<?xml version="1.0" standalone="yes"?>
<Paper uid="P86-1013">
  <Title>PARSING CONJUNCTIONS DETERMINISTICALLY</Title>
  <Section position="7" start_page="80" end_page="80" type="evalu">
    <SectionTitle>
RESULTS
</SectionTitle>
    <Paragraph position="0"> Using the rules described above, NEXUS can successfully parse all the conjunction examples given in all the papers, with two exceptions. It cannot parse: * conjoined adverbs, e.g., Slowly and stealthily, he crept toward his victim.</Paragraph>
    <Paragraph position="1"> * embedded clausal complement gaps, e.g., Max wants to try to begin to write a novel and Alex a play.</Paragraph>
    <Paragraph position="2"> The problem with these forms lies not so much in the conjunction rules as in the rules for adverbs and clausal complements in general. These latter rules simply aren't very well developed yet. It is instructive to compare the NEXUS parser to that of Lesmo &amp; Toraseo. Like theirs, NEXUS solves the first problem mentioned in the introduction by using transition rules rather than a more conventional declarative grammar. Also like theirs, NEXUS solves the third problem by means of special rules which detect gaps in conjuncts and which fill those gaps by copying constituents from the other conjunct. Unlike theirs, however, NEXUS delays recombination decisions as long as it can and so does not have to search for possible attachments in some situations where theirs does. For instance, in processing Henry repeated the story John told Mary and Bob told Ann his opinion.</Paragraph>
    <Paragraph position="3"> their parser would first mis.attach \[and Bob\] to \[Mary\], then misattach \[and Bob told Ann\] to \[John told Mary\]. Each time, a search would be made to find a new attachment when the next word of the input was read. NEXUS can parse this sentence successfully without any mis-attachments at all.</Paragraph>
    <Paragraph position="4"> It is also instructive to compare NEXUS to the work of Church. His thesis \[3\] gives a detailed specification of a some fairly elegant rules for conjunction (and several other constructions) along with their linguistic and psycholinguistic justification. While most of the rules are not actually exhibited, their specification suggests that they are similar in many ways to those in NEXUS.</Paragraph>
    <Paragraph position="5"> However, Church was primarily concerned with the implications of determinism and limited memory, and so his parser, YAP, does not defer decisions as long as NEXUS does. Hence, YAP could not find, or ask for resolution of, the ambiguity in a sentence like: I know Bob and Bill left. YAP parses this as \[I know Bob\] and \[Bill left\]. NEXUS would find both parses because the third and fifth verb rules both apply when the verb left is processed. Note that these two parses are required not because of the conjunction, but because of the verb know, which can take either a noun phrase or a clause as its object. Only one parse would be needed for unambiguous variations such as I know that Bob and Bill left and I know Bob and Bill knows me. In general, the conjunction rules do not introduce any additional nondeterminism into the grammar beyond that which was there already.</Paragraph>
    <Paragraph position="6"> With respect to efficiency, the table below gives the execution times in milliseconds for NEXUS's parsing of the sample sentences tabulated in \[5\]. For comparison, the times from \[5\] for MSG and RPM are also shown. All three systems were executed on a Dec.20 and the times shown for each are just the time taken to build parse trees: time spent on morphological analysis and post-parse transformations is not included. MSG and RPM are written in Prolog and NEXUS is written in Maclisp (compiled).</Paragraph>
    <Paragraph position="7"> NEXUS was run with the 'no-interpreter' switch turned on.</Paragraph>
    <Paragraph position="8"> Sample Sentences MSG RPM NEXUS Each man ate an apple and a pear. 662 292 112 John ate an apple and a pear. 613 233 95 A man and a woman saw each train. 319 506 150 Each man and each woman ate an apple. 320 503 129 John saw and the woman heard a man that laughed. 788 834 275 John drove the car through and completely demolished a window. 275 1032 166 The woman who gave a book to John and drove a car through a window laughed. 1007 3375 283 John saw the man that Mary saw and Bill gave a book to laughed. 439 311 205 John saw the man that heard the woman that laughed and saw Bill. 636 323 289 The man that Mary saw and heard gave an apple to each woman. 501 982 237 John saw a and Mary saw the red pear. 726 770 190 In all cases, NEXUS is faster, and in the majority, it is more that twice as fast as either other system. Averaging over all the sentences, NEXUS is about 4 times faster than RPM and 3 times faster than MSG.</Paragraph>
  </Section>
class="xml-element"></Paper>