<?xml version="1.0"?><!DOCTYPE article SYSTEM "/project/take/software/searchbench_offline_processing/paperxml_generator/aclextractor/src/python/../resource/dtd/paperxml.dtd"><article><header><firstpageheader><page local="1" global="9"/><title>Automatic Extraction of English-Chinese Transliteration Pairs using Dynamic Window and Tokenizer</title><pubinfo>Sixth SIGHAN Workshop on Chinese Language Processing</pubinfo><author surname="Jin" givenname="Chengguo"><org  name="Language Engineering Institute" country="China"/></author><author surname="Kim" givenname="Dong-Il"><org  name="Language Engineering Institute" country="China"/></author><author surname="Na" givenname="Seung-Hoon"><org  name="Language Engineering Institute" country="China"/></author><author surname="Lee" givenname="Jong-Hyeok"><org  name="Language Engineering Institute" country="China"/></author></firstpageheader><frontmatter><p><b>Automatic Extraction of English-Chinese Transliteration Pairs using Dynamic Window and Tokenizer</b></p><p><b>Chengguo Jin</b></p><p>Dept. of Graduate School for Information</p><p>Technology, POSTECH, Korea</p><p>chengguo@postech.ac.kr</p><p><b>Dong-Il Kim</b></p><p>Language Engineering Institute, YUST, China</p><p>dongil@ybust.edu.cn</p><p><b>Seung-Hoon Na</b></p><p>Dept. of Computer Science &amp; Engineering POSTECH, Korea nsh197 9@postech.ac.kr <b>Jong-Hyeok Lee </b>Dept. of Computer Science &amp; Engineering</p><p>POSTECH, Korea</p><p>j hlee@postech.ac.kr</p></frontmatter><abstract>Recently, many studies have been focused on extracting transliteration pairs from bi­lingual texts. Most of these studies are based on the statistical transliteration mod­el. The paper discusses the limitations of previous approaches and proposes novel approaches called dynamic window and to­kenizer to overcome these limitations. Ex­perimental results show that the average rates of word and character precision are 99.0% and 99.78%, respectively. </abstract></header><body><section number="1" title="Introduction"><p>Machine transliteration is a type of translation based on phonetic similarity between two lan­guages. Chinese Named entities including foreign person names, location names and company names, etc are usually transliterated from foreign words. The main problem of transliteration resulted from complex relations between Chinese phonetic sym­bols and characters. Usually, a foreign word can be transliterated into various Chinese words, and sometimes this will lead to transliteration complex­ity. In addition, dozens of Chinese characters cor­respond to each pinyin which uses the Latin alphabet to represent sounds in Standard Mandarin. In order to solve these problems, Chinese government published the "Names of the world's peoples"[12] containing 630,000 entries in 1993, which took about 40 years. However, some new foreign names still cannot be found in the diction­ary. Constructing an unknown word dictionary is a difficult and time consuming job, so in this paper we propose a novel approach to automatically con­struct the resource by efficiently extracting trans­literation pairs from bilingual texts.</p><p>Recently, much research has been conducted on machine transliteration. Machine transliteration is classified into two types. One is automatic genera­tion of transliterated word from the source lan­guage [6]; the other one is extracting transliteration pairs from bilingual texts [2]. Generally, the gen­eration process performs worse than the extraction process. Especially in Chinese, people do not al­ways transliterate foreign words only by sound but also consider the meanings. For example, the word 'blog' is not transliterated into ' <b>^ 5</b><b>j </b><b>M</b>' (Bu-LaoGe) which is phonetically equivalent to the source word, but transliterated into <b>'ff</b>^'(BoKe) which means 'a lot of guests'. In this case, it is too difficult to automatically generate correct translit­eration words. Therefore, our approach is based on the method of extracting transliteration pairs from bilingual texts.</p><p>The type of extraction of transliteration pairs can also be further divided into two types. One is ex­tracting transliteration candidates from each lan­guage respectively, and then comparing the pho­netic similarities between those candidates of two languages [2, 8]. The other one is only extracting transliteration candidates from the source language, and using the candidates to extract corresponding transliteration words from the target language [1]. In Chinese, there is no space between two words and no special character set to represent foreign words such as Japanese; hence the candidate ex­traction is difficult and usually results in a low pre­cision. Therefore, the method presented in [2] which extracted transliteration candidates from both English and Chinese result in a poor perform­ance.<page local="2" global="10"/> Compared to other works, Lee[1] only ex­tracts transliteration candidates from English, and finds equivalent Chinese transliteration words without extracting candidates from Chinese texts. The method works well, but the performance is required to be improved. In this paper we present a novel approaches to obtain a remarkable result in extracting transliteration word pairs from parallel texts.</p><p>The remainder of the paper is organized as fol­lows: Section 2 gives an overview of statistical machine transliteration and describes proposed approaches. Section 3 describes the experimental setup and a quantitative assessment of performance of our approaches. Conclusions and future work are presented in Section 4.</p></section><section number="2" title="Extraction of English-Chinese translit­eration pairs"><p>In this paper, we first extract English named en­tities from English-Chinese parallel texts, and se­lect only those which are to be transliterated into Chinese. Next we extract Chinese transliteration words from corresponding Chinese texts. [Fig. 1] shows the entire process of extracting translitera­tion word pairs from English-Chinese parallel texts.</p><p><b>English-Chinese Parallel Corpus</b> <b>English named entity recognition model</b> <b>Chinese sentence</b> <b>Transliteration pair extraction model</b></p><p><b>English-Chinese transliteration pair</b> [Fig 1]. The process of extracting transliteration pairs from English-Chinese parallel corpus</p><doubt alpha="0.0" length="3" tooSmall="False" monospace="0.0">2.1</doubt><p><b>Statistical machine transliteration model</b></p><p>Generally, the Chinese Romanization system pin­yin which is used to represent the pronunciation of each Chinese character is adopted in Chinese transliteration related studies. For example, the Chinese word <b>'jj</b><b>L</b><b>#TT' </b>is first transformed to piny in 'Ke Lin Dun', and we compare the phonetic similarities between 'Clinton' and 'KeLinDun'. In this paper, we assume that E is written in English, while C is written in Chinese, and TU represents translitera­tion units. So P(C|E), P(^# <b>W </b>|Clinton) can be transformed to P(KeLinDun|Clinton). In this paper we define English TU as unigram, bigram, and tri-gram; Chinese TU is pinyin initial, pinyin final and the entire pinyin. With these definitions we can further write the probability, P(j<b>^#W</b>|Clinton), as follows:</p><doubt alpha="55.6" length="90" tooSmall="False" monospace="0.0">P(jKttTT!|Clinton)= P(kelindunl Clinton) = P(ke| C)P(l| /)P(in | in)P(t | d)P(un | on) (1)</doubt><p>[Fig 2]. TU alignment between English and Chinese pinyin</p><p>[Fig 2] shows the possible alignment between Eng­lish word 'Clinton' and Chinese word <b>'jj</b><b>L</b><b>#TTT</b>'s pinyin 'KeLinDun'.</p><p>In [1], the authors add the match type informa­tion in Eq. (1). The match type is defined with the lengths of TUs of two languages. For example, in the case of <b>p </b>(<b>ke </b>| <b>C</b>) the match type is 2-1, be­cause the size of Chinese TU ke is 2 and the size of English TU C is 1. Match type is useful when estimating transliteration model's parameters with­out a pronunciation dictionary. In this paper, we use the EM algorithm to estimate transliteration model's parameters without a pronunciation dic­tionary, so we applied match type to our model. Add Match type(M) to Eq.(1) to formulate as fol­lows:</p><doubt alpha="61.5" length="26" tooSmall="False" monospace="0.0">P(CIE) « maxP(CIM,E)P(MIE)</doubt><doubt alpha="100.0" length="1" tooSmall="False" monospace="0.0">M</doubt><doubt alpha="47.6" length="21" tooSmall="False" monospace="0.0">« maxP (CIM,E)P(M)(2)</doubt><doubt alpha="60.0" length="40" tooSmall="False" monospace="0.0">logP(CIE)« maxY((logP(ViIu,)+logP(m))(3)</doubt><doubt alpha="100.0" length="2" tooSmall="False" monospace="0.0">io</doubt><table class="main" frame="box" rules="all" border="1" regular="False"><tr class="row"><td class="cell"></td><td class="cell"></td><td class="cell"></td><td class="cell"></td><td class="cell"></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p>c</p></td><td class="cell"><p>1 in</p></td><td class="cell"><p>t</p></td><td class="cell"><p>on</p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p>ke</p></td><td class="cell"><p>1 in</p></td><td class="cell"><p>d</p></td><td class="cell"><p>un</p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p><i>%</i></p></td><td class="cell"><p></p></td><td class="cell"><p><i>\</i></p></td><td class="cell"><p>■M</p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"></td><td class="cell"></td><td class="cell"></td><td class="cell"></td><td class="cell"></td></tr></table><table class="main" frame="box" rules="all" border="1" regular="False"><tr class="row"><td class="cell"></td><td class="cell"></td><td class="cell"></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p><b>English</b></p></td><td class="cell"><p></p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p><b>word</b></p></td><td class="cell"><p></p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p><b>t</b></p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"></td><td class="cell"></td><td class="cell"></td></tr></table><page local="3" global="11"/><p>where u, v are English TU and Chinese TU, re[Fig 3] shows how to extract the correct Chinese transliteration <b>"]S</b><b> </b><b>W</b>^"(KeLinDun) with the given English word "Clinton" from a Chinese sentence.</p><subsection number="2.2" title="Proposed methods"><p>When the statistical machine transliteration is used to extract transliteration pairs from a parallel text, the problems arise when there is more than one Chinese character sequence that is phonetically similar to the English word. In this paper we pro­pose novel approaches called dynamic window and tokenizer to solve the problems effectively.</p><subsubsection number="2.2.1" title="Dynamic window method"><p>The dynamic window approach does not find the transliteration at once, but first sets the window size range according to the English word candi­dates, and slides each window within the range to find the correct transliterations.</p><p>[Fig 4]. Alignment result between English word "Clinton" and correct Chinese transliteration, add a character into correct Chinese transliteration, and eliminate a character from correct Chinese transliteration.</p><p>If we know the exact Chinese transliteration's size, then we can efficiently extract Chinese trans­literations by setting the window with the length of the actual Chinese transliteration word. For exam­ple, in [Fig 4] we do alignment between the Eng­lish word "Clinton" and correct Chinese translit­eration "j^i#it!j5l"(KeLinDun), add a character into correct Chinese transliteration " J^L # <i>M </i>^"(KeLinYiDun), and eliminate a character from correct Chinese transliteration " <b>W </b><i>W, </i>"(LinDun) respectively. The result shows that the highest score is the alignment with correct Chinese trans­literation. This is because the alignment between the English word and the correct Chinese translit­eration will lead to more alignments between Eng­lish TUs and Chinese TUs, which will result in highest scores among alignment with other Chi­nese sequences. This characteristic does not only exist between English and Chinese, but also exists between other language pairs.</p><p>However, in most circumstances, we can hardly determine the correct Chinese transliteration's length. Therefore, we analyze the distribution be­tween English words and Chinese transliterations to predict the possible range of Chinese translitera­tion's length according to the English word. We spec<u>tively and m is the match type of u and v.</u><u></u></p><p>[Fig 3]. The alignment of the English word and the ( sentence containing corresponding transliteration<page local="4" global="12"/></p><table class="main" frame="box" rules="all" border="0" regular="False"><tr class="row"><td class="cell"></td><td class="cell"></td><td class="cell"></td><td class="cell"></td><td class="cell"></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p><b><i>English</i></b></p></td><td class="cell"><p><b><i>Chinese</i></b></p></td><td class="cell"><p><b><i>Match</i></b></p></td><td class="cell"><p><b><i>Chinese</i></b></p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p><b><i>TU</i></b></p></td><td class="cell"><p><b><i>TU</i></b></p></td><td class="cell"><p><b><i>type</i></b></p></td><td class="cell"><p><b><i>characters</i></b></p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p><b><i>C</i></b></p></td><td class="cell"><p>ke</p></td><td class="cell"><p>[1,2]</p></td><td class="cell"><p></p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p>li</p></td><td class="cell"><p>li</p></td><td class="cell"><p>[2.2]</p></td><td class="cell"><p>**</p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p>n</p></td><td class="cell"><p>n</p></td><td class="cell"><p>[1,1]</p></td><td class="cell"><p></p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p>t</p></td><td class="cell"><p>d</p></td><td class="cell"><p>[1,1]</p></td><td class="cell"><p></p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p>on</p></td><td class="cell"><p>un</p></td><td class="cell"><p>[2.2]</p></td><td class="cell"><p></p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p></p></td><td class="cell"><p>Score:</p></td><td class="cell"><p>-4.29</p></td><td class="cell"><p></p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p>C</p></td><td class="cell"><p>ke</p></td><td class="cell"><p>[1,2]</p></td><td class="cell"><p></p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p>li</p></td><td class="cell"><p>li</p></td><td class="cell"><p>[2,2]</p></td><td class="cell"><p>**</p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p>n</p></td><td class="cell"><p>n</p></td><td class="cell"><p>[1,1]</p></td><td class="cell"><p></p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p></p></td><td class="cell"><p>yi</p></td><td class="cell"><p>[0,2]</p></td><td class="cell"><p></p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p>t</p></td><td class="cell"><p>d</p></td><td class="cell"><p>[1,1]</p></td><td class="cell"><p>li!</p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p>on</p></td><td class="cell"><p>un</p></td><td class="cell"><p>[2,2]</p></td><td class="cell"><p></p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p></p></td><td class="cell"><p>Score:</p></td><td class="cell"><p>-6.94</p></td><td class="cell"><p></p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p>C</p></td><td class="cell"><p></p></td><td class="cell"><p>[1,0]</p></td><td class="cell"><p></p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p>li</p></td><td class="cell"><p>li</p></td><td class="cell"><p>[2,2]</p></td><td class="cell"><p><i>#</i></p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p>n</p></td><td class="cell"><p>n</p></td><td class="cell"><p>[1,1]</p></td><td class="cell"><p></p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p>t</p></td><td class="cell"><p>d</p></td><td class="cell"><p>[1,1]</p></td><td class="cell"><p></p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p>on</p></td><td class="cell"><p>un</p></td><td class="cell"><p>[2,2]</p></td><td class="cell"><p></p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p></p></td><td class="cell"><p>Score:</p></td><td class="cell"><p>-7.28</p></td><td class="cell"><p></p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"></td><td class="cell"></td><td class="cell"></td><td class="cell"></td><td class="cell"></td></tr></table><table class="main" frame="box" rules="all" border="0" regular="False"><tr class="row"><td class="cell"></td><td class="cell"></td><td class="cell"></td><td class="cell"></td><td class="cell"></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p>English Word:</p></td><td class="cell"><p>Clinton</p></td><td class="cell"><p></p></td><td class="cell"><p></p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p>Chinese Transliteration word: ;£}#!)! (KeLinDun)</p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p>Chinese sentence:</p></td><td class="cell"><p></p></td><td class="cell"><p></p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p></p></td><td class="cell"><p></p></td><td class="cell"><p></p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p>(Clinton will</p></td><td class="cell"><p>visit Kore</p></td><td class="cell"><p>i tomorrow. )</p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p><b><i>English</i></b></p></td><td class="cell"><p><b><i>Chinese</i></b></p></td><td class="cell"><p><b><i>Match</i></b></p></td><td class="cell"><p><b><i>Chinese</i></b></p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p><b><i>TU</i></b></p></td><td class="cell"><p><b><i>TU</i></b></p></td><td class="cell"><p><b><i>type</i></b></p></td><td class="cell"><p><b><i>characters</i></b></p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p></p></td><td class="cell"><p>mi</p></td><td class="cell"><p>[0,2]</p></td><td class="cell"><p><i>m</i></p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p></p></td><td class="cell"><p>n</p></td><td class="cell"><p>[0,1]</p></td><td class="cell"><p></p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p></p></td><td class="cell"><p>g</p></td><td class="cell"><p>[0,1]</p></td><td class="cell"><p></p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p></p></td><td class="cell"><p>t</p></td><td class="cell"><p>[0,1]</p></td><td class="cell"><p></p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p></p></td><td class="cell"><p>i</p></td><td class="cell"><p>[0,1]</p></td><td class="cell"><p></p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p></p></td><td class="cell"><p>an</p></td><td class="cell"><p>[0,2]</p></td><td class="cell"><p></p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p>C</p></td><td class="cell"><p>ke</p></td><td class="cell"><p>[1,2]</p></td><td class="cell"><p>-£</p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p>li</p></td><td class="cell"><p>li</p></td><td class="cell"><p>[2,2]</p></td><td class="cell"><p></p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p>n</p></td><td class="cell"><p>n</p></td><td class="cell"><p>[1,1]</p></td><td class="cell"><p></p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p>t</p></td><td class="cell"><p>d</p></td><td class="cell"><p>[1,1]</p></td><td class="cell"><p><i>m</i></p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p>on</p></td><td class="cell"><p>un</p></td><td class="cell"><p>[2,2]</p></td><td class="cell"><p></p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p></p></td><td class="cell"><p>ji</p></td><td class="cell"><p>[0,2]</p></td><td class="cell"><p><i>m</i></p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p></p></td><td class="cell"><p>an</p></td><td class="cell"><p>[0,2]</p></td><td class="cell"><p></p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p></p></td><td class="cell"><p>g</p></td><td class="cell"><p>[0,1]</p></td><td class="cell"><p></p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p></p></td><td class="cell"><p>f</p></td><td class="cell"><p>[0,1]</p></td><td class="cell"><p><b><i>ill</i></b></p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p></p></td><td class="cell"><p>an</p></td><td class="cell"><p>[0,2]</p></td><td class="cell"><p></p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p></p></td><td class="cell"><p>g</p></td><td class="cell"><p>[0,1]</p></td><td class="cell"><p></p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p></p></td><td class="cell"><p>we</p></td><td class="cell"><p>[0,2]</p></td><td class="cell"><p><i>u</i></p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p></p></td><td class="cell"><p>n</p></td><td class="cell"><p>[0,1]</p></td><td class="cell"><p></p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p></p></td><td class="cell"><p>h</p></td><td class="cell"><p>[0,1]</p></td><td class="cell"><p><i>m</i></p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p></p></td><td class="cell"><p>an</p></td><td class="cell"><p>[0,2]</p></td><td class="cell"><p></p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p></p></td><td class="cell"><p>g</p></td><td class="cell"><p>[0,1]</p></td><td class="cell"><p><i>m</i></p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p></p></td><td class="cell"><p>uo</p></td><td class="cell"><p>[0,2]</p></td><td class="cell"><p></p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"></td><td class="cell"></td><td class="cell"></td><td class="cell"></td><td class="cell"></td></tr></table><p>present the algorithm for the dynamic window ap­proach as follows:</p><p>Step 1: Set the range of Chinese transliteration's length according to the extracted English word candidate.</p><p>Step 2: Slide each window within the range to calculate the probability between an English word and a Chinese character sequence contained in the current window using Eq 3.</p><p>Step 3: Select the Chinese character sequence with highest score and back-track the alignment result to extract the correct transliteration word. [Fig 5] shows the entire process of using the dy­namic window approach to extract the correct transliteration word.</p><p>[Fig 5]. Extract the correct transliteration using the dynamic window method</p><p>The dynamic window approach can effectively solve the problem shown in [Fig 5] which is the most common problem that arises from using sta­tistical machine transliteration model to extract a transliteration from a Chinese sentence. However, it can not handle the case that a correct translitera­tion with correct window size can not be extracted. Moreover, when the dynamic window approach is used, the processing time will increase severely. Hence, the following approach is presented to deal with the problem as well as to improve the per­formance.</p></subsubsection><subsubsection number="2.2.2" title="Tokenizer method"><p>The tokenizer method is to divide a sentence with characters which have never been used in Chinese transliterations and applies the statistical transliteration model to each part to extract a cor­rect transliteration.</p><p>There are certain characters that are frequently used for transliterating foreign words, such as"Jfe (shi), #-(de), $fi(le), f#(he) On the other hand, there are other characters, such as "^(shi), fft(de), T(le), ^P(he),...", that have never been used for Chinese transliteration, while they are phonetically equivalent with the above characters. These characters are mainly particles, copulas and non-Chinese characters etc., and always come with named entities and sometimes also cause some problems. For example, when the English word "David" is transliterated into Chinese, the last pho­neme is omitted and transliterated into " <i>~k.</i></p><p><b>ni</b>"(DaWei). In this case of a Chinese character</p><p>such as " (ft "(De) which is phonetically similar with the omitted syllable 'd', the statistical translit­eration model will incorrectly extract " <i>~k. </i>HZ fft"(DaWeiDe) as transliteration of "David". In [1], the authors deal with the problem through a post-process using some linguistic rules. Lee and Chang [1] merely eliminate the characters which have never been used in Chinese transliteration such as "(ft"(De) from the results. Nevertheless, the ap­proach cannot solve the problem shows in [Fig 6], because the copula "^"(Shi) combines with the other character "^f "(zhe) to form the character sequence " <i>^ </i><b>te </b>" (ZheShi) which is phonetically similar with the English word "Jacey", and is in­correctly recognized as a transliteration of "Jacey". Thus, in this case, although the copula "^"(Shi) is eliminated from the result through the post-process method presented in [1], the remaining part is not the correct transliteration.<page local="5" global="13"/> Compared with the method in [1], our tokenizer approach eliminates copula "^"(Shi) at pre-processing time and then the phonetic similarity between "Jacey" and the remaining part "^t"(Zhe) becomes very low; hence our approach overcomes the problem prior to the entire process. In addition, the tokenizer approach also reduces the processing time dramatically due to separating a sentence into several parts. [Fig 6] shows the process of extracting a correct translit­eration using the tokenizer method.</p><table class="main" frame="box" rules="all" border="1" regular="False"><tr class="row"><td class="cell"></td><td class="cell"></td><td class="cell"></td><td class="cell"></td><td class="cell"></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p>English Word</p></td><td class="cell"><p>Ziegler</p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p>Chinese Sentence</p></td><td class="cell"><p></p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p><b><i>mi </i></b>1963 <b>%i£jn</b>^b<b>^3t,</b></p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p>English Sentence</p></td><td class="cell"><p>Ziegler and Italian Chemist Julio re­ceived the Nobel prize of 1963 together.</p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p>Extracted translit­eration without using dynamic window</p></td><td class="cell"><p>tg^I (JiaJuLiAo)</p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p>Correct translitera­tion</p></td><td class="cell"><p><i>Jx</i><b><i>^m </i></b>(QiGeLe)</p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p>Steps</p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p>1. Set Chinese transliteration's range according to English word "Ziegler" to [2, 7] (After analyzing the distribution be­tween an English word and a Chinese transliteration word, we found that if the English word length is L , then the Chinese transliteration word is between L <i>13 </i>and L .)</p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p>2. Slide each window to find sequence with highest score.</p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p>3 Select the Chinese character sequence with highest score and back-track the alignment result to extract a correct translitera­tion word.</p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p>Win­dow size</p></td><td class="cell"><p>Chinese character sequence with high­est score of each window (underline the back-tracking result)</p></td><td class="cell"><p>Score (normal­ize with window</p><p>size)</p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p>2</p></td><td class="cell"><p><b>JHI</b><b> </b>(QiGe)</p></td><td class="cell"><p>-9.327</p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p>3</p></td><td class="cell"><p><b><i>ftmh</i></b><b><i> </i></b>(OiGeLe)</p></td><td class="cell"><p>-6.290</p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p>4</p></td><td class="cell"><p>5R£fJ^(QiGeLeYu)</p></td><td class="cell"><p>-8.433</p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p>5</p></td><td class="cell"><p>^FteU^jS (QiGeLeYuYi)</p></td><td class="cell"><p>-9.719</p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p>6</p></td><td class="cell"><p>^cg^HSIl (JiaJuLiAoGongTong)</p></td><td class="cell"><p>-10.458</p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p>7</p></td><td class="cell"><p>(QiGeLeYuYiDaLi)</p></td><td class="cell"><p>-10.721</p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"></td><td class="cell"></td><td class="cell"></td><td class="cell"></td><td class="cell"></td></tr></table><p>[Fig 6]. Extracting the correct transliteration using the to­kenizer method.</p><p>In conclusion, the two approaches complement each other; hence using them together will lead to a better performance.</p></subsubsection></subsection></section><section number="3" title="Experiments"><p>In this section, we focus on the setup for the ex­periments and a performance evaluation of the proposed approaches to extract transliteration word pairs from parallel corpora.</p><subsection number="3.1" title="Experimental setup"><p>We use 300 parallel English-Chinese sentences containing various person names, location names, company names etc. The corpus for training con­sists of 860 pairs of English names and their Chi­nese transliterations. The performance of translit­eration pair extraction was evaluated based on pre­cision and recall rates at the word and character levels. Since we consider exactly one proper name in the source language and one transliteration in the target language at a time, the word recall rates are the same as the word precision rates. In order to demonstrate the effectiveness of our approaches, we perform the following experiments: firstly, only use STM(Statistical transliteration model) which is the baseline of our experiment; secondly, we apply the dynamic window and tokenizer method with STM respectively; thirdly, we apply these two methods together; at last, we perform experiment presented in [1] to compare with our methods.</p></subsection><subsection number="3.2" title="Evaluation of dynamic window and to­kenizer methods"><p><b>[table 1 </b><b>].</b><b> The experimental results of extracting transliteration pairs using proposed methods</b></p><p>As shown in table 1, the baseline STM achieves a word precision rate of 75%. The STM works relatively well with short sentences, but as the length of sentences increases the performance sig­nificantly decreases. The dynamic window ap­proach overcomes the problem effectively. If the dynamic window method is applied with STM, the model will be tolerant with the length of sentences. The dynamic window approach improves the per­formance of STM around 21%, and reaches the average word precision rate of 96% (STM+DW). In order to estimate the highest performance that the dynamic window approach can achieve, we apply the correct window size which can be ob­tained from the evaluation data set with STM. The result (STM+CW) shows around 98% word precision rate and about 23% improvement over the baseline. Therefore, dynamic window approach is remarkably efficient; it shows only 2% difference with theoretically highest performance.<page local="6" global="14"/> However, the dynamic window approach increases the proc­essing time too much.</p><table class="main" frame="box" rules="all" border="1" regular="False"><tr class="row"><td class="cell"></td><td class="cell"></td><td class="cell"></td><td class="cell"></td><td class="cell"></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p>English Word</p></td><td class="cell"><p>Jacey</p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p>Chinese Sentence</p></td><td class="cell"><p></p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p>English Sentence</p></td><td class="cell"><p>The authors of this book are Peni-nah Thomson and Jacey Grahame.</p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p>Incorrectly extracted transliteration</p></td><td class="cell"><p>^(ZheShi)</p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p>Correct transliteration</p></td><td class="cell"><p>*ffi(JieXi)</p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p>Steps</p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p>1. Separate the Chinese sentence with characters,       61), <i>Ml, </i>fP" (including non-Chinese characters such as punctuation, number, English characters etc.), which have never been used in Chinese transliteration as follows:</p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p><i>2.</i><i> </i>Apply statistical transliteration model to each part and se­lect the part with highest score, and back-track the part to ex­tract a correct transliteration.</p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p>No.</p></td><td class="cell"><p>Chinese character sequence of each part (underline the back­tracking result)</p></td><td class="cell"><p>Score (normalize with window size)</p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p>1</p></td><td class="cell"><p>^ (BenShu)</p></td><td class="cell"><p>-24.79</p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p>2</p></td><td class="cell"><p>fF(ZuoZhe)</p></td><td class="cell"><p>-15.83</p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p>3</p></td><td class="cell"><p>(PeiNiNaTangMuShen)</p></td><td class="cell"><p>-16.32</p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p>4</p></td><td class="cell"><p><i>fcWffimipM </i>(JieXi)</p></td><td class="cell"><p>-10.29</p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"></td><td class="cell"></td><td class="cell"></td><td class="cell"></td><td class="cell"></td></tr></table><table class="main" frame="box" rules="all" border="1" regular="False"><tr class="row"><td class="cell"></td><td class="cell"></td><td class="cell"></td><td class="cell"></td><td class="cell"></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p>Methods</p></td><td class="cell"><p>Word precision</p></td><td class="cell"><p>Character precision</p></td><td class="cell"><p>Character recall</p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p>STM (baseline)</p></td><td class="cell"><p>75.33%</p></td><td class="cell"><p>86.65%</p></td><td class="cell"><p>91.11%</p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p>STM+DW</p></td><td class="cell"><p>96.00%</p></td><td class="cell"><p>98.51%</p></td><td class="cell"><p>99.05%</p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p>STM+TOK</p></td><td class="cell"><p>78.66%</p></td><td class="cell"><p>85.24%</p></td><td class="cell"><p>86.94%</p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p>STM+DW+TOK</p></td><td class="cell"><p>99.00%</p></td><td class="cell"><p>99.78%</p></td><td class="cell"><p>99.72%</p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p>STM+CW</p></td><td class="cell"><p>98.00%</p></td><td class="cell"><p>98.81%</p></td><td class="cell"><p>98.69%</p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p>STM+CW+TOK</p></td><td class="cell"><p>99.00%</p></td><td class="cell"><p>99.89%</p></td><td class="cell"><p>99.61%</p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"></td><td class="cell"></td><td class="cell"></td><td class="cell"></td><td class="cell"></td></tr></table><p>When using tokenizer method (STM+TOK), only about 3% is approved over the baseline. Al­though the result is not considerably improved, it is extremely important that the problems that the dy­namic window method cannot solve are managed to be solved. Thus, when using both dynamic win­dow and tokenizer methods with STM (STM+ DW+TOK), it is found that around 3% improve­ment is achieved over using only the dynamic win­dow (STM+DW), as well as word precision rates of 99%.</p><p>[t<u>able 2]. Processing time evaluation of proposed method</u>s</p><p>Table 2 shows the evaluation of processing time of dynamic window and tokenizer methods. Using the dynamic window leads to 27 times more proc­essing time than STM, while using the tokenizer method with the dynamic window method reduces the processing time around 5 times than the origi­nal. Hence, we have achieved a higher precision as well as less processing time by combining these two methods.</p></subsection><subsection number="3.3" title="Comparing experiment"><p>In order to compare with previous methods, we perform the experiment presented in [1]. Table 3 shows using the post-processing method presented in [1] achieves around 87% of word precision rates, and about 12% improvement over the baseline. However, our methods are 11% superior to the method in [1].</p><p>[Table 3] Comparing experiment with previous work</p></subsection></section><section number="4" title="Conclusions and future work"><p>In this paper, we presented two novel approaches called dynamic window and tokenizer based on the statistical machine transliteration model. Our ap­proaches achieved high precision without any post­processing procedures. The dynamic window ap­proach was based on a fundamental property, which more TUs aligned between correct translit­eration pairs. Also, we reasonably estimated the range of correct transliteration's length to extract transliteration pairs in high precision. The token-izer method eliminated characters that have never been used in Chinese transliteration to separate a sentence into several parts. This resulted in a cer­tain degree of improvement of precision and sig­nificantly reduction of processing time. These two methods are both based on common natures of all languages; thus our approaches can be readily port to other language pairs.</p><p>In this paper, we only considered the English words that are to be transliterated into Chinese. Our work is ongoing, and in near future, we will extend our works to extract transliteration pairs from large scale comparable corpora. In compara­ble corpora, there are many uncertainties, for ex­ample, the extracted English word may be not transliterated into Chinese or there may be no cor­rect transliteration in Chinese texts. However, with large comparable corpora, a word will appear sev­eral times, and we can use the frequency or entropy information to extract correct transliteration pairs based on the proposed perfect algorithm.</p><p>Acknowl edgem ent</p><p>This work was supported by the Korea Science and Engineer­ing Foundation (KOSEF) through the Advanced Information Technology Research Center (AITrc), also in part by the BK 21 Project and MIC &amp; IITA through IT Leading R&amp;D Support Project in 2007.</p><p>Reference</p><p>[1] C.-J. Lee, J.S. Chang, J.-S.R. Jang, Extraction of translit­eration pairs from parallel corpora using a statistical translit­eration model, in: Information Sciences 176, 67-90 (2006) [2] Richard Sproat, Tao Tao, ChengXiang Zhai, Named Entity Transliteration with Comparable Corpora, in: Proceedings of the 21st International Conference on Computational Linguis­tics. (2006) [3] J.S. Lee and K.S. Choi, "English to Korean statistical transliteration for information retrieval," International Journal of Computer Processing of Oriental Languages, pp.17-37, (1998).</p><table class="main" frame="box" rules="all" border="1" regular="False"><tr class="row"><td class="cell"></td><td class="cell"></td><td class="cell"></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p><b>Methods</b></p></td><td class="cell"><p><b>Processing time</b></p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p>STM (baseline)</p></td><td class="cell"><p>5 sec (5751 milisec)</p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p>STM+DW</p></td><td class="cell"><p>2min 34sec (154893 milisec)</p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p>STM+TOK</p></td><td class="cell"><p>4sec (4574 milisec)</p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p>STM+DW+TOK</p></td><td class="cell"><p>32sec (32751 milisec)</p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"></td><td class="cell"></td><td class="cell"></td></tr></table><table class="main" frame="box" rules="all" border="1" regular="False"><tr class="row"><td class="cell"></td><td class="cell"></td><td class="cell"></td><td class="cell"></td><td class="cell"></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p>Methods</p></td><td class="cell"><p>Word Precision</p></td><td class="cell"><p>Character Precision</p></td><td class="cell"><p>Character Recall</p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p>STM (baseline)</p></td><td class="cell"><p>75.33%</p></td><td class="cell"><p>86.65%</p></td><td class="cell"><p>91.11%</p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p>STM+DW+TOK</p></td><td class="cell"><p>99.00%</p></td><td class="cell"><p>99.78%</p></td><td class="cell"><p>99.72%</p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p>STM+[1]'s method</p></td><td class="cell"><p>87.99%</p></td><td class="cell"><p>90.17%</p></td><td class="cell"><p>91.11%</p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"></td><td class="cell"></td><td class="cell"></td><td class="cell"></td><td class="cell"></td></tr></table><page local="7" global="15"/><p>[4] K. Knight, J. Graehl, Machine transliteration, Computa­tional Linguistics 24 (4), 599-612, (1998). [5] W.-H. Lin, H.-H. Chen, Backward transliteration by learn­ing phonetic similarity, in: CoNLL-2002, Sixth Conference on Natural Language Learning, Taipei, Taiwan, (2002). [6] J.-H. Oh, K.-S. Choi, An English-Korean transliteration model using pronunciation and contextual rules, in: Proceed­ings of the 19th International Conference on Computational Linguistics (COLING), Taipei, Taiwan, pp. 758-764, (2002). [7] C.-J. Lee, J.S. Chang, J.-S.R. Jang, A statistical approach to Chinese-to-English Backtransliteration, in: Proceedings of the 17th Pacific Asia Conference on Language, Information, and Computation (PACLIC), Singapore, pp. 310-318, (2003). [8] Jong-Hoon Oh, Sun-Mee Bae, Key-Sun Choi, An Algo­rithm for extracting English-Korean Transliteration pairs using Automatic E-K Transliteration In Proceedings of Korean In­formation Science Socieity (Spring). (In Korean), (2004). [9] Jong-Hoon Oh, Jin-Xia Huang, Key-Sun Choi, An Align­ment Model for Extracting English-Korean Translations of Term Constituents, Journal of Korean Information Science Society, SA, 32(4), (2005) [10] Chun-Jen Lee, Jason S. Chang, Jyh-Shing Roger Jang: Alignment of bilingual named entities in parallel corpora us­ing statistical models and multiple knowledge sources. ACM Trans. Asian Lang. Inf. Process. 5(2): 121-145 (2006) [11] Lee, C. J. and Chang, J. S., Acquisition of English-Chinese Transliterated Word Pairs from Parallel-Aligned Texts Using a Statistical Machine Transliteration Model, In. Proceedings of HLT-NAACL, Edmonton, Canada, pp. 96-103, (2003).</p><p>[12] Xinhua Agency, Names of the world's peoples: a com­prehensive dictionary of names in Roman-Chinese (ttt-W-A^a S»*S$*),(1993)</p></section></body></article>