<?xml version="1.0" standalone="yes"?>
<Paper uid="A92-1043">
  <Title>Learning a Scanning Understanding for &amp;quot;Real-world&amp;quot; Library Categorization</Title>
  <Section position="3" start_page="0" end_page="251" type="metho">
    <SectionTitle>
3 The Architecture of the Recurrent
</SectionTitle>
    <Paragraph position="0"> The semantic classification was learned by using a connectionist recurrent plausibility network. A recurrent plausibility network is similar to a simple recurrent network (Elman 89) but instead of learning to predict words, recurrent connections support the assignment of plausible classes (see figure 1). The recurrent plausibility network was trained in a supervised mode using  the backpropagation learning algorithm (Rumelhart et al. 86). In each training step the feature representation of a word and its preceding context was presented to the network in the word bank and context bank together with the desired class. A unit in the output layer received the value 1 if the unit represented the particular class of the title, otherwise the unit received the value 0. The real-valued hidden layer represented the context of preceding words. At the beginning of a title the context bank was initialized with values of 0 since there was no preceding context. After the first word had been presented the context bank was initialized with the values of the hidden layer that encoded the reduced preceding context.</Paragraph>
  </Section>
class="xml-element"></Paper>