<?xml version="1.0" standalone="yes"?>
<Paper uid="C90-2012">
  <Title>The E-Framework: Emerging Problems</Title>
  <Section position="3" start_page="0" end_page="0" type="relat">
    <SectionTitle>
3 Consolidation and Ambi-
</SectionTitle>
    <Paragraph position="0"> guity Consolidation has a peculiar property, that in certain restricted circumstances it throws away disambiguation results, recreating earlier ambiguities.</Paragraph>
    <Paragraph position="1"> Suppose there is a sentence S that according to grammar G has a set of representations rt = {r:, r2,...r,}.</Paragraph>
    <Paragraph position="2"> In the EI&amp;quot;W these are trees, but consider them as labelled bracketings. Every rl will have a &amp;quot;stretch set&amp;quot; S(r ) = {,' I r c rt r has at least the same brackets as ri} Suppose then that some ri is consolidated according to a G ~ that also assigns R as the representations of S. ri will consolidate ambiguously, yielding S(rl).</Paragraph>
    <Paragraph position="3"> The linguistic claim that this embodies is obviously false, if G and G ~ are similar enough to yield the same set of representations for S, then each ri of G is in truth equivalent to a single representation, identical to ri~ in G': it is not equivalent to a set of G ~ representations that partly recapitulates the ambiguities that are already identified.</Paragraph>
    <Paragraph position="4"> An example is co-ordination. Consider the co-ordinated NP Bob and Carol and Ted and Alice. This is 11 ways ambiguous and might be  Plausible t-rules into some target language will map these onto descriptors identical to tile source representation. Each local tree of these descriptors will then be parsed. It is obvious that this process will ill the case of (6)(a) will produce 11 consolidations identical to the original 11 parses. (6)(b) will consolidate unambiguously into something identical to (6)(b). This is because (6)(a) is a case where S(ri) = R, and (6)(b) a case where S(r,) = {r,}. Less fiat representations have a smaller stretch set and in this example will consolidate into 3 or 1 translations.</Paragraph>
    <Paragraph position="5"> The claim that (6)(a) is 11 ways ambiguous in any target language is false. (6)(a) represents one interpretation of the surface string, and the one interpretation has one translation into any other language.</Paragraph>
    <Paragraph position="6"> It is obvious that this weakness also affects the EFW as a formalism for NLP. Suppose (6)(a) were a representation at a predicate-argument level, to be mapped to a surface syntax level of the sarne language. The EFW embodies a claim that (6)(a) is 11 ways arnbiguous on the surface, which is false, just as the claim that (6)(a) has \]1 translations is false. hi fact, re-parsing descriptors adds another dimension to the normal problem of the parsing ambiguity of conjoined structures: the first surface parse will in general be ambiguous, and each of its representations will in general map 4 69 onto many representations at the next level, each of which will in general breed again at the next level, and so on. Some sample figures are  Sadler (1989)). This is plausible: if w in language A translates into x, y, and z in language B, it is surely correct to say that w will appear in the set of possible translations in A of each of x, y and z.</Paragraph>
    <Paragraph position="7"> Many MT notations, including CAT, fail to embody this observation in a reversible notation, and thus fail to force linguists to implement a symmetrical translation relation. The EFW makes it impossible to implement a fully symmetrical relation. (6)(a) translates into 11 things in any target language, but (6)(a) does not appear in the set of possible backtranslations of any of those 11 except one, the one identical to (6)(a).</Paragraph>
  </Section>
class="xml-element"></Paper>