<?xml version="1.0" standalone="yes"?>
<Paper uid="C86-1033">
  <Title>A Stochastic Approach t O Parsing</Title>
  <Section position="2" start_page="154" end_page="154" type="abstr">
    <SectionTitle>
6. Simulated annealing appeals
</SectionTitle>
    <Paragraph position="0"> strongly to some writers (e.g. Bridle &amp; Moore 1984: 315) as a model of psychological perception mechanisms. In the case of grammatical parsing, though, there is one respect in which the model presented so far is quite implausible psychologically: it ignores the left-to-right sequential manner in which humans procesS written as well as spoken language. There is a natural way to incorporate time into an annealing parser which not only is psychologically plausible but promises greatly to increase its efficiency as a practical automatic system.</Paragraph>
    <Paragraph position="1"> Rather than a whole sentence being submitted to the annealing system at once, in a &amp;quot;dynamic&amp;quot; annealing system parsing would proceed in a series of rounds. The input to the nth round would be an annealed parsing of the first n-1 words of the sentence, followed by the nth word; annealing would begin anew at melting temperature on this input. The opportunity for efficiency would arise from the fact that NL grammar only rarely forces the reader to backtrack -- the insight on which Mitchell Marcus's Parsifal system was founded (Marcus 1980). Marcus's strategy involved a total exclusion of backtracking from his central parsing system, with &amp;quot;garden path&amp;quot; sentences being handed over to a quite separate &amp;quot;higher level problem solver&amp;quot; for processing. However, Marcus's predictions about a sharp categorization of NL sentences into garden-paths and non-garden-paths have provoked considerable criticism. In a dynamic annealing parser, all parts of the curreDtr tree would at all stages be available to revision, but the relative rarity of the need for backtracking could be exploited by adding a bias to the function which randomly selects nodes for reconsideration, so that nodes are reconsidered less frequently as they become &amp;quot;older&amp;quot;. Since the bulk of computing time in an annealing parser would undoubtedly be consumed in calculating gains and losses for candidate tree-changes, this system of concentrating the search for profitable tree-changes on the areas of the tree where such changes are most likely to be found could be a good means of saving processing time by reducing the total number of moves considered.</Paragraph>
    <Paragraph position="2"> 7. A problem that will not have escaped the reader's attention is that I have discussed parsing purely in terms of finding surface parse-trees (which hapEens to be the task which the UCREL group are engaged on). It is not obvious how to extend the annealing approach so as to yield deep parses. However, there is nothing about simulated annealing that makes it intrinsically inapplicable to the task of deep parsing. What needs to be done is to define a class of logically-possible deep parse-trees and a class of moves between them, and to find an evaluation function which takes any pairing of a deep structure with a surface word-sequence into a likelihood-value. This task is very different in kind from the work currently done by theoretical linguists and AI researchers interested in underlying or logical grammar, who tend to have little time for statistical thinking, but that is not to say that the task is necessarily senseless or impossible. Deep parsing, if possible at all, will presumably need to exploit semantic/&amp;quot;inferencing&amp;quot; consideratJons as well as information about grammar in the narrow sense, but nothing says that these matters might not be built into the evaluation function.</Paragraph>
    <Paragraph position="3"> 8 Finaiiy, it may be that annealing is:useless as a parsing technique because the geometry of NL parsing space is wrong.</Paragraph>
    <Paragraph position="4"> Perhaps the space of English parse-trees (whether surface or deep) resembles the Witwatersrand rather than the Cotswolds, being an upland plateau riddled with deep goldmines rather than a rolling landscape whose treasures lie exposed in valley bottoms. I conjecture that NLs are Cotswold-like rather than Rand-like, and that, if they were not, humans could not understand them. Only empirical research using authentic data can settle the question.</Paragraph>
  </Section>
class="xml-element"></Paper>