<?xml version="1.0" standalone="yes"?>
<Paper uid="C90-2029">
  <Title>Constraining Tree Adjoining Grammars by Unification</Title>
  <Section position="5" start_page="0" end_page="0" type="evalu">
    <SectionTitle>
SUTAG and IUTAG
</SectionTitle>
    <Paragraph position="0"> Finally, let's weigh the two approaches. A first impression in reading both definitions can be that the differences are marginal. It is simply a decision in ordering processes. But a second look offers that there ,are more differences. The ideas of what should be represented by an adjoining are different. A SUTAG supports the monotone idea behind unification and the parallel and absolute presence of propagated knowledge (what is written in tile '1&amp;quot; and ,\[, sets should be propagated exactly fiom that node all over the tree withot,t any more changes). In a UTAG, the idea of filtering propagated information has the highest priority. Specification information should be revisable by new adjoinings. Each of the two approaches works more efficient for different problem classes, which cannot be characterized in more detail here.</Paragraph>
    <Paragraph position="1"> For SUTAGs, the main disadvantage is the late interpretation of the specifcation rules. Especially here lies the motivation for the definition of IUTAGs. In the above mentioned example, the focus in describing the advantage of online unification is demonstated by the representation of the intermediate states of the DAGs, but it is clear as well, how the lexical readings with case equal dative or accusative for &amp;quot;Marie&amp;quot;,&amp;quot;Pict&amp;quot; and &amp;quot;Jan&amp;quot; are eliminated.</Paragraph>
    <Paragraph position="2"> Another distinguishing feature of IUTAGs is that unification is not further a monotone process as it is in SUTAGs (same as for UGs). Links, which have been installed, can be eliminated. This is the more formally circumscribed effect of a filter, which is the intuitive idea behind that definition. It is obvious that the realization of that fact is more expensive than to wait until all adjoinings are realized. But it comes up with a reduction of execution time for invalid readings.</Paragraph>
    <Paragraph position="3"> Actually, this is a trade off, which has more or less effect on the time complexity in the relation in which specification rules are used to subcategorize and restrict the structural descriptions. Therefore, both definitions probably have more lxzrfect application domains.</Paragraph>
    <Paragraph position="4"> A point which was mentioned as a disadvantage of the Unification formalism, was the Turing capacity. For both definitions of UTAGs, the power is restricted in a 5 i71 way t!hat no path can be introduced via specifications without adding a piece of structure, because the definition of an elementary tree requires a non-empty leaf. Therefore, the defined unification process underlies the constant growth property. At the moment, it is not yet clear, what class of languages can exactly be described with that formalism, but it is obvious that it is less powerful than Turing machines (e.g., languages such as a 2n are not describable). But a result for an upper time bound of an algorithm for that formalism is yet unknown.</Paragraph>
    <Paragraph position="5"> by 1'(01) reaching 01 in cc by 1'(11) reaching 11 in ~1: and reintroduced at 11 (and reintroduced at 20 ~r yd!oining ,of 131) ' , after adjoining of \[32) ntax- veroform- in! ~set'~&amp;quot; ...... fset t-syntax- verbform- inf y\[It \[ule vero-zwemmen , ~, .^, ...... a ~,^~</Paragraph>
  </Section>
class="xml-element"></Paper>