<?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"/><title>Refinements in BTG-based Statistical Machine Translation</title><author surname="Xiong" givenname="Deyi"><org  name="Institute for Infocomm Research" country="Singapore"/></author><author surname="Zhang" givenname="Min"><org  name="Institute for Infocomm Research" country="Singapore"/></author><author surname="Aw" givenname="AiTi"><org  name="Institute for Infocomm Research" country="Singapore"/></author><author surname="Mi" givenname="Haitao"><org  name="Chinese Academy of Sciences" country="China" city="Beijing"/></author><author surname="Liu" givenname="Qun"><org  name="Chinese Academy of Sciences" country="China" city="Beijing"/></author><author surname="Lin" givenname="Shouxun"><org  name="Chinese Academy of Sciences" country="China" city="Beijing"/></author></firstpageheader><frontmatter><p><b>Reinements in BTG-based Statistical Machine Translation</b></p><p><b>Deyi Xiong, Min Zhang, Aiti </b><b>Aw</b><b>      Haitao Mi, Qun Liu and Shouxun Lin</b></p><p>Human Language Technology Key Lab of Intelligent Information Processing Institute for Infocomm Research Institute of Computing Technology</p><p>21 Heng Mui Keng Terrace Chinese Academy of Sciences</p><p>Singapore 119613 Beijing China, 100080</p><p>{dyxiong, mzhang,  aaiti}@i2r.a-star.edu.sg   {htmi,  liuqun, sxlin}@ict.ac.cn</p></frontmatter><abstract>Bracketing Transduction Grammar (BTG) has been well studied and used in statistical machine translation (SMT) with promising results. However, there are two major issues for BTG-based SMT. First, there is no effec­tive mechanism available for predicting or­ders between neighboring blocks in the orig­inal BTG. Second, the computational cost is high. In this paper, we introduce two re­finements for BTG-based SMT to achieve better reordering and higher-speed decod­ing, which include (1) reordering heuristics to prevent incorrect swapping and reduce search space, and (2) special phrases with tags to indicate sentence beginning and end­ing. The two refinements are integrated into a well-established BTG-based Chinese-to-English SMT system that is trained on large-scale parallel data. Experimental results on the NIST MT-05 task show that the proposed refinements contribute significant improve­ment of 2% in BLEU score over the baseline system. </abstract></header><body><section number="1" title="Introduction"><p>Bracket transduction grammar was proposed by Wu (1995) and firstly employed in statistical machine translation in (Wu, 1996). Because of its good trade­off between efficiency and expressiveness, BTG re­striction is widely used for reordering in SMT (Zens et al., 2004). However, BTG restriction does not provide a mechanism to predict final orders between two neighboring blocks.</p><p>To solve this problem, Xiong et al. (2006) proposed an enhanced BTG with a maximum en­tropy (MaxEnt) based reordering model (MEBTG). MEBTG uses boundary words of bilingual phrases as features to predict their orders. Xiong et al. (2006) reported significant performance im­provement on Chinese-English translation tasks in two different domains when compared with both Pharaoh (Koehn, 2004) and the original BTG us­ing flat reordering. However, error analysis of the translation output of Xiong et al. (2006) reveals that boundary words predict wrong swapping, espe­cially for long phrases although the MaxEnt-based reordering model shows better performance than baseline reordering models.</p><p>Another big problem with BTG-based SMT is the high computational cost. Huang et al. (2005) re­ported that the time complexity of BTG decoding with m-gram language model is <i>O(n3+4(m~<footnote anchor="1"/>"<footnote anchor="1"/></i>). If a 4-gram language model is used (common in many current SMT systems), the time complexity is as high as O(n<footnote anchor="15"/>). Therefore with this time complexity translating long sentences is time-consuming even with highly stringent pruning strategy.</p><p>To speed up BTG decoding, Huang et al. (2005) adapted the hook trick which changes the time complexity from <i>O(n3+4(m-1)) </i>to <i>O(n3+3(m~1)). </i>However, the implementation of the hook trick with pruning is quite complicated. Another method to in­crease decoding speed is cube pruning proposed by Chiang (2007) which reduces search space signifi­cantly.</p><p>In this paper, we propose two refinements to ad­dress the two issues, including (1) reordering heuristics to prevent incorrect swapping and reduce search space using swapping window and punctuation re­striction, and (2) phrases with special tags to indicate beginning and ending of sentence.<page local="2" global="506"/> Experimental re­sults show that both refinements improve the BLEU score significantly on large-scale data.</p><doubt alpha="100.0" length="3" tooSmall="False" monospace="0.0">SOS</doubt><p>The above refinements can be easily implemented and integrated into a baseline BTG-based SMT sys­tem. However, they are not specially designed for BTG-based SMT and can also be easily integrated into other systems with different underlying trans­lation strategies, such as the state-of-the-art phrase-based system (Koehn et al., 2007), syntax-based sys­tems (Chiang et al., 2005; Marcu et al., 2006; Liu et al., 2006).</p><p>The rest of the paper is organized as follows. In section 2, we review briefly the core elements of the baseline system. In section 3 we describe our proposed refinements in detail. Section 4 presents the evaluation results on Chinese-to-English trans­lation based on these refinements as well as results obtained in the NIST MT-06 evaluation exercise. Fi­nally, we conclude our work in section 5.</p></section><section number="2" title="The Baseline System"><p>In this paper, we use Xiong et al. (2006)'s sys­tem Bruin as our baseline system. Their system has three essential elements which are (1) a stochastic BTG, whose rules are weighted using different fea­tures in log-linear form, (2) a MaxEnt-based reorder­ing model with features automatically learned from bilingual training data, (3) a CKY-style decoder us­ing beam search similar to that of Wu (1996). We describe the first two components briefly below.</p><subsection number="2.1" title="Model"><p>The translation process is modeled using BTG rules which are listed as follows two rules (1) and (2) are used to merge two consec­utive blocks into a single larger block in a straight or inverted order.</p><doubt alpha="25.0" length="12" tooSmall="False" monospace="0.0">A — [A1, A2]</doubt><doubt alpha="27.3" length="11" tooSmall="False" monospace="0.0">A -(A1, A2)</doubt><doubt alpha="42.9" length="7" tooSmall="False" monospace="0.0">A — x/y</doubt><doubt alpha="0.0" length="3" tooSmall="False" monospace="0.0">(1)</doubt><doubt alpha="0.0" length="3" tooSmall="False" monospace="0.0">(2)</doubt><doubt alpha="0.0" length="3" tooSmall="False" monospace="0.0">(3)</doubt><p>To construct a stochastic BTG, we calculate rule probabilities using the log-linear model (Och and Ney, 2002). For the two merging rules <i>(1) </i>and <i>(2), </i>the assigned probability <i>Prm(A) </i>is defined as fol­lows where Q, the reordering score of block <i>A<footnote anchor="1"/> </i>and <i>A<footnote anchor="2"/>, </i>is calculated using the MaxEnt-based reordering model (Xiong et al., 2006) described in the next sec­tion, <b>Aq </b>is the weight of Q, and <i>^</i><i>P</i><i>LM</i><b>(</b>a<b>i</b>,a<b><footnote anchor="2"/>) </b>is the increment of language model score of the two blocks according to their final order, <i>A</i><i>LM </i>is its weight.</p><doubt alpha="54.5" length="11" tooSmall="False" monospace="0.0">Prm(A) =QA«</doubt><doubt alpha="100.0" length="4" tooSmall="False" monospace="0.0">AXLM</doubt><doubt alpha="45.5" length="11" tooSmall="False" monospace="0.0">^PLM(A1,A2)</doubt><doubt alpha="0.0" length="3" tooSmall="False" monospace="0.0">(4)</doubt><p>For the lexical rule (3), it is applied with a proba­bility <i>Prl (A)</i> where <i>p() </i>are the phrase translation probabilities in both directions, <i>plex( ) </i>are the lexical translation probabilities in both directions, <i>exp(1) </i>and <i>exp(\y\) </i>are the phrase penalty and word penalty, respec­tively and As are weights of features. These features are commonly used in the state-of-the-art systems (Koehn et al., 2005; Chiang et al., 2005).</p><doubt alpha="50.0" length="42" tooSmall="False" monospace="0.0">Prl(A)   = p(x\y)Xl• p(y\x)X2• Viex(x\y)Xi</doubt><doubt alpha="50.0" length="34" tooSmall="False" monospace="0.0">■Viex(y\x)Xi• exp(1)X5• exp(\y\)X6</doubt><doubt alpha="40.0" length="10" tooSmall="False" monospace="0.0">•pLm(y)(5)</doubt></subsection><subsection number="2.2" title="MaxEnt-based Reordering Model"><p>The MaxEnt-based reordering model is defined on two consecutive blocks <i>A<footnote anchor="1"/> </i>and <i>A<footnote anchor="2"/> </i>together with their order <i>o £ {straight, inverted} </i>according to the maximum entropy framework.</p><doubt alpha="45.5" length="11" tooSmall="False" monospace="0.0">pe (o\A\A2)</doubt><p>The lexical rule <i>(3) </i>is used to translate source phrase <i>x </i>into target phrase <i>y </i>and generate a block <i>A. </i>The where the functions <i>h</i><i>i £ </i><i>{0,1} </i>are model features and <i>0i </i>are weights of the model features trained au­tomatically (Malouf, 2002).</p><doubt alpha="0.0" length="3" tooSmall="False" monospace="0.0">(6)</doubt><p>There are three steps to train a MaxEnt-based re­ordering model. First, we need to extract reordering examples from unannotated bilingual data, then gen­erate features from these examples and finally esti­mate feature weights.</p><doubt alpha="57.1" length="21" tooSmall="False" monospace="0.0">exp(J2iOjhj(o,A1,A2))</doubt><doubt alpha="59.1" length="22" tooSmall="False" monospace="0.0">Eoexp(Ei0ihi(o,A1,A2))</doubt><page local="3" global="507"/><p>For extracting reordering examples, there are two points worth mentioning:</p><p>1. In the extraction of useful reordering examples, there is no length limitation over blocks com­pared with extracting bilingual phrases.</p><p>2. When enumerating all combinations of neigh­boring blocks, a good way to keep the number of reordering examples acceptable is to extract smallest blocks with the <i>straight</i><i> </i>order while largest blocks with the <i>inverted</i><i> </i>order .</p></subsection></section><section number="3" title="Refinements"><p>In this section we describe two refinements men­tioned above in detail. First, we present fine­grained reordering heuristics using swapping win­dow and punctuation restriction. Secondly, we inte­grate special bilingual phrases with sentence begin­ning/ending tags.</p><subsection number="3.1" title="Reordering Heuristics"><p>We conduct error analysis of the translation out­put of the baseline system and observe that Bruin sometimes incorrectly swaps two large neighboring blocks on the target side. This happens frequently when inverted order successfully challenges straight order by the incorrect but strong support from the language model and the MaxEnt-based reordering model. The reason is that only boundary words are used as evidences by both language model and MaxEnt-based reordering model when the decoder selects which merging rule (straight or inverted) to be used <footnote anchor="1"/>. However, statistics show that bound­ary words are not reliable for predicting the right order between two larger neighboring blocks. Al-Onaizan and Papineni (2006) also proved that lan­guage model is insufficient to address long-distance word reordering. If a wrong inverted order is se­lected for two large consecutive blocks, incorrect long-distance swapping happens.</p><p>Yet another finding is that many incorrect swap-pings are related to punctuation marks. First, the source sequence within a pair of balanced punctua­tion marks (quotes and parentheses) should be kept</p><p>'In (Xiong et al., 2006), the language model uses the left­most/rightmost words on the target side as evidences while the MaxEnt-based reordering model uses the boundary words on both sides.</p><doubt alpha="50.0" length="26" tooSmall="False" monospace="0.0">Chinese:% $:   r&amp; &amp;^ftj ft</doubt><doubt alpha="71.4" length="7" tooSmall="False" monospace="0.0">tête-j_</doubt><p><b>Bruin: </b>urgent action , he said : "This is a very serious situation , we can only hope that there will be a possibility ." <b>Bruin+RH:</b><b> </b>he said : "This is a very serious sit­uation , we can only hope that there will be the possibility to expedite action ." <b>Ref: </b>He said: "This is a very serious situa­tion. We can only hope that it is possible to speed up the operation."</p><p>Figure 1: An example of incorrect long-distance swap. The underlined Chinese words are incorrectly swapped to the beginning of the sentence by the original Bruin. RH means reordering heuristics.</p><p>within the punctuation after translation. However, it is not always true when reordering is involved. Sometime the punctuation marks are distorted with the enclosed words sequences being moved out. Secondly, it is found that a series of words is fre­quently reordered from one side of a structural mark, such as commas, semi-colons and colons, to the other side of the mark for long sentences contain­ing such marks. Generally speaking, on Chinese-to-English translation, source words are translated monotonously relative to their adjacent punctuation marks, which means their order relative to punctua­tion marks will not be changed. In summary, punctu­ation marks place a strong constraint on word order around them.</p><p>For example, in Figure 1, Chinese words <b>"J</b><b>P</b><b>'tfe tT^t</b><b>l" </b>are reordered to sentence beginning. That is an incorrect long-distance swapping, which makes the reordered words moved out from the balanced punctuation marks " <b>T"</b><b> </b>and <b>"J</b><b> </b>", and incorrectly precede their previous mark "<b>, </b>".</p><p>These incorrect swappings definitely jeopardize the quality of translation. Here we propose two straightforward but effective heuristics to control and adjust the reordering, namely swapping window and punctuation restriction.</p><p><b>Swapping Window (SW): </b>It constrains block swapping in the following way where <i>\Ais\ </i>denotes the number of words on the source side <i>A</i><i>S </i>of block Ai, <i>sws </i>is a pre-defined swapping window size. Any inverted reordering be­yond the pre-defined swapping window size is pro­hibited.<page local="4" global="508"/></p><doubt alpha="48.7" length="39" tooSmall="False" monospace="0.0">ACTIVATEA — (A1,A2)IF\A]\ + \A2S\ &lt; sws</doubt><p><b>Punctuation Restriction (PR): </b>If two neighbor­ing blocks include any of the punctuation marks <i>p </i><i>£ {</i><b>, </b><b><i>,</i></b><b><i> </i></b><b><i>:</i></b><b><i> </i></b><b>; r J </b><b><i>())</i></b><b><i> </i></b>( <b>) " " </b>}, the two blocks will be merged with straight order.</p><p>Punctuation marks were already used in pars­ing (Christine Doran, 2000) and statistical machine translation (Och et al., 2003). In (Och et al., 2003), three kinds of features are defined, all re­lated to punctuation marks like quotes, parentheses and commas. Unfortunately, no statistically signifi­cant improvement on the BLEU score was reported in (Och et al., 2003). In this paper, we consider this problem from a different perspective. We em­phasize that words around punctuation marks are reordered ungrammatically and therefore we posi­tively use punctuation marks as a hard decision to restrict such reordering around punctuations. This is straightforward but yet results in significant im­provement on translation quality.</p><p>The two heuristics described above can be used together. If the following conditions are satisfied, we can activate the inverted rule:</p><doubt alpha="34.3" length="35" tooSmall="False" monospace="0.0">\A]\ + \A2S\ &lt;sws &amp;&amp; Pf](A]JA2S) =0</doubt><p>where <i>P </i>is the set of punctuation marks mentioned above.</p><p>The two heuristics can also speed up decoding be­cause decoding will be monotone within those spans which are not in accordance with both heuristics. For a sentence with <i>n </i>words, the total number of spans is <i>O(n<footnote anchor="2"/>). </i>If we set <i>sws = m (m &lt; </i>n), then the number of spans with monotone search is <i>O((n </i><i>— </i><i>m)<footnote anchor="2"/>). </i>With punctuation restriction, the non­monotone search space will reduce further.</p></subsection><subsection number="3.2" title="Phrases with Sentence Beginning/Ending"><doubt alpha="100.0" length="4" tooSmall="False" monospace="0.0">Tags</doubt><p>We observe that in a sentence some phrases are more likely to be located at the beginning, while other phrases are more likely to be at the end. This kind of location information with regard to the phrase posi­tion could be used for reordering. A straightforward way to use this information is to mark the begin­ning and ending of word-aligned sentences with (s) and <b><i>( </i></b>/s) respectively. This idea is borrowed from language modeling (Stolcke, 2002). The corre­sponding tags at the source and target sentences are aligned to each other, i.e, the beginning tag of source sentences is aligned to the beginning tag of target sentences, similarly for the ending tag. Figure 2 shows a word-aligned sentence pair annotated with the sentence beginning and ending tag.</p><p>During training, the sentence beginning and end­ing tags ((s) and (/s)) are treated as words. There­fore the phrase extraction and MaxEnt-based re­ordering training algorithm need not to be modified. Phrases with the sentence beginning/ending tag will be extracted and MaxEnt-based reordering features with such tags will also be generated. For example, from the word-aligned sentence pair in Figure 2, we can extract tagged phrases like and generate MaxEnt-based reordering features with tags like <i>i</i><i> </i><i>0,</i><i> </i><i>otherwise</i> where <i>b<footnote anchor="1"/>,b<footnote anchor="2"/> </i>are blocks, <i>t1 </i>denotes the last source word, <i>o </i><i>= </i><i>s </i>means the order between two blocks is straight. To avoid wrong alignments, we remove tagged phrases where only the beginning/ending tag is extracted on either side of the phrases, such as</p><doubt alpha="47.1" length="17" tooSmall="False" monospace="0.0">(s)\ \\(s)Tibet's</doubt><doubt alpha="60.0" length="30" tooSmall="False" monospace="0.0">J&amp;ifi(/s)\ \\achievements (/s)</doubt><doubt alpha="28.6" length="21" tooSmall="False" monospace="0.0">h(ob1b2)={1=(/s),(°=s</doubt><doubt alpha="47.1" length="17" tooSmall="False" monospace="0.0">(s)\ \\ (s)Thoseo</doubt><doubt alpha="22.2" length="9" tooSmall="False" monospace="0.0">(/s) (/s)</doubt><p>During decoding, we first annotate source sen­tences with the beiginning/ending tags, then trans­late them as what Bruin does. Note that phrases with sentence beginning/ending tags will be used in the same way as ordinary phrases without such tags during decoding. With the additional support oflan-guage model and MaxEnt-based reordering model, we observe that phrases with such tags are always moved to the beginning or ending of sentences cor­rectly.</p><page local="5" global="509"/></subsection></section><section number="4" title="Evaluation"><p>In this section, we report the performance of the en­hanced Bruin on the NIST MT-05 and NIST MT-06 Chinese-to-English translation tasks. We describe the corpus, model training, and experiments related to the refinements described above.</p><subsection number="4.1" title="Corpus"><p>The bilingual training data is derived from the fol­lowing various sources: the FBIS (LDC2003E14), Hong Kong Parallel Text (Hong Kong News and Hong Kong Hansards, LDC2004T08), Xinhua News (LDC2002E18), Chinese News Translation Text</p><p>Part1 (LDC2005T06), Translations from the Chi­nese Treebank (LDC2003E07), Chinese English sentence pairs in total (68.1M Chinese words and 73.8M English words).</p><doubt alpha="61.4" length="44" tooSmall="False" monospace="0.0">News Magazine (LDC2005E47). It contains 2.4M</doubt><p>For the efficiency of minimum-error-rate training, we built our development set using sentences not ex­ceeding 50 characters from the NIST MT-02 evalu­ation test data (580 sentences).</p></subsection><subsection number="4.2" title="Training"><p>We use exactly the same way and configuration de­scribed in (He et al., 2006) to preprocess the training data, align words and extract phrases.</p><p>We built two four-gram language models using Xinhua section of the English Gigaword corpus (181.1M words) and the English side of the bilin­gual training data described above respectively. We applied modified Kneser-Ney smoothing as imple­mented in the SRILM toolkit (Stolcke, 2002).</p><p>The MaxEnt-based reordering model is trained using the way of (Xiong et al., 2006). The difference is that we only use lexical features generated by tail words of blocks, instead of head words, removing features generated by the combination of two bound­ary words.</p><p>Table 1: Effect of reordering heuristics. <i>RH<footnote anchor="1"/> </i>de­notes swapping window while <i>RH2<footnote anchor="1"/></i><i> </i>denotes swap­ping window with the addition of punctuation re­striction.</p></subsection><subsection number="4.3" title="Translation Results"><p>Table 1 compares the BLEU scores <footnote anchor="2"/> and the speed in seconds/sentence of the baseline system Bruin and the enhanced system with reordering heuristics applied. The second row gives the BLEU score and the average decoding time of Bruin. The rows be­low row 3 show the BLEU scores and speed of the enhanced Bruin with different combinations of re­ordering heuristics. We can clearly see that the re­ordering heuristics proposed by us have a two-fold effect on the performance: improving the BLEU score and decreasing the average decoding time. The example in Figure 1 shows how reordering heuristics prevent incorrect long-distance swapping which is not in accordance with the punctuation re­striction.</p><p>Table 1 also shows that a 15-word swapping win­dow is an inflexion point with the best tradeoff be­tween the decoding time and the BLEU score. We speculate that in our corpus most reorderings hap­pen within a 15-word window. We use the FBIS corpus to testify this hypothesis. In this corpus, we extract all reordering examples using the algorithm of Xiong et al. (2006). Figure 3 shows the reorder­ing length distribution curve in this corpus. Accord-<page local="6" global="510"/></p><footnote label="2">In this paper, all BLEU scores are case-sensitive and evalu­ated on the NIST MT-05 Chinese-to-English translation task if there is no special note.</footnote><table caption="Figure 2: A word-aligned sentence pair annotated with the sentence beginning and ending tag." 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><td class="cell"></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p>(s)</p></td><td class="cell"><p></p></td><td class="cell"><p></p></td><td class="cell"><p></p></td><td class="cell"><p></p></td><td class="cell"><p><b>Mi</b></p></td><td class="cell"><p></p></td><td class="cell"><p>(/s)</p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p>(s)</p></td><td class="cell"><p>Tibet's</p></td><td class="cell"><p>financial</p></td><td class="cell"><p>work</p></td><td class="cell"><p>has gained</p></td><td class="cell"><p>remarkable</p></td><td class="cell"><p>achievements</p></td><td class="cell"><p>(/s)</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><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></tr><tr class="row"><td class="cell"></td><td class="cell"><p></p></td><td class="cell"><p>Bleu(%)</p></td><td class="cell"><p>Secs/sent</p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p>Bruin</p></td><td class="cell"><p>29.96</p></td><td class="cell"><p>54.3</p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p><i>sws</i></p></td><td class="cell"><p><i>RH</i><footnote anchor="1"/></p></td><td class="cell"><p></p></td><td class="cell"><p><i>RH</i><footnote anchor="1"/></p></td><td class="cell"><p>RH2<footnote anchor="1"/></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>29.65</p></td><td class="cell"><p>29.95</p></td><td class="cell"><p>42.6</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>10</p></td><td class="cell"><p>30.55</p></td><td class="cell"><p>31.27</p></td><td class="cell"><p>46.2</p></td><td class="cell"><p>41.8</p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p>15</p></td><td class="cell"><p>30.26</p></td><td class="cell"><p>31.40</p></td><td class="cell"><p>48.0</p></td><td class="cell"><p>42.2</p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p>20</p></td><td class="cell"><p>30.19</p></td><td class="cell"><p>31.42</p></td><td class="cell"><p>49.1</p></td><td class="cell"><p>43.2</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></tr></table><doubt alpha="0.0" length="1" tooSmall="True" monospace="0.0">5</doubt><doubt alpha="0.0" length="2" tooSmall="True" monospace="0.0">10</doubt><doubt alpha="0.0" length="2" tooSmall="True" monospace="0.0">15</doubt><doubt alpha="0.0" length="2" tooSmall="True" monospace="0.0">20</doubt><p>Reordering Length</p><p>Figure 3: Reordering length distribution. The hor­izontal axis (reordering length) indicates the num­ber of words on the source side of two neighboring blocks which are to be swapped. The vertical axis represents what proportion of reorderings with a cer­tain length is likely to be in all reordering examples with an inverted order.</p><p>ing to our statistics, reorderings within a window not exceeding 15 words have a very high proportion, 97.29%. Therefore we set <i>sws = 15 </i>for later exper­iments.</p><p>Table 2 shows the effect of integrating special phrases with sentence beginning/ending tags into Bruin. As special phrases accounts for only 1.95% of the total phrases used, an improvement of 0.6% in BLEU score is well worthwhile. Further, the im­provement is statistically significant at the 99% con­fidence level according to Zhang's significant tester (Zhang et al., 2004). Figure 4 shows several exam­ples translated with special phrases integrated. We can see that phrases with sentence beginning/ending tags are correctly selected and located at the right place.</p><table caption="Table 3 shows the performance of two systems on"></table><p>the NIST MT-05 Chinese test data, which are (1)</p><p>Table 3: Results of different systems. The refine­ments RH, SP represent reordering heuristics and special phrases with the sentence beginning/ending tag, respectively.</p><p>Bruin, trained on the large data described above; and (2) enhanced Bruin (EBruin) with different refine­ments trained on the same data set. This table also shows the evaluation result of the enhanced Bruin with reordering heuristics, obtained in the NIST MT-06 evaluation exercise. <footnote anchor="3"/></p></subsection></section><section number="5" title="Conclusions"><p>We have described in detail two refinements for BTG-based SMT which include reordering heuris­tics and special phrases with tags. The refinements were integrated into a well-established BTG-based system Bruin introduced by Xiong et al. (2006). Re­ordering heuristics proposed here achieve a twofold improvement: better reordering and higher-speed decoding. To our best knowledge, we are the first to integrate special phrases with the sentence be­ginning/ending tag into SMT. Experimental results show that the above refinements improve the base­line system significantly.</p><p>For further improvements, we will investigate possible extensions to the BTG grammars, e.g. learning useful nonterminals using unsupervised learning algorithm.</p></section><section title="Acknowledgements"><p>We would like to thank the anonymous review­ers for useful comments on the earlier version of this paper. The first author was partially sup­ported by the National Science Foundations of China (No. 60573188) and the High Technology Research and Development Program of China (No. 2006AA010108) while he studied in the Institute of Computing Technology, Chinese Academy of Sci­ences.</p><footnote label="3">Full results are available at http://www.nist.gov/ speech/tests/mt/doc/mtG6eval_official _results.html.</footnote><doubt alpha="0.0" length="2" tooSmall="True" monospace="0.0">25</doubt><doubt alpha="100.0" length="1" tooSmall="True" monospace="0.0">c</doubt><doubt alpha="75.0" length="4" tooSmall="True" monospace="0.0">CD U</doubt><doubt alpha="100.0" length="2" tooSmall="True" monospace="0.0">CD</doubt><doubt alpha="0.0" length="1" tooSmall="True" monospace="0.0">0</doubt><doubt alpha="0.0" length="2" tooSmall="True" monospace="0.0">30</doubt><doubt alpha="0.0" length="2" tooSmall="True" monospace="0.0">40</doubt><doubt alpha="0.0" length="2" tooSmall="True" monospace="0.0">50</doubt><doubt alpha="0.0" length="2" tooSmall="True" monospace="0.0">60</doubt><doubt alpha="0.0" length="2" tooSmall="True" monospace="0.0">70</doubt><doubt alpha="0.0" length="2" tooSmall="True" monospace="0.0">80</doubt><doubt alpha="100.0" length="3" tooSmall="False" monospace="0.0">SlO</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>System</p></td><td class="cell"><p>Refine</p></td><td class="cell"><p>MT-G5</p></td><td class="cell"><p>MT-G6</p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p>Bruin</p></td><td class="cell"><p>-</p></td><td class="cell"><p>29.96</p></td><td class="cell"><p>-</p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p>EBruin</p></td><td class="cell"><p>RH</p></td><td class="cell"><p>31.4G</p></td><td class="cell"><p>3G.22</p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p>EBruin</p></td><td class="cell"><p>RH+SP</p></td><td class="cell"><p>32.G1</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 caption="Table 2: Effect of integrating special phrases with the sentence beginning/ending tag." 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></p></td><td class="cell"><p>Bleu(%)</p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p>Without Special Phrases</p></td><td class="cell"><p>31.4G</p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p>With Special Phrases</p></td><td class="cell"><p>32.G1</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="7" global="511"/><p><b>With Special Phrases </b>I <b>Without Special Phrases</b></p><p>Figure 4: Examples translated with special phrases integrated. The bold underlined words are special phrases with the sentence beginning/ending tag.</p><table class="main" frame="box" rules="all" border="1" regular="True"><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>(s) Japan had </b>already pledged to provide 30 mil­lion US dollars of aid due to the tsunami victims of the country . ( /s)</p></td><td class="cell"><p><b>originally has </b>pledged to provide 30 million US dollars of aid from Japan tsunami victimized coun­tries .</p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p><b>(s) the results of the survey is based on </b>the re­sults of the chiefs of the Ukrainian National 50.96% cast by chiefs . (/s)</p></td><td class="cell"><p><b>is based on </b>the survey findings Ukraine 50.96% cast by the chiefs of the chiefs of the country .</p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p>(s) and at the same time, the focus of the world have been transferred to <b>other areas . (/s)</b></p></td><td class="cell"><p>and at the same time , <b>the global focus has shifted he.</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></section><references><p>Yaser Al-Onaizan, Kishore Papineni. 2006. Distortion Models for Statistical Machine Translation. In <i>Pro­ceedings ofACL-COLING 2006.</i></p><p>David Chiang, Adam Lopez, Nitin Madnani, Christof Monz, Philip Resnik, Michael Subotin. 2005. The Hiero Machine Translation System: Extensions, Eval­uation, and Analysis. In <i>Proceedings ofHLT/EMNLP, </i>pages 779 - 786, Vancouver, October 2005.</p><p>David Chiang. 2007. Hierarchical Phrase-based Transla­tion. In <i>computational linguistics, </i>33(2).</p><p>Christine Doran. 2000. Punctuation in a Lexicalized Grammar. In <i>Proceedings of Workshop TAG+5, </i>Paris.</p><p>Zhongjun He, Yang Liu, Deyi Xiong, Hongxu Hou, Qun</p><doubt alpha="63.0" length="46" tooSmall="False" monospace="0.0">Liu. 2006. ICT System Description for the 2006</doubt><p>TC-STAR Run #2 SLT Evaluation. In <i>Proceedings of TC-STAR Workshop on Speech-to-Speech Translation, </i>Barcelona, Spain.</p><p>Liang Huang, Hao Zhang and Daniel Gildea. 2005. Ma­chine Translation as Lexicalized Parsing with Hooks. In <i>Proceedings of the 9th International Workshop on Parsing Technologies (IWPT-05), </i>Vancouver, BC,</p><p>Canada, October 2005.</p><p>Philipp Koehn. 2004. Pharaoh: a beam search decoder for phrase-based statistical machine translation mod­els. In <i>Proceedings of the Sixth Conference of the Association for Machine Translation in the Americas, </i>pages 115 - 124.</p><p>Philipp Koehn, Amittai Axelrod, Alexandra Birch Mayne, Chris Callison-Burch, Miles Osborne and David Talbot. 2005. Edinburgh System Description for the 2005IWSLT Speech Translation Evaluation. In <i>International Workshop on Spoken Language Transla­tion.</i><i></i></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 Con­stantin, Evan Herbst. 2007. Moses: Open Source Toolkit for Statistical Machine Translation. ACL 2007, demonstration session, Prague, Czech Republic,</p><p>June 2007.</p><p>Yang Liu, Qun Liu, Shouxun Lin. 2006. Tree-to-String Alignment Template for Statistical Machine Transla­tion. In <i>Proceedings ofACL-COLING 2006.</i></p><p>Robert Malouf. 2002. A comparison of algorithms for maximum entropy parameter estimation. In <i>Proceed­ings ofCoNLL-2002.</i></p><p>Daniel Marcu, Wei Wang, Abdessamad Echihabi, and Kevin Knight. 2006. SPMT: Statistical Ma­chine Translation with Syntactified Target Language Phraases. In <i>Proceedings of EMNLP.</i></p><p>Franz Josef Och and Hermann Ney. 2002. Discrimina­tive training and maximum entropy models for statisti­cal machine translation. In <i>Proceedings ofACL 2002, </i>pages 295 - 302.</p><p>Franz Josef Och, Daniel Gildea, Sanjeev Khudanpur, Anoop Sarkar, Kenji Yamada, Alex Fraser, Shankar Kumar, Libin Shen, David Smith, Katherine Eng, Viren Jain, Zhen Jin, Dragomir Radev. 2003. Final Report of Johns Hopkins 2003 Summer Workshop on Syntax for Statistical Machine Translation.</p><p>Andreas Stolcke. 2002. SRILM - an extensible lan­guage modeling toolkit. In <i>Proceedings of Interna­tional Conference on Spoken Language Processing, </i>volume 2, pages 901-904.</p><p>Dekai Wu. 1995. Stochastic inversion transduction grammars, with application to segmentation, bracket­ing, and alignment of parallel corpora. In <i>Proceedings oflJCAL 1995, </i>pages 1328-1334, Montreal, August.</p><page local="8" global="512"/><p>Dekai Wu. 1996. A Polynomial-Time Algorithm for Sta­tistical Machine Translation. In <i>Proceedings ofACL 1996.</i></p><p>Deyi Xiong, Qun Liu and Shouxun Lin. 2006. Maxi­mum Entropy Based Phrase Reordering Model for Sta­tistical Machine Translation. In <i>Proceedings ofACL-COLING 2006, </i>pages 521 - 528.</p><p>R. Zens, H. Ney, T. Watanabe, and E. Sumita. 2004. Re­ordering Constraints for Phrase-Based Statistical Ma­chine Translation. In <i>Proceedings of CoLing 2004, </i>Geneva, Switzerland, pp. 205-211.</p><p>Ying Zhang, Stephan Vogel, and Alex Waibel. 2004. In­terpreting BLEU/NIST scores: How much improve­ment do we need to have a better system? In <i>Proceed­ings ofLREC 2004, </i>pages 2051 - 2054.</p></references></body></article>