<?xml version="1.0" standalone="yes"?>
<Paper uid="A00-2007">
  <Title>Noun Phrase Recognition by System Combination</Title>
  <Section position="6" start_page="53" end_page="53" type="relat">
    <SectionTitle>
4 Related work
</SectionTitle>
    <Paragraph position="0"> (Abney, 1991) has proposed to approach parsing by starting with finding correlated chunks of words.</Paragraph>
    <Paragraph position="1"> The chunks can be combined to trees by a second processing stage, the attacher. (Ramshaw and Marcus, 1995) have build a chunker by applying transformation-based learning to sections of the Penn Treebank. Rather than working with bracket structures, they have represented the chunking task as a tagging problem. POS-like tags were used to account for the fact that words were inside or outside chunks. They have applied their method to two segments of the Penn Treebank and these are still being used as benchmark data sets.</Paragraph>
    <Paragraph position="2"> Several groups have continued working with the Ramshaw and Marcus data sets for base noun phrases. (Argamon et al., 1998) use Memory-Based Sequence Learning for recognizing both NP chunks and VP chunks. This method records POS tag sequences which contain chunk boundaries and uses these sequences to classify the test data. Its performance is somewhat worse than that of Ramshaw and Marcus (F~=1=91.6 vs. 92.0) but it is the best result obtained without using lexical information 6.</Paragraph>
    <Paragraph position="3"> (Cardie and Pierce, 1998) store POS tag sequences that make up complete chunks and use these sequences as rules for classifying unseen data. This approach performs worse than the method of Argamon et al. (F~=1=90.9).</Paragraph>
    <Paragraph position="4"> Three papers mention having used the memory-based learning method IBI-IG. (Veenstra, 1998) introduced cascaded chunking, a two-stage process in which the first stage classifications are used to improve the performance in a second processing stage.</Paragraph>
    <Paragraph position="5"> This approach reaches the same performance level as Argamon et al. but it requires lexical information. (Daelemans et al., 1999a) report a good performance for baseNP recognition but they use a different data set and do not mention precision and recall rates. (Tjong Kim Sang and Veenstra, 1999) compare different data representations for this task.</Paragraph>
    <Paragraph position="6"> Their baseNP results are slightly better than those of Ramshaw and Marcus (F~=1=92.37).</Paragraph>
    <Paragraph position="7"> (XTAG, 1998) describes a baseNP chunker built from training data by a technique called supertagging. The performance of the chunker was an improvement of the Ramshaw and Marcus results (Fz=I =92.4). (Mufioz et al., 1999) use SNOW, a network of linear units, for recognizing baseNP phrases 6We have applied majority voting of five data representations to the Ramshaw and Marcus data set without using lexical information and the results were: accuracy O: 97.60%, accuracy C: 98.10%, precision: 92.19%, recall: 91.53% and F~=I: 91.86.</Paragraph>
    <Paragraph position="8"> and SV phrases. They compare two data representations and report that a representation with bracket structures outperforms the IOB tagging representation introduced by (Ramshaw and Marcus, 1995). SNoW reaches the best performance on this task (Fz=I =92.8).</Paragraph>
    <Paragraph position="9"> There has been less work on identifying general noun phrases than on recognizing baseNPs. (Osborne, 1999) extended a definite clause grammar with rules induced by a learner that was based upon the maximum description length principle. He processed other parts of the Penn Treebank than we with an F~=I rate of about 60. Our earlier effort to process the CoNLL data set was performed in the same way as described in this paper but without using the combination method for baseNPs. We obtained an F~=I rate of 82.98 (CoNLL-99, 1999).</Paragraph>
  </Section>
class="xml-element"></Paper>