<?xml version="1.0" standalone="yes"?>
<Paper uid="W04-2613">
  <Title>Generating Linear Orders of Text-Based Events</Title>
  <Section position="3" start_page="0" end_page="0" type="intro">
    <SectionTitle>
1 Introduction
</SectionTitle>
    <Paragraph position="0"> The world is a dynamic place and the activities and actions that are part of our everyday experience include such phenomena as the flow of traffic on the morning drive to work, a person walking across a street, or the fluctuation of water bodies due to seasonal change. An interest in developing computational models that convey the dynamic aspects of the world has resulted in a more explicit focus on modeling events, i.e., actions that result in some change to an entity over time, such as the creation, elimination, or transmission of entities (Al-Taha and Barrera, 1994; Claramunt and Theriault, 1995; Claramunt and Theriault, 1996; Medak, 1999; Hornsby and Egenhofer, 2000). The segmentation of real-world happenings into events facilitates the processing and conveying of natural language information to a user (Zacks and Tversky, 2001). An understanding of text is improved if the event structure of the text can be reconstructed. Alfonseca and Manandhar (2002) extract events based on verbs that represent actions, verbs associated with the verb to be which represent states, and occurrences of nouns that are specifications of verbs, such as arrival or accident. Events are anchored in time relative to either the time the text is written or to the main event, and reordered based on this anchoring and verb tenses. Indexing and information retrieval for current and historical reports on infectious disease outbreaks is improved through an approach where events relating to occurrences of infectious diseases are automatically extracted from news sources on the Internet (Grishman et al., 2002). These events are then used to populate a database with the view that coupling the text-based event extraction with a database approach, offers better indexing for reports on infectious disease outbreaks. null Reasoning about events commonly requires assembling the events into a sequence or order of events such that a temporal pattern of events becomes distinguishable (Frank, 1998) and more understandable. For many reasoning tasks involving events, users require a simple, total order of events, where for every pair of events, A and B, either A is before B or B is before A (or both) (Frank, 1998). The case where A is before B and B is before A describes the case where A occurs at the same time as B. This linear sequencing helps us to understand and communicate in a simpler fashion how events occur over time.</Paragraph>
    <Paragraph position="1"> In a typical database, the values of one or more attributes of data may be ordered through queries to the database using a database query language such as SQL.</Paragraph>
    <Paragraph position="2"> Dynamic scenarios as captured in text-based narratives, for example, require alternative approaches to ordering, where orders are based on events or the relations among events even though the knowledge about these relations may be incomplete or uncertain (Pustejovsky et al., 2003). This paper presents methods to generate linear orders of events from more complex orderings of events in text. These linear orders provide simpler, summarized views of the events in a narrative as well as a basis for event-based question and answer systems. Automatic text summarization strategies are necessary to support decision making from large textual narratives as well as the large number of information resources available via the Web (Mani and Maybury, 2001). Temporal semantics and events have been considered in the context of time-based summarization strategies that assist users to monitor changes in news coverage over time (Allan et al., 2001).</Paragraph>
    <Paragraph position="3"> In this paper, entities refer to phenomena in the real world, and an event results in some change to an entity over time. Events are assumed to be linear with both a start point and an end point. Events that are ongoing, i.e., have no end, are not treated here, nor are pre-existing events, i.e., those that have no known start point.</Paragraph>
    <Paragraph position="4"> The objective of this paper is to introduce an approach for automatically generating plausible linear orders of events from partially-ordered sets of event intervals drawn from text descriptions. The remainder of this paper is structured as follows: Section 2 describes events modeled as event intervals and the relations that hold between event intervals. Section 3 presents an approach to generating linear orders of event intervals where the set of 13 possible event interval relations are reduced to either before or equals. In Section 4 an example is introduced to demonstrate this approach. The next section further refines the ordering process by incorporating constraints derived from the semantics of the original event interval relations, and Section 6 uses the example scenario to illustrate the use of these semantics in the ordering process. Section 7 presents the conclusions and discusses future work.</Paragraph>
    <Paragraph position="5"> 2 Events and relations among events Events are often modeled as being instantaneous such as, for example, an update to a bank account or the transmission of an electronic message, i.e., changes of state having no duration (Hinze and Voisard, 2002).</Paragraph>
    <Paragraph position="6"> Alternatively, events may be modeled as occurring over a period of time and therefore have duration. These events are typically associated with a specific point in time (Motakis and Zaniolo, 1995), usually the point at which the event finishes (Galton and Augusto, 2002). In linguistics and cognitive psychology, an event is most often modeled as occurring over a period of time (Larson, 1999; Pedersen and Wright, 2002), and human perceptions of this event time are of particular interest to researchers. For example, in a court case, a suspect's claim about events during a particular time period versus a witness' perception of the same events may affect the jury and make a difference to the outcome of a trial (Pedersen and Wright, 2002).</Paragraph>
    <Paragraph position="7"> The events described in this paper are treated as intervals with a start point and end point, and are assumed to have some duration. For example, LowPressure-Moves refers to an event-the movement of a low pressure system-modeled as an interval. Text, such as a paragraph describing the weather over the past twenty-four hours, can be processed to yield a set of event-relation combinations. Our focus in this paper is not so much on event extraction from text but rather the methods to automatically generate a plausible linear order of events. As event intervals are assumed to be linear, Allen's temporal interval relations (Allen, 1983) describe the set of possible relations that hold among event intervals (Figure 1). For example, scattered showers occur this afternoon before drier air arrives from the west relates two events, ScatteredShowersOccur and DrierAirArrives, by the temporal interval relation before.</Paragraph>
    <Paragraph position="8"> This approach assumes an underlying linear model of time and excludes cyclic or branching time (Frank, 1998; Hornsby et al., 1999).</Paragraph>
    <Paragraph position="9"> Figure 1. Thirteen event interval relations (after Allen 1983).</Paragraph>
    <Paragraph position="10"> Retrieving information about event intervals that shows how events relate to each other is required for an understanding of how entities described in a narrative evolve over space and time. Generating orders of events is necessary such that useful information about events is presented for users, either as a summarization tool for large information resources or as a method for question answering.</Paragraph>
  </Section>
class="xml-element"></Paper>