<?xml version="1.0" standalone="yes"?>
<Paper uid="P02-1019">
  <Title>Pronunciation Modeling for Improved Spelling Correction</Title>
  <Section position="9" start_page="0" end_page="0" type="evalu">
    <SectionTitle>
4.2 Results
</SectionTitle>
    <Paragraph position="0"> We tested our system and compared it to the Brill and Moore model on a dataset of around BDBCBNBCBCBC pairs of misspellings and corresponding correct words, split into training and test sets. The exact data sizes are BJBNBFBKBH word pairs in the training set and BDBNBKBDBE word pairs in the test set. This set is slightly different from the dataset used in Brill and Moore's experiments because we removed from the original dataset the pairs for which we did not have the correct word in the pronunciation dictionary. Both models LTR and PH were trained on the same training set. The interpolation weight that the combined model CMB uses is also set on the training set to maximize the classification accuracy.</Paragraph>
    <Paragraph position="1"> At test time we do not search through all possible words D6 in the dictionary to find the one maximizing</Paragraph>
    <Paragraph position="3"> B4DBCYD6B5. Rather, we compute the combination score only for candidate words D6 that are in the top C6 according to the C8  of the pronunciations of D6 from the dictionary and any of the pronunciations for DB that were proposed by the letter-to-phone model. The letter-to-phone  model returned for each DB the BF most probable pronunciations only. Our performance was better when we considered the top BF pronunciations of DB rather than a single most likely hypothesis. That is probably due to the fact that the BF-best accuracy of the letter-to-phone model is significantly higher than its BD-best accuracy.</Paragraph>
    <Paragraph position="4"> Table 3 shows the spelling correction accuracy when using the model LTR, PH, or both in combination. The table shows C6-best accuracy results. The C6-best accuracy figures represent the percent test cases for which the correct word was in the top C6 words proposed by the model. We chose the context size of BF for the LTR model as this context size maximized test set accuracy. Larger context sizes neither helped nor hurt accuracy.</Paragraph>
    <Paragraph position="5"> As we can see from the table, the phone-based model alone produces respectable accuracy results considering that it is only dealing with word pronunciations. The error reduction of the combined model compared to the letters-only model is substantial: for 1-Best, the error reduction is over BEBFB1; for 2-Best, 3-Best, and 4-Best it is even higher, reaching over BGBIB1 for 4-Best.</Paragraph>
    <Paragraph position="6"> As an example of the influence of pronunciation modeling, in Table 4 we list some misspellingcorrect word pairs where the LTR model made an incorrect guess and the combined model CMB guessed accurately.</Paragraph>
  </Section>
class="xml-element"></Paper>