<?xml version="1.0" standalone="yes"?>
<Paper uid="W06-1623">
  <Title>Inducing Temporal Graphs</Title>
  <Section position="4" start_page="0" end_page="189" type="intro">
    <SectionTitle>
1 Introduction
</SectionTitle>
    <Paragraph position="0"> Understanding the temporal flow of discourse is a significant aspect of text comprehension. Consequently, temporal analysis has been a focus of linguistic research for quite some time. Temporal interpretation encompasses levels ranging from the syntactic to the lexico-semantic (Reichenbach, 1947; Moens and Steedman, 1987) and includes the characterization of temporal discourse in terms of rhetorical structure and pragmatic relations (Dowty, 1986; Webber, 1987; Passonneau, 1988; Lascarides and Asher, 1993).</Paragraph>
    <Paragraph position="1"> Besides its linguistic significance, temporal analysis has important practical implications. In multidocument summarization, knowledge about the temporal order of events can enhance both the content selection and the summary generation processes (Barzilay et al., 2002). In question answering, temporal analysis is needed to determine when a particular event occurs and how events relate to each other. Some of these needs can be addressed by emerging technologies for temporal analysis (Wilson et al., 2001; Mani et al., 2003; Lapata and Lascarides, 2004; Boguraev and Ando, 2005).</Paragraph>
    <Paragraph position="2"> This paper characterizes the temporal flow of discourse in terms of temporal segments and their ordering. We define a temporal segment to be a fragment of text that does not exhibit abrupt changes in temporal focus (Webber, 1988). A segment may contain more than one event or state, but the key requirement is that its elements maintain temporal coherence. For instance, a medical case summary may contain segments describing a patient's admission, his previous hospital visit, and the onset of his original symptoms. Each of these segments corresponds to a different time frame, and is clearly delineated as such in a text.</Paragraph>
    <Paragraph position="3"> Our ultimate goal is to automatically construct a graph that encodes ordering between temporal segments. The key premise is that in a coherent document, temporal progression is reflected in a wide range of linguistic features and contextual dependencies. In some cases, clues to segment ordering are embedded in the segments themselves.</Paragraph>
    <Paragraph position="4"> For instance, given a pair of adjacent segments, the temporal adverb next day in the second segment is a strong predictor of a precedence relation. In other cases, we can predict the right order between a pair of segments by analyzing their relation to other segments in the text. The interaction between pairwise ordering decisions can easily be formalized in terms of constraints on the graph topology. An obvious example of such a constraint is prohibiting cycles in the ordering graph. We show how these complementary sources of information can be incorporated in a model using global inference.</Paragraph>
    <Paragraph position="5"> We evaluate our temporal ordering algorithm on a corpus of medical case summaries. Temporal  analysis in this domain is challenging in several respects: a typical summary exhibits no significant tense or aspect variations and contains few absolute time markers. We demonstrate that humans can reliably mark temporal segments and determine segment ordering in this domain. Our learning method achieves 83% F-measure in temporal segmentation and 84% accuracy in inferring temporal relations between two segments.</Paragraph>
    <Paragraph position="6"> Our contributions are twofold: Temporal Segmentation We propose a fully automatic, linguistically rich model for temporal segmentation. Most work on temporal analysis is done on a finer granularity than proposed here.</Paragraph>
    <Paragraph position="7"> Our results show that the coarse granularity of our representation facilitates temporal analysis and is especially suitable for domains with sparse temporal anchors.</Paragraph>
    <Paragraph position="8"> Segment Ordering We introduce a new method for learning temporal ordering. In contrast to existing methods that focus on pairwise ordering, we explore strategies for global temporal inference.</Paragraph>
    <Paragraph position="9"> The strength of the proposed model lies in its ability to simultaneously optimize pairwise ordering preferences and global constraints on graph topology. While the algorithm has been applied at the segment level, it can be used with other temporal annotation schemes.</Paragraph>
  </Section>
class="xml-element"></Paper>