<?xml version="1.0" standalone="yes"?>
<Paper uid="W04-3213">
  <Title>Unsupervised Semantic Role Labelling</Title>
  <Section position="8" start_page="5" end_page="5" type="concl">
    <SectionTitle>
7 Conclusions and Future Work
</SectionTitle>
    <Paragraph position="0"> Using an unsupervised algorithm for semantic role labelling, we have achieved 90% correct on identified arguments, well over an informed baseline of 77%, and have achieved 87% correct on all slots (64% baseline). On PP objects, our conservative role assignment shows promise at leaving adjuncts unlabelled. However, PP objects also have the lowest performance (of 78% correct on identified arguments, compared to 93% for subjects or objects).</Paragraph>
    <Paragraph position="1"> More work is required on our frame matching approach to determine appropriate roles for PP objects given the specification in the lexicon, which (in the case of VerbNet) often over-constrains the allowable prepositions for a slot.</Paragraph>
    <Paragraph position="2"> Although these results are promising, they are only a first step in demonstrating the potential of the approach. We need to test more verbs, from a wider variety of verb classes (or even a different kind of predicate classification, such as FrameNet), to determine the generalizability of our findings. Using FrameNet would also have the advantage of providing large amounts of labelled test data for our evaluation. We also hope to integrate some processing of adjunct roles, rather than limiting ourselves to the specified arguments.</Paragraph>
    <Paragraph position="3"> A unique aspect of our method is the probability model, which is novel in its generalizations over verb, slot, and noun classes for role labelling. However, these have room for improvement--our noun classes are coarse, and prepositions clearly have the potential to be divided into more informative subclasses, such as spatial or time relations. Our on-going work is investigating better class models to make the backoff process even more effective.</Paragraph>
  </Section>
class="xml-element"></Paper>