Rethinking Parameter Theory in Language Acquisition: Towards a Complex Learning Approach, Papers of Education Planning And Management

Download Rethinking Parameter Theory in Language Acquisition: Towards a Complex Learning Approach and more Papers Education Planning And Management in PDF only on Docsity! MIND THE GAP: EPISTEMOLOGY AND DEVELOPMENT OF NATURAL LANGUAGE Teresa Satterfield, University of Michigan Anjum Saleemi, National Chi Nan University I. INTRODUCTION This paper is part of a newly-initiated research program, as such it is an effort at resolving a longstanding tension between certain related epistemological conceptions of natural language and the real-time evidence from L1 language acquisition. We hope to bridge this gap by exploring a unique account of syntactic development that departs in critical ways from standard conceptions of child grammatical development. Specifically, in the most prominent generative-grammar based acquisition studies Universal Grammar (UG) is viewed as an overarching innate linguistic system constituting a unit of universal principles, some which contain parameters (or p-parameters, subsets of grammatical principles, such pro-drop as described in Hermon (1990)) as choice points which can be fixed in one of a limited number of ways. A particular grammar is assumed to be immediately derived from UG by ‘fixing’ the parameters in a certain way on the basis of experience: Spanish, Russian, etc., are direct expressions of UG under particular, and distinct, sets of parametric values. No theory of language learning is explicitly postulated; instead, most final state grammatical structures are assumed to be directly attainable by UG principles, under the parametric options. Under this conception acquiring a language becomes merely a trivial endeavor: the initially novel assumption of language variation adopted in the Principles and Parameters approach (Chomsky 1981, etc.) now is subsumed in the hypothesis that the mechanisms of language variation and learning are identical--one and the same--and that children really do ‘set’ parameters, in the sense of selecting, among the options generated by the mind, those which match experience, and somehow discarding all others. In this paper, we entertain the hypothesis that the parameters associated with UG principles, as characterized in the standard acquisition literature are not conceptually desirable. These assumptions remain dominant in the literature because, yes, parameter theory has provided an entirely new way to think about language acquisition; and yes, in the process it has yielded certain interesting ideas about language uniformity. However, as we will argue, the presupposition of such a type of parameter leads to a number of difficulties. Notably, when attempts are made to explicitly formalize the standard claims using mathematically-based models, failures arise because parameter theory incorrectly assumes that language learning mechanisms must be impoverished just because those of language variation appear to be (Clark 1992; Gibson and Wexler 1994; also contributions to Bertolo 2001). Secondly, because the standard view of language variation adopted is grossly oversimplified in the first place (e.g., [+/- value]), methodological and interpretive difficulties emerge in the most basic analyses (Culicover 2000). For instance, there exist an empirically significant number of adult grammars whose syntactic systems are partially ‘split’ or internally ambiguous with respect to the presumably mutually exclusive values of parameters represented by the ‘core’ phenomena (Satterfield 1999, 2003; Saleemi 2002). Take examples in movement, such as wh-questions in French, or scrambling in West Germanic; or in word order, as in the SVO/VSO alternation and asymmetrical agreement in Standard Arabic, where often neither variation is considered more ‘basic’. Standard Arabic manifests two different and co-existing word orders in addition to alternate inflectional patterns. SVO order elicits verb agreement in person, number and gender with its DP 2 subject, whereas VSO order yields verb agreement of person and gender only with the subject. It is important to note that these ‘mixed paradigm’ facts are not epiphenomenal, but rather are systematic variations that present several problems for standard parameter theory, and in turn for learnability theory, in their current conceptions.1 Using the ‘mixed’ paradigm data as support, this paper addresses the issues of determining the relevant characteristics of syntactic variation and distinguishing universal principles of grammar from optional features in order to begin to provide a plausible account of the acquisition of L1 syntax that reflects real-time data. We begin with two working hypotheses, stated, below: (1) DISTRIBUTED UG HYPOTHESIS: UG is made up of a multi-level, separable system of learning mechanisms. Within UG, the L1 learning system need not be an undifferentiated “blackbox” which is uniformly impoverished, as is commonly assumed. Language is a complex cognitive domain permitting a degree of variation that is by no means insignificant. Given our conception of variation below, a grammar is not, perhaps normally cannot, be a completely unified (in Jackendoff’s 2002 terms, a ‘monolithic’) system. (2) VARIATION WITHOUT PARAMETERS HYPOTHESIS (VWP): Syntactic variation is not parametric. We contend that linguistic variation is a cumulative outcome, the interplay between the highly restricted tenets of UG with lexical items and, by extension, with the classes of derivative linguistic constructions. In the remainder of this paper we present some in-depth arguments to support the Distributed UG and the Impoverished Parametric Variation Hypotheses. We will then further argue against parameter theory in its standard formulations as a mechanism of grammatical acquisition; that is, we will demonstrate that certain “contradictory” options are maintained even after the parameter has presumably been fixed for the alternate value, using cases across various languages as examples. We then turn to the alternative learning architecture itself, hypothesizing that within the language faculty (FL) the human language learning system comprises a collective of both domain-specific and domain-general (as opposed to domain-neutral) learning mechanisms (see Culicover 1999, Saleemi 2002, Yang 1999, and Satterfield and Pérez-Bazán (in preparation) for related but by no means identical views).2 In the proposed account, the learner is endowed with two types of UG principles, both generic and ‘systematic’ but there are none of the 1 And while we agree with the research strategy of limiting the data, somehow extracting from the wide-range diversity found across human languages to provide the linguist with an idealized decontextualized system within which hypotheses can be tested and manipulated, the object of inquiry cannot be simplified to the point of not representing the true structure of the grammar. To what extent is one justified in extrapolating from the analysis of one language to another: related languages typically contain subsets which are deceptively identical in some (many?) respects, and a closer look is usually likely to reveal significant differences in the underlying grammars. We think in some of the current linguistic theory the practice of generalizing from one language to another is indulged in rather too quickly. 2 As will be seen, in order to develop an account of a domain-specific learning mechanism, we need to do two things. First, we need to come up with a principled way to individuate domains. Second, we need to say what makes a learning mechanism specific to a domain (or not). 5 language, yielding an infinite number of interpretable expressions. That is, if the strong ‘No Dead End Condition’ obtains, then the task of determining the possible languages is mitigated, as minimalist assumptions hold that you cannot have linguistic variation in such a way that you get a system that will fail to have an infinite satisfaction of the interface conditions, including in the initial state. The question of convergence to a possible human language is then redirected to the question of how exactly the learner converges to the ambient variations. In addition, there is a more specific dimension of the learnability problem, which is how to delimit the hypothesis space such that it can be efficiently searched by means of a memory-less enumeration function with access to solely positive input. Models within the standard Principles and Parameters approach (Chomsky 1981; Chomsky and Lasnik 1993) support the notion of a deductive learner who, driven by input data from the environment, searches the finite parameter space as defined by UG to select the appropriate binary-valued parametric option (and associated cluster properties) as suitable for the target grammar. At any point in the process, a single grammar is made available to the L1 learner. Recall that for Chomsky, parameters are set merely based on a preponderance of evidence, with no implicit learning mechanisms in place. Saleemi (2002) has pointed out that this conception is fraught with learnability problems. For example, even by constraining choices to only yes or no options, a linguistically simple 10-parameter space still yields a daunting number of grammar combinations to be considered by the learner (210 = 1024 possible options). Moreover, in the worst case, if certain parameters are linked, setting a sequence of parameters in the face of independent cluster properties would soon become an insurmountable task. Several well-known accounts maintain that a triggering model can allay these difficulties; however Gibson and Wexler’s (1994) analysis appears to inherit all of the conceptual problems of the account proposed by Chomsky, in that the GW model similarly operates on the learner moving through a [+finite] hypothesis space to select a grammar based on highly idealized input, based on a class of Comprehensive Triggering Algorithms. Moreover, to the degree that this model reflects real-time acquisition, questions arise regarding GW’s rationale for implementing certain default parameter values for the potential success of V2 acquisition in a three-parameter learning space. Even granted the viability of standard triggering models to provide a reasonable solution to the learnability problem, a related difficulty still concerns ambiguous input, in the form of hypothesis sentences that correspond to several possible grammars. The logic follows that if the binary-valued parameters are fixed via very limited linguistic evidence (Chomsky 2000), then an undesirable outcome is that a triggering model is able to drastically alter the learner’s grammar based on one input sentence. Thus, as a result of ambiguous incoming data, the all-or- nothing constraint imposed on triggering could easily lead the learner to converge to local maxima, and inevitably to a grammar that did not correspond to the target. If conditions of learnability could actually be met under trigger-parameter theory, another important question that comes up is the possibility of inducing unambiguous triggers. Fodor (1998) not only demonstrates clear-cut evidence of unambiguity for each of the grammars in the three-parameter domain, as proposed by GW (1994), she then formulates an alternate triggering model that implements only unambiguous input. On a more trivial note, parameters are presumably fixed based solely on their correct “match” with the ambient data. At a very fundamental level then, parameter-setting, especially when viewed as a fine-tuning of the invariant grammar endowed in UG, is merely an exercise in associative learning. Since the nature of the information that the learner extracts from the input data and how it is represented is as yet unknown to us, how can we have a significant discussion of “triggers” at this point? In light of the important projects recently initiated (Fodor and Sakas, (pc)), hopefully we can begin to address notions of triggers, by 6 exploring probabilistic learning mechanisms with natural language input that provide both an empirical basis for assessing how innate constraints interact with information derived from the environment, and a source of hypotheses for experimental tests. Nevertheless, the obstacle remaining for a one-grammar trigger theory of parameter-setting is how it might include the often attested presence of mixed- paradigm input. As mentioned, mixing occurs with non-marginal structures not associated with the particular target grammar in the literature, even though they actually appear frequently within native speech patterns. Briefly, there are the well-know cases used to substantiate triggering models such as V2 in German; however, the verb-second rule applies not across the board, but instead excludes certain verbs. In English, V-raising applies only to Aux and not to lexical verbs, giving rise to two distinct but co-existing patterns; the Case and agreement system of Hindi-Urdu is split between ergative-absolutive and nominative-accusative patterns depending on a combination of aspectual, lexical, and structural factors (Saleemi 2002, 2003). The above arguments can be repeated, for many different languages and many different structures. To begin to address the conspicuous failure of many parametric accounts to accommodate mixed-paradigms found naturally across languages, it is necessary to subject the assumed properties of UG to greater scrutiny. As noted by Jackendoff (2002), perhaps the main snag lies in the conception of UG---both by its proponents and its detractors---in terms of a “grammar box cut off from the rest of the mind. This is clearly an oversimplification (cf. Jackendoff 2002: 79).” We therefore concur, with the spirit of Tomasello’s (1995) pronouncement that we should be judicious about how much linguistic structure is ascribed exclusively to the initial state UG. As Chomsky (1965: 207) concedes: “In fact, it is an important problem for psychology to determine to what extent other aspects of cognition share properties of language acquisition and language use, and to attempt, in this way, to develop a richer and more comprehensive theory of mind.” This said, we are by no means advocating general mechanisms alone to provide the kind of coverage to explain the unique architecture of language and the repertoire of rule types that governs it. In our view, any general mechanisms of cognition do not dispense with the specific ones; instead, they crucially presuppose them. When dealing with the countless complexities of linguistic structure, Skinner’s box cannot afford to be empty: it must at least contain a random assortment of various specific entities. In other words, if UG is to be the biologically-endowed basis from which L1 development commences, it had better be minimally equipped to supply the learner with a specific knowledge of fixed word order, grammatical functions, and Case systems, etc., just in case the child encounters these properties in his primary linguistic data. If instead of viewing UG as a single, unified “grammar box,” we follow Jackendoff’s (2002) lead to envisage it instead as a “toolbox,” then beyond the most universal “no-frills” minimum (perhaps this being syntax, the equivalent of Merged constituents), the learner must utilize many resources of his “toolbox.” In this light, languages cannot constrained to simply activating a ‘yes or no’ value of a parameter to obtain the maximal advantages of the grammar. They pick which apparatuses they use, more specialized or less-specialized tools, and to what degree in a flexible and adaptive manner, but only in case UG is not conceived of as an indivisible black box. Keeping this postulation in mind, consider the critical period in L1 acquisition as further support for the DUG Hypothesis. It has been successfully argued that not all aspects of language display critical period effects. Specifically, the acquisition of lexical items, the concatenation of words, and the nuts and bolts of semantically-based principles of word order seem immune to age-related “atrophy” of language acquisition (Hudson and Newport 1999). However, the capacity to acquire other parts of language in late language learning, such as 7 the inflectional system or real intricacies in phrase structure, appear to be largely diminished. As we claim, these common L1 versus L2 conditions become at once isolatable by viewing UG as a multi-layered construction with distinct learning mechanisms. III. ALTERNATIVE PROPOSAL We now put forward an alternative account as a preliminary attempt to avoid the bottlenecks of parameter theory, and to provide a more conceptually attractive story for the acquisition data to be discussed. The current analysis is drawn up with an eye to the non-parametric adaptive learning theory hypothesized in Culicover (1995). In our approach, the FL consists minimally of a Domain-general learning (DGL) device that serves as the interface between other brain systems that interface with FL and Domain- specific learning mechanisms (DSL) internal to the FL. Since we claim that maturational constraints also come to bear on the FL, it is reasonable to assume that no learning mechanism is fully operative in the initial state, nor are all mechanisms necessarily functioning in all knowledge states.5 We will also assume without further discussion that a mechanism is domain specific if it is used only to learn about the things within its domain. Thus, if a learning mechanism M were used just for learning about a grammatical property, it would be a domain-specific language learning mechanism. A learning mechanism is more domain neutral, the more things outside its domain it is used to learn about. Thus, if M were used both for learning about statistical regularities and language when its domain was just language, it would, to that extent, be domain general.6 5 Please note previous discussion (page 4) regarding Chomsky’s “No Dead Ends Assumption.” 6 Clearly, our views about what the domain of a learning process is must be revisable in the light of what is discovered about how that process actually works and interacts with other processes. 10 Along with the CLM, there is another DSL mechanism termed the Structural Learning Mechanism (SLM). SLM represents the moderately “specialized” tools in the FL toolbox, in terms of the systematic UG principles that are implemented in the structures, and translate mainly to the computational system in the FL. SLM mediates between the information in the CLM and the general structural descriptions mapped from the DGL. It may be the case that a sufficient threshold of lexical items must be accumulated in the CLM and thus, then when the GSDs are mapped to SLM, they can be re-analyzed for very specific syntactic operations such as movement operations in terms of wh-move, or scrambling, or V-to-I movement, etc. Once a GSD has been analyzed by the DSL mechanisms, it is again mapped to the DGL where it is maintained now as structural description, (SD) in the child’s grammar. Various data sequences of particular structures found in the linguistic environment can be stored by the DGL and organized into “like” pools of grammar which increasingly correspond to the target L1. As these pools of SDs are motivated by the classes of SDs characteristic of the target both in terms of systematic and “arbitrary” structures, over time those SDs of high frequency are stored as grammatical representations in the mind/brain of the L1 speaker. Given this formulation, the learner generalizes many aspects of the developing grammar in a gradual manner, moving from generic DGL to systematic SLM as linguistic evidence presented to him supports this step. It is important to recognize that while each FL learning mechanism is specialized, none of them can function independently of the other, and none is causally efficient (in the classical Aristotelian sense) on its own. For instance, SLM cannot carry out lexical category assignment which can only be treated by CLM; and without the output of the lexical projections provided by CLM, SLM cannot make a decision about how to analyze V-raising in this particular knowledge state. This interdependence of the UG learning mechanisms in the absence of parameters seems to be the right strategy. Many problems in language acquisition are difficult because no single feature of the input correlates with the relevant aspects of language structure. Although it is a natural starting point for computational and empirical research to study input cues in isolation, it may be that the problem of acquisition is easier when multiple information sources are taken into account. Figure 3 below shows abstractly how three constraints A, B and C, represented by regions of the hypothesis space, are insufficient to identify the correct hypothesis when considered in isolation. It is only by combining these input cues that the hypothesis space can be substantially narrowed down. Thus, as the quantity of cues that learner considers increases, the difficulty of the learning problem the observation that if all occurrences of word A can be replaced by word B, without loss of syntactic well-formedness, then they share the same syntactic category. For example, dog can be substituted freely for boy, in phrases such as the boy sat on the mat, nine out of ten boys prefer ..., indicating that these items have the same category. The distributional test is not a foolproof method of grouping words by their syntactic category, because distribution is a function of many factors other than syntactic category (e.g., word meaning). Thus, for example, boy and thunder might appear in very different contexts in some corpora, although they have the same word class. Nevertheless, it may be possible to exploit the general principle underlying the distributional test to obtain useful information about word classes. One common method records the contexts in which the words to be classified appear in a corpus of language, and then groups together words with similar distributions of contexts. 11 may decrease. This suggests that the cognitive system may aim to exploit as many sources as possible in FL: Figure 3. A conceptual illustration of 3 hypothesis spaces given the information provided by input A,B, and C (x’s correspond to hypotheses consistent with all 3 inputs) Moreover, it is possible that input cues only be useful when considered together. For example, in the sequences in Figure 4 below, each cue X and Y seems completely random with respect to the target Z; but when considered together X and Y determine Z perfectly (specifically, Z has value 1 exactly when just one of X and Y have the value 1). Considering input information in isolation implicitly assumes that there is a simple additive relation between cues: X: 1 0 0 1 1 1 0 1 0 1 1 0 0 0 1 1 0 1 0 1 0 1 Y: 0 1 1 0 1 0 0 1 1 0 1 1 0 0 0 0 1 1 1 0 0 1 Z: 1 1 1 1 0 1 0 0 1 1 0 1 0 0 1 1 1 0 1 1 0 0 Figure 4. A sequence of cues. IV. INFORMAL ANALYSIS Let us walk through a well-known example using some Null Subject facts. There is actually a striking difference between the types of languages that 12 are classified as so-called Null Subject Languages, since they are observed to license null subjects, according to the conditions of the standard NSP analyses. Languages such as Standard Italian, Castilian Spanish, and European Portuguese (EP) are held to represent the non-mixed paradigm. Neapolitan (Italian), Caribbean Spanish, and Brazilian Portuguese (BP) represent the mixed paradigm which can allow the same surface distribution and often same semantic functions to co-exist between overt and null subjects. Inventory of Pronominals 1) ENGLISH: S/he ate pizza: EUROPEAN PORTUGUESE (EP) BRAZILIAN PORTUGUESE (BP) Pro comeu pizza. = [+referential pro] *Pro comeu pizza = Topic delete 2) EP AND BP: Expletives and arbitrary subjects, same pattern for EP and BP a. Proarb Esquiam muito bem na Suíça = Theyimpersonal ski very well in Switzerland. ARBITRARY b. Ontem *ele/*aquele/pro fez muito frio. = It was very cold yesterday. QUASI ARGUMENT c. *Ele/*Aquele/ Pro é certo que (ele/ela/pro) fala bem = It is true that s/he speaks well EXPLETIVE 3) OPC FACTS (Montalbetti 1984: overt pronouns cannot link to formal variables) a. EP: Todo alunoi achou que proi/ele*i/j ia a passar de ano. b. BP: Todo alunoi achou que proi/elei? ia a passar de ano. (Negrão 1997) “Every student thought that he was going to pass.” c. EP: O Joãoi disse que *elei/ proi vai trazer uma garafa de vinho. d. BP: O Joãoi disse que elei/ proi vai trazer uma garafa de vinho. “João said that he is going to bring a bottle of wine.” e. EP: Alguns alunosi disseram que proi/*elesi acham que elesi são inteligentes. “Some students said that they think that they are intelligent.” f. BP: *Quemi acha que proi disse que elei é inteligente? (Negrão 1997) “Who thinks that (pro) said that he is intelligent? 15 seems a reasonable working assumption that, given the immense difficulty of the language acquisition problem, the cognitive system is likely to exploit such simple and useful sources of information, and worth exploring. Another possible project would be to determine whether certain surface configurations of syntactic development can appear to resemble very early parameter-setting (VEPS in the sense of Wexler 1998), due to intermittent, albeit observable, adult-like features which appear at certain points in the child’s speech production. However it may be the case that the child’s grammar may still not be in a final (steady) state in terms of the interplay of all the deeper derivative information necessary for the truly stable emergence of a given linguistic property. This is an interesting question to pursue. Perhaps the more intriguing aspect of this program is to determine how it would fare in the case of truly multilingual speakers, who must organize two or more distinct grammars from disparate inputs. Lastly, a problem with this account is that while domain specificity is well defined, domain generality comes in degrees. That is, we can say that a mechanism is domain specific, period, so long as it is only used for learning within its domain, that is, so long as it only does what it evolved to do. But we can only say that it is more or less domain general: a mechanism is the more domain general the more its operations generalize to other learning tasks. Now, this might not be a problem in principle – lots of perfectly good distinctions, after all, involve matters of degree. A computational simulation must be designed, since it will oblige us to make explicit such inner-workings of the model (and of the learner) that we are assuming. REFERENCES Belletti, A. & Rizzi, L., (eds.). (2002). Noam Chomsky: On Nature and Language. Cambridge: Cambridge University Press. Bertolo, S. (ed.). (2001). Language Acquisition and Learnability. Cambridge: Cambridge University Press. Bosch, L. & Sebastian-Galles, N. (1997). Native language recognition abilities in 4- month-old infants from monolingual and bilingual environments. Cognition 65, 33-69. Chomsky, N. (2000a). Minimalist inquiries: the framework. In R. Martin, D. Michaels, & J. Uriagereka, (eds.), Step by step-Essays in minimalist syntax in honor of Howard Lasnik. Cambridge, MA: MIT Press. Chomsky, N. (1995). The minimalist program. Cambridge, MA: MIT Press. Chomsky, N. & Lasnik, H. (1993). The theory of principles and parameter. In J. Jacobs, (ed.), Syntax: An International Handbook of Contemporary Research. Berlin: Walter de Gruyter. Chomsky, N. (1988). Language and problems of knowledge: The Managua Lectures. Cambridge, MA: MIT Press. Chomsky, N. (1986). Knowledge of language: its nature, origin, and use. New York: Praeger. Chomsky, N. (1981). Lectures on government and binding. Dordrecht: Foris. Chomsky, N. (1965). Aspects of the Theory of Syntax. Cambridge, MA: MIT Press. Clark, R. (1992). Selection of syntactic knowledge. Language Acquisition 2(2):83-149. Crain, S. & Wexler, K. (1999). A modular approach to methodology. In W. Ritchie & T. Bhatia (eds.), Handbook of Child Language Acquisition, pp. 387-425. London: Academic Press. Crain, S. & Thorton, R. (1998). Investigations in Universal Grammar. Cambridge, MA: MIT Press. Crocker, M. (1996). Computational Psycholinguistics: An Interdisciplinary Approach to the Study of Language. Dordrecht: Kluwer. 16 Culicover, P. (1995). Adaptive Learning and concrete minimalism, Ms. Ohio State University. de Boysson-Bardies, B. & Vihman, M. (1991). Adaptation to language: Evidence from babbling and first words in two languages. Language 67, 297-319. Felix, S. (1992). Language acquisition as a maturational process. In J. Weissenborn, H. Goodluck, & T. Roeper (eds.), Theoretical Issues in Language Acquisition: continuity and change in development, pp.25-52. Hillsdale, NJ: Lawrence Erlbaum. Fodor, J.D.. (1998). Unambiguous triggers.Linguistic Inquiry 29, 1-36. Gibson, E. & Wexler, K. (1994). Triggers. Linguistic Inquiry 25, 407-454. Hermon, G. (1990). Syntactic theory and language acquisition. Studies in the Linguistic Sciences 20:2, 139-163. Hudson, C. & Newport, E. (1999). Creolization: Could adults really have done it all? In A. Greenhill, A. Littlefield, & C. Tano, (eds.), Proceedings of the Boston University Conference on Language Development 23(1), 265-276. Somerville, MA: Cascadilla Press. Hallé, P.A. & de Boysson-Bardies, B.(1996). The format of representation of recognized words in infants’ early receptive lexicon. Infant Behavioral Development 19, 463-481. Hyams, N. (1986). Language acquisition and the theory of parameters. Dordrecht: Reidel. Jackendoff, R. (2002). Foundations of Language. Oxford: Oxford University Press. Jackendoff, R. (1997). The Architecture of the Language Faculty. Cambridge, MA: MIT Press. Jackendoff, R. (1992a). Languages of the mind. Cambridge, MA: MIT Press. Jenkins, L. Biolinguistics. Cambridge: Cambridge University Press. Jusczyk, P.W. (1997). The Discovery of Spoken Language. Cambridge, MA: MIT Press. Lasnik, H. (2000). Syntactic structures revisited. Cambridge, MA: MIT Press. Lust, B, S. Flynn, C. Foley, & Y. Chien. (1999). How Do We Know What Children Know? Problems and Advances in Establishing Scientific Methods for the Study of Language Acquisition and Linguistic Theory. In W. Ritchie & T. Bhatia (eds.), Handbook of Child Language Acquisition, pp. 427-456. London: Academic Press. Montalbetti, M. (1984). After binding: on the interpretation of pronouns. PhD thesis, MIT. Moon, C., Cooper, R.P., Fifeer, W.P. (1993). Two-day old infants prefer native language. Infant Behavioral Development. 16, 495-500. Negrão, E. (1997). Asymmetries in the distribution of overt pronouns and empty categories in Brazilian Portuguese. In J. Black & V. Montapanyane, (eds.), Current Issues in LInguistic Theory 140. Amsterdam: John Benjamins. Negrão, E. & Müller, A. (1996). As mudanças no sistema pronominal do Portugu s Brasileiro. DELTA 12(1):125-152. Newport, E. (1990). Maturational Constraints on Language Learning. Cognitive Science 14, 11-28. Newport, E. (1988). Constraints on learning and their role in language acquisition: Studies of the acquisition of American Sign Language. Language Sciences 10, 147-172. Pinker, S. (1995). Why the Child Holded the Baby Rabbits: A Case Study in Language Acquisition. In L. Gleitman & M. Liberman (eds.), An Invitation to Cognitive Science, Volume One: Language, pp. 199-241. Cambridge, MA: MIT Press. Poeppel, D. & Wexler, K. (1993). The Full Competence Hypothesis of clause structure in Early German. Language 69, 1-33. 17 Roeper, T. & Rohrbacher, B. (1995). Null subjects in early child English and the theory of Economy of Projection. Penn Working Papers in Linguistics, 2, 83-117. Saffran, J., Aslin, R., & Newport, E. (1996). Statistical Learning by 8-month-old Infants. Science 274 (13 December): 1926-1928. Sakas, W.G. (2000). Modeling the Effect of Cross-Language Ambiguity on Human Syntax Acquisition. In Proceedings of CoNLL-2000 and LLL-2000, pp. 61-66. Saleemi, A. (2002). Syntax Learnability: the problem that won't go away. In R. Singh, (ed.), The Yearbook of South Asian Languages and Linguistics Saleemi, A. (1992). Universal grammar and language learnability. Cambridge: Cambridge University Press. Satterfield, T. & Pérez-Bazán. (in preparation). The dynamic emergence of early bilingualism. Satterfield, T. (2003). Economy of interpretation: patterns of pronoun selection in transitional bilinguals. In V. Cook, (ed.), Effects of L1 on L2. Clevedon: Multilingual Matters. Satterfield, T. (1999). Bilingual Selection of Syntactic Knowledge: Extending the Principles and Parameters Approach. Dordrecht: Kluwer. Werker, J. F. & Tees, R. C. 1999. Influences on infant speech processing: toward a new synthesis. Annual Review of Psychology. 50, 509-535. Wexler, K. (1999). Maturation and Growth of Grammar. In W. Ritchie & T. Bhatia (eds.), Handbook of Child Language Acquisition, pp. 55-110. London: Academic Press. Wexler, K. (1998). Very early parameter setting and the unique checking constraint: A new explanation of the optional infinitive stage. Lingua 106, 23-79. Wexler, K. (1990a). Innateness and maturation in linguistic development. Developmental Psychobiology 23(3):645-660. Yang, C.D. (1999). A selectionist theory of language acquisition. In Proceedings of the 37th Annual Meeting of the Association for Computational Linguistics.