<?xml version="1.0" standalone="yes"?>
<Paper uid="W91-0212">
  <Title>Redefining the &amp;quot;Level&amp;quot; of the</Title>
  <Section position="2" start_page="0" end_page="127" type="intro">
    <SectionTitle>
1 Introduction
</SectionTitle>
    <Paragraph position="0"> In a recent paper Boguraev and Levin (1990) point out inadequacies in common conceptions of what a Lexical Knowledge Base (LKB) should be, inadequacies which stem from the assumption that a machine-readable dictionary (MRD) is not only the right source for acquisition of a lexicon but also the right model for its form. Their points about the kinds of generalizations that the standardly conceived LKB does not incorporate, especially if it is built from a dictionary source, are well-taken. They cite a need for the representation of various kinds of dynamic potential--specifically, the capacity to participate in predictable syntactic alternations and regular meaning extensions--which words of particular semantic classes possess in common, and which constitute the basis for lexical productivity.</Paragraph>
    <Paragraph position="1"> Representation of this kind of dynamic potential is missing from standard dictionaries, they suggest, because it is seen as being more predictable to a native speaker than are specific -nym relations. The same representation is missing from most LKBs both because it was not in the source dictionary and because it cannot be encoded in the usual LKB format. Boguraev and Levin conclude that tile conception of the LKB must be extended to include both representation of tile semantic classes whose members typically participate in these productive alternations and extensions, and sets of inference rules which capture the productive potential of members of those classes. They see this as a partial answer to the open-endedness problem; when a system of the sort they envision is given a new word, we need only specify its semantic class, and we will then have many aspects of both its syntactic and its meaning behavior in hand.</Paragraph>
    <Paragraph position="2"> However perhaps a more serious difficulty arising from the adoption of the dictionary as model for the lexicon, and one which Boguraev and Levin do not discuss, is the implicit assumption that the level of the word, as structurally defined, is a semantically well-defined level also. 1 In the sections that follow, I will examine this assumption from several angles. Section 2 reviews the semantic behavior of what are structurally words,  illustrating cases where it seems to take more than one &amp;quot;word&amp;quot; to make a single meaning, and where a single &amp;quot;word&amp;quot; is semantically decomposable into several meaning components. Section 3 discusses the problem of extricating a functor's argument requirements from other aspects of its meaning. In Section 4 the discussion turns to difficulties inherent in distinguishing individual word senses. Sections 5 explores some possibilities for escaping from the dictionary model and solving some of these problems.</Paragraph>
  </Section>
class="xml-element"></Paper>