<?xml version="1.0" standalone="yes"?>
<Paper uid="W04-3221">
  <Title>Attribute-Based and Value-Based Clustering: An Evaluation</Title>
  <Section position="4" start_page="3" end_page="4" type="metho">
    <SectionTitle>
3 First Experiment: Using a Set of Concepts
</SectionTitle>
    <Paragraph position="0"> from Lund and Burgess One limitation of using Google is that even with an increased daily limit of 20,000, it wouldn't really be feasible to attempt to cluster, say, all of WordNet 100,000 noun concepts. For this reason, we used much smaller sets of concepts in our two experiments. The first set allowed us to compare our results with those obtained by Lund and Burgess (1996); the second set consisted of a larger number of concepts from WordNet.</Paragraph>
    <Paragraph position="1"> Lund and Burgess (1996) used a set of 34 concepts belonging to 3 different classes (animals, body parts, and geographical locations) to evaluate their method for acquiring lexical representations, HAL (Hyperspace Analogue to Language). Lund and Burgess were able to correctly cluster all of the concepts except for one body part, tooth, which was incorrectly clustered with animals. In this first experiment, we used the 34 Lund and Burgess concepts plus Italy, horse, and tongue (37 in total) to compare value-based and attribute-based description when used for clustering, using concept descriptions collected using the methods described above.</Paragraph>
    <Paragraph position="2"> The input to clustering is a frequency table with concepts as rows and values, attributes, or both attributes and values as columns. Each cell in the table contains the frequency of co-occurrence between the concept and corresponding value or attribute. Before clustering, the frequencies are transformed into weighted values using the t test (Manning and Schutze, 1999). (The t test was found by Curran and Moens (2002) to be the best weighting method.) The t test formula we used for attributes is shown below:</Paragraph>
    <Paragraph position="4"> where N is the total number of relations, and C is a count function. The values formula is similar.</Paragraph>
    <Paragraph position="5"> We use the CLUTO vcluster command for clustering, with parameters: similarity function =  Here, we choose the top k features by their overall frequency.</Paragraph>
    <Paragraph position="6"> Table 4 shows the accuracy of the produced clusters when using values, attributes, and the combination with different vector sizes. The results show that with concept descriptions of length 500, attributes (97.30%) are much more accurate than values (64.86%). With vectors of size 1522, the accuracy with attributes remains the same, while the accuracy with values improves, but is still lower than the accuracy with attributes (94.59%). This indicates that attributes have more discriminatory power than values: an attribute vector of size 500 is sufficient to produce a more accurate result than using a value vector of three times the size. But perhaps the most interesting result is that even though further increasing the size of pure attribute- and value- descriptions (to 4753 and 4969, respectively) does not improve accuracy, perfect accuracy can be obtained by using vectors of length 3044, including the 1522 best attributes and the 1522 best values. This suggests that while attributes are a good way of generalizing across properties, not all properties of concepts can be viewed as attribute/value pairs (section 5; also (Poesio and Almuhareb, submitted)).</Paragraph>
  </Section>
class="xml-element"></Paper>