<?xml version="1.0" standalone="yes"?>
<Paper uid="W05-0805">
  <Title>Revealing Phonological Similarities between Related Languages from Automatically Generated Parallel Corpora</Title>
  <Section position="2" start_page="0" end_page="33" type="intro">
    <SectionTitle>
1 Introduction
</SectionTitle>
    <Paragraph position="0"> German and Dutch are languages that exhibit a wide range of similarities. Beside similar syntactic features like word order and verb subcategorization frames, the languages share phonological features which are due to historical sound changes. These similarities are one reason why it is easier to learn a closely historically related language than languages from other language families: the learner's native language provides a valuable resource which can be used in learning the new language. Although English also belongs to the West Germanic languages, German and Dutch share more lexical entries with a common root than German and English.</Paragraph>
    <Paragraph position="1"> The knowledge about language similarities on the lexical level is exploited in various fields. In machine translation, some approaches search for similar words (cognates) which are used to align parallel texts (e.g., Simard et al. (1992)). The word triple Text-tekst-text ([tEkst] in German, Dutch and English) can be easily recognized as a cognate; recognizing Pfeffer-peper-pepper ([pfE][f@r][pe:][p@r])-[pE][p@r*]), however, requires more knowledge about sound changes within the languages. The algorithms developed for machine translation search for similarities on the orthographic level, whereas some approaches to comparative and synchronic linguistics put their focus on similarities of phonological sequences.</Paragraph>
    <Paragraph position="2"> Covington (1996), for instance, suggests different algorithms to align the phonetic representation of words of historical languages. Kondrak (2000) presents an algorithm to align phonetic sequences by computing the similarities of these words.</Paragraph>
    <Paragraph position="3"> Nerbonne and Heeringa (1997) use phonetic transcriptions to measure the phonetic distance between different dialects. The above mentioned approaches presuppose either parallel texts of different languages for machine translation or manually compiled lists of transcribed cognates/words for analyzing synchronic or diachronic word pairs. Unfortunately, transcribed bilingual data are scarce and it  is labor-intensive to collect these kind of corpora.</Paragraph>
    <Paragraph position="4"> Thus, we aim at exploiting electronic pronunciation dictionaries to overcome the lack of data.</Paragraph>
    <Paragraph position="5"> In our approach, we automatically generate data as input to an unsupervised training regime and with the aim of automatically learning similar structures from these data using Expectation Maximization (EM) based clustering. Although the generation of our data introduces some noise, we expect that our method is able to automatically learn meaningful sound correspondences from a large amount of data.</Paragraph>
    <Paragraph position="6"> Our main assumption is that certain German/Dutch and German/English phoneme pairs from related stems occur more often and hence will appear in the same class with a higher probability than pairs not in related stems. We assume that the historical sound changes are hidden information in the classes.</Paragraph>
    <Paragraph position="7"> The paper is organized as follows: Section 2 presents related research. In Section 3, we describe the creation of our bilingual pronunciation dictionaries. The outcome is used as input to the algorithm for automatically deriving phonological classes described in Section 4. In Section 5, we apply our classes to a transcribed cognate list and measure the similarity between the two language pairs. A qualitative evaluation is presented in Section 6, where we interpret our best models. In Sections 7 and 8, we discuss our results and draw some final conclusions.</Paragraph>
  </Section>
class="xml-element"></Paper>