<?xml version="1.0" standalone="yes"?>
<Paper uid="W00-0605">
  <Title>A Question Answering System Developed as a Project in a Natural Language Processing Course*</Title>
  <Section position="6" start_page="34" end_page="34" type="concl">
    <SectionTitle>
4 Future Directions
</SectionTitle>
    <Paragraph position="0"> As a next step, we will try to tame our feature set.</Paragraph>
    <Paragraph position="1"> One possibility is to use a rule-based classifier that is less impacted by the serious imbalance between negative and positive instances than C5.0 in order to learn more effective feature sets for answer candidate discrimination corresponding to different question types. We could then use the classifier as a pre-processing filter to discard those less relevant comparison vector elements before inputting them into the classifiers, instead of inputting comparison results based on the complete feature sets. This should help to reduce noise generated by irrelevant features.</Paragraph>
    <Paragraph position="2"> Also, we will perform additional data analysis on the classification results to gain further insight into the noise sources.</Paragraph>
    <Paragraph position="3"> The classifiers we developed covered a wide range of approaches. To optimize the classification performance, we would like to implement a voting module to process the answer candidates from different classifiers. The confidence rankings of the classifiers would be determined from their corresponding answer selection accuracy in the training set, and will be used horizontally over the classifiers to provide a weighted confidence measure for each sentence, giving a final ordered list, where the head of the list is the proposed answer sentence. We propose to use a voting neural network to train the confidence weights on different classifiers based on different question types, since we also want to explore the relationship of classifier performance with question types. We believe this voting scheme will optimize the bagging of different classifiers and improve the hypothesis accuracy.</Paragraph>
  </Section>
class="xml-element"></Paper>