<?xml version="1.0" standalone="yes"?>
<Paper uid="P86-1013">
  <Title>PARSING CONJUNCTIONS DETERMINISTICALLY</Title>
  <Section position="8" start_page="80" end_page="82" type="concl">
    <SectionTitle>
CONCLUSIONS
</SectionTitle>
    <Paragraph position="0"> The most innovative feature in NEXUS is its use of only two kinds of stack structures, one for clauses and one for everything else. When a structure is at the top of the stack, it represents a top.down prediction of constituents yet to come, and words from the input simply drop into the slots that are open to that class of word. When a word is encountered that cannot be inserted into the top structure nor into any structure lower in the stack, a new structure is built bottom-up, the new word inserted in it, and the parse goes on. When a word can both be inserted somewhere in the stack and also in a new structure, all possible parses are pursued in parallel. Thus, NEXUS seems to be a unique member of the wait-and-see family since it is not always deterministic and hence need not disembiguate until all information it could get from the sentence is available.</Paragraph>
    <Paragraph position="1"> The general efficiency of the parser is due primarily to its separation of segmentation from recombination. This is a divide and conquer strategy which reduces a large search space -- grammatical patterns for words in sentences -- into two smaller ones: (1) the set of grammatical patterns for simple phrases and clause nuclei, and (2) the set of allowable combinations of stack structures. Of course, search is still required to resolve structural ambiguity, but the total number of combinations is much less.</Paragraph>
    <Paragraph position="2"> It is not clear whether the parser's speed in the particular cases above comes from divide and conquer or from the differences between Prolog and Maclisp. Nevertheless, as systems are built that require larger, more comprehensive grammars, and that must deal with longer, more complicated sentences, the efficiency of wait-and-see methods like those presented here should become increasingly important.</Paragraph>
  </Section>
class="xml-element"></Paper>