| { | |
| "paper_id": "2005", | |
| "header": { | |
| "generated_with": "S2ORC 1.0.0", | |
| "date_generated": "2023-01-19T07:21:17.068194Z" | |
| }, | |
| "title": "Nobody is Perfect: ATR's Hybrid Approach to Spoken Language Translation", | |
| "authors": [ | |
| { | |
| "first": "Michael", | |
| "middle": [], | |
| "last": "Paul", | |
| "suffix": "", | |
| "affiliation": { | |
| "laboratory": "ATR Spoken Language Communication Research Laboratories", | |
| "institution": "Keihanna Science City", | |
| "location": { | |
| "addrLine": "Hikaridai 2-2-2", | |
| "postCode": "619-0288", | |
| "settlement": "Kyoto" | |
| } | |
| }, | |
| "email": "" | |
| }, | |
| { | |
| "first": "Takao", | |
| "middle": [], | |
| "last": "Doi", | |
| "suffix": "", | |
| "affiliation": { | |
| "laboratory": "ATR Spoken Language Communication Research Laboratories", | |
| "institution": "Keihanna Science City", | |
| "location": { | |
| "addrLine": "Hikaridai 2-2-2", | |
| "postCode": "619-0288", | |
| "settlement": "Kyoto" | |
| } | |
| }, | |
| "email": "" | |
| }, | |
| { | |
| "first": "Youngsook", | |
| "middle": [], | |
| "last": "Hwang", | |
| "suffix": "", | |
| "affiliation": { | |
| "laboratory": "ATR Spoken Language Communication Research Laboratories", | |
| "institution": "Keihanna Science City", | |
| "location": { | |
| "addrLine": "Hikaridai 2-2-2", | |
| "postCode": "619-0288", | |
| "settlement": "Kyoto" | |
| } | |
| }, | |
| "email": "" | |
| }, | |
| { | |
| "first": "Kenji", | |
| "middle": [], | |
| "last": "Imamura", | |
| "suffix": "", | |
| "affiliation": { | |
| "laboratory": "ATR Spoken Language Communication Research Laboratories", | |
| "institution": "Keihanna Science City", | |
| "location": { | |
| "addrLine": "Hikaridai 2-2-2", | |
| "postCode": "619-0288", | |
| "settlement": "Kyoto" | |
| } | |
| }, | |
| "email": "" | |
| }, | |
| { | |
| "first": "Hideo", | |
| "middle": [], | |
| "last": "Okuma", | |
| "suffix": "", | |
| "affiliation": { | |
| "laboratory": "ATR Spoken Language Communication Research Laboratories", | |
| "institution": "Keihanna Science City", | |
| "location": { | |
| "addrLine": "Hikaridai 2-2-2", | |
| "postCode": "619-0288", | |
| "settlement": "Kyoto" | |
| } | |
| }, | |
| "email": "" | |
| }, | |
| { | |
| "first": "Eiichiro", | |
| "middle": [], | |
| "last": "Sumita", | |
| "suffix": "", | |
| "affiliation": { | |
| "laboratory": "ATR Spoken Language Communication Research Laboratories", | |
| "institution": "Keihanna Science City", | |
| "location": { | |
| "addrLine": "Hikaridai 2-2-2", | |
| "postCode": "619-0288", | |
| "settlement": "Kyoto" | |
| } | |
| }, | |
| "email": "" | |
| }, | |
| { | |
| "first": "{michael", | |
| "middle": [], | |
| "last": "Paul", | |
| "suffix": "", | |
| "affiliation": { | |
| "laboratory": "ATR Spoken Language Communication Research Laboratories", | |
| "institution": "Keihanna Science City", | |
| "location": { | |
| "addrLine": "Hikaridai 2-2-2", | |
| "postCode": "619-0288", | |
| "settlement": "Kyoto" | |
| } | |
| }, | |
| "email": "" | |
| }, | |
| { | |
| "first": "", | |
| "middle": [], | |
| "last": "Jp", | |
| "suffix": "", | |
| "affiliation": { | |
| "laboratory": "ATR Spoken Language Communication Research Laboratories", | |
| "institution": "Keihanna Science City", | |
| "location": { | |
| "addrLine": "Hikaridai 2-2-2", | |
| "postCode": "619-0288", | |
| "settlement": "Kyoto" | |
| } | |
| }, | |
| "email": "" | |
| } | |
| ], | |
| "year": "", | |
| "venue": null, | |
| "identifiers": {}, | |
| "abstract": "This paper describes ATR's hybrid approach to spoken language translation and it's application to the IWSLT 2005 translation task. Multiple corpus-based translation engines are used to translate the same input, whereby the best translation among the element MT outputs is selected according to statistical models. The evaluation results of the Japanese-to-English and Chinese-to-English translation tasks for different training data conditions showed the potential of the proposed hybrid approach and revealed new directions in how to improve the current system performance.", | |
| "pdf_parse": { | |
| "paper_id": "2005", | |
| "_pdf_hash": "", | |
| "abstract": [ | |
| { | |
| "text": "This paper describes ATR's hybrid approach to spoken language translation and it's application to the IWSLT 2005 translation task. Multiple corpus-based translation engines are used to translate the same input, whereby the best translation among the element MT outputs is selected according to statistical models. The evaluation results of the Japanese-to-English and Chinese-to-English translation tasks for different training data conditions showed the potential of the proposed hybrid approach and revealed new directions in how to improve the current system performance.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Abstract", | |
| "sec_num": null | |
| } | |
| ], | |
| "body_text": [ | |
| { | |
| "text": "Corpus-based approaches to machine translation (MT) have achieved much progress over the last decades. There are two main strategies used in corpus-based translation:", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Introduction", | |
| "sec_num": "1." | |
| }, | |
| { | |
| "text": "1. Example-Based Machine Translation (EBMT) [1] :", | |
| "cite_spans": [ | |
| { | |
| "start": 44, | |
| "end": 47, | |
| "text": "[1]", | |
| "ref_id": "BIBREF0" | |
| } | |
| ], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Introduction", | |
| "sec_num": "1." | |
| }, | |
| { | |
| "text": "EBMT uses the corpus directly. EBMT retrieves the translation examples that are best matched to an input expression and then adjusts the examples to obtain the translation.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Introduction", | |
| "sec_num": "1." | |
| }, | |
| { | |
| "text": "SMT learns statistical models for translation from corpora and dictionaries and then searches for the best translation at run-time according to the statistical models for language and translation.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Statistical Machine Translation (SMT) [2]:", | |
| "sec_num": "2." | |
| }, | |
| { | |
| "text": "Despite a high performance on average, these approaches can often produce translations with severe errors. However, different MT engines not always do the same error. Due to the particular output characteristics of each MT engine, quite different translation hypotheses are produced. Thus, combining multiple MT systems can boost the system performance by exploiting the strengths of each MT engine. We propose a corpus-based approach that uses multiple translation engines in parallel. All engines translate the same input, whereby the best translation among the multiple MT outputs is selected according to multiple statistical language and translation models. The outline of our hybrid approach is given in Section 2.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Statistical Machine Translation (SMT) [2]:", | |
| "sec_num": "2." | |
| }, | |
| { | |
| "text": "The proposed system was applied to two translation directions (Japanese-to-English, Chinese-to-English) and three data tracks (Supplied Data Track, Supplied+Tools Data Track, C-STAR Track) . The evaluation of the obtained results is given in Section 3.", | |
| "cite_spans": [ | |
| { | |
| "start": 126, | |
| "end": 188, | |
| "text": "(Supplied Data Track, Supplied+Tools Data Track, C-STAR Track)", | |
| "ref_id": null | |
| } | |
| ], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Statistical Machine Translation (SMT) [2]:", | |
| "sec_num": "2." | |
| }, | |
| { | |
| "text": "We use an architecture in which multiple EBMT and SMT engines work in parallel and their outputs are passed to a post-process that selects the best candidate according to SMT models (cf. Figure 1 ).", | |
| "cite_spans": [], | |
| "ref_spans": [ | |
| { | |
| "start": 187, | |
| "end": 195, | |
| "text": "Figure 1", | |
| "ref_id": null | |
| } | |
| ], | |
| "eq_spans": [], | |
| "section": "System Description", | |
| "sec_num": "2." | |
| }, | |
| { | |
| "text": "This section is structured as follows: (1) eight corpusbased MT engines are introduced in Section 2.1; (2) the SMT-based approach to select the best translation out of multiple hypotheses is explained in Section 2.2; (3) the resources utilized for the IWSLT 2005 translation task are described in Section 2.3; and (4) a summary of the data tracks we participated in and an overview on which MT engines were utilized for the respective tracks is given in Section 2.4.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "System Description", | |
| "sec_num": "2." | |
| }, | |
| { | |
| "text": "We employed the following four SMT and four EBMT systems: An SMT engine that uses an example-based decoding method [ The translation knowledge of the eight MT systems is automatically acquired from a parallel corpus. The characteristics of the element MTs are summarized in Table 1 .", | |
| "cite_spans": [ | |
| { | |
| "start": 115, | |
| "end": 116, | |
| "text": "[", | |
| "ref_id": null | |
| } | |
| ], | |
| "ref_spans": [ | |
| { | |
| "start": 274, | |
| "end": 281, | |
| "text": "Table 1", | |
| "ref_id": "TABREF1" | |
| } | |
| ], | |
| "eq_spans": [], | |
| "section": "MT Engines", | |
| "sec_num": "2.1." | |
| }, | |
| { | |
| "text": "SAT is an SMT system [3] . The decoder searches for an optimal translation by using SMT models starting from a decoder seed, i.e., the source language input paired with an initial translation hypothesis. In SAT, the search is initiated from similar translation examples retrieved from a parallel corpus.", | |
| "cite_spans": [ | |
| { | |
| "start": 21, | |
| "end": 24, | |
| "text": "[3]", | |
| "ref_id": "BIBREF2" | |
| } | |
| ], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "SAT", | |
| "sec_num": "2.1.1." | |
| }, | |
| { | |
| "text": "The similarity measure used here is a combination of an editdistance and tf/idf criteria as seen in the information retrieval framework [4] . The retrieved translations are modified by using a greedy search approach to find better translations [5] .", | |
| "cite_spans": [ | |
| { | |
| "start": 136, | |
| "end": 139, | |
| "text": "[4]", | |
| "ref_id": "BIBREF3" | |
| }, | |
| { | |
| "start": 244, | |
| "end": 247, | |
| "text": "[5]", | |
| "ref_id": "BIBREF4" | |
| } | |
| ], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "SAT", | |
| "sec_num": "2.1.1." | |
| }, | |
| { | |
| "text": "PBHMTM is a statistical MT system that is based on a phrase-based HMM translation model [6] . The model directly structures the phrase-based SMT approach in a Hidden Markov structure. The probability P (f |e) of translating a foreign source sentence f into a target language sentence e using noisy channel modeling is approximated by introducing two new hidden variables,f and\u0113, to explicitly capture the phrase translation relationship:", | |
| "cite_spans": [ | |
| { | |
| "start": 88, | |
| "end": 91, | |
| "text": "[6]", | |
| "ref_id": "BIBREF5" | |
| } | |
| ], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "PBHMTM", | |
| "sec_num": "2.1.2." | |
| }, | |
| { | |
| "text": "EQUATION", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [ | |
| { | |
| "start": 0, | |
| "end": 8, | |
| "text": "EQUATION", | |
| "ref_id": "EQREF", | |
| "raw_str": "P (f |e) = f ,\u0113 P(f |f ,\u0113, e)P(f |\u0113, e)P(\u0113|e)", | |
| "eq_num": "(1)" | |
| } | |
| ], | |
| "section": "PBHMTM", | |
| "sec_num": "2.1.2." | |
| }, | |
| { | |
| "text": "The first term represents the probability that a phrasesegmented source language sentencef can be reordered and generated as the source text of f (Phrase Segmentation Model). The second term indicates the translation probability of the two phrase sequences of\u0113 andf (Phrase Translation Model). The last term is the likelihood that the phrasesegmented target language sentence\u0113 is generated from e (Phrase Ngram Model).", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "PBHMTM", | |
| "sec_num": "2.1.2." | |
| }, | |
| { | |
| "text": "If the phrase segmented sentences\u0113 andf are expanded into corresponding lattice structures\u0112 andF, then the approximation of the proposed models can be regarded as a Hidden Markov Model in which each source phrase in the latticeF is treated as an observation emitted from a state, a target phrase, in the lattice\u0112.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "PBHMTM", | |
| "sec_num": "2.1.2." | |
| }, | |
| { | |
| "text": "The decoder is a word-graph-based decoder [7] , which allows the multi-pass decoding strategies to incorporate complicated submodel structures. The first pass of the decoding procedure generates the word-graph, or the lattice, of translations for an input sentence by using a beam search. On the first pass, the submodels of all phrase-based HMM translation models were integrated with the word-based trigram language model and the class 5-gram model. The second pass uses the A* strategy to search for the best path for translation on the generated word-graph.", | |
| "cite_spans": [ | |
| { | |
| "start": 42, | |
| "end": 45, | |
| "text": "[7]", | |
| "ref_id": "BIBREF6" | |
| } | |
| ], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "PBHMTM", | |
| "sec_num": "2.1.2." | |
| }, | |
| { | |
| "text": "MSEP is a phrase-based SMT system that utilizes morphosyntactic information such as part-of-speech and chunk information [8] . It exploits a phrase translation lexicon that is created using word-alignment results and chunk boundary information in a target language sentence. Reliable bilingual phrase pairs are identified using the statistical \u03c7 2 -test at a significance level \u03b1=0.05 over a given frequency threshold.", | |
| "cite_spans": [ | |
| { | |
| "start": 121, | |
| "end": 124, | |
| "text": "[8]", | |
| "ref_id": "BIBREF7" | |
| } | |
| ], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "MSEP", | |
| "sec_num": "2.1.3." | |
| }, | |
| { | |
| "text": "For selecting the most probable translation of a given source sentece, a log-linear model is used:", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "MSEP", | |
| "sec_num": "2.1.3." | |
| }, | |
| { | |
| "text": "EQUATION", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [ | |
| { | |
| "start": 0, | |
| "end": 8, | |
| "text": "EQUATION", | |
| "ref_id": "EQREF", | |
| "raw_str": "P r \u039b (e I 1 |f J 1 ) = exp( m \u03bb m h m (e I 1 , f J 1 , a J 1 )) e I 1 ,a J 1 exp( m \u03bb m h m (e I 1 , f J 1 , a J 1 ))", | |
| "eq_num": "(2)" | |
| } | |
| ], | |
| "section": "MSEP", | |
| "sec_num": "2.1.3." | |
| }, | |
| { | |
| "text": "where", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "MSEP", | |
| "sec_num": "2.1.3." | |
| }, | |
| { | |
| "text": "h m (e I 1 , f J 1 , a J 1 )", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "MSEP", | |
| "sec_num": "2.1.3." | |
| }, | |
| { | |
| "text": "is the m-th feature and \u03bb m is the weight of the feature.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "MSEP", | |
| "sec_num": "2.1.3." | |
| }, | |
| { | |
| "text": "In addition to the IBM model 4 features (word-based ngram language model P r(e I 1 ), lexicon model t(f |e), fertility model n(\u03c6|e), distortion probability d, and NULL translation model p 1 ), we incorporate the following features into the loglinear translation model:", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "MSEP", | |
| "sec_num": "2.1.3." | |
| }, | |
| { | |
| "text": "\u2022 Class-based n-gram model:", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "MSEP", | |
| "sec_num": "2.1.3." | |
| }, | |
| { | |
| "text": "P r(e I 1 ) = i Pr(e i |c i )P r(c i |c i\u22121 1 )", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "MSEP", | |
| "sec_num": "2.1.3." | |
| }, | |
| { | |
| "text": "\u2022 Length model: P r(l|e I 1 , f J 1 ), whereby l is the length (number of words) of a translated target sentence.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "MSEP", | |
| "sec_num": "2.1.3." | |
| }, | |
| { | |
| "text": "\u2022 Phrase matching score: The translated target sentence is matched with phrase translation examples that are extracted from a parallel corpus based on bidirectional word alignment of phrase translation pairs. A score is derived based on the number of matches.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "MSEP", | |
| "sec_num": "2.1.3." | |
| }, | |
| { | |
| "text": "HPATR2 is a statistical MT system based on syntactic transfer [9] . The translation model of HPATR2 is defined as an inside probability of two parse trees, which is used to create probabilistic context-free grammar rules.", | |
| "cite_spans": [ | |
| { | |
| "start": 62, | |
| "end": 65, | |
| "text": "[9]", | |
| "ref_id": "BIBREF8" | |
| } | |
| ], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "HPATR2", | |
| "sec_num": "2.1.4." | |
| }, | |
| { | |
| "text": "The system searches for the best translation that maximizes the product of the following probabilities, where F, E are a source and a target parse trees, and \u03b8, \u03c0 are context-free grammar rules of the source and the target language, respectively.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "HPATR2", | |
| "sec_num": "2.1.4." | |
| }, | |
| { | |
| "text": "\u2022 Probability of Source Tree Model", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "HPATR2", | |
| "sec_num": "2.1.4." | |
| }, | |
| { | |
| "text": "EQUATION", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [ | |
| { | |
| "start": 0, | |
| "end": 8, | |
| "text": "EQUATION", | |
| "ref_id": "EQREF", | |
| "raw_str": "P (F) = \u03b8:\u03b8\u2208F P (\u03b8)", | |
| "eq_num": "(3)" | |
| } | |
| ], | |
| "section": "HPATR2", | |
| "sec_num": "2.1.4." | |
| }, | |
| { | |
| "text": "\u2022 Probability of Target Tree Model", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "HPATR2", | |
| "sec_num": "2.1.4." | |
| }, | |
| { | |
| "text": "EQUATION", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [ | |
| { | |
| "start": 0, | |
| "end": 8, | |
| "text": "EQUATION", | |
| "ref_id": "EQREF", | |
| "raw_str": "P (E) = \u03c0:\u03c0\u2208E P (\u03c0)", | |
| "eq_num": "(4)" | |
| } | |
| ], | |
| "section": "HPATR2", | |
| "sec_num": "2.1.4." | |
| }, | |
| { | |
| "text": "\u2022 Probability of Tree-mapping Model", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "HPATR2", | |
| "sec_num": "2.1.4." | |
| }, | |
| { | |
| "text": "EQUATION", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [ | |
| { | |
| "start": 0, | |
| "end": 8, | |
| "text": "EQUATION", | |
| "ref_id": "EQREF", | |
| "raw_str": "P (F|E)P (E|F) = \u03b8:\u03b8\u2208F , \u03c0:\u03c0\u2208E P (\u03b8|\u03c0)P (\u03c0|\u03b8)", | |
| "eq_num": "(5)" | |
| } | |
| ], | |
| "section": "HPATR2", | |
| "sec_num": "2.1.4." | |
| }, | |
| { | |
| "text": "A characteristic of HPATR2 is that not only word translations but also the translation of multi-word sequences is carried out by the syntactic transfer. Parsing hypotheses, which are multi-word sequences connected by context-free grammar rules, are created. The best hypothesis (parse tree and translation) is selected based on the above models.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "HPATR2", | |
| "sec_num": "2.1.4." | |
| }, | |
| { | |
| "text": "Therefore, HPATR2 is an MT system that contains features of phrase-based SMT as well as syntax-based SMT.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "HPATR2", | |
| "sec_num": "2.1.4." | |
| }, | |
| { | |
| "text": "HPATR is an extension of the example-based HPAT system (cf. Section 2.1.6) that incorporates a word-based statistical MT system [10] . Similar to HPAT, an EBMT module based on syntactic transfer is used to generate translation candidates that have minimum semantic distances. However, word selection is not performed during transfer, but all possible word translation candidates are generated.", | |
| "cite_spans": [ | |
| { | |
| "start": 128, | |
| "end": 132, | |
| "text": "[10]", | |
| "ref_id": "BIBREF9" | |
| } | |
| ], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "HPATR", | |
| "sec_num": "2.1.5." | |
| }, | |
| { | |
| "text": "In a second step, an SMT module using a lexicon model and an n-gram language model is exploited to search for the best translation that maximizes the product of the probabilities. Therefore, HPATR selects the best translation among the output of example-based MT using models of statistical MT from the viewpoints of adequacy of word translation and fluency of the target sentence.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "HPATR", | |
| "sec_num": "2.1.5." | |
| }, | |
| { | |
| "text": "HPAT is an example-based MT system based on syntactic transfer [11] . The most important knowledge in HPAT are transfer rules, which define the correspondences between source and target patterns. The transfer rules can be regarded as synchronized context-free grammar rules.", | |
| "cite_spans": [ | |
| { | |
| "start": 63, | |
| "end": 67, | |
| "text": "[11]", | |
| "ref_id": "BIBREF10" | |
| } | |
| ], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "HPAT", | |
| "sec_num": "2.1.6." | |
| }, | |
| { | |
| "text": "When the system translates an input sentence, the sentence is first parsed by using the source side of the transfer rules. Next, a tree structure of the target language is generated by mapping the source grammar rules to the corresponding target rules. When non-terminal symbols remain in the target tree, target words are inserted by referring to a translation dictionary.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "HPAT", | |
| "sec_num": "2.1.6." | |
| }, | |
| { | |
| "text": "Ambiguities, which occur during parsing or mapping, are resolved by selecting the rules that minimize the semantic distance between the input words and source examples of the transfer rules. In general, the automatic acquisition process generates many redundant rules. To avoid this problem, HPAT optimizes the transfer rules by removing redundant rules (feedback cleaning, [12] ) in order to increase an automatic evaluation score.", | |
| "cite_spans": [ | |
| { | |
| "start": 374, | |
| "end": 378, | |
| "text": "[12]", | |
| "ref_id": "BIBREF11" | |
| } | |
| ], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "HPAT", | |
| "sec_num": "2.1.6." | |
| }, | |
| { | |
| "text": "D 3 (DP-match Driven transDucer) is an EBMT system that exploits DP-matching between word sequences [13] . In the translation process, D 3 retrieves the most similar source sentence from a parallel corpus for an input sentence.", | |
| "cite_spans": [ | |
| { | |
| "start": 100, | |
| "end": 104, | |
| "text": "[13]", | |
| "ref_id": "BIBREF12" | |
| } | |
| ], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "D 3", | |
| "sec_num": "2.1.7." | |
| }, | |
| { | |
| "text": "The similarity is calculated based on the counts of insertion, deletion, and substitution operations, where the total is normalized by the sum of the lengths of the word sequences. Substitution considers the semantic distance between two substituted words and is defined as the division of K, the level of the least common abstraction in the thesaurus of two words, by N, the height of the thesaurus [14] .", | |
| "cite_spans": [ | |
| { | |
| "start": 400, | |
| "end": 404, | |
| "text": "[14]", | |
| "ref_id": "BIBREF13" | |
| } | |
| ], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "D 3", | |
| "sec_num": "2.1.7." | |
| }, | |
| { | |
| "text": "According to the difference between the input sentence and the retrieved source sentence, the translation of the retrieved source sentence is modified by using dictionaries.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "D 3", | |
| "sec_num": "2.1.7." | |
| }, | |
| { | |
| "text": "EM is a translation memory system that matches a given source sentence against the source language parts of translation examples extracted from a parallel corpus. In case an exact match can be achieved, the corresponding target language sentence will be used. Otherwise, the system fails to output a translation.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "EM", | |
| "sec_num": "2.1.8." | |
| }, | |
| { | |
| "text": "We use an SMT-based method of automatically selecting the best translation among outputs generated by multiple MT systems [15] . This approach scores MT outputs by using multiple language (LM) and translation model (TM) pairs trained on different subsets of the training data. It uses a statistical test to check whether the average TM\u2022LM score of one MT output is significantly higher than those of another MT output. The SELECTOR algorithm is summarized in Figure 2 . In order to detect a significant difference, the proposed method first prepares multiple subsets of the full parallel text corpus according to n-fold 1 cross validation [16] and trains pairs of language and translation models on the resepective subsets.", | |
| "cite_spans": [ | |
| { | |
| "start": 122, | |
| "end": 126, | |
| "text": "[15]", | |
| "ref_id": "BIBREF14" | |
| }, | |
| { | |
| "start": 639, | |
| "end": 643, | |
| "text": "[16]", | |
| "ref_id": "BIBREF15" | |
| } | |
| ], | |
| "ref_spans": [ | |
| { | |
| "start": 459, | |
| "end": 467, | |
| "text": "Figure 2", | |
| "ref_id": "FIGREF1" | |
| } | |
| ], | |
| "eq_spans": [], | |
| "section": "Selection of the Best MT Engine Output", | |
| "sec_num": "2.2." | |
| }, | |
| { | |
| "text": "EQUATION", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [ | |
| { | |
| "start": 0, | |
| "end": 8, | |
| "text": "EQUATION", | |
| "ref_id": "EQREF", | |
| "raw_str": "(1) proc SELECTOR( Input, Corpus, n, M T 1 , . . . , M T m ) ; (2) begin (3) ( * initalize statistical models * ) (4) for each i in {1, . . . , n} do (5) Corpus i \u2190 subset(Corpus) ; (6) T M i \u2190 translation-model(Corpus i ) ; (7) LM i \u2190 language-model(Corpus i ) ;", | |
| "eq_num": "(" | |
| } | |
| ], | |
| "section": "Selection of the Best MT Engine Output", | |
| "sec_num": "2.2." | |
| }, | |
| { | |
| "text": "The algorithm assumes a priority order in the given MT engines, i.e., the indices i of M T i should indicate an increasing level (i=1 -high, . . ., i=n -low) of translation quality of the utilized MT systems. Such a priority order can be obtained by evaluating the respective system performance using a separate development set 2 .", | |
| "cite_spans": [ | |
| { | |
| "start": 328, | |
| "end": 329, | |
| "text": "2", | |
| "ref_id": "BIBREF1" | |
| } | |
| ], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Selection of the Best MT Engine Output", | |
| "sec_num": "2.2." | |
| }, | |
| { | |
| "text": "Given an input sentence, m translation hypotheses are produced by the element MT engines, whereby n different statistical scores (TM\u2022LM i ) are assigned to each hypothesis. In order to check whether the highest score is significantly different from the others, a multiple comparison test based on the Kruskal-Wallis test [17] is used. If the MT output is significantly better, this output is selected. Otherwise, the output of the best performing MT engine according to the given priority order is selected.", | |
| "cite_spans": [ | |
| { | |
| "start": 321, | |
| "end": 325, | |
| "text": "[17]", | |
| "ref_id": "BIBREF16" | |
| } | |
| ], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Selection of the Best MT Engine Output", | |
| "sec_num": "2.2." | |
| }, | |
| { | |
| "text": "The performance of the selection method depends on the predefined priority order of the MT engines. For the experiments described in Section 3, the respective priority order for each data track were determined as follows:", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Selection of the Best MT Engine Output", | |
| "sec_num": "2.2." | |
| }, | |
| { | |
| "text": "1. The highest priority was given to the EM engine.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Selection of the Best MT Engine Output", | |
| "sec_num": "2.2." | |
| }, | |
| { | |
| "text": "2. The rest of the MT engines were sorted according to the WER scores obtained for the development set.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Selection of the Best MT Engine Output", | |
| "sec_num": "2.2." | |
| }, | |
| { | |
| "text": "3. For each subset {MT 1 ,. . .,MT j } (1\u2264j\u2264n-1) of topscoring j MT engines, all order combinations were used to translate the development set. The combination achieving the lowest WER score was chosen for the translation of the official test set.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Selection of the Best MT Engine Output", | |
| "sec_num": "2.2." | |
| }, | |
| { | |
| "text": "The participants were supplied with 20,000 sentence pairs for each translation direction that are subsets of the BTEC corpus introduced in Section 2.3.1. For the Japanese and Chinese parts, word segmentations were provided based on the output of the speech recognition engine utilized in this years' workshop. In addition, two development sets of additional 506 and 500 sentences, respectively, including up to 16 reference translations, were provided to the participants.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Resources", | |
| "sec_num": "2.3." | |
| }, | |
| { | |
| "text": "The Basic Travel Expressions Corpus (BTEC) is a collection of sentences that bilingual travel experts consider useful for people going to or coming from another country and cover utterances for every potential subject in travel situations [18] . The original Japanese-English corpus has been translated into several languages by members of the Consortium for Speech Translation Advanced Research (C-STAR) 3 resulting in a multilingual spoken language corpus consisting of 172K sentences/language. Moreover, we used additional training data from the same domain for the Japanese-to-English translation task. In total, 541K of aligned sentence pairs were available for training purposes. Besides the data sets of the Supplied Data Track (cf. Section 2.4), all training data were preprocessed using in-house morphological analysis tools. Details of the utilized corpus are given in Table 2 , where word token refers to the number of words in the corpus and word type refers to the vocabulary size.", | |
| "cite_spans": [ | |
| { | |
| "start": 239, | |
| "end": 243, | |
| "text": "[18]", | |
| "ref_id": "BIBREF17" | |
| }, | |
| { | |
| "start": 405, | |
| "end": 406, | |
| "text": "3", | |
| "ref_id": "BIBREF2" | |
| } | |
| ], | |
| "ref_spans": [ | |
| { | |
| "start": 879, | |
| "end": 886, | |
| "text": "Table 2", | |
| "ref_id": "TABREF3" | |
| } | |
| ], | |
| "eq_spans": [], | |
| "section": "Basic Travel Expressions Corpus", | |
| "sec_num": "2.3.1." | |
| }, | |
| { | |
| "text": "Besides the parallel text corpus, we also used in-house NLP tools (tagger, chunker, parser) for the preprocessing of the training data and additional knowledge resources like bilingual dictionaries and a thesaurus [19] as summarized in Table 1.", | |
| "cite_spans": [ | |
| { | |
| "start": 214, | |
| "end": 218, | |
| "text": "[19]", | |
| "ref_id": null | |
| } | |
| ], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "NLP Tools and Dictionaries", | |
| "sec_num": "2.3.2." | |
| }, | |
| { | |
| "text": "We took part in the following translation tasks using the manual transcriptions as the input of the MT engines.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Track Participation", | |
| "sec_num": "2.4." | |
| }, | |
| { | |
| "text": "Translation Direction: (JE) Japanese-to-English (CE) Chinese-to-English Data Track:", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Track Participation", | |
| "sec_num": "2.4." | |
| }, | |
| { | |
| "text": "(C) C-STAR Track (T) Supplied+Tool Data Track (S) Supplied Data Track", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Track Participation", | |
| "sec_num": "2.4." | |
| }, | |
| { | |
| "text": "Each MT engine was trained on the resources permitted for the respective data track. In the case of the Supplied Data Track, the training data was limited to the supplied corpus only, i.e., preprocessing tools like tagger, chunker, parser and external dictionaries were not allowed. Therefore, only three element MT engines (PBHMTM, SAT, EM) were used for the Supplied Data Track. Moreover, the HPAT engine for CE could not be prepared in time for the workshop. In addition, the SELECTOR engine was tuned to obtain the best WER performance as described in Section 2.2. Therefore, not all MT engines were used for the official run submissions. Table 3 gives an overview of the used MT engines for each data track, whereby the respective priority orders are given in Table 4 .", | |
| "cite_spans": [], | |
| "ref_spans": [ | |
| { | |
| "start": 643, | |
| "end": 650, | |
| "text": "Table 3", | |
| "ref_id": null | |
| }, | |
| { | |
| "start": 765, | |
| "end": 772, | |
| "text": "Table 4", | |
| "ref_id": "TABREF4" | |
| } | |
| ], | |
| "eq_spans": [], | |
| "section": "Track Participation", | |
| "sec_num": "2.4." | |
| }, | |
| { | |
| "text": "For the automatic evaluation of the MT outputs, we used the online evaluation server 4 , whereby the following scoring metrics were used:", | |
| "cite_spans": [ | |
| { | |
| "start": 85, | |
| "end": 86, | |
| "text": "4", | |
| "ref_id": "BIBREF3" | |
| } | |
| ], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Evaluation", | |
| "sec_num": "3." | |
| }, | |
| { | |
| "text": "\u2022 BLEU: the geometric mean of n-gram precision for the translation results found in reference translations. \u2022 NIST: a variety of BLEU using the arithmetic mean of weighted n-gram precision. \u2022 METEOR which scores unigram matches using different criteria (exact matching, stemmed matching, and synonym matching). 4 http://penance.is.cs.cmu.edu/iwslt2005/eval servers.html Table 3: MT engines used for official run submissions MT JE CE engine C T S C T S SAT \u2022 Word Error Rate (WER) which penalizes edit operations for the translation output against reference translations. \u2022 Position independent WER (PER) which penalizes without considering positional disfluencies. \u2022 GTM which measures the similarity between texts by using a unigram-based F-measure.", | |
| "cite_spans": [ | |
| { | |
| "start": 311, | |
| "end": 312, | |
| "text": "4", | |
| "ref_id": "BIBREF3" | |
| } | |
| ], | |
| "ref_spans": [ | |
| { | |
| "start": 370, | |
| "end": 467, | |
| "text": "Table 3: MT engines used for official run submissions MT JE CE engine C T S C T S SAT", | |
| "ref_id": null | |
| } | |
| ], | |
| "eq_spans": [], | |
| "section": "Evaluation", | |
| "sec_num": "3." | |
| }, | |
| { | |
| "text": "In contrast to WER/PER, higher BLEU/NIST/METEOR/ GTM scores indicate better translations. Besides for GTM (one reference translation), up to 16 human reference translations were used for the automatic scoring.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Evaluation", | |
| "sec_num": "3." | |
| }, | |
| { | |
| "text": "In addition, the translation quality was judged by an English native speaker based on the fluency and adequacy criteria, whereby fluency refers to the degree to which the translation is well-formed according to the grammar of the target language and adequacy refers to the degree to which the translation communicates the information present in the reference output. The evaluator assigns a fluency score in the range from 5 (=flawless) to 1 (=incomprehensible) and an adequacy score in the range from 5 (=all information) to 1 (=none of it) to each translation. The system scores for fluency and adequacy are obtained as the average of all sentence scores.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Evaluation", | |
| "sec_num": "3." | |
| }, | |
| { | |
| "text": "The evaluation results of the official run submissions for the IWSLT 2005 translation task are summarized in Table 5 . The best performing system for each of the translation directions is marked with bold-face.", | |
| "cite_spans": [], | |
| "ref_spans": [ | |
| { | |
| "start": 109, | |
| "end": 116, | |
| "text": "Table 5", | |
| "ref_id": "TABREF5" | |
| } | |
| ], | |
| "eq_spans": [], | |
| "section": "Official Run Submissions", | |
| "sec_num": "3.1." | |
| }, | |
| { | |
| "text": "At submission time, 541K sentence pairs of the same domain as the supplied corpus were available for JE, but only 172K sentence pairs for CE. These were used for the official run submissions of the C-STAR track (cf. Table 5 ). Afterwards the CE corpus was extended to 541K and additional runs for CE using 541K sentence pairs and JE using 172K sentence pairs were evaluated using the online evaluation server. In order to illustrate the similarity between the training data used for the respective data tracks and the test data, the language perplexity 5 (PPL) of the source language input and target language translations is summarized in Table 6 . In addition, the average sentence length in words, the total word counts and the percentage of out-of-vocabulary (OOV) words of our official run submissions are given.", | |
| "cite_spans": [], | |
| "ref_spans": [ | |
| { | |
| "start": 216, | |
| "end": 223, | |
| "text": "Table 5", | |
| "ref_id": "TABREF5" | |
| }, | |
| { | |
| "start": 640, | |
| "end": 647, | |
| "text": "Table 6", | |
| "ref_id": "TABREF6" | |
| } | |
| ], | |
| "eq_spans": [], | |
| "section": "Effects of Training Data Size", | |
| "sec_num": "3.2." | |
| }, | |
| { | |
| "text": "The results show different tendencies for CE and JE. An analysis of the data sets revealed, that the increase in language perplexity for the CE data tracks is due to Chinese text style differences between the utilized training corpus and the IWSLT05 test set. ", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Effects of Training Data Size", | |
| "sec_num": "3.2." | |
| }, | |
| { | |
| "text": "For JE, a large difference in system performance can be seen for the Supplied Data Track and the Supplied+Tools Data Track. This gain is partially due to the usage of morphological analyzing tools that normalize variations of surface words, e.g., usage of hiragana or kanji in Japanese. In order to investigate the effects of prepocessing tools, we compared the Supplied Data Track submission (selection of three element MT outputs) with the SELECTOR output when applied to the respective three MT engines (3MT = SAT, PBHMTM, EM) of the Supplied+Tool Data Track. The evaluation results summarized in Table 7 show that a medium improvement of 3.5% in WER is achieved for JE. However, word segmentation differences and the lower coverage of our inhouse Chinese tagging tool resulted in a degradation in performance for the CE task. The effects of other NLP tools, like chunker and parser, cannot be compared directly to the results obtained for the Supplied Data Track. However, they are important in our hybrid approach, because they cover morpho-syntactic information that cannot be obtained from plain text and thus lead to the generation of translation hypotheses with quite different characteristics. ", | |
| "cite_spans": [], | |
| "ref_spans": [ | |
| { | |
| "start": 600, | |
| "end": 607, | |
| "text": "Table 7", | |
| "ref_id": "TABREF7" | |
| } | |
| ], | |
| "eq_spans": [], | |
| "section": "Effects of NLP tools", | |
| "sec_num": "3.3." | |
| }, | |
| { | |
| "text": "In order to investigate the effects of our hybrid approch, we compared the SELECTOR output towards the system performance of each element MT engine (cf. Table 8 ).", | |
| "cite_spans": [], | |
| "ref_spans": [ | |
| { | |
| "start": 153, | |
| "end": 160, | |
| "text": "Table 8", | |
| "ref_id": null | |
| } | |
| ], | |
| "eq_spans": [], | |
| "section": "Effects of Multi-Engine Approach", | |
| "sec_num": "3.4." | |
| }, | |
| { | |
| "text": "The results of the element MT engines show large variations in the WER scores, whereby the SMT engines performed best. However, this does not mean that translation hypotheses produced by EBMT engines are not useful at all. Table 9 gives the percentage of translation hypotheses of each MT engine selected for the official run submissions.", | |
| "cite_spans": [], | |
| "ref_spans": [ | |
| { | |
| "start": 223, | |
| "end": 230, | |
| "text": "Table 9", | |
| "ref_id": "TABREF9" | |
| } | |
| ], | |
| "eq_spans": [], | |
| "section": "Effects of Multi-Engine Approach", | |
| "sec_num": "3.4." | |
| }, | |
| { | |
| "text": "Comparing the results of the best performing MT engines (marked in bold-face in Table 8 ) for all data tracks towards the respective SELECTOR outputs reveales that the selection algorithm outperforms all element MT engines (cf. Table 10 ) gaining 4-5% in WER for the C-STAR Track submissions and even up to 10% in WER for the Supplied+Tools Data Track.", | |
| "cite_spans": [], | |
| "ref_spans": [ | |
| { | |
| "start": 80, | |
| "end": 87, | |
| "text": "Table 8", | |
| "ref_id": null | |
| }, | |
| { | |
| "start": 228, | |
| "end": 236, | |
| "text": "Table 10", | |
| "ref_id": "TABREF1" | |
| } | |
| ], | |
| "eq_spans": [], | |
| "section": "Effects of Multi-Engine Approach", | |
| "sec_num": "3.4." | |
| }, | |
| { | |
| "text": "In order to investigate in an upper boundary for the proposed selection algorithm, we performed and \"oracle\" translation experiment. Each input sentence was translated by all element MT engines and the translation hypothesis with the lowest WER (compared to the reference translations) was output as the translation, i.e., the ORACLE system simulates an optimal selection method for the given element MT engines based on the WER criterion.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Effects of Multi-Engine Approach", | |
| "sec_num": "3.4." | |
| }, | |
| { | |
| "text": "The comparison of the achieved gains 6 of the SELEC- 6 The WER figures in parantheses represent the gain of the respective TOR and the ORACLE method given in Table 10 reveals that despite a significant improvement in the overall system performance, the proposed SMT-based selection method still underachieved its task.", | |
| "cite_spans": [ | |
| { | |
| "start": 37, | |
| "end": 38, | |
| "text": "6", | |
| "ref_id": "BIBREF5" | |
| }, | |
| { | |
| "start": 53, | |
| "end": 54, | |
| "text": "6", | |
| "ref_id": "BIBREF5" | |
| } | |
| ], | |
| "ref_spans": [ | |
| { | |
| "start": 158, | |
| "end": 166, | |
| "text": "Table 10", | |
| "ref_id": "TABREF1" | |
| } | |
| ], | |
| "eq_spans": [], | |
| "section": "Effects of Multi-Engine Approach", | |
| "sec_num": "3.4." | |
| }, | |
| { | |
| "text": "In this paper, a hybrid corpus-based approach to spoken language translation was evaluated on the IWSLT 2005 translation task for Japanese-to-English and Chinese-to-English. Each input sentence was translated using up to eight MT engines, whereby the best translation was selected based on statistical models. High performances were achieved for both translation directions. The analysis of the evaluation results revealed, that:", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Conclusion", | |
| "sec_num": "4." | |
| }, | |
| { | |
| "text": "For this years' workshop, we randomly divided the training corpus into three subsets (Corpus i ; 1\u2264i\u22643) and trained three different translation and language model pairs on all pairwise combinations of the subsets (Corpus 1 \u222a Corpus 2 , Corpus 1 \u222a Corpus 3 , Corpus 2 \u222a Corpus 3 ).2 We used a development set of the supplied corpus and sorted the MT engines according to an automatic evaluation score (WER, cf. Section 3).", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "http://www.c-star.org/", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "We used the SRI Language Modeling Toolkit (version 1.4.4), which is available at http://www.speech.sri.com/projects/srilm/download.html.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "", | |
| "sec_num": null | |
| } | |
| ], | |
| "back_matter": [], | |
| "bib_entries": { | |
| "BIBREF0": { | |
| "ref_id": "b0", | |
| "title": "A framework of a mechanical translation between japanese and english by analogy principle", | |
| "authors": [ | |
| { | |
| "first": "M", | |
| "middle": [], | |
| "last": "Nagao", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 1984, | |
| "venue": "Proc. of the International NATO Symposium on Artificial and Human Intelligence", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "173--180", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "M. Nagao, \"A framework of a mechanical translation between japanese and english by analogy principle,\" in Proc. of the International NATO Symposium on Artifi- cial and Human Intelligence. North Holland: Elsevier, 1984, pp. 173-180.", | |
| "links": null | |
| }, | |
| "BIBREF1": { | |
| "ref_id": "b1", | |
| "title": "A statistical approach to machine translation", | |
| "authors": [ | |
| { | |
| "first": "P", | |
| "middle": [], | |
| "last": "Brown", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "S", | |
| "middle": [ | |
| "D" | |
| ], | |
| "last": "Pietra", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "V", | |
| "middle": [ | |
| "D" | |
| ], | |
| "last": "Pietra", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "R", | |
| "middle": [], | |
| "last": "Mercer", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 1990, | |
| "venue": "Computational Linguistics", | |
| "volume": "16", | |
| "issue": "2", | |
| "pages": "79--85", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "P. Brown, S. D. Pietra, V. D. Pietra, and R. Mercer, \"A statistical approach to machine translation,\" Computa- tional Linguistics, vol. 16(2), pp. 79-85, 1990.", | |
| "links": null | |
| }, | |
| "BIBREF2": { | |
| "ref_id": "b2", | |
| "title": "Example-based decoding for statistical machine translation", | |
| "authors": [ | |
| { | |
| "first": "T", | |
| "middle": [], | |
| "last": "Watanabe", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "E", | |
| "middle": [], | |
| "last": "Sumita", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2003, | |
| "venue": "Proc. of MT Summit IX", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "410--417", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "T. Watanabe and E. Sumita, \"Example-based decoding for statistical machine translation,\" in Proc. of MT Sum- mit IX, New Orleans, USA, 2003, pp. 410-417.", | |
| "links": null | |
| }, | |
| "BIBREF3": { | |
| "ref_id": "b3", | |
| "title": "Foundations of Statistical Natural Language Processing. Cambridge, Massachusetts", | |
| "authors": [ | |
| { | |
| "first": "C", | |
| "middle": [], | |
| "last": "Manning", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "H", | |
| "middle": [], | |
| "last": "Sch\u00fctze", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 1999, | |
| "venue": "", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "C. Manning and H. Sch\u00fctze, Foundations of Statisti- cal Natural Language Processing. Cambridge, Mas- sachusetts: The MIT Press, 1999.", | |
| "links": null | |
| }, | |
| "BIBREF4": { | |
| "ref_id": "b4", | |
| "title": "Fast decoding and optimal decoding for machine translation", | |
| "authors": [ | |
| { | |
| "first": "U", | |
| "middle": [], | |
| "last": "Germann", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "M", | |
| "middle": [], | |
| "last": "Jahr", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "K", | |
| "middle": [], | |
| "last": "Knight", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "D", | |
| "middle": [], | |
| "last": "Marcu", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "K", | |
| "middle": [], | |
| "last": "Yamada", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2001, | |
| "venue": "Proc. of ACL", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "U. Germann, M. Jahr, K. Knight, D. Marcu, and K. Ya- mada, \"Fast decoding and optimal decoding for ma- chine translation,\" in Proc. of ACL, Toulouse, France, 2001.", | |
| "links": null | |
| }, | |
| "BIBREF5": { | |
| "ref_id": "b5", | |
| "title": "EBMT, SMT, Hybrid and More: ATR Spoken Language Translation System", | |
| "authors": [ | |
| { | |
| "first": "E", | |
| "middle": [], | |
| "last": "Sumita", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Y", | |
| "middle": [], | |
| "last": "Akiba", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "T", | |
| "middle": [], | |
| "last": "Doi", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "A", | |
| "middle": [], | |
| "last": "Finch", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "K", | |
| "middle": [], | |
| "last": "Imamura", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "H", | |
| "middle": [], | |
| "last": "Okuma", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "M", | |
| "middle": [], | |
| "last": "Paul", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "M", | |
| "middle": [], | |
| "last": "Shimohata", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "T", | |
| "middle": [], | |
| "last": "Watanabe", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2004, | |
| "venue": "Proc. of International Workshop on Spoken Language Translation", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "13--20", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "E. Sumita, Y. Akiba, T. Doi, A. Finch, K. Imamura, H. Okuma, M. Paul, M. Shimohata, and T. Watan- abe, \"EBMT, SMT, Hybrid and More: ATR Spoken Language Translation System,\" in Proc. of Interna- tional Workshop on Spoken Language Translation, Ky- oto, Japan, 2004, pp. 13-20.", | |
| "links": null | |
| }, | |
| "BIBREF6": { | |
| "ref_id": "b6", | |
| "title": "Generation of word graphs in statistical machine translation", | |
| "authors": [ | |
| { | |
| "first": "H", | |
| "middle": [ | |
| "N" | |
| ], | |
| "last": "Nicola Ueffing", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Franz", | |
| "middle": [ | |
| "Josef" | |
| ], | |
| "last": "Och", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2002, | |
| "venue": "Proc. of EMNLP", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "156--163", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "H. N. Nicola Ueffing, Franz Josef Och, \"Generation of word graphs in statistical machine translation,\" in Proc. of EMNLP, Philadelphia, USA, 2002, pp. 156-163.", | |
| "links": null | |
| }, | |
| "BIBREF7": { | |
| "ref_id": "b7", | |
| "title": "Weighted probabilistic sum model based on decision tree decomposition for text chunking", | |
| "authors": [ | |
| { | |
| "first": "Y", | |
| "middle": [], | |
| "last": "Hwang", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "H", | |
| "middle": [], | |
| "last": "Chung", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "H", | |
| "middle": [], | |
| "last": "Rim", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2003, | |
| "venue": "IJCPOL", | |
| "volume": "16", | |
| "issue": "1", | |
| "pages": "1--20", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Y. Hwang, H. Chung, and H. Rim, \"Weighted proba- bilistic sum model based on decision tree decomposi- tion for text chunking,\" IJCPOL, vol. 16(1), pp. 1-20, 2003.", | |
| "links": null | |
| }, | |
| "BIBREF8": { | |
| "ref_id": "b8", | |
| "title": "Practical approach to syntax-based statistical machine translation", | |
| "authors": [ | |
| { | |
| "first": "K", | |
| "middle": [], | |
| "last": "Imamura", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "H", | |
| "middle": [], | |
| "last": "Okuma", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "E", | |
| "middle": [], | |
| "last": "Sumita", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2005, | |
| "venue": "Proc. of Machine Translation", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "K. Imamura, H. Okuma, and E. Sumita, \"Practical ap- proach to syntax-based statistical machine translation,\" in Proc. of Machine Translation Summit X, Phuket, Thailand, 2005, p. (to appear).", | |
| "links": null | |
| }, | |
| "BIBREF9": { | |
| "ref_id": "b9", | |
| "title": "Example-based machine translation based on syntactic transfer with statistical models", | |
| "authors": [ | |
| { | |
| "first": "K", | |
| "middle": [], | |
| "last": "Imamura", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "H", | |
| "middle": [], | |
| "last": "Okuma", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "T", | |
| "middle": [], | |
| "last": "Watanabe", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "E", | |
| "middle": [], | |
| "last": "Sumita", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2004, | |
| "venue": "Proc. of COLING", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "99--105", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "K. Imamura, H. Okuma, T. Watanabe, and E. Sumita, \"Example-based machine translation based on syntactic transfer with statistical models,\" in Proc. of COLING, Geneva, Switzerland, 2004, pp. 99-105.", | |
| "links": null | |
| }, | |
| "BIBREF10": { | |
| "ref_id": "b10", | |
| "title": "Application of translation knowledge acquired by hierarchical phrase alignment for patternbased MT", | |
| "authors": [ | |
| { | |
| "first": "K", | |
| "middle": [], | |
| "last": "Imamura", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2002, | |
| "venue": "Proc. of TMI, Keihanna", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "74--84", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "K. Imamura, \"Application of translation knowledge ac- quired by hierarchical phrase alignment for pattern- based MT,\" in Proc. of TMI, Keihanna, Japan, 2002, pp. 74-84.", | |
| "links": null | |
| }, | |
| "BIBREF11": { | |
| "ref_id": "b11", | |
| "title": "Feedback cleaning of machine translation rules using automatic evaluation", | |
| "authors": [ | |
| { | |
| "first": "K", | |
| "middle": [], | |
| "last": "Imamura", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "E", | |
| "middle": [], | |
| "last": "Sumita", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Y", | |
| "middle": [], | |
| "last": "Matsumoto", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2003, | |
| "venue": "Proc. of ACL", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "447--454", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "K. Imamura, E. Sumita, and Y. Matsumoto, \"Feedback cleaning of machine translation rules using automatic evaluation,\" in Proc. of ACL, Sapporo, Japan, 2003, pp. 447-454.", | |
| "links": null | |
| }, | |
| "BIBREF12": { | |
| "ref_id": "b12", | |
| "title": "Example-based machine translation using DP-matching between word sequences", | |
| "authors": [ | |
| { | |
| "first": "E", | |
| "middle": [], | |
| "last": "Sumita", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2001, | |
| "venue": "Proc. of ACL, Workshop: Data-Driven Methods in Machine Translation", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "1--8", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "E. Sumita, \"Example-based machine translation us- ing DP-matching between word sequences,\" in Proc. of ACL, Workshop: Data-Driven Methods in Machine Translation, Toulouse, France, 2001, pp. 1-8.", | |
| "links": null | |
| }, | |
| "BIBREF13": { | |
| "ref_id": "b13", | |
| "title": "Experiments and prospects of example-based machine translation", | |
| "authors": [ | |
| { | |
| "first": "E", | |
| "middle": [], | |
| "last": "Sumita", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "H", | |
| "middle": [], | |
| "last": "Iida", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 1991, | |
| "venue": "Proc. of ACL", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "185--192", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "E. Sumita and H. Iida, \"Experiments and prospects of example-based machine translation,\" in Proc. of ACL, 1991, pp. 185-192.", | |
| "links": null | |
| }, | |
| "BIBREF14": { | |
| "ref_id": "b14", | |
| "title": "Using language and translation models to select the best among outputs from multiple mt systems", | |
| "authors": [ | |
| { | |
| "first": "Y", | |
| "middle": [], | |
| "last": "Akiba", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "T", | |
| "middle": [], | |
| "last": "Watanabe", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "E", | |
| "middle": [], | |
| "last": "Sumita", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2002, | |
| "venue": "Proc. of COLING", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "8--14", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Y. Akiba, T. Watanabe, and E. Sumita, \"Using language and translation models to select the best among outputs from multiple mt systems,\" in Proc. of COLING, Taipei, Taiwan, 2002, pp. 8-14.", | |
| "links": null | |
| }, | |
| "BIBREF15": { | |
| "ref_id": "b15", | |
| "title": "Machine Learning", | |
| "authors": [ | |
| { | |
| "first": "T", | |
| "middle": [], | |
| "last": "Mitchell", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 1997, | |
| "venue": "", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "T. Mitchell, Machine Learning. New York, USA: The McGraw-Hill Companies Inc., 1997.", | |
| "links": null | |
| }, | |
| "BIBREF16": { | |
| "ref_id": "b16", | |
| "title": "Multiple Comparison Procedures", | |
| "authors": [ | |
| { | |
| "first": "Y", | |
| "middle": [], | |
| "last": "Hochberg", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "A", | |
| "middle": [], | |
| "last": "Tamhane", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 1987, | |
| "venue": "", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Y. Hochberg and A. Tamhane, Multiple Comparison Procedures. New York, USA: Wiley, 1987.", | |
| "links": null | |
| }, | |
| "BIBREF17": { | |
| "ref_id": "b17", | |
| "title": "Creating corpora for speech-to-speech translation", | |
| "authors": [ | |
| { | |
| "first": "G", | |
| "middle": [], | |
| "last": "Kikui", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "E", | |
| "middle": [], | |
| "last": "Sumita", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "T", | |
| "middle": [], | |
| "last": "Takezawa", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "S", | |
| "middle": [], | |
| "last": "Yamamoto", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2003, | |
| "venue": "Proc. of EUROSPEECH03", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "381--384", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "G. Kikui, E. Sumita, T. Takezawa, and S. Yamamoto, \"Creating corpora for speech-to-speech translation,\" in Proc. of EUROSPEECH03, Geneve, Switzerland, 2003, pp. 381-384.", | |
| "links": null | |
| } | |
| }, | |
| "ref_entries": { | |
| "FIGREF0": { | |
| "type_str": "figure", | |
| "num": null, | |
| "uris": null, | |
| "text": "System outline" | |
| }, | |
| "FIGREF1": { | |
| "type_str": "figure", | |
| "num": null, | |
| "uris": null, | |
| "text": "SELECTOR algorithm" | |
| }, | |
| "FIGREF2": { | |
| "type_str": "figure", | |
| "num": null, | |
| "uris": null, | |
| "text": "used \u00d7 = not used for official runs N/A = output not available" | |
| }, | |
| "FIGREF3": { | |
| "type_str": "figure", | |
| "num": null, | |
| "uris": null, | |
| "text": "illustrates how the system performance improves for larger amounts of training data. Effects of variable amounts of training data" | |
| }, | |
| "FIGREF4": { | |
| "type_str": "figure", | |
| "num": null, | |
| "uris": null, | |
| "text": "Evaluation of element MT engines (WER)" | |
| }, | |
| "TABREF1": { | |
| "num": null, | |
| "type_str": "table", | |
| "html": null, | |
| "content": "<table><tr><td/><td/><td>SMT</td><td/><td/><td/><td colspan=\"2\">EBMT</td><td/></tr><tr><td/><td>SAT</td><td colspan=\"4\">PBHMTM MSEP HPATR2 HPATR</td><td>HPAT</td><td>D 3</td><td>EM</td></tr><tr><td>U nit</td><td>sentence&word</td><td>phrase</td><td>phrase</td><td>phrase</td><td>phrase</td><td>phrase</td><td>sentence</td><td>sentence</td></tr><tr><td>Coverage</td><td>wide</td><td>wide</td><td>wide</td><td>wide</td><td>wide</td><td>wide</td><td>narrow</td><td>narrow</td></tr><tr><td>Quality</td><td>excellent</td><td>good</td><td>good</td><td>good</td><td>good</td><td>good</td><td>excellent</td><td>excellent</td></tr><tr><td>Speed</td><td>modest</td><td>slow</td><td>slow</td><td>modest</td><td>modest</td><td>fast</td><td>fast</td><td>fast</td></tr><tr><td>Resources</td><td>corpus</td><td>corpus</td><td>corpus,</td><td>corpus,</td><td>corpus,</td><td>corpus,</td><td>corpus,</td><td>corpus</td></tr><tr><td/><td/><td/><td>chunker</td><td>parser</td><td>parser</td><td>parser,</td><td>thesaurus,</td><td/></tr><tr><td/><td/><td/><td/><td/><td/><td>thesaurus</td><td>bilingual</td><td/></tr><tr><td/><td/><td/><td/><td/><td/><td/><td>dictionary</td><td/></tr></table>", | |
| "text": "Features of element MT engines" | |
| }, | |
| "TABREF3": { | |
| "num": null, | |
| "type_str": "table", | |
| "html": null, | |
| "content": "<table><tr><td>data</td><td>lang</td><td colspan=\"2\">sentence count</td><td>avg.</td><td>word</td><td>word</td></tr><tr><td>track</td><td>uage</td><td>total</td><td colspan=\"3\">unique length tokens</td><td>types</td></tr><tr><td/><td colspan=\"3\">Japanese 541,665 382,446</td><td colspan=\"3\">9.6 5,564,159 40,064</td></tr><tr><td colspan=\"4\">C Chinese 172,170 87,845</td><td colspan=\"3\">6.7 1,158,039 17,570</td></tr><tr><td/><td colspan=\"3\">English 541,665 343,764</td><td colspan=\"3\">8.3 4,520,340 21,737</td></tr><tr><td/><td colspan=\"3\">Japanese 20,000 19,078</td><td colspan=\"3\">9.8 196,513 8,376</td></tr><tr><td>T</td><td>Chinese</td><td colspan=\"2\">20,000 18,668</td><td colspan=\"3\">8.1 162,808 9,040</td></tr><tr><td/><td>English</td><td colspan=\"2\">20,000 19,914</td><td colspan=\"3\">9.2 183,995 5,464</td></tr><tr><td/><td colspan=\"3\">Japanese 20,000 18,906</td><td colspan=\"3\">8.6 171,259 9,251</td></tr><tr><td>S</td><td>Chinese</td><td colspan=\"2\">20,000 19,267</td><td colspan=\"3\">8.8 176,174 8,683</td></tr><tr><td/><td>English</td><td colspan=\"2\">20,000 19,902</td><td colspan=\"3\">7.7 154,483 6,937</td></tr></table>", | |
| "text": "Training corpus statistics" | |
| }, | |
| "TABREF4": { | |
| "num": null, | |
| "type_str": "table", | |
| "html": null, | |
| "content": "<table><tr><td/><td colspan=\"2\">Priority order of MT engines</td></tr><tr><td colspan=\"2\">lang data</td><td>priority</td></tr><tr><td colspan=\"2\">uage track</td><td>order</td></tr><tr><td/><td colspan=\"2\">C EM>D3>HPATR2>SAT>PBHMTM</td></tr><tr><td>JE</td><td colspan=\"2\">T EM>D3>HPAT>HPATR2>PBHMTM>SAT>MSEP</td></tr><tr><td/><td>S EM>PBHMTM>SAT</td></tr><tr><td/><td colspan=\"2\">C EM>D3>HPATR2>HPATR>MSEP>PBHMTM>SAT</td></tr><tr><td colspan=\"3\">CE T EM>MSEP>D3>HPATR>PBHMTM>HPATR2>SAT</td></tr><tr><td/><td>S EM>PBHMTM>SAT</td></tr></table>", | |
| "text": "" | |
| }, | |
| "TABREF5": { | |
| "num": null, | |
| "type_str": "table", | |
| "html": null, | |
| "content": "<table><tr><td colspan=\"2\">lang data</td><td/><td/><td colspan=\"2\">Automatic Evaluation</td><td/><td>Subjective Evaluation \u2020</td></tr><tr><td colspan=\"3\">uage track BLEU</td><td>NIST</td><td colspan=\"2\">METEOR WER</td><td>PER</td><td>GTM</td><td>Fluency</td><td>Adequacy</td></tr><tr><td/><td>C</td><td colspan=\"2\">0.6873 10.7375</td><td>0.8102</td><td colspan=\"3\">0.2768 0.2286 0.6934</td><td>4.52</td><td>4.27</td></tr><tr><td>JE</td><td>T</td><td colspan=\"2\">0.4774 8.1720</td><td>0.6658</td><td colspan=\"3\">0.4349 0.3742 0.5520</td><td>3.68</td><td>3.33</td></tr><tr><td/><td>S</td><td colspan=\"2\">0.3744 7.7368</td><td>0.6008</td><td colspan=\"3\">0.5568 0.4570 0.4822</td><td>3.23</td><td>2.68</td></tr><tr><td/><td>C</td><td colspan=\"2\">0.5031 8.6875</td><td>0.6845</td><td colspan=\"3\">0.4389 0.3727 0.5898</td><td>4.14</td><td>3.26</td></tr><tr><td>CE</td><td>T</td><td colspan=\"2\">0.3804 6.7540</td><td>0.5819</td><td colspan=\"3\">0.5439 0.4624 0.4950</td><td>3.61</td><td>2.57</td></tr><tr><td/><td>S</td><td colspan=\"2\">0.3938 8.0004</td><td>0.6291</td><td colspan=\"3\">0.5235 0.4276 0.5533</td><td>3.54</td><td>2.73</td></tr><tr><td>\u2020</td><td/><td/><td/><td/><td/><td/></tr></table>", | |
| "text": "Evaluation results of offical run submissions This subjective evaluation is carried-out in-house according to the evaluation specifications of IWSLT." | |
| }, | |
| "TABREF6": { | |
| "num": null, | |
| "type_str": "table", | |
| "html": null, | |
| "content": "<table><tr><td/><td/><td colspan=\"2\">Run Submission Statistics</td></tr><tr><td colspan=\"2\">lang data</td><td>source</td><td>target</td></tr><tr><td colspan=\"4\">uage track PPL OOV PPL OOV words length</td></tr><tr><td/><td>C</td><td colspan=\"2\">13.2 0.3 14.3 0.1 3006 5.9</td></tr><tr><td>JE</td><td>T</td><td colspan=\"2\">19.5 1.7 23.0 0.3 2925 5.7</td></tr><tr><td/><td>S</td><td colspan=\"2\">40.6 7.1 50.7 15.8 3139 6.2</td></tr><tr><td/><td colspan=\"3\">C 165.4 1.8 17.8 0.1 3050 6.0</td></tr><tr><td>CE</td><td>T</td><td colspan=\"2\">80.5 3.8 16.6 0.4 2979 5.8</td></tr><tr><td/><td>S</td><td colspan=\"2\">52.8 3.8 39.5 15.6 3141 6.2</td></tr></table>", | |
| "text": "" | |
| }, | |
| "TABREF7": { | |
| "num": null, | |
| "type_str": "table", | |
| "html": null, | |
| "content": "<table><tr><td/><td colspan=\"2\">Effects of tagging tools</td></tr><tr><td colspan=\"3\">language data track WER</td></tr><tr><td>JE</td><td>T (3MT)</td><td>0.5221</td></tr><tr><td/><td>S</td><td>0.5568</td></tr><tr><td>CE</td><td>T (3MT)</td><td>0.5913</td></tr><tr><td/><td>S</td><td>0.5235</td></tr></table>", | |
| "text": "" | |
| }, | |
| "TABREF8": { | |
| "num": null, | |
| "type_str": "table", | |
| "html": null, | |
| "content": "<table/>", | |
| "text": "" | |
| }, | |
| "TABREF9": { | |
| "num": null, | |
| "type_str": "table", | |
| "html": null, | |
| "content": "<table><tr><td/><td colspan=\"6\">Element MT engines of selected hypotheses (%)</td></tr><tr><td>MT</td><td/><td>JE</td><td/><td/><td>CE</td><td/></tr><tr><td>engine</td><td>C</td><td>T</td><td>S</td><td>C</td><td>T</td><td>S</td></tr><tr><td>SAT</td><td>9.9</td><td>3.0</td><td>5.9</td><td>5.3</td><td>1.8</td><td>19.2</td></tr><tr><td>PBHMTM</td><td>16.4</td><td>23.3</td><td>84.4</td><td>8.9</td><td>13.7</td><td>76.1</td></tr><tr><td>MSEP</td><td>\u00d7</td><td>10.9</td><td>N/A</td><td>39.1</td><td>68.8</td><td>N/A</td></tr><tr><td>HPATR2</td><td>17.2</td><td>12.0</td><td>N/A</td><td>23.3</td><td>3.3</td><td>N/A</td></tr><tr><td>HPAT</td><td>\u00d7</td><td>17.4</td><td>N/A</td><td>N/A</td><td>N/A</td><td>N/A</td></tr><tr><td>HPATR</td><td>\u00d7</td><td>\u00d7</td><td>N/A</td><td>4.7</td><td>6.1</td><td>N/A</td></tr><tr><td>D3</td><td>12.1</td><td>19.4</td><td>N/A</td><td>11.1</td><td>2.8</td><td>N/A</td></tr><tr><td>EM</td><td>44.4</td><td>14.0</td><td>9.7</td><td>7.6</td><td>3.5</td><td>4.7</td></tr></table>", | |
| "text": "" | |
| }, | |
| "TABREF10": { | |
| "num": null, | |
| "type_str": "table", | |
| "html": null, | |
| "content": "<table><tr><td>lang data</td><td>best</td><td>SELECTOR</td><td>ORACLE</td></tr><tr><td colspan=\"2\">uage track MT</td><td>(gain)</td><td>(gain)</td></tr><tr><td>C</td><td/><td/><td/></tr></table>", | |
| "text": "Effects of SELECTOR approach (WER)" | |
| } | |
| } | |
| } | |
| } |