<?xml version="1.0" standalone="yes"?>
<Paper uid="I05-6001">
  <Title>The TIGER 700 RMRS Bank: RMRS Construction from Dependencies</Title>
  <Section position="7" start_page="8" end_page="9" type="concl">
    <SectionTitle>
5 Conclusion
</SectionTitle>
    <Paragraph position="0"> We presented a method for semantics construction that converts dependency structures to RMRSs as they are output by HPSG grammars. By applying this method to the TIGER Dependency Bank, we construct an RMRS Bank that allows cross-framework parser evaluation for German. Our method for RMRS construction can betransposedto dependency banks for other languages, such as the PARC 700 Dependency Bank for English (King et al., 2003). The choice of RMRS also ensures that the semantic bank can be used for comparative evaluation of HPSG grammars with low-level parsers that output partial semantics in terms of RMRS, such as the RASP parser of Carroll and Briscoe (2002).</Paragraph>
    <Paragraph position="1"> While the formalism of (R)MRS has its origins in HPSG, we have shown that RMRS semantics construction can be carried over to dependency-based frameworks like LFG. In future research, we will investigate how the semantic algebra of Copestake et al. (2001)  compares to Glue Semantics (Dalrymple, 1999). Our construction rules may in fact be modified and extended to yield semantics construction along the lines of Glue Semantics, with hooks as resources and Rels, Hcons andIngsets as meaning language. Inthis scenario, the composition rules would consume the hook of the semantic argument, so that resource-sensitivity is assured. Scope ambiguities would not result in alternative derivations, since RMRS makes use of scope under-specification in the meaning language.</Paragraph>
    <Paragraph position="2"> Related work in Dyvik et al. (2005) presents MRS construction from LFG grammars in a correspondence architecture, where semantics is defined as a projection in individual syntactic rules. Our architecture follows a description-by-analysis (DBA) approach, where semantics construction applies to fully resolved syntactic structures. This architecture is especially suited for the present task of treebank creation, where grammars for a given language may not have full coverage.</Paragraph>
    <Paragraph position="3"> Also, in a DBA architecture, incomplete rule sets can still yield partially annotated, i.e., unconnected semantic structures. Likewise, this construction method can deal with partially analysed syntactic input.</Paragraph>
    <Paragraph position="4"> Finally, our method can be extended to a full parsing architecture with deep semantic output, where care should be taken to preserve structural or categorial information that we identified as crucial for the purpose of principle-driven semantics construction.</Paragraph>
  </Section>
class="xml-element"></Paper>