<?xml version="1.0" standalone="yes"?>
<Paper uid="C92-3164">
  <Title>A SPOKEN LANGUAGE TRANSLATION SYSTEM : SL-TRANS2</Title>
  <Section position="3" start_page="0" end_page="0" type="metho">
    <SectionTitle>
TRANS \[Morimoto-90\]
</SectionTitle>
    <Paragraph position="0"> mechanism is given. In the following three sections, distinctive technical features of each component of the translation system are described. Experiment results are shown in section 6.</Paragraph>
  </Section>
  <Section position="4" start_page="0" end_page="0" type="metho">
    <SectionTitle>
2. System Overview
</SectionTitle>
    <Paragraph position="0"> A block diagram of the system is shown in Fig.1. Using a speaker adaptation technique, the  speech recognizer can accept any speaker's speech. Phone level Hidden Marker Models (HMM) and syntactic rules for Japanese are defined in the recognizer \[Kita-90\]. By referring to these rules, the recognizer predicts the next possible phones and verifies their existence by matching them with corresponding HMMs. The process is invoked for many hypotheses in parallel until the end of the utterance is reached. Finally, the n-best sentential hypotheses are output, with their respective recognition scores. The output hypotheses are tagged with word information such as a part-of-speech label, then AcI~ DE COLING-92, NANTES. 23-28 AOC'r I992 1 0 4 8 PROC. OF COLING-92, NArcrEs, AUG. 23-28, 1992 the speech recognizer works as a kind of morphological analyzer for the following analysis component. These hypotheses are all well-formed syntactically, hut not always semantically, pragmatically or contextually.</Paragraph>
    <Paragraph position="1"> The next analysis component checks the validity of each hypothesis and selects the most plausible one*2). After analysis, some zero anaphora (noun phrase ellipses) are supplemented using pragmatics such as honorific expressions in Japanese. Then, the semantics of the sentence is output in the form of a feature structure. This feature structure is generally composed of two parts: an intentional content and a propositional content. The former indicates the speaker's intention and is expressed in terms of language-independent conccpts. The latter is expressed in terms of language-dependent concepts. The subsequent transfer system transfers only the propositional content to their target language concepts. During Um generation process, tim two components are merged and a f'mal surface expression in the target language is generated. Finally, synthesized English speech is output from the synthesizer. Currently, a commercial English language speech synthesizer is used in the system.</Paragraph>
  </Section>
  <Section position="5" start_page="0" end_page="0" type="metho">
    <SectionTitle>
3. Analysis
</SectionTitle>
    <Paragraph position="0"/>
    <Section position="1" start_page="0" end_page="0" type="sub_section">
      <SectionTitle>
3.1 Grammar Formalization
</SectionTitle>
      <Paragraph position="0"> The grammar formalisnl adopted was originally based on HPSG ( and its Japanese version JPSG) \[Kogure-90\]. In each lexical entry, syntactic constraints, semantic constraints and even pragmatic constraints are defined as a feature structure (Fig.2).</Paragraph>
      <Paragraph position="1"> Parsing is basically accomplished using a unification operation between lexical items or successively derived constituents. This is effective in parsing Japanese spoken sentences which have a variety of expressions. According to the JPSG theory, a few principles (the head feature principle, the subeategorization feature principle, etc.) and one grammar rule (a mother is composed of a daughter and a head) are *2) The contextual evaluation function is not yet implemented in the current system  sufficient to cover all linguistic phenomena. ltowever, naive implementation of the theory as a practical system brings an explosion of processing time and memory consumption, even for a simple sentence. To solve this problem, medium grained context free grammar (CFG) rules are introduced \[Nagata-92\]. The grammar rules are constructed to maintain declarative description of lexieal constraints and also to suppress unnecessary unification execution. For instance, the concatenation conditions between Japanese verbs and auxiliaries are defined explicitly by the rules.</Paragraph>
    </Section>
    <Section position="2" start_page="0" end_page="0" type="sub_section">
      <SectionTitle>
3,2 Parsing Algorithm
</SectionTitle>
      <Paragraph position="0"> Parsing is guided by CFG rules and accomplished by the unification operation as described above. Generally, most of the processing time in a unification-based parsing method is consumed for the unification operation. In this system, besides dividing CFG rules as mentioned above, other efficiency improving technologies have been introduced. For instance, unification execution is delayed until all CFG rules have been applied. Another approach is to improve the unification procedure itself. In our system, an efficient unification mechanism using several techniques such as the quasi-destructive graph unification algorithm \[Tomabechi-91\] has been implemented. Using these improvements, tiffs system can analyze an input utterance in a fairly short time.</Paragraph>
    </Section>
    <Section position="3" start_page="0" end_page="0" type="sub_section">
      <SectionTitle>
3.3 Zero Anaphora Resolution
</SectionTitle>
      <Paragraph position="0"> ACRES DE COLING-92, NANTES, 23-28 hotter 1992 i 0 4 9 Paoc. OF COLING-92. NANTES, AUO. 23-28, 1992 Some zero anaphora are resolved and supplemented using pragmatic information in Japanese \[Dohsaka-90l. In general, pronouns indicating the participants such as &amp;quot;I&amp;quot; or &amp;quot;You&amp;quot; are seldom explicit in spoken Japanese. On the other hand, Japanese is abundant in honorific expressions, and such information can be used to interpret some zero pronouns. For instance, in the following example, the agent of the predicate &amp;quot;okuru&amp;quot; (send) in the last phrase can be inferred to be the speaker because he (she) is stating a promise and expressing it politely. Then, the indirect object of&amp;quot;okuru&amp;quot; is decided as the hearer, if the dialogue only concerns two participants.</Paragraph>
      <Paragraph position="2"> Other zero anaphora, especially a subject, in a sentence without such information is interpreted as the speaker.</Paragraph>
    </Section>
  </Section>
  <Section position="6" start_page="0" end_page="0" type="metho">
    <SectionTitle>
4. Transfer
</SectionTitle>
    <Paragraph position="0"> Input to the transfer process is a feature structure indicating the semantics of an utterance. The feature structure is represented recursively using relationships. A relationship consists of a relationship name and its case roles. A relationship name is essentially a concept. In the analysis system, the surface illocutionary force type for the utterance is calculated from the surface expression. This is converted to an appropriate illocutionary force type by analyzing and taking into consideration the speaker's plan in that situation. In the current system, however, this conversion is done straight from the surface illocutionary force type because a contextual processing mechanism has not yet been integrated into the system. Typical illocutionary force types established for goal-oriented dialogues, such as those in the target domain, are shown in Table 1.</Paragraph>
    <Paragraph position="1"> The transfer system transfers only the feature structure of the propositional content using a feature structure rewriting system \[Hasegawa-90\]. The rewriting system traverses an input feature structure and rewrites it according to a set of rewriting rules. There are  many kinds of rules such as concept-to-concept transfer rules or structural transfer rules from Japanese-to-English; or even Japanese-to-Japanese paraphrasing rules which make transferring easier. A rule is defined declaratively and composed of three subspecifications as in Fig.4: an environment condition, an input condition and an output specification. The environment condition is used on &lt;rein&gt; ~ts=9 in :phase J-E  to control the application of rules instead of encoding explicit relationships between rules; when some condition is given, only the rules satisfying it are applied. It could also be used to transfer the input properly based on a given context.</Paragraph>
    <Paragraph position="2"> Another important problem in the transfer process is how to disambiguate polysemous words and how to choose a proper target concept. In this system, a thesaurus of concepts is defined and used for this purpose. This thesaurus is ACRES DE COTING-92, NANTES, 23-28 AOt~T 1992 1 0 S 0 PREC. OF COLING-92, NANTES, AUO. 23-28, 1992 implemented as a type system and referred to by related rules.</Paragraph>
  </Section>
  <Section position="7" start_page="0" end_page="0" type="metho">
    <SectionTitle>
5. Generation
</SectionTitle>
    <Paragraph position="0"> The basic method used in the generation system is also a unification algorithm. However, unlike the analysis system, each linguistic generation rule is defined for a comparatively large unit. This is because the variety of sentences to be generated is not as great as that in analysis, e.g. an idiomatic expression can be used in certain cases. A generation rule is defined as a phrase definition \[Kikui-92\]. A phrase definition is basically composed of three subspecifications as shown in Fig.4: a syntactic phrase structure, syntactic constraints and semantics, and an application environment.</Paragraph>
    <Paragraph position="1"> structure (S-TOP (S AUX (NP PRON) VP) SIGN)  In principle, a phrase definition is equivalent to a syntactic derivation rule augmented with semantics, other linguistic constraints and environmental constraints. Generation is executed by activating related phrase definitions successively which can subsume the whole semantics of the input feature structure. The validity of a combination of phrase definitions is examined using the unification algorithm.</Paragraph>
    <Paragraph position="2"> Finally, a set of phrase definitions is determined and their combined syntactic structure is produced as a result. An environment description is not used in the current system, but will be used to generate a more appropriate expression in a given context.</Paragraph>
  </Section>
class="xml-element"></Paper>