<?xml version="1.0" standalone="yes"?>
<Paper uid="W06-3208">
  <Title>Morphology Induction from Limited Noisy Data Using Approximate String Matching</Title>
  <Section position="3" start_page="0" end_page="60" type="relat">
    <SectionTitle>
2 Related Work
</SectionTitle>
    <Paragraph position="0"> Much of the previous work on morphology learning has been reported on automatically acquiring affix lists. Inspired by works of Harris (1955), Dejean (1998) attempted to find a list of frequent affixes for several languages. He used successor and predecessor frequencies of letters in a given sequence of letters in identifying possible morpheme bound- null aries. The morpheme boundaries are where the predictability of the next letter in the letter sequence is the lowest.</Paragraph>
    <Paragraph position="1"> Several researchers (Brent, 1993; Brent et al., 1995; Goldsmith, 2001) used Minimum Description Length (MDL) for morphology learning. Snover and Brent (2001) proposed a generative probability model to identify stems and suffixes. Schone and Jurafsky (2001) used latent semantic analysis to find affixes. Baroni et al. (2002) produced a ranked list of morphologically related pairs from a corpus using orthographic or semantic similarity with minimum edit distance and mutual information metrics. Creutz and Lagus (2002) proposed two unsupervised methods for word segmentation, one based on maximum description length, and one based on maximum likelihood. In their model, words consisted of lengthy sequences of segments and there is no distinction between stems and affixes. The Whole Word Morphologizer (Neuvel and Fulop, 2002) uses a POS-tagged lexicon as input, induces morphological relationships without attempting to discover or identify morphemes. It is also capable of generating new words beyond the learning sample.</Paragraph>
    <Paragraph position="2"> Mystem (Segalovich, 2003) uses a dictionary for unknown word guessing in a morphological analysis algorithm for web search engines. Using a very simple idea of morphological similarity, unknown word morphology is taken from all the closest words in the dictionary, where the closeness is the number of letters on its end.</Paragraph>
    <Paragraph position="3"> The WordFrame model (Wicentowski, 2004) uses inflection-root pairs, where unseen inflections are transformed into their corresponding root forms.</Paragraph>
    <Paragraph position="4"> The model works with imperfect data, and can handle prefixes, suffixes, stem-internal vowel shifts, and point-of-affixation stem changes. The WordFrame model can be used for co-training with low-accuracy unsupervised algorithms.</Paragraph>
    <Paragraph position="5"> Monson (2004) concentrated on languages with limited resources. The proposed language-independent framework used a corpus of full word forms. Candidate suffixes are grouped into candidate inflection classes, which are then arranged in a lattice structure.</Paragraph>
    <Paragraph position="6"> A recent work (Goldsmith et al., 2005) proposed to use string edit distance algorithm as a bootstrapping heuristic to analyze languages with rich morphologies. String edit distance is used for ranking and quantifying the robustness of morphological generalizations in a set of clean data.</Paragraph>
    <Paragraph position="7"> All these methods require clean and most of the time large amounts of data, which may not exist for languages with limited electronic resources. For such languages, the morphology induction is still a problem. The work in this paper is applicable to noisy and limited data. String searching algorithms are used with information found in dictionaries to extract the affixes.</Paragraph>
  </Section>
class="xml-element"></Paper>