<?xml version="1.0" standalone="yes"?>
<Paper uid="P99-1005">
  <Title>Distributional Similarity Models: Clustering vs. Nearest Neighbors</Title>
  <Section position="6" start_page="39" end_page="39" type="concl">
    <SectionTitle>
4 Conclusion
</SectionTitle>
    <Paragraph position="0"> In our experiments, the performances of distributional clustering and nearest-neighbors averaging proved to be in general very similar: only in the unorthodox AP89 setting did nearest-neighbors averaging clearly yield better error rates. Overall, both methods achieved peak performances at relatively small values of k, which is gratifying from a computational point of view.</Paragraph>
    <Paragraph position="1"> Some questions remain. We observe that distributional clustering seems to suffer higher variance. It is not clear whether this is due to poor estimates of the KL divergence to centroids, and thus cluster membership, for rare nouns, or to noise sensitivity in the search for cluster splits. Also, weighted-average clustering never seems to outperform the nearest-centroid method, suggesting that the advantages of probabilistic clustering over &amp;quot;hard&amp;quot; clustering may be computational rather than in modeling elfectiveness (Boolean clustering is NP-complete (Brucker, 1978)). Last but not least, we do not yet have a principled explanation for the similar performance of nearest-neighbors averaging and distributional clustering. Further experiments, especially in other tasks such as language modeling, might help tease apart the two methods or better understand the reasons for their similarity. null</Paragraph>
  </Section>
class="xml-element"></Paper>