<?xml version="1.0" standalone="yes"?>
<Paper uid="W03-1315">
  <Title>An Investigation of Various Information Sources for Classifying Biological Names</Title>
  <Section position="8" start_page="6" end_page="6" type="concl">
    <SectionTitle>
6 Conclusions and Future Work
</SectionTitle>
    <Paragraph position="0"> We have considered a few name internal and external sources of information that help in the classification of names. Despite using fairly simple methods to classify the names, we have obtained encouraging results which we take to suggest that that our intuitions about them are on the right track. We feel that the effectiveness of f-terms and suffixes that generalize the idea of f-terms, the matching algorithm that places more emphasis on partial matches of words to the right vindicates our stance that the classification of names is a task sufficiently distinct from the name identification process and warrants an independent investigation. Even the use of context is different for the two tasks as in the latter task only the immediately neighboring words are used and that too for purpose of demarking the extremities of the name string.</Paragraph>
    <Paragraph position="1"> While the high precision of f-terms and suffix based classification was expected, the recall of these methods was higher than expected. It is also clear that these methods do not help much with the chemicals class. We believe that in addition to suffix, the knowledge of other chemical root forms (such as &amp;quot;meth&amp;quot;), e.g., used in (Narayanaswamy et al., 2003), would be useful.</Paragraph>
    <Paragraph position="2"> We would like to focus more on the matching part of the work. In particular, rather than hand-coding our intuitions in terms of weights for the different parameters, we would like to automatically, e.g., using a held-out validation set, have these set and see to what extent the automated choice of parameters show the bias for the rightmost words in the matching. We would also like to generalize our work further by not limiting the classes to the ones chosen here but allow a wider range of classes. To do this, we would like to consider the GENIA classes and collapse classes at various levels of their ontology and try to see at what level of fine-grained distinctions can classification still be done satisfactorily. In regards to the use of the contextual method, while we have some preliminary ideas, we need to investigate further why the use of a single strong clue, as exemplified by the decision list method, does not work as well as it seems to for the WSD task.</Paragraph>
  </Section>
class="xml-element"></Paper>