<?xml version="1.0" standalone="yes"?>
<Paper uid="W06-1623">
  <Title>Inducing Temporal Graphs</Title>
  <Section position="5" start_page="189" end_page="189" type="relat">
    <SectionTitle>
2 Related Work
</SectionTitle>
    <Paragraph position="0"> Temporal ordering has been extensively studied in computational linguistics (Passonneau, 1988; Webber, 1988; Hwang and Schubert, 1992; Lascarides and Asher, 1993; Lascarides and Oberlander, 1993). Prior research has investigated a variety of language mechanisms and knowledge sources that guide interpretation of temporal ordering, including tense, aspect, temporal adverbials, rhetorical relations and pragmatic constraints. In recent years, the availability of annotated corpora, such as TimeBank (Pustejovsky et al., 2003), has triggered the use of machine-learning methods for temporal analysis (Mani et al., 2003; Lapata and Lascarides, 2004; Boguraev and Ando, 2005). Typical tasks include identification of temporal anchors, linking events to times, and temporal ordering of events.</Paragraph>
    <Paragraph position="1"> Since this paper addresses temporal ordering, we focus our discussion on this task. Existing ordering approaches vary both in terms of the ordering unit -- it can be a clause, a sentence or an event -- and in terms of the set of ordering relations considered by the algorithm. Despite these differences, most existing methods have the same basic design: each pair of ordering units (i.e., clauses) is abstracted into a feature vector and a supervised classifier is employed to learn the mapping between feature vectors and their labels. Features used in classification include aspect, modality, event class, and lexical representation. It is important to note that the classification for each pair is performed independently and is not guaranteed to yield a globally consistent order.</Paragraph>
    <Paragraph position="2"> In contrast, our focus is on globally optimal temporal inference. While the importance of global constraints has been previously validated in symbolic systems for temporal analysis (Fikes et al., 2003; Zhou et al., 2005), existing corpus-based approaches operate at the local level. These improvements achieved by a global model motivate its use as an alternative to existing pairwise methods.</Paragraph>
  </Section>
class="xml-element"></Paper>