<?xml version="1.0" standalone="yes"?>
<Paper uid="P06-1129">
  <Title>Exploring Distributional Similarity Based Models for Query Spelling Correction</Title>
  <Section position="4" start_page="1025" end_page="1025" type="relat">
    <SectionTitle>
2 Related Work
</SectionTitle>
    <Paragraph position="0"> The method for web query spelling correction proposed by Cucerzan and Brill (2004) is essentially based on a source channel model, but it requires iterative running to derive suggestions for very-difficult-to-correct spelling errors. Word bigram model trained from search query logs is used as the source model, and the error model is approximated by inverse weighted edit distance of a correction candidate from its original term.</Paragraph>
    <Paragraph position="1"> The weights of edit operations are interactively optimized based on statistics from the query logs.</Paragraph>
    <Paragraph position="2"> They observed that an edit distance-based error model only has less impact on the overall accuracy than the source model. The paper reports that un-weighted edit distance will cause the overall accuracy of their speller's output to drop by around 2%. The work of Ahmad and Kondrak (2005) tried to employ an unsupervised approach to error model estimation. They designed an EM (Expectation Maximization) algorithm to optimize the probabilities of edit operations over a set of search queries from the query logs, by exploiting the fact that there are more than 10% misspelled queries scattered throughout the query logs. Their method is concerned with single character edit operations, and evaluation was performed on an isolated word spelling correction task.</Paragraph>
    <Paragraph position="3"> There are two lines of research in conventional spelling correction, which deal with non-word errors and real-word errors respectively. Non-word error spelling correction is concerned with the task of generating and ranking a list of possible spelling corrections for each query word not found in a lexicon. While traditionally candidate ranking is based on manually tuned scores such as assigning weights to different edit operations or leveraging candidate frequencies, some statistical models have been proposed for this ranking task in recent years. Brill and Moore (2000) presented an improved error model over the one proposed by Kernigham et al. (1990) by allowing generic string-to-string edit operations, which helps with modeling major cognitive errors such as the confusion between le and al. Toutanova and Moore (2002) further explored this via explicit modeling of phonetic information of English words. Both these two methods require misspelled/correct word pairs for training, and the latter also needs a pronunciation lexicon. Real-word spelling correction is also referred to as context sensitive spelling correction, which tries to detect incorrect usage of valid words in certain contexts (Golding and Roth, 1996; Mangu and Brill, 1997).</Paragraph>
    <Paragraph position="4"> Distributional similarity between words has been investigated and successfully applied in many natural language tasks such as automatic semantic knowledge acquisition (Dekang Lin, 1998) and language model smoothing (Essen and Steinbiss, 1992; Dagan et al., 1997). An investigation on distributional similarity functions can be found in (Lillian Lee, 1999).</Paragraph>
  </Section>
class="xml-element"></Paper>