<?xml version="1.0" standalone="yes"?>
<Paper uid="W04-0110">
  <Title>Segment Predictability as a Cue in Word Segmentation: Application to Modern Greek</Title>
  <Section position="4" start_page="9" end_page="9" type="evalu">
    <SectionTitle>
3 Results
</SectionTitle>
    <Paragraph position="0"> The four model variants (global MI, global TP, local MI, and local TP) were each evaluated on three metrics: word boundaries, word tokens, and word types. Note that the first metric reported, simple boundary placement, considers only utterance-internal word-boundaries, rather than including those word boundaries which are detected 'for free' by virtue of being utteranceboundaries also. This boundary measure may be more conservative than that reported by other authors, but is easily convertible into other metrics. The second metric, the percentage of word tokens detected, is the same as Brent (1999a). In order for a word to be counted as correctly found, three conditions must be met: (a) the word's beginning (left boundary) is correctly detected, (b) the word's ending (right boundary) is correctly detected, and (c) these two are consecutive (i.e., no false boundaries are posited within the word).</Paragraph>
    <Paragraph position="1"> The last metric (word type) is slightly more conservative than Brent's (1999a) in that the word type must have been actually spoken in the same utterance (not the same block of 500 utterances) in which it was detected to count as a match. This lessens the possibility that a mismatch that happens to be segmentally identical to an actual word (but whose semantic context may not be conducive to learning its correct meaning) is counted as a match.</Paragraph>
    <Paragraph position="2"> However, this situation is presumably rather rare.</Paragraph>
    <Paragraph position="3"> Tables 2 and 3 present the results over the test set for both the global and the local comparisons of the predictability statistics proposed by Saffran et al. (1996) and Brent (1999a).</Paragraph>
  </Section>
class="xml-element"></Paper>