Modeling co-evolution of language and social groups in a dynamic environment

ly, a grammar is a function (a set of rules, constraints, or the like) that produces language by turning information drawn from the mental lexicon into the basic units that the articulatory system turns into a linguistic expression understandable to a hearer. In language comprehension, this process is inverted; the grammar turns information received from the perceptual system into units that represent knowledge. In our simulations a grammar can be something quite complex, like the rules that define the grammatical sentences of English, or something simpler, like the set of rules that govern the allowable co-occurrences of sounds in English words (e.g., words can start with the sound sequence ‘bl’, as in blue, but may not start with the sound sequence ‘bn’, as in the nonsense word *bnick). Our system will be able to encode theories that range from “categorical” models of grammar, in which every expression gets a binary evaluation as either grammatical or ungrammatical, to “soft” models of grammar that evaluate expressions probabilistically. The structures in our models can range from simple strings of sounds to complex things like syntactic trees indicating connections between verbs and their arguments or sentences paired with meaning representations. For our simulations, we will implement and test theories of grammar that include both those that are currently popular in academia and those that are used in industrial Natural Language Processing. In the domain of syntax, we will test two kinds of models: (i) models that represent the syntactic systems of various languages as settings of binary parameters (cf. Chomsky 1981, Manzini and Wexler 1987, Gibson and Wexler 1994) and (ii) models that encode the syntactic system of a language using sets of “features” specified for the items in the lexicon (cf. Chomsky 1995 and the computational implementation in Stabler 1997). These kinds of models have been studied extensively in theoretical linguistics and both have already been given computational implementations complete with learning algorithms. Furthermore, both systems can be represented as either categorical or soft models. In the domain of phonology—the study of the sound patterns permitted in a language—we will test three kinds of models: (i) models that represent phonology as sets of rules that map entries from the lexicon to words to be sent to the articulation system (cf. Chomsky and Halle 1968 and the computational implementation in Kaplan and Kay 1981, 1984); (ii) models that represent phonology as optimization over competing constraints (cf. Prince and Smolensky 1993 and the computational implementation in Riggle 2004); and (iii) models that represent phonology as sets of overlapping probability fields (cf. Pierrehumbert 2003 or Goldsmith and Riggle forthcoming). Though adapting the existing computational implementations of grammar models to fit into our system will be somewhat arduous, nothing truly novel needs to be created to make this work. Especially at the outset of our project, we will focus on implementing and testing existing models rather than creating new ones. For