<?xml version="1.0" standalone="yes"?>
<Paper uid="W06-3809">
  <Title>Random-Walk Term Weighting for Improved Text Classification</Title>
  <Section position="2" start_page="0" end_page="53" type="intro">
    <SectionTitle>
1 Introduction
</SectionTitle>
    <Paragraph position="0"> Term frequency has long been adapted as a measure of term significance in a specific context (Robertson and Jones, 1997). The logic behind it is that the more a certain term is encountered in a certain context, the more it carries or contributes to the meaning of the context. Due to this belief, term frequency has been a major factor in estimating the probabilistic distribution of features using maximum likelihood estimates and hence has been incorporated in a broad spectrum of tasks ranging from feature selection techniques (Yang and Pedersen, 1997; Schutze et al., 1995) to language models (Bahl et al., 1983).</Paragraph>
    <Paragraph position="1"> In this paper we introduce a new measure of term weighting, which integrates the locality of a term and its relation to the surrounding context. We model this local contribution using a co-occurrence relation in which terms that co-occur in a certain context are likely to share between them some of their importance (or significance). Note that in this model the relation between a given term and its context is not linear, since the context itself consists of a collection of other terms, which in turn have a dependency relation with their own context, which might include the original given term. In order to model this recursive relation we use a graph-based ranking algorithm, namely the PageRank random-walk algorithms (Brin and Page, 1998), and its TextRank adaption to text processing applications (Mihalcea and Tarau, 2004). TextRank takes as input a set of textual entities and relations between them, and uses a graph-based ranking algorithm (also known as random walk algorithm) to produce a set of scores that represent the accumulated weight or rank for each textual entity in their context. The TextRank model was so far evaluated on three natural language processing tasks: document summarization, word sense disambiguation, and keyword extraction, and despite being fully unsupervised, it has been shown to be competitive with other sometime supervised state-of-the-art algorithms.</Paragraph>
    <Paragraph position="2"> In this paper, we show how TextRank can be used to model the probabilistic distribution of word features in a document, by making further use of the scores produced by the random-walk model.</Paragraph>
    <Paragraph position="3">  Through experiments performed on a text classification task, we show that these random walk scores outperform the traditional term frequencies typically used to model the feature weights for this task.</Paragraph>
  </Section>
class="xml-element"></Paper>