| { | |
| "paper_id": "F14-2003", | |
| "header": { | |
| "generated_with": "S2ORC 1.0.0", | |
| "date_generated": "2023-01-19T10:22:41.122609Z" | |
| }, | |
| "title": "On-going Cooperative Research towards Developing Economy-Oriented Chinese-French SMT Systems with a New SMT Framework", | |
| "authors": [ | |
| { | |
| "first": "Yidong", | |
| "middle": [], | |
| "last": "Chen", | |
| "suffix": "", | |
| "affiliation": {}, | |
| "email": "[email protected]" | |
| }, | |
| { | |
| "first": "Lingxiao", | |
| "middle": [], | |
| "last": "Wang", | |
| "suffix": "", | |
| "affiliation": { | |
| "laboratory": "Laboratoire d'Informatique Grenoble (LIG)", | |
| "institution": "Universit\u00e9 Joseph Fourier", | |
| "location": { | |
| "settlement": "Grenoble", | |
| "country": "France" | |
| } | |
| }, | |
| "email": "" | |
| }, | |
| { | |
| "first": "Christian", | |
| "middle": [], | |
| "last": "Boitet", | |
| "suffix": "", | |
| "affiliation": { | |
| "laboratory": "Laboratoire d'Informatique Grenoble (LIG)", | |
| "institution": "Universit\u00e9 Joseph Fourier", | |
| "location": { | |
| "settlement": "Grenoble", | |
| "country": "France" | |
| } | |
| }, | |
| "email": "" | |
| }, | |
| { | |
| "first": "Xiaodong", | |
| "middle": [], | |
| "last": "Shi", | |
| "suffix": "", | |
| "affiliation": {}, | |
| "email": "" | |
| } | |
| ], | |
| "year": "", | |
| "venue": null, | |
| "identifiers": {}, | |
| "abstract": "R\u00e9 sum\u00e9. Nous pr\u00e9 sentons un projet collaboratif en cours men\u00e9 par l'universit\u00e9 de Grenoble et l'universit\u00e9 de Xiamen, et visant \u00e0 cr\u00e9 er des instances d'un nouveau type de syst\u00e8 me de traduction automatique statistique utilisant des ressources lexico-s\u00e9 mantiques et discursives. Le but concret est de d\u00e9 velopper des syst\u00e8 mes de TAS chinois-fran\u00e7 ais pour des sites boursiers et \u00e9 conomiques. Comme tr\u00e8 s peu de corpus et de dictionnaires bilingues chinois-fran\u00e7 ais sont disponibles sur Internet, l'anglais est utilis\u00e9 comme \"pivot\" pour construire les \u00e9 quivalents chinois-fran\u00e7 ais par transitivit\u00e9. Outre la description g\u00e9 n\u00e9 rale de ce projet, nous d\u00e9 crivons les progr\u00e8 s sur deux axes de recherche li\u00e9 s \u00e0 ce projet. Pour cela, nous utilisons une m\u00e9 thode, propos\u00e9 e par XMU, d'induction de probabilit\u00e9 fond\u00e9 e sur la similarit\u00e9 th\u00e9 matique, qui produit des tables de traduction C-F \u00e0 partir de tables de traduction C-E et E-F. Pour disposer de bons corpus parall\u00e8 les C-F, nous utilisons un syst\u00e8 me Web de post-\u00e9 dition collaborative qui peut d\u00e9 clencher l'am\u00e9 lioration incr\u00e9 mentale du syst\u00e8 me de TA en utilisant des m\u00e9 triques d'\u00e9 valuation de TA et en extrayant la \"meilleure partie\" de la m\u00e9 moire de traductions courante.", | |
| "pdf_parse": { | |
| "paper_id": "F14-2003", | |
| "_pdf_hash": "", | |
| "abstract": [ | |
| { | |
| "text": "R\u00e9 sum\u00e9. Nous pr\u00e9 sentons un projet collaboratif en cours men\u00e9 par l'universit\u00e9 de Grenoble et l'universit\u00e9 de Xiamen, et visant \u00e0 cr\u00e9 er des instances d'un nouveau type de syst\u00e8 me de traduction automatique statistique utilisant des ressources lexico-s\u00e9 mantiques et discursives. Le but concret est de d\u00e9 velopper des syst\u00e8 mes de TAS chinois-fran\u00e7 ais pour des sites boursiers et \u00e9 conomiques. Comme tr\u00e8 s peu de corpus et de dictionnaires bilingues chinois-fran\u00e7 ais sont disponibles sur Internet, l'anglais est utilis\u00e9 comme \"pivot\" pour construire les \u00e9 quivalents chinois-fran\u00e7 ais par transitivit\u00e9. Outre la description g\u00e9 n\u00e9 rale de ce projet, nous d\u00e9 crivons les progr\u00e8 s sur deux axes de recherche li\u00e9 s \u00e0 ce projet. Pour cela, nous utilisons une m\u00e9 thode, propos\u00e9 e par XMU, d'induction de probabilit\u00e9 fond\u00e9 e sur la similarit\u00e9 th\u00e9 matique, qui produit des tables de traduction C-F \u00e0 partir de tables de traduction C-E et E-F. Pour disposer de bons corpus parall\u00e8 les C-F, nous utilisons un syst\u00e8 me Web de post-\u00e9 dition collaborative qui peut d\u00e9 clencher l'am\u00e9 lioration incr\u00e9 mentale du syst\u00e8 me de TA en utilisant des m\u00e9 triques d'\u00e9 valuation de TA et en extrayant la \"meilleure partie\" de la m\u00e9 moire de traductions courante.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Abstract", | |
| "sec_num": null | |
| } | |
| ], | |
| "body_text": [ | |
| { | |
| "text": "The desire and need for cross-cultural communication between China and Europe, both officially and nongovernmentally, are on the increase. Especially, France is an important strategic partner of China and has deep relationships with China in many fields such as economy, science and technology, culture and education, etc. But the language barrier between Chinese and French is impassable in most situations.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Introduction", | |
| "sec_num": "1" | |
| }, | |
| { | |
| "text": "Concerning the economy, the cooperation between China and France grew rapidly in the past decades. France is now China's fourth largest trade partner in the EU, behind Germany, the Netherlands and the UK. According to data from the National Bureau of Statistics of China, bilateral trade between China and France increased from $13.4 billion in 2003 to $51 billion in 2012. Two-way direct investments are also on the rise. With the continuous improvement of China's investment environment, more and more French investors intend to invest in China. However, most portals of China, such as the website of Shanghai Stock Exchange and the website of Shenzhen Stock Exchange, only provide a Chinese version and an English version, but no French version. Moreover, Chinese investors are also more and more on the look for opportunities in France and other francophone areas, notably in Africa.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Introduction", | |
| "sec_num": "1" | |
| }, | |
| { | |
| "text": "Although French-Chinese bilingual applications are urgently needed in many situations, not much work has been done yet on French-Chinese MT, and French-Chinese is still an under-resourced language pair as there are no large good YIDONG CHEN, LINGXIAO WANG, CHRISTIAN BOITET ET XIAODONG SHI quality freely available bilingual lexico-semantic resources, and no sufficiently good MT systems. Most French-Chinese MT systems, in particular GT (GoogleTranslate 1 ) use English text as a pivot, which introduces more errors, often degrading translation quality to uselessness, as illustrated below. Results may be useful for guessing the topic of a text, but are otherwise bad or misleading, while investors need precise translations to decide whether to invest and where. Here are two examples, both selected from the announcements of the Shanghai Stock Exchange. We show the result of GT: output of GT-zh-en and then of en-fr (calling GT on zh-fr gives the same outputs) and en-PE-fr (post-edited en).", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Introduction", | |
| "sec_num": "1" | |
| }, | |
| { | |
| "text": "Ex1 (Source): \u6062\u590d\u4ea4\u6613\u540e\uff0c\u5982\u8be5\u8bc1\u5238\u4ea4\u6613\u4e2d\u518d\u6b21\u51fa\u73b0\u5f02\u5e38\u60c5\u51b5\uff0c\u672c\u6240\u53ef\u5b9e\u65bd\u7b2c\u4e8c\u6b21\u505c\u724c\uff0c\u505c\u724c\u65f6\u95f4\u6301\u7eed\u81f3\u4eca\u65e5 \u6536\u76d8\u524d\u4e94\u5206\u949f\u3002 (GT-zh-en) After the resumption of trading in the securities trading as abnormal situation occurs again, this can be implemented by the second suspension, suspension lasted until today the closing five minutes. (GT-zh-en-PE) If the trading of this security appears to be abnormal again after resumption of its trading, we may perform a second suspension upon it and this suspension will last until five minutes before today's closing. (GT-zh-en-fr = GT-zh-fr): Apr\u00e8 s la reprise de la n\u00e9 gociation dans le n\u00e9 goce de titres comme situation anormale se produit de nouveau, ce qui peut \u00eatre mis en oeuvre par la deuxi\u00e8 me suspension, la suspension a dur\u00e9 jusqu'\u00e0 aujourd'hui les cinq derni\u00e8 res minutes. (GT-zh-en-PE-fr): Si la n\u00e9 gociation de ce titre semble \u00ea tre anormal \u00e0 nouveau apr\u00e8 s la reprise de ses op\u00e9 rations, nous pouvons effectuer une deuxi\u00e8 me suspension sur elle et cette suspension durera jusqu'\u00e0 cinq minutes avant la cl\u00f4ture d'aujourd'hui. Examples above show that GT uses English as a pivot for the zh-fr direction. However, it is not the case for the fr-zh direction: outputs are different. Google Translate is claimed to be one of the \"state-of-the-art\" MT system. The truth is however that \"state-of-the-art\" for general MT systems means \"very bad to useless\": not knowing Chinese, one is at a loss to guess the exact meaning. Also, even if an output looks good (due to the use of a good target language model), its reliability may be very low, especially when negation appears or disappears at random (ex. 2 in French), or, as in these examples, when essential words are badly translated (\"If\" \u2192 \"apr\u00e8 s\" [after] and not \"si\", \"bond\" \u2192 \"lien\" [link] and not \"obligation\", \"last\" \u2192 \"derni\u00e8 re\" instead of \"durer\"), grammar is wrong leading to ununderstandability (ex. 2 in French, at several points), dependencies have been modified (\"its resumption of\" should be \"resumption of its\"), and bad handling of time in English leads to incorrect tenses in French (ex. 1 and 2). Actually, Example 2 is total gibberish in French. No sense can be extracted from it. Example 1 has actually a negative adequacy 2 in French, as it gives counterfactual and misleading information (the suspension \"has lasted until\u2026\" instead of \"will last until\u2026\").", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Introduction", | |
| "sec_num": "1" | |
| }, | |
| { | |
| "text": "This illustrates the need of building direct MT systems between French and Chinese, specialized to the domains (stock option markets, economy) and to the corresponding sublanguages. That will certainly considerably improve MT quality. But, even if one augments the BLEU by 25% or more 3 , the problems above remain. In situations where exactness of meaning is crucial and post-editing would require professionals working round the clock (because flash reports have a life expectancy of only a few hours). To improve quality, semantic and discourse information should be used. Section 2 will now give an overall description of this project. Then two aspects where progress has been made will be described in Section 3 in more detail.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Introduction", | |
| "sec_num": "1" | |
| }, | |
| { | |
| "text": "WITH A NEW SMT FRAMEWORK", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "ON-GOING COOPERATIVE RESEARCH TOWARDS DEVELOPING ECONOMY-ORIENTED CHINESE-FRENCH SMT SYSTEMS", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "As mentioned in Section 1, the overall goal of this project is to construct economy-oriented Chinese-French computational linguistic resources, e.g. bilingual semantic resources, parallel corpora, and discourse structure annotation banks, etc., then propose a new SMT model making use of semantic and discourse information, and finally build a webbased collaborative post-editing platform. We distinguish three levels: data development, fundamental research, and tool building. FIGURE 1 shows the overall organization of this project, in which three points should be noted.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Overview of the Project", | |
| "sec_num": "2" | |
| }, | |
| { | |
| "text": "Firstly, a pivot-based method is used for extracting the Chinese-French translation equivalents. We choose to use English as a pivot, because it is much easier to gather large-scale Chinese-English and English-French parallel texts than to directly collect Chinese-French ones. Moreover, pivot-based methods (Wu and Wang, 2007; Paul et al., 2009; Wu and Wang, 2009) have been proven to be effective when building SMT systems for under-resourced language pairs.", | |
| "cite_spans": [ | |
| { | |
| "start": 308, | |
| "end": 327, | |
| "text": "(Wu and Wang, 2007;", | |
| "ref_id": "BIBREF9" | |
| }, | |
| { | |
| "start": 328, | |
| "end": 346, | |
| "text": "Paul et al., 2009;", | |
| "ref_id": "BIBREF6" | |
| }, | |
| { | |
| "start": 347, | |
| "end": 365, | |
| "text": "Wu and Wang, 2009)", | |
| "ref_id": "BIBREF10" | |
| } | |
| ], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "FIGURE 1 : Overall picture of the project organization", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "Secondly, the translation model uses discourse-level information. Actually, recent research has shown that discourselevel information is quite important for machine translation systems, especially those concerning Chinese (Gong et al., 2011 (Gong et al., , 2012 Tu et al., 2013) . As Chinese does not have grammatical markers of tense, the time at which an action takes place usually has to be inferred from the context. Consider again example 1 in Section 1. The source sentence contains three clauses, namely \"\u6062\u590d\u4ea4\u6613\u540e\uff0c\u5982\u8be5\u8bc1\u5238\u4ea4\u6613\u4e2d\u518d\u6b21\u51fa\u73b0\u5f02\u5e38\u60c5\u51b5 (If the trading of this security appears to be abnormal again after resumption of its trading,)\", \"\u672c\u6240\u53ef\u5b9e\u65bd\u7b2c\u4e8c\u6b21\u505c\u724c (we may perform a second suspension upon it)\" and \"\u505c\u724c\u65f6\u95f4\u6301\u7eed\u81f3\u4eca\u65e5\u6536\u76d8\u524d\u4e94\u5206\u949f (and this suspension will last until five minutes before today's closing.)\". Because of the omission of the object in the second clause, we don't know what the second suspension will be executed for, and the unclearness of the tense in the third clause also contributed to the inadequacy of the translation. Suppose now that we have the discourse graph shown in FIGURE 2 below. Both the omitted information and the tense information could then be inferred. Therefore, we decide to conduct research on Chinese-French MT making use of discourse-level information. Thirdly, semantic knowledge is used when extracting the translation equivalents. We would do so because we found that the previous pivot-based approaches, which do not take semantic knowledge into account, may produce incorrect translation equivalents due to the lack of sufficient context. For example, suppose we have a phrase pair \"\u94f6\u884c \uf0ab bank\" in the Chinese-English bilingual phrase table, and phrase pairs \"bank \uf0ab la banque\" and \"bank \uf0ab la rive\" in the English-French bilingual phrase table. Then, using the transfer method (Paul et al., 2009; Wu and Wang, 2009) , we get two Chinese-French phrase pairs, i.e. \"\u94f6\u884c \uf0ab la banque\" and \"\u94f6\u884c \uf0ab la rive\". However, the latter is obviously wrong. We expect that, by incorporating semantic information, this problem could be overcome.", | |
| "cite_spans": [ | |
| { | |
| "start": 222, | |
| "end": 240, | |
| "text": "(Gong et al., 2011", | |
| "ref_id": "BIBREF2" | |
| }, | |
| { | |
| "start": 241, | |
| "end": 261, | |
| "text": "(Gong et al., , 2012", | |
| "ref_id": "BIBREF3" | |
| }, | |
| { | |
| "start": 262, | |
| "end": 278, | |
| "text": "Tu et al., 2013)", | |
| "ref_id": "BIBREF7" | |
| }, | |
| { | |
| "start": 1801, | |
| "end": 1820, | |
| "text": "(Paul et al., 2009;", | |
| "ref_id": "BIBREF6" | |
| }, | |
| { | |
| "start": 1821, | |
| "end": 1839, | |
| "text": "Wu and Wang, 2009)", | |
| "ref_id": "BIBREF10" | |
| } | |
| ], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "FIGURE 1 : Overall picture of the project organization", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "We distinguish seven subtasks in this project (see FIGURE 1), briefly described below.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "FIGURE 1 : Overall picture of the project organization", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "This subtask aims at creating bilingual lexico-semantic data by enriching HowNet (Dong and Dong, 2006) , a Chinese-English conceptual database, with French data. In this subtask, economy-related bilingual terms will be integrated into the semantic resource.", | |
| "cite_spans": [ | |
| { | |
| "start": 81, | |
| "end": 102, | |
| "text": "(Dong and Dong, 2006)", | |
| "ref_id": "BIBREF1" | |
| } | |
| ], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "1) Constructing Chinese-French bilingual lexico-semantic resources", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "In order to carry out SMT research, parallel corpora are needed. Since very few Chinese and French bilingual data are freely available on Internet, it is not easy to construct large-scale Chinese-French parallel corpora, especially economyoriented ones. Therefore, it is reasonable to build the resources of Chinese-French SMT systems using English as a pivot.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "2) Constructing Chinese-French/Chinese-English/English-French parallel corpora", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "To this end, we are building Chinese-English and English-French bilingual corpora related to the domain of economy. Then, to support learning structure transformation knowledge between Chinese and French, a medium-scale Chinese-French parallel corpus will also be created.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "2) Constructing Chinese-French/Chinese-English/English-French parallel corpora", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "In order to incorporate discourse-level information in our Chinese-French SMT system, we first need a Chinese discourse structure analysis module trained on Chinese discourse structure annotation banks. However, the Chinese discourse structure annotation banks, currently under construction to support discourse-based Chinese-related MT research, are not yet available now (Li et al., 2012) . Hence, this subtask aims at constructing a Chinese discourse structure annotation bank based on the Chinese part of the economy-oriented parallel corpora constructed in Subtask 2. The course of constructing the annotation bank could be summarized as two steps. Firstly, sentences are partitioned into sentence groups according to their topic similarities. Secondly, the relationships among the sentences in the same groups, as well as the time information of the sentences or the time relationships between sentence pairs are humanly tagged. The annotation result for example 1 in Section 1 may look like as follows:", | |
| "cite_spans": [ | |
| { | |
| "start": 373, | |
| "end": 390, | |
| "text": "(Li et al., 2012)", | |
| "ref_id": "BIBREF5" | |
| } | |
| ], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "3) Constructing Chinese discourse structure annotation bank", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "(Discourse (Sentences (S1, 0-21), (S2, 22-32), (S3, 33-48)), (Sentence-relationships (Condition S1, S2), (Sequence S2, S3)), (Time-information Time(S1)=present, Time(S2)>Time(S1), Time(S3)=Time(S2)))", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "3) Constructing Chinese discourse structure annotation bank", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "Given the constructed Chinese-English and English-French parallel corpora, the objective of the subtask is to construct Chinese-French translation equivalents. In the course of equivalents extraction, semantic information based on the lexico-semantic resources constructed in Subtask 1 will be considered.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "4) Chinese-French translation equivalents construction via English", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "The objective of this subtask is to research and propose a Chinese discourse structure analysis model based on the Chinese discourse structure annotation bank constructed in Subtask 3.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "5) Chinese discourse structure analysis", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "The goal of this subtask is to propose a new Chinese-French SMT model using semantic and discourse information. In the course of translation, the tense and semantic consistency across sentences will be considered, based on the discourse graphs produced by the analysing module in Subtask 5 and the time model trained with the annotation bank in Subtask 3.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "6) Chinese-French SMT models using semantic and discourse information", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "The objective of this subtask is to build a collaborative web-based post-editing platform. This platform is built based on iMAG/SECTra (Wang and Boitet, 2013) , and incorporates the economy-oriented SMT system developed in Subtask 6.", | |
| "cite_spans": [ | |
| { | |
| "start": 135, | |
| "end": 158, | |
| "text": "(Wang and Boitet, 2013)", | |
| "ref_id": "BIBREF8" | |
| } | |
| ], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "7) Web-based collaborative post-editing platform", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "ON-GOING COOPERATIVE RESEARCH TOWARDS DEVELOPING ECONOMY-ORIENTED CHINESE-FRENCH SMT SYSTEMS WITH A NEW SMT FRAMEWORK", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "7) Web-based collaborative post-editing platform", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "This project started in September 2013. Since then, we progressed on two fronts: data collection and pivot-base construction of bilingual equivalents.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Some preliminary progresses", | |
| "sec_num": "3" | |
| }, | |
| { | |
| "text": "From websites related to stock exchange, about 1000 bilingual web pages have been crawled and handled so far ( Table 1 shows the statistics of this dataset). At the same time, with the participation of 2 Chinese students, we started the process of creating a Chinese-French parallel corpus via post-editing, using an iMAG collaborative gateway, as in (Wang and Boitet, 2013) .", | |
| "cite_spans": [ | |
| { | |
| "start": 352, | |
| "end": 375, | |
| "text": "(Wang and Boitet, 2013)", | |
| "ref_id": "BIBREF8" | |
| } | |
| ], | |
| "ref_spans": [ | |
| { | |
| "start": 111, | |
| "end": 119, | |
| "text": "Table 1", | |
| "ref_id": null | |
| } | |
| ], | |
| "eq_spans": [], | |
| "section": "1) Data collection", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "In previous pivot-based methods, such as (Wu and Wang, 2007) , the translation probability induction may become inaccurate if the semantic tendencies of the source-pivot (SP) corpus and the pivot-target (PT) corpus are different. To overcome this problem, an effective method is to use context information to measure the semantic similarity of the rule pairs. To solve this problem, we have proposed and validated a pivot probability induction method based on topic similarity information. It consists of two steps.", | |
| "cite_spans": [ | |
| { | |
| "start": 41, | |
| "end": 60, | |
| "text": "(Wu and Wang, 2007)", | |
| "ref_id": "BIBREF9" | |
| } | |
| ], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "2) Research on the pivot-based construction of bilingual equivalents", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "Firstly, we use the topic model (Blei, 2003) to discover the latent topic structure for the pivot language document and each synchronous pivot phrase rule is then attached with the topic distribution according to the document it comes from. Formula 1 is used to assign topic distribution to a given rule. Secondly, the probability induction is executed by computing the topic similarity for the rule pairs instead of using simple phrase translation probability multiplication between the SP and PT translation models. Formulas 2-3 are used: ", | |
| "cite_spans": [ | |
| { | |
| "start": 32, | |
| "end": 44, | |
| "text": "(Blei, 2003)", | |
| "ref_id": "BIBREF0" | |
| } | |
| ], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "2) Research on the pivot-based construction of bilingual equivalents", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "EQUATION", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [ | |
| { | |
| "start": 0, | |
| "end": 8, | |
| "text": "EQUATION", | |
| "ref_id": "EQREF", | |
| "raw_str": "( ( | ), ( | )) ( | ) ( ( | ), (", | |
| "eq_num": "| )" | |
| } | |
| ], | |
| "section": "2) Research on the pivot-based construction of bilingual equivalents", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "where s , p , t and Z are the source phrase, pivot phrase, target phrase and topic distribution, respectively. sim (x, y) is the similarity function, and is defined as the cosine (pseudo-distance) of vectors x and y (Formula 4).", | |
| "cite_spans": [ | |
| { | |
| "start": 115, | |
| "end": 121, | |
| "text": "(x, y)", | |
| "ref_id": null | |
| } | |
| ], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "2) Research on the pivot-based construction of bilingual equivalents", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "xy sim x y", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": ", ( , ) cos( , )", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "x y xy \uf03d\uf03d", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": ", ( , ) cos( , )", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "YIDONG CHEN, LINGXIAO WANG, CHRISTIAN BOITET ET XIAODONG SHI", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": ", ( , ) cos( , )", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "The experimental results showed that our method greatly outperforms the baseline. A paper related to this work was published in the Journal of Computational Information Systems in 2013 (Huang et al., 2013) .", | |
| "cite_spans": [ | |
| { | |
| "start": 185, | |
| "end": 205, | |
| "text": "(Huang et al., 2013)", | |
| "ref_id": "BIBREF4" | |
| } | |
| ], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": ", ( , ) cos( , )", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "This paper gives an introduction of a joint project between GETALP and the NLP lab of XMU on building economyoriented Chinese-French SMT systems, as well as its preliminary progress. We demonstrate the need of building direct Chinese-French MT systems specialized field to that field and its sublanguages, and also explain why it seems to be necessary to incorporate semantic and discourse-based information to obtain a quality (of raw MT) sufficient for the needs of users, in a situation when usually there is no time for post-editing. This project is still in its initial stages and many efforts are required in the future. However, specialized versions not yet using the new semantic-and discourserelated features should be demonstrable at the time of the conference. We hope and believe that the future results of this project will be useful in overcoming the language barrier of the communications between China and France in the economy-related field.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Conclusion", | |
| "sec_num": "4" | |
| }, | |
| { | |
| "text": "http://translate.google.com/ 2 Adequacy is usually measured on a scale from 0 to 5, but should rather be similar to a kappa coefficient, ranging from -1 to +1.3 In an experiment conducted at the EU, a Moses system was built on 50,000 sentences of the corpus of the Council. Compared to the unspecialized system built on 20 M sentences (400 times more!), it showed an increase by 25% of the BLEU measure.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "", | |
| "sec_num": null | |
| } | |
| ], | |
| "back_matter": [], | |
| "bib_entries": { | |
| "BIBREF0": { | |
| "ref_id": "b0", | |
| "title": "Latent Dirichlet Allocation", | |
| "authors": [ | |
| { | |
| "first": "D", | |
| "middle": [ | |
| "M" | |
| ], | |
| "last": "Blei", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2003, | |
| "venue": "Journal of Machine Learning", | |
| "volume": "3", | |
| "issue": "", | |
| "pages": "993--1022", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "BLEI D. M. (2003). Latent Dirichlet Allocation. Journal of Machine Learning 3, 993-1022.", | |
| "links": null | |
| }, | |
| "BIBREF1": { | |
| "ref_id": "b1", | |
| "title": "HowNet and the Computation of Meaning", | |
| "authors": [ | |
| { | |
| "first": "Z", | |
| "middle": [], | |
| "last": "Dong", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Q", | |
| "middle": [], | |
| "last": "Dong", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2006, | |
| "venue": "", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "DONG, Z.,AND DONG, Q. (2006). HowNet and the Computation of Meaning. World Scientific Publishing Co.Pte.Ltd.", | |
| "links": null | |
| }, | |
| "BIBREF2": { | |
| "ref_id": "b2", | |
| "title": "Cache-based Document-level Statistical Machine Translation", | |
| "authors": [ | |
| { | |
| "first": "Z", | |
| "middle": [], | |
| "last": "Gong", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "M", | |
| "middle": [], | |
| "last": "Zhang", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "G", | |
| "middle": [], | |
| "last": "Zhou", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2011, | |
| "venue": "Proceedings of the 2011 Conference on Empirical Methods in Natural Language Processing", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "909--919", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "GONG, Z., ZHANG, M., ZHOU, G. (2011). Cache-based Document-level Statistical Machine Translation. Proceedings of the 2011 Conference on Empirical Methods in Natural Language Processing (EMNLP 2011), Edinburgh, UK, pp. 909- 919.", | |
| "links": null | |
| }, | |
| "BIBREF3": { | |
| "ref_id": "b3", | |
| "title": "N-gram-based Tense Models for Statistical Machine Translation", | |
| "authors": [ | |
| { | |
| "first": "Z", | |
| "middle": [], | |
| "last": "Gong", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "M", | |
| "middle": [], | |
| "last": "Zhang", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "C", | |
| "middle": [], | |
| "last": "Tan", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "G", | |
| "middle": [], | |
| "last": "Zhou", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2012, | |
| "venue": "Proceedings of the 2012 Joint Conference on Empirical Methods in Natural Language Processing and Computational Natural Language Learning", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "276--285", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "GONG, Z., ZHANG, M., TAN, C., ZHOU, G. (2012). N-gram-based Tense Models for Statistical Machine Translation. Proceedings of the 2012 Joint Conference on Empirical Methods in Natural Language Processing and Computational Natural Language Learning, Jeju, Korea, pp. 276-285.", | |
| "links": null | |
| }, | |
| "BIBREF4": { | |
| "ref_id": "b4", | |
| "title": "Pivot Probability Induction for Statistical Machine Translation with Topic Similarity", | |
| "authors": [ | |
| { | |
| "first": "Y", | |
| "middle": [], | |
| "last": "Huang", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "X", | |
| "middle": [], | |
| "last": "Shi", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "J", | |
| "middle": [], | |
| "last": "Su", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Y", | |
| "middle": [], | |
| "last": "Chen", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "G", | |
| "middle": [], | |
| "last": "Huang", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2013, | |
| "venue": "Journal of Computational Information Systems", | |
| "volume": "20", | |
| "issue": "9", | |
| "pages": "8351--8359", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "HUANG, Y., SHI, X., SU, J., CHEN Y., HUANG, G. (2013). Pivot Probability Induction for Statistical Machine Translation with Topic Similarity. Journal of Computational Information Systems, 20(9), 8351-8359.", | |
| "links": null | |
| }, | |
| "BIBREF5": { | |
| "ref_id": "b5", | |
| "title": "Summary of research on English discourse parsing", | |
| "authors": [ | |
| { | |
| "first": "Y", | |
| "middle": [], | |
| "last": "Li", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "K", | |
| "middle": [], | |
| "last": "Zhu", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "G", | |
| "middle": [], | |
| "last": "Zhou", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2012, | |
| "venue": "Application Research of Computers", | |
| "volume": "", | |
| "issue": "6", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "LI, Y., ZHU, K., ZHOU, G. (2012). Summary of research on English discourse parsing. Application Research of Computers, 29(6).", | |
| "links": null | |
| }, | |
| "BIBREF6": { | |
| "ref_id": "b6", | |
| "title": "On the Importance of Pivot Language Selection for Statistical Machine Translation", | |
| "authors": [ | |
| { | |
| "first": "M", | |
| "middle": [], | |
| "last": "Paul", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "H", | |
| "middle": [], | |
| "last": "Yamamoto", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "E", | |
| "middle": [], | |
| "last": "Sumita", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "", | |
| "middle": [], | |
| "last": "Nakamura S", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2009, | |
| "venue": "Proceedings of NAACL-HLT'09", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "221--224", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "PAUL, M., YAMAMOTO, H., SUMITA, E., AND NAKAMURA S. (2009). On the Importance of Pivot Language Selection for Statistical Machine Translation. Proceedings of NAACL-HLT'09, 2009, pp. 221-224.", | |
| "links": null | |
| }, | |
| "BIBREF7": { | |
| "ref_id": "b7", | |
| "title": "A Novel Translation Framework Based on Rhetorical Structure Theory", | |
| "authors": [ | |
| { | |
| "first": "M", | |
| "middle": [], | |
| "last": "Tu", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Y", | |
| "middle": [], | |
| "last": "Zhou", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "", | |
| "middle": [], | |
| "last": "Zong C", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2013, | |
| "venue": "The 51st Annual Meeting of the Association for Computational Linguistics (ACL 2013)", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "TU, M., ZHOU Y., AND ZONG C. (2013). A Novel Translation Framework Based on Rhetorical Structure Theory. The 51st Annual Meeting of the Association for Computational Linguistics (ACL 2013), Sofia, Bulgaria.", | |
| "links": null | |
| }, | |
| "BIBREF8": { | |
| "ref_id": "b8", | |
| "title": "Online production of HQ parallel corpora and permanent task-based evaluation of multiple MT systems: both can be obtained through iMAGs with no added cost", | |
| "authors": [ | |
| { | |
| "first": "L", | |
| "middle": [], | |
| "last": "Wang", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "C", | |
| "middle": [], | |
| "last": "Boitet", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2013, | |
| "venue": "Proceedings of MT Summit XIV, The 2nd Workshop on Post-Editing Technologies and Practice", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "WANG, L., AND BOITET, C. (2013), Online production of HQ parallel corpora and permanent task-based evaluation of multiple MT systems: both can be obtained through iMAGs with no added cost. Proceedings of MT Summit XIV, The 2nd Workshop on Post-Editing Technologies and Practice. Nice, France 2 -6 September 2013.", | |
| "links": null | |
| }, | |
| "BIBREF9": { | |
| "ref_id": "b9", | |
| "title": "Pivot Language Approach for Phrase-Based Statistical Machine Translation", | |
| "authors": [ | |
| { | |
| "first": "H", | |
| "middle": [], | |
| "last": "Wu", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "", | |
| "middle": [], | |
| "last": "Wang H", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2007, | |
| "venue": "Proceedings of ACL'07", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "856--863", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "WU, H. AND WANG H. (2007). Pivot Language Approach for Phrase-Based Statistical Machine Translation. Proceedings of ACL'07, 2007, pp. 856-863.", | |
| "links": null | |
| }, | |
| "BIBREF10": { | |
| "ref_id": "b10", | |
| "title": "Revisiting Pivot Language Approach for Machine Translation", | |
| "authors": [ | |
| { | |
| "first": "H", | |
| "middle": [], | |
| "last": "Wu", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "H", | |
| "middle": [], | |
| "last": "Wang", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2009, | |
| "venue": "Proceeding of ACL'09", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "154--162", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "WU, H. AND WANG, H. (2009). Revisiting Pivot Language Approach for Machine Translation. Proceeding of ACL'09, 2009, pp. 154-162.", | |
| "links": null | |
| } | |
| }, | |
| "ref_entries": { | |
| "FIGREF0": { | |
| "text": "Ex2 (Source): \u56db\u3001\u51e1 2013 \u5e74 12 \u6708 31 \u65e5\u524d\u5728\u672c\u6240\u4e0a\u5e02\u7684\u516c\u53f8\u503a\u5238\uff0c\u4ee5\u53ca 2014 \u5e74\u5728\u672c\u6240\u4e0a\u5e02\u65f6\u672a\u62ab\u9732 2013 \u5e74 \u5e74\u5ea6\u62a5\u544a\u7684\u516c\u53f8\u503a\u5238\uff0c\u5e94\u4e8e 2014 \u5e74 4 \u6708 30 \u65e5\u524d\u5b8c\u6210\u672c\u6b21\u5e74\u5ea6\u62a5\u544a\u7684\u62ab\u9732\u5de5\u4f5c\u3002 (GT-zh-en) Fourth, where the corporate bond December 31, 2013 are listed in this, as well as the 2014 listed in the 2013 annual report of undisclosed corporate bonds should be April 30, 2014 disclosed in the annual report is completed work. (GT-zh-en-PE) Fourth, the bonds which were listed before December 31, 2013 or listed in 2014, but have not yet disclosed their annual reports of 2013, should complete the disclosure of their annual reports before April 30, 2014. (GT-zh-en-fr = GT-zh-fr): Quatri\u00e8 mement, lorsque le lien de l'entreprise 31 D\u00e9 cembre , 2013 figurent dans cet , ainsi que de 2014 figurant dans le rapport annuel 2013 d'obligations de soci\u00e9 t\u00e9 s non divulgu\u00e9 s devrait \u00ea tre de 30 Avril , 2014 communiqu\u00e9 es dans le rapport annuel est termin\u00e9 travail. (GT-zh-en-PE-fr): Quatri\u00e8 mement, les liens qui ont \u00e9 t\u00e9 r\u00e9 pertori\u00e9 s avant le 31 D\u00e9 cembre 2013 r\u00e9 pertori\u00e9 s en 2014, mais n'ont pas encore divulgu\u00e9 leurs rapports annuels de 2013, devrait compl\u00e9 ter la divulgation de leurs rapports annuels avant le 30 Avril 2014.", | |
| "type_str": "figure", | |
| "uris": null, | |
| "num": null | |
| }, | |
| "FIGREF1": { | |
| "text": "The discourse graph of the source sentence of Example 1 trading of this security appears to be abnormal again after resumption of its trading,) \u672c\u6240\u53ef\u5b9e\u65bd\u7b2c\u4e8c\u6b21\u505c\u724c\uff0c (we may perform a second suspension upon it) \u505c\u724c\u65f6\u95f4\u6301\u7eed\u81f3\u4eca\u65e5\u6536\u76d8\u524d\u4e94\u5206\u949f\u3002 (and this suspension will last until five minutes before today's closing.)ConditionSequenceYIDONG CHEN, LINGXIAO WANG, CHRISTIAN BOITET ET XIAODONG SHI", | |
| "type_str": "figure", | |
| "uris": null, | |
| "num": null | |
| }, | |
| "FIGREF2": { | |
| "text": "a bilingual document in the given bilingual corpus * \uf0e5 . ( | ) ip P z d represents the probability of the ith topic of the pivot language document d p in D. In addition, the function count(I, D) denotes the frequency of the rule instance I in D.", | |
| "type_str": "figure", | |
| "uris": null, | |
| "num": null | |
| } | |
| } | |
| } | |
| } |