<?xml version="1.0" standalone="yes"?>
<Paper uid="W05-1520">
  <Title>Statistical Shallow Semantic Parsing despite Little Training Data</Title>
  <Section position="4" start_page="186" end_page="186" type="metho">
    <SectionTitle>
4 Conclusions
</SectionTitle>
    <Paragraph position="0"> This work illustrates that one can achieve fair success in building a statistical NLU engine for a restricted domain using relatively little training data and surprisingly using a rather simple voting model.</Paragraph>
    <Paragraph position="1"> The consistently good results obtained from all the systems on the task clearly indicate the feasibility of using using only word/ngram level features for parsing. null</Paragraph>
  </Section>
  <Section position="5" start_page="186" end_page="186" type="metho">
    <SectionTitle>
5 Future Work
</SectionTitle>
    <Paragraph position="0"> Having successfully met the initial challenge of building a statistical NLU with limited training data, we have identified multiple avenues for further exploration. Firstly, we wish to build an hybrid system that will combine the strengths of all the systems to produce a much more accurate system. Secondly, we wish to see the effect that ASR output has on each of the systems. We want to test the robustness of systems against an increase in the ASR word error rate. Thirdly, we want to build a multi-clause utterance chunker to integrate with our systems. We have identified that complex multi-clause utterances have consistently hurt the system performances. To handle this, we are making efforts along with our colleagues in the speech community to build a real-time speech utterance-chunker. We are eager to discover any performance benefits. Finally, since we already have a corpus containing sentence and their corresponding semantic-frames, we want to explore the possibility of building a Statistical Generator using the same corpus that would take a frame as input and produce a sentence as output. This would take us a step closer to the idea of building a Reversible System that can act as a parser when used in one direction and as a generator when used in the other.</Paragraph>
  </Section>
class="xml-element"></Paper>