Constructing grammar:
A computational model of the emergence of early constructions
This thesis explores and formalizes the view that grammar learning is
driven by meaningful language use in context. We describe a
usage-based computational model of how children acquire their earliest
multi-unit constructions, or relational form-meaning mappings. The
model incorporates key features of child language learning, including
rich conceptual structure and pragmatic abilities. We define a
construction-based grammar formalism to serve as the hypothesis space,
a model of how such constructions can be used to interpret utterances
in context, and a model of how such constructions can be learned from
a sequence of utterance-situation pairs. Learning biases are designed
to facilitate successful communication: new constructions are directly
motivated by the language understanding process, either to capture
relational mappings present in context but not predicted by the
current grammar, or to reorganize the existing set of
constructions. Candidate constructions are evaluated using a minimum
description length criterion that balances a prior structural bias
toward simpler grammars against a data-driven bias toward simpler
analyses of the data. When applied to input corpora based on annotated
child-directed language, the model exhibits learning behavior
compatible with crosslinguistic developmental patterns. It thus offers
a cognitively motivated and computationally precise account of how
diverse constraints can be integrated to account for the transition
from single words to complex utterances.