<?xml version="1.0" standalone="yes"?>
<Paper uid="P00-1043">
  <Title>Extracting Causal Knowledge from a Medical Database Using Graphical Patterns</Title>
  <Section position="7" start_page="0" end_page="0" type="concl">
    <SectionTitle>
6. Conclusion
</SectionTitle>
    <Paragraph position="0"> We have described a method for performing automatic extraction of cause-effect information from textual documents. We use Conexor's FDG parser to construct a syntactic parse tree for each target sentence. The parse tree is matched with a set of graphical causality patterns that indicate the presence of a causal relation. When a match is found, various attributes of the causal relation (e.g. the cause, the effect, and the modality) can then be extracted and entered in a cause-effect template.</Paragraph>
    <Paragraph position="1"> The accuracy of our extraction system is not yet satisfactory, with an accuracy of about 0.51 (F-measure) for extracting the cause and 0.58 for extracting the effect that are explicitly e x pressed. If both implicit and explicit causal rel a tions are included, the accuracy is 0.41 for cause and 0.48 for effect. We were heartened to find that when the extraction patterns were applied to 2 new medical areas, the extraction precision was the same as for the original 4 medical areas.</Paragraph>
    <Paragraph position="2"> Future work includes:  formation extracted can be chained together to synthesize new knowledge.</Paragraph>
    <Paragraph position="3"> Two aspects of discourse pro cessing is being studied: co-reference resolution and hypothesis confirmation. Co-reference resolution is impo r tant for two reasons. The first is the obvious re a son that to extract complete cause-effect info r mation, pronouns and references have to be resolved and replaced with the information that they refer to. The second reason is that quite o f ten a causal relation between two events is e x pressed more than once in a medical abstract, each time providing new information about the causal situation. The extraction system thus needs to be able to recognize that the different causal expressions refer to the same causal situation, and merge the information extracted from the different sentences.</Paragraph>
    <Paragraph position="4"> The second aspect of discourse processing being investigated is what we refer to as h y pothesis confirmation. Sometimes, a causal rel a tion is hypothesized by the author at the begi n ning of the abstract. This hypothesis may be confirmed or disconfirmed by another sentence later in the abstract. The information extraction system thus has to be able to link the initial h y pothetical cause-effect expression with the co n firmation or disconfirmation expression later in the abstract.</Paragraph>
    <Paragraph position="5"> Finally, we hope eventually to develop a system that not only extracts cause-effect info r mation from medical abstracts accurately, but also synthesizes new knowledge by chaining the extracted causal relations. In a series of studies, Swanson (1986) has demonstrated that logical connections between the published literature of two medical research areas can provide new and useful hypotheses. Suppose an article reports that A causes B, and another article reports that B causes C, then there is an implicit logical link between A and C (i.e. A causes C). This relation would not become explicit unless work is done to extract it. Thus, new discoveries can be made by analysing published literature automatically (Finn, 1998; Swanson &amp; Smalheiser, 1997).</Paragraph>
  </Section>
class="xml-element"></Paper>