<?xml version="1.0" standalone="yes"?>
<Paper uid="C86-1149">
  <Title>Anothe r St ride Towa rds Knowledge-Based Machine Translation</Title>
  <Section position="6" start_page="635" end_page="637" type="concl">
    <SectionTitle>
7. Future Directions
</SectionTitle>
    <Paragraph position="0"> The twin advantages of the KBMT approach and the reversible functional grammars, applied to f-structures and semantic entity definitions, are 1) to provide a measure of extensibility that cannot be achieved via the conventional transfer grammar approach, and 2) to enable efficient real-time parsing via multi-stage precompilation. A further advantage over traditional transfer grammars becomes evident when one considers the translation problem from a more global perspective. In order to translate between any pair of N languages, our approach requires .the development of only N bi-directienal grammars (one per language). On the other hand, the conventional transfer approach requires that a new grammar be developed for each pair of languages and for each direction of the translation, Thus, to achieve the same number of bi-directional translations, requires on the order of N 2 transfer grammars. This calculation yields over 5,000 transfer grammars vs 72 functional/entity grammars to translate among the 72 most commonly spoken languages today.</Paragraph>
    <Paragraph position="1"> Recall that in addition to the economy of developrnent argument, the KBMT paradigm produces meaning.invariant translations for those domains whose semantics have been successfully codified.</Paragraph>
    <Paragraph position="2"> Although we have made significant inroads in the establishment of knowledge-based machine translation as a viable and superior alternative to the transfer grammar methodology, much of the difficult work remains before us. The integration of entity-oriented semantic representations and a generalized functional grammar, coupled with grammar precompilers, on-line parsers and generator provide a significant improvement over the first successful attempts to perform knowledge.based machine translation \[10, 5\]. The improvements are based on extensibility and uniformity of the semantic and syntactic knowledge sources, providing static separation and dynamic &amp;quot;run-time&amp;quot; integration, Our initial implementations convince us that this approach may hold the key to practical KBMT.</Paragraph>
    <Paragraph position="3"> Our pilot system operates in a subdomain ()f doctor.patient communications, selected for its relative syntactic richness, but fairly self-contained semantics. We have selected English and Japanese as our initial source and target languages, although we are also starting to investigate Spanish, French, German and Italian. Moreover, we are-striving to produce a system requir!ng mihimal if any clarification from the source-language user in his or 3 re be exact, eaeh rule has two Lisp programs; one lot parsing and the other for gel~erntion, Thec~e programs ale syt~thesized automatically by the ptecenlpiler in oldel \[0 tesl ~t~ltlOOtJc aRd Sylltac\[iC ~:onstraiuts. including as Iong-dislaoce dependencies alibi to assign const\[iucllts their appropriate sornantic and syntactic  her own language, and no aid whatsoever from a human translator or &amp;quot;posteditor&amp;quot; who knows both languages. We intend to grow this pilot system in several dimensions, including achieving a measure of completeness in subdomain coverage, adding one or two more languages, moving to a second and perhaps a third domain, and tailoring our implementation for relative efficiency of operation by completing the development of our multi-phase precompilers.</Paragraph>
    <Paragraph position="4"> In additio!\] to continued construction and extension of the pilot system -- the vehicle through which we are testing our theoretical tenets -- we are pursing the following objectives: * Bi-directionality -- As discussed above, functional grammars are theoretically bi-directional, but such a property has not yet been proven in practice for large scale systems. Our approach is not to interpret the bi-directienal grammars directly, but rather to compile them into much more efficient (and different) parsing and generation grarnmars, The latter endeavor still requires empirical wtlidation.</Paragraph>
    <Paragraph position="5"> o hlcrenlerltal Compilation .. In order to expedite the grammar development and testing cycle, we are contemplating incremental compilation for new additions or recent changes into large existing gn~mmars rapidly. Although the compilation process has proven successful in earlier parsers we have built \[3,27\], incremental compilation introduces new teuhnical problems.</Paragraph>
    <Paragraph position="6"> ~, User ex'tensJbility -- A longer range research topic is to provide zt structured interface whereby a user of th(~ KI\]MT system couM add donlain knowledge (entities) sad dictionary envies without r(.'quiring any kno~Icdgc of the internal struciure of the system.</Paragraph>
    <Paragraph position="7"> Extendir~ 9 th,!~ lexicon i'.~, of course, milch simpler them extending the domain semantics. All such extensions would work in concert with existing domain knowledge, lexicon, and grammar.</Paragraph>
    <Paragraph position="8"> e Robustness -- The recognition of ill-structured language is very important, especially for the shorttext domains we envision for our system (telex messages, banking transactions, doctor.patient dialogs, etc.). We have built selective-relaxation methods that integrate semantic and syntactic constrains before in the MULTIPAR system \[7, 12\], but have not yet investigated their application or extension into the functional/entity paradigm selected here.</Paragraph>
    <Paragraph position="9"> * Speech Compatibility -- A long-term objective is to integrate speech recognition and generation with on-line real-time machine translation. A parallel project at CMU is integrating speaker-independent continuousospeech recognition with a case.frame semantic parser of English \[15\]. We expect results of that investigation, which is already moving towards the precornpilation parsers discussed here, to pave the way towards eventual translation of spoken language.</Paragraph>
    <Paragraph position="10"> We expect that these and other developments will require a continued focused research effort over the coming years. 4 We claim only to have taken one more stride in the long march towards the theoretical and practical development of fully-automated knowledge-based machine translation.</Paragraph>
    <Paragraph position="11"> 4Since the deve\]oDieent of this new.generation technology for knowledge-based machine translation promises Io be a lYlajer ilew direction to the field, but requires sLibslantial resources to grow frolri theoretical conceptiun to large-sea.to application, we are starting the International Center for Machine Translation al CMU. &amp;quot;file center is dedicat~,,d to research and devnlopmenl of new techl3iques for machine Iraaslatien and theil eegineerinc into substantial denlonstratierl systems. )Is first major project is the investigation at the lunctionat grammar/entily-orierlted approach to 1(\[3M\], as fJlosenled irl this hOPer. Persor, s ieteresled in acquiring more inlellnStiorJ aboel tlln center elay rio slJ I.)y COl~taclitlg one of the atl~l;OlS.</Paragraph>
  </Section>
class="xml-element"></Paper>