<?xml version="1.0" standalone="yes"?>
<Paper uid="P92-1026">
  <Title>HANDLING LINEAR PRECEDENCE CONSTRAINTS BY UNIFICATION</Title>
  <Section position="9" start_page="206" end_page="207" type="concl">
    <SectionTitle>
CONCLUSION
</SectionTitle>
    <Paragraph position="0"> We have presented a formal method for the treatment of LP constraints, which requires no addition to standard feature unification formalisms. It should be emphasized that our encoding only affects the compiled grammar used for the processing. The linguist does not lose any of the descriptive means nor the conceptual clarity that an ID/LP formalism offers.</Paragraph>
    <Paragraph position="1"> Yet he gains an adequate computational interpretation of LP constraints.</Paragraph>
    <Paragraph position="2"> Because of the declarative specification of LP constraints, this encoding is neutral with respect to processing direction (parsing-generation). It does not depend on specific strategies (top-down vs. bottom-up) although, as usual, some combinations are more efficient than others. This is an advantage over the formalization of unification ID/LP grammars in Seiffert (1991) and the approach by Erbach (1991).</Paragraph>
    <Paragraph position="3"> Seiffert's approach, in which LP constraints operate over siblings, requires an addition to the parsing algorithm, by which LP constraints are checked during processing to detect violations as early as possible, and again after processing, in case LP-relevant information has been added later by unification. Erbach's approach can handle LP constraints in head domains by building up a list of constituents over which the LP constraints are enforced, but also requires an addition to the parsing algorithm for checking LP constraints during as well as after processing.</Paragraph>
    <Paragraph position="4"> Our encoding of LP constraints does not require any particular format of the grammar, such as left- or right-branching structures. Therefore it can be incorporated into a variety of linguistic analyses.</Paragraph>
    <Paragraph position="5"> There is no need to work out the formal semantics of LP constraints because feature unification formalisms already have a well-defined formal semantics.</Paragraph>
    <Paragraph position="6"> Reape (1989) proposes a different strategy for treating partially free word order. His approach also permits the application of LP constraints across local trees. This is achieved by separating word order variation from the problem of building a semantically motivated phrase structure. Permutation across constituents can be described by merging the fringes (terminal yields) of the constituents using the operation of sequence union. All orderings imposed on the two merged fringes by LP constraints are preserved in the merged fringe.</Paragraph>
    <Paragraph position="7"> Reape treats clause union and scrambling as permutation that does not affect constituent structure. Although we are intrigued by the elegance and descriptive power of Reape's approach, we keep our bets with our more conservative proposal. The main problem we see with Reape's strategy is the additional  burden for the LP component of the grammar. For every single constituent that is scrambled out of some clause into a higher clause, the two clauses need to be sequence-unioned. A new type of LP constraints that refer to the position of the constituents in the phrase or dependency structure is employed for ensuring that the two clauses are not completely interleaved. Hopefully future research will enable us to arrive at better judgements on the adequacy of the different approaches. Pollard (1990) proposes an HPSG solution to German word order that lets the main verb first combine with some of its arguments and adjuncts in a local tree. The resulting constituent can be fronted.</Paragraph>
    <Paragraph position="8"> The remaining arguments and adjuncts are raised to the subcategorization list 7 of the auxiliary verb above the main verb. Yet, even if a flat structure is assumed for both the fronted part of the clause and the part remaining in situ as in (Pollard 1990), LP constraints have to order major constituents across the two parts.</Paragraph>
    <Paragraph position="9"> For a discussion, see Uszkoreit (1991b).</Paragraph>
    <Paragraph position="10"> Uszkoreit (1991b) applies LP principles to head domains but employs a finite-state automaton for the encoding of LP constraints. We are currently still investigating the differences between this approach and the one presented here.</Paragraph>
    <Paragraph position="11"> Just as most other formal appraoches to linear precedence, we treat LP-rules as absolute constraints whose violation makes a string unacceptable. Sketchy as the data may be, they suggest that violation of certain LP-eonstraints merely makes a sentence less acceptable. Degrees of acceptability are not easily captured in feature structures as they are viewed today. In terms of our theory, we must ensure that the unification of the complement's or adjunct's left or right context restriction with the head's LP-STORE does not fail in case of a value clash, but rather results in a feature structure with lower acceptability than the structure in which there is no feature clash. But until we have developed a well-founded theory of degrees of acceptability, and explored appropriate formal means such as weighted feature structures, as proposed in (Uszkoreit 1991a), we will either have to ignore ordering principles or treat them as absolute constraints.</Paragraph>
  </Section>
class="xml-element"></Paper>