<?xml version="1.0" standalone="yes"?>
<Paper uid="C92-1030">
  <Title>An Empirical Study on Rule Granularity and Unification Interleaving Toward an Efficient Unification-Based Parsing System</Title>
  <Section position="3" start_page="0" end_page="0" type="metho">
    <SectionTitle>
2 The Granularity of Phrase
</SectionTitle>
    <Paragraph position="0"/>
    <Section position="1" start_page="0" end_page="0" type="sub_section">
      <SectionTitle>
Structure Rules
2.1 Granularity
</SectionTitle>
      <Paragraph position="0"> Phrase structure rule granularity has been introduced to refer to the amount of linguistic constraints specified in the atomic CFG phrase structures rules without annotations. The rule granularity spectrum has been classified into four categories as shown in Table 1, using the number of grammar rules ms a ruessure. null Unification-based grammars, in general, are characterized by a few general annotated pbrase structure rules, and a lexicon with specific linguistic descriptions. This is especially true for HPSG \[Pollard and Sag, 87\] and JPSG \[Gunji, 87\], which are to be categorized as extremely-coarse grained, as they drastically reduce the nmnber of phrase structure rules into two for English and one for Japanese, respectively. In these frameworks, the only role of the phrase structure rules is to provide a device for combining a head with its complement. Most linguistic constraints are stored in the feature descriptions.</Paragraph>
      <Paragraph position="1"> Coarse-grained rules have been characterized as a grammar consisting of atonfic phrase structure rules with medium constraints, and feature descrip tions with strong constraints. Medium-grained rules have been characterized as a grammar consisting of atomic phrase structure rules witb strong constraints, and feature descriptions with mediuln constraints. Medium-grained rules differ from coarse-grained rules in that they include morpho~syntax in the phrase structure rules, while coarse-grained rules include them in the feature descriptions. This means that medium-grained rules are strong enough to derive syntactic structures from atomic phrase structure rules without feature descriptions.</Paragraph>
      <Paragraph position="2"> Grammars for conventional NLP systems using simple or augmented CFG fall into the category of fine-grained rules, which represent most of linguistic constraints as CFG phrase structure rules, and the number of rules usually amounts to an intractable number of several thousands for practical applications. null</Paragraph>
    </Section>
    <Section position="2" start_page="0" end_page="0" type="sub_section">
      <SectionTitle>
2.2 Maintainability and Efficiency
</SectionTitle>
      <Paragraph position="0"> In unification-based framework, a linguistic constraint cart either be described as atomic context-flee phrase structure rules, or as feature descriptions in annotations and lexical entries. As the number of atomic phrase structure rules decreases, the number of feature descriptions increases.</Paragraph>
      <Paragraph position="1"> It is true that the lexieo-syntactic approach makes tile granunar modular and improves its maintainability by reducing the number of rules. However, it must be noted that the computational cost of disjunctive feature structure unification, in the worst ease, is exponential in the nmnber of disjunctions \[Kasper, 87\], whereas tile cost of CFG parsing is o(N s) in the input length N. Therefore, extreme rule reduction results in inefficiency. This overwhelms the benefits of the maintainability of the reduced number of rules since grammar development is essentially a trial-and-error process and requires a short turn-around time. However, the cost for CFG parsing also increases as the number of rules increases. Therefore, we must chose tile granularity so that the reduction in unification cost outweighs tile increase in CFG parsing cost, in order to gain overall etfieiency.</Paragraph>
    </Section>
  </Section>
  <Section position="4" start_page="0" end_page="0" type="metho">
    <SectionTitle>
3 The HPSG-Based Japanese
Grammars
</SectionTitle>
    <Paragraph position="0"> In this section, we illustrate the difference between &amp;quot;coarse-grained&amp;quot; rules and &amp;quot;medium-grained&amp;quot; rules using our HPSG-based spoken-style Japanese grarnlnal'S as an exaluple, We have developed two unification-based grammars with different granularity l, which are essentially based on tIPSG and its application to Japanese (JPSG), for the analysis module \[Nagata and Kogure, 90\] of an experimental Japanese-to-English speech-to-speech translation system (SL-TRANS) \[Morimoto et al., 90\].</Paragraph>
    <Paragraph position="1"> We have selected the &amp;quot;secretarial service of an international conference registration&amp;quot; as our task domain, in which a conversation between a secretary and a questioner is carried out. Tile Japanese grauunars~ however, ~tre not task-specific, but rather general-purpose OlleSj which cover a wide range of pllenonl-I Historically speaking, we fil~t developed coarse-grained rules &amp;lid then we nlallllally tl'al|sfonned them into mediumgrldned rules for e|licicncy.</Paragraph>
    <Paragraph position="2">  ACTES DE COL1NG-92, NAmes, 23-28 hotyr 1992 1 7 8 PRoc. OF COLING-92, NANTES, AUG. 23-28, 1992 ena at ruazly linguistic levels from syntax, and seman tics, to pragmatics using typcd feature structure descriptions. The linguistic phenomena covered in these grammars include: * l,'undamental Constructions: causative, passive, benefactive, negation, interrogative, etc., * Control and Gaps: subject/object control, * Unbounded Dependencies: topic, relative, * Word Order Variation and Ellipsis.</Paragraph>
    <Section position="1" start_page="0" end_page="0" type="sub_section">
      <SectionTitle>
3.1 Coarse-Grained Rules vs.
Medium-Grained Rules
</SectionTitle>
      <Paragraph position="0"> The coarse-grained HPSG-based Japanese grammar has about 20 generalized phrase structure rules, while the medium-grained grarmnar has about 200 phrase structure rules. Both gra, lnmars use the same lexicon with a vocabulary of about 400. ~ In the coarse-grained grarmnar, phrase structure rules only refer to the relative position l)etween the five basic syntactic categories for Japanese: verb (V), noun (N), adverb (ADV), postposition (P), and attributive (ATT). Most of the specific linguistic information is encoded as feature descriptions in either the annotation of the l)hrase structure rules or the lexical entries. In principle, there is no distinction as to whether a constituent is lexical or phrasal, and no subcategories of the 5 basic categories. This contributes greatly to the reduction in the numbcr of phrase structure rules, which results in better grammar maintainability. We present all the phrase structure rules of the coarse-grained Japanese grammar in Appendix A.</Paragraph>
      <Paragraph position="1"> It has been noticed that the extensive use of dis-junctions in feature descriptions, which results from the reduction of the number of phrase structure rules, is the main cause of incfficieney in the coarse-grained version of the grammar. The three major sources of disjunctions are, lnorpho-syntactic specifications for diverse expressions in the final part of the sentence, frec word order and ellipsis of verb complements (subeat slash scrambling), and semantic interpretation of deep case and aspect, where the first two particularly are the problems in spoken-style Japanese.</Paragraph>
      <Paragraph position="2"> We have manually converted the coarse-grained phrase structure rules into medium-grained rules to reduce thc computational cost of unilication. First, we divided each of the basic categories into several subcategories. Then, we divided the coarse-grained phrase structure rules according to the subcategorics.</Paragraph>
      <Paragraph position="3"> qb kee I) the grammar readable, however, we choose to leave the subcat slash scrambling and the semantic  tion module \[Takez.awa el, ~xl. , 911. It only imea atoaalc CFG la31esp a31d the \]lulll}~r of rules ~llOUll{S to Inol~ thall 2,0(~, It is, thcrefore~ categolJzed ~-s a tin,grained gr~Htnar in our defiltition.</Paragraph>
      <Paragraph position="4"> interpretation undone, and nmde extensive efforts to expand the morpho-syntactic speeificatioas.</Paragraph>
    </Section>
    <Section position="2" start_page="0" end_page="0" type="sub_section">
      <SectionTitle>
3.2 Example: Medium-Grained Rules
for Predicate Verb Phrases
</SectionTitle>
      <Paragraph position="0"> In this section, we illustrate the process of transfof marion using a predicate verb l)hrasc production rulc as an example. Japanese predicate phrases consist of a main verb followed by a sequence of auxiliaries azld sentence final particles. There is an ahnost ottodimensional order of verbal constituents such as in Figure l, which reflects the basic hierarchy of the J apanese sentence structure.</Paragraph>
      <Paragraph position="1"> Kernel verbs occur first in a predicate phrase sequence. Voice auxiliaries precede all other auxiliaries, and within this category, the causative auxiliary (sa)se,'u precedes the passive auxiliary (ra)re~t. Aspect auxiliaries, such a.s the progressive auxiliary (Ie)ivu precede modal auxiliaries ;rod follow voice auxiliaries. Modal auxiliaries are classified into two groups with respect to the relative order of negative and tense auxiliaries. Mood1 iuehldes the optative arlxiliaries, such as tai (want), beki (should/must), etc. Mood2 includes the evidential or inferential auxiliarics such as rashii (seem/look), kamoshirenai (may), etc. Negative auxiliaries uai, u (not) follow voice, aspect, and mood l auxiliaries, and precede tense and mood2 auxiliaries. Tease auxiliaries la, da (-ed) show irregular behavior. They follow the voice, aspect, mood1, and negative auxiliaries, and precede the mood2 auxiliaries. They also can tollow the mood2 auxiliaries.</Paragraph>
      <Paragraph position="2"> in the coarse-grained grammar, we provide a single phrase structure rule for the phcnomena.</Paragraph>
      <Paragraph position="3"> v~(v AUXV) O) The order constraints between auxiliaries are specified in the annotution of rule (1) and each lexi cal entry by the combination of tile syntactic features, such as the synlheadlsubcat for preceding constituents, tile synlheadlcoh for following constituents, and the syn\[headlroodl for the position of the con stituent~ in the verb phrase hierarchy. For example, the causative auxiliary verb sern has the following, feature bundles in its syn{headlmodl feature.</Paragraph>
      <Paragraph position="5"> In converting the rule, first we have claasitied the verbal phrasal categories according to the hierar.</Paragraph>
      <Paragraph position="6"> ehy, e.g. V-kernel, V-aspect, V-moodl, V-negt, Vmood2~ and V-tease, then we have subcategorized the auxiliari~ as shown in Table 2. Thus, the coarse-grained phrase structure rule (1) is converted to the 32 medium-grained grammar rules in Appendix B ACRES DE COLING-92. NANTES, 23-28 AOt~W 1992 1 7 9 Prtoc. OF COLING-92, NANTI~S, AUG. 23-28, 1992 kernel &lt; voice &lt; aspect &lt; moodl &lt; negate &lt; tense &lt; mood2 &lt; tense (sa)seru (te)iru tai nai ta rasii ta (ra)reru (te)morau tagaru n da desu u  causative auxiliary: (sa)sei'u passive auxiliary: (ra)reru aspect auxiliary: (~e)iru, Oe)a(u benefactive auxiliary: (le)tnorau AUXV-optt optative auxiliary: tai, beki AUX%negt negative auxiliary: nai, n AUXV-tense tense auxiliary: la, da AUXV-evid evidential auxiliary: rashii, darou AUXV-copl copulative auxiliary: da, desu</Paragraph>
    </Section>
  </Section>
  <Section position="5" start_page="0" end_page="0" type="metho">
    <SectionTitle>
4 Interleaving CFG Parsing
</SectionTitle>
    <Paragraph position="0"/>
    <Section position="1" start_page="0" end_page="0" type="sub_section">
      <SectionTitle>
and Unification
4.1 Strategies for Evaluating Feature
Descriptions
</SectionTitle>
      <Paragraph position="0"> Unification is an expensive operation, so the point of evaluating feature descriptions during CFG parsing has serious affects on the overall performance. We have implemented two strategies for feature descrip- null descriptions are evaluated when a complete CFG parse is found. The &amp;quot;well-formedness&amp;quot; of a parse derived from atomic CFG rules is verified by evaluating associated feature descriptions.</Paragraph>
      <Paragraph position="1"> The granularity of the phrase structnre rules is closely related to the proper selection of the evaluation strategy. Since the atomic phrase structure rules ill the coarse-grained grammar are not so strong as to constrain syntactic structures, we have to employ the early unification to avoid a nnmber of irrelevant subparses which should have been eliminated by the evaluation of annotations. IIowever, since the atomic rules in the medium-grained grammar have detailed morpho-syntax specifications, they should be able to avoid irrelevant copies by using the late unification.</Paragraph>
    </Section>
    <Section position="2" start_page="0" end_page="0" type="sub_section">
      <SectionTitle>
4.2 Implementing the Evaluation
Strategies
</SectionTitle>
      <Paragraph position="0"> We have implemented the various evaluation strategies by doing additional housekeeping in the underlying parser. The parser used here is called the Typed</Paragraph>
    </Section>
    <Section position="3" start_page="0" end_page="0" type="sub_section">
      <SectionTitle>
Chart Parsing Algorithm
</SectionTitle>
      <Paragraph position="0"> Feature Structure Propagation Parser (TFSP Parser) \[Kogure, 89\], which is based on the active chart parsing algorithm \[Kay, 80\] and typed feature structure unification \[Ait-Kaci, 86\].</Paragraph>
      <Paragraph position="1"> Tile active chart parser and the unification algorithm are implemented in C on Sun4, which is a 10-MIPS work station. The unification algorithm is based oil nondestructive graph unification \[Wroblewski, 87\], which we extend to treat negation, loop, type symbol subsumption relationships, and disjunctiou. Successive approximation \[Kasper, 87\] is used for disjunctive feature structure unification.</Paragraph>
      <Paragraph position="2"> The Active chart parsing algorithm basically consists of chart initialization and iterative rule invocation. The basic part of the iterative rule invocation is shown in Figure 2. AcpContinue checks the suspending condition and calls rule invocation recursivcly. AcpOneStep carries out a cycle of rule invocation which consists of getting a new pending edge (GetPendingEdge), adding it to the chart (AddEdge), combining active and inactive edges (TryToContinue-ActiveEdge/TryToContinuelnactiveEdge), and proposing new edges (ProposeProductions). The parser stops (SatisfySuspendingCondition?) when it finds an inactive edge whose starting and ending vertex are the left-most and right-most vertex of the chart, respectively, and whose label is the start symbol of the granunar.</Paragraph>
      <Paragraph position="3"> PROC. OF COLING-92, NANTES, AUG. 23-28, 1992 In early unification, the feature descriptions are evaluated when the edges are combined, while in late unification, they are evaluated iu the chart suspending condition check only if tile clmrt suspending condition bolds. Delaying unification is implemented by adding a slot edge.parse to the edge structure, which keeps a list of the pair of active aud inactive edges constructing the edge. If either or both of the argument feature structures of the unification have not been evaluated, they are recursively evaluated to get the target feature structure.</Paragraph>
      <Paragraph position="4"> It has to bc j~oted that some derivations that termhmte when feature descriptions are evaluated, may not terminate if they are ignored. For example, it is possible to write a rule for unbounded dependency like (2), in which m~ element in tile subcat feature is moved to the slash feature, to introduce slashed categories dynamically 3.</Paragraph>
      <Paragraph position="6"> Ignoring feature descriptions in the rule may cause aal infinite loop. Therefore, feature descriptions arc forced to be evaluated, when rules that cause a loop are encountered in late unification.</Paragraph>
    </Section>
  </Section>
class="xml-element"></Paper>