<?xml version="1.0" standalone="yes"?>
<Paper uid="C94-2135">
  <Title>ACHIEVING FLEXIBILITY IN UNIFICATION FORMALISMS</Title>
  <Section position="3" start_page="0" end_page="842" type="intro">
    <SectionTitle>
1 MOTIVATION
</SectionTitle>
    <Paragraph position="0"> Unification based-formalisms are in common use within natural language processing and many different formalisms have been defined. PATR-II (Shieber et al., 1983) is the most basic and a kind of common denominator which other formalisms are extensions of. Other formalisms are STUF (Beierle et al., 1988), TFS (Emele &amp; Zajac, 1990), CUF (D6rre &amp; Eisele, 1991) and, ALE (Carpenter, 1992). These formalisms include, for example, disjunction, various variants of negation and typing. When vaa'ious grammatical theories, such as LFG (Kaplan &amp; Bresnan, 1983) or HPSG (Polku'd &amp; Sag, 1987) are included, the range of extensions suggested to unification-based grammars becomes very wide. There are also many variant proposals on how the same extension should be used and interpreted.</Paragraph>
    <Paragraph position="1"> When using these formalisms for a particular problem, it is oRen the case that the constructions provided do not con'espond to the needs of your problem. It could either be the case that you want an additional construction or that you need a slight modification of an existing one. Since the extensions are nmnerous it seems hard to include everything in one single formalism.</Paragraph>
    <Paragraph position="2"> In some formalisms, especially TFS and CUF, the user is allowed to define new constructions. This is an interesting property that I will develop further to achieve flexibility. In a flexible formalism, the user can define all the constructions he needs or modify definitions provided by the formalism. With this kind of formalism problems such as those mentioned above would not arise.</Paragraph>
    <Paragraph position="3"> 1. Email address: lestr@ida.liu.se A flexible formalism would be a useful tool for delining various kinds of knowledge needed at different levels in a natural language system. It would be a great advantage to be able to use the same system for all levels, but adjusting it to suit the various structures that are needed at each level since the relations between the different levels would be clearer and it would be easier to share structures between the levels (cf. Seiffert (1992) for more motiwttion).</Paragraph>
    <Paragraph position="4"> Another advantage with such a formalism is that it can be used to define and test extensions and w~rious grammatical formalisms for the purpose of comparison. null Flexible formalisms allow the user to defne an expensive extension and use it for the cases where he really needs it. Thus an extension that is considered too expensive to be provided by a general formalism, can be included as it can provide a more efficient representation of some particular phenomenon. This is particularly important since, in natural language processing, it is common that expensive constructions are required for few and limited cases.</Paragraph>
    <Paragraph position="5"> However, if the user is allowed to define expensive constructions, an additional goal, predictability, must be considered. By this I mean that inefficient computations should be necessary only when the construction causing the inefficiency really participates in the computation. This idea was discussed in StrOmb/ick (1992) where I give a predictable algorithm for unification of feature structures containing disjunction. The goal of predictability is closely related to modularity, since if it is possible to make different constructions independent of each other, it is easier to find predictable algorithms for them. Since this paper will discuss properties of a flexible formalism rather than unification algorithms there, will be no further discussion of predictability.</Paragraph>
    <Paragraph position="6"> In the following I first discuss the most important properties of a flexible formalism. I then present a flexible formalism, FLUF, by using it to define PATR-I\[. The size of this paper does not admit a thorough description of FLUF and its semantics. This is given in StrOmbfick (l 994a, 1994b).</Paragraph>
  </Section>
class="xml-element"></Paper>