<?xml version="1.0" standalone="yes"?>
<Paper uid="P01-1055">
  <Title>Using Machine Learning to Maintain Rule-based Named-Entity Recognition and Classification Systems Georgios Petasis +, Frantz Vichot SS, Francis Wolinski SS</Title>
  <Section position="4" start_page="0" end_page="0" type="relat">
    <SectionTitle>
2 Related Work
</SectionTitle>
    <Paragraph position="0"> As mentioned above, the exploitation of learning techniques to support the domain adaptation of NERC systems has recently attracted the attention of several researchers. Some of these approaches are briefly discussed in this section.</Paragraph>
    <Paragraph position="1"> Nymble (Bikel et al., 1997) uses statistical learning to acquire a Hidden Markov Model (HMM) that recognises NEs in text. Nymble did particularly well in the MUC-7 competition (DARPA, 1998), due mainly to the use of the correct features in the encoding of words, e.g.</Paragraph>
    <Paragraph position="2"> capitalisation, and the probabilistic modelling of the recognition system.</Paragraph>
    <Paragraph position="3"> Named-entity recognition in Alembic (Vilain and Day, 1996) uses the transformation-based rule learning approach introduced in Brill's work on part-of-speech tagging (Brill, 1993). An important aspect of this approach is the fact that the system learns rules that can be freely intermixed with hand-engineered ones.</Paragraph>
    <Paragraph position="4"> The RoboTag system presented in (Bennett et al., 1997) constructs decision trees that classify words as being start or end points of a particular named-entity type. A variant of this approach was used in the system presented by the  (Cuchiarelli et al., 1998), uses unsupervised learning to expand a manually constructed system and improve its performance. The learning algorithm tries to supplement the manually constructed system by classifying recognised but unclassified NEs. In (Petasis et al., 2000) the manually constructed system was replaced by the supervised tree induction algorithm C4.5 (Quinlan, 1993), reaching very good performance on the MUC-6 corpora.</Paragraph>
    <Paragraph position="5"> The partially supervised multi-level bootstrapping approach presented in (Riloff and Jones, 1999) induces a set of information extraction patterns, which can be used to identify and classify NEs. The system starts by generating exhaustively all candidate extraction patterns, using an earlier system called AutoSlog (Riloff, 1993). Given a small number of seed examples of NEs, the most useful patterns for recognising the seed examples are selected and used to expand the set of classified NEs. The end result is a dictionary of NEs and the extraction patterns that correspond to them.</Paragraph>
    <Paragraph position="6"> Our method follows an alternative innovative approach to the use of learning for NERC. Instead of using ML to construct a NERC system that will be used autonomously, the system constructed by ML, according to our approach is used to monitor the performance of an existing rule-based NERC system. In this manner, the new system provides feedback on whether the rule-based system under control has become obsolete and needs to be updated. An important advantage of this approach is that no manual tagging of training data is needed, despite the use of a supervised learning algorithm.</Paragraph>
    <Paragraph position="7"> Our method bears some similarities with systems based on active learning (Thompson et al., 1999). According to this technique, multiple classifiers performing the same task are used in order to actively create training data, through their disagreements. Usually, this involves an iterative procedure. First a few initial labelled examples are used to train the classifiers and then, unlabelled examples are presented to the classifiers. Examples that cause the classifiers to disagree are good candidates to retrain the classifiers on. The difference of active learning to our method is the use of a manually-constructed rule-based NERC system as the basic system.</Paragraph>
    <Paragraph position="8"> The ML method is used only to identify when the rule-based NERC system should be updated, but not for creating new training instances. Another approach, which bears some similarity to ours, is presented in (Kushmerick, 1999) where a heuristic algorithm is used to monitor the performance of web-page wrappers.</Paragraph>
  </Section>
class="xml-element"></Paper>