Thomas G. Bever, University of Arizona

As neurological and genetic methodologies improve, it becomes even more important to isolate true linguistic universals from those universal properties of language that can be ascribed to other constraints and sources. We consider two kinds of universals closely related to neurological and genetic investigations of language (my presentation will include specific examples in most cases).

Such considerations give us perspective on today's searches for the correct (range of) neurological and genetic substrates for language. In some cases, we may find evidence that some language universals are truly caused by extra-linguistic structures and processes. At the same time, the possibility that extra linguistic factors (have) shape(d) certain language universals, does not mean that those universals are not genetically and neurologically coded in unique and distinct ways. Even more important, the existence of certain kinds of universals of language may carry over in the opposite direction and themselves account for many other properties of human cognition and behavior. The primary goal of my considerations is to create some intellectual options in the complexity of how we think about meeting the goal of better understanding the foundations of language and their relation to the human mind.

1. Brain localization
   1. It is generally claimed that the left hemisphere is the seat of essential linguistic knowledge, suggesting an intimate evolutionary relationshi between the existence of asymmetries and language. Yet, current imaging techniques suggest that many other brain structures are involved in language: and 35 years of research has solidified the claim that the left hemisphere of humans and other mammals is specialized for computationally complex behaviors in general.
   2. The neurological localizationist assumption in general, presupposes that there is a single form of "normal" representation for language, within the left hemisphere, at least for right handed people. Yet, several decades of research show a consistent differentiation in the organizational relation between lexical and syntactic knowledge, in right-handers with and without left handed family members. Since almost 50% of right handers have left handed family members, this difference is not trivial and shows that there is considerable and profound lability in the "normal" organization for linguistic knowledge. This in turn suggests that the individual maturational/ neurolophysiological foundations for language may not be rooted in localized brain structures, but rather may emerge in important ways from maturational dynamics.
2. Architectural and computational structures
   1. Language appears universally to have levels of representation organized such that more abstract levels cannot be induced from more concrete levels: additional (innate) knowledge is required. This is puzzling prima facie, since in many cases, the expressive power of known languages could be entirely captured by levels of representation that ARE successively inducible. Indeed, each level is characterized by overwhelmingly dominant generalizations that DO map onto more abstract levels reliably - e.g., semantic gender -> syntactic gender, surface canonical constructions (e.g., NVN in English) -> underlying thematic relations. The universal presence of such generalizations in syntax, often motivates eccentric organization of rules, which tend to guarantee that sentences have a consistent superficial form, regardless of their underlying thematic organization.

      Reliance on surface templates, with dominant relations to internal relations fit with current theories of language use and learning, which critically depend on some form of initial "scaffolding". It is also consistent with perceptual and motor behavior in many non-linguistic domains. But, why do they not fully account for relations between levels of representation in language? On our view, the need for deduction, or "displacement" between levels of representation is an essential part of human cognition: it underlies many psychodynamic structures, aesthetic processes, and problem solving in general. We have suggested that the role of this in language, is to turn language learning into a problem solving activity which is FUN. This offers an explanation to why children even bother to learn structures that go beyond the powerful generalizations they acquire initially.

   2. Upward ho! It appears to be a universal of syntactic movement that it is from more to less embedded parts of trees - this was implicit even in some transformations suggested by Harris, and has been consistent during the last 60 years in often very different versions of transformational grammar. Yet, the descriptive and expressive power of sentences in SOME kind of language would not require this, indeed might not require movement at all. One explanation may lie in the fact that neurologically, language has usurped brain structures that evolved over many millennia for vision. Thus, the existence of linguistic movement (and displacement in general) may rely on neurological mechanisms underlying the processing of actual movement.

      One consequence of this idea is that the direction of linguistic movement may be telling us something hitherto unknown about the perception of real movement - in particular, that it is easier going from more embedded to less embedded parts of a visual field, from figure to ground, rather than the reverse. (This might follow from the simple fact that figures move around IN their ground, by definition). We have subjected this prediction to various psychophysical tests, and it is indeed true!

      At the moment, such a demonstration counts only as a suggestive analogy: further research with humans in other domains, and potentially with other mammals is needed to show that movement is easiest to perceive going from figure to ground, and that this is evolutionarily prior to its expression in language.

Lila Gleitman, UPenn

Word learning seems to be an odd choice of topic in a conference about universals. After all, the most obvious non-universal aspect of language is in the mapping between word-meaning and sound: the concept 'cat' isn't uniformly rendered as /kat/. However, I will try to show that word learners exhibit unlearned biases in both the kinds of evidence they deem relevant to this mapping problem, and in how they organize and weight input evidence.

In this talk I will concentrate attention on three kinds of Hard Words: verbs in general, mental-content verbs (such as think), and perspective-verb pairs whose environmental contingencies are heavily overlapping if not identical (e.g,, chase and flee). These words are identified in the presence of cooperating cues from observational and linguistic structure. Two kinds of 'universal' are perhaps at work in the mapping procedure for words. The first has to do with unlearned biases in the kinds of evidence that a learner will use (Lidz & Gleitman, in press); the second has to do with how humans compare and weight evidential cues on the basis of experience (Trueswell & Gleitman, in press).

John A. Hawkins, USC & MPI EvAn

Comparison of a wide range of grammars reveals patterns. These patterns can suggest principles that underlie the variation. I have argued (Hawkins 1994, 2004) that a large number of patterns reflect efficiencies of various sorts in language use. The efficiency principles formulated make predictions for numerous grammatical phenomena and for corresponding patterns in performance in languages with variation, because on this view performance actually explains the grammatical variation: grammars have "fixed" or "conventionalized" the preferred structures of performance. A challenge for typology has always been to account for minority types and apparent exceptions and to try to link unusual properties to other phenomena in the relevant languages. But do general principles that hold for the majority patterns still apply to the puzzling cases? In this paper I explore patterns and puzzles in some universals of word order, relativization and morpho-syntax and I shall argue that principles of efficiency can explain both. Grammar-only approaches, by contrast, can often derive the general patterns from stipulated axioms, but this is a shallow form of explanation (what explains the axioms?) and these approaches do not predict, let alone explain, the puzzling phenomena.

Norbert Hornstein, Maryland

Modern generative grammar makes one very big claim; that humans are biologically built to acquire, use and understand natural language (NL). This biological construction is modular in the sense that it is task specific and functionally independent of other endowments Darwin and his minions have provided to us. We name this part of the mind/brain the Faculty of Language (FL). The main aim of linguistics so construed is to describe the fine structure of FL, figure out how it interacts with other cognitive modules, and, hopefully, figure out how it is incarnated in brain wet-ware.

This general picture has been considerably elaborated over the last 45 years. To date, most progress has come by considering the structure of FL against the backdrop of the logical problem of language acquisition. The problem, as all know by now, is that the a speakers attained linguistic capacity outstrips the information contained in the evidence that the child has access to during the acquisition period. This gap between the information in the input and the knowledge attained is bridged by innate properties of FL. One limns the specific features of FL by studying these sorts of inductive gaps and postulating properties of FL that given the kinds of input generally available would permit the child to acquire the relevant language. Now, several empirical boundary conditions characterize this process. First, any child can learn any language simply if exposed sufficiently to "bits" of the relevant language. Second, the information need not be systematically presented nor carefully monitored or supervised. Third, exposure suffices, instruction and training are irrelevant. Forth, the attained capacity vastly exceeds both qualitatively and quantitatively that of the input. These features together indicate that whatever properties FL has are general in the sense of being applicable to any NL and reflex like in kicking in when the appropriate input occurs. The forth property indicates that what is attained is in part at least a set of recursive rules able to generate a practically infinite set of expressions of any given NL. Thus, at least one feature of FL is that it can take well formed, relatively simple linguistic inputs and from this create a set of rules (a grammar) of the relevant language. In other words, part of FL is a Universal Grammar, a recipe for building language specific grammars for any language based on "bits" of that language.

This, to repeat, is all conventional wisdom. Universals on this conception are simply those features of FL that underlie this process. They are, in other words, the principles that describe the structure and functioning of FL. So, speaking most properly, Universals are the principles of Universal Grammar which are just the design features of FL.

Note, on this conception universals are quite abstract. They need not be observable even were one to survey thousands of languages looking for commonalities. In fact, on this conception, the mere fact that every language displayed some property P does not imply that P is a universal of NL. Put more paradoxically, the fact that P holds universally does not imply that P is a universal. Conversely, some property can be a universal even if only manifested in a single NL. The only thing that makes a principle a Universal on this view is that it is a property of FL. Universals so conceived are the laws of FL and the aim of the study of language is to uncover these laws.

How is this done? The short answer is any way one can. The longer answer involves identifying the research programs devised to advance this end. My paper will be an attempt to survey three such research programs in the history of modern linguistics. We will concentrate on the two most recent, the GB program theory and the more recent Minimalist Program.

James R. Hurford, Edinburgh

D'Arcy Thomson's classic work On Growth and Form identified a kind of process, distinct from natural selection, contributing to the symmetries and patterns observable in nature. Examples are the the hexagonal cells in honeycombs, and the spiral shapes of snail shells. A well-known quotation from Chomsky is "It could be that when the brain reached a certain level of complexity it simply automatically had certain properties because that's what happens when you pack 10^10 neurons into something the size of a basketball." This has been interpreted as an appeal to "as-yet unknown laws of growth and form" by Pinker and Bloom. Both D'Arcy Thomson and Chomsky generally take laws of growth and form to provide an alternative kind of explanation to any kind of selection. It now seems clear that selection and growth-and-form are not contrasting explanations for patterned phenomena, but complementary. Laws of growth and form mathematically determine the overall topology of the landscape with which selection works.

Both D'Arcy Thomson and Chomsky appeal to processes of GROWTH WITHIN SINGLE INDIVIDUALS to explain their eventual patterned forms. In this paper, I will explore the extent to which another process, CULTURAL TRANSMISSION, contributes to the patterned form of its products. Languages, while existing within an envelope constrained biologically by what is acquirable by individuals, are also cultural products constructed over many generations by a cycle of iterated cultural transmission. In the course of this cycle, languages exist in two life-phases, as I-Language (representations in individual minds) and as E-Language (public behaviour). In practice, what linguists describe is guided both by intuitive judgements about internal representations and by observation of public behaviour.

An individual's I-Language defines an infinite set of forms. This internal representation is acquired on the basis of a necessarily finite set of examples. The internalized grammar, moreover, generalizes powerfully over the observed examples, extrapolating well beyond the finite experience that triggers it. Thus there is, every generation, a repeated process of data-compression, and re-production of a new wave of public expressions from the compressed representations. Given this repeated cycle of induction and selective expression, certain patterns naturally emerge, by what one can call "Laws of Transmission and Form". Many such patterns can be observed in all languages, and all can be observed in many languages, thus deserving the label "linguistic universal". Examples include: the typical patterning of vowel systems; the symmetry found in consonant systems; the compositionality of syntax; and Zipfian frequency laws. The existence of such historical processes eases the explanatory burden on the innate human Language Acquisition Device. To be sure, one still needs to postulate some innate specific predispositions in language learners, but these can be relatively weak, compared to the power which must be attributed to a device which has only one generation in which to make its impact (the Chomskyan LAD). In the Chomskyan mode of explanation for linguistic universals, the set of possible languages is the set of acquirable languages. In the present extended view, we identify HISTORICALLY POSSIBLE languages, that is languages standing some way along in a cultural tradition; these are a proper subset of the languages acquirable (given suitable data) by an individual.

Background reading:

Paul Kiparsky, Stanford

A classic project of generative grammar is to ground the properties of language change in principles of UG. Recent examples from syntax are the explanations for the constant rate effect (Kroch 1989) and for the unidirectionality of grammaticalization (Roberts and Roussou 2003).

An older research program, increasingly popular again, pursues the reverse direction of explanation, by seeking the causes of typological generalizations in recurrent historical processes (Anderson 1988, Garrett 1990, Aristar 1991, Givon 1994).

I argue that the two programs can coexist without contradiction or circularity. This requires making a principled distinction between true universals, which constrain both synchronic grammars and language change, and typological generalizations, which are the results of typical paths of change. The following criteria are relevant. (1) Universals have no exceptions (for what does not arise by change cannot be subverted by it either). (2) Universals are process-independent. (3) Universals can be manifested in "emergence of the unmarked" effects. (4) Universals are embedded in grammars as constraints and can interact with other grammatical constraints. (5) Universals determine pathways for analogical change.

A review of anaphora and ergative case assignment suggests that these criteria converge rather cleanly. The core constraints of binding theory and the hierarchy of binding domains are universals (Kiparsky 2002); Talmy's and Pica's generalization about the binding properties of simple vs. complex anaphors and Everaert's generalization that there are no nominative anaphors are not universals, but diachronically explicable typological generalizations. The Animacy Hierarchy that governs split case assignment, number marking, and agreement is a universal; tense-based split ergativity is an epiphenomenon of diachrony.

Chris Manning, Stanford University

Today it may appear as if linguistic theory and computational linguistics have never been further apart. But I think we are just around the corner from them coming together again, as they have twice in the past: with the formalization of classes of languages by Chomsky in the late 1950s, and in the 1980s era of declarative 'unification-based' grammars.

They need to come together, because the central problems in linguistics are computational problems: what mechanisms are needed to process languages in real time, what intermediate representations allow human beings to transduce between surface linguistic forms and their contextual meanings, and what mechanisms enable children to aquire languages, apparently quite effortlessly.

Language understanding involves problems of evidence combination and reasoning from uncertain and incomplete inputs. Language learning involves abstracting a productive language system from noisy, uncertain, and uncontrolled input. In the 1950s when the current linguistic orthodoxy developed, almost nothing was known about how to treat these problems computationally, and a claim that it was impossible to solve these problems using the input data alone was perhaps not unreasonable. In the intervening fifty years, the fields of reasoning under uncertainty and machine learning have blossomed, and we now have a good understanding of how both these operations can be performed over complex and large representations and datasets, of the sort found in linguistics. Central to this has been the development of large scale applications of probabilistic models: Not only is probability theory in some sense the right way to resolve ambiguities and to learn, but there is solid evidence that humans use statistical methods to learn and resolve ambiguities in vision, concept learning, causal reasoning, and language. I will look in particular at computational work on language learning and language processing, and show how computational work is addressing in a concrete way the logical problem of language acquisition and issues of human language understanding, and examine the connections to linguistic questions.

The word 'formal' in my title is unfortunately very ambiguous; I do not mean it in the sense of form opposing function, but rather in the sense of being formalized with precise models. The natural venue for linguistics and computational linguistics to link up again is in the areas of linguistics where people are dealing with variable, empirical data: sociolinguistics, corpus linguistics, language acquisition, psycholinguistics and typological linguistics. In all these areas, functional concerns are important, and predictive formal probabilistic models can be usefully used. A new formalized linguistics will develop out of building complex models that work over noisy, empirical data, whereas, I suspect that a considerable amount of the 'formal' generative linguistics done in recent decades will be seen in retrospect as rather like drawing epicycles.

Ralph-Axel Müller, San Diego State University

Edward Stabler, UCLA

Already in 1961 it was clear that the fundamental properties of human languages are not independent; many of the logically independent properties are not empirically independent. Some of the most interesting hypotheses about language universals aim to explain this fact.

1. The fact that we see patterns in a small sample of sentences, patterns that extend well beyond the sample, has the consequence that many languages cannot be learned (as we see, in different ways, in the formal results of Gold, Angluin, Blumer et al, for example), but this same fact allows us to understand many sentences never encountered before.
2. Many different kinds of grammars have been proposed to characterize the kinds of patterns we seem to find in language: transformational grammars, categorial grammars, tree adjoining grammars, head-driven phrase structure grammars, and even tupled pregroup grammars, among many others. In this great and fractious diversity, careful study in the past couple of decades has revealed an astounding convergence: very roughly, (formalized versions of) all the mentioned frameworks have been shown to be expressively equivalent or closely related (Joshi, Weir, Vijay-Shanker, Michaelis, Harkema, Stabler). Considering a more precise statement of these results, we find, for example, that certain kinds of categorial grammars, tree adjoining grammars, and `minimalist' transformational grammars not only define exactly the same languages, but they define them so similarly that it is very easy to translate grammars from one framework into any of the others, for reasons that are fairly easy to state. We argue that some of the most fundamental universal properties of human languages, including some that seemed quite surprising in early studies, are shared with all the grammars in all of these frameworks.

3. The particular languages we know are learned, but when we characterize the structure of these languages and consider the entire class of languages with fundamentally similar structures, we can consider the more abstract question of whether that entire class is learnable. Do the structural properties of human language guarantee that they are learnable? Interpreting `structural' in something like the traditional sense, the answer to this question must be: no. Nevertheless we can see that some fundamental properties of human languages are consequences of the fact that they are learned.

Michael K. Tanenhaus, University of Rochester

Most psycholinguists would probably agree that there are at least three classes of constraints that might have shapes the form of languages. These are constraints arising from the demands of acquisition, i.e., considerations of learnability, and the demands of language processing, in particular demands on production and comprehension. I'll focus on constraints that might arise from language processing, in particular the need for hierarchical structure in production, and the need for a separation of form and meaning in comprehension and learning. I'll also suggest that the study of face-to face conversation, arguably the most basic universal site of language use might (a) provide insights into why comprehension and production are so incremental, and (b) help shed light on some classic concerns about learnability. In particular, the demands of interactive conversation offer a possible explanation for why language might need to be produced and understood incrementally. Moreover, studies of interactive conversation suggest that the link between eye movements and utterance generation might provide important information that could guide learning of grammatical constraints.