<?xml version="1.0" standalone="yes"?>
<Paper uid="C94-2162">
  <Title>Towards a Proper Linguistic and Computational Treatment of Scrambling: An Analysis of Japanese</Title>
  <Section position="5" start_page="7002" end_page="7002" type="concl">
    <SectionTitle>
3 Scrambling and Comput.a-
</SectionTitle>
    <Paragraph position="0"> lion Although J~qsmw.sc scra.mbling ix comph~x, we haw~ seen that by lever,tging existing principles, ma, ny examples of short a.ml long disl;ance scrmnb/ing can he ~cconmlodal.ed almost wit.lloul, cha.ugc t.o i.he cxist.ing l.heory. At first glance, the same seems I.o I&gt;c t,\[l(~ c~/se for COlil\[)llttl,tiO\[l. (~ellel'&amp;\] l)\[lr~l,se Sl, l'llCl, lll'(~ h~cluding the addit, ional a, djtlnctioil to IP ~md VP is cowa:ed by the exist.iug M{(I) based I)oLgoln-Ul)shift.-I'edtlce |5;Lrser (1(11111.tl, 1965). The relation between a scramlded object mid i~s hmneh site ix computed as one p~rt o|' the gCller~-t\[ l'lll(z of lllOVClIlelll.,, move-(~. 12 degTwo poinl, s: (1) Saito refers t,o I, he Propeu Binding Condi|,lOll rather t, han Ihe I';CP. (2) do~o h.o~-v lowers for the same i'e;tstm iL raises in ((h:). See \[tl)|,e 7.</Paragraph>
    <Paragraph position="1"> lO~/ve llo{e ltel'e (,h&amp;t. |.he illl.el'atcl;iiiiI betwoc'll 1,1&amp;quot; l'D:const,rucl.i.m ett'ecls and scrambling is also discussed in (Tad~L, 1903), \] I In 0m iml)hmmntal, ion, NQ adjoins I,o NI' and both NQ-NP and NP-NQ Ol'H(~rs ;Ire ilia, de avalbdfle.</Paragraph>
    <Paragraph position="2"> |2More precisely, t.h0 relal, ion is .recovered by a t'ule of (',haln I&amp;quot;ormaLion. bee (1,'mtg, 1991) for l, he del.ails ~f this and el, bin&amp;quot;  The problem that arises is that the parser produces extra parses not eliminated by the existing principles, involving vacuous scrambling of the form shown in figtire 2.</Paragraph>
    <Paragraph position="3"> Let us define the obvious notion of &amp;quot;non-vacuous&amp;quot; or visible scrambling as scrambling that &amp;quot;passes over&amp;quot; one or more overt elements at S-strncture. Adopting this definition immediately eliminates cases (a) (c), and more importantly, the repeated application of case (a) which causes parsing non-termination, la In partitular, this rules out vacnons scrambling over empty subjects, e.g. pro or PRO. As far the data from (Saito, 1985),(1992) and ('I'ada, 1993) goes, this is no restriction at all. This leaves case (d) which is vacuous only when considered as a &amp;quot;pair&amp;quot;, i.e. each sub-instance is clearly non-vacuous.</Paragraph>
    <Paragraph position="4"> We will make use of the following assumptions:  1. Scrambling is always to the left.</Paragraph>
    <Paragraph position="5"> 2. Empty NPs don't scramble.</Paragraph>
    <Paragraph position="6">  Again, judging from the data, these appear to be no restriction all. Now, it, is simple to implement the non-vacuity (NV) constraint as a licensing condition on phrasal adjunction: IPi --+ sNP, IPi(x), {NV(x)} VP; ~sNP, VPi(x), {NV(z)} ltere, we assume that there are two phrasM adjnnction rules, for IP attd VP, that introduce scrambled NPs (sNPs). 14 Here, {NV(x)} is a semantic aelion which checks the frontier of x, the tP or VP phrase, for non-vacuity using the following left-to-right, two state machine: State 0: (Start) See an overt node, go to State 1.</Paragraph>
    <Paragraph position="7"> State 1: (End) Skip until see an NP gap, halt.</Paragraph>
    <Paragraph position="8"> Note this is potentially inefficient since the NV constraint is only checked when the LtUmachine completes the RHS of a rule, i.e. when it completes an IP or VP phrase. By operating at the lew~l of the terminal string, instead of waiting for the complete IP or VP, laNote that the string w'tcuous empty operator movement showIl in (C) do(Is not count as an instance of scrambling. It's not adjunction at VP or \[P.</Paragraph>
    <Paragraph position="9"> 14The tricky case of (d) shown earlier can be handled hy restricting sNP to overt NPs only.</Paragraph>
    <Paragraph position="10"> we (:an take advantage of the fact; that scrambling is always to tile left to implement the non-vacuity check in ~ strictly left-to-right fashion. As before, when we see a potentially scrambled NP, we start the two state machine. Ilere, the basic idea is that a shift action (read a terminM) corresponds to the state 0 to state 1 transition. Similarly, a NP-~ ,~ reduce action corresponds to the &amp;quot;see an NP gap and halt&amp;quot; part. Con,paring the two algorithms on 50 Japanese sentences (25 scrambling examples) fi'om (Saito, 1985) &amp; (Lasnik &amp; Saito, 1984), an average speed-up factor of 2.3 per sentence and a total of 3.2 over the 50 sentences was observed for the second algorithm over just the phrase structure recovery portion. Due to the varying effects from the participation of other principh'.s, the improvement for the total parsing time was le.ss clearcut, generally varying between no speed-up at all l;o a factor of two depending on the sentence. Using the 25 non-scrambling examples fl'om the test set, an additional experiment hetween two variants of the parser, one with and one without the ability to handle scram bling, showed that the ability to handle scrambling exacts a penalty of about 30 50% in total parsing time, In conclusion, given tit(; pertlaps disproportionate el'feet of scrambling on parsing time, we suggest that although scrambling comes virtually for free linguistically speaking, the same certainly does not N)pear to be the case for computation.</Paragraph>
  </Section>
class="xml-element"></Paper>