<?xml version="1.0" standalone="yes"?>
<Paper uid="W03-0613">
  <Title>Learning Word Meanings and Descriptive Parameter Spaces from Music</Title>
  <Section position="10" start_page="0" end_page="0" type="concl">
    <SectionTitle>
9 Conclusions
</SectionTitle>
    <Paragraph position="0"> We show that we can derive the most semantically significant description spaces automatically, and also form them into a knob for future classification, retrieval and even synthesis. Our next steps involve user studies of music description, an attempt to discover if the meaning derived by community metadata matches up with individual description, and a way to extract a user model from language to specify results based on prior experience.</Paragraph>
    <Paragraph position="1"> We are also currently working on new automatic lexical relation discovery techniques. For example, from the set of audio observations, we can infer antonymial relations without the use of an expert by finding optimally statistically separable observations. As well, meronymy, hyponymy and synonymy can be inferred by studying artificial combinations of observation (the mixture of 'loud' and 'peaceful' might not resolve but the mixture of 'sexy' and 'romantic' might.) From the perspective of computational linguistics, we see a rich area of future exploration at the boundary of perceptual computing and lexical semantics. We have drawn upon WordNet to strengthen our perceptual representations, but we believe the converse is also true. These experiments are a step towards grounding WordNet in machine perception.</Paragraph>
  </Section>
class="xml-element"></Paper>