<?xml version="1.0" standalone="yes"?>
<Paper uid="W97-0215">
  <Title>Combining Knowledge Sources for Automatic Semantic Tagging</Title>
  <Section position="2" start_page="0" end_page="88" type="abstr">
    <SectionTitle>
2 Issues for the Workshop
</SectionTitle>
    <Paragraph position="0"> The main issue for discussion will be the advantages of various methods of combining evidence.</Paragraph>
    <Paragraph position="1"> Other issues that could be discussed include:  * Do we assume there is always a single correct tag, or do we allow a set of equally correct tags? * Do we rank or assign probabilities for all senses? * Do we tag phrases/collocations/idioms (or just individual tokens)? If so, this complicates evidence combination.</Paragraph>
    <Paragraph position="2"> * What preprocessing do we assume as input to the taggers in the dynamic scenario? * DO we approach evidence combination for homograph distictions differently than for polysemy? Are there other types of differences among senses that might affect evidence combination? * What are the implications of a sequential combination of evidence vs. a paralel approach for the dynamic scenario? * How do we map word senses/semantic tags from multiple knowledge sources into a single set in the static knowledge acquisition scenario?  Possible sources of evidence that could be considered for dynamic combination include: domain tags (e.g., LDOCE box codes), collocational and corpus co-occurrence approaches, frequency (domain-specific or domain-independent), selectional restrictions, decision trees, part of speech and subcategorization, Lesk et al dictionary approaches, semantic distance approaches over ontologles, spreading activation/marker passing over semantic nets, scripts/MOPs, word experts.</Paragraph>
    <Paragraph position="3"> Possible sources for static combination include: MRD entries, WordNet, Levin verb classes, corpus statistics, and other lexical resources.</Paragraph>
    <Paragraph position="4"> In order to constrain the discussion, we will make the following assumptions: Senses for each word have been pre-enumerated (Compare, for example, Pustejovsky or Nunberg, and the references cited in these works which point out difficulties in enumerating senses.) In the dynamic case, we are combining compatible knowledge sources, i.e., they share the a semantic tagset. (Contrast work in combining WordNet and Levin Classes.)</Paragraph>
  </Section>
class="xml-element"></Paper>