<?xml version="1.0" standalone="yes"?>
<Paper uid="W97-1509">
  <Title>Head-Driven Generation and Indexing in ALE</Title>
  <Section position="3" start_page="62" end_page="62" type="intro">
    <SectionTitle>
2 Head-Driven Generation
</SectionTitle>
    <Paragraph position="0"> Our head-driven generator uses essentially the same control strategy as proposed in (SNMP90), which was first used in the BUG system of (Noo89). This algorithm is quite well suited to large-scale HPSG generation, as it avoids the termination problems inherent to top-down processing of strongly lexicocentric theories, and, at the same time, does not require of its grammar rules the same naive form of compositionality, known as semantic monotonicity, as Earley-based strategies do. A semantically monotonic grammar rule is one in which the semantic contribution of every daughter category subsumes a portion of the contribution of the mother category.</Paragraph>
    <Paragraph position="1"> In general, wide-coverage theories cannot guarantee this.</Paragraph>
    <Paragraph position="2"> Control in this algorithm is guided by meaning rather than a particular direction over a string, and thus requires the user to distinguish two classes of rules: those in which a mother has the same semantics as some daughter (a chain rule), and those in which it does not (non-chain rule). The strategy is a combination of bottom-up and top-down steps based on the location of a pivot, the lowest node in a derivation tree which has the same semantics as the root goal. Once a pivot is located, one can recursively process top-down from there with non-chain rules (since the pivot must be the lowest such node), and attach the pivot to the root bottom-up with chain rules. A pivot can either be a lexical entry or empty category (the base cases), or the mother category of a non-chain rule. The base case for bottom-up processing is when the pivot and root are taken to be the same node, and thus unified. The reader is referred to (SNMP90) for the complete algorithm.</Paragraph>
    <Paragraph position="3"> What we will be concerned with here is the adaptation of this algorithm to grammars based on a logic of typed feature structures, such as HPSG.</Paragraph>
    <Paragraph position="4"> (SNMP90) uses definite clause grammars, while (Noo89) uses a Prolog-based extension of PATR-II, which has features and atoms, but no featurebearing types, and thus no appropriateness. Unlike both of these approaches, our goal is also to compile the grammar itself into lower-level code which is specifically suited to the particular requirements of head-driven generation, very much as ALE already does for its parser, and much as one would compile a Prolog program for a particular set of mode specifications. null</Paragraph>
  </Section>
class="xml-element"></Paper>