MoL-2010-16: Exploiting Systematicity: a Connectionist Model of Bootstrapping in Language Acquisition

MoL-2010-16: Tourigny, Hélène (2010) Exploiting Systematicity: a Connectionist Model of Bootstrapping in Language Acquisition. [Report]

[thumbnail of Full Text]
Preview
Text (Full Text)
MoL-2010-16.text.pdf

Download (672kB) | Preview
[thumbnail of Abstract] Text (Abstract)
MoL-2010-16.abstract.txt

Download (1kB)

Abstract

This thesis presents a connectionist model of syntactic bootstrapping
processes in language acquisition. According to the Syntactic
Bootstrapping hypothesis, children acquiring language can learn (part
of) the meaning of new words based on the syntactic context in which
they appear. Psycholinguistic research has shown that children can
indeed use morphosyntactic cues to guide their interpretation of novel
words.
This project investigates whether a connectionist network can exploit
systematicity in language to acquire novel words over the course of
development. The network is trained according to a semi-supervised
algorithm. The model learns a sentence-interpretation task from both
labelled and unlabelled data. Specifically, it learns to output a
semantic representation (roughly corresponding to ‘who did what to
whom’) for given sentences. To investigate whether syntactic
bootstrapping can successfully lead to lexical development, some
vocabulary items are only presented in unlabelled sentences. To
correctly process these examples and learn these words, the network
must infer the novel words’ properties (e.g. grammatical category,
animacy features) based on the context in which they appear. The
network must then use its own output (i.e. its interpretation of the
sentences) to train itself.
The system’s ability to rely on syntactic cues for vocabulary
acquisition is tested in a number of experiments. Although the network
is able to acquire the language and shows very good generalization,
its ability to rely on syntactic bootstrapping to learn novel words
does not meet expectations.

Item Type: Report
Report Nr: MoL-2010-16
Series Name: Master of Logic Thesis (MoL) Series
Year: 2010
Date Deposited: 12 Oct 2016 14:38
Last Modified: 12 Oct 2016 14:38
URI: https://eprints.illc.uva.nl/id/eprint/841

Actions (login required)

View Item View Item