<?xml version="1.0" standalone="yes"?>
<Paper uid="E89-1036">
  <Title>A Descriptive Framework for Translating Speaker's Meaning - Towards a Dialogue Translation System between Japanese and English -</Title>
  <Section position="6" start_page="0" end_page="0" type="metho">
    <SectionTitle>
REQUESTING
COMPLAINING
ADVISING
CONFIRMING
</SectionTitle>
    <Paragraph position="0"> ...etc.</Paragraph>
    <Paragraph position="1"> Conversely, the same intention can be conveyed through various surface expressions, as in the following variations of (2-1):</Paragraph>
  </Section>
  <Section position="7" start_page="0" end_page="0" type="metho">
    <SectionTitle>
REQUESTING
</SectionTitle>
    <Paragraph position="0"> (2-2) gakusei waribiki o site kudasai.</Paragraph>
    <Paragraph position="1"> student discount OBJ make do-GIVFAV-POL-IMP Please make me a student discount.</Paragraph>
    <Paragraph position="2"> (2-3) gakusei waribiki o site itadaki tai student discount OBJ make do-RECFAV-PO|, want nodesu ga.</Paragraph>
    <Paragraph position="3"> EXI'I,-I'OL MODEl{ 1 wonder ifyou could make me a student discount. (2-4) watasi wa gakusei na no desu ga.</Paragraph>
    <Paragraph position="4"> ! TOP student COPL EXPL-POL MODER  I am a student, you know.</Paragraph>
    <Paragraph position="5"> N.B. Concerning a 'discount' request, (2-2) seems a bit strong for a real situation although there is no specific contexttml condition to decide definitely if it is or not. (2-1) (2-3) and (2-4) are seen in our data.</Paragraph>
    <Paragraph position="6"> These examples clearly show that intention is context-dependent, and that to understand the speaker's meaning correctly, an inference mechanism is necessary.</Paragraph>
    <Paragraph position="7"> Various surface expression patterns give clues for ascertaining illocutionary forces (Wierzbicka 1986).</Paragraph>
    <Paragraph position="8"> (2-5) t~rokuydsi o o-okuri negae masu ks? registration form OBJ send-POL desire POL QUEST Can you please send me a registration form7 (2-6) Could you kindly send them all together? Hegau in (2-5), a verb for request, and ks, the sentence-final particle of questions, indicate request. Kindly in (2-6) signals a request in English. In other words, even without knowledge of the context of an utterance, knowledge of communicstive strategies of language and their expression patterns allow the derivation of intentions from utterances.</Paragraph>
    <Paragraph position="9"> In the above examples, we can see there are various ways of expressing requests. This indirectness derives from social patterns in requesting things common to all cultures to some degree. On the other hand, however, it depends on each specific society. In this paper we accept indirectness as an unavoidable and basic feature of spoken utterances, and deal with indirect patterns such in (2-1) and (2-3) that will be called speech-act indirectness. Indirect expressions such as (2-4), which are called propositional indirectness, are not treated for the reason given in the next subsection. We use the term speaker's meaning to refer to intention expressed by speech-act indirectness. Using this notion, we try to capture syntactically the major portion of speech-act-related expressions in spoken Japanese.</Paragraph>
    <Paragraph position="10"> 2.2. Translation of speaker's meaning We assume that for machine translation it is sufficient to understand utterances on the level of speech-act indirectness, without referring to propositional indirectness. On the one hand, when there is a large degree of indirectness such as the omission of propositional content in (2-4) where the topic &amp;quot;discount fee for students&amp;quot; is not actually mentioned, we must be content with a direct translation of what has been stated. This is because a sentence-based translation cannot compensate for the missing content. In addition, since the hearer will no doubt be able to infer something about the omitted content anyway, the speaker is best served by a direct translation closest to the original. On the other hand, when the propositional content is explicitly phrased but requires indirectness to make an appropriate translation into the target language, a system that concentrates on speech-act indirectness will again be tile most useful, because socio-linguistic differences will be expressed typically in speech-act indirectness as in (2-1) and (2-3).</Paragraph>
    <Paragraph position="11"> Consequently, we develop a framework aimed at extracting speaker's meaning in terms of speech-act indirectness.</Paragraph>
  </Section>
  <Section position="8" start_page="0" end_page="0" type="metho">
    <SectionTitle>
3. I FTs
</SectionTitle>
    <Paragraph position="0"> 3.1.Classiflcatlon of IFTs An experiment has been carried out on collected data of spoken-style inter-terminal dialogues to extract illocutionary acts. The subject of the conversations was limited to - 265 application for an international conference, and the content was mainly on inquiry, request, and confirmation about the conference between a secretary and an applicant.</Paragraph>
    <Paragraph position="1"> We classify surface IFTs into six types (Table 1). This is the immediate result of the analysis made intrasententially by means of Head-Driven Phrase Structure Grammar (HPSG)/Japanese Phrase Structure Grammar (JPSG). The six types are differentiated from each other only by means of the uppermost predicate value that is the result of the surface-based analysis. For example, an indirect request with an interrogative sentence pattern such as (2-5) t6rokuySsi o o-okuri negae masu ha? Could you please send me a registration form? is classified simply as an INTERROGATIVE type, though it is OPTATIVE at the deep IFT level. Also, a sentence with an active, present- null tense verb such as (3-1) tdrohuyOsi o o-ohuri si masu registration form OBJ send-POL do-POL I will send you a registration form.</Paragraph>
    <Paragraph position="2"> is analyzed as INFORMATIVE, though it is PROMISE at the deep level.</Paragraph>
    <Paragraph position="3"> Table 1. SurfacelFTs surface surface IFT instances predicate value F,X PRESS! VI~ arlgat6 (thanks) arlgal6- sumimasen (sorry) THANKS, etc. mosimosi (hello) mosimosi- I'liN\['lC say, nora (goodbye) ltELLO, etc.</Paragraph>
    <Paragraph position="4"> negau (wish) x-REQU EST OPTATIVE kudasai (please) QUESTIONIF I NTI~RI~,OGATI VE ha, ne QUESTIONREF tai (want) SU BJ ECTIVE hosii (want...to) x-WISH  procedure for translating speaker's meaning. In contrast to a conventional machine translation procedure, speaker's meaning can be analyzed and generated, without passing through transfer, by means of IFTs and DPs. Here, we do not pursue machine translation problems concerning propositional content. The processing of speaker's meaning consists of two stages, unification-based syntactico-semantic analysis and plan inference. We will now give a more precise description of these two stages.</Paragraph>
    <Paragraph position="6"> As a grammar for surface-level analysis, we have adopted HPSG (Pollard and Sag 1987) and JSPG (Gunji 1987), that is a modification of the former for dealing with Japanese. On the basis of a unification parser developed at ATR (Kogure et al. 1988), the grammar has been written and proven capable of analyzing all fundamental sentence patterns in spoken-style Japanese conversation (Yoshimoto, Kogure and Iida 1989).</Paragraph>
    <Paragraph position="7"> This grammar analyzes sentence (3-2) as (3-3) by means of syntactic rules and lexical descriptions, of which only those for the subsidiary verb morau are given as (3-4).</Paragraph>
    <Paragraph position="8">  (3-2) t6rokuy~si o Okutte morse rnasu ka? registration form OBJ send RECFAV-POSS POL QUEST (lit.) Could I have the favor of your sending me a registration form?  '?' is a prefix for a tag name representing a token identity of feature structures. In (3-4), the third member of the SUBCAT value specifies the conjugational form and modality type of the complement verb. The feature MODL imposes  conditions on the modality type that plays a key role in Japanese syntax by dominating mutual predicate component subcategorization and subordination. In order to handle the unorderedhess of Japanese case phrases, the SUB\[AT value is a set, following JPSG, instead of an ordered list in the HPSG for English. The set is expanded by a rule reader into its corresponding possible ordered list descriptions. Since Japanese case phrases are always postposed by a caseindicator, they are assigned to the part-of-speech category P. The PRAG feature stipulates here that the speaker empathizes more with the subject (?X1 in (3-4)) than with the indirect object (?X2).</Paragraph>
    <Paragraph position="9"> This pragmatic information is further utilized with a discourse model to identify omitted subjects and objects, because they are mostly omitted in honorific or empathy-related sentences.</Paragraph>
  </Section>
  <Section position="9" start_page="0" end_page="0" type="metho">
    <SectionTitle>
4. Identification of IFTs
</SectionTitle>
    <Paragraph position="0"> The surface analysis result such as (3-3) serves as an input to plan schemata called IFT-Schemata that identify deep IFTs (or merely IFTs) syntactically by means of predicateinternal collocation, adjunction, tense, and modal information. An IFT-Schema consists of a goal whose value is a partial description of a deep IFT, and a decomposition whose value is a disjunction of partial descriptions of surface IFTs, preconditions, and effects as in (4-1), (4-2) and (43). A surface IFT is searched for which unifies with one of the descriptions in the decompostion.</Paragraph>
    <Paragraph position="1"> The goal in the same schema is the resulting deep IFT. Adoption of the unification method enables hi-directional flow of information between the deep speech act type and the decomposition. This leads to an easier disambiguation and supplementation of surface analysis results by linguistically specifying IFTs (Kogure et el.</Paragraph>
    <Paragraph position="2"> 1988).</Paragraph>
    <Paragraph position="3"> The difference between surface analyses and deep IFTs is absorbed by a &amp;quot;thesaurus&amp;quot;, as in (44), that relates the two. This specifies that MORAU-RECEIVE-FAVOR is a subtype of RECEIVE-FAVOR. (4-5) is the result of the IFT inference.</Paragraph>
    <Paragraph position="4">  In identifying deep IFTs, syntactic constraints in Japanese are fully utulized.</Paragraph>
    <Paragraph position="5"> On the one hand, IFTs SUBJECTIVE and OPTATIVE are universally limited to expressions with first person singular subject and  present tense and without modal information, and Japanese surface predicates reflect these restrictions very well. Also, OPTATIVE is limited to second person recipient. For example, (4-6) C/SBJ kaigi ni mdsikomi tai.</Paragraph>
    <Paragraph position="6"> conference OBJ2 reserve want I would like to register for the conference.</Paragraph>
    <Paragraph position="7"> (4-7) ~SBJ kaigi ni mdsikomi tai sd do.</Paragraph>
    <Paragraph position="8"> conferenceOBJ2 reserve want l-hear  I hear (someone) wants to register for the conference. While sentence (4-6) with the present, non-modal auxiliary tai (want to) belongs to the SUBJECTIVE type, (4-7) with the evidential modality belongs to the ASSER'ITVE type. This fact is utilized, by means of two lexical descriptions of tai and IFT-Schemata restricting the decomposition members' person, tense, and modal information, to identify the omitted subject of(4-6) as the first person, and that of(4-7) as the third person.</Paragraph>
    <Paragraph position="9"> On the other hand, adverbials that exclusively modify deep IFTs are also utilized in disambiguating IFTs, For example, a sentence with O-Regal simasu (request, implore) is ambiguous among OPTATIVE, ASSERTIVE, and PROMISE. If it is modified by dEzo (please), however, the sentence is always an OPI'ATIVE type.</Paragraph>
    <Paragraph position="10"> Deep IFTs with their corresponding syntactic constraints are diagramed by Table 2. Instances in the Table indicate each of the corresponding deep IFTs, but the opposite is not necessarily true. For example, a deep IFT OPTATIVE can be indicated by complex predicates that belong to the surface category INTERROGATIVE or ASSERTIVE. Table 3 illustrates the relation between the deep IFT OPTATIVE and its corresponding surface IFT with instances.</Paragraph>
    <Paragraph position="11">  ...re hure masu ha? (will you do me the favor of...) ...re kure masen ks? (won't you do me the favor of...?) te morae masu ha? (can I receive the favor of...?) ...tain desu ga (I would like to...) ...re morai tain desu ga (I 'd like to receive the favor of...) ...re morai masu (I will receive the favor of...) ...to arigatai n desu ga (I would be happy if you...) By so specifying the IFT, information absent in surface utterances such as zero anaphora are compensated for and in some cases multiple analyses are disambiguated. (3-3), the surface analysis of (3-2), is analyzed as (4-5). This enables an adequate English translation (4-8) instead of an inappropriate literal translation (4 null 9). Note that at the same time the subject and indirect object missing in the surface sentence are compensated for by the IFT specification of the agent and recipient.</Paragraph>
    <Paragraph position="12"> (4-8) Could you send me a registration form? (4-9) *Can I receive a favor of your sending me a registration form? 5. Dl's 5.1. Necessity of DPs  We can summarize the difference between Japanese and English communication behavior as follows: Japanese interpersonal relation is the most essential factor English interpersonal relation is essential, but how to convey or read intentions is more important For example, (5-1) is an utterance from a boss to a secretary to request him to work overtime. This Japanese utterance is not an order because it is expressed in a polite way using the negative interrogative. This kind of request is not unusual in Japanese because of the priority given to social standing. Because Japanese think a request phrased like this is normal, the English translation shown in (5-1) using can and sorry seems appropriate to them, too. But actually an appropriate translation requires a more polite expression that addresses the secretary's inconvenience, as in (5-1)'. Thus, to get an appropriate translation of (5-1), we must reconsider from the viewpoint of the target language interpersonal relations between the speaker and the hearer and the inconvenience of requested action for the hearer.</Paragraph>
    <Paragraph position="13"> (5-1) sumanaiga, zangyd site syorui o sorry work overtime documentsOBJ taipu site kure nai ha na? type do-GIVFEV NEG QUEST Sorry, but can you stay late to type these documents? (5-1)' Do you think you could possibly stay late to type these documents? To resolve these communicative differences between Japanese and English, we assume four kinds of parameterlzed factors, which we call Decision Parameters (DPs). These are: interpersonal relation, cost-benefit relation, definiteness of propositional content, and topicality of propositional content. Interpersonal relation indicates the situational relationship between utterance participants as constituted by age, social status, familiarity, gender, and the other factors governing use of Japanese honorifics. Cost-benefit relation indicates whether the action intended by the speaker's utterance is convenient to the speaker or to the hearer. Definiteness of propositional content means whether propositional content is routine or easily performed work, or whether it requires additional or unusual work. Topicality of propositional content is related to the position of an utterance in discourse, which means whether or not the speaker's intention is already implied. Table 4 shows these four parameters and their values. In particular, DP4 or topicality presents discourse information which affects the politeness level of surface expressions. In the present experimental situation, extraction of speaker's meaning is limited to isolated utterances separate from discourse structure, but - 269 to get appropriate expressions in generation, we need DP4 in connection with a discourse model.  In the plan inference method of generation, we use DPs in order to get appropriate English surface IFTs to convey IFTs in English. Since we are limiting the input to a task-oriented domain like conferences, we can re-state input in terms of propositional content. This propositional content is then measured in terms of the three DP values as a default (Table 5).</Paragraph>
    <Paragraph position="14">  (from a client to a secretary)* (1) send a registration form HR SP ROU (2) inform about the conference HR SP ROU (3) assist a hotel accomodation HR SP ROU (4) provide an interpreter HR SP UNS (5) give a student discount HR SP UNS (6) reimburse a fee HR SP UNS (7) come for to the station HR SP U NS \[S\] Request (from a secretariy to a client) (8) send back the registration form HR SP ROU (9) tell one's name and address HR SP ROU (10) make a registration procedure HR SP ROU (11) pay by bank transfer HR SP UNS (12) take part in the party HR SP UNS (13) be informed about persons HR SP UNS  who wish to participate *In bt, siness telephone conversations in English, the hearer is always considered to be in a higher position, even in the case of a boss to a secretary. So the value of DP1 for \[A\] is always IlR.</Paragraph>
    <Paragraph position="15"> We suppose that differences between Japanese and English consist in the different amount of DPs we should refer to when extracting surface IFTs. Japanese surface IFTs will be concerned with DP1 and DP2 since Japanese expressions do not stress speaker's intention, whereas English surface IFTs will range over all four DPs and produce a larger range of appropriate translation choices.</Paragraph>
    <Paragraph position="16"> For example, (1) and (7) ofTable 5 which differ in definiteness of propositional content (i.e. routine or unusual), can be generated in the same way in Japanese, which involves only DP1 and DP2. That is, (5-2) t6rokuydsi o okut-te moral tai registration form OBJ send do-RBCFAV want no desu ga. ---(1)</Paragraph>
  </Section>
  <Section position="10" start_page="0" end_page="0" type="metho">
    <SectionTitle>
EXPL-POL MODER
</SectionTitle>
    <Paragraph position="0"> (5-3) eki made mukaeni hi-re moral tai station LOC come for do-RECFAV want no desu ga. ---(7)</Paragraph>
  </Section>
class="xml-element"></Paper>