<?xml version="1.0" standalone="yes"?>
<Paper uid="P06-3006">
  <Title>Semantic Discourse Segmentation and Labeling for Route Instructions</Title>
  <Section position="4" start_page="0" end_page="31" type="intro">
    <SectionTitle>
2 Task
</SectionTitle>
    <Paragraph position="0"/>
    <Section position="1" start_page="0" end_page="31" type="sub_section">
      <SectionTitle>
2.1 Input and Output
</SectionTitle>
      <Paragraph position="0"> Three inputs are required for the task:  The output is the location of the office the directions aim to reach.</Paragraph>
    </Section>
    <Section position="2" start_page="31" end_page="31" type="sub_section">
      <SectionTitle>
2.2 Corpus Collection
</SectionTitle>
      <Paragraph position="0"> In an experiment to collect the corpus, (Haas, 1995) created a simulated office building modeled after the actual computer science department at SUNY/Albany. This environment was set up like a popular first person shooter game such as Doom, and the subject saw a demonstration of the route he/she was asked to describe. The subject wrote directions and sent them to the experimenter, who sat at another computer in the next room. The experimenter tried to follow the directions; if he reaches the right destination, the subject got $1.</Paragraph>
      <Paragraph position="1"> This process took place 10 times for each subject; instructions that the experimenter could not follow correctly were not added to the corpus. In this manner, they were able to elicit 427 route instructions from the subject pool of 44 undergraduate students.</Paragraph>
    </Section>
    <Section position="3" start_page="31" end_page="31" type="sub_section">
      <SectionTitle>
2.3 Abstract Map
</SectionTitle>
      <Paragraph position="0"> To simplify the learning task, the map of our computer science department was abstracted to a graph. Imagine a track running down the halls of the virtual building, with branches into the office doors. The nodes of the graph are the intersections, the edges are the pieces of track between them. We assume this map can either be prepared ahead of time, or dynamically created as a result of solving Simultaneous Localization and Mapping (SLAM) problem in robotics (Montemerlo et al, 2003).</Paragraph>
    </Section>
    <Section position="4" start_page="31" end_page="31" type="sub_section">
      <SectionTitle>
2.4 System Components
</SectionTitle>
      <Paragraph position="0"> Since it is difficult to jump ahead and learn the whole input-output association as described in the task section, we will break down the system into two components.</Paragraph>
    </Section>
    <Section position="5" start_page="31" end_page="31" type="sub_section">
      <SectionTitle>
Front End:
RouteInstruction-ActionList
Back End:
ActionListxMapxStart-Goal
</SectionTitle>
      <Paragraph position="0"> The front-end is an information extraction system, where the system extracts how one should move from a route instruction. The back-end is a reasoning system which takes a sequence of moves and finds the destination in the map. We will first describe the front-end, and then show how to integrate the back-end to it.</Paragraph>
      <Paragraph position="1"> One possibility is to keep the semantic representation close to the surface structure, including under-specification and ambiguity, and leaving the back-end to resolve the ambiguity. We will pursue a different route. The disambiguation will be done in the front-end; the representation that it passes to the back-end will be unambiguous, describing at most one path through the building. The task of the back-end is simply to check the sequence of moves the front-end produced against the map and see if there is a path leading to a point in the map or not. The reason for this is two fold. One is to have a minimal annotation scheme for the corpus, and the other is to enable the learning of the whole task including the disambiguation as an IE problem.</Paragraph>
    </Section>
  </Section>
class="xml-element"></Paper>