<?xml version="1.0" standalone="yes"?>
<Paper uid="W05-0106">
  <Title>Making Hidden Markov Models More Transparent</Title>
  <Section position="3" start_page="0" end_page="34" type="intro">
    <SectionTitle>
2 Displays
</SectionTitle>
    <Paragraph position="0"> Figure 1 shows a snapshot of our first display. It contains three kinds of information: most likely path for input, transition probabilities, and history of most likely prefixes for each observation index in the Viterbi lattice. The user can input text at the bottom of the display, e.g., Pelham pointed out that Georgia voters rejected the bill. The system then runs Viterbi and animates the search through all possible state sequences and displays the best state sequence prefix as it works its way through the observation  most likely path for &amp;quot;Pelman pointed out that Georgia voters ...&amp;quot;; Middle pane: a mouse-over-triggered bar graph of out transition probabilities for a state; Bottom pane: a history of most likely prefixes for each observation index in the Viterbi lattice. Below the panes is the input text field.  sequence from left to right (these are lines connecting the states in Figure 1). At any point, the student can mouse-over a state to see probabilities for transitions out of that state (this is the bar graph in Figure 1). Finally, the history of most likely prefixes is displayed (this history appears below the bar graph in Figure 1). We mentioned that students often falsely believe that the most likely prefix is extended monotonically. By seeing the path through the states reconfigure itself in the middle of the observation sequence and by looking at the prefix history, a student has a good chance of dispelling the false belief of monotonicity.</Paragraph>
    <Paragraph position="1"> The second display allows the user to contrast two state sequences for the same observation sequence.</Paragraph>
    <Paragraph position="2"> See Figure 2. For each contrasting state pairs, it shows the ratio of the corresponding transition to each state and it shows the ratio of the generation of the observation conditioned on each state. For example, in Figure 2 the transition DT-JJ is less likely than DT-NNP. The real culprit is generation probability P(Equal|JJ) which is almost 7 times larger than P(Equal|NNP). Later in the sequence we see a similar problem with generating opportunity from a NNP state. These generation probabilities seem to drown out any gains made by the likelihood of NNP runs.</Paragraph>
    <Paragraph position="3"> To use this display, the user types in a sentence in the box above the graph and presses enter. The HMM is used to tag the input. The user then modifies (e.g., corrects) the tag sequence and presses enter and the ratio bars then appear.</Paragraph>
    <Paragraph position="4"> Let us consider another example: in Figure 2, the mis-tagging of raises as a verb instead of a noun at the end of the sentence. The display shows us that although NN-NNS is more likely than NN-VBZ, the generation probability for raises as a verb is over twice as high as a noun. (If this pattern of mis-taggings caused by high generation probability ratios was found repeatedly, we might consider smoothing these distributions more aggressively.)</Paragraph>
  </Section>
class="xml-element"></Paper>