<?xml version="1.0" standalone="yes"?>
<Paper uid="W06-1103">
  <Title>investigations</Title>
  <Section position="3" start_page="0" end_page="0" type="intro">
    <SectionTitle>
1 Introduction
</SectionTitle>
    <Paragraph position="0"> The task of word sense disambiguation has been at the heart of Natural Language Processing (NLP) for many years. Recent Senseval competitions (Mihalcea and Edmonds, 2004; Preiss and Yarowsky, 2001) have stimulated the development of algorithms to tackle different lexical disambiguation tasks. Such tasks require at their core a judgment of similarity as a word's multiple definitions and its contexts of occurrences are compared. Similarity judgment algorithms come in many different forms. One angle of this article is to analyze the assumptions behind such similarity metrics by looking at different shared or non-shared properties. Among the interesting properties we note symmetry and transitivity, which are fundamental to the understanding of similarity. This angle is investigated in Section 4 and 5, looking respectively at two broad classes of mathematical models of similarity and then more closely at different similarity metrics.</Paragraph>
    <Paragraph position="1"> As Senseval and other similar competitions need a gold standard for evaluating the competing systems, the second angle of our research looks into literature in philosophy and psychology to gain insight on the human capability in performing a similarity judgment. From the first discipline explored in Section 2, we discover that philosophers have divergent views on concept identification, ranging from scientific definitions to human perception of concepts. From the second discipline, explored in Section 3, we discover different psychological models for concept identification and implicitly concept comparison, this time ranging from continuous concepts being positioned in multi-dimensional spaces to concrete concepts being grasped as entities.</Paragraph>
    <Paragraph position="2"> The two angles (metrics and humans) converge in the conclusion of Section 6 with general observations and future work.</Paragraph>
  </Section>
class="xml-element"></Paper>