JP2007323476A - Mechanical translation device and computer program - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To improve final translation quality by reducing the level of ambiguity in selection of a best translation candidate from translation candidates obtained by a plurality of mechanical translation parts. <P>SOLUTION: The mechanical translation system 130 comprises mechanical translation devices 30A and 30B each of which outputs an English translation candidate to a Chinese common input sentence; and a selector 182 which receives the translation candidates, calculates, based on a determination reference using a language model group 150A and a translation model group 152 and on a determination reference using a language model group 150B and a translation model group 152B, which are obtained from parallel text corpuses 142 and 144 differed in property, both scores of each translation candidate, and selects and outputs, based on the scores, either one of the translation candidates outputted by the mechanical translation devices 30A and 30B as a translation output 160. <P>COPYRIGHT: (C)2008,JPO&INPIT

Description

  The present invention relates to a machine translation system, and more particularly to a machine translation system that selects one of hypotheses output from a plurality of translation engines for the same input.

  Machine translation (MT) is the application of a computer to the task of translating text from one natural language to another. The latest approach to machine translation is to use a massively parallel text corpus, ie a set of parallel translations of source and target languages. This approach is called “corpus-based machine translation”.

  There are two main strategies employed in corpus-based machine translation. The first is example machine translation (EBMT). The second is statistical machine translation (SMT).

  EBMT uses the corpus directly. That is, given an input expression, EBMT finds a translation example (parallel translation) having a sentence most similar to the input expression from the corpus, transforms the sentence of the target language in the parallel translation, and translates the input expression. produce.

  On the other hand, the SMT first learns a translation statistical model for translation from the corpus, and creates a large number of translation sentence candidates at the time of execution, from which the best translation according to the language model and statistical model, that is, Search for the most probable thing.

  The SMT operation framework reduces the problem of translating the source language sentence S to the target language sentence T into the following conditional probability maximization problem.

ただしここでｐ（Ｔ）は言語モデル（ＬＭ）と呼ばれる統計的なモデルであって、ターゲット言語において各単語の出現する確率を表す。Ｔハットはソース言語の文Ｓに対する翻訳である。ｐ（Ｓ｜Ｔ）は翻訳モデル（ＴＭ）と呼ばれ、ソース言語の文Ｓからターゲット言語の文Ｔが生成される確率を表す。

Here, p (T) is a statistical model called a language model (LM), and represents the probability that each word appears in the target language. T hat is a translation for sentence S in the source language. p (S | T) is called a translation model (TM) and represents a probability that a target language sentence T is generated from a source language sentence S.

The probabilities of LM and TM can be automatically learned from a monolingual corpus and a bilingual corpus, respectively, and tools for that purpose are also popular. Based on such probabilities, scores can be calculated for all translation candidates generated by machine translation. The candidate showing the highest score among the candidates is selected as the translation output. If a bilingual corpus exists, a collection of sentences in one of the languages (source language or target language) can be used as a monolingual corpus and used to calculate a language model for that language.
JP 2006-39616 A

  It is known that the performance of a machine translation system depends on the size of monolingual and bilingual training data. For example, according to a survey on the translation of Arabic newswire text into English, doubling the size of monolingual training data used for statistical language model training resulted in a BLEU score (machine translation field). , Which is an objective evaluation scale for translation, calculated by comparing the output of MT with a reference translation), shows a 0.5% improvement in the size of parallel training data. If doubled, the same improvement of 2.5% can be seen.

  This indicates that the larger the training data, the better the translation quality of the MT engine used. However, increasing the amount of training data increases the level of ambiguity of the translation model and increases the complexity of how to identify the best translation from the many translation candidates generated. Therefore, more mistranslations can occur.

  Most of the countermeasures that have been proposed so far on the problem of how to select the best translation from a large number of translation candidates use statistical techniques. Of the prior art methods, the one described in Patent Document 1 seems to be the most promising.

  FIG. 1 shows the overall structure of a machine translation system 30 disclosed in Patent Document 1. Referring to FIG. 1, the machine translation system 30 includes an MT block 42 including different MT engines 70A, 70B,..., 70M for translating an input sentence 40 into respective hypotheses, and a hypothesis output from the MT block 42. A selection block 52 for selecting one of them based on a statistical model, a parallel text corpus 44, and a training module 46 for building and training appropriate data for the MT engines 70A, 70B,. including.

  The machine translation system 30 further generates N sub-corpora 80A-80N by dividing the parallel text corpus 44 into N subsets and removing each one of the subsets (1 / N subsets) from the entire corpus. And an N-fold cross-validation model creation module 48 for generating language models 90A to 90N and translation models 92A to 92N, respectively, using the sub-corpora 80A to 80N. The language models 90A to 90N and the translation models 92A to 92N are used when selecting a translation candidate in the selection module 52.

  The selection module 52 calculates N statistical scores for each hypothesis output from the MT engines 70A, 70B,..., 70M by N sets of language models and translation models and obtains for one MT output. A multiple comparison test module 100 for determining whether the obtained statistical score is significantly higher than that for another MT output, and in response to the output of the multiple comparison test module 100, the output 54 of the machine translation system 30 And a selector 102 for selecting one of the hypotheses.

  When there is no hypothesis having a score significantly higher than that of other hypotheses, the selector 102 selects one of the hypotheses in the MT engines 70A, 70B,. The priorities are pre-calculated using an appropriate scoring technique such as BLEU and a development set corpus.

  The multiple comparison test module 100 uses a pairwise comparison test technique called “Kruskal-Wallis test” in calculating the significance of a hypothesis.

  In operation, the N-fold cross-validated model creation module 48 creates N sub-corpora 80A-80N by removing different 1 / N subsets from the parallel text corpus 44. The N-fold cross-validation model creation module 48 further creates language models and translation models sets 90A and 92A, 90B and 92B,... From the sub-corpora 80A to 80N, respectively.

  The training module 46 trains each of the MT engines 70A, 70B, ..., 70M using the parallel text corpus 44. That is, in accordance with the specifications of these MT engines, a language model and a statistical model are created by the parallel text corpus 44, and translation rules are extracted.

  When the input sentence 40 is given, each of the MT engines 70A, 70B,..., 70M translates it into a target language and outputs a hypothesis (translation result). The hypothesis is given to the multiple comparison test module 100 and the selector 102.

  For each hypothesis, the selector 102 uses the language model and translation model sets 90A and 92A, 90B and 92B,..., 90N and 92N to calculate a set of scores. After calculating all the scores, the multiple comparison test module 100 applies the Kruskal-Wallis test to the hypothetical score, and whether any MT engine has a significantly higher score than the others Check. If applicable, the multiple comparison test module 100 gives a control signal to the selector 102, selects a hypothesis that the selector 102 gives a significantly high score, and outputs it as an output 54.

  In this approach, training data (parallel text corpus 44) is randomly divided into sub-corpora 80A-80N. These are used to train statistical language models 90A-90N and translation models 92A-92N. Thus, the level of ambiguity inherent in the original corpus cannot be reduced and sometimes an undesirable output is obtained.

  Accordingly, one of the objects of the present invention is a machine translation device that selects the best translation candidate from translation candidates obtained by a plurality of machine translation units, and reduces the level of ambiguity at the time of translation. It is to provide a machine translation apparatus capable of improving a general translation quality.

  The machine translation device according to the first aspect of the present invention provides a first and a second for outputting translation candidates in a predetermined second language for a common input sentence in a predetermined first language. Machine translation means and translation candidates respectively output by the first and second machine translation means are received, and translation candidates that are determined to have higher quality than the other according to a predetermined first criterion The first determination means for outputting the information to be displayed and the translation candidates output by the first and second machine translation means, respectively, and from the other according to a predetermined second determination criterion different from the first determination criterion Based on outputs of a second determination means for outputting information indicating translation candidates determined to have excellent quality, and outputs of the first determination means and the second determination means. Input any of translation candidates output by the second machine translation means Select the translation of and a selection means for outputting.

  The first and second machine translation units receive a common input sentence in the first language and output translation candidates in the second language, respectively. The first determination means outputs information indicating translation candidates determined to have quality superior to the other according to the first determination criterion. Similarly, the second determination means outputs information indicating translation candidates determined to have a quality superior to the other according to the second determination criterion. The selection means selects and outputs one of translation candidates output from the first and second machine translation means as a translation sentence for the input based on the outputs of the first determination means and the second determination means. The determination criteria used by the first and second determination means are different from each other, and translation candidates are evaluated from different aspects. Since the selection means is selected based on these multifaceted evaluations, the ambiguity when selecting translation candidates is reduced, and the quality of the selected translation candidates is improved.

  Preferably, the first determination means calculates scores of translation candidates output from the first machine translation means and the second machine translation means by a predetermined scoring method using a predetermined first statistical model. Whether there is a significant difference between the scores calculated by the first score calculation means for the translation and the translation candidates output by the first machine translation means and the second machine translation means by the first score calculation means Information indicating translation candidates determined to have a score superior to the other with a significant difference according to the determination result of the first score determination unit and the first score determination unit First translation candidate selection means.

  The first score calculation means calculates scores of translation candidates output from the first and second machine translation means by a predetermined scoring method. If there is a significant difference between these scores, a translation candidate showing an excellent score is selected by the first candidate selection means and becomes a selection target of the selection means. On the other hand, since the translation candidate determined not only to have a high score but also to be superior with a significant difference becomes a selection target by the selection means, the quality of the finally selected translation is also improved.

  More preferably, the second determination means uses the predetermined scoring method using a predetermined second statistical model different from the first statistical model, and uses the first machine translation means and the second machine translation means. The second score calculating means for calculating the score of the translation candidate output by the second score calculating means, and the second score calculating means calculated for the translation candidate output by the first machine translation means and the second machine translation means The second score determination means for determining whether or not there is a significant difference in the score, and the determination result of the second score determination means, it is determined that it has a score superior to the other with a significant difference Second translation candidate selection means for outputting information indicating the translated translation candidates.

  The second score calculation means uses a predetermined scoring technique using a statistical model different from that used by the first score calculation means, and outputs the translation candidate scores output by the first and second machine translation means. Is calculated. If there is a significant difference between these scores, a translation candidate showing an excellent score is selected by the second translation candidate selection means and is selected by the selection means. On the other hand, since the translation candidate determined not only to have a high score but also to be superior with a significant difference becomes a selection target by the selection means, the quality of the finally selected translation is also improved. When the same processing is performed in the first score calculation means, the possibility that the quality of translation finally selected by the selection means will be further increased.

  More preferably, the machine translation device obtains the first statistical language model group of the second language from the second language sentence of the first parallel text corpus composed of the parallel translation sentences of the first and second languages. A first translation model for creating a first statistical translation model group from the second language to the first language from the first language model creating means for creating and the first parallel text corpus Creating means, wherein the first score calculating means uses a predetermined scoring technique using the first statistical language model group and the first translation model group, and the first machine translation means and the second machine translation means. Means for calculating a translation candidate score output by the machine translation means are included.

  The machine translation device includes a second statistic of a second language from a sentence of a second language of a second parallel text corpus different from the first parallel text corpus, which is composed of a bilingual sentence of the first and second languages. A second statistical translation model group from the second language to the first language from the second language model creating means for creating the static language model group and the second parallel text corpus Second translation model creation means may be further included. The second score calculation means is a translation output from the first machine translation means and the second machine translation means by a predetermined scoring method using the second statistical language model group and the second translation model group. Means for calculating a score for the candidate.

  Preferably, the machine translation apparatus further includes means for training the first machine translation means using the first parallel text corpus as learning data.

  The first machine translation means is trained using the first parallel text corpus as learning data. The translation result is evaluated not only using the statistical model obtained from the first parallel text corpus, but also using the statistical model obtained from the second parallel text corpus. Since the translation candidates are evaluated based on different criteria by these different corpora, the reliability of selection of the translation candidates is increased, and the quality of the translation sentence to be finally output is improved.

  More preferably, the machine translation apparatus further includes means for training the second machine translation means using the second parallel text corpus as learning data.

  The second machine translation means is trained using the second parallel text corpus as learning data. The translation result is evaluated not only using the statistical model obtained from the second parallel text corpus, but also using the statistical model obtained from the first parallel text corpus. Since the translation candidates are evaluated based on different criteria by these different corpora, the reliability of selection of the translation candidates is increased, and the quality of the translation sentence to be finally output is improved.

  More preferably, the machine translation apparatus further includes priority order storage means for storing priorities assigned to the first and second machine translation means. When there is no translation candidate that is determined to have statistical significance and superior quality to the other, the first and second determination means replace the information indicating the translation candidate with a condition. Information indicating that there are no satisfactory translation candidates is output. The selection means selects and outputs one of the translation candidates output from the first and second machine translation means based on the outputs of the first and second determination means and the information stored in the priority order storage means. Means for.

  When there is no translation candidate that is determined to have quality superior to the other with statistical significance, the first and second determination means include information indicating that there is no translation candidate that satisfies the condition. Output. The means for outputting can select which translation candidate is to be output using the information stored in the priority storage means even when there is no translation candidate satisfying the condition by the first and second determination means. By registering the first and second machine translation means in a desired priority order in the priority order storage means in advance, even if there is no superior quality translation candidate with a significant difference from the other, a certain degree of trust A translation candidate can be selected with certainty.

  Means for outputting, in response to the outputs of the first and second determination means being coincident to indicate one translation candidate, means for selecting the translation candidate, and the first and second Means for selecting translation candidates output from the machine translation means determined by the priority order storage means in response to the output of the determination means competing; and outputs of the first and second determination means Means for selecting one translation candidate in response to indicating one translation candidate and the other indicating the absence of a translation candidate that satisfies the condition.

  When the outputs of the first and second determination means match to indicate one translation candidate, the translation candidate is selected as an output. In this case, there is a very high possibility that the quality of translation candidates will be high. Only one determination means identifies a translation candidate that is determined to have a higher quality than the other with a significant difference, and when there is no such translation candidate in the other determination means, a translation candidate that is determined to have high quality on one side Selected. In this case, the reliability may be inferior as compared to the case where both the determination means match, but the same reliability as that of the prior art can be realized. When the outputs of both determination means are in conflict, the candidate output from the machine translation means determined by the priority storage means is selected. Even in this case, a certain degree of reliability can be expected depending on how the priorities are assigned. Therefore, the quality of the translation candidate finally selected can be improved as a whole.

  Preferably, each of the first and second machine translation means receives a plurality of machine translation engines that each output a translation candidate of the second language, and translation candidates output by the plurality of machine translation engines And means for selecting according to a predetermined criterion.

  As the first and second machine translation means, one that selects candidates selected according to a predetermined criterion from outputs of a plurality of machine translation engines is used. By using a multi-stage mechanism for selecting one of a plurality of translation candidates in this way, determination based on a plurality of criteria is performed before final selection, and the quality of the output translation result is improved. be able to.

  More preferably, the machine translation apparatus further includes priority order storage means for storing the priority order assigned to each of the machine translation engines included in the first and second machine translation means. Information indicating which machine translation engine is output is attached to each translation candidate output by the first and second machine translation means. When there is no translation candidate that is determined to have statistical significance and superior quality to the other, the first and second determination means replace the information indicating the translation candidate with a condition. Information indicating that there are no satisfactory translation candidates is output. The selection means includes any one of the machine translation engines attached to the outputs of the first and second determination means, the information stored in the priority order storage means, and the translation candidates output from the first and second machine translation means. And means for selecting and outputting one of the translation candidates output by the first and second machine translation means according to the information indicating whether the output is output.

  When there is no translation candidate that is determined to have quality superior to the other with statistical significance, the first and second determination means include information indicating that there is no translation candidate that satisfies the condition. Output. The means for outputting is the translation candidate output by any machine translation engine using the information stored in the priority storage means even when there is no translation candidate satisfying the condition by the first and second judging means. It is possible to select which translation candidate is output on the basis of whether or not it exists. By registering in advance the machine translation engine used for the first and second machine translation means with a desired priority in the priority order storage means, an excellent quality translation candidate can be obtained with a significant difference from the other. Even without it, translation candidates can be selected with a certain degree of reliability.

  More preferably, at least some of the first language sentences of the first and second parallel text corpora are common to each other, and the second language sentences of the first and second parallel text corpora are mutually Independent.

  Thus, if the sentences in the second language are independent of each other, both corpuses represent the same meaning as a whole, but the grammar, vocabulary, notation, and first language used in the sentence in the second language Are different from each other in terms of their literal translation. Thus, the statistical models obtained from the corpus having different personalities are both used, the translation candidates from the first and second machine translation means are evaluated, and one of them is selected. It is possible to evaluate the quality of the translation candidate that is finally selected.

  When the computer program according to the second aspect of the present invention is executed by a computer, the computer program causes the computer to function as any one of the machine translation devices described above. Therefore, it is possible to obtain the same effect as the machine translation device described above.

  In the following description, the same parts or devices are denoted by the same reference numerals. The name and function are also the same. Therefore, detailed description thereof will not be repeated. The following embodiment relates to translation from English to Chinese. However, the present invention is applicable to any combination of source and target languages.

  As will be described later, the translation system of the present invention is usually realized by computer hardware and computer software executed on the computer hardware. In the following description, the software structure is represented in block diagram form.

<Configuration>
FIG. 2 is a diagram showing a configuration of the Chinese-English machine translation system 130 according to the embodiment of the present invention. The machine translation system 130 is an extension of the machine translation system 30 described in the prior art. Referring to FIG. 2, machine translation system 130 includes an MT block 140 that translates input source language sentence 40 into a target language and outputs two hypotheses (translation candidates).

  The MT block 140 includes two machine translation devices 30A and 30B, both of which are configured in accordance with the machine translation system 30 described in the prior art section. Machine translation apparatuses 30A and 30B according to this embodiment include a combination of the same MT engines each including a plurality of MT engines. As described below, the language model and translation model of the machine translation device 30A are trained using a specific corpus, and the language model and translation model of the machine translation device 30B are trained using another corpus. Therefore, the output of the machine translation devices 30A and 30B is generally different for the same input sentence.

  In the present embodiment, the MT block 140 attaches a tag that identifies the MT engine that outputs the hypothesis to each hypothesis that is output. This tag is hypothesized throughout the subsequent processing. Therefore, it is possible to confirm which MT engine has output the hypothesis at any stage of the process.

  The machine translation system 130 further includes two corpora, a Chinese-English parallel text corpus 142 and 144. In this embodiment, the parallel text corpus 144 is a modification of the parallel text corpus 142. That is, referring to FIG. 3, parallel text corpus 142 and parallel text corpus 144 are generated from the same English-Japanese parallel text corpus 200. The English-Japanese parallel text corpus 200 includes a number of pairs of English sentences and their Japanese equivalents (translations), that is, parallel translations.

  Each English sentence in the text corpus 200 is translated from English to Chinese (202). A parallel text corpus 142 is obtained by pairing the Chinese obtained by translation with the original English text. On the other hand, each Japanese sentence of the English-Japanese parallel text corpus 200 is translated from Japanese to Chinese (204). The parallel text corpus 144 is obtained by replacing the Japanese text in the parallel text corpus 200 with the Chinese text obtained as a result of translation. Here, the parallel text corpora 142 and 144 are referred to as “deformed corpora”. These are semantically equivalent parallel corpora, but have different vocabulary, grammatical structure, notation, and direct translation (strength correspondence between English and Chinese) to represent the same content. Corpus.

  Machine translation system 130 further includes training modules 146 and 148 for training statistical models used in the MT engine in machine translation devices 30A and 30B, respectively. As described above, the machine translation devices 30A and 30B each include a plurality of MT engines. The training module 146 trains a model used in the MT engine in the machine translation apparatus 30A by using the parallel text corpus 142, and the training module 148 trains a model used in the MT engine in the machine translation apparatus 30B in the parallel text corpus 144. To train.

  Needless to say, this training is not performed for a rule-based MT engine that does not need to be trained.

  Further, the machine translation system 130 is configured to train the language model group 150A and the translation model group 152A using the parallel text corpus 142 in the same form as the N-fold cross-validation model creation module 48 shown in FIG. A double cross-validation model creation module 162 and an N-fold cross-validation model creation module 164 for training the language model group 150B and the translation model group 152B using the parallel text corpus 144 are included.

  In the present embodiment, the MT engine in the MT block 140 is trained using the parallel text corpus 142 and 144 used for training by the N-fold cross-validation model creation module 164. However, the present invention is not limited to such an embodiment, and different corpora may be used for this training.

  Because training corpus 142 and 144 with the same content and different sources are used, the resulting language models and translation models focus on specific linguistic aspects, both of which are translated as described below. Testing the significance of candidate quality differences can reduce the ambiguity of selection of translation candidates compared to using a model trained on a corpus obtained by integrating all deformation corpora.

  The machine translation system 130 further selects one of the outputs of the MT block 140 based on the priority table 156 that stores the priorities of the MT engines used in the MT block 140 and the score difference between the hypotheses. Output of the MT block 140 based on the rule table 158 that stores the rules for and the significance of the difference in score between hypotheses, the rules stored in the rule table 158, and the priorities stored in the priority table 156 A selection module 154 for selecting one of the translation candidates to be translated.

  The contents of the priority table 156 are prepared by evaluating the output of the MT engine in the MT block 140 according to an evaluation criterion such as BLEU or word error rate using a test set sentence. The higher the MT engine rating, the higher the priority.

  The selection module 154 determines, as an output 160, one of the outputs of the MT block 140 according to the determination of the modified corpus-based verification processing module 180 and the verification corpus-based verification processing module 180 for determining which hypothesis should be selected. And a selector 182 to select.

  FIG. 4 is a block diagram of the verification processing module 180 using the modified corpus shown in FIG. Referring to FIG. 4, the modified corpus-based verification processing module 180 uses the N language model groups 150A and the N translation model groups 152A trained using the parallel text corpus 142 (FIG. 2), and performs MT processing. Language models trained using first and second scoring modules 220 and 222 for outputting score vectors scoring the two outputs of block 140, respectively, and parallel text corpus 144 (FIG. 2) The third and fourth scoring modules 224 and 226 for scoring the two outputs of the MT block 140 and outputting the respective scoring vectors using the 150B and the translation model group 152B, And from the second scoring modules 220 and 222 In order to check whether one of the two scoring vectors given is significantly higher than the other using the Kruskal-Wallis test and output the result, as in the prior art The first significant difference test module 230 of the first and the second scoring modules 224 and 226 output from the third and fourth scoring modules 224 and 226 are similar to each other in terms of whether the score is significantly higher than the other. And a second significance test module 232 for outputting the result. If there is a scoring vector having a score that is significantly higher than the other, the significance test modules 230 and 232 provide information indicating translation candidates corresponding to the vector to the table lookup module 240. If there is no such, a signal is provided to the table lookup module 240 indicating that none meets the condition.

  In the present embodiment, the product of the score LM calculated from the language model for the hypothesis H and the score TM calculated from the translation model is used as the score to be evaluated by the Kruskal-Wallis test. This product is calculated for each set of language model and translation model, and the scoring vector is created by using the values calculated for all the sets as vectors, and used for evaluation in the Kruskal-Wallis test.

  The modified corpus test processing module 180 further determines whether the results from the first and second significance test modules 230 and 232 clearly indicate one of the two outputs of the MT block 140. If so, determine which output should be selected by examining the appropriate entry in the rule table 158 corresponding to the combination of results from the first and second significance test modules 230 and 232. One of the two outputs of the MT block 140 is determined based on the table lookup module 240, the output 246 of the table lookup module 240, and the contents of the priority table 156, and the selector 182 (see FIG. 2). And a final selection module 242 for outputting a signal identifying it.

  FIG. 5 shows an example of the contents of the rule table 158. Referring to FIG. 5, the top row corresponds to the output of first significant difference test module 230, and the leftmost column corresponds to the output of second significant difference test module 232.

  According to the rule table 158, hypotheses are selected by the following sequence.

  (1) A column is selected according to the output value from the first significant difference test module 230. If the output of the first significance test module 230 indicates that a significant difference has been found for the first hypothesis, the first column is selected. If a significant difference is found for the second hypothesis, the second column is selected. If no significant difference is found in the output, the third column is selected.

  (2) Similarly, a row is selected according to the output value from the second significance test module 232.

  (3) View the contents in the frame at the intersection of the selected row and column. The contents in this frame indicate which hypothesis to select. If there is a conflict between the outputs of the first and second significance test modules 230 and 232, or if there is no significant difference, the contents of the frame should output a hypothesis determined according to the priority table. It shows that.

<Operation>
The machine translation system 130 of this embodiment operates as follows. Referring to FIG. 2, it is assumed that a priority table 156, a rule table 158, a parallel text corpus 142, and a parallel text corpus 144 are prepared prior to processing.

  The machine translation system has two phases of operation. The first phase is a training phase where the MT engine in MT block 140 is trained using parallel text corpus 142 and parallel text corpus 144. The second phase is the application phase, where the input sentence is applied to the machine translation system 130 and the machine translation system 130 outputs a preferred translation as output 160.

-Training Phase In the training phase, the machine translation system 130 operates as follows. The training module 146 trains the language model and translation model of the MT engine of the machine translation apparatus 30 </ b> A using the parallel text corpus 142. Since the MT engines in the machine translation apparatus 30A are of different types, the performance of these MT engines will be different even if they are trained with the same corpus. Similarly, the training module 148 trains the language model and translation model of the MT engine in the machine translation apparatus 30B using the parallel text corpus 144.

  The set of MT engines in the machine translation apparatus 30A is the same as that in the machine translation apparatus 30B. However, both are trained by different corpora. Therefore, their performance will be different.

  After completing the training of the MT engine in the machine translation devices 30A and 30B in this manner, the MT included in the machine translation device 30A and the machine translation device 30B using a predetermined test set and a predetermined evaluation method (such as BLEU). Evaluate the engine. Then, the higher priority is assigned in order from the highest evaluation, and the priority order is stored in the priority table 156.

  Further, the language model group 150A and the translation model group 152A are trained using the parallel text corpus 142 by the N-fold cross validation model creation module 162, respectively. Similarly, the language model group 150B and the translation model group 152B are trained by the N-fold cross-validation model creation module 164 using the parallel text corpus 144, respectively.

Application Phase As described in the section of the prior art, when the input sentence 40 is given to the machine translation device 30A, the MT engine in the machine translation device 30A outputs the respective hypotheses. One of the hypotheses is selected by the selection module 52 (FIG. 1) and output as a hypothesis of the machine translation apparatus 30A.

  The same procedure is performed in the machine translation apparatus 30B, and a hypothesis is output.

  Both hypotheses are given to the selector 182 and the verification processing module 180 using a modified corpus shown in FIG.

  Referring to FIG. 4, scoring modules 220 and 222 use the language model group and translation model group created from parallel text corpus 142 (corpus A), and are the same as the processing in the conventional multiple comparison test module 100. Then, scoring is performed on the hypotheses output from the machine translation devices 30A and 30B, respectively, and the result is given to the significance test module 230 as a scoring vector. The significance test module 230 determines whether there is a significant difference between the scoring vector from the scoring module 220 and the scoring vector from the scoring module 222, and if so, scores more than the other. A value indicating a good hypothesis is provided to the table lookup module 240 and a value indicating that it did not exist is provided to the table lookup module 240.

  Similarly, the scoring modules 224 and 226 use the language model group and the translation model group created from the parallel text corpus 144 (Corpus B) to score the hypotheses output from the machine translation devices 30A and 30B, respectively. The result is given to the significant difference test module 232 as a scoring vector. The significance test module 232 determines whether there is a significant difference between the scoring vector from the scoring module 224 and the scoring vector from the scoring module 226, and if so, scores more than the other. A value indicating a good hypothesis is provided to the table lookup module 240 and a value indicating that it did not exist is provided to the table lookup module 240.

  The table lookup module 240 finds a frame corresponding to the combination of the value given from the significance test module 230 and the value given from the significance test module 232 in the rule table 158, and reads the contents. If the read content is “first”, the table lookup module 240 decides to select the hypothesis output by the machine translation device 30A, and sets the signal 246 to the final selection module 242 as the first value. If the read content is “second”, it is decided to select the hypothesis output by the machine translation device 30B, and the value of the signal 246 to the final selection module 242 is set to a second value different from the first value. If the read content is “by priority”, the value of the signal 246 is set to a predetermined third value that is different from both the first and second values.

  If the value of the signal 246 is the first or second value, the final selection module 242 outputs the signal 246 as it is to the selector 182. If the value of the signal 246 is the third value, the final selection module 242 refers to the priority table 156, and translates the hypothesis selected by the significance test module 230 (the machine translation device 30A in FIG. 2). And the translation engine (in the machine translation apparatus 30B in FIG. 2) that outputs the hypothesis selected by the significance test module 232, according to the priority in the priority table 156, and selects them. A signal indicating the machine translation device including the signal is supplied to the selector 182.

  Referring to FIG. 2, selector 182 selects one of hypotheses output from machine translation devices 30A and 30B according to the signal given from final selection module 242 of verification processing module 180 using the modified corpus, and outputs 160 Output as.

  In the machine translation system 130 according to the present embodiment, two Chinese-English parallel text corpuses having different personalities are created from one English-Japanese parallel text corpus. In this case, the “character” means the direct translation (whether the translation is faithfully translated into the original sentence or whether it is translated), the grammatical tendency and vocabulary of the translation, the difference in notation, etc. Differences in various grammatical aspects of sentences. For example, in machine translation using a model trained using a corpus having many sentences with high literal translation, the obtained translation may not be a natural translation, but it is often not a complete error. On the other hand, machine translation using a model trained with a corpus that has a low level of literal translation and a large number of translated sentences may result in a very natural translation. Often not available.

  In particular, in the present embodiment, as shown by reference numeral 202 in FIG. 3, the original English structure is maintained relatively well when translated from English to Chinese, whereas it is shown by reference numeral 204. Thus, when Japanese is translated into Chinese, Japanese expressions are reflected in the Chinese translation, and a Chinese translation that is different from the original English structure tends to be obtained. Therefore, the parallel text corpora 142 and 144 shown in FIG. 3 are corpora having different “characters”.

  In the present embodiment, as described in the prior art in the machine translation devices 30A and 30B, the translations obtained by using the respective corpora are those that show better scores than other translations. It is. However, this is only an evaluation by a single corpus. On the other hand, the selection module 154 again determines which one of the hypotheses output from the MT block 140 is to be selected. The evaluation here is performed using not only models obtained from one corpus but also models obtained from both corpora. That is, although the corpora express the same meaning as a whole, hypotheses are evaluated from different aspects by using models obtained from two corpuses that are greatly different in the character of the translation.

  According to the evaluation result, if there is a hypothesis having a score that is significantly higher than that of the other in any aspect, that hypothesis is selected. The hypothesis quality in this case can be expected to be very high. If the significance is confirmed in one evaluation but the significance is not confirmed in the other evaluation, a hypothesis in which the significance is confirmed only in one evaluation is selected. In this case, the reliability of the hypothesis is slightly lowered, but a certain level of reliability can be expected. If such a hypothesis does not exist (if the evaluation results are mutually conflicting or contradictory, or if no significance has been confirmed), a favorable result will be obtained according to the evaluation results that were previously performed using the test set. The hypothesis by the MT engine determined to be obtained is selected. In this case, the evaluation of the translation candidate cannot be used for the selection, but the hypothesis is selected by a more general evaluation criterion (ie information about the general superiority or inferiority of the translation engine used for the translation) The results can be expected to have a certain level of reliability.

  That is, in this embodiment, hypotheses are evaluated using corpus having different personalities, and the hypotheses are selected by combining the results. Therefore, as compared with the case where the same corpus is divided without setting any standard as in the system described in Patent Document 1, and the hypothesis is evaluated with a model obtained from the corpus, the obtained result is more reliable. It seems that the nature becomes high. In fact, as will be described later, good results have been obtained in experiments conducted in accordance with the system configuration of the present embodiment.

<Experimental result>
FIG. 6 shows the results of an experiment performed using the configuration of the above-described embodiment. In the experiment, seven types of MT engines were used. These MT engines include four types of SMT (MT1 to MT4) and three types of EBMT (MT5 to MT7). The corpus used was an English-Japanese corpus of conversation in the travel domain, called BTEC, created by the applicant. Based on the English-Japanese parallel text corpus, two English-English parallel text corpora were obtained by extracting only English sentences and independently translating them into Chinese twice, unlike the method shown in FIG. . Here, for convenience, one is called an original corpus and the other is called a modified corpus, which are represented by BTEC (O) and BTEC (V), respectively. In addition, MT trained by BTEC (O) or BTEC (V) is also denoted by (O) or (V), respectively.

  In FIG. 6, each of the seven lines in which the leftmost column is “SMT” and “EBMT” is a single MT as the machine translation device 30A and the machine translation device 30B in the MT block 140 shown in FIG. The BLEU score when using the engine and configuring the machine translation system 130 and its improvement from the BLEU score by the same MT engine trained using a corpus obtained by merging BTEC (O) and BTEC (V) Rate. In this case, selection of a hypothesis within the machine translation apparatus 30A or 30B is not necessary, and thus is not performed.

  For example, the result of mutually testing either the hypothesis from MT1 (O) or the hypothesis from MT1 (V) using the language model and translation model obtained from BTEC (O) and BTEC (V). When used, the BLEU score was 0.5010. This is a value indicating an improvement of 3.8% compared to the case where the merged BTEC (O) and BTEC (V) are used.

  As can be seen from FIG. 6, in any of MT1 to MT7, the BLEU score obtained by the embodiment of the present invention is a better value than the comparative example. This fully demonstrates the effectiveness of the present invention.

  The bottom row of FIG. 6 shows the results obtained by using all of MT1 to MT7 simultaneously in each of the machine translation devices 30A and 30B. Each MT engine of the machine translation apparatus 30A was trained using BTEC (O), and each MT engine of the machine translation apparatus 30B was trained using BTEC (V).

  The result was a BLEU score of 0.5765 as shown on the right side of the bottom line of FIG. This result is a considerably high numerical value as in the case where MT1 to MT7 are used individually. The BLEU score showed a 4.2% improvement over the use of a corpus that merged BTEC (O) and BTEC (V).

<Realization by computer>
As described above, the machine translation system 130 of this embodiment includes computer hardware, a computer program executed by the computer hardware, a corpus, a language model, a translation model, and the like stored in the computer hardware. Realized with data. FIG. 7 shows the external appearance of the computer system 330, and FIG. 8 shows the internal configuration of the computer system 330.

  Referring to FIG. 7, the computer system 330 includes a computer 340 having a memory port 352 and a DVD (Digital Versatile Disc) drive 350, a keyboard 346, a mouse 348, and a monitor 342.

  Referring to FIG. 8, in addition to the memory port 352 and the DVD drive 350, the computer 340 boots up a CPU (Central Processing Unit) 356, a bus 366 connected to the CPU 356, the memory port 352, and the DVD drive 350, and A read only memory (ROM) 358 that stores programs and the like, and a random access memory (RAM) 360 that is connected to the bus 366 and stores program instructions, system programs, work data, and the like.

  Although not shown here, the computer 340 may further include a network adapter board that provides a connection to a local area network (LAN).

  A computer program for causing the computer system 330 to operate as the machine translation system 130 is stored in the DVD 362 or the portable memory 364 attached to the DVD drive 350 or the memory port 352, and further transferred to the hard disk 354. Alternatively, the program may be transmitted to the computer 340 through a network (not shown) and stored in the hard disk 354. The program is loaded into the RAM 360 when executed. The program may be loaded into the RAM 360 directly from the DVD 362, from the portable memory 364, or via a network.

  This program includes a plurality of instructions for causing the computer 340 to operate as the machine translation system 130 according to this embodiment. Some of the basic functions required to perform this operation are provided by operating system (OS) or third party programs running on the computer 340 or various toolkit modules installed on the computer 340. Therefore, this program does not necessarily include all functions necessary for realizing the system and method of this embodiment. This program operates as the machine translation system 130 described above by calling an appropriate function or “tool” available in the computer 340 in a controlled manner so as to obtain a desired result. Need only contain instructions to execute The operation of computer system 330 is well known and will not be repeated here.

<Possible modification>
In the embodiment described above, the machine translation devices included in the MT block 140 are the two machine translation devices 30A and 30B. However, the present invention is not limited to such an embodiment. The MT block 140 may be provided with three or more machine translation devices. In that case, the comparison in the verification processing module 180 by the modified corpus shown in FIG. 4 cannot be performed using the method in the above-described embodiment as it is. However, all the arbitrary two combinations of the machine translation apparatus are processed in the test processing module 180 using the modified corpus described above, and hypotheses are selected based on the total number of times each hypothesis is selected in those processes. In this way, implementation in a program is relatively simple with only a double loop.

  Further, in the above embodiment, the modified corpus used in the selection module 154 is also used for training each MT engine in the MT block 140. In this way, each hypothesis output from the MT block 140 strongly reflects the character of one corpus. For this reason, the translation results are expected to be considerably different from each other, and it is likely that a significant difference occurs in the determination in the selection module 154. However, the present invention is not limited to such an embodiment. The corpus for training the MT block 140 and the corpus used for the determination in the selection module 154 may be completely different. However, among the corpora for training the MT block 140, for example, the machine translation device 30A and the machine translation device 30B must have different personalities as described above, and are selected. The corpus used for the determination in the module 154 also needs to have different personalities.

  As is apparent from the above, the number of corpuses for training the MT engine in the MT block 140 and the number of corpuses used for evaluation in the selection module 154 need not match each other. In this embodiment, since there is a limit to the corpus that can be used, both are made common, but a completely different number of corpuses may be used for both.

  Further, in the above-described embodiment, as the parallel text corpora 142 and 144, a Chinese-English corpus having a common English sentence and a different Chinese sentence is used. That is, the English sentences are consistent in both. When such a corpus is used, the domains covered by the corpus are almost the same. Therefore, the translation output from the MT block 140 is likely to clearly reflect the character of the corpus. Therefore, the evaluation from various aspects performed in the selection module 154 will be meaningful.

  However, the present invention is not limited to one using a pair of corpora in which one of the parallel translations is matched. Parallel text corpora with the same language combination and different personalities will have various aspects of evaluation compared to using a single corpus, regardless of the combination. It can be expected that the final result will be better.

  In the above-described embodiment, the hypothesis score is calculated using both the language model group and the translation model group. However, the present invention is not limited to such an embodiment. For example, the hypothesis score may be calculated using only the language model group or only the translation model group.

  The embodiment disclosed herein is merely an example, and the present invention is not limited to the above-described embodiment. The scope of the present invention is indicated by each claim in the claims after taking into account the description of the detailed description of the invention, and all modifications within the meaning and scope equivalent to the wording described therein are intended. Including.

It is a block diagram of the machine translation system 30 by a prior art. It is a figure which shows the structure of the machine translation system 130 which concerns on one embodiment of this invention. It is a block diagram which shows the example of the process in which the parallel text corpus 142 and 144 used in one embodiment of this invention are produced. It is a block diagram which shows the detail of a structure of the test | inspection processing module 180 by the deformation corpus in the machine translation system 130 concerning one embodiment of this invention. It is a figure which shows the detail of a structure of the rule table 158 used with the machine translation system 130 which concerns on one embodiment of this invention in a table format. It is a figure which shows the improvement of evaluation with respect to the translation sentence obtained by the machine translation system 130 which concerns on one embodiment of this invention in a table format. 2 is a front view of a computer system for realizing a machine translation system 130. FIG. It is a block diagram which shows the internal structure of the computer system 330 shown in FIG.

Explanation of symbols

30, 130 Machine translation system 30A, 30B Machine translation device 140 MT block 142, 144 Parallel text corpus 156 Priority table 158 Rule table 180 Modified corpus-based test processing module 220, 222, 224, 226 Scoring module 230, 232 Significant difference Test module

Claims (13)

First and second machine translation means for outputting translation candidates in a predetermined second language for a common input sentence in a predetermined first language;
The translation candidates output by the first and second machine translation units are received, and information indicating translation candidates determined to have a quality superior to the other according to a predetermined first determination criterion is output. First determination means for
The translation candidates output by the first and second machine translation units are received, respectively, and determined to have a quality superior to the other according to a predetermined second criterion different from the first criterion. Second determination means for outputting information indicating translation candidates
Based on the outputs of the first determination means and the second determination means, one of the translation candidates output by the first and second machine translation means is selected and output as a translation for the input. A machine translation device including selection means. The first determination means includes
First score calculation for calculating scores of translation candidates output from the first machine translation unit and the second machine translation unit by a predetermined scoring method using a predetermined first statistical model Means,
A first score for determining whether or not there is a significant difference between the scores calculated by the first score calculation means for the translation candidates output by the first machine translation means and the second machine translation means; A score determination means,
First translation candidate selection means for outputting information indicating translation candidates determined to have a score superior to the other with a significant difference according to the determination result of the first score determination means The machine translation apparatus according to claim 1. The second determination means includes
By using a predetermined scoring method using a predetermined second statistical model different from the first statistical model, the scores of translation candidates output from the first machine translation unit and the second machine translation unit are obtained. A second score calculating means for calculating;
A second for determining whether or not there is a significant difference between the scores calculated by the second score calculation means for the translation candidates output by the first machine translation means and the second machine translation means; A score determination means,
Second translation candidate selection means for outputting information indicating translation candidates determined to have a score superior to the other with a significant difference based on the determination result of the second score determination means The machine translation apparatus according to claim 1 or 2. A first statistical language model group for the second language is created from the second language sentence of the first parallel text corpus comprising the parallel translation sentences of the first and second languages. Language model creation means,
First translation model creating means for creating a first statistical translation model group from the second language to the first language from the first parallel text corpus;
The first score calculation means uses the first machine translation means and the second machine translation by a predetermined scoring method using the first statistical language model group and the first translation model group. 4. The machine translation apparatus according to claim 1, further comprising means for calculating a score of translation candidates output by the means. A second statistical of the second language from a sentence of the second language of a second parallel text corpus different from the first parallel text corpus, comprising bilingual sentences of the first and second languages. A second language model creating means for creating a language model group;
A second translation model creating means for creating a second statistical translation model group from the second language to the first language from the second parallel text corpus;
The second score calculating means uses the first machine translation means and the second machine translation by a predetermined scoring method using the second statistical language model group and the second translation model group. The machine translation apparatus according to claim 4, comprising means for calculating a score of translation candidates output by the means. 6. The machine translation apparatus according to claim 4 or 5, further comprising means for training the first machine translation means using the first parallel text corpus as learning data. The machine translation apparatus according to claim 6, further comprising means for training the second machine translation means using the second parallel text corpus as learning data. Furthermore, priority order storage means for storing priority given to the first and second machine translation means,
When there is no translation candidate that is determined to have statistical significance and superior quality to the other, the first and second determination means, instead of the information indicating the translation candidate, Output information indicating that there are no translation candidates that satisfy the condition,
The selection means selects one of the translation candidates output from the first and second machine translation means based on the output of the first and second determination means and the information stored in the priority order storage means. The machine translation apparatus according to claim 1, comprising means for selecting and outputting. The means for outputting is:
Means for selecting the translation candidate in response to the output of the first and second determining means being matched and indicating one translation candidate;
Means for selecting translation candidates output from the machine translation means determined by the priority storage means in response to conflicting outputs of the first and second determination means;
Means for selecting the certain translation candidate in response to one of the outputs of the first and second determination means indicating a translation candidate and the other indicating the absence of a translation candidate that satisfies the condition. The machine translation apparatus according to claim 8. Each of the first and second machine translation means includes:
A plurality of machine translation engines each receiving the input and outputting translation candidates for the second language;
The machine translation device according to claim 1, further comprising: means for selecting any one of translation candidates output from the plurality of machine translation engines according to a predetermined determination criterion. And further comprising priority order storage means for storing priorities assigned to each of the machine translation engines included in the first and second machine translation means,
Each of the translation candidates output by the first and second machine translation means has information indicating which machine translation engine is output,
When there is no translation candidate that is determined to have statistical significance and superior quality to the other, the first and second determination means, instead of the information indicating the translation candidate, Output information indicating that there are no translation candidates that satisfy the condition,
The selection means is attached to the output of the first and second determination means, the information stored in the priority order storage means, and the translation candidate output from the first and second machine translation means, 11. The machine translation according to claim 10, further comprising means for selecting and outputting one of translation candidates output by the first and second machine translation means according to information indicating which machine translation engine output. apparatus. At least some of the sentences in the first language of the first and second parallel text corpora are common to each other;
6. The machine translation device according to claim 5, wherein sentences in the second language of the first and second parallel text corpora are independent of each other. A computer program that, when executed by a computer, causes the computer to function as the machine translation device according to any one of claims 1 to 12.

Images