<?xml version="1.0" standalone="yes"?>
<Paper uid="C92-4182">
  <Title>PRI~;DICTING NOUN~PIIRASE SURFACh; I~'ORMS USING Q~ONTEXTUAL \[NFORMA'PION</Title>
  <Section position="3" start_page="0" end_page="0" type="metho">
    <SectionTitle>
2 Dialogue Interpretation and Predicting
</SectionTitle>
    <Paragraph position="0"> the Next Utterance The next. utterance call be predicted after understanding the previous utterances, because predicted information must be affected by tile dialogue struc-AcrEs DE COLING-92, NANTES, 23-28 Aour 1992 1 1 5 2 PROC. OF COLING-92, NANTES, AUO. 23-28, 1992 lure. This section brielly describes tile model for interpreting a dialogue\[7\] and tile ulethod of pre(lictng the next uttcrance\[ll\]\[12\].</Paragraph>
    <Paragraph position="1"> In tile lnodel, an utterance is represented, by a predicate form. An typical Japanese sentence, &amp;quot;(;0juusyo we ouegai-shi-masu.&amp;quot; (May 1 have your ad dress?), uttered by the secretariat in a inquiry dialogue, is shown below: (ASK-VALUE u q (address q) (IS (address q) ?va\].)) where constant s denotes the secretariat, q the questioner, address tile concel)t of an a.ddress, and the variable, ?val, is the value for the address of q.</Paragraph>
    <Paragraph position="2"> The dialogue interpretation model h;us \[our types of plan and can interpret input utterances as the dialogue proceeds, using au extended plain inference mechanism\[7\]. Thus, a dialogue structure can be constructed. null lit order to provide contextual ieforlnation about discourse entities we use typed variable notation\[2\] to describe a discourse entity in a plan schema, l';ach type in this notation corresponds to a particular concept node ill the domaiu-del)eudent NP knowledge base (described in Section a.3) '\['he following description is an example of a Domailt Pla~J to send something:  The state of understanding is managed nsieg two pushdown stacks. The mlderstanding llst stores completed plans as the current understanding state, and tile goal list maintains incomplete plans ~Ls possibilities and expectations tbr fltture goals. By rc ferring to tile goal list., the next utl.erance can I)e predicted on an abstract, level as the dialogue proceeds, using the two generalized rules: eximctation mid prefere.nee\[12\].</Paragraph>
    <Paragraph position="3"> Predicted utterances are tel)resented in the slttnc style as intmt utterances. As a result,, we can predict two types of infornlatiou, one about the COlllllltlUiCalive act, types and the other about discourse entities in the propositional contents (or in the topic slot,) (t\[&amp;quot; the ilext uttcrarlce, lnforrnatioll abotlt a discourse entity lnay appear in the forlu of ;ill particular ex pression if it is ill a prvvious utterance that can be related to the. current atterauce. Othe.rwise infornl~ ties will be in the tbrm of a type represcuting a particular concept in tile related domain plan. We call such information conteztual information in tile t~sk of selecting ttle constituents of tile next utterance.</Paragraph>
  </Section>
  <Section position="4" start_page="0" end_page="0" type="metho">
    <SectionTitle>
3 NP Identification Model
</SectionTitle>
    <Paragraph position="0"/>
    <Section position="1" start_page="0" end_page="0" type="sub_section">
      <SectionTitle>
3.1 Chauge to NP Linguistic Expressions
</SectionTitle>
      <Paragraph position="0"> hi general, when we are talking about a concept X, there are many possible surface expressions and fornls to represent X. In llarticular, Japanese ha.s several possible SEFs for a given X, one from ttle Chinese reading and another based on tile original Japanese language (e.g. &amp;quot;oka~sakf' and &amp;quot;atesakf' for 'destination' in Fig. 1). In addition, there are particular phenomena of expression variations depending hi)on the particular dialogue. For exmnple, if a speaker is uttering his/her own address for tile concept 'address', he/she will use &amp;quot;3uusyo&amp;quot;(\[my\] address), e.g. &amp;quot;Juusyowa Oosaka-sht desu.&amp;quot;(My address is in Osaka~city.). On the other hand, if he/she is uttering tile other participant's address, he/she will use &amp;quot;.qo.juusyo&amp;quot;(\[your\] address (polite fornl)), e.g. &amp;quot;Go-juusyo-wo onegaishi-masu.&amp;quot;(Your address, please?) These facts lead rlS to ilnplenmltt knowledge SOllrces elf s/Idl vari~.</Paragraph>
      <Paragraph position="1"> lions(we call them changes) in a computational l)rc~ cessilr g systenl.</Paragraph>
      <Paragraph position="2"> Only by liltcriug using any intra-sentential knowledge sources, several candidates may remain ~m syntactically and semantically correct senteuces. For example, &amp;quot;9ojuu-shichi'(fifl.y-seven) sounds like &amp;quot;yo3uusyo', and the sentence &amp;quot;Gojuu-shichi-wo onegaishi-masu.&amp;quot;(Fifty-seveu, i)lea-ne.) is not only well-formed but also correct in a particular context. It is possible to select the correct candidate by referring to both the context and the situation of the ongoing dialogue. Ewm so, to pick the surface form, we nmst kllow wily tile speaker haq used ~t given expression to represel~t a COIIgellt, If we can determine llow these NPs ehallge, ~,lld what effect they bawl then we can choose the speech recognition candidates more accurately.</Paragraph>
    </Section>
    <Section position="2" start_page="0" end_page="0" type="sub_section">
      <SectionTitle>
3.2 Analysis of NP Changes
</SectionTitle>
      <Paragraph position="0"> In order to analyze NI' changes in a dialogue we. inspected 50 dialogues in a corpus. As a result of the analysis, NP changes arc categorized into three main classes: 1) Change by lexical e.ohesion: (tiffs class corresponds to reiteration\[5\]), 2) Change by differe.nt viewpoints: (described in detail in the next paragraph), and 3) Change by misrecognition.</Paragraph>
      <Paragraph position="1"> There are two aspects of viewpoint, which are the standpoinl of the agent and the node of the concept.</Paragraph>
      <Paragraph position="2"> in an inquiry dialogue, the standpoints of the agents are. always different. Thus, this class has only two stlb cla~ssl!s : 2(a) iioint keeping: both agents see the s;uue t~odc of a concept, and this subclass is divided according to the SEI.' : i. ditferent expressione.g. &amp;quot;walashf' and &amp;quot;Yama0ka-san&amp;quot;. il. additimt of prefix e.g. &amp;quot;lUUSyO&amp;quot; and &amp;quot;9o-juusyo&amp;quot;. iii. COml)lex a mixture of 2(a)i and 2(a)ii, 2(I)) shifting: the viewpoint of one of the agents shifts from thc node of a conccpt to a node of a related concept, and this is divided into: i. shortening e.g. &amp;quot;Kokusai-kaigf'(lntenrational Confereuee) and &amp;quot;kaiqf' (tire conference).</Paragraph>
      <Paragraph position="3"> it. unitinge.g. &amp;quot;ryousyuu-syo-to saNka-touroku-syo&amp;quot;(a receipt and an application form) and '&amp;quot;2syurnt-uo syorltf'(two types of forms).</Paragraph>
    </Section>
    <Section position="3" start_page="0" end_page="0" type="sub_section">
      <SectionTitle>
3.3 Domain-dependent Knowledge
</SectionTitle>
      <Paragraph position="0"> Configuration: q-'he domain-dependent knowledge base consists of a network of nodes and links. Basic nodes are divided into three types: a concept node representing a particular thing or concept retained in human memory, a lexleal node representing a partieular word or phrase used when expressing something, and a variable node representing a particular value corresponding to a valuable concept, which can have a specified value. A variable node can be instantiated by executiug tbe effect of a completcd plan (usually by GET-VALUE-UNIT ill Inleractiou plan\[11\]), so that it can have a particular SEF as the valnc of the node. For example, &amp;quot;Yamaoka&amp;quot; could be the value of a variable node corresponding to a concept node of 'name' in a sentence like &amp;quot;My name is Yamaoka &amp;quot;.</Paragraph>
      <Paragraph position="1"> The following types of links are defined: is-a link, representing a superordinate/subordiuate relation between two concept nodes, part-of link, representing a whole/part relation between two concept nodes, causal link, representing a causal relation between two concept nodes, prag llnk, representing a pragmatic relation to connect a particular concept node to a lexical node representing the tyllical SEFfor the concept, value link, representing an instance value relation between a particular valuable concept node and a variable node which has been bound to the SEF of its value, and eq llnk, representing the same meaning between two lexical nodes.</Paragraph>
      <Paragraph position="2"> Extension of eq link: In order to make the knowledge base sensitive to the changes considered in Section 3.2, tile eq link is extended. This lets us to add applicable conditions to eq links as sub-types of tile link. Applicable conditions are defined based ou classes of the categorization in 3.2. For example, if one lexical node is a polite SEF of another, the two lexical nodes can be emmected with an eq-if-polite llnk, e.g. &amp;quot;juusyo&amp;quot; and &amp;quot;go-juusyo&amp;quot;(see Fig. 1).</Paragraph>
    </Section>
  </Section>
  <Section position="5" start_page="0" end_page="0" type="metho">
    <SectionTitle>
4 Selnctlon Strategy
</SectionTitle>
    <Paragraph position="0"> Ill the dialogue, a speaker chooses an expression according to tile situation, the preceding context, and his/her beliefs. Assuming that tile system has recognized such conditions, we can efficiently choose the correct speech recognition candidate by searching the SEFs that are appropriate under the conditions.</Paragraph>
    <Section position="1" start_page="0" end_page="0" type="sub_section">
      <SectionTitle>
4.1 Rules of Applicable Conditions
</SectionTitle>
      <Paragraph position="0"> \[terc, two terms are defined for explanation: seed: if the predicted contextual information is bolmd to a particular SEF, then the seed of tile contextual information is tile SEI&amp;quot;, otherwise the seed is the value of the lexieal node linked by tbe l)rag llnk to the concept node corresponding to the contextual information, preferable set: a set of SEFs derived from a seed by an applicable rnlc, whicb then takes first priority for selecting the candidate.</Paragraph>
      <Paragraph position="1"> The basic rule for making a preferable set is: collect tile SEFs by following the eq link from the sced. Because in this paper we are focusing on dialogue situations and contcxts ratber than the speaker's beliefs, we only cover rules regarding changes by different viewpoints.</Paragraph>
      <Paragraph position="2"> For a predicted contextual information I, considering the dialogue situations ill Class 2(a): I. if I is in the territory of information of the other agent, then make a preferable set by following the eq-if-polite link from the seed, additionally, considering the preceding context: 2. if \[ has an antecedent which denotes the status of the other agent, i.e., there is an instantiated variallle node corresponding to I , then replace the seed with the antecedent, i.e., the SEF of I.he variable node, and make a preferable sct by following the eq-if-pollte llnk from the seed.</Paragraph>
      <Paragraph position="3"> Considering the contexts ill Class 2(b): 3. if 1 is a compound noun, (it's obviously the antecedent) then shift the seed to the concept one-level up t and make a prefcrablc set using the basic rule, 4. if l includes two or more concepts or SEFs and there is a concept node which is the upper node of both of these concepts, then shift the seed to the upper concept and make a preferable set using the basic rule 2, I Precisely, shifting a seed to a concept metals an operation to replace the seed with the SI~F of the lexical node followed by the prag llnk from the concept node.</Paragraph>
      <Paragraph position="4"> In this case an auxilisa'y word is usually added.</Paragraph>
      <Paragraph position="5"> AcrEs DE COLING-92. NA,'CrE.s, 23-28 AOt~r 1992 1 I 5 4 PRec. ot: COLING-92, NANTES. AUG. 23-28. 1992 In daily dialogue, speakers apply combinations of the above rules and other rules, but in this study we arc concentrating on simpler cases.</Paragraph>
    </Section>
    <Section position="2" start_page="0" end_page="0" type="sub_section">
      <SectionTitle>
4.2 Seleetion Algorithm
</SectionTitle>
      <Paragraph position="0"> Our ultimate goal is to select the correct speech recognition candidate from the predicted contextual information. An algorithm to do this is roughly de fined by following the three steps:  1. provide contextual information, 2. make a prefcrable set from i by the rules, 3. compare speech recognition outputs with 2, and if all equivalent is found then pick it ms the ai)propriate candidate,  else goto 2.</Paragraph>
      <Paragraph position="1"> Steps 1 and 2 above are I)acktracking points. For details of Step l, see \[1I\],\[12\]. l,'urthcr large-scale experiments may deternfinc hcuristically how many times Step 2 should be iterated.</Paragraph>
    </Section>
  </Section>
class="xml-element"></Paper>