<?xml version="1.0" standalone="yes"?>
<Paper uid="N03-3001">
  <Title>Semantic Language Models for Topic Detection and Tracking</Title>
  <Section position="8" start_page="90" end_page="90" type="concl">
    <SectionTitle>
7 Conclusions and Future work
</SectionTitle>
    <Paragraph position="0"> In this work, we have presented a novel approach for link detection task of TDT. The new approach has three key ideas, namely modeling the relative importance of terms by their semantic classes through a new semantic language modeling approach, casting the link detection task as a two-class classi cation problem and learning the optimum linear discriminant function using the perceptron learning algorithm. We believe this is one of the earliest works that attempts incorporating semantic information into the language modeling framework. Although we have built the model speci cally for the link detection task, it is general enough to be extended to the other tasks of TDT such as Tracking, New Event Detection and Clustering.</Paragraph>
    <Paragraph position="1"> The results on train and test sets indicate that there is a little or no improvement in the performance from the new model as compared to the unigram approach.</Paragraph>
    <Paragraph position="2"> As part of our future work, we would like to understand the reasons behind the unsatisfactory performance of the new model and try out a few improvements suggested in section 6. The possible improvements could consist of nding the optimal smoothing parameters for each semantic class and using better non-linear classi ers like SVM. Another possible area of improvement is to consider more semantic classes such as dates, numbers, etc.</Paragraph>
    <Paragraph position="3"> We would also like to build systems for other tasks in TDT based on semantic language models and test their performance. We believe that semantic information is more critical in tasks such as New Event Detection which involves identifying the rst story that discusses a particular event. New events are typically characterized by mentions of new persons, locations or actions and our semantic models are capable of capturing exactly such information. null Additionally, it has been suggested that statistical models such as the aspect model (Hoffman, 1999) and the latent Dirichlet allocation (Blei et al., 2001) which generate words from a mixture of aspect-models can be exploited by modeling semantic classes as the aspects. We will be studying the applicability of these ideas to the current task as part of our future work.</Paragraph>
    <Paragraph position="4"> We believe the main contribution of our work lies in our attempt at incorporating semantic information in the language modeling framework and combining scores in a principled way. We believe we have only taken a rst step in this direction and much remains to be done as part of future work.</Paragraph>
  </Section>
class="xml-element"></Paper>