<?xml version="1.0" standalone="yes"?>
<Paper uid="C82-2029">
  <Title>CONSTRAINTS ON NOUN PHRASE CONJUNCTION: A DOMAIN-INDEPENDENT MECEAZI~</Title>
  <Section position="1" start_page="0" end_page="0" type="metho">
    <SectionTitle>
CONSTRAINTS ON NOUN PHRASE CONJUNCTION: A DOMAIN-INDEPENDENT
</SectionTitle>
    <Paragraph position="0"> expressing complex information. To provide a habitable language processing system, most applications require some treatment of conjunction. When a rich conjunction mechanism is introduced to account for the wide variety of conJotnings found in continuous text (as opposed to queries), it also becomes crucial to restrict the parses obtained for the o0nJoinings, since con~nction is a major source of ambiguity in parsing. For example, the sequence of syntactic classes:</Paragraph>
  </Section>
  <Section position="2" start_page="0" end_page="0" type="metho">
    <SectionTitle>
NO~N PREPOSITION NOUN and NOUN
</SectionTitle>
    <Paragraph position="0"> has two possible pareeez</Paragraph>
  </Section>
  <Section position="3" start_page="0" end_page="0" type="metho">
    <SectionTitle>
NOUN I~%EPOSITION (NOUN and NOUN)
</SectionTitle>
    <Paragraph position="0"> swelling of (hands and feet) I NOUN PREPOSITION NOUN) and NOUN swelling of hands) and fever The choice of the correct parse depends on applyin~ domain-specific constrs/nts to determine which words &amp;quot;go together&amp;quot; best. The concept of oonJunotional compatibility depends on semanti~ psxalleliam between the conjoined elements! thus, in the first example, hands and feet are more semantically parallel than swellln~ and feet. Although it is possible to write special rules to define allowable con~oln~ngs for a partlaular domain of application, this would mean that the - 129 system would not be portable: the conjunction constraints would have to be rewritten for each new domain. It is, however, possible to formulate a general domain-independent principle to capture conjunction constraints. This paper will describe such a principle, implemented as &amp;quot;conjunction restrictions&amp;quot; within the general treatment of conjunction in the natural language processing system of the Linguistic String Project parser. The illustrative examples are drawn from actual text occurrences in hospital discharge sunnnaries, processed using the mechanisms described below, as well as from some building code specifications, also processed with the LSP system.</Paragraph>
    <Paragraph position="1"> The first rule of conJunctional compatibility is simply: Two conjoined nouns must belong to the same (domainspecific) semantic class, This rule will produce the correct parse for the two previous examples, swellin~ of hands and feet and swelli~ of hands and fever. Hands and fee_._~t are both BODY-PART words in the first example, while swellin~ is not; in the second example, swellin~ and feve.._...rr are both SIGN.SYMPTOM words.</Paragraph>
    <Paragraph position="2"> In order to determine conJunctional compatibility, the noun conjuncts are compared pairwise! each conjunct is initially assigned a list consisting of its semantic class(es). (Most words belong to Just one semantic class! however, certain lexically ambiguous wo~ds belong to two or more classes, as illustrated below.) The intersection of the two lists of semantic classes is then computed. If the intersection is not empty, then the conjoining is allowed, and the intersection replaces the original list of classes associated with each conjunct. This serves as a record of a~v disambignation that occurs as a result of computing oonJunctionsl compatibility.</Paragraph>
    <Paragraph position="3"> For example, discharge is lexically ambiguous in medical narrative: it belongs to the ~DICAL-ACT class, as in:  Thus oomputin~ conjunction compatibility also provides a record of the correct reading when a conjunct is lexical!y ambiguous.</Paragraph>
    <Paragraph position="4"> The conJuBction rule does not account for the fact that certain noun phrases have a semantic class different fr.-.</Paragraph>
    <Paragraph position="5"> that of their head noun. Consider the following phrase from a building code text: bulldin~s and portions of buildings</Paragraph>
  </Section>
  <Section position="4" start_page="0" end_page="0" type="metho">
    <SectionTitle>
BLDG PART BLDG
</SectionTitle>
    <Paragraph position="0"> In this case, the head noun~ portion is &amp;quot;transparent&amp;quot; to its prepositional modifier, of buildln~se That is, the phrase portion of buildln~s really has a distribution characteristic of the word buildlnK. If we could disregard the word portions of and look instead at the noun in the prepositional phrase, namely buildin~ with class BLDG, then the conjunction rule would apply correctly. We call this type of construction the &amp;quot;computed attribute&amp;quot; oonstructlon, because, in its most general form, the semantic class for an entire phrase is computed from the interaction of the semantic class of the head noun and the olass(es) of its modifiers.</Paragraph>
    <Paragraph position="1"> The computed attribute construction turns out to be characteristic of a set of English semantic classes common to ma~7 domains. These classes include the classes AMOUNT, PERIOD, BEGINNING, ENDING, FREQUENCY, etc. To handle these - 131constructions, an additional mechanism has been added as a restriction to the grammar. It operates on a word ~ belo~LUg to a transparent word class modified by an appropriate prenomiual noun or adjective modifier (edegg. buildin~ portion), or by an appropriate prepositional phrase modifier (e.g,, portions of ~ldin~s). It assigns to the phrase as a whole the class of the modifier and binds the modifier to the head noun, so that the modifier does not distribute over the other con~oined noun(s)deg This produces only the ps.Tse buildings and (portions of buildings)</Paragraph>
  </Section>
  <Section position="5" start_page="0" end_page="0" type="metho">
    <SectionTitle>
.BLDG P~RT BLDG
BLDG 4= COMPUTED ATTRIBUTE
</SectionTitle>
    <Paragraph position="0"> &amp;quot;and eliminates the parse (buildings and portions) of buildings, The computer attribute is recorded as a special COMPUTED-ATTRIBUTE list on the head noun; it lists the semantic olass(es) associated with the entire ~ The COMPUTED-ATTRIBUTE llst is assigned before conJunetional compatibility is checked, whloh allows the conjunction meahanism to make use Of a COMPUTED-ATTRIBUTE, if one is presentdeg On the basis of a computed attribute and con~unctlon compatibility, the following sentence received exactly one analysis: There was no cyanosis and no history of prior seizures (SIaW-S~PTOM) (PERIOD) (SIOS-S~-</Paragraph>
    <Paragraph position="2"> A preliminary test on a 22-sentence p~ph from a hospital discharge s~ was run to compare the. parses obtained with and without the above con~unction and computed attribute mechanisms. Five sentences in the paragraph conteined conjoined noun phrases. These mechanisms reduced the total number of parses obtained for the five sentences almost in half: 8 parses total with the conjunction and computed attribute mechanisms compared to 14 parses total without theme - 132 The com~unction and computed attribute mechanisms do not resolve all of the ambiguities of oomJoined noun phrases. In pe~tieular, they do not always resolve the issue of distribution of right and left modifiers. However, they substantially reduce the number of incorrect parses and they require no modificstion for application to new domains.</Paragraph>
    <Paragraph position="3"> - 133 -</Paragraph>
  </Section>
class="xml-element"></Paper>