JP2013054608A - Translation device, method and program, and translation model learning device, method and program - Google Patents

Abstract

PROBLEM TO BE SOLVED: To perform high-speed and high-accuracy translation in which words are easily associated.SOLUTION: A learning preprocessing part 22 generates translation intermediate language in which translation source language is replaced with vocabularies of translation destination language using a word order close to the translation source language. A first-translation learning part 24 learns a first-translation model 34 for translating the translation source language into an intermediate translation sentence using a parallel corpus of the translation source language and the translation intermediate language. A latter-translation learning part 26 learns a latter-translation model 36 for translating the intermediate translation sentence into the translation destination language using a parallel corpus of the translation intermediate language and the translation destination language. A first-translation part 54 refers to the first-translation model 34 and translates an input sentence into the intermediate translation sentence. A latter-translation part 56 refers to the latter-translation model 36 and translates the intermediate translation sentence into a sentence of the translation destination language.

Description

  The present invention relates to a translation apparatus, method, and program, and a translation model learning apparatus, method, and program.

  Conventionally, a sentence in an input language (hereinafter referred to as “translation source language”) is machine-translated into a sentence in an output side language (hereinafter referred to as “translation destination language”). Difficulties in word rearrangement due to differences in word order between different languages are a common issue for all machine translation technologies. In particular, in statistical machine translation technology based on learning from a translation corpus as a translation method, word rearrangement causes an increase in the amount of calculation, so it is practical to search for word rearrangement efficiently. This is an important issue. In conventional statistical translation technology, especially translation technology based on phrases that can be said to be representative examples, the constraints when phrases are rearranged are statistically modeled, and beam search and rearrangement search range restrictions are limited. Search efficiency is improved by such an approximate search means (see, for example, Non-Patent Document 1 and Non-Patent Document 2).

On the other hand, since the number of essential sorting candidates is enormous even by such a method, syntax (or syntactic) analysis for the sentence to be translated (input sentence) and sorting for the obtained syntax tree are performed. A pre-processing method called “pre-ordering” in which words are rearranged before translation processing by the processing and brought closer to the word order of the translation destination language may be used. According to this method, in principle, the factorial n! The number of rearrangement candidates can be reduced to 2 ^n-1 when the syntax tree is a binary tree. Furthermore, very high translation accuracy may be obtained by using information related to the syntax (see, for example, Non-Patent Document 3).

  Other techniques for using information about syntax for statistical machine translation include input sentences such as translation from a syntax tree to a string, translation from a string to a syntax tree, and translation from a syntax tree to a syntax tree. On the other hand, either or both of the output sentence side and the output sentence side are syntax trees, and a technique using the partial structure as a unit of translation is also widely known (see Non-Patent Document 4, for example). In this case, not only the source language but also the target language syntax analysis is used.

Phillip Koehn, Franz Josef Och, and Daniel Marcu, "Statistical Phrase-based Translation." In Proc. HLT- NAACL, 2003. pages 263-270. David Chiang, "Hierarchical Phrase-based Translation." Computational Linguistics, 2007. 33 (2): 201-228. Hideki Isozaki, Katsuhito Sudoh, Hajime Tsukada, and Kevin Duh, "A Simple Reordering Rule for SOV Languages." In Proc. WMT-MetricsMATR. 2010. Michel Galley, Mark Hopkins, Kevin Knight, and Daniel Marcu, "What ’s in a translation rule?” In Proc. NAACL, 2004, pages 273-280.

  However, the method described in Non-Patent Document 3 has a problem that a technique and a language resource for performing a syntax analysis of the translation source language are essential.

  Further, as described in Non-Patent Document 4, the correspondence between the syntax tree and the character string or the partial structure between the syntax trees is compared with the correspondence between the character strings because the constraints on the tree structure and the word order are mixed. As a result, there is a problem that a large translation phrase pair having no generality is extracted, and as a result, many translation rules that hardly contribute to translation of other sentences may be obtained.

  The present invention has been made in view of the above-described problems, a translation device, a method, and a program, a parallel translation model learning device, a method, and a translation device that can easily perform word association and perform high-speed and high-accuracy translation. The purpose is to provide a program.

  In order to achieve the above object, the translation apparatus of the present invention converts an input sentence described in a translation source language into a translation destination language in a word order in which each word included in the input sentence conforms to the word order of the translation source language. Pre-translation means for translating into an intermediate translation sentence replaced with a word, and the intermediate translation sentence translated by the pre-translation means by translating the word order of the intermediate translation sentence into the word order of the translation destination language And a post-translation means for translating the sentence described in the above.

  According to the translation apparatus of the present invention, the pre-translation means converts the input sentence described in the translation source language into words in the translation destination language with each word included in the input sentence remaining in the word order according to the word order of the translation source language. Translated into the replaced intermediate translated sentence, and the post-translation means replaces each word included in the input sentence with the word in the target language in the word order according to the word order in the source language. By rearranging the word order to the word order of the translation destination language, the intermediate translation sentence translated by the pre-translation means is translated into a sentence described in the translation destination language.

  In this way, in the pre-translation, the translation of the words from the source language to the intermediate translation sentence replaced with the words in the target language remains the same according to the word order of the source language, so that the rearrangement of the words is considered almost or not at all. In post-translation, translating from the intermediate translation sentence to the target language, it is not necessary to consider word replacement or not at all. It can be carried out.

  The pre-translation means translates the input sentence into the intermediate translation sentence based on a pre-translation model for translating a sentence described in the translation source language into the intermediate translation sentence, and the post-translation means Can translate the intermediate translation sentence into a sentence described in the translation destination language based on a post-translation model for translating the intermediate translation sentence into a sentence described in the translation destination language.

  In addition, as the pre-translation model, learning was performed using a parallel corpus of the translation source language and a translation intermediate language in which each word included in the sentence described in the translation destination language is rearranged according to the word order of the translation source language. A model learned using a parallel corpus of the translation intermediate language and the translation target language can be used as the post-translation model.

  In addition, after the words included in the sentence described in the first language are arranged in advance in the order of the words in the second language, the sentence described in the first language is translated into the sentence described in the second language. The accuracy of the first language is such that the words included in the sentence described in the second language are sorted in advance in the order of the words in the first language, and then the sentence described in the second language is converted into the first language. The first language can be the translation target language and the second language can be the translation source language when the accuracy is higher than the translation accuracy when the sentence is described. In the prior art based on “pre-ordering”, when the first language is translated into the second language, the first language can be successfully rearranged in the order of the words in the second language, thereby obtaining translation accuracy. However, when the second language is translated into the first language, it is known that the translation accuracy may not be obtained because it is not easy to rearrange the second language in the order of the words of the first language. . For example, the first language is English and the second language is Japanese. In this way, there is no way to rearrange the source language according to the word order of the destination language (for example, Japanese to English), or the rearrangement accuracy is not sufficient, and the word order is first changed with high accuracy by preprocessing. The present invention is particularly effective when it is difficult. In the present invention, the translation intermediate language for learning the pre-translation model and the post-translation model is created by rearranging the translation destination language according to the word order of the translation source language. In the translation from the translation destination language to the translation source language, the word order switching method can be used in the translation method in which the word order switching is performed as a preprocessing and the phrase is translated thereafter.

  In addition, the translation model learning device of the present invention includes an intermediate language sentence creation unit that creates a translation intermediate language in which each word included in a sentence described in a translation destination language is rearranged according to the word order of the translation source language; Using a parallel corpus of the translation intermediate language created by the language sentence creation means and the translation source language, each word of the sentence described in the translation source language is converted into a sentence described in the translation source language. Pre-learning means for learning a pre-translation model for translating into an intermediate translation sentence replaced with a word in the translation destination language in the word order according to the word order of the language, and the translation intermediate language created by the intermediate language sentence creation means And a post-learning means for learning a post-translation model for translating the intermediate translation sentence into a sentence described in the translation-destination language using a parallel corpus with the translation-destination language. it can.

  The translation method of the present invention is a translation method in a translation apparatus including a pre-translation unit and a post-translation unit, and the pre-translation unit converts an input sentence described in a source language into the input sentence. Each word included is translated into an intermediate translation sentence replaced with a word in the translation destination language in the word order according to the word order of the translation source language, and the post-translation means, each word contained in the input sentence is translated into the translation source language By rearranging the word order of the intermediate translation sentence that has been replaced with the words in the translation destination language in the word order according to the word order of This is a method of translating into sentences written in the target language.

  In the translation method of the present invention, the pre-translation means converts the input sentence into the intermediate translation sentence based on a pre-translation model for translating a sentence described in the translation source language into the intermediate translation sentence. The post-translation means translates the intermediate translation sentence into a sentence described in the translation destination language based on a post-translation model for translating the intermediate translation sentence into a sentence described in the translation destination language. Can be translated.

  In the translation method, the pre-translation means includes a parallel corpus of a translation source language and a translation intermediate language in which each word included in a sentence described in the translation destination language is rearranged according to the word order of the translation source language. Using the pre-translation model learned using, the post-translation means can use the post-translation model learned using a parallel corpus of the translation intermediate language and the translation target language.

  The translation model learning method of the present invention is a translation model learning method in a translation model learning device including an intermediate language sentence creation means, a pre-learning means, and a post-learning means, wherein the creation means includes a translation destination language A translation intermediate language is created by rearranging the words included in the sentence described in accordance with the word order of the translation source language, and the pre-learning means includes a translation intermediate language created by the intermediate language sentence creation means, and a translation Using a parallel corpus with the source language, a sentence described in the source language is converted into a word in the target language in which each word of the sentence described in the source language remains in the word order according to the word order of the source language. A pre-translation model for translating into an intermediate translation sentence replaced with the intermediate translation sentence, and the post-learning means uses the parallel corpus of the translation intermediate language created by the creation means and the translation destination language, Before translation It is a way to learn the translation model after for translation to the statement written in the target language.

  The translation program of the present invention is a program for causing a computer to function as each means constituting the translation apparatus.

  The translation model learning program of the present invention is a program for causing a computer to function as each means constituting the above translation model learning apparatus.

  As described above, according to the translation device, method, and program of the present invention, and the translation model learning device, method, and program, in the pre-translation, the word order according to the word order from the source language to the source language is maintained. It is not necessary to consider the rearrangement of words by translating to an intermediate translation sentence that has been replaced with a word in the target language. In post-translation, word replacement is performed by translating from the intermediate translation sentence to the target language. Since there is no need to consider, it is possible to obtain an effect that word association is easy and high-speed and highly accurate translation can be performed.

It is a block diagram which shows the functional structure of the translation apparatus of this Embodiment. It is a flowchart which shows the content of the learning process routine in the translation apparatus of this Embodiment. It is a flowchart which shows the content of the translation processing routine in the translation apparatus of this Embodiment. It is a figure which shows an example of the Japanese part of a parallel corpus. It is a figure which shows an example of the English part of a parallel corpus. It is a figure which shows an example of a Japanese word division corpus. It is a figure which shows an example of an English word division corpus. It is a figure which shows an example of a head postfix type English corpus. It is a figure which shows an example of a translation table file. It is a figure which shows an example of a setting file. It is a figure which shows an example of a language model file. It is a figure which shows an example of the setting file after parameter adjustment. It is a figure which shows an example of an input file. It is a figure which shows an example of an intermediate translation sentence. It is a figure which shows an example of a translation result file.

  Hereinafter, embodiments of the present invention will be described in detail with reference to the drawings. In the present embodiment, translation and learning of a translation model when the translation source language is Japanese and the translation destination language is English will be described.

  First, an outline of the present embodiment and definitions of terms will be described. In this embodiment, the input sentence is translated into the target language in the word order according to the word order of the source language, that is, the word order close to the source language, instead of the pre-ordering of the input sentences by the syntax analyzer of the source language. After that, "post-ordering" translation is realized in which the words are rearranged in the order of the target language. In the following description, this “sentence translated into the translation destination language in the order of words close to the translation source language” is referred to as “intermediate translation sentence”, and this translation process is referred to as “pretranslation”. In addition, the intermediate translation sentence is considered as another virtual language, and is called “translation intermediate language”. In the intermediate translation, the vocabulary is that of the target language, but the word order is close to that of the corresponding source language. Replacement, insertion, and deletion are required. Since this word order rearrangement and word replacement, insertion, and deletion processes can also be regarded as translation, this process is called “post-translation”.

  The translation apparatus 10 according to the present embodiment is configured by a computer including a CPU, a RAM, and a ROM that stores a program for executing a learning process and a translation process routine described later. Further, an HDD as a storage unit may be included. Functionally, this computer can be represented by a configuration including a learning unit 20 and a translation unit 50, as shown in FIG.

  The learning unit 20 can be represented by a configuration that further includes a pre-learning processing unit 22, a pre-translation learning unit 24, and a post-translation learning unit 26.

  The pre-learning processing unit 22 pre-processes the parallel corpus of the translation source language and the translation destination language used in the pre-translation learning unit 24 and the post-translation learning unit 26, which will be described later. Parallel corpus (parallel corpus (translation source language + translation intermediate language) 30) ”and“ parallel corpus of translation intermediate language and translation destination language (parallel corpus (translation intermediate language + translation destination language) 32) ”are generated. Specifically, a sentence in the translation intermediate language is generated by word rearrangement processing for the sentence in the translation destination language. As the word rearrangement process, a method based on a postfix rearrangement rule for English parsing results (for example, Non-Patent Document 3), a word rearrangement method for parsing results (Non-Patent Document 5, “Michael Collins, Philipp Koehn, and Ivona Kucerova, “Clause Restructuring for Statistical Machine Translation.” In Proc. ACL, 2005. pages 531-540. is an example of a rule-based method, Non-Patent Document 6 “Chi-Ho Li, Dongdong Zhang, Mu Li, Ming Zhou, Minghui Li, and Yi Guan, "A Probabilistic Approach to Syntax-based Reordering for Statistical Machine Translation." In Proc. ACL, 2007. pages 720-727. Use existing techniques such as word rearrangement methods that do not use vocabulary (for example, Non-Patent Document 7 “Roy Tromble and Jason Eisner,“ Learning Linear Ordering Problems for Better Translation. ”In Proc. EMNLP, 2009. pages 1007-1016.”) You can sort specific words It is not limited to reason. In addition, the pre-learning processing unit 22 also performs word segmentation processing for each of the sentence in the translation source language, the sentence in the translation intermediate language, and the sentence in the translation destination language.

  The pre-translation learning unit 24 uses the parallel corpus of the translation source language and the translation intermediate language generated by the learning pre-processing unit 22 to translate the sentence of the translation source language into the translation intermediate language. To learn. In the present embodiment, since no special function is required for the pre-translation model 34, a known statistical translation model learning method (for example, Non-Patent Document 1) can be used as the learning method. In addition to the statistical method, a translation model using a translation dictionary describing translation rules may be learned. In the pre-translation learning unit 24, it is not necessary to consider word rearrangement almost or not at all.

  The post-translation learning unit 26 uses the parallel corpus of the translation intermediate language and the translation target language generated by the pre-learning processing unit 22 to translate the sentence in the translation intermediate language into the translation target language 36. To learn. As with the pre-translation learning unit 24, the learning method need not be a specific method. In the post-translation learning unit 26, since the translation intermediate language is the rearrangement result of the translation destination language, the correspondence between the two words is self-explanatory, and there is no need to re-estimate by a statistical method. The post-translation model 36 can be learned. In addition, when a translation intermediate language sentence is generated, re-estimation may be performed if a word is inserted, replaced, or deleted from the translation destination language.

  The translation unit 50 can be represented by a configuration including an input unit 52, a pre-translation unit 54, a post-translation unit 56, and an output unit 58.

  The input unit 52 receives an input sentence (a sentence in a translation source language, a Japanese sentence in this case) from the outside via an input device such as a keyboard (not shown) or a network, performs necessary preprocessing, and then translates it. Pass to the department. In the preprocessing, the input sentence is shaped into a format accepted by the pretranslation unit 54. The shaping here means, for example, performing word breaks and syntax analysis in accordance with the pre-translation model 34, removing unnecessary tags, adding necessary tags, and the like.

  The pretranslation unit 54 refers to the pretranslation model 34 and translates the input sentence of the translation source language passed from the input unit 52 into an “intermediate translation sentence” described in the translation intermediate language. Here, the intermediate translation sentence is a sentence in which each Japanese word is replaced with an English word in a word order close to Japanese. The pre-translation unit 54 operates with little or no consideration of word rearrangement, and can be realized only by translating words and rearranging the neighborhood.

  The post-translation unit 56 refers to the post-translation model 36 and translates the intermediate translation sentence of the translation intermediate language delivered from the pre-translation part 54 into an output sentence of the translation destination language. The post-translation unit 56 operates with little or no consideration of translation of words and phrases, and can be realized by rearranging phrases and replacing, inserting, and deleting a small number of words.

  The output unit 58 performs post-processing necessary for the output sentence, and outputs it to the outside via a display device such as a display (not shown) or a network. In the post-processing, the output of the post-translation unit 56 is shaped into an output format. Here, shaping means removing word breaks, removing unnecessary tags, adding necessary tags, etc. according to the output specifications.

  Next, the operation of translation apparatus 10 of the present embodiment will be described. First, the pre-translation model 34 and the post-translation model 36 are learned by executing a learning process. Then, translation processing is executed using the learned pre-translation model 34 and post-translation model 36, and the input sentence of the input source language is translated into the target language.

  Here, with reference to FIG. 2, the learning process routine executed in the translation apparatus 10 of the present embodiment will be described.

  In step 100, the parallel corpus 28 of the translation source language and the translation destination language is read. Next, in step 102, the sentence in the translation source language of the read parallel corpus is subjected to morphological analysis to perform word segmentation processing.

  Next, in step 104, the sentence of the translation destination language of the read parallel corpus is subjected to form analysis to perform word segmentation processing and to syntax analysis.

  Next, in step 106, a sentence in the translation intermediate language is generated by word rearrangement processing for the sentence in the translation target language using a technique such as a postfix rearrangement.

  Next, in step 108, using the parallel corpus 30 of the translation source language read in step 100 and the translation intermediate language generated in step 106, the sentence in the translation source language is translated into the translation intermediate language. The pre-translation model 34 for learning is learned.

  Next, in step 110, using the parallel corpus 32 of the translation destination language read in step 100 and the translation intermediate language generated in step 106, the sentence in the translation intermediate language is translated into the translation destination language. Therefore, the post-translation model 36 is learned, and the process ends.

  Next, a translation processing routine executed in translation apparatus 10 of the present embodiment will be described with reference to FIG.

  In step 200, an input sentence described in the translation source language is read, morphological analysis is performed on the input sentence, and word separation processing is performed.

  Next, in step 202, the pre-translation model 34 learned in step 108 of the learning process is referred to, and the input sentence in the source language is translated into an intermediate translation sentence.

  Next, in step 204, the post-translation model 36 learned in step 110 of the learning process is referred to, the intermediate translation sentence pre-translated in step 202 is translated into a translation target language sentence, and post-processing is performed. Output, and the process ends.

  Processing of the translation apparatus according to the present embodiment will be described using a more specific example. Also in this embodiment, the case where the translation source language is Japanese and the translation destination language is English will be described.

  FIG. 4 shows an example of a Japanese part of a file in which a parallel corpus of Japanese and English used in this embodiment is recorded, and FIG. 5 shows an example of an English part. In the figure, the symbol “\\” indicates that the line is continued. The same applies to the following drawings.

  The Japanese sentence of the parallel corpus is subjected to word division processing, and the English sentence is subjected to word division processing, syntax analysis processing, and rearrangement processing into a translation intermediate language (steps 102 to 106). A known morphological analysis program MeCab was used for Japanese word division processing. A well-known English syntax analysis program Enju was used for English word division processing and syntax analysis processing. Further, the rearrangement of the translation destination language into the translation intermediate language was performed according to the postfix rearrangement rule described in Non-Patent Document 3.

  By these processes, three files are created: a Japanese word division corpus shown in FIG. 6, an English word division corpus shown in FIG. 7, and a main postfix English (translation intermediate language) corpus shown in FIG. The combination of this "Japanese word division corpus" and the "postfix postfix English (translation intermediate language) corpus" is a parallel corpus of the source language and the translation intermediate language. A combination of the “intermediate language) corpus” and “English word segmentation corpus” is a parallel corpus of the translation intermediate language and the translation target language, and is used by the pre-translation learning unit 24 and the post-translation learning unit 26 in the next stage. .

  Next, for learning in the pre-translation learning unit 24 (translation from the translation source language to the translation intermediate language), Moses, which is a known statistical translation tool kit, and a known statistical language model creation tool kit. SRILM was used. A translation model learning tool included in Moses generates a translation table file as shown in FIG. 9 and a setting file as shown in FIG. Further, the N-gram language model file as shown in FIG. 11 is generated by SRILM (step 108).

  In the translation table file, phrases on the translation source language side, translation intermediate language phrases translated therefrom, and scores at the time of translation are listed with a delimiter “|||”. The setting file is a file for designating the storage location of the translation table file and N-gram language model file to be read when executing the translation program, the model parameters, and the like. The N-gram language model file records a known N-gram language model, that is, a conditional probability model in which words are generated after (N-1) word histories.

  The parameters described in the setting file are initial values, and for practical use, it is necessary to adjust these values separately using a parallel corpus. For parameter adjustment, a known technique called “Minimum Error Rate Training (MERT)” (for example, Non-Patent Document 8 “Franz Josef Och,“ Minimum Error Rate Training in Statistical Machine Translation. ”In Proc ACL, 2003. pages 160-167 "). The MERT repeatedly executes a process of temporarily translating the translation source language side of the parallel corpus for parameter adjustment and updating the parameters so that some translation accuracy is maximized by comparison with the translation destination language side of the parallel corpus. This is a method for obtaining optimum parameters. In this example, the MERT program included in Moses was also used. The setting file obtained as a result is shown in FIG.

  Next, the post-translation learning unit 26 implements the same processing as the pre-translation learning unit 24 except that the translation language pair is changed from the translation intermediate language to the translation destination language (step 110). Detailed description is omitted. If the pre-translation model 34 used by the pre-translation unit 54 and the post-translation model 36 used by the post-translation unit 56 are created by the learning pre-processing unit 22, the pre-translation learning unit 24, and the post-translation learning unit 26, Translation from Japanese to English is possible as follows.

  First, the input sentence file shown in FIG. 13 is read by the input unit 52, and word division is performed in advance by the Japanese morpheme analysis program MeCab. It is also possible to perform word division processing with the input unit 52.

  Then, the pre-translation unit 54 translates the input sentence read from the input file into a translation intermediate language using the pre-translation model 34 (translation table file, N-gram language model file and its parameters are specified in the setting file). Thus, an intermediate translation is obtained (step 202). An example of the intermediate translation is shown in FIG. The intermediate translation has an vocabulary in English, but the word order is not appropriate for the English word order and is close to the Japanese word-by-word translation.

  Next, the post-translation unit 56 further translates the intermediate translation sentence pre-translated by the pre-translation unit 54 into English (step 204). The post-translation unit 56 does not need much translation of words, and performs translation mainly by rearranging words. The result of the post-translation is passed to the output unit 58.

  The output unit 58 receives the output of the post-translation unit 56, erases unnecessary word breaks, and writes the result of post-processing for converting the first letter to upper case in a translation result file. The translation result file is a text file describing English as shown in FIG. 15, for example.

  In the publicly known automatic translation evaluation scale BLEU, the accuracy of Japanese-English translation by the conventional statistical translation technique was 0.2688, whereas in this example, a high accuracy of 0.2960 was obtained.

  As described above, according to the translation apparatus of the present embodiment, after the pre-translation into the intermediate translation sentence in which the translation destination language is replaced with the vocabulary of the translation destination language in the word order close to the translation source language, the intermediate translation sentence is After rearranging the words in the order of the target language, perform the translation. For this reason, in the pre-translation, it is not necessary to consider the word rearrangement or not at all, and in the post-translation, it is not necessary to consider the translation of the words. In the statistical machine translation method, considering that the time and space complexity for solution search increase exponentially with respect to the length of the input sentence, the translation apparatus according to the present embodiment performs translation processing. Speeding up can be realized. Further, in learning the pre-translation model used for pre-translation, it is not necessary to consider word rearrangement substantially or at all, so that the correspondence between words becomes easy. Also, in learning the post-translation model used for post-translation, the translation intermediate language is the result of sorting the translation-destination languages, so the correspondence between the two words is self-explanatory, and it is necessary to reestimate it using statistical methods. It is possible to learn a post-translation model with high accuracy.

  In the translation apparatus according to the present embodiment, the input sentence is not rearranged by analyzing the source language, but is sorted on the target language, thereby eliminating the difference in the word order between the source language and the target language. In addition, it is possible to easily associate partial structure pairs and to extract translation units therefrom, thereby improving the accuracy of machine translation.

  Further, when the pre-ordering (translation from language A to language B is performed after language A is rearranged in the order of language B), that is, a pre-ordering method for accurately translating from language A to language B (And when translation from language B to language A does not have an accurate pre-ordering method), translation intermediate language A ′ (language A is rearranged in the order of language B) Can be created from a sentence of language A, it is easy to create a parallel corpus of language A and translation intermediate language A ′ using a large number of sentences of language A. Therefore, according to the present invention, post-translation from the translation intermediate language A ′ to the language A, which has been improved in accuracy by using the large number of parallel corpora, is performed from the efficient language B that minimizes the rearrangement search of the word order. By combining with pre-translation to the translation intermediate language A ′, the accuracy of machine translation can be improved.

  In the above embodiment, the case where the translation source language is Japanese and the translation destination language is English has been described, but the present invention can also be applied to other languages.

  Moreover, although the case where the learning unit and the translation unit are configured by one computer has been described in the above embodiment, the learning unit and the translation unit may be configured by separate computers.

  The present invention is not limited to the above-described embodiment, and various modifications and applications are possible without departing from the gist of the present invention.

  Moreover, although the above translation apparatus has a computer system, the “computer system” includes a homepage providing environment (or display environment) if the WWW system is used.

  Further, in the present specification, the embodiment has been described in which the program is installed in advance. However, the program can be provided by being stored in a computer-readable recording medium.

DESCRIPTION OF SYMBOLS 10 Translation apparatus 20 Learning part 22 Pre-learning processing part 24 Pre-translation learning part 26 Post-translation learning part 34 Pre-translation model 36 Post-translation model 50 Translation part 52 Input part 54 Pre-translation part 56 Post-translation part 58 Output part

Claims (11)

Pre-translation means for translating the input sentence described in the translation source language into an intermediate translation sentence in which each word included in the input sentence is replaced with a word in the translation destination language in the word order according to the word order of the translation source language; ,
By rearranging the word order of the intermediate translation sentence in which the words included in the input sentence are replaced with the words of the translation destination language in the word order according to the word order of the translation source language, Post-translation means for translating the intermediate translation sentence translated by the pre-translation means into a sentence described in a translation-destination language;
Translation device including The pre-translation means translates the input sentence into the intermediate translation sentence based on a pre-translation model for translating a sentence described in the translation source language into the intermediate translation sentence,
The post-translation means translates the intermediate translation sentence into a sentence described in the translation-destination language based on a post-translation model for translating the intermediate translation sentence into a sentence described in the translation-destination language. Item 3. The translation device according to item 1. The pre-translation model is learned using a parallel corpus of a translation source language and a translation intermediate language in which each word included in a sentence described in the translation destination language is rearranged according to the word order of the translation source language,
The translation apparatus according to claim 2, wherein the post-translation model is learned using a parallel corpus of the translation intermediate language and a translation destination language.   When the words included in the sentence described in the first language are rearranged in advance in the order of the words in the second language, and then the sentence described in the first language is translated into a sentence described in the second language The accuracy of the is described in the first language after the words included in the sentence described in the second language are sorted in advance in the order of the words in the first language. The said 1st language is made into the said translation destination language, and the said 2nd language is made into the said translation origin language, when it is higher than the precision at the time of translating into the sentence. Translation device. Intermediate language sentence creating means for creating a translation intermediate language in which each word included in the sentence described in the translation target language is rearranged according to the word order of the source language;
Using a parallel corpus of the translation intermediate language created by the intermediate language sentence creation means and the translation source language, each sentence of the sentence described in the translation source language is changed into a sentence described in the translation source language. A pre-learning means for learning a pre-translation model for translating into an intermediate translation sentence replaced with a word in the translation target language in the word order according to the word order of the translation source language;
Learning a post-translation model for translating the intermediate translation sentence into a sentence described in the translation destination language using a parallel corpus of the translation intermediate language created by the intermediate language sentence creation means and the translation destination language After-learning means,
Translation model learning device including A translation method in a translation apparatus including a pre-translation means and a post-translation means,
The pretranslation means converts the input sentence described in the translation source language into an intermediate translation sentence in which each word included in the input sentence is replaced with a word in the translation destination language in the word order according to the word order of the translation source language. Translate
The post-translation means converts the word order of the intermediate translation sentence in which the words included in the input sentence are replaced with the words of the translation destination language in the word order according to the word order of the translation source language. A translation method for translating the intermediate translation sentence translated by the pre-translation means into a sentence described in the translation destination language by rearranging to The pre-translation means translates the input sentence into the intermediate translation sentence based on a pre-translation model for translating a sentence described in the translation source language into the intermediate translation sentence,
The post-translation means translates the intermediate translation sentence into a sentence described in the translation-destination language based on a post-translation model for translating the intermediate translation sentence into a sentence described in the translation-destination language. Item 7. The translation method according to Item 6. The pre-translation means is a pre-translation trained using a parallel corpus of a translation language and a translation intermediate language in which each word included in a sentence described in the translation-destination language is rearranged according to the word order of the translation-source language. Using the model,
8. The translation method according to claim 7, wherein the post-translation means uses a post-translation model learned using a parallel corpus of the translation intermediate language and the translation target language. A translation model learning method in a translation model learning device including an intermediate language sentence creation means, a pre-learning means, and a post-learning means,
The intermediate language sentence creation means creates a translation intermediate language in which each word included in the sentence described in the translation destination language is rearranged according to the word order of the translation source language,
The pre-learning means uses a parallel corpus of the translation intermediate language created by the creating means and the translation source language to convert a sentence described in the translation source language into a sentence described in the translation source language. Learn a pre-translation model to translate each word into an intermediate translation sentence that has been replaced by a word in the target language in the order of the word in the source language,
The post-learning means uses the parallel corpus of the translation intermediate language created by the creation means and the translation destination language to post-translate to translate the intermediate translation sentence into a sentence described in the translation destination language Learning a model A translation model learning method.   The translation program for functioning a computer as each means which comprises the translation apparatus of any one of Claims 1-4.   A translation model learning program for causing a computer to function as each means constituting the translation model learning device according to claim 5.

Images