<?xml version="1.0" standalone="yes"?>
<Paper uid="E99-1023">
  <Title>Representing Text Chunks</Title>
  <Section position="5" start_page="176" end_page="177" type="relat">
    <SectionTitle>
4 Related work
</SectionTitle>
    <Paragraph position="0"> The concept of chunking was introduced by Abney in (Abney, 1991). He suggested to develop a chunking parser which uses a two-part syntactic analysis: creating word chunks (partial trees) and attaching the chunks to create complete syntactic trees. Abney obtained support for such a chunking stage from psycholinguistic literature.</Paragraph>
    <Paragraph position="1"> Ramshaw and Marcus used transformation-based learning (TBL) for developing two chunkers (Ramshaw and Marcus, 1995). One was trained to recognize baseNPs and the other was trained to recognize both NP chunks and VP chunks.</Paragraph>
    <Paragraph position="2"> Ramshaw and Marcus approached the chunking task as a tagging problem. Their baseNP training and test data from the Wall Street Journal corpus are still being used as benchmark data for current chunking experiments. (Ramshaw and Marcus, 1995) shows that baseNP recognition (Fz=I =92.0) is easier than finding both NP and VP chunks (Fz=1=88.1) and that increasing the size of the training data increases the performance on the test set.</Paragraph>
    <Paragraph position="3"> The work by Ramshaw and Marcus has inspired three other groups to build chunking algorithms.</Paragraph>
    <Paragraph position="4"> (Argamon et al., 1998) introduce Memory-Based Sequence Learning and use it for different chunking experiments. Their algorithm stores sequences of POS tags with chunk brackets and uses this information for recognizing chunks in unseen data.</Paragraph>
    <Paragraph position="5"> It performed slightly worse on baseNP recognition than the (Ramshaw and Marcus, 1995) experiments (Fz=1=91.6). (Cardie and Pierce, 1998) uses a related method but they only store POS tag sequences forming complete baseNPs. These sequences were applied to unseen tagged data aIter which post-processing repair rules were used for fixing some frequent errors. This approach performs worse than othe.r reported approaches (Fo=I =90.9).</Paragraph>
    <Paragraph position="6">  fourth experiment series. The accuracy rate contains the fraction of chunk tags that was correct. The other three rates regard baseNP recognition. The bottom part of the table shows some other reported results with this data set. With all but two formats IBI-IG achieves better FZ=l rates than the best published result in (Ramshaw and Marcus, 1995).</Paragraph>
    <Paragraph position="7"> (Veenstra, 1998) uses cascaded decision tree learning (IGTree) for baseNP recognition. This algorithm stores context information of words, POS tags and chunking tags in a decision tree and classifies new items by comparing them to the training items. The algorithm is very fast and it reaches the same performance as (Argamon et al., 1998) (F,~=1=91.6). (Daelemans et al., 1999) uses cascaded MBL (IBI-IG) in a similar way for several tasks among which baseNP recognition. They do not report F~=~ rates but their tag accuracy rates are a lot better than accuracy rates reported by others. However, they use the (Ramshaw and Marcus, 1995) data set in a different training-test division (10-fold cross validation) which makes it (tifficult to compare their results with others.</Paragraph>
  </Section>
class="xml-element"></Paper>