<?xml version="1.0" standalone="yes"?>
<Paper uid="N06-1009">
  <Title>Role of Local Context in Automatic Deidentification of Ungrammatical, Fragmented Text</Title>
  <Section position="3" start_page="65" end_page="66" type="intro">
    <SectionTitle>
2 Related Work
</SectionTitle>
    <Paragraph position="0"> In the literature, named entities such as people, places, and organizations mentioned in news articles have been successfully identified by various approaches (Bikel et al., 1999; McCallum et al., 2000; Riloff and Jones, 1996; Collins and Singer, 1999; Hobbs et al., 1996). Most of these approaches are tailored to a particular domain, e.g., understanding disaster news; they exploit both the characteristics of the entities they focus on and the contextual clues related to these entities.</Paragraph>
    <Paragraph position="1"> In the biomedical domain, NER has focused on identification of biological entities such as genes and proteins (Collier et al., 2000; Yu et al., 2002).</Paragraph>
    <Paragraph position="2"> Various statistical approaches, e.g., a maximum entropy model (Finkel et al., 2004), HMMs and SVMs (GuoDong et al., 2005), have been used with various feature sets including surface and syntactic features, word formation patterns, morphological patterns, part-of-speech tags, head noun triggers, and coreferences.</Paragraph>
    <Paragraph position="3"> Deidentification refers to the removal of identifying information from records. Some approaches to deidentification have focused on particular categories of PHI, e.g., Taira et al. focused on only patient names (2002), Thomas et al. focused on proper names including doctors' names (2002). For full deidentification, i.e., removal of all PHI, Gupta et al. used &amp;quot;a complex set of rules, dictionaries, pattern-matching algorithms, and Unified Medical Language System&amp;quot; (2004). Sweeney's Scrub system employed competing algorithms that used patterns and lexicons to find PHI. Each of the algorithms included in her system specialized in one kind of PHI, each calculated the probability that a given word belonged to the class of PHI that it specialized in, and the algorithm with the highest precedence and the highest probability labelled the given word. This system identified 99-100% of all PHI in the test corpus of patient records and letters to physicians (1996).</Paragraph>
    <Paragraph position="4"> We use a variety of features to train a support vector machine (SVM) that can automatically extract local context cues and can recognize PHI (even when some PHI are ambiguous between PHI and non-PHI, and even when PHI do not appear in dictionaries). We compare this approach with three others: a heuristic rule-based approach (Douglass, 2005), the SNoW (Sparse Network of Winnows) system's NER component (Roth and Yih, 2002), and IdentiFinder (Bikel et al., 1999). The heuristic rule-based system relies heavily on dictionaries. SNoW and IdentiFinder consider some representation of the local context of words; they also rely on information about global context. Local context helps them recognize stereotypical names and name structures.</Paragraph>
    <Paragraph position="5"> Global context helps these systems update the probability of observing a particular entity type based on the other entity types contained in the sentence. We hypothesize that, given the mostly fragmented and ungrammatical nature of discharge summaries, local context will be more important for deidentification than global context. We further hypothesize that local context will be a more reliable indication of PHI than dictionaries (which can be incomplete). The results presented in this paper show that SVMs trained with a statistical representation of local context out-perform all baselines. In other words, a classifier that relies heavily on local context (very little on dictionaries, and not at all on global context) out-performs classifiers that rely either on global context or dictionaries (but make much less use of local context). Global context cannot contribute much to deidentification when the language of documents is fragmented; dictionaries cannot contribute to deidentification when PHI are either missing from dictionaries or are ambiguous between PHI and non-PHI. Local context remains a reliable indication of PHI under these circumstances.</Paragraph>
    <Paragraph position="6"> The features used for our SVM-based system can be enriched in order to automatically acquire more and varied local context information. The features discussed in this paper have been chosen because of their simplicity and effectiveness on both grammatical and ungrammatical free text.</Paragraph>
  </Section>
class="xml-element"></Paper>