<?xml version="1.0" standalone="yes"?>
<Paper uid="W01-1313">
  <Title>Assigning Time-Stamps to Event-Clauses</Title>
  <Section position="4" start_page="2" end_page="2" type="metho">
    <SectionTitle>
3 Identifying Events
</SectionTitle>
    <Paragraph position="0"> To divide sentences into event-clauses we use CONTEX (Hermjakob, 1997), a parser that produces a syntactic parse tree augmented with semantic labels. CONTEX uses machine learning techniques to induce a grammar from a given treebanks. A sample output of CONTEX is given in Appendix 1.</Paragraph>
    <Paragraph position="1"> To divide a sentence into event-clauses the parse tree output by CONTEX is analyzed from left to right (root to leaf). The ::CAT field for each node provides the necessary information about whether the node under consideration forms a part of its upper level event or whether it introduces a new event. ::CAT features that indicate new events are: S-CLAUSE, S-SNT, S-SUB-CLAUSE, S-PART-CLAUSE, S-REL-CLAUSE. These features mark clauses which contain both subject (one or several NPs) and predicate (VP containing one or several verbs). The above procedure classifies a clause containing more than one verb as a simple clause. Such clauses are treated as one event and only one time-point will be assigned to them.</Paragraph>
    <Paragraph position="2"> This is fine when the second verb is used in the same tense as the first, but may be wrong in some cases, as in He lives in this house now and will stay here for one more year. There are no such clauses in the analyzed data, so we ignore this complication for the present.</Paragraph>
    <Paragraph position="3"> The parse tree also gives information about the tense of verbs, used later for time assignment. In order to facilitate subsequent processing, we wish to rephrase relative clauses as full independent sentences. We therefore have to replace pronouns where it is possible by their antecedents. Very often the parser gives information about the referential antecedents (in the example below, Russia). Therefore we introduced the rule: if it is possible to identify the referent, put it into the event-clause:  But sometimes the antecedent is identified incorrectly.</Paragraph>
    <Paragraph position="4"> Qulle charged that the United Nations and non-governmental organizations involved in the relief were poorly coordinated, which was costing lives.</Paragraph>
    <Paragraph position="5"> Here the antecedent for which is identified as the relief, and gives which &lt;the relief&gt; was costing lives instead of which &lt;poor coordination&gt; was costing lives. Fortunately, in most cases our rule works correctly.</Paragraph>
    <Paragraph position="6"> Although the event-identifier works reasonably well, breaking text into event-clauses needs further investigation. Table 1 shows the performance of the system. Two kinds of mistakes are made by the event identifier: those caused by CONTEX (it does not identify clauses with omitted predicate, etc.) and those caused by the fact that our clause identifier does too shallow analysis of the parse tree.</Paragraph>
  </Section>
class="xml-element"></Paper>