<?xml version="1.0" standalone="yes"?>
<Paper uid="W06-1669">
  <Title>Sydney, July 2006. c(c)2006 Association for Computational Linguistics Two graph-based algorithms for state-of-the-art WSD</Title>
  <Section position="8" start_page="591" end_page="591" type="concl">
    <SectionTitle>
6 Conclusions and further work
</SectionTitle>
    <Paragraph position="0"> This paper has explored the use of two graph algorithms for corpus-based disambiguation of nominal senses. We have shown that the parameter optimization learnt over a small set of nouns signi cantly improves the performance for all nouns, and produces a system which (1) in a lexical-sample setting (Senseval 3 dataset) is 10 points over the Most-Frequent-Sense baseline, 1 point over a supervised system using the same kind of information (i.e. bag-of-words features), and 8 points below the best supervised system, and (2) in the all-words setting is a la par the best supervised system. The performance of PageRank is statistically the same as that of HyperLex, with the advantage of PageRank of using less parameters.</Paragraph>
    <Paragraph position="1"> In order to compete on the same test set as supervised systems, we do use hand-tagged data, but only to do the mapping from the induced senses into the gold standard senses. In fact, we believe that using our WSD system as a purely unsupervised system (i.e. returning just hubs), the perfomance would be higher, as we would avoid the information loss in the mapping. We would like to test this on Information Retrieval, perhaps on a setting similar to that of (Schcurrency1utze, 1998), which would allow for an indirect evaluation of the quality and a comparison with supervised WSD system on the same grounds.</Paragraph>
    <Paragraph position="2"> We have also shown that the optimization according to purity and entropy values (which does not need the supervised mapping step) yields very good parameters, comparable to those obtained in the supervised optimization strategy. This indicates that we are able to optimize the algorithm in a completely unsupervised fashion for a small number of words, and then carry over to tag new text with the induced senses.</Paragraph>
    <Paragraph position="3"> Regarding ef ciency, our implementation of HyperLex is extremely fast. Trying the 6700 combinations of parameters takes 5 hours in a 2 AMD Opteron processors at 2GHz and 3Gb RAM. A single run (building the MST, mapping and tagging the test sentences) takes only 16 sec. For this reason, even if an on-line version would be in principle desirable, we think that this batch version is readily usable as a standalone word sense disambiguation system.</Paragraph>
    <Paragraph position="4"> Both graph-based methods and vector-based clustering methods rely on local information, typically obtained by the occurrences of neighbor words in context. The advantage of graph-based techniques over over vector-based clustering might come from the fact that the former are able to measure the relative importance of a vertex in the whole graph, and thus combine both local and global cooccurrence information.</Paragraph>
    <Paragraph position="5"> For the future, we would like to look more closely the micro-senses induced by HyperLex, and see if we can group them into coarser clusters. We would also like to integrate different kinds of information, specially the local or syntactic features so successfully used by supervised systems, but also more heterogeneous information from knowledge bases.</Paragraph>
    <Paragraph position="6"> Graph models have been very successful in some settings (e.g. the PageRank algorithm of Google), and have been rediscovered recently for natural language tasks like knowledge-based WSD, textual entailment, summarization and dependency parsing. Now that we have set a robust optimization and evaluation framework we would like to test other such algorithms (e.g.</Paragraph>
    <Paragraph position="7"> HITS (Kleinberg, 1999)) in the same conditions.</Paragraph>
  </Section>
class="xml-element"></Paper>