<?xml version="1.0" standalone="yes"?>
<Paper uid="H86-1023">
  <Title>The Lexicon in Text Generation</Title>
  <Section position="2" start_page="253" end_page="254" type="ackno">
    <SectionTitle>
5.4. Metaphor
</SectionTitle>
    <Paragraph position="0"> A large range of phenomena which have been treated as idiosyncratic to individual lexical items -- i.e. as idioms or collocations -- could perhaps be treated in a more motivated way in a system which had a notion of standard metaphor. (This proposal is cogently stated in \[--- 85\]; the sense of metaphor involved here is that presented in e.g. \[Lakoff &amp; Johnson 80\].) For example, consider the metaphor &amp;quot;time is money&amp;quot;. In a system which had a way of representing this association, a number of collocations involving time -- &amp;quot;spend time&amp;quot;, &amp;quot;waste time&amp;quot;, &amp;quot;lose time&amp;quot; etc. -- are not random, but can be predicted from the collocations involving money. Another set of expressions involving time, e.g. &amp;quot;time passed&amp;quot;, &amp;quot;time flies&amp;quot;, &amp;quot;the days marched by in weary succession&amp;quot; etc., are derived from another standard metaphor for time, namely &amp;quot;time is a moving object&amp;quot;. While some of Mel'chuk's lexical functions have to do with standard metaphors of this sort, as far as I know his is the only system that treats them systematically as such, although any system based on a taxonomic hierarchy with inheritance can simulate metaphor after a fashion. For example, there is a popular metaphor &amp;quot;a computer is a conscious being&amp;quot;, which is involved when we refer to computers as agents of processes that normally only take conscious agents, e.g. &amp;quot;the computer deleted my files&amp;quot;. In the Janus system, the only convenient way to represent this is by classifying the concept COMPUTER under CONSCIOUS BEING in the semantic taxonomy. Ideally, however, it would be preferable not to commit one's taxonomy to the claim that a computer is literally a conscious being, since we also talk about computers as unconscious objects; e.g. we usually say &amp;quot;the computer that just went down&amp;quot;, not &amp;quot;the computer who just went down&amp;quot;.</Paragraph>
    <Paragraph position="1"> 5.5. Choice Ideally, a system should have some way of choosing between lexical items on other than purely grammatical and denotational grounds. Human speakers take a variety of factors into consideration when making lexical decisions. We use different words for the same things depending on who we're talking to, what we're talking about, where we are, and what role we're playing. A simple example is the observation that in more formal contexts English speakers tend to use Latinate words such as &amp;quot;expunge, remove, infer&amp;quot; instead of Anglo-Saxon phrasal verbs like &amp;quot;wipe out, take off, figure out&amp;quot;. In addition to simply responding to social context in the way we choose words, we can use words in a way which evoke or create a context for our utterances; for instance, we can use borrowings from French in order to sound suave, or surfer slang in order to sound cool. We use more general or more specific terms for the same thing depending on which of its characteristics we're interested in: ff we see a friend careening towards a tree, we're more likely to say &amp;quot;watch out for that tree!&amp;quot; than &amp;quot;watch out for that eucalyptus!&amp;quot; or &amp;quot;watch out for that plant!&amp;quot; And so on. We're a long way from having natural language generators that have the degree of control over any level of linguistic choice, grammatical or lexical, that a serious treatment of these considerations would entail; but we can design our systems such that such distinctions could be accomodated when we have the analyses to support them.</Paragraph>
    <Section position="1" start_page="253" end_page="254" type="sub_section">
      <SectionTitle>
5.6. Conclusion
</SectionTitle>
      <Paragraph position="0"> Lexicons play a wide varieties of roles in text generation systems, from the very central one of providing the primary link between form and meaning, to the quite peripheral one of finishing up after the grammar is done. Lexical phenomena such as semantic relationships, syntactic classes, collocation and idioms have received vastly different amounts of attention in different systems, while other phenomena such as metaphor and non-denotational meaning have received virtually none. Looking at the capabilities of a wide range of generation lexicons provides an exhilirating sense of the potential for future systems, both from the variety of phenomena that existing systems have dealt with, and from the challenges that still remain. I hope that bringing a few of these phenomena to light in this paper will succeed in sparking the interest necessary to ensure the lexicon the attention it warrants in text generation research.</Paragraph>
    </Section>
  </Section>
class="xml-element"></Paper>