<?xml version="1.0" standalone="yes"?>
<Paper uid="P06-2027">
  <Title>Automatic Creation of Domain Templates</Title>
  <Section position="4" start_page="207" end_page="208" type="intro">
    <SectionTitle>
2 Related Work
</SectionTitle>
    <Paragraph position="0"> Our system automatically generates a template that captures the generally most important information for a particular domain and is reusable across multiple instances of that domain. Deciding what slots to include in the template, and what restrictions to place on their potential fillers, is a knowledge representation problem (Hobbs and Israel, 1994). Templates were used in the main IE competitions, the Message Understanding Conferences (Hobbs and Israel, 1994; Onyshkevych, 1994; Marsh and Perzanowski, 1997). One of the recent evaluations, ACE,2 uses pre-defined frames connecting event types (e.g., arrest, release) to a set of attributes. The template construction task was not addressed by the participating systems.</Paragraph>
    <Paragraph position="1"> The domain templates were created manually by experts to capture the structure of the facts sought.</Paragraph>
    <Paragraph position="2"> Although templates have been extensively used in information extraction, there has been little work on their automatic design. In the Conceptual Case Frame Acquisition project (Riloff and Schmelzenbach, 1998), extraction patterns, a domain semantic lexicon, and a list of conceptual roles and associated semantic categories for the domain are used to produce multiple-slot case frames with selectional restrictions. The system requires two sets of documents: those relevant to the domain and those irrelevant. Our approach does not require any domain-specific knowledge and uses only corpus-based statistics.</Paragraph>
    <Paragraph position="3"> The GISTexter summarization system (Harabagiu and Maiorano, 2002) used statistics over an arbitrary document collection together with semantic relations from WordNet.</Paragraph>
    <Paragraph position="4"> The created templates heavily depend on the topical relations encoded in WordNet. The template models an input collection of documents. If there is only one domain instance described in the input than the template is created for this particular instance rather than for a domain. In our work, we learn domain templates by cross-examining several collections of documents on the same topic, aiming for a general domain template. We rely on relations cross-mentioned in different instances of the domain to automatically prioritize roles and relationships for selection.</Paragraph>
    <Paragraph position="5"> Topic Themes (Harabagiu and LVacVatus,u, 2005) used for multi-document summarization merge various arguments corresponding to the same se- null mantic roles for the semantically identical verb phrases (e.g., arrests and placed under arrest).</Paragraph>
    <Paragraph position="6"> Atomic events also model an input document collection (Filatova and Hatzivassiloglou, 2003) and are created according to the statistics collected for co-occurrences of named entity pairs linked through actions. GISTexter, atomic events, and Topic Themes were used for modeling a collection of documents rather than a domain.</Paragraph>
    <Paragraph position="7"> In other closely related work, Sudo et al. (2003) use frequent dependency subtrees as measured by TF*IDF to identify named entities and IE patterns important for a given domain. The goal of their work is to show how the techniques improve IE pattern acquisition. To do this, Sudo et al. constrain the retrieval of relevant documents for a MUC scenario and then use unsupervised learning over descriptions within these documents that match specific types of named entities (e.g., Arresting Agency, Charge), thus enabling learning of patterns for specific templates (e.g., the Arrest scenario). In contrast, the goal of our work is to show how similar techniques can be used to learn what information is important for a given domain or event and thus, should be included into the domain template. Our approach allows, for example, learning that an arrest along with other events (e.g., attack) is often part of a terrorist event. We do not assume any prior knowledge about domains. We demonstrate that frequent subtrees can be used not only to extract specific named entities for a given scenario but also to learn domain-important relations. These relations link domain actions and named entities as well as general nouns and words belonging to other syntactic categories.</Paragraph>
    <Paragraph position="8"> Collier (1998) proposed a fully automatic method for creating templates for information extraction. The method relies on Luhn's (1957) idea of locating statistically significant words in a corpus and uses those to locate the sentences in which they occur. Then it extracts Subject-Verb-Object patterns in those sentences to identify the most important interactions in the input data. The system was constructed to create MUC templates for terrorist attacks. Our work also relies on corpus statistics, but we utilize arbitrary syntactic patterns and explicitly use multiple domain instances. Keeping domain instances separated, we cross-examine them and estimate the importance of a particular information type in the domain.</Paragraph>
  </Section>
class="xml-element"></Paper>