<?xml version="1.0" standalone="yes"?>
<Paper uid="E87-1023">
  <Title>A MODEL FOR PREFERENCE</Title>
  <Section position="4" start_page="137" end_page="137" type="metho">
    <SectionTitle>
5. Future directions
</SectionTitle>
    <Paragraph position="0"> The implementation of this preference model has been written in Prolog. To facilitate experimentation, a mechanism is provided for tracing the preference rules application to observe their behaviour.</Paragraph>
    <Paragraph position="1"> The model described above is very flexible. We are currently studying the implementation of variants of the basic comparison algorithm: We are investigating algorithms that would: - reduce the number of comparisons, by aiming at extracting only the maximal (minimal) elements, without trying to order all elements.</Paragraph>
    <Paragraph position="2"> - calculate the transitive closure of the directed graph, and then remove all contradictory relationships, thereby removing all cycles. This amounts to saying that two interpretations are not comparable if their comparison leads to contradictory decisions.</Paragraph>
    <Paragraph position="3"> - compare the competing interpretations stepwise, that is all comparisons are performed with the first rule in a list, then only the pairs for which there is no decision yet are compared with the second rule, and so on.</Paragraph>
  </Section>
  <Section position="5" start_page="137" end_page="137" type="metho">
    <SectionTitle>
ACKNOWLEDGEMENTS
</SectionTitle>
    <Paragraph position="0"> We would like to thank Paul Bennett, Maghi King, Gertjan Van Noord, Mike Rosner and Susan Warwick for their fruitful comments and their support.</Paragraph>
  </Section>
  <Section position="6" start_page="137" end_page="137" type="metho">
    <SectionTitle>
APPENDIX
</SectionTitle>
    <Paragraph position="0"> In the current framework of EUROTRA (Arnold and des Tombe 1987), representation of interpretations are derivation trees, containing at each node a set of attribute-value pairs. Here is a very sketchy and intuitive description of the syntax used in the patterns: - The identifiers s, np, vp etc. are values of the distinguished attribute of the node (in these examples, the syntactic category).</Paragraph>
    <Paragraph position="1"> - Curly brackets delimit a set of conditions to be satisfied by a node. For example (s,f=declarative} indicate the required conditions on the node for the distinguished attribute (should have value s) and for an f attribute (should  have value declarative).</Paragraph>
    <Paragraph position="2"> - SA, SB, etc. are variable identifiers. - s.\[np,vp\] indicates a tree with root s and two daughters np and vp.</Paragraph>
    <Paragraph position="3"> - ? or (?) indicates an unspecified node. - * indicates a list of unspecified nodes.</Paragraph>
    <Paragraph position="4"> - SAiPattern indicates that the variable $A is instantiated to the sub-tree that matches Pattern - $more branches (and $1ess_branches) is  a predefined preference rule that prefer the argument that has more (less) branches than the other.</Paragraph>
    <Paragraph position="5"> - The first rule declared becomes the distinguished rule applied to the competing interpretations.</Paragraph>
    <Paragraph position="7"> This set of preference rules will explore, in parallel, two trees, from top to bottom, always taking the 's' branch, and prefer the tree in which it finds a declarative sentence (opposed to an interrogative).If one inverts the order of pl and p2 in the distinguished composite rule p0 the trees would be explored from bottom to top.</Paragraph>
    <Paragraph position="8"> Rule p0 just passes its arguments to pl or p2~ Rule pl prefers a declarative s over an interrogative s.</Paragraph>
    <Paragraph position="9"> Rule p2 identifies the embedded s in each argument and passes them to pl or p2.</Paragraph>
    <Paragraph position="11"> p3($A,$B)).</Paragraph>
    <Paragraph position="12"> Given two sentences, this set of rules will prefer the one that has the pp attached deeper in the structure than the other (right attachment). This example is restricted to explore only embedded nps. For both arguments, rule p0 identifies the last daughters of the vp of a sentence s, and passes them to preference rules pl or p2 or p3.</Paragraph>
    <Paragraph position="13"> Rule pl will prefer a pp attached under an np to a pp (which was attached higher in the structure).</Paragraph>
    <Paragraph position="14"> Rule p2 will be tried only if pl was not applicable. It is there for the case the pp is imbedded deeper in the np.</Paragraph>
    <Paragraph position="15"> Rule p3 is similar to rule p0, except that it takes the last daughters of a np. It is tried only if pl and p2 are not applicable. null</Paragraph>
  </Section>
class="xml-element"></Paper>