Off-campus UMass Amherst users: To download campus access dissertations, please use the following link to log into our proxy server with your UMass Amherst user name and password.
Non-UMass Amherst users: Please talk to your librarian about requesting this dissertation through interlibrary loan.
Dissertations that have an embargo placed on them will not be available to anyone until the embargo expires.
Author ORCID Identifier
Open Access Dissertation
Doctor of Philosophy (PhD)
Year Degree Awarded
Month Degree Awarded
Computational Linguistics | Phonetics and Phonology
This dissertation shows how a theory of grammatical representations and a theory of learning can be combined to generate gradient typological predictions in phonology, predicting not only which patterns are expected to exist, but also their relative frequencies: patterns which are learned more easily are predicted to be more typologically frequent than those which are more difficult.
In Chapter 1 I motivate and describe the specific implementation of this methodology in this dissertation. Maximum Entropy grammar (Goldwater & Johnson 2003) is combined with two agent-based learning models, the iterated and the interactive learning model, each of which mimics a type of learning dynamic observed in natural language acquisition.
In Chapter 2 I illustrate how this system works using a simplified, abstract example typology, and show how the models generate a bias away from patterns which rely on cumulative constraint interaction ("gang effects"), and a bias away from variable patterns. Both of these biases match observed trends in natural language typology and psycholinguistic experiments.
Chapter 3 further explores the models' bias away from cumulative constraint interaction using an empirical test case: the typology of possible patterns of contrast between two fricatives. This typology yields five possible patterns, the rarest of which is the result of a gang effect. The results of simulations performed with both models produce a bias against the gang effect pattern.
Chapter 4 further explores the models' bias away from variation using evidence from artificial grammar learning experiments, in which human participants show a bias away from variable patterns (e.g. Smith & Wonnacott 2010). This test case was chosen additionally to disambiguate between variable behavior within a lexical item (variation), and variable behavior across lexical items (exceptionality). The results of simulations performed with both learning models are consistent with the observed bias away from variable patterns in humans.
The results of the iterated and interactive learning models presented in this dissertation provide support for the use of this methodology in investigating the typological predictions of linguistic theories of grammar and learning, as well as in addressing broader questions regarding the source of gradient typological trends, and whether certain properties of natural language must be innately specified, or might emerge through other means.
Hughto, Coral, "Emergent Typological Effects of Agent-Based Learning Models in Maximum Entropy Grammar" (2020). Doctoral Dissertations. 2028.
Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.