<?xml version="1.0" standalone="yes"?>
<Paper uid="W06-1648">
  <Title>Language Modeling, and Shallow Morphology</Title>
  <Section position="6" start_page="411" end_page="412" type="evalu">
    <SectionTitle>
4 Experimental Results
</SectionTitle>
    <Paragraph position="0"> Table 1 reports on the percentage of words for which a proper correction was found in the top n generated corrections using different models.</Paragraph>
    <Paragraph position="1"> The percentage of words for which a proper correction exists in the top 10 proposed correction is the upper limit accuracy we can achieve given than we can rerank the correction using language modeling. Table 2 reports the word error rate for the 1:1 and m:n models with and without CP, UP, LM, and SM. Further, the before and after stemming error rates are reported for setups that use language modeling. Table 3 reports on the stem error rate when using the stem trigram language model.</Paragraph>
    <Paragraph position="2"> The best model was able to find the proper correction within the top 10 proposed correction for 90% of the words. The failure to find a proper correction within the proposed corrections was generally due to grossly misrecognized words and was rarely due to words that do not exist in web-mined collection. Perhaps, more training examples for the character based models would improve correction.</Paragraph>
    <Paragraph position="3">  trigram language model The results indicate that the m:n character model is better than the 1:1 model in two ways. The first is that the m:n model yielded a greater percentage of proper corrections in the top 10 generated corrections, and the second is that the scores of the top 10 corrections were better which led to better results compared to the 1:1 model when used in combination with language modeling. For the m:n model with language modeling, the language model properly picked the proper correction from the proposed correction 98% of the time (for the cases where a proper correction was within the proposed corrections). null Also the use of smoothing, UP, produced better corrections, while accounting for character positions had an adverse effect on correction. This might be an indication that the character segment correction training data was sparse. Using the 6-gram language model on the segmented words had a severely negative impact on correction accuracy. Perhaps is due to insufficient training data for the model. This situation lends itself to using a factored language model using the surface form of words as well as other linguistic features of the word such as part of speech tags, prefixes, and suffixes.</Paragraph>
    <Paragraph position="4"> As for training a language model on words versus stems, the results suggest that word based correction is slightly better than stem based correction. The authors' intuition is that this resulted from having a sufficiently large corpus to train the language model and that this might have been reversed if the training corpus for the language model was smaller. Perhaps further investigation would prove or disprove the authors' intuition.</Paragraph>
  </Section>
class="xml-element"></Paper>