<?xml version="1.0" standalone="yes"?>
<Paper uid="W00-0728">
  <Title>A Context Sensitive Maximum Likelihood Approach to Chunking</Title>
  <Section position="4" start_page="136" end_page="136" type="metho">
    <SectionTitle>
4 Discussion
</SectionTitle>
    <Paragraph position="0"> The memory-based approach seemingly postulates innate tags in the processing machinery.</Paragraph>
    <Paragraph position="1"> The author has found very little discussion on how the tags are thought to correspond to reality, a fact that was also pointed out, not so long ago, by Palmeri (1998). However, a few papers aiming towards automatic 'label', 'feature' or 'tag' creation are available (Miikkulainen and Dyer, 1991; Johansson, 1999).</Paragraph>
    <Paragraph position="2"> It is undeniable that, from a practical perspective, it is possible to reach very high performance on tasks, such as tagging, that demand a choice from a known set of alternatives by estimating statistical properties (e.g., the most likely label) from a large enough training set.</Paragraph>
    <Paragraph position="3"> This makes the method extremely useful for quick development of tools, which can be used in practical applications such as text retrieval and machine translation; but also in linguistic research; e.g., finding examples of specific grammatical constructions in large collections of data.</Paragraph>
    <Paragraph position="4"> A challenge for future research is how tags could be constructed automatically, and what kind of information would be necessary to detect the relevant tag dimensions for some linguistically motivated task.</Paragraph>
  </Section>
class="xml-element"></Paper>