Below you will find the schedule of speakers, along with more detailed information about their presentations. A powerpoint projector will be available in the room. All talks will take place in SSPA 2112 on UCI campus.
The Schedule
Friday, Sept 11
| 8:45am-9:00am | Coffee Fuel-Up |
| 9:00am-9:15am | Lisa Pearl & Jon Sprouse, University of California, Irvine |
| Welcome & Introduction | |
| 9:20am-10:15am | Anne Hsu, University of California, Berkeley |
| Quantitative analysis of the problem of no negative evidence in Language Acquisition | |
| Children learn what their language does not allow, apparently from observing only what is allowed. Yet natural languages are full of "restrictions"---i.e., patterns that apparently fit with general linguistic rules, but are ungrammatical. Without negative feedback, how do children learn not to overgeneralize and fill in these restrictions? This central puzzle of language acquisition has been viewed both as an argument for and against Universal Grammar: putative innate language-specific knowledge, guiding language acquisition. We reframe this problem of learning "restrictions" in language as a matter of probabilistic inference using the minimum description length principle. This provides a simple practical methodology for estimating how much linguistic data is required to learn a particular linguistic "restriction." We apply this method to a range of classic linguistic puzzles. We find that some linguistic rules appear easily learnable from language experience only while others appear learnable only with the help of innate constraints. | |
| Presentation slides: (pdf) (ppt) | |
| 10:20am-11:15am | Kamil Ud Deen, University of Hawai'i at Manoa |
| Binding Beyond the Input | |
| Thai is a language famous for supposedly violating Principle C of the Binding Theory. In this talk, I show that adult Thai speakers do not in fact violate Principle C, but that what counts as an R-expression in Thai differs from English, thus giving the illusion of Principle C violations. I then show that Thai children, despite input that appears to violate Principle C, behave like children acquiring languages like English - they adhere to Principle C in contexts in which adults violate it. This suggests that Principle C, and the expressions over which it applies, are predetermined, and as the Thai child matures towards adulthood, she must learn the idiosyncratic nature of nominals in her language. | |
| Presentation slides: (pdf) (ppt) | |
| 11:15am-11:30am | Coffee Break |
| 11:30am-12:25pm | Jeff Lidz, University of Maryland |
| Selective Statistical Learning | |
| While research in the acquisition of syntax has largely focused on the necessity of abstract representations and the poverty of the stimulus with respect to these representations, very little research has asked how learners use the input to identify these representations. I present several experiments illustrating the role of statistical learning in a selective theory of syntax acquisition. I show that statistical generalizations to be found in the input have consequences for morphosyntax that go beyond what can be inferred simply from the distributions. Hence, to the extent that learners use statistical information in learning syntax, they are doing so by comparing that information against the predictions of precise alternative syntactic representations. | |
| Presentation slides: (pdf) | |
| 12:30pm-2:00pm | Lunch Break |
| 2:00pm-2:55pm | Charles Yang, University of Pennsylvania |
| On Usage and Grammar | |
| We provide statistical evidence that contrary to popular views, early child grammar is not item based but fully productive. Moreover, early generativity seems necessary given the statistical distributions of natural language expressions and their combinations. | |
| Presentation slides: (pdf) | |
| 3:00pm-3:55pm | Julian Pine, University of Liverpool |
| Simulating the developmental pattern of finiteness marking in English, Dutch, German, French and Spanish using MOSAIC | |
| One of the challenges facing computational approaches to language development is to develop models whose behaviour can be directly compared with that of language-learning children. Another is to show that these models can be extended beyond English-speaking children to simulate the behaviour of children learning a variety of different languages. MOSAIC (Model of Syntax Acquisition in Children) is a computational model of language acquisition that attempts to meet these challenges by using exactly the same learning mechanism to simulate the behaviour of children learning several different languages. MOSAIC takes as input corpora of orthographically-transcribed child-directed speech and learns to produce as output ‘child-like’ utterances that become progressively longer as learning proceeds. As a result of these characteristics, MOSAIC can be used to generate corpora of utterances at different stages of development, and hence to model the behaviour of children learning different languages across a range of MLU values. extended abstract (pdf) | |
| Presentation slides: (pdf) | |
| 3:55pm-4:10pm | Coffee Break |
| 4:10pm-5:05pm | Terry Regier, University of Chicago |
| Language acquisition and the poverty of the stimulus | |
| A classic argument holds that certain aspects of language structure cannot be learned from the impoverished linguistic input that children receive - and that children therefore must rely in part on innate knowledge of language. I will argue for another possibility: that children succeed at learning language because of domain-general biases favoring simplicity and a close fit to the data. I will show that an idealized learner with such biases, given realistic child-directed speech, can acquire aspects of linguistic knowledge that have been held to be necessarily innate. This is joint work with Stephani Foraker, Naveen Khetarpal, Amy Perfors, and Joshua Tenenbaum. | |
| 5:10pm-6:05pm | William Sakas, City University of New York |
| Disambiguating Syntactic Triggers | |
|
Contrary to the trend toward statistical modeling of language acquisition, we present data from an artificial language domain which suggest that a deterministic approach to modeling syntactic parameter setting by children may be feasible after all. Deterministic learning systems need unambiguous input information. This may seem to conflict with the observation that there is a great deal of parametric ambiguity between natural languages. However, since ambiguity and unambiguity can co-exist, there is no problem for deterministic learning as long as there are sufficient unambiguous triggers for all natural language parameters, which can be recognized by the learning device. It is not practical to establish this for the full domain of natural languages, some of whose properties remain to be understood. Our research strategy is to estimate the incidence of unambiguity in natural language by searching for unambiguous triggers in a constructed domain of languages whose structural properties are fully and precisely specified. extended abstract (pdf) |
Saturday, Sept 12
| 10:00am-10:30am | Speaker Administrivia & Coffee Fuel-Up |
| 10:30am-12:30pm | General Discussion |