<?xml version="1.0" standalone="yes"?>
<Paper uid="P06-2022">
  <Title>Automatically Extracting Nominal Mentions of Events with a Bootstrapped Probabilistic Classifier[?]</Title>
  <Section position="3" start_page="0" end_page="168" type="intro">
    <SectionTitle>
1 Introduction
</SectionTitle>
    <Paragraph position="0"> The goal of information extraction is to generate a set of abstract information objects that represent the entities, events, and relations of particular types mentioned in unstructured text. For example, in a judicial domain, relevant event types might be ARREST, CHARGING, TRIAL, etc.</Paragraph>
    <Paragraph position="1"> Although event extraction techniques usually focus on extracting mentions textually anchored by verb phrases or clauses, e.g. (Aone and Ramos[?] This work was supported in part by SBIR grant FA8750-05-C-0187 from the Air Force Research Laboratory (AFRL)/IFED.</Paragraph>
    <Paragraph position="2"> Santacruz, 2000), many event mentions, especially subsequent mentions of events that are the primary topic of a document, are referred to with nominals. Because of this, detecting nominal event mentions, like those in (1), can increase the recall of event extraction systems, in particular for the most important events in a document.1 (1) The slain journalist was a main organizer of the massive demonstrations that forced Syria to withdraw its troops from Lebanon last April, after Assad was widely  accusedofplanningHariri'sassassinationinaFebruary car bombing that was similar to today's blast. Detecting event nominals is also an important step in detecting relations between event mentions, as in the causal relation between the demonstrations and the withdrawal and the similarity relation between the bombing and the blast in (1).</Paragraph>
    <Paragraph position="3"> Finally, detecting nominal events can improve detection and coreference of non-named mentions of non-event entities (e.g. persons, locations, and organizations) by removing event nominals from consideration as mentions of entities.</Paragraph>
    <Paragraph position="4"> Current extraction techniques for verballyanchored events rest on the assumption that most verb phrases denote eventualities. A system to extract untyped event mentions can output all constituents headed by a non-auxiliary verb with a filter to remove instances of to be, to seem, etc.</Paragraph>
    <Paragraph position="5"> A statistical or rule-based classifier designed to detect event mentions of specific types can then be applied to filter these remaining instances.</Paragraph>
    <Paragraph position="6"> Noun phrases, in contrast, can be used to denote anything--eventualities, entities, abstractions, and only some are suitable for event-type filtering.</Paragraph>
    <Section position="1" start_page="168" end_page="168" type="sub_section">
      <SectionTitle>
1.1 Challenges of nominal event detection
</SectionTitle>
      <Paragraph position="0"> Extraction of nominal mentions of events encompasses many of the fundamental challenges of natural language processing. Creating a general purpose lexicon of all potentially event-denoting terms in a language is a labor-intensive task. On top of this, even utilizing an existing lexical resource like WordNet requires sense disambiguation at run-time because event nominals display the full spectrum of sense distinction behaviors (Copestake and Briscoe, 1995), including idiosyncratic polysemy, as in (2); constructional polysemy, as in (3); coactivation, (4); and copredication, as in (5).</Paragraph>
      <Paragraph position="1">  (2) a. On May 30 a group of Iranian mountaineers hoisted the Iranian tricolor on the summit.</Paragraph>
      <Paragraph position="2"> b. EU Leaders are arriving here for their two-day summit beginning Thursday.</Paragraph>
      <Paragraph position="3"> (3) Things are getting back to normal in the Baywood Golf Club after a chemical spill[=event]. Clean-up crews said the chemical spill[=result] was 99 percent water and shouldn't cause harm to area residents.</Paragraph>
      <Paragraph position="4"> (4) Managing partner Naimoli said he wasn't concerned about recent media criticism.</Paragraph>
      <Paragraph position="5"> (5) The construction lasted 30 years and was inaugurated  in the presence of the king in June 1684.</Paragraph>
      <Paragraph position="6"> Given the breadth of lexical sense phenomena possible with event nominals, no existing approach can address all aspects. Lexical lookup, whether using a manually- or automaticallyconstructed resource, does not take context into consideration and so does not allow for vagueness or unknown words. Purely word-cooccurrence-based approaches (e.g. (Sch&amp;quot;utze, 1998)) are unsuitable for cases like (3) where both senses are possible in a single discourse. Furthermore, most WSD techniques, whether supervised or unsupervised, must be retrained for each individual lexical item, a computationally expensive procedure both at training and run time. To address these limitations, we have developed a technique which combines automatic lexical acquisition and sense disambiguation into a single-pass weakly-supervised algorithm for detecting event nominals.</Paragraph>
      <Paragraph position="7"> The remainder of this paper is organized as follows: Section 2 describes our probabilistic classifier. Section 3 presents experimental results of this model, assesses its performance when bootstrapped to increase its coverage, and compares it to a lexical lookup technique. We describe related work in Section 4 and present conclusions and implications for future work in Section 5.</Paragraph>
    </Section>
  </Section>
class="xml-element"></Paper>