Date of Completion

7-22-2014

Embargo Period

7-22-2014

Keywords

noun-noun compounds, dynamical systems theory, recurrent neural networks, relation priming, garden path effects, local coherence effects, self-organization

Major Advisor

Whitney Tabor

Associate Advisor

Jay Rueckl

Associate Advisor

James S. Magnuson

Associate Advisor

William Snyder

Field of Study

Psychology

Degree

Doctor of Philosophy

Open Access

Open Access

Abstract

By noun-noun compound, we mean any combination of two nouns that native speakers of a language can understand. Native speakers can easily generate and understand novel, transparent compounds (e.g., mountain magazine), suggesting compositional processing. Relation-based theories of compound meaning (e.g., Levi, 1978) provide an explanation for this apparent productivity by assuming a set of semantic/thematic relations (e.g., ABOUT) that can bind two component nouns. Inspired by relation-based theories, we propose a self-organizing network model according to which (1) the meaning of a noun-noun compound M H (e.g., mountain magazine) corresponds to a tree-like constituent structure whose daughter nodes represent constituent meanings and whose mother node represents a relation R such as [R M H], (2) each of the mother node and the daughter nodes is represented as a vector in a similarity space (in which similar relations are placed close together) that consists of multiple units (each of which is not necessarily interpretable) associated with activation levels, (3) connection strengths between units are assumed to be learned from linguistic experience, and (4) a particular structure (e.g., [ABOUT mountain magazine]) is realized from the interactive activation between the two groups of constituent units via the relational units. The model integrates prior efforts to model processing difficulty (Gagné & Shoben, 1997) and relational similarity (Devereux & Costello, 2006) to provide a comprehensive understanding of compound representation and processing dynamics. Furthermore, the model is applied to opaque compounds (e.g., seahorse) that have idiosyncratic meanings by assuming that they are represented in the same space as transparent compounds. We describe a free card sorting study that suggests that the relation space is a similarity space and hierarchically organized. In Experiment 1, we demonstrate that two kinds of interpretation revision phenomena, known from the sentence processing literature—garden path and local coherence effects—both occur in compound processing, suggesting an interactive constraint-satisfaction process. In Experiment 2, we observed positive priming between two transparent compounds that instantiate similar relations. In Experiment 3, we observed negative priming between an opaque and a transparent compound (and vice versa) regardless of relational similarity. Results in Experiment 2 and 3 together suggest that processing difficulty is not simply correlated with distance in similarity space: processing dynamics on the space must be considered. We report two simulation studies to explain experimental data and discuss why complex dynamics is observed in the relation similarity space. We contrast our model with symbolic treatments of the relation between regular and exceptional morphology (Pinker, 1999) and other similarity space models (Devereux & Costello, 2006), arguing that the critical distinguishing property of the dynamical models is the inclusion of feedback dynamics.

Download

Share

COinS