<?xml version="1.0" standalone="yes"?>
<Paper uid="W03-0605">
  <Title>An Architecture for Word Learning using Bidirectional Multimodal Structural Alignment</Title>
  <Section position="9" start_page="6" end_page="6" type="concl">
    <SectionTitle>
8 Contributions
</SectionTitle>
    <Paragraph position="0"> Effectively learning the meanings of words from non-linguistic input requires the development of representations and algorithms to determine correspondences between the linguistic and non-linguistic domains. Through this research, our contributions to this goal include: * We propose a general architecture, based on structural alignment, for employing linguistic and non-linguistic context in word learning. The system bootstraps itself by using acquired words to learn new words. We define the necessary properties of semantic representations used in such a system. We also define the modules this system will require.</Paragraph>
    <Paragraph position="1"> * We outline a system which implements this architecture for the specific semantic domain of vision. We identify LCS structures as an appropriate semantic representation, and we demonstrate techniques for extracting LCS from video. We also show a bidirectional approach to the parsing and alignment problem. null We currently have the components described in our implementation functional in isolation. The true merit of the system will be determined as we bring together all the pieces; thus our final contribution is the actual implementation of the systems described herein. It is our hope that our research will act as a springboard for the development of model refinement algorithms which have the advantage of support from semantic alignment systems such as ours.</Paragraph>
  </Section>
class="xml-element"></Paper>