Charles Yang

Universal Grammar, statistics or both?

Recent demonstrations of statistical learning in infants have reinvigorated the innateness versus learning debate in language acquisition. This article addresses these issues from both computational and developmental perspectives. First, I argue that statistical learning using transitional probabilities cannot reliably segment words when scaled to a realistic setting (e.g. child-directed English). To be successful, it must be constrained by knowledge of phonological structure. Then, turning to the bona fide theory of innateness--the Principles and Parameters framework--I argue that a full explanation of childrens grammar development must abandon the domain-specific learning model of triggering, in favor of probabilistic learning mechanisms that might be domain-general but nevertheless operate in the domain-specific space of syntactic parameters.

Morphosyntactic Learning and the Development of Tense

In this article, we propose that the Root Infinitive (RI) phenomenon in child language is best viewed and explained as the interaction between morphological learning and syntactic development. We make the following specific sugges tions: The optionality in RI reflects the presence of a grammar such as Chinese which does not manifest tense marking. The gradual elimination of the non tense-marking grammar is facilitated by the learning of the morphosyntactic system of the target language. Quantitative differences in the input data among morphosyntactic systems result in the cross-linguistic variation in the RI phe nomenon. More broadly, we aim to demonstrate that quantitative aspects of lan guage learning data and concrete mechanisms of the language learning process can play an important role in the generative approach to language acquisition. In Section 2, we give a brief overview of the RI literature along with some methodological remarks regarding the explanation of the phenomenon. In Sec tion 3, we lay out our theory of morphosyntactic learning and the broader vari ational approach to language acquisition. Our empirical work focuses on the development of tense in Spanish, French, and English. In Section 4 we show through corpus study of child-directed speech that differences in the morphosyn tactic systems of these three languages explain the brief RI stage in Spanish acquisition, the prolonged RI stage in English acquisition, as well the inter mediate status of the RI stage in French. In Section 5, we discuss how our

The mystery of language evolution

Understanding the evolution of language requires evidence regarding origins and processes that led to change. In the last 40 years, there has been an explosion of research on this problem as well as a sense that considerable progress has been made. We argue instead that the richness of ideas is accompanied by a poverty of evidence, with essentially no explanation of how and why our linguistic computations and representations evolved. We show that, to date, (1) studies of nonhuman animals provide virtually no relevant parallels to human linguistic communication, and none to the underlying biological capacity; (2) the fossil and archaeological evidence does not inform our understanding of the computations and representations of our earliest ancestors, leaving details of origins and selective pressure unresolved; (3) our understanding of the genetics of language is so impoverished that there is little hope of connecting genes to linguistic processes any time soon; (4) all modeling attempts have made unfounded assumptions, and have provided no empirical tests, thus leaving any insights into languages origins unverifiable. Based on the current state of evidence, we submit that the most fundamental questions about the origins and evolution of our linguistic capacity remain as mysterious as ever, with considerable uncertainty about the discovery of either relevant or conclusive evidence that can adjudicate among the many open hypotheses. We conclude by presenting some suggestions about possible paths forward.

Internal and External Forces in Language Change.

If every productive form of linguistic expression can be described by some idealized human grammar, an individualss variable linguistic behavior (Weinreich, Labov, & Herzog, 1968) can be modeled as a statistical distribution of multiple idealized grammars. The distribution of grammars is determined by the interaction between the biological constraints on human grammar and the properties of linguistic data in the environment during the course of language acquisition. Such interaction can be formalized precisely and quantitatively in a mathematical model of language learning. Consequently, we model language change as the change in grammar distribution over time, which can be related to the statistical properties of historical linguistic data. As an empirical test, we apply the proposed model to explain the loss of the verb-second phenomenon in Old French and Old English based on corpus studies of historical texts.

Empirical re-assessment of stimulus poverty arguments

Abstract It is a fact that the child learner does not entertain logically possible but empirically impossible linguistic hypotheses, despite the absence of sufficient disconfirming evidence. While Pullum & Scholz claim to have shown the existence of disconfirming evidence, they fail to demonstrate its sufficiency. By situating the acquisition problem in a quantitative and comparative framework, we show that the evidence is, after all, insufficient. Hence the argument from the poverty of the stimulus, and the innateness of linguistic knowledge, stand unchallenged.

Estimation of software reliability by stratified sampling

A new approach to software reliability estimation is presented that combines operational testing with stratified sampling in order to reduce the number of program executions that must be checked manually for conformance to requirements. Automatic cluster analysis is applied to execution profiles in order to stratify captured operational executions. Experimental results are reported that suggest this approach can significantly reduce the cost of estimating reliability.

Ontogeny and phylogeny of language

How did language evolve? A popular approach points to the similarities between the ontogeny and phylogeny of language. Young children’s language and nonhuman primates’ signing both appear formulaic with limited syntactic combinations, thereby suggesting a degree of continuity in their cognitive abilities. To evaluate the validity of this approach, as well as to develop a quantitative benchmark to assess children’s language development, I propose a formal analysis that characterizes the statistical profile of grammatical rules. I show that very young children’s language is consistent with a productive grammar rather than memorization of specific word combinations from caregivers’ speech. Furthermore, I provide a statistically rigorous demonstration that the sign use of Nim Chimpsky, the chimpanzee who was taught American Sign Language, does not show the expected productivity of a rule-based grammar. Implications for theories of language acquisition and evolution are discussed.

Partition testing, stratified sampling, and cluster analysis

We present a new approach to reducing the manual labor required to estimate software reliability. It combines the ideas of partition testing methods with those of stratified sampling to reduce the sample size necessary to estimate reliability with a given degree of precision. Program executions are stratified by using automatic cluster analysis to group those with similar features. We describe the conditions under which stratification is effective for estimating software reliability, and we present preliminary experimental results suggesting that our approach may work well in practice.

The Great Number Crunch

A hard look in the mirror, as they say, is good for fitness and vitality. The time seems ripe, then, fifty years after the birth of modern linguistics, to reexamine its foundations. Or rather, the rubble, as the editors of Probabilistic linguistics suggest : corpus statistics, Markov chains, information theory, and the very notion of probability that were supposedly buried by the Chomskyan landslide.

Computational models of syntactic acquisition

The computational approach to syntactic acquisition can be fruitfully pursued by integrating results and perspectives from computer science, linguistics, and developmental psychology. In this article, we first review some key results in computational learning theory and their implications for language acquisition. We then turn to examine specific learning models, some of which exploit distributional information in the input while others rely on a constrained space of hypotheses, yet both approaches share a common set of characteristics to overcome the learning problem. We conclude with a discussion of how computational models connects with the empirical study of child grammar, making the case for computationally tractable, psychologically plausible and developmentally realistic models of acquisition. WIREs Cogn Sci 2012, 3:205-213. doi: 10.1002/wcs.1154 For further resources related to this article, please visit the WIREs website.