<?xml version="1.0" standalone="yes"?>
<Paper uid="P06-3006">
  <Title>Semantic Discourse Segmentation and Labeling for Route Instructions</Title>
  <Section position="7" start_page="34" end_page="35" type="evalu">
    <SectionTitle>
6 Results
</SectionTitle>
    <Paragraph position="0"> We evaluated the performance of the systems using three measures: overlap match, exact match, and instruction follow through, using 6-fold crossvaliadation on 427 samples. Only the action chunks were considered for exact match and overlap match. Overlap match is a lenient measure that considers a segmentation or labeling to be cor- null rect if it overlaps with any of the annotated labels. Instruction follow through is the success rate for reaching the destination, and the most important measure of the performance in this domain. Since the baseline algorithm does not identify the token labeled with B-prefix, no exact match comparison is made. The result (Table 4) shows that CRFs perform better with a 73.7% success rate.</Paragraph>
  </Section>
class="xml-element"></Paper>