In this talk, I address the question of how general cognitive capabilities help shape what human languages look like. I do so by showing that the acquisition of certain aspects of language depends on specific types of memory using both behavioral and neuroimaging data. In particular, I demonstrate that rules and analogy are supported by procedural memory (memory for sequences) and declarative memory (memory for facts), respectively. This, combined with data from my colleagues on tone, suggests that some people may be better at acquiring certain types of languages depending on non-linguistic cognitive skills, neuroanatomical structure and, ultimately, genetics. I outline how we are currently exploring some of these questions as well as what alternatives there are to genetic determinism. Of interest to computational researchers, I will also present the early stages of a project exploring how these general cognitive biases may exert pressure on language via iterated learning.

If we conceptualize a theory of human sentence comprehension as a combination of (1) a grammar (2) a strategy for using the rules of the grammar and (3) some architectural facilities like memory we still have a huge space of possible theories. It would be nice to narrow this class down to just those that somehow made sense in relation to the communicative function sentence-comprehension often serves.

This talk examines a smaller class of comprehension theories that strive to finish parsing as soon as possible. These theories would be ``rational'' on a view of the comprehender as doing his or her best to understand what the speaker means. I shall argue that they correctly derive well-known garden pathing phenomena along with the puzzling Local Coherence effects studied by Tabor, Galantuccia and Richardson (2004). Time permitting, I will discuss the relationship between this class of theories and the Entropy Reduction Hypothesis revived in Hale (2006).

Tabor, W., Galantuccia, B., & Richardson, D. (2004). Effects of merely local syntactic coherence on sentence processing. Journal of Memory and Language, 50 (4), 355-370.

Hale, J. (2006). Uncertainty about the rest of the sentence. Cognitive Science, 30 (4), 609-642.

I will argue that listeners modify the way they process language according to the circumstances, and concentrate on the case of processing non-native speech. I will describe the results of an eye-tracking study that demonstrates that listeners rely on contextual cues to a greater degree, even at the expense of the speech input, when they listen to non-native speakers. I will also show findings that suggest that this ability to modify the way one processes language depends on one's Working Memory Capacity.

I will present ongoing work in close collaboration with Morgan Sonderegger (Computer Science, U Chicago) and Peter Graff (Linguistics, MIT) on the dynamics of sociophonetic variation among the contestants of "Big Brother", a British reality-TV program. The premise of the show is that sixteen contestants, or "housemates", of diverse linguistic background must live together in one house for 13 weeks under 24-hour audio/video surveillance (continually broadcast to viewers), completely isolated from the outside world; each week, the housemates vote to eliminate one of their number from the house, the object being to remain in the house for all 13 weeks, and to finally be selected as "winner" by a vote of viewers. Because of the continual surveillance and total linguistic isolation of the participants, Big Brother offers a unique natural experiment for sociophoneticians.

I will discuss two kinds of preliminary data that we have collected: measurements of the degree of social interaction between pairs of housemates over time, and measurements of several phonetic variables taken at regular intervals from a subset of the housemates. I will describe our tentative solutions to some of the methodological and statistical difficulties inherent in working with a large, spontaneous sociophonetic corpus of this nature, as well as the preliminary results and conclusions we have gleaned so far, including:

• a novel method of quantifying social affinities between individuals;
• the finding that some housemates exhibit significant phonetic change over the 13 weeks, while others do not;
• a suggestion that some changes may correlate with socially significant events.

A large body of sound structure research suggests that highly modular frameworks--which assume a strict separation between discrete phonological and continuous phonetic representations--cannot account for the interdependence of gradient variation and phonological structure. This has lead to the development of novel theoretical frameworks with non-symbolic primitives (e.g., dynamical systems, exemplars). In this work we seek to develop an alternative approach that incorporates gradience while allowing for a formal integration with symbolic theories of sound structure.

Building on work in connectionist theory (Smolensky and Legendre, 2006), we embed symbolic structures and symbolic well-formedness constraints within a continuous representational space. Numerical simulations show that this framework can capture gradient speech production data. Additional analytical results and simulations reveal important implications for grammatical theory as well. Incorporating gradient structure allows for languages that cannot be specified within standard Optimality Theoretic or Harmonic Grammar frameworks.

Joint work with Don Mathis and Paul Smolensky, Johns Hopkins University. Supported by NSF BCS 0846147.

The phonologization model (Hyman, 1976) conceptualizes a sound change such as the emergence of phonemic tone as a three-stage process. In the first stage, a language displays different degrees of automatic, physiologically-based pitch perturbations on vowels following voiced and voiceless consonants. In the second stage, these perturbations have become exaggerated to such a degree that they are no longer predictable based on the voicing of the preceding consonant, and are said to have been phonologized. In the third stage, the voicing contrast has been neutralized, leaving only the pitch differences on the vowel to indicate the phonemic contrast, and the language is said to have undergone phonemicization of tone.

While this model has proven conceptually useful, certain details remain critically underspecified. In particular, the model fails to provide an account of why one cue to a contrast should be targeted for phonologization and not another, or why the phonologization of a secondary cue appears to trigger the neutralization of a primary cue. In this talk, I present an account of phonologization as subphonemic restructuring by modeling contrasts as Gaussian mixtures (Nearey & Hogan, 1986). I argue that phonologization is the result of an adaptive enhancement strategy, driven by loss of contrast precision, that optimizes both listener- and speaker-oriented constraints (Lindblom et al., 1995). Subphonemic cues are targeted and enhanced in a probabilistic fashion, proportional to their informativeness and the precision of the contrast they help to cue. I will present the details of this model and show how it can account for the ongoing phonologization of pitch in Seoul Korean (Kang & Guion, 2008).

Kirby, J. (To appear). The role of probabilistic enhancement in phonologization. In A. Yu (ed.), Origins of sound patterns: approaches to phonologization. Oxford: OUP.

Statistical learning for language has been extensively studied using the artificial language learning paradigm (Gomez, 2002; Thompson & Newport, 2007; Wonnacott, Newport & Tanenhaus, 2008). Implicit in these studies is the learning of the appropriate 'units' on which to keep statistics. In my talk I will present a unified computational model in which both the grammatical units and usage statistics are learned simultaneously from naturalistic, contextually-grounded Mandarin Chinese input. Using Construction Grammar, this model learns argument structure constructions in a language with argument omission. Statistics are gathered by the model as learned constructions are used to interpret new utterances. I will demonstrate how grammatical structures and statistics increase the combinatorial power of the learned grammar when the model generalizes across syntactically and semantically aligned constructions.

This work was carried out as part of the dissertation work while the author was at the University of California, Berkeley and the International Computer Science Institute.