<?xml version="1.0" standalone="yes"?>
<Paper uid="W06-0906">
  <Title>Extending TimeML with Typical Durations of Events</Title>
  <Section position="5" start_page="42" end_page="43" type="metho">
    <SectionTitle>
3 TimeML and Its Event Classes
</SectionTitle>
    <Paragraph position="0"> TimeML (Pustejovsky et al., 2003) is a rich specification language for event and temporal expressions in natural language text. Unlike most previous attempts at event and temporal specification, TimeML separates the representation of event and temporal expressions from the anchoring or ordering dependencies that may exist in a given text.</Paragraph>
    <Paragraph position="1"> TimeML includes four major data structures: EVENT, TIMEX3, SIGNAL, AND LINK.</Paragraph>
    <Paragraph position="2"> EVENT is a cover term for situations that happen or occur, and also those predicates describing states or circumstances in which something obtains or holds true. TIMEX3, which extends TIMEX2 (Ferro, 2001), is used to mark up explicit temporal expressions, such as time, dates, and durations. SIGNAL is used to annotate sections of text, typically function words that indicate how temporal objects are related to each other (e.g., &amp;quot;when&amp;quot;, &amp;quot;during&amp;quot;, &amp;quot;before&amp;quot;). The set of LINK tags encode various relations that exist between the temporal elements of a document, including three subtypes: TLINK (temporal links), SLINK (subordination links), and ALINK (aspectual links).</Paragraph>
    <Paragraph position="3"> Our event duration annotations can be integrated into the EVENT tag. In TimeML each event belongs to one of the seven event classes, i.e., reporting, perception, aspectual, I-action, Istate, state, occurrence. TimeML annotation guidelines  give detailed description for each of the classes: Reporting. This class describes the action of a person or an organization declaring something, narrating an event, informing about an event, etc (e.g., say, report, tell, explain, state).</Paragraph>
    <Paragraph position="4"> Perception. This class includes events involving the physical perception of another event (e.g., see, watch, view, hear).</Paragraph>
    <Paragraph position="5"> Aspectual. In languages such as English and French, there is a grammatical device of aspectual predication, which focuses on different facets of event history, i.e., initiation, reinitiation, termination, culmination, continuation (e.g., begin, stop, finish, continue).</Paragraph>
    <Paragraph position="6"> I-Action. An I-Action is an Intensional Action. It introduces an event argument (which must be in the text explicitly) describing an action or situation from which we can infer something given its relation with the I-Action (e.g., attempt, try, promise).</Paragraph>
    <Paragraph position="7"> I-State. This class of events are similar to the previous class. This class includes states that refer to alternative or possible worlds (e.g., believe, intend, want).</Paragraph>
    <Paragraph position="8"> State. This class describes circumstances in which something obtains or holds true (e.g., on board, kidnapped, peace).</Paragraph>
    <Paragraph position="9"> Occurrence. This class includes all the many other kinds of events describing something that happens or occurs in the world (e.g., die, crash, build, sell).</Paragraph>
  </Section>
  <Section position="6" start_page="43" end_page="43" type="metho">
    <SectionTitle>
4 Integrating Event Duration Annota-
</SectionTitle>
    <Paragraph position="0"> tions into TimeML Our event duration annotations can be integrated into TimeML by adding two more attributes to the EVENT tag for the lower bound and upper bound duration annotations (e.g., &amp;quot;lowerBound-Duration&amp;quot; and &amp;quot;upperBoundDuration&amp;quot; attributes). To minimize changes of the existing TimeML specifications caused by the integration, we can try to share as much as possible our event classes as described in Section 2.2 with the existing ones in TimeML as described in Section 3.</Paragraph>
    <Paragraph position="1"> We can see that four event classes are shared with very similar definitions, i.e., reporting, aspectual, state, and action/occurrence. For the other three event classes that only belong to TimeML (i.e., perception, I-action, I-state), the I-action and perception classes can be treated as special subclasses of the action/occurrence class, and the I-state class as a special subclass of the state class.</Paragraph>
    <Paragraph position="2"> However, there are still three classes that only belong to the event duration annotations (i.e., multiple, negation, and positive infinite). The positive infinite class can be treated as a special subclass of the state class with a special duration annotation for positive infinity.</Paragraph>
    <Paragraph position="3"> Each multiple event has two annotations. For example, for Iraq has destroyed its long-range missiles.</Paragraph>
    <Paragraph position="4"> there is the time it takes to destroy one missile and the duration of the interval in which all the individual events are situated - the time it takes to destroy all its missiles.</Paragraph>
    <Paragraph position="5"> Since the single event is usually more likely to be encountered in multiple documents, and thus the duration of the single event is usually more likely to be shared and re-used, to simplify the specification, we can take only the duration annotation of the single events for the multiple event class, and the single event can be assigned with one of the seven TimeML event classes. For example, the &amp;quot;destroyed&amp;quot; event in the above example is assigned with the occurrence class in TimeBank.</Paragraph>
    <Paragraph position="6"> The events involving negation can be simplified similarly. Since the event negated is usually more likely to be encountered in multiple documents, we can take only the duration annotation of the negated event for this class. For example, in He was willing to withdraw troops in exchange for guarantees that Israel would not be attacked.</Paragraph>
    <Paragraph position="7"> the event negated is the &amp;quot;being attacked&amp;quot; event and it is assigned with the occurrence class in TimeBank. Alternatively, TimeML could be extended to treat negations of events as states. The format used for annotated durations is consistent with that for the value of the DURA-TION type in TimeML. For example, the sentence null The official said these sites could only be visited by a special team of U.N. monitors and diplomats. null can be marked up in TimeML as: The official &lt;EVENT eid=&amp;quot;e63&amp;quot; class=&amp;quot;REPORTING&amp;quot;&gt; said &lt;/EVENT&gt; these sites &lt;SIGNAL sid=&amp;quot;s65&amp;quot; &gt;could&lt;/SIGNAL&gt; only be &lt;EVENT eid=&amp;quot;e64&amp;quot; class=&amp;quot;OCCURRENCE&amp;quot;&gt; visited &lt;/EVENT&gt; by a special team of &lt;ENAMEX TYPE=&amp;quot;ORGANIZATION&amp;quot;&gt; U.N.</Paragraph>
    <Paragraph position="8"> &lt;/ENAMEX&gt; monitors and diplomats.</Paragraph>
    <Paragraph position="9"> If we annotate the &amp;quot;said&amp;quot; event with the duration annotation of [5 seconds, 5 minutes], and the &amp;quot;visited&amp;quot; event with [10 minutes, 1 day], the ex- null special team of &lt;ENAMEX TYPE=&amp;quot;ORGANIZATION&amp;quot;&gt; U.N. &lt;/ENAMEX&gt; monitors and diplomats.</Paragraph>
  </Section>
class="xml-element"></Paper>