<?xml version="1.0" standalone="yes"?>
<Paper uid="P06-1056">
  <Title>Sydney, July 2006. c(c)2006 Association for Computational Linguistics Semi-Supervised Learning of Partial Cognates using Bilingual Bootstrapping</Title>
  <Section position="5" start_page="441" end_page="442" type="relat">
    <SectionTitle>
3 Related Work
</SectionTitle>
    <Paragraph position="0"> As far as we know there is no work done to disambiguate partial cognates between two languages. null Ide (2000) has shown on a small scale that cross-lingual lexicalization can be used to define and structure sense distinctions. Tufis et al. (2004) used cross-lingual lexicalization, wordnets alignment for several languages, and a clustering algorithm to perform WSD on a set of polysemous English words. They report an accuracy of 74%.</Paragraph>
    <Paragraph position="1"> One of the most active researchers in identifying cognates between pairs of languages is Kondrak (2001; 2004). His work is more related to the phonetic aspect of cognate identification. He used in his work algorithms that combine different orthographic and phonetic measures, recurrent sound correspondences, and some semantic similarity based on glosses overlap.</Paragraph>
    <Paragraph position="2"> Guy (1994) identified letter correspondence between words and estimates the likelihood of relatedness. No semantic component is present in the system, the words are assumed to be already matched by their meanings. Hewson (1993), Lowe and Mazadon (1994) used systematic sound correspondences to determine proto-projections for identifying cognate sets.</Paragraph>
    <Paragraph position="3"> WSD is a task that has attracted researchers since 1950 and it is still a topic of high interest. Determining the sense of an ambiguous word, using bootstrapping and texts from a different language was done by Yarowsky (1995), Hearst (1991), Diab (2002), and Li and Li (2004).</Paragraph>
    <Paragraph position="4"> Yarowsky (1995) has used a few seeds and untagged sentences in a bootstrapping algorithm based on decision lists. He added two constrains - words tend to have one sense per discourse and one sense per collocation. He reported high accuracy scores for a set of 10 words. The monolingual bootstrapping approach was also used by Hearst (1991), who used a small set of hand-labeled data to bootstrap from a larger corpus for training a noun disambiguation system for English. Unlike Yarowsky (1995), we use automatic collection of seeds. Besides our monolingual bootstrapping technique, we also use bilingual bootstrapping.</Paragraph>
    <Paragraph position="5"> Diab (2002) has shown that unsupervised WSD systems that use parallel corpora can achieve results that are close to the results of a supervised approach. She used parallel corpora in French, English, and Spanish, automatically-produced with MT tools to determine cross-language lexicalization sets of target words. The major goal of her work was to perform monolingual English WSD. Evaluation was performed on the nouns from the English all words data in Senseval2.</Paragraph>
    <Paragraph position="6"> Additional knowledge was added to the system  from WordNet in order to improve the results. In our experiments we use the parallel data in a different way: we use words from parallel sentences as features for Machine Learning (ML). Li and Li (2004) have shown that word translation and bilingual bootstrapping is a good combination for disambiguation. They were using a set of 7 pairs of Chinese and English words. The two senses of the words were highly distinctive: e.g. bass as fish or music; palm as tree or hand.</Paragraph>
    <Paragraph position="7"> Our work described in this paper shows that monolingual and bilingual bootstrapping can be successfully used to disambiguate partial cognates between two languages. Our approach differs from the ones we mentioned before not only from the point of human effort needed to annotate data - we require almost none, and from the way we use the parallel data to automatically collect training examples for machine learning, but also by the fact that we use only off-the-shelf tools and resources: free MT and ML tools, and parallel corpora. We show that a combination of these resources can be used with success in a task that would otherwise require a lot of time and human effort.</Paragraph>
  </Section>
class="xml-element"></Paper>