<?xml version="1.0" standalone="yes"?>
<Paper uid="C86-1149">
  <Title>Anothe r St ride Towa rds Knowledge-Based Machine Translation</Title>
  <Section position="3" start_page="0" end_page="633" type="intro">
    <SectionTitle>
2. Background
</SectionTitle>
    <Paragraph position="0"> Automating various forms of syntactic analysis has been a central concern of Computational Linguistics, producing methods ranginq from context-free grarnmar interpreters \[11,25, 13\], to ATNs\[28\], to unification grammars \[18\], and lexical-functional grammars \[4\]. The problem is that the production of accurate, uoarnbigaous parses of \[he source text, tolerant of minor grammatical deviations; requires a fairly complete semantic model of the domain, and a method for bringing the semantic knowledge to bear in the parsing process. Semantically-oriented parsers have succeeded at integrating semantics with syntax, but only at the cost of intertwining both knowledge suurces into ttlo program 1The riathors would like to acknowledge the other members of the the machine tlanslation laboratory at CMU who contributed in various ways to the research described in this paper: Peggy Anderson, Philip Franklin, Alex Ilcuphuaml, Marion Kee, I liroaki Sails, Yuko Tomita and Teruka Watanabe.</Paragraph>
    <Paragraph position="1"> 2The analyzer needs to comprehend all possible syntactic variants of any semantic messafjo in the analysis phase because it cannot contrel the form of its input, but to produce acceptable output, the generator need only render the the meaning in a well-d~.~fined staridald surface form. Of course, to I)reduco more expressive text, end to preserve syr,lactic as well as semantic Invari..u'~ce in the translation process, tile generalor must he expanded into a one-to-many mappirlg process compel able in complexity to !hat of the analyzer.</Paragraph>
    <Paragraph position="2">  ~// +..e1,,+|o,, i.</Paragraph>
    <Paragraph position="3"> Humell Ueel! Figu re 1 -1 : Knowledge-Based Interactive Machine Translation itself in fairly non-extensible ways \[23, 17, 2, 6\]. Subsequent improvements have succeeded in factoring out much of the domain semantics, but leaving the syntactic interpret;.{tion as part of the recognition program rather than as an explicit external grammar \[9, 14, 16\].</Paragraph>
    <Paragraph position="4"> In order to overcome these problems we have sought a method for static separation of the syntactic and semantic knowledge sources in the data structures, and dynanlic integration to bring all relevant knowledge to bear in the process of parsing. Static separation has the advantage that as linguistic coverage increases, or new languages are added to the system, parsing (and translation) still function for all previous semantic domains. Conversely, if the semantic domains are extended, or new ones added, parsing and translation of texts in these domains will function for all previously entered languages. In contrast, earlier methods that mixed semantic and syntactic information required hand-crafted updates to all previous structures in order to integrate new grammatical extensions or new languages. With the possible exception of Lytinen \[21J, who attempted a rudimentary form of static separation and dynamic integration, this rather appealing principle has not heretofore been a primary design criterion of natural language parsers in general, much less full machine-translation systems.</Paragraph>
    <Paragraph position="5"> Many of the syntactic analysis methods do not integrate well with semantic knowledge, especially knowledge that must be kept in separate data structures and integrated only by the preeompiler at the run:time language intepretation process. Similarly, many of the semantic representation formalisms do not lend themselves well to dynamic integration with syntactic constraints at parse time. The best fit we have been able to achieve comes from precompiling syntactic and semantic knowledge into a single knowledge base which is used only at run.time, as described in the subsequent sections.</Paragraph>
  </Section>
class="xml-element"></Paper>