<?xml version="1.0" standalone="yes"?>
<Paper uid="W06-3403">
  <Title>Computational Measures for Language Similarity across Time in Online Communities</Title>
  <Section position="2" start_page="0" end_page="0" type="intro">
    <SectionTitle>
1 Introduction
</SectionTitle>
    <Paragraph position="0"> While document similarity has been a concern in computational linguistics for some time, less attention has been paid to change in similarity across time. And yet, while historical linguists have long addressed the issue of divergence or convergence among language groups over long periods of time, there has also been increasing interest in convergence (also referred to as entrainment, speech accommodation, or alignment) in other areas of Linguistics, with the realization that we have little understanding of change in very short periods of time, such as months, in a particular conversational setting, between two people, or in a large group.</Paragraph>
    <Paragraph position="1"> The Internet provides an ideal opportunity to examine questions of this sort since all texts persevere for later analysis, and the diversity in kinds of online communities ensures that the influence of social behavior on language can be examined. Yet there has been very little work on language similarity in online communities.</Paragraph>
    <Paragraph position="2"> In this paper we compare the use of three separate tools to measure document or message similarity in a large data set from an online community of over 3,000 participants from 140 different countries. Based on a review of related work on corpus similarity measures and document comparison techniques (Section 2.2), we chose Spearman's Correlation Coefficient, a comparison algorithm that utilizes GZIP (which we will refer to as &amp;quot;Zipping&amp;quot;) and Latent Semantic Analysis. These three tools have all been shown effective for document comparison or corpus similarity, but never to our knowledge have any of them been used for document similarity over time, nor have they been compared to one another. Even though each of these tools is quite different in what it specifically measures and how it is used, and each has been used by quite different communities of researchers, they are all fairly well-understood (Section 4).</Paragraph>
  </Section>
class="xml-element"></Paper>