<?xml version="1.0" standalone="yes"?>
<Paper uid="W99-0628">
  <Title>PP-Attachment: A Committee Machine Approach</Title>
  <Section position="4" start_page="231" end_page="232" type="intro">
    <SectionTitle>
ENTITY
OBJECT
SUBSTANCE
FOOD
GREEN GOODS
EDIBLE FRUIT
</SectionTitle>
    <Paragraph position="0"> Most of the statistical methods that have used classes do not carry out a prior disambiguation of the words \[Brill, Resnick 1994\], \[Ratnaparkhi et. al 1994\] and others, nor do they determine the adequate level of abstraction. Some that do make the determination have a poor level of efficiency.</Paragraph>
    <Paragraph position="1"> Table 1 shows the accuracy of the results reported in previous work. The worst results were obtained when only classes were used.</Paragraph>
    <Paragraph position="2"> Stettina and Nagao used the Ratnaparkhi data set but they eliminated 3,224 4-tuples (15~) from the training set containing contradicting examples.</Paragraph>
    <Paragraph position="3"> For reasons of complexity, the complete 4-tuple has not been considered simultaneously except in \[Ratnaparkhi et. al 1994\].</Paragraph>
    <Paragraph position="4"> Classes of a given sense and classes of different senses of different words can have complex interactions and the preceding methods cannot take such interactions into account.</Paragraph>
    <Paragraph position="5"> Neural networks (NNs) are appropriates in dealing with this complexity. A very impor- null reported in previous works.</Paragraph>
    <Paragraph position="6"> tant characteristic of NNs is their capacity to deal with multidimensional inputs. They need much fewer parameters to achieve the same result than traditional numerical methods. Recently \[Barton, 1993\] has shown that feedforward networks with one layer of sigmoidal nonlinearities achieve an integrated squared error of order O(1/4) for input spaces of dimension d, where n is the number of units of the network.</Paragraph>
    <Paragraph position="7"> Traditional methods (series expansions) with n terms can only achieve an integrated squared er0( (1~2d~ ror of order ~j j, for functions satisfying the same smoothness assumption. NNs are surprisingly advantageous in high dimensional inputs since the integrated squared error is independent of the input dimension. They can compute very complex statistical functions, they are model free, and compared to the current methods used by the statistical approach to NLP, NNs offer the possibility of dealing with a more complex (non-linear and multivariant) approach.</Paragraph>
    <Paragraph position="8"> In the next section we describe a PP attachment disambiguation system based on neural networks that takes better advantage of the use of classes.</Paragraph>
  </Section>
class="xml-element"></Paper>