<?xml version="1.0" standalone="yes"?>
<Paper uid="W00-1431">
  <Title>The.,CLEF= semi-~recursive generation algori:thm</Title>
  <Section position="5" start_page="235" end_page="236" type="metho">
    <SectionTitle>
MARY TA RT//~N,, SLEEP I
PETER TART PETER
</SectionTitle>
    <Paragraph position="0"> Fimlre 'q. The 1 st nha~e.</Paragraph>
    <Paragraph position="1"> .~&lt; Mary~cctoke&amp;a:fatt;;then..:. }});;~hen- ::-COOK will be computed before EAT and SLEEP. On the contrary, ifa tree anchoring &lt;&lt; S1 after $2, is used (like in &lt;&lt; Peter ate the tart and fell asleep after... &gt;&gt;), EAT and SLEEP will be computed first, then COOK.</Paragraph>
    <Paragraph position="2"> This important property of the algorithm  The graph walk has the following properties : o It is carried out in three phases, one for each conceptual level.</Paragraph>
    <Paragraph position="3"> o The walk is done depth-first, but following the surfaeie linear order of the elements. Thus, the walk order for the lSLorder elements depends on the lexical choices for R0 and R1 (see fig.3) : if the lexicalized tree selected for R1 situates E2 before RI, then E2 should be computed before El I and El2.</Paragraph>
    <Paragraph position="4"> The same way, for El 1 and El2, the order of the processing will depend on the tree selected by R1. On our example (fig. 5 to 7), this means that the order of walk of the second phase (regarding the l SLorder relations) will depend on the lexical choices took during the first phase. For instance, if the SUCCESSION concept is lexicalized using a tree anchoring &lt;~ S1 then $2 ~&gt; (like in The qrd nh,qRe allows to make lexical choices according to previous ones.</Paragraph>
    <Paragraph position="5"> * A stack is added, and allows the storage of every lexical choice according to their linear order. This stack stores the history of the choices carried out, and thus allows to backtrack when needed.</Paragraph>
    <Paragraph position="6"> o During the tree walk, some constraints are propagated towards the lower elements of the tree. So, a lexical-syntactic selection of a concept would be able to add constraints over the lexical choices of_the lexical bases lowerdependent. For example, such a concept could select a particular Form feature, or a particular set ofT-Features, for one of its dependent.</Paragraph>
    <Paragraph position="7"> Unlike the semi-recursive algorithm, the global context choice also carries out according to local choices. Tile stylistic rules can not only use the  . .information given by. the:conceptsofthe same level, but also the information given by the dependent nodes, which allows the retrieval of some predictable information. This information is, of course, limited as the lexical-syntactic choice of the dependent lexical bases if not performed at this point.</Paragraph>
    <Paragraph position="8"> For example, if E2 and E 12 refers to the same ~ concept, some constraints could be.e0mputed as. soon as the lexical selection for R0 is done, and added to the lexical bases of the dependent nodes of R0, in particular RI.</Paragraph>
    <Paragraph position="9"> 3.4.2 Minimum constraint propagation : controlling the backtracking The backtrack, when the processing comes to an impossibility for a lexical base to get a satisfactory lexical choice, is not excluded, although it is inherently limited by the nature of the graph walk.</Paragraph>
    <Paragraph position="10"> The handling of the backtracking can also take advantage of both the walk mechanism and the data structures used.</Paragraph>
    <Paragraph position="11"> For example, if the algorithm fails to find a valid choice for the El2 element (see figure 3), the backtracking can be performed directly on R1, to find an alternative choice compatible with all the dependent nodes (that is E 11 and R0, in our example). Ifa modification is both available and compatible with the related lexicai bases, it will be validated. For a lexical base, a modification is considered compatible if it does not imply any modification in the set of implicit constraints added to the related lexical bases, except the lexical base from which the processing backtracks.</Paragraph>
    <Paragraph position="12"> In our example, the backtracking process can ask the LB for RI to find an alternative choice that modifies the constraints on El2, without modifying any other, that is letting the implicit constraints for E11 et R0 untouched. In other words, it consists in adding a new constraint on * R1 that will imply the selection of a new lexical-syntactic structure so that the implicit constraints on E11 and R0 remain identical, and so that the implicit constraints on El2 impose a new, compatible, choice.</Paragraph>
    <Paragraph position="13"> The algorithm can therefore carry out a minimum propagation of the constraints, without necessarily doing a full backtracking. In that case. ......... it_attem pts .to ..find z~the~.best.compram ise .w ith.:the.. related lexical *bases, before falling back to the normal backtracking mechanism.</Paragraph>
  </Section>
class="xml-element"></Paper>