<?xml version="1.0" standalone="yes"?>
<Paper uid="E06-1018">
  <Title>Word Sense Induction: Triplet-Based Clustering and Automatic Evaluation</Title>
  <Section position="3" start_page="0" end_page="137" type="intro">
    <SectionTitle>
2 Related work
</SectionTitle>
    <Paragraph position="0"> A substantial number of different approaches to WSI has been proposed so far. They are all based on co-occurrence statistics, albeit using different context representations such as co-occurrence of words within phrases (Pantel and Lin, 2002; Dorow and Widdows, 2003; Velldal, 2005), bi-grams (Sch&amp;quot;utze, 1998; Neill, 2002; Udani et al., 2005), small windows around a word (Gauch and Futrelle, 1993), or larger contexts such as sentences (Bordag, 2003; Rapp, 2004) or large windows of up to 20 words (Ferret, 2004). Moreover they all employ clustering methods to partition the co-occurring words into sets describing concepts or senses. Some algorithms aim for a global clustering of words into concepts (Yarowski, 1995; Pantel and Lin, 2002; Velldal, 2005). But the majority of algorithms are based on a local clustering: Words co-occurring with the target word are grouped into the various senses the target word has. It is not immediately clear which approach to favor, however aiming at global senses has the inherent property to produce a uniform granularity of distinctions between senses that might not be desired (Rapp, 2004).</Paragraph>
    <Paragraph position="1"> Graph-based algorithms differ from the majority of algorithms in several aspects. Words  can be taken as nodes and co-occurrence of two words defines an edge between the two respective nodes. Activation spreading on the resulting graph can be employed (Barth, 2004) in order to obtain most distinctly activated areas in the vicinity of the target word. It is also possible to use graph-based clustering techniques to obtain sense representations based on sub-graph density measures (Dorow and Widdows, 2003; Bordag, 2003).</Paragraph>
    <Paragraph position="2"> However, it is not yet clear, whether this kind of approach differs qualitatively from the standard clustering approaches. Generally though, the notion of sub-graph density seems to be more intuitive compared to the more abstract clustering.</Paragraph>
    <Paragraph position="3"> There are different types of polysemy, the most significant distinction probably being between syntactic classes of the word (e.g. to plant vs. a plant) and conceptually different senses (e.g.</Paragraph>
    <Paragraph position="4"> power plant vs. green plant). As known from work on unsupervised part-of-speech tagging (Rohwer and Freitag, 2004; Rapp, 2005), the size of the window in which words will be found similar to the target word plays a decisive role. Using most significant direct neighbours as context representations to compare words results in predominantly syntactical similarity to be found. On the other hand, using most significant sentence cooccurrencesresultsinmostlysemanticalsimilarity null (Curran, 2003). However, whereasvariouscontext representations, similarity measures and clustering methods have already been compared against each other (Purandare, 2004), there is no evidence so far, whether the various window sizes or other parameters have influence on the type of ambiguity found, see also (Manning and Sch&amp;quot;utze, 1999, p. 259).</Paragraph>
    <Paragraph position="5"> Pantel &amp; Lin (2002) introduced an evaluation method based on comparisons of the obtained word senses with senses provided in Word-Net. This method has been successfully used by other authors as well (Purandare, 2004; Ferret, 2004) because it is straightforward and produces intuitive numbers that help to directly estimate whether the output of a WSI algorithm is meaningful. On the other hand, any gold standard such  asWordNetisbiasedandhencealsolacksdomainspecific sense definitions while providing an abundance of sense definitions that occur too rarely in most corpora. For example in the British National Corpus (BNC), the sense #2 of MALE ([n] the capital of Maldives) from WordNet is represented by a single sentence only. Furthermore, comparing results of an algorithm to WordNet automatically implies another algorithm that matches the found senses with the senses in WordNet. This is very similar to the task of WSD and therefore can be assumed to be similarly error prone. These reasons have led some researchers to opt for a manual evaluation of their algorithms (Neill, 2002; Rapp, 2004; Udani et al., 2005). Manual evaluation, however, has its own disadvantages, most notably the poor reproducability of results. In this  workapseudowordbasedevaluationmethodsimilar to Sch&amp;quot;utze's (1992) pseudoword method is employed. It is automatic, easy to reproduce and adapts well to domain specificity of a given corpus. null</Paragraph>
  </Section>
class="xml-element"></Paper>