<?xml version="1.0" standalone="yes"?>
<Paper uid="C90-2039">
  <Title>Sign NonEmpty Empty V N P ADV</Title>
  <Section position="4" start_page="225" end_page="226" type="metho">
    <SectionTitle>
4. The Lazy Incremental Copy Graph Unification Method
</SectionTitle>
    <Paragraph position="0"> In Wroblewski's method, copying unique label arc values whole in order to treat cases like \]Pig. 5 disables structure sharing, ttowever, this whole copying is not necessary if a lazy evaluation method is used. With such a method, it is possible to delay copying a node until either its own contents need to change (e.g., node G3/Ka c  !7&gt;) or until it is found to have an arc (sequence) to a node t, hat needs to be copied (e.g., node X G3/&lt;a c&gt; in Fig. 5 due to a change of node Y G3/&lt;a c g&gt;). To achieve this, I, he LING unification method, which uses copy dependency information, was developed.</Paragraph>
    <Paragraph position="1"> The LING unification procedure uses a revised CopyNode procedure which does not copy structures immediately. The revised procedure uses a newly introduced slot COPY-DEPENDENCY. The slot has pairs consisting of nodes and arcs as its value. The revised CopyNode procedure takes as its inputs the node to be copied node I and the arc arc I with node I as its value and node2 as its immediate ancestor node (i.e., the arc's initial node), and does the following (set Fig. 7):  (1) if nodel ', the dereference result of node/, is current, then CopyNode returns node l&amp;quot; to indicate that the ancestor node node2 must be coiffed immediately; (2)otherwise, CopyArcs is applied to node1&amp;quot; and if it returns ,~;everal arc copies, CopyNode creates a new copy node. It then adds the arc copies and arcs of node/' that are not copied to the new node, and returns the new node; (3) otherwise, CopyNode adds the pair consisting of the ancestor node node2 and the are arcl into the COPY-DEPENDENCY slot of node 1&amp;quot; and returns Nil_.  ,',:opyArcs applies CopyNode to each arc value with node l' as the new ancestor node and returns the set of new arcs for non-Nil_ CopyNode results.</Paragraph>
    <Paragraph position="2"> When a new copy of a node is needed later, the LING unification procedure will actually copy structures using the COPY-DEPENDENCY slot value of the node (in GetOutNode procedure in lJ'ig. 6). It substitutes arcs with newly copied nodes for existing arcs. That is, antecedent nodes in the COPY-DEPENDENCY values are also copied.</Paragraph>
    <Paragraph position="3"> In the above explanation, both COPY-DEPENDENCY and COPY slots are used for the sake of simplicity.</Paragraph>
    <Paragraph position="4"> \]lowever, this method can be achieved with only the COPY slot because a node does not have non-NIL COPY-I)EPENDENCY and COPY values simultaneously.</Paragraph>
    <Paragraph position="5"> The data in the COPY-DEPENDENCY slot are I;emporary and they are discarded during an extensive process such as analyzing a sentence, ttowever, this does not result in any incompleteness or in any partial analysis structure being test. Moreover, data can be accessed in a constant order time relative to the number of DG nodes and need not be reconstructed because this method does not use a data structure consisl, ing of ,';keleton and environments as does Pereira's method.</Paragraph>
    <Paragraph position="6"> The efficiency of the LING unification method depends on the proportion of newly created structures in the unification result structures. Two worst eases can be considered: (t) If there are no arcs whose labels are unique to an input node witlh respect to each other, the procedure in LING unification method behaves in the same way as the procedure in the Wroblewski's method.</Paragraph>
    <Paragraph position="7"> (2) In the worst eases, in which there are unique label arcs but all result structures are newly created, the method  has the disadvantage of treating copy dependency information.</Paragraph>
    <Paragraph position="8"> However, these two cases are very rare. Usually, the number of features in two input structures is relatively small and the sizes of the two input structures are often very different. For example, in Kasper's disjunctive feature description unification, a definite part \[&amp;quot;S is larger than a disjunet definite part t&amp;quot;S.</Paragraph>
  </Section>
  <Section position="5" start_page="226" end_page="227" type="metho">
    <SectionTitle>
5. The Strategic Incremental Copy Graph Unification
</SectionTitle>
    <Paragraph position="0"/>
    <Section position="1" start_page="226" end_page="227" type="sub_section">
      <SectionTitle>
Method
</SectionTitle>
      <Paragraph position="0"> In a system where FS unification is applied, there are features whose values fail relatively often in unification with other values and there are features whose values do not fail so often. For example, in Japanese sentence analysis, unification of features for conjugation forms, case markers, and semantic selectional restrictions tends to fail but unification of features for semantic representations does not fail. In such cases, application of the EFF strategy, that is, treating features tending to fall in unification first, reduces unnecessary computation when the unification finally fails. For example, when unification of features for case markers does fail, treating these features first avoids treating features for senmntic representations. The SING unification method uses this failure tendency infornmtion.</Paragraph>
      <Paragraph position="1"> These unification failure tendencies depend on systems such as analysis systems or generation systems.</Paragraph>
      <Paragraph position="2">  Unlike the analysis case, unification of features for semantic representations tends to fail. in this method, theretbre, the failure tendency information is acquired by a learning process. That is, the SING unification method applied in an analysis system uses the failure tendency information acquired by a learning analysis process.</Paragraph>
      <Paragraph position="3"> in the learning process, when FS unification is applied, feature treatment orders are randomized for the sake of random extraction. As in TFS unification, failure tendency information is recorded in terms of a triplet consisting of the greatest lower bound type symbol of the input TFSs' type symbols, a feature and success/failure flag. This is because the type symbol of a 'rFS represents salient information on the whole TFS.</Paragraph>
      <Paragraph position="4"> By using learned failure tendency information, feature value unification is applied in an order that first treats features with the greatest tendency to fail. This is achieved by the sorting procedure of common label arc pairs attached to the meet type symbol. The arc pairs obtained by the SharedArcs procedure are sorted before treating arcs.</Paragraph>
      <Paragraph position="5"> The efficiency of the SING unification method depends on the following factors: (1) The overall FS unification failure rate of the process: in extreme cases, if Go unification failure occurs, the method has no advantages except the overhead of feature unification order sorting. However, such cases do not occur in practice.</Paragraph>
      <Paragraph position="6"> (2) Number of features FSs have: if each FS has only a small number of features, the efficiency gain from the SING unification method is small.</Paragraph>
      <Paragraph position="7"> (3) Unevenness of FS unification failure tendency: in extreme cases, if every feature has the same unification failure tendency, this method has no advantage. However, such cases do not occur or are very rare, and for example, in many cases of natural language analysis, FS unification failures occur in treating only limited kinds of features related to grammatical agreement such as number and/or person agreement and semantic selectional constraints. In such cases, the SING unification method obtains efl\]ciency gains.</Paragraph>
      <Paragraph position="8"> The above factors can be examined by inspecting failure tendency information, from which the efficiency gain from the SING method can be predicted. Moreover, it is possible for each type symbol to select whether to apply feature unification order sorting or not.</Paragraph>
    </Section>
  </Section>
class="xml-element"></Paper>