<?xml version="1.0" standalone="yes"?>
<Paper uid="W05-0606">
  <Title>Computing Word Similarity and Identifying Cognates with Pair Hidden Markov Models</Title>
  <Section position="3" start_page="0" end_page="40" type="intro">
    <SectionTitle>
1 Introduction
</SectionTitle>
    <Paragraph position="0"> The computation of surface similarity between pairs of words is an important task in many areas of natural language processing. In historical linguistics phonetic similarity is one of the clues for identifying cognates, that is, words that share a common origin (Oakes, 2000). In statistical machine translation, cognates are helpful in inducing translation lexicons (Koehn and Knight, 2001; Mann and Yarowsky, 2001), sentence alignment (Melamed, 1999), and word alignment (Tiedemann, 2003). In dialectology, similarity is used for estimating distance between dialects (Nerbonne, 2003). Other applications include cross-lingual information retrieval (Pirkola et al., 2003), detection of confusable drug names (Kondrak and Dorr, 2004), and lexicography (Brew and McKelvie, 1996).</Paragraph>
    <Paragraph position="1"> Depending on the context, strong word similarity may indicate either that words share a common origin (cognates), a common meaning (synonyms), or are related in some way (e.g. spelling variants). In this paper, we focus on cognates. Genetic cognates are well-suited for testing measures of word similarity because they arise by evolving from a single word in a proto-language. Unlike rather indefinite concepts like synonymy or confusability, cognation is a binary notion, which in most cases can be reliably determined.</Paragraph>
    <Paragraph position="2"> Methods that are normally used for computing word similarity can be divided into orthographic and phonetic. The former includes string edit distance (Wagner and Fischer, 1974), longest common subsequence ratio (Melamed, 1999), and measures based on shared character n-grams (Brew and Mc-Kelvie, 1996). These usually employ a binary identity function on the level of character comparison. The phonetic approaches, such as Soundex (Hall and Dowling, 1980) and Editex (Zobel and Dart, 1996), attempt to take advantage of the phonetic characteristics of individual characters in order to estimate their similarity. All of the above methods are static, in the sense of having a fixed definition that leaves little room for adaptation to a specific context. In contrast, the methods proposed by Tiedemann (1999) automatically construct weighted string similarity measures on the basis of string segmentation and bitext co-occurrence statistics.</Paragraph>
    <Paragraph position="3"> We have created a system for determining word similarity based on a Pair Hidden Markov Model.</Paragraph>
    <Paragraph position="4"> The parameters of the model are automatically learned from training data that consists of word  pairs that are known to be similar. The model is trained using the Baum-Welch algorithm (Baum et al., 1970). We examine several variants of the model, which are characterized by different training techniques, number of parameters, and word length correction method. The models are tested on a cognate recognition task across word lists representing several Indo-European languages. The experiments indicate that our system substantially outperforms the most commonly used approaches.</Paragraph>
    <Paragraph position="5"> The paper is organized as follows. Section 2 gives a more detailed description of the problem of word similarity. Section 3 contains an introduction to Pair Hidden Markov Models, while section 4 describes their adaptation to our domain. Sections 5 and 6 report experimental set-up and results.</Paragraph>
  </Section>
class="xml-element"></Paper>