<?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="1031"/><title>Mining and Modeling Relations between Formal and Informal Chinese Phrases from Web Corpora</title><author surname="Li" givenname="Zhifei"><org  name="Johns Hopkins University" country="USA" city="Baltimore"/></author><author surname="Yarowsky" givenname="David"><org  name="Johns Hopkins University" country="USA" city="Baltimore"/></author></firstpageheader><frontmatter><p><b>Mining and Modeling Relations between Formal and Informal Chinese Phrases from Web Corpora</b></p><p><b>Zhifei Li  </b>and  <b>David Yarowsky</b></p><p>Department of Computer Science and Center for Language and Speech Processing Johns Hopkins University, Baltimore, MD 21218, USA</p><p>zhifei.work@gmail.com and yarowsky@cs.jhu.edu</p></frontmatter><abstract></abstract></header><body><section title="Formal"><p><b>Informal</b></p><p>We present a novel method for discovering and modeling the relationship between in­formal Chinese expressions (including collo­quialisms and instant-messaging slang) and their formal equivalents. Specifically, we pro­posed a bootstrapping procedure to identify a list of candidate informal phrases in web corpora. Given an informal phrase, we re­trieve contextual instances from the web us­ing a search engine, generate hypotheses of formal equivalents via this data, and rank the hypotheses using a conditional log-linear model. In the log-linear model, we incorpo­rate as feature functions both rule-based intu­itions and <i>data co-occurrence </i>phenomena (ei­ther as an explicit or indirect definition, or through formal/informal usages occurring in free variation in a discourse). We test our system on manually collected test examples, and find that the <i>(formal-informal) </i>relation­ship discovery and extraction process using our method achieves an average 1-best preci­sion of 62%. Given the ubiquity of informal conversational style on the internet, this work has clear applications for text normalization in text-processing systems including machine translation aspiring to broad coverage.</p></section><section number="1" title="Introduction"><p>Informal text (e.g., newsgroups, online chat, blogs, etc.) is the majority of all text appearing on the Inter­net. Informal text tends to have very different style from formal text (e.g., newswire, magazine, etc.). In particular, they are different in vocabulary, syn­tactic structure, semantic interpretation, discourse</p><doubt alpha="58.6" length="29" tooSmall="False" monospace="0.0">W%(BaiBai)[bye-bye] 88 (BaBa)</doubt><p>MX (XiHuan)[like] <i>MÏR </i>(XiFan)[gruel]</p><p>W (GeGe)[elder brother] GG</p><p>tfôzË (GeMi)[fans] fôiÉ (FenSi)[a food]</p><p>Table 1: Example Chinese <i>Formal-informal </i>Relations. The <i>PinYin </i>pronunciation is in parentheses and an op­tional literal gloss is in brackets.</p><p>structure, and so on. On the other hand, certain re­lations exist between the informal and formal text, and informal text often has a viable formal equiva­lent. Table 1 shows several naturally occurring ex­amples of informal expressions in Chinese, and Ta­ble 2 provides a more detailed inventory and charac­terization of this phenomena<footnote anchor="1"/>. The first example of informal phrase "88" is used very often in Chinese on-line chat when a person wants to say "bye-bye" to the other person. This can be explained as fol­lows. In Chinese, the standard equivalent to "bye-bye" is whose <i>PinYin </i>is "BaiBai". Coincidentally, the <i>PinYin </i>of "88" is "BaBa". Because "BaBa" and "BaiBai" are near homophones, people often use "88" to represent either for input convenience or just for fun. The other relations in Table 1 are formed due to similar processes as will be described later.</p><p>Due to the often substantial divergence between</p><footnote label="1">For clarity, we represent Chinese words in the format: Chi­nese characters (optional PinYin equivalent in parentheses and optional English gloss in brackets).</footnote><doubt alpha="0.0" length="4" tooSmall="False" monospace="0.0">1031</doubt><p><i>Proceedings of the 2008 Conference on Empirical Methods in Natural Language Processing, </i>pages 1031-1040, Honolulu, October 2008. ©2008 Association for Computational Linguistics<page local="2" global="1032"/></p><p>informal and formal text, a text-processing system trained on formal text does not typically work well on informal genres. For example, in a machine translation system (Koehn et al., 2007), if the bilin­gual training data does not contain the word "M W" (the second example in Table 1), it leaves the word untranslated. On the other hand, if the word "M W" does appear in the training data but it has only a translation "gruel" as that is the meaning in the formal text, the translation system may wrongly translate "M W" into "gruel" for the <i>informal </i>text where the word "M W" is more likely to mean "like". Therefore, as a text-normalization step, it is desirable to transform the informal text into its standard formal equivalent before feeding it into a general-purpose text-processing system. Unfortu­nately, there are many processes for generating in­formal expressions in common use today. Such transformations are highly flexible/diverse, and new phrases are invented on the Internet every day due to major news events, popular movies, TV shows, ra­dio talks, political activities, and so on. Therefore, it is of great interest to have a <i>data-driven </i>method that can <i>automatically </i>find the relations between in­formal and formal expressions.</p><p>In this paper, we present a novel method for dis­covering and modeling the relationship between in­formal Chinese expressions found in web corpora and their formal equivalents. Specifically, we im­plement a bootstrapping procedure to identify a list of candidate informal phrases. Given an indi­vidual informal phrase, we retrieve contextual in­stances from the web using a search engine (in this case, www.baidu.com<i>), </i>generate hypotheses of for­mal equivalents via this data, and rank the hypothe­ses using a conditional log-linear model. In the log­linear model, we incorporate as feature functions both rule-based intuitions and <i>data co-occurrence </i>phenomena (either as an explicit or indirect defini­tion, or through formal/informal usages occurring in free variation in a discourse). We test our system on manually collected test examples<footnote anchor="2"/>, and find that the <i>(formal-informal) </i>relationship discovery and extrac­tion process using our method achieves an average precision of more than 60%. This work has applications for text normalization in many general-purpose text-processing tasks, e.g., machine translation.</p><footnote label="2">The training and test examples are freely available at http://www.cs.jhu.edu/~zfli .</footnote><p>To the best of our knowledge, our work is the first published machine-learning approach to pro­ductively model the broad types of relationships be­tween informal and formal expressions in Chinese using web corpora.</p></section><section number="2" title="Formal to Informal: Phenomena and Examples"><p>In this section, we describe the phenomena and pro­vide examples of the relations between formal and informal expressions in Chinese (we refer to the relation as <i>formal-informal </i>phrases hereafter, even in the case of single-word expressions). We man­ually collected 908 <i>formal-informal </i>relations, and classified these relations into four categories. We collected these pairs by investigating multiple web­pages where the <i>formal-informal </i>relations are man­ually compiled, and then merged these seed relations and removed duplicates. In this way, the 908 exam­ples should give good coverage on the typical cat­egories in the <i>formal-informal </i>relations. Also, the distribution of the categories found in the 908 exam­ples should be representative of the actual distribu­tion of the <i>formal-informal </i>relations occurring in the real text. Table 2 presents these categories and ex­amples in each category. In the last column, the table also shows the relative frequency of each category, computed based on the 908 examples. Recall that we represent Chinese words in the format: Chinese characters (optional <i>PinYin </i>equivalent in parenthe­ses and optional English gloss in brackets).</p><subsection number="2.1" title="Homophone"><p>In general, a homophone is a word that is pro­nounced the same as another word but differs in meaning and/or written-form. Here, we use the word "homophone" in a loose way. In particular, we re­fer an informal phrase as a homophone of a formal phrase if its pronunciation is the same or <i>similar </i>to the formal phrase. In the three examples belonging to the homophone category in Table 2, the first ex­ample is a true homophone, while the other two are loose homophones. The third example represents a major sub-class where the informal phrase is a <i>num­ber </i>(e.g., 88).</p><doubt alpha="0.0" length="4" tooSmall="False" monospace="0.0">1032</doubt><page local="3" global="1033"/><doubt alpha="100.0" length="6" tooSmall="False" monospace="0.0">Formal</doubt><p><b>Informal</b> <b>Category</b></p><p>For illustrative purposes, we can present the <i>transformation path </i>showing how the informal phrase is obtained from the formal phrase. In par­ticular, the <i>transformation path </i>for this category is "Formal — PinYin — Informal (similar or same PinYin as the formal phrase)".</p></subsection><subsection number="2.2" title="Abbreviation and Acronym"><p>A Chinese <i>abbreviation </i>of a formal phrase is ob­tained by selecting <i>one or more </i>characters from this formal phrase, and the selected characters can be at <i>any </i>position in the formal phrase (Li and Yarowsky, 2008; Lee, 2005; Yin, 1999). In comparison, an <i>acronym </i>is a special form of abbreviation, where only the first character of each word in the formal phrase is selected to form the informal phrase. Table 2 presents three examples belonging to this category. While the first example is an abbreviation, and the other two examples are acronyms.</p><p>The <i>transformation path </i>for the second exam­ple is "Formal — PinYin — Acronym", and the <i>transformation path </i>for the third example is "For­mal — English — Acronym". Clearly, they differ in whether <i>PinYin </i>or English is used as a bridge.</p></subsection><subsection number="2.3" title="Transliteration"><p>A transliteration is transcribing a word or text writ­ten in one writing system into another writing sys­tem. Table 2 presents examples belonging to this category. In the first example, the Chinese infor­mal phrase "f&amp;li. (FenSi)[a Chinese food]" can be thought as a transliteration of the English phase "fans" as the pronunciation of "fans" is quite sim­ilar to the <i>PinYin </i>"FenSi".</p><p>The <i>transformation path </i>for this category is "For­mal — English — Chinese Transliteration".</p></subsection><subsection number="2.4" title="Others"><p>Due to the inherently informal and flexible nature of expressions in informal genre, the formation of an informal phrase can be very complex or ad-hoc. For example, an informal phrase can be generated by ap­plying the above transformation rules <i>jointly. </i>More importantly, many relations cannot be described us­ing a simple set of rules. Table 2 presents three such examples, where the first two examples are gener­ated by applying rules <i>jointly </i>and the third example is created by decomposing the Chinese characters in the formal form. The statistics collected from the 904 examples tells us that about 45% of the relations belonging to this category. This motivates us to use a <i>data-driven </i>method to automatically discover the relations between informal and formal phrases.</p></subsection></section><section number="3" title="Data Co-occurrence"><p>In natural language, related words tend to appear to­gether (i.e., co-occurrence). For example, <i>Bill Gates</i> tends to appear together with <i>Microsoft </i>more of­ten than expected by chance.<page local="4" global="1034"/> Such co-occurrence may imply the existence of a relationship, and is ex­ploited in <i>formal-informal </i>relation discovery under different conditions.</p><doubt alpha="0.0" length="4" tooSmall="False" monospace="0.0">1033</doubt><table caption="Table 2: ChineseFormal-informalRelations: Categories and Examples. Literal glosses in brackets." 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>Homophone</b></p></td><td class="cell"><p><i>¥ij</i><i>£3</i><i>i </i>(BanZhu) [system administrator]</p></td><td class="cell"><p>ffift (BanZhu) [bamboo]</p></td><td class="cell"><p>4.2</p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p></p></td><td class="cell"><p>MX (XiHuan)[like]</p></td><td class="cell"><p>MW (XiFan)[gruel]</p></td><td class="cell"><p>4.4</p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p></p></td><td class="cell"><p>(BaiBai)[bye-bye]</p></td><td class="cell"><p>88 (BaBa)</p></td><td class="cell"><p>21</p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p><b>Abbreviation</b></p></td><td class="cell"><p>(MeiGuoJunDui)[american army]</p></td><td class="cell"><p>(MeiJun)[american army]</p></td><td class="cell"><p>3.8</p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p><b>Acronym</b></p></td><td class="cell"><p>Iflf (GeGe)[elder brother]</p></td><td class="cell"><p>GG</p></td><td class="cell"><p>12.3</p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p></p></td><td class="cell"><p>(NuPengYou)[girl friend]</p></td><td class="cell"><p>GF</p></td><td class="cell"><p>7.2</p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p><b>Transliteration</b></p></td><td class="cell"><p>(GeMi)[fans]</p></td><td class="cell"><p>föÜ (FenSi)[a Chinese food]</p></td><td class="cell"><p>2.3</p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p></p></td><td class="cell"><p>(XieXie)[thank you]</p></td><td class="cell"><p>3Q (SanQiu)</p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p><b>Others</b></p></td><td class="cell"><p>#IÈJI$jviÉ (XiLaLiFenSi)[fans of Hilary]</p></td><td class="cell"><p>MW (XiFan)[gruel]</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>lE^fôË (AoBaMaFenSi)[fans of Obama]</p></td><td class="cell"><p>Üfft (OuFen)[a food]</p></td><td class="cell"><p>44.8</p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p></p></td><td class="cell"><p>II'jS (ChaoQiang)[super strong]</p></td><td class="cell"><p>7t"i=î ^ S (ZouZhaoGongXu)</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><subsection number="3.1" title="Data Co-occurrence in Definitions"><p>In general, for many informal phrases in popular use, there is likely to be an explicit definition somewhere that provides or paraphrases its meaning for an unfa­miliar audience. People have created dedicated def­inition web-pages to explain the relations between formal and informal phrases. For example, the first example in Table 3 is commonly explained in many dedicated definition web-pages on the Internet. On the other hand, in some formal text (e.g., research papers), people tend to define the informal phrase before it is used frequently in the later part of the text. The second example of Table 3 illustrates this phenomena. Clearly, the definition text normally contains salient patterns. For example, the first ex­ample follows the <i>"infonnal/kformalifi) </i>" defi­nition pattern, while the second example follows the pattern <i>"formal (informal)". </i>This gives us a reliable way to seed and bootstrap a list of informal phrases as will be discussed in Section 4.1.</p><p><b>Relation Definition Text</b></p><p>to)        £ s m (ffl <i>n</i><i>mm</i></p><table caption="Table 3: Data Co-occurrence in Definitions"></table></subsection><subsection number="3.2" title="Data Co-occurrence in Online Chat"><p>Informal phrases appear in online chat very often for input convenience or just for fun. Since different people may have different ways or traditions to ex­press semantically-equivalent phrases, one may find many nearby <i>data co-occurrence </i>examples in chat text. For example, in Table 4, after a series of mes­sage exchanges, person A wants to end the conver­sation and types "WW" (meaning "bye-bye"), per­son B later includes the same semantic content, but in a different (more or less formal) expression (e.g.</p><p>"88").</p><p>PersonA: <i>ÄM^Tt^T </i>Person A: WW</p><doubt alpha="58.3" length="12" tooSmall="False" monospace="0.0">Person B: 88</doubt><table caption='Table 4: Data Co-occurrence in Online Chat for Relation(WW,88) meaning "bye-bye"'></table></subsection><subsection number="3.3" title="Data Co-occurrence in News Articles"><p>For some <i>formal-informal </i>relations, since both of the informal and formal phrases have been used in public very often and people are normally aware of these relations, an author may use the informal and formal phrases interchangeably without bother­ing to explain the relations. This is particularly true in news articles for some well-known relations. Ta­ble 5 shows an example, where the abbreviation j* 1z?" (meaning "winter olympics") appears in the title and its full-form "^gpjiljg;^" appears in the text of the same document. In general, the relative distance between an informal phrase and its formal phrase varies. For example, they may appear in the same sentence, or in neighboring sentences.</p><doubt alpha="38.2" length="68" tooSmall="False" monospace="0.0">Textif^m f2n9h&amp;(is *t^ mm)M20mm * &amp;#w n ¥ ^ # ^ a *m^10h^8ä# m f * #</doubt><table caption='Table 5: Data Co-occurrence in News Article forRelation ^*#) meaning "winter olympics"'></table><p><b>4  Mining Relations between Informal and Formal Phrases from Web</b></p><p>In this section, we describe an approach that auto­matically discovers the relation between a formal phrase and an informal phrase from web corpora. Specifically, we propose a bootstrapping procedure to identify a list of candidate informal phrases. Given a target informal phrase, we retrieve a large set of of instances in context from the Web, generate candidate hypotheses (i.e, candidate formal phrases) from the data, and rank the hypotheses by using a conditional log-linear model. The log-linear model is very flexible to incorporate both the rule- and datadriven intuitions (described in Sections 2 and 3, re­spectively) into the model as feature functions.<page local="5" global="1035"/></p><doubt alpha="0.0" length="4" tooSmall="False" monospace="0.0">1034</doubt></subsection><subsection number="4.1" title="Identifying Informal Phrases"><p>Before finding the formal phrase corresponding to an informal phrase, we first need to identify infor­mal phrases of interest. For example, one can collect informal phrases manually. However, this is too ex­pensive as new relations between informal and for­mal phrases emerge every day on the Internet. Alter­natively, one can employ a large amount of formal text (e.g., newswire) and informal text (e.g., Inter­net blogs) to derive such a list as follows. Specifi­cally, from the informal corpus we can extract those phrases whose frequency in the informal corpus is significantly different from that in the formal cor­pus. However, such a list may be quite noisy, i.e., many of them are not informal phrases at all.</p><p>An alternative approach to extracting the infor­mal phrases is to use a bootstrapping algorithm (e.g., Yarowsky (1995)). Specifically, we first manually collect a small set of example relations. Then, using these relations as a <i>seed set, </i>we extract the text <i>pat­terns </i>(e.g., the definition pattern showing how the informal and formal phrases co-occur in the data as discussed in Section 3.1). With these patterns, we identify many more new relations from the data and augment them into the seed set. The procedure it­erates. Using such an approach, we should be able to extract a large list of <i>formal-informal </i>relations. Clearly, the list extracted in this way may be quite noisy, and thus it is important to exploit both the data- and rule-driven intuitions to rank these rela­tions properly.</p></subsection><subsection number="4.2" title="Retrieving Data from Web"><p>Given an informal phrase, we retrieve training data from the web on the fly. Specifically, we first use a search engine to identify a set of hyper-links that point to web pages containing contexts relevant to the informal phrase, and then follow the hyper-links to download the web pages. The input to the search engine is a text query. One can simply use the infor­mal phrase as a query. However, this may lead to a set of pages that have nothing to do with the infor­mal phrase. For example, if we search the informal phrase "88" (the third example in Table 2) using the well-known Chinese search engine www.baidu.com<i>,</i> none of the top-10 pages are related to the infor­mal phrase "88". To avoid this situation, one can use a search engine that is dedicated to informal text search (e.g., blogsearch.baidu.com<i>). </i>Alternatively, one can use the general-purpose search engine but expanding the query with domain information. For example, for the informal phrase "88", we can use a query "88 where means <i>internet language.</i><i></i></p></subsection><subsection number="4.3" title="Generating Candidate Hypotheses"><p>Given an informal phrase, we generate a set of hy­potheses which are candidate formal phrases corre­sponding to the informal phrase. We considered two general approaches to the generation of hypotheses.</p><p><b>Rule-driven Hypothesis Generation: </b>One can use the rules described in Section 2 to generate a set of hypotheses. However, with this approach, one may generate an exponential number of hypotheses. For example, assuming the number of English words starting with a given letter is O(| V|), we can generate O(| V |n) hypotheses given an <i>acronym </i>containing n letters. Another problem with this approach is that a relation between an informal phrase and a formal phrase may not be explained by a specific rule. In fact, as shown in the last row of Table 2, such rela­tions consist of 44.8% of all corpus instances.</p><p><b>Data-driven Hypothesis Generation: </b>With data retrieved from the Web, we can generate hypotheses by enumerating the frequent <i>n</i>-grams co-occurring with the informal phrase within certain distance. This exploits the <i>data co-occurrence </i>phenomena de­scribed in Section 3, that is, the formal phrase tends to co-occur with the informal phrase nearby in the data, for the multiple reasons described above. This can deal with the cases where the relation between an informal phrase and a formal phrase cannot be explained by a rule. However, it also suffers from the over-generation problem as in the rule-driven ap­proach.</p><p>In this paper, we use the data-driven method to generate hypotheses, and rank the hypotheses using a conditional log-linear model that incorporates both the rule and data intuitions as feature functions.</p><doubt alpha="0.0" length="4" tooSmall="False" monospace="0.0">1035</doubt><page local="6" global="1036"/></subsection><subsection number="4.4" title="Ranking Hypotheses: Conditional Log-linear Model"><p>Log-linear models are known for flexible incorpora­tion of features into the model. Each feature func­tion reflects a hint/intuition that can be used to rank the hypotheses. In this subsection, we develop a conditional log-linear model that incorporates both the rule and data intuitions as feature functions.</p><subsubsection number="4.4.1" title="Conditional Log-linear Model"><p>Given an informal phrase (say x) and a candidate formal phrase (say y), the model assigns the pair a score (say s(x, y)), which will be used to rank the hypothesis y. The score <i>s(x,</i><i> </i><i>y)</i><i> </i>is a linear combina­tion of the feature scores (say $i(x, y)) over a set of feature functions indexed by i. Formally, where K is the number of feature functions defined and a<i>i </i>is the weight assigned to the i-th feature func­tion (i.e., To learn the weight vector <i>a</i>, we first define a probability measure, where Z (x, <i>a ) </i>is a normalization constant. Now, we define the regularized log-likelihood <i>(LLR)</i><i> </i>of the training data (i.e, a set of pairs of (x, y)), as follows, where N is the number of training examples, and the regularization term <i>^2 </i>is a Gaussian prior with a variance <i>a<footnote anchor="2"/> </i>(Roark et al., 2007). The optimal weight vector <i>a * </i>is obtained by maximizing the regularized log-likelihood <i>(LLR),</i><i> </i>that is, arg max <i>LLR(a)</i></p><doubt alpha="100.0" length="1" tooSmall="False" monospace="0.0">K</doubt><doubt alpha="50.0" length="18" tooSmall="False" monospace="0.0">s(x,y) =$i(x,y)xai</doubt><doubt alpha="0.0" length="3" tooSmall="False" monospace="0.0">(1)</doubt><doubt alpha="66.7" length="3" tooSmall="False" monospace="0.0">i=l</doubt><doubt alpha="44.4" length="9" tooSmall="False" monospace="0.0">Pa(y|x) =</doubt><doubt alpha="0.0" length="1" tooSmall="False" monospace="0.0">1</doubt><doubt alpha="60.0" length="5" tooSmall="False" monospace="0.0">Z(x,a</doubt><doubt alpha="0.0" length="3" tooSmall="False" monospace="0.0">(2)</doubt><doubt alpha="100.0" length="1" tooSmall="False" monospace="0.0">N</doubt><doubt alpha="58.8" length="17" tooSmall="False" monospace="0.0">LLR(a)= ^ logPa(y</doubt><doubt alpha="0.0" length="3" tooSmall="False" monospace="0.0">3=1</doubt><doubt alpha="100.0" length="1" tooSmall="False" monospace="0.0">a</doubt><doubt alpha="20.0" length="5" tooSmall="False" monospace="0.0">3|x3)</doubt><doubt alpha="33.3" length="3" tooSmall="False" monospace="0.0">2a2</doubt><doubt alpha="0.0" length="3" tooSmall="False" monospace="0.0">(3)</doubt><doubt alpha="100.0" length="2" tooSmall="False" monospace="0.0">aa</doubt><doubt alpha="0.0" length="3" tooSmall="False" monospace="0.0">(4)</doubt><p>To maximize the above function, we use a limited-memory variable method (Benson and More, 2002) that is implemented in the TAO package (Benson et al., 2002) and has been shown to be very effective in various natural language processing tasks (Malouf, 2002).</p><p>During test time, the following <b>decision rule </b>is normally used to predict the optimal formal phrase <i>y* </i>for a given informal phrase <i>x,</i></p><doubt alpha="50.0" length="2" tooSmall="False" monospace="0.0">y*</doubt><p>arg max <i>y</i> s(x,y).</p><doubt alpha="0.0" length="3" tooSmall="False" monospace="0.0">(5)</doubt></subsubsection><subsubsection number="4.4.2" title="Feature Functions"><p>As mentioned before, we incorporate both the rule- and data-driven intuitions as feature functions in the log-linear model.</p><p><b>Rule-driven feature functions: </b>Clearly, if a pair <i>(x,</i><i> </i><i>y)</i><i> </i>matches the rule patterns described in Table 2, the pair has a high possibility to be a true <i>formal-informal </i>relation. To reflect this intuition, we de­velop several feature functions as follows.</p><p>• LD-PinYin(x, y): the <i>Levenshtein distance </i>on PinYin of <i>x</i><i> </i>and <i>y.</i><i> </i>The distance between two PinYin characters is weighted based on the similarity of pronunciation, for example, the weight <i>w(l,n)</i><i> </i>is smaller than the weight <i>w(a,</i><i> </i><i>z).</i></p><p>• LEN-PinYin<i>(x,</i><i> </i><i>y):</i><i> </i>the difference in the num­ber of PinYin characters between <i>x</i><i> </i>and <i>y.</i></p><p>• Is-PinYin-Acronym<i>(x,</i><i> </i><i>y):</i><i> </i>is  <i>x</i><i> </i>a PinYin acronym of y? For example, Is-PinYin-Acronym(GG, ffff)=1, Is-PinYin-Acronym(GG, JnLI^)=0.</p><p>• Is-CN-Abbreviation<i>(x,</i><i> </i><i>y):</i><i> </i>is <i>x</i><i> </i>a Chinese ab­breviation of <i>y? </i>For example, Is-CN-Abbreviation(§t¥,§tH¥PA)=1, Is-CN-Abbreviation(^¥,^ H¥PA)=0.</p><p><b>Data-driven feature functions: </b>As described in Section 3, the informal and formal phrases tends to co-occur in the data. Here, we develop several fea­ture functions to reflect this intuition.</p><p>• <i>n</i>-gram co-occurrence relative frequency: we collect the <i>n</i>-grams that occur in the data within a window of the occurrence of the informal phrase, and compute their relative frequency as feature values. Since different orders of grams will have quite different statistics, we define 7 features in this category: 1-gram, 2-gram, 3-gram, 4-gram, 5-gram, 6to10-gram, and 11to15-gram. Note that the <i>order n </i>of a <i>n</i>-gram is in terms of number of Chinese char­acters instead of words.</p><doubt alpha="0.0" length="4" tooSmall="False" monospace="0.0">1036</doubt><page local="7" global="1037"/><p>• Features on a definition pattern: we have dis­cussed definition patterns in Section 3.1. For each definition pattern, we can define a feature function saying that if the co-occurrence of <i>x </i>and y satisfies the definition pattern, the feature value is one, otherwise is zero.</p><p>• Features on the number of relevant web-pages: another interesting feature function can be de­fined as follows. For each candidate relation <i>(x,</i><i> </i><i>y),</i><i> </i>we use the pair as a query to search the web, and treat the <i>number </i>of pages returned by the search engine as a feature value.<footnote anchor="3"/> However, these features are quite expensive as millions of queries may need to be served.</p></subsubsection></subsection></section><section number="5" title="Experimental Results"><p>Recall that in Section 2 we categorize the <i>formal-informal </i>relations based on the manually collected relations. In this section, we use a subset of them for training and testing. In particular, we use 252 exam­ples to train the log-linear model that is described in Section 4, and use 249 examples as test data to compute the precision.<footnote anchor="4"/></p><p>Table 6 shows the weights<footnote anchor="5"/> learned for the var­ious feature functions described in Section 4.4. Clearly, different feature functions get quite differ­ent weights. This is intuitive as the feature functions may differ in the scale of the feature values or in their importance in ranking the hypotheses. In fact, this shows the importance of using the log-linear model to learn the optimal weights in a principled and automatic manner, instead of manually tuning the weights in an ad-hoc way.</p><p>Tables 7-9 show the precision results for different categories as described in Section 2, using the rule-driven, data-driven, or both rule and data-driven fea­tures, respectively. In the tables, the precision corre­sponding to the "top<i>-N" </i>is computed in the follow­ing way: if the true hypothesis is among the top<i>-N</i><i> </i>hypotheses ranked by the model, we tag the classi­fication as correct, otherwise as wrong. Clearly, the larger the <i>N </i>is, the higher the precision is. Comput­ing the top<i>-N</i><i> </i>precision (instead of just computing the usual top-1 precision) is meaningful especially when we consider our relation extractor as an inter­mediate step in an end-to-end text-processing sys­tem (e.g., machine translation) since the final deci­sion can be delayed to later stage based on more ev­idence. In general, our model gets quite respectably high precision for such a task (e.g., more than 60% for top-1 and more than 85% for top-100) when us­ing both data and rule-driven features, as shown in Table 9. Moreover, the data-driven features are more helpful than the rule-driven features (e.g, 25.3% ab­solute improvement in 1-best precision), while the combination of these features does boost the perfor­mance of any individual feature set (e.g., 10.4% ab­solute improvement in 1-best precision over the case using data-driven features only).</p><footnote label="3">Note that the number of pages relevant to a query can be easily obtained as most search engines return this number.</footnote><footnote label="4">Again, the training and test examples are freely available at http://www.cs.jhu.edu/~zfli .</footnote><footnote label="5">Note that we do not use the features on definition patterns and on the number of relevant web pages, for efficiency.</footnote><p>We also carried out experiments (see Table 10) in the bootstrapping procedure described in Section 4.1. In particular, we start from a seed set having 130 relations. We identify the frequent patterns from the data retrieved from the web for these seed exam­ples. Then, we use these patterns to identify many more new possible <i>formal-informal </i>relations. After the first iteration, we select the top 3000 pairs of re­lations matched by the patterns. The recall of a man­ually collected test set (having 750 pairs) on these 3000 pairs is around 30%, which is quite promising given the highly noisy data.</p><doubt alpha="0.0" length="4" tooSmall="False" monospace="0.0">1037</doubt><table caption="Table 6: Optimal Weights in the Log-linear Model" 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></tr><tr class="row"><td class="cell"></td><td class="cell"><p><b>Category</b></p></td><td class="cell"><p><b>Feature</b></p></td><td class="cell"><p><b>Weight</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>LD-PinYin</p></td><td class="cell"><p>0.800</p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p><b>Rule-driven</b></p></td><td class="cell"><p>Len-PinYin Is-PinYin-Acronym</p></td><td class="cell"><p>0.781 7.594</p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p></p></td><td class="cell"><p>Is-CN-Abbreviation</p></td><td class="cell"><p>7.464</p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p></p></td><td class="cell"><p>1-gram</p></td><td class="cell"><p>14.506</p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p></p></td><td class="cell"><p>2-gram</p></td><td class="cell"><p>108.193</p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p></p></td><td class="cell"><p>3-gram</p></td><td class="cell"><p>82.975</p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p><b>Data-driven</b></p></td><td class="cell"><p>4-gram</p></td><td class="cell"><p>66.872</p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p></p></td><td class="cell"><p>5-gram</p></td><td class="cell"><p>42.258</p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p></p></td><td class="cell"><p>6to10-gram</p></td><td class="cell"><p>21.229</p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p></p></td><td class="cell"><p>11to15-gram</p></td><td class="cell"><p>0.985</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></tr></table><page local="8" global="1038"/><p>Table 7: <b>Rule-driven Features only: </b>Precision on Chinese <i>Formal-informal </i>Relation Extraction</p><p>Table 8: <b>Data-driven Features only: </b>Precision on Chinese <i>Formal-informal </i>Relation Extraction</p><p>Table 9: <b>Both Data and Rule-drive Features: </b>Precision on Chinese <i>Formal-informal </i>Relation Extraction<page local="9" global="1039"/></p><doubt alpha="0.0" length="4" tooSmall="False" monospace="0.0">1038</doubt><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><td class="cell"></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p></p></td><td class="cell"><p><b>Category</b></p></td><td class="cell"><p>Top-1</p></td><td class="cell"><p><b>Precision (%)</b></p><p>Top-10 Top-50</p></td><td class="cell"><p>Top-100</p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p><b>Homophone</b></p></td><td class="cell"><p>Same PinYin</p></td><td class="cell"><p>31.6</p></td><td class="cell"><p>47.4</p></td><td class="cell"><p>68.4</p></td><td class="cell"><p>73.7</p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p></p></td><td class="cell"><p>Similar PinYin</p></td><td class="cell"><p>15.0</p></td><td class="cell"><p>35.0</p></td><td class="cell"><p>45.0</p></td><td class="cell"><p>50.0</p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p></p></td><td class="cell"><p>Number</p></td><td class="cell"><p>31.6</p></td><td class="cell"><p>64.2</p></td><td class="cell"><p>84.2</p></td><td class="cell"><p>90.5</p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p><b>Abbreviation</b></p></td><td class="cell"><p>Chinese abbreviation</p></td><td class="cell"><p>11.8</p></td><td class="cell"><p>35.3</p></td><td class="cell"><p>41.2</p></td><td class="cell"><p>41.2</p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p><b>Acronym</b></p></td><td class="cell"><p>PinYin Acronym</p></td><td class="cell"><p>39.3</p></td><td class="cell"><p>82.1</p></td><td class="cell"><p>91.1</p></td><td class="cell"><p>92.9</p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p></p></td><td class="cell"><p>English Acronym</p></td><td class="cell"><p>3.1</p></td><td class="cell"><p>6.3</p></td><td class="cell"><p>9.4</p></td><td class="cell"><p>28.1</p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p><b>Transliteration</b></p></td><td class="cell"><p>10.0</p></td><td class="cell"><p>20.0</p></td><td class="cell"><p>20.0</p></td><td class="cell"><p>20.0</p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p><b>Average</b></p></td><td class="cell"><p></p></td><td class="cell"><p><i>26.1</i></p></td><td class="cell"><p><i>53.4</i></p></td><td class="cell"><p><i>66.3</i></p></td><td class="cell"><p><i>72.3</i></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><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><td class="cell"></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p></p></td><td class="cell"><p><b>Category</b></p></td><td class="cell"><p>Top-1</p></td><td class="cell"><p><b>Precision (%)</b></p><p>Top-10 Top-50</p></td><td class="cell"><p>Top-100</p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p><b>Homophone</b></p></td><td class="cell"><p>Same PinYin</p></td><td class="cell"><p>52.6</p></td><td class="cell"><p>73.7</p></td><td class="cell"><p>73.7</p></td><td class="cell"><p>78.9</p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p></p></td><td class="cell"><p>Similar PinYin</p></td><td class="cell"><p>45.0</p></td><td class="cell"><p>65.0</p></td><td class="cell"><p>75.0</p></td><td class="cell"><p>75.0</p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p></p></td><td class="cell"><p>Number</p></td><td class="cell"><p>66.3</p></td><td class="cell"><p>86.3</p></td><td class="cell"><p>94.7</p></td><td class="cell"><p>96.8</p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p><b>Abbreviation</b></p></td><td class="cell"><p>Chinese abbreviation</p></td><td class="cell"><p>0.0</p></td><td class="cell"><p>23.5</p></td><td class="cell"><p>47.1</p></td><td class="cell"><p>47.1</p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p><b>Acronym</b></p></td><td class="cell"><p>PinYin Acronym</p></td><td class="cell"><p>58.9</p></td><td class="cell"><p>78.6</p></td><td class="cell"><p>85.7</p></td><td class="cell"><p>87.5</p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p></p></td><td class="cell"><p>English Acronym</p></td><td class="cell"><p>25.0</p></td><td class="cell"><p>46.9</p></td><td class="cell"><p>68.6</p></td><td class="cell"><p>68.8</p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p><b>Transliteration</b></p></td><td class="cell"><p>50.0</p></td><td class="cell"><p>50.0</p></td><td class="cell"><p>50.0</p></td><td class="cell"><p>50.0</p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p><b>Average</b></p></td><td class="cell"><p></p></td><td class="cell"><p><i>51.4</i></p></td><td class="cell"><p><i>71.1</i></p></td><td class="cell"><p><i>81.1</i></p></td><td class="cell"><p><i>82.7</i></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><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><td class="cell"></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p></p></td><td class="cell"><p><b>Category</b></p></td><td class="cell"><p>Top-1</p></td><td class="cell"><p><b>Precision (%)</b></p><p>Top-10 Top-50</p></td><td class="cell"><p>Top-100</p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p><b>Homophone</b></p></td><td class="cell"><p>Same PinYin</p></td><td class="cell"><p>63.2</p></td><td class="cell"><p>73.7</p></td><td class="cell"><p>84.2</p></td><td class="cell"><p>84.2</p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p></p></td><td class="cell"><p>Similar PinYin</p></td><td class="cell"><p>40.0</p></td><td class="cell"><p>60.0</p></td><td class="cell"><p>70.0</p></td><td class="cell"><p>80.0</p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p></p></td><td class="cell"><p>Number</p></td><td class="cell"><p>81.1</p></td><td class="cell"><p>91.6</p></td><td class="cell"><p>95.8</p></td><td class="cell"><p>96.8</p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p><b>Abbreviation</b></p></td><td class="cell"><p>Chinese abbreviation</p></td><td class="cell"><p>11.8</p></td><td class="cell"><p>41.2</p></td><td class="cell"><p>52.9</p></td><td class="cell"><p>52.9</p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p><b>Acronym</b></p></td><td class="cell"><p>PinYin Acronym</p></td><td class="cell"><p>82.1</p></td><td class="cell"><p>94.6</p></td><td class="cell"><p>96.4</p></td><td class="cell"><p>96.4</p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p></p></td><td class="cell"><p>English Acronym</p></td><td class="cell"><p>21.9</p></td><td class="cell"><p>46.9</p></td><td class="cell"><p>56.3</p></td><td class="cell"><p>59.4</p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p><b>Transliteration</b></p></td><td class="cell"><p>20.0</p></td><td class="cell"><p>40.0</p></td><td class="cell"><p>50.0</p></td><td class="cell"><p>50.0</p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p><b>Average</b></p></td><td class="cell"><p></p></td><td class="cell"><p><b>61.8</b></p></td><td class="cell"><p><b>77.1</b></p></td><td class="cell"><p><b>83.1</b></p></td><td class="cell"><p><b>84.7</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><td class="cell"></td><td class="cell"></td><td class="cell"></td><td class="cell"></td></tr></table><doubt alpha="65.0" length="20" tooSmall="False" monospace="0.0">Size of seed set 130</doubt><p>Size of candidate set 3000</p><doubt alpha="65.0" length="20" tooSmall="False" monospace="0.0">Size of test set 750</doubt><doubt alpha="66.7" length="9" tooSmall="False" monospace="0.0">Recall30%</doubt><table caption="Table 10: Recall of Test Set on a Candidate Set Extracted"></table><p>by a <i>Bootstrapping </i>Procedure</p></section><section number="6" title="Related Work"><p>Automatically extracting the relations between full-form Chinese phrases and their abbreviations is an interesting and important task for many NLP appli­cations (e.g., machine translation, information re­trieval, etc.). Recently, Chang and Lai (2004), Lee (2005), Chang and Teng (2006), Li and Yarowsky (2008) have investigated this task. Specifically, Chang and Lai (2004) describes a hidden markov model (HMM) to model the relationship between a full-form phrase and its abbreviation, by treat­ing the abbreviation as the <i>observation </i>and the full-form words as <i>states </i>in the model. Using a set of manually-created <i>full-abbreviation </i>relations as training data, they report experimental results on a <i>recognition </i>task (i.e., given an abbreviation, the task is to obtain its full-form, or the vice versa). Chang and Teng (2006) extends the work in Chang and Lai (2004) to automatically extract the relations between full-form phrases and their abbreviations, where both the full-form phrase and its abbrevia­tion are not given. Clearly, the method in (Chang and Lai, 2004; Chang and Teng, 2006) is <i>super­vised </i>because it requires the <i>full-abbreviation </i>rela­tions as training data. Li and Yarowsky (2008) pro­pose an <i>unsupervised </i>method to extract the relations between full-form phrases and their abbreviations. They exploit the <i>data co-occurrence </i>phenomena in the newswire text, as we have done in this paper. Moreover, they augment and improve a statistical machine translation by incorporating the extracted relations into the baseline translation system.</p><p>Other interesting work that addresses a similar task as ours includes the work on homophones (e.g., Lee and Chen (1997)), abbreviations with their defi­nitions (e.g., Park and Byrd (2001)), abbreviations and acronyms in the medical domain (Pakhomov, 2002), and transliteration (e.g., (Knight and Graehl, 1998; Virga and Khudanpur, 2003; Li et al., 2004;</p><p>Wu and Chang, 2007)).</p><p>While all the above work deals with the rela­tions occurring within the <i>formal </i>text, we consider the <i>formal-informal </i>relations that occur across both formal and informal text, and we extract the rela­tions from the web corpora, instead from just formal text. Moreover, our method is <i>semi-supervised </i>in the sense that the weights of the feature functions are tuned in a <i>supervised </i>log-linear model using a small number of seed relations while the generation and ranking of the hypotheses are <i>unsupervised </i>by exploiting the <i>data co-occurrence </i>phenomena.</p></section><section number="7" title="Conclusions"><p>In this paper, we have first presented a taxonomy of the <i>formal-informal </i>relations occurring in Chinese text. We have then proposed a novel method for discovering and modeling the relationship between informal Chinese expressions (including colloqui­alisms and instant-messaging slang) and their formal equivalents. Specifically, we have proposed a boot­strapping procedure to identify a list of candidate informal phrases in web corpora. Given an infor­mal phrase, we retrieved contextual instances from the web using a search engine, generated hypothe­ses of formal equivalents via this data, and ranked the hypotheses using a conditional log-linear model. In the log-linear model, we incorporated as feature functions both rule-based intuitions and <i>data co­occurrence </i>phenomena (either as an explicit or in­direct definition, or through formal/informal usages occurring in free variation in a discourse). We tested our system on manually collected test examples, and found that the <i>(formal-informal) </i>relationship discovery and extraction process using our method achieves an average 1-best precision of 62%. Given the ubiquity of informal conversational style on the internet, this work has clear applications for text nor­malization in text-processing systems including ma­chine translation aspiring to broad coverage.</p></section><section title="Acknowledgments"><p>We would like to thank Yi Su, Sanjeev Khudanpur, and the anonymous reviewers for their helpful com­ments. This work was partially supported by the De­fense Advanced Research Projects Agency's GALE program via Contract No HR0011-06-2-0001.</p><doubt alpha="0.0" length="4" tooSmall="False" monospace="0.0">1039</doubt><page local="10" global="1040"/></section><references><p>S. J. Benson, L. C. McInnes, J. J. More, and J. Sarich.</p><p>2002. Tao users manual, Technical Report ANL/MCS-</p><p>TM-242-Revision 1.4, Argonne National Laboratory.</p><p>S. J. Benson and J. J. More. 2002. A limited memory vari­able metric method for bound constrained minimiza­tion. preprint ANL/ACSP909-0901, Argonne National Laboratory.</p><p>Jing-Shin Chang and Yu-Tso Lai. 2004. A preliminary study on probabilistic models for Chinese abbrevia­tions. <i>In Proceedings of the 3rd SIGHAN Workshop on Chinese Language Processing, </i>Barcelona, Spain (2004),pages 9-16.</p><p>Jing-Shin Chang and Wei-Lun Teng. 2006. Mining Atomic Chinese Abbreviation Pairs: A Probabilistic Model for Single Character Word Recovery. <i>In Pro­ceedings ofthe 5rd SIGHAN Workshop on Chinese Language Processing, </i>Sydney, Australia (2006), pages 17-24.</p><p>Kevin Knight and Jonathan Graehl. 1998. Machine Transliteration. <i>Computational Linguistics, </i>24(4):599612.</p><p>Philipp Koehn, Hieu Hoang, Alexandra Birch, Chris Callison-Burch, Marcello Federico, Nicola Bertoldi, Brooke Cowan,Wade Shen, Christine Moran, Richard Zens, Chris Dyer, Ondrej Bojar, Alexandra Con-strantin, and Evan Herbst. 2007. Moses: Open source toolkit for statistical machine translation. <i>In Proceed­ings ofACL, </i>Demonstration Session, pages 177-180.</p><p>H.W.D Lee. 2005. A study of automatic expansion of Chinese abbreviations. MA Thesis, The University of Hong Kong.</p><p>Yue-Shi Lee and Hsin-Hsi Chen. 1997. Applying Repair Processing in Chinese Homophone Disambiguation.</p><p><i>In Proceedings ofthe Fifth Conference on AppliedNat-ural Language Processing, </i>pages 57-63.</p><p>Haizhou Li, MinZhang, andJianSu. 2004. A jointsource channel model for machine transliteration. <i>In Proceed­ings of ACL 2004, </i>pages 159-166.</p><p>Zhifei Li and David Yarowsky. 2008. Unsupervised Translation Induction for Chinese Abbreviations using Monolingual Corpora. <i>In Proceedings ofACL 2008,</i> pages 425-433.</p><p>R. Malouf. 2002. A comparison of algorithms for maxi­mum entropy parameter estimation. <i>In Proceedings of</i> <i>CoNLL 2002, </i>pages 49-55.</p><p>Serguei Pakhomov. 2002. Semi-Supervised Maximum Entropy Based Approach to Acronym and Abbrevia­tion Normalization in Medical Texts. <i>In Proceedings of ACL 2002, </i>pages 160-167.</p><p>Youngja Park and Roy J. Byrd. 2001. Hybrid text min­ing for finding abbreviations and their definitions. <i>In Proceedings ofEMNLP 2001, </i>pages 126-133.</p><p>Brian Roark, Murat Saraclar, and Michael Collins. 2007. Discriminative n-gram language modeling. <i>Computer Speech and Language, </i>21(2):373-392.</p><p>Paola Virga and Sanjeev Khudanpur. 2003. Transliter­ation of Proper Names in Cross lingual Information Retrieval. <i>In Proceedings ofthe ACL 2003 Workshop on Multilingual and Mixed-language Named Entity Recognition.</i></p><p>Jian-Cheng Wu and Jason S. Chang. 2007. Learning to Find English to Chinese Transliterations on the Web.</p><p><i>In Proceedings ofEMNLP-CoNLL 2007, </i>pages 996­1004.</p><p>David Yarowsky. 1995. Unsupervised word sense disam­biguation rivaling supervised methods. <i>In Proceedings of ACL 1995, </i>pages 189-196.</p><p>Z.P. Yin. 1999. Methodologies and principles of Chi­nese abbreviation formation. In Language Teaching and Study, No.2 (1999) 73-82.</p><doubt alpha="0.0" length="4" tooSmall="False" monospace="0.0">1040</doubt></references></body></article>