<?xml version="1.0" standalone="yes"?>
<Paper uid="W06-3812">
  <Title>Chinese Whispers an Efficient Graph Clustering Algorithm and its Application to Natural Language Processing Problems</Title>
  <Section position="7" start_page="77" end_page="79" type="concl">
    <SectionTitle>
5 Conclusion
</SectionTitle>
    <Paragraph position="0"> Chinese Whispers, an efficient graph-clustering algorithm was presented and described in theory and practice. Experiments with synthetic graphs showed that for small graphs, results can be inconclusive due to its non-deterministic nature.</Paragraph>
    <Paragraph position="1"> But while there exist plethora of clustering approaches that can deal well with small graphs, the power of CW lies in its capability of handling very large graphs in reasonable time. The  application field of CW rather lies in size regions, where other approaches' solutions are intractable. On the NLP data discussed, CW performs equally or better than other clustering algorithms. As CW - like other graph clustering algorithms chooses the number of classes on its own and can handle clusters of different sizes, it is especially suited for NLP problems, where class distributions are often highly skewed and the number of classes (e.g. in WSI) is not known beforehand.</Paragraph>
    <Paragraph position="2"> To relate the partitions, it is possible to set up a hierarchical version of CW in the following way: The nodes of equal class are joined to hyper-nodes. Edge weights between hyper-nodes are set according to the number of inter-class edges between the corresponding nodes. This results in flat hierarchies.</Paragraph>
    <Paragraph position="3"> In further works it is planned to apply CW to other graphs, such as the co-citation graph of Citeseer, the co-citation graph of web pages and the link structure of Wikipedia.</Paragraph>
  </Section>
class="xml-element"></Paper>