<?xml version="1.0" standalone="yes"?>
<Paper uid="W98-1422">
  <Title>Fully Lexicalized Head-Driven Syntactic Generation</Title>
  <Section position="5" start_page="214" end_page="215" type="evalu">
    <SectionTitle>
6 Conclusion and Comparison
</SectionTitle>
    <Paragraph position="0"> We have shown how preprocessing an HPSG grammar can be used to avoid the costly on-line application (unification) of HPSG schemata in a modularized generation system with a microplanner and a separate syntactic generator. The compilation of an HPSG grammar to TAG grammar allows the use of an efficient syntactic generator without sacrificing the declarative nature of the HPSG grammar.</Paragraph>
    <Paragraph position="1"> It is important to compare the generation strategy presented here with Semantic-head-driven generation \[Shieber et al. 1990, van Noord 1990\] which is a direct generation algorithm froni logical form encodings. It improves previous algorithms in efficiency and in imposing less restrictions on the type of grammar. It is also applicable to HPSG and proceeds by applying the HPSG schemata in a bottom-up fashion, driven from the lexical heads of the schemata.</Paragraph>
    <Paragraph position="2"> To a large ex.tend, the TAG-based generation algorithm presented here goes through the same steps as semantic-head-driven generation. However, most of *these steps will have been made during the off-line preprocessing and are encoded in the elementary trees of the TAG grammar thns resulting 6Note that the node labels shown in Figures 7 are only a concession to readability. The TAG requirement that in an auxiliary tree the foot node must have the same category label as the root node is formally fulfilled in our implementation.</Paragraph>
    <Paragraph position="3">  in an important gain in efficiency. Note though, that the generation task in the algorithm presented here is shared between the micr0planner and the syntactic generator,-so a formal comparison must include both components.</Paragraph>
    <Paragraph position="4"> Work on generation with TAG generally assumes that there is a one, to~-one mapping between the information in the generator input and the choice of elementary tree \[Mcdonald and Pustejovsky 1985, Yang, McCoy, and Vijay-Shanker 1991, D0ran and Stone 1997\]. In general, this will not be the case. In particular, in our system the input is not always sufficiently analyzed and the preprocessing froman HPSG grammar potentially *creates more than one elementary tree that fits the input parameters.</Paragraph>
    <Paragraph position="5"> One possible approach are choice nets-see \[Yang, McCoy, and Vijay-Shanker 1991\] who interpret systemic grammar in this way. Our approach has some similarity, though we have provided a more general algorithm that does not require the specification of grammar specific choice nets but rather executes tree Selection and combination from more declarative knowledge bases. Tree selection is implemented mainly by unification (adding feature values from the input specification to the trees where unifiable) and the best-first search algorithm is a general framework for handling sets of possible elementary trees, including backtracking steps when non-local tests (e.g. unification in the resulting derived tree) fail. This approach is also a precondition in our system since we have no direct access to the TAG grammar as it is automatically preprocessed from an HPSG grammar.</Paragraph>
    <Paragraph position="6"> VM-GECO is fully implemented (in Common Lisp) and integrated into the speech-to-speech translation system Verbmobil for Enghsh and German. For example, the underlying English HPSG grammar has almost 3000 iexical entries with over 200 lexical types. The resulting lexicalized TAG consists of about 2800 trees. The average overall generation time per sentence (up to length 24) is 0.7 cpu *seconds on a SUN ULTRA-1 machine, 68% of the runtime are used for tile microplanning while tile remaining 32% of the runtime are used for syntactic generation.</Paragraph>
  </Section>
class="xml-element"></Paper>