<?xml version="1.0" standalone="yes"?>
<Paper uid="C94-2147">
  <Title>Table-driven Neural Syntactic Analysis of Spoken Korean</Title>
  <Section position="3" start_page="0" end_page="913" type="metho">
    <SectionTitle>
2. Features of spoken Korean
</SectionTitle>
    <Paragraph position="0"> Korean, a SOV and an agghmating language, has the following characteristics: 1) A Korean word, Eojeol, consists of more than one ntorphemes with clear-cut boundaries in between. For example, an Eojeol &amp;quot;pka.il-dulul(fiies\[objl)&amp;quot; consists of 3 ,norl~hemes: pha-il = il + dul + ul filcslobjl file \[pl. suffix\] \[obj markerl 2) Korean is a postposifional language with nounendings, verb-endings and prefinal verb-endings. These functional morphemes determine the noun's case role, verb's tense, modality, and modification relations between phrases. For example, in &amp;quot;e-cey pha-il(the file :i~:~: ::.:,.</Paragraph>
    <Paragraph position="1"> :: ~ii!~i :~:~ yesterday)&amp;quot; the verb &amp;quot;swu-ceng (edit)&amp;quot; is of past tense and modifies &amp;quot;pha il (file)&amp;quot;: su-ceng-ha yess ten pha-il edit \[past\] \[adnominal\] file 3) Korean has relatively fi'ce word order compared to SVO hmguages, such as English. For example, the sentence ~ ~ ~-ha-vess-ten pha-il-tul-ul/tn~pok-sa-ha-ye-la (Copy the files that was edited by me yesterday to /trap.)&amp;quot; may be written as &amp;quot;e-cey. na!~dca swu-cen~ess-ten &amp;quot; or &amp;quot;~10 e-cey_ naok-ka swu-cen ,-hgzhaeyess-ten .., .&amp;quot;  Besides these characteristics of written Korean, spoken Korean has the following more characteristics: 4) The unit of pause in a speech (Eonjeol) may be different from that of a written text (Eojeol). For example, in speaking &amp;quot;nay-ka e-cey swu-ceng-ha-yess-ten phaoil-tul-ul ~trap lo pok-sa-ha-ye-la (spaces delimit Eojeols), a person may pause after &amp;quot;nay-ka&amp;quot; and after &amp;quot;e-cey swu-ceng-ha-yess-ten pha-il-tul-ul&amp;quot; and after &amp;quot;~trap lo pok-sa-ha-ye-la.&amp;quot; 5) Phonological changes occur in a morpheme, between morphemes in an Eojeol, and between Eojeols in an Eonjeol. These changes include assimilation, dissimilation, and contraction. For example, a morpheme &amp;quot;pok-sa&amp;quot; is pronounced as /pok-ssa/ and &amp;quot;yess&amp;quot; is pronounced as /yet/. An Eojeol &amp;quot;su-ceng-ha-yess-ten&amp;quot; is pronounced as/suceng-ha-yet-tten/. null 3. Table driven neural syntactic analysis This section explains interactive relaxation parsing of spoken Korean using neural network, its underlying grammar, and control mechanism.</Paragraph>
    <Paragraph position="2"> A sequence of candidate phonemes in phoneme lattice \[figure 1\] is assumed to be the output of the speech recognizer.</Paragraph>
    <Paragraph position="3"> \[Figure 1. Phoneme lattice\] A CYK-based morphological analyzer is used to extract a morpheme lattice from the phoneme lattice. In the morphological analysis, special procedural attachments resolve the phonological changes. The use of phoneme lattice gives the problem of exponential number of Eonjeol candidates. For this problem, trie data structure is used for the phonetic transcription-toorthographic morpheme dictionary (morphemelevel phonetic dictionary).</Paragraph>
    <Paragraph position="4"> 3.1. Extending the Categorial Grammar To model the syntax of Korean, we extended the Categorial Grammar in two ways (Zeevat  1988; Uszkoreit 1986).</Paragraph>
    <Paragraph position="5"> A (directional) Categorial Grammar is an ordered quintuple G = &lt;V, C, S, R, f&gt;, where 1) V: the vocabulary set, 2) C: a finite set of basic categories which generates a full set C' of categories via recursive  application of the following category furmation rules: if a~ C, then a~ C' and if a~C' and b~ C', then a/b~ C' and akbc C',  3) S: the category for sentences, 4) R: a set of functional application rules such as left cancellation * A B\A __&gt; B right cancellation&amp;quot; B/A A _.&gt; B 5) f: an assignment flmction from elements of V into the subsets of C'.</Paragraph>
    <Paragraph position="6"> To treat the free word-order in Korean, we extended the category formation rules and the application rules: 2') Extended category formation rules: ifa~C, then a~C' and ifa~C' and ScC', then a/s~ C' and a~ C' and 4') Extended functional application rules : left cancellation : Ai I~{AI,...,An}-&gt; ~{Al,...,Ai-l,Ai+l,...,Atl} right cancellation : B/{AI,...,An} Ai --&gt; B/{AI,...,Ai_ l,Ai+l,...,An} 3.2. Interactive relaxation parsing  (Howells 1988) developed an interactive relaxation parsing method which used a dynamic network building scheme, and decay over time with competition instead of explicit inhibitory links, which is similar to the (Reggia 1987)'s approach.</Paragraph>
    <Paragraph position="7"> The interactive relaxation algorithm consists of the following steps (Howells 1988): 1) add nodes, 2) spread actiwltion and 3) decay. Bottom-up information gathering and top-down expectations occur during the parsing.</Paragraph>
    <Paragraph position="8"> 1) to add a node: A grammar node is added tot&amp;quot; each sense of morphemes when the parsing begins. Statistical information on the senses of a morpheme determines the initial activation value of the senses.</Paragraph>
    <Paragraph position="9"> A grammar node which has more activation than the predefined threshold 6) makes new nodes (expectations). The newly generated nodes represent candidate parse trees containing the generator node.</Paragraph>
    <Paragraph position="10">  2) to spread actiwltion: A predefined portion, P, of a node's total activation, A, is passed upward, to bigger parse trees. When more than one destination nodes exist, they compete to get more actiwttions. A higher node with actiwltion a i gets the following amount of actiw~tions:</Paragraph>
    <Paragraph position="12"> A higher level node (with total actiwltion A) spreads a prcdefined portion (Q) of its actiwltion wdue equally to the constituents. When there are n constituents, a constituent gets the following</Paragraph>
    <Paragraph position="14"> A llOde's actiwttion value (A) after decay is A times (1 D), where I) is the decay ratio.</Paragraph>
    <Paragraph position="15"> Moreover, a node with less constituents than needed is penalized by the number of actual constituents (Ca) divided by the numher of required constituents (Cr). After all, a node's activation value changed to</Paragraph>
    <Paragraph position="17"> And finally a node whose actiwttion value is less than the predefmed threshold * is removed.</Paragraph>
    <Paragraph position="18"> 3.3. CYK-table-driven control The interactive relaxation parsing scheme (Howells 1988) lacks efficient control structures for constituent searching and expectation generation. We provided the positional information through the CYK-tablc and the structuring information through the Categorial Grammar formalism. Using Categorial Grammar makes the parse tree be a binary tree, not a general n-ary tree.</Paragraph>
    <Paragraph position="19"> All the grammar nodes reside in a CYK-table.</Paragraph>
    <Paragraph position="20"> The position (i,j) in the table explicitly says where to find the constituents, where to add new expectations and what the uode there stands for. A node in CYK(i,j) represents a parse tree for the input segmeut from i to j. A node in CYK(i,j) with category P, called P(i,j), can be nsed in 3 ways to construct larger parse trees:</Paragraph>
    <Paragraph position="22"> becomes a constituent of Q(k,j). Or it combins with Q\p(j+l,k) for some k &gt; j and becomes a constituent of Q(i,k). In these two cases P(i,j) is used as an argument of the limctional categories.</Paragraph>
    <Paragraph position="24"> creates a larger parse tree A(i,k), when P = A/B. In this case, P(i,j) is a flmctional category A/B and searches ti)r an argument B on the right side.</Paragraph>
    <Paragraph position="26"> creates A(k,j) , when P=A\ B. In this case, P(i,j) is a functional category A\ B and searches for an argument B on the left side.</Paragraph>
    <Paragraph position="27"> 0 1 2 3 4 \[Figure 2. A/B( l,l)'s expectation generation\] The following scenarios can explain the CYK-table driven interactive relaxation parsing of Korean.</Paragraph>
    <Paragraph position="28"> A node A/B(i,j) whose actiwltion wdue is greater than the threshold (0 makes new nodes (A(i,k)'s for all j &lt; k &lt; input-length), and each A(i,k) looks for a constituent B(j+l,k). \[Figure 2\] shows A/B(I,I)'s expectation generations. There are 3 possible A's which can have A/B(1,1 ) as their constituents: A(I,2), A(I,3) and A(1,4). A(I,2) looks for 1/(2,2), A(I,3) for B(2,3), and A(1,4) for B(2,4).</Paragraph>
    <Paragraph position="29"> Synnnetrically, a node A\B(i,j) whose activation value is greater than the threshold O makes new nodes (A(k,j)'s for all 0 &lt; k &lt; i) and each A(k,j) looks for a constituent B(k,i-l).</Paragraph>
    <Paragraph position="30"> Only the nodes with enough actiw~tions generate hypotheses, and the hypotheses which lack of constituents disappear rapidly by the decay with penalty mechanism. Each node (hypothesis) which looks for a constituent drives  the parsing, and the efficiency of the constituent searching is guaranteed by the CYK-table. We call the parsing technique as Connectionist-CYK parsing.</Paragraph>
  </Section>
  <Section position="4" start_page="913" end_page="913" type="metho">
    <SectionTitle>
4. System architecture
</SectionTitle>
    <Paragraph position="0"> The Connectionist-CYK parser is incorporated into the DINX (Dialog Interface to UNIX) system which is under development in POSTECH (Lee, W. I. and Lee, G.B. 1993). \[Figure 3\] shows the part of D1NX system architecture. There are two modules in the system (for language analysis): morphological analyzer and the CCYK parser.</Paragraph>
    <Paragraph position="1">  A phoneme lattice for each Eonjeol is analyzed by an extension of the CYK-based morphological analyzer (Lee, E. C. 1992). The result of the analysis is a morpheme lattice (see \[Figure 4\]) for each Eonjeol. These morpheme lattices are linked, and the morpheme lattice for entire sentence is constructed. Grammar nodes for the senses of each morpheme are created in the CYK-table and the following four steps are repeated for a fixed number of iterations: 1) making hypotheses (new nodes), 2) constituent searching, 3) computing out-going activations and 4) updating activation values. Functional category nodes A/B(i,j) and A\B(i,j) whose activation values are greater than the threshold generate the expectations A(i,k)'s and A(k,j)'s respectively. All the hypotheses (nodes which lack of a constituent) search for their constituents. Outgoing bottom-up and top-down activations are computed for each node. Each node gathers incoming activations and decays.</Paragraph>
  </Section>
  <Section position="5" start_page="913" end_page="914" type="metho">
    <SectionTitle>
5. Sample run
</SectionTitle>
    <Paragraph position="0"> In this section, a detailed example of the Connectionist-CYK parsing is given. The system parameters are as follows: bottom-up activation ratio P = 0.87, top-down activation ratio Q = 0.39, decay ratio D = 0.495, expectation threshold (r) = 6.66 and remove threshold * = 0.66.</Paragraph>
    <Paragraph position="1"> The sentence &amp;quot;ci-wul su iss-nun pha-il-dul-ul po-ye-la (List the files which can be removed.)&amp;quot; is assumed to be spoken with two pauses, &amp;quot;ci-wul su iss-nun / pha-il-dul-ul / po-ye-la.&amp;quot; \]Figure 1\] showed the first phoneme lattice of the three.</Paragraph>
    <Paragraph position="2"> The phoneme lattices are analyzed one by one, and a morpheme lattice for the sentence is created by merging the morpheme lattice for each Eonjeol \]Figure 4\].</Paragraph>
    <Paragraph position="3"> \[Figure 4. A morpheme lattice\] Grammar nodes for the senses morpheme are created \[Figure 5 (a)\].</Paragraph>
    <Paragraph position="4">  Each dot represents a node and the darkness of a dot denotes the degree of activation. The morpheme lattice is embedded in the CYK-table, and the senses of each morpheme are created in the corresponding position in the table. Each nodc with functional category generates hypotheses, and each hypothesis searches for thcir constituents. After the 1-st iteration, 209 nodes are in the table. The number of nodes change to 282, 302, 289, and 265 along the iterations. After 6-th iteration, the number of nodes decreases to 253, and the correct parse tree for the whole morpheme latticc is created (\[Figure 5 (b)\] shows a part of the parse tree). After the 7-th iteration, the number of nodes decreases 191, 180, 163, 103, 98 ..... 78.</Paragraph>
    <Paragraph position="5"> Aftcr 30-th, thc correct parse trcc which covers the entire sentence is extracted.</Paragraph>
  </Section>
class="xml-element"></Paper>