<?xml version="1.0" standalone="yes"?>
<Paper uid="W04-1307">
  <Title>Statistics Learning and Universal Grammar: Modeling Word Segmentation</Title>
  <Section position="5" start_page="51" end_page="51" type="concl">
    <SectionTitle>
5 Conclusion
</SectionTitle>
    <Paragraph position="0"> Further work, both experimental and computational, will need to address a few pressing questions, in order to gain a better assessment of the relative contribution of SL and UG to language acquisition. These include, more pertinent to the problem of word segmentation: null Can statistical learning be used in the acquisition of language-specific phonotactics, a pre-requisite to syllabification and a prelude to word segmentation? Given that prosodic constraints are critical for the success of SL in word segmentation, future work needs to quantify the availability of stress information in spoken corpora.</Paragraph>
    <Paragraph position="1"> Can further experiments, carried over realistic linguistic input, further tease apart the multiple strategies used in word segmentation [14]? What are the psychological mechanisms (algorithms) that integrate these strategies? How does word segmentation, statistical or otherwise, work for agglutinative (e.g., Turkish) and polysynthetic languages (e.g. Mohawk), where the division between words, morphology, and syntax is quite different from more clear-cut cases like English? Computational modeling can make explicit the balance between statistics and UG, and are in the same vein as the recent findings [24] on when/where SL is effective/possible. UG can help SL by providing specific constraints on its application, and modeling may raise new questions for further experimental studies. In related work [6,7], we have augmented traditional theories of UGderivational phonology, and the Principles and Parameters framework-with a component of statistical learning, with novel and desirable consequences. Yet in all cases, statistical learning, while perhaps domain-general, is constrained by what appears to be innate and domain-specific knowledge of linguistic structures, such that learning can operate on specific aspects of the input evidence</Paragraph>
  </Section>
class="xml-element"></Paper>