii) and (iii)] will follow the same implicational logic but in a reverse direction (if low, then high for gaps, and if high, then low for pronouns). A subject-gap relative will occur before or simultaneously with an object-gap relative, an object resumptive pronoun will occur before or simultaneously with a SU resumptive pronoun, and so on.

If in L2 Japanese, therefore, SU, DO, and oblique (OBL) relatives all arise simultaneously and are equally correct, this is fully compatible with the NPAH. If SU relatives arise before DO and OBL or are more frequent and error-free than DO and OBL, this is also compatible with the NPAH. The former scenario means that Japanese learners are behaving like native English speakers with many relativizable positions, the latter means that they are like Malagasy speakers with just one. What is predicted not to happen is DO relatives arising in the absence of, and earlier than, SU relatives, a pattern for which no natively spoken languages are attested.

The significance of the Cantonese-English bilingual data in Yip and Matthews' (this issue) article and of the Japanese L2 data with Mandarin learners in Kanno (this issue) is that there are apparent counterexamples of this type. Object relatives occur first, and more productively, in both the early Cantonese and the early English of Cantonese-English bilinguals and in the Japanese L2 data of Mandarin learners.

Faced with this situation, we must look for possible causes. Yip and Matthews (this issue) point to structural properties of Chinese that make object relatives easier to learn by virtue of their greater similarity to main (nonrelative) clauses, a consequence ultimately of the mixed word order typology of this language, in which clauses are SVO and noun phrases are head-final and premodifying. Relativizations on object positions preserve this basic SVO structure, although relativizations on SU positions do not. The authors also show that there is considerable transfer of the premodifying structures of Chinese into the early English of the bilinguals.

In other words, Yip and Matthews (this issue) appeal to general learning strategies and to transfer effects that apply to the structural details of these languages, and they see these factors as interacting with the predictions of a set of (processing-driven) universals. This results in a partially different set of acquisition and interlanguage data than would be predicted by the universals alone. The object relative advantage for Mandarin learners in Kanno's (this issue) Japanese L2 data might also reflect the same factors for which Yip and Matthews argue in explaining the Cantonese data.

On the other hand, recent experimental work by Lau (2006) found that, in accordance with the NPAH, Cantonese monolingual children performed better on SU than on object relatives. Yip and Matthews' study (this issue) analyzes spontaneous production as recorded in diary data, whereas Lau used elicited imitation and act-out tasks, both of which included an element of comprehension, and these task differences could be significant (Lau argued that they are). Yip and Matthews also point to the fact that the early object relatives in their data could be analyzed as internally headed relative clauses (IHRCs). As in Korean (Jeon & Kim, this issue), IHRCs in Cantonese might emerge before head-final RCs. Because the possibility of an IHRC structure arises in Cantonese only with object relatives, not with SU relatives (Matthews & Yip, 2002), this might explain why object relatives appear to emerge earlier (S. Matthews, personal communication, October 2006).

In sum, it remains to be seen whether there are genuine counterexamples to universal predictions in these data. Whatever the outcome, there is an important moral we need to draw from this special issue in the current state of the art. We need to view L1 acquisition and SLA as complex adaptive systems (Gell-Mann, 1992) in which multiple factors interact with different strengths at different periods. In a similar vein, Diessel (2004) and Diessel and Tomasello (2005) point to a number of factors that facilitate the learning of RCs in English and German. The kinds of factors that have emerged from these and other studies include the early occurrence of relatives on propositionally simple predicate nominals (see also Jisa & Kern, 1998, for French), the similarity of early relatives to main clause word order (Bever, 1970, for English, and Yip & Matthews, this issue, for Cantonese), the preference for relatives on the single argument of an intransitive clause prior to multiple-argument transitive and ditransitive clauses (Diessel & Tomasello), animacy effects as temporary generalizations that facilitate the acquisition of SU versus object relatives (Jeon & Kim, this issue, for Korean, and Ozeki & Shirai, this issue, for Japanese), sensitivities to structural frequency effects in acquisition (MacDonald, 1999), and the operating principles and cognitive prerequisites of Slobin (1973, 1985). The outcomes of these general learning strategies will be partially different in the different language types, and the transfer possibilities will vary for different pairs of L1 and L2, as shown in Kanno's (this issue) data for head ordering transfers from different L1s into L2 Japanese.

Putting all of this together, we can say that the predictions made by language universals for acquisition and interlanguages appear to hold, all things being equal. However, things might not always be equal! General learning strategies and transfer effects can impact learners' output in different language development situations. We also have to be careful about the structural analyses we give to these early RCs and about differences between the kinds of data we collect. The challenge for future work, as I see it, is to better understand how the different factors—universals, learning, and transfer—work together. We need to develop a model that actually predicts the outcomes for given pairs of L1 and L2 instead of just observing their effects after the fact, as we are doing at the moment. In short, we need a predictive complex adaptive model of language acquisition in the spirit of Gell-Mann (1992). Currently, we rely primarily on the universals for our predictions, and we take note of the exceptions and try to make sense of them on a case-by-case basis. However, the predictive factors must be more coherently integrated. We need to be able to predict that general learning strategies, when applied to a particular structure, will yield a particular acquisition sequence, that transfer from L1 to L2 will be favored when L1 has certain properties and L2 has certain other properties (e.g., parsing-favored vs. disfavored RC orders in Kanno, this issue), and that universal predictions will or will not be temporarily sidetracked for a given structure in a given acquisition context, and so on.

Progress in this multifactor direction is just getting underway. It requires interdisciplinary research environments in which psychologists working on processing and learning collaborate with typologists and generative grammarians and with computational modelers. It requires lots of empirical data from different languages and language pairs as well as theoretical and computational sophistication. The articles of this issue are an important contribution to where we need to go.