JP2014170296A - Word order rearranging device, translation device, translation model learning device, method, and program - Google Patents

Abstract

PROBLEM TO BE SOLVED: To accurately rearrange a word order of an input sentence.SOLUTION: A language analysis section 30 classifies each element in a clause of an input sentence into two or more kinds of elements, and a rearrangement section 40 rearranges an order of two or more kinds of the classified elements in accordance with a rearrangement rule determined in advance for every clause, so that a word order of the input sentence is rearranged.

Description

  The present invention relates to a word order rearrangement device, a translation device, a translation model learning device, a method, and a program, and more particularly to a word order rearrangement device, a translation device, a translation model learning device, a method, and a program for rearranging the word order of input sentences. About.

  The process of machine translation from language A to language B is roughly divided into two: translation from language A phrases to language B phrases, and appropriate reordering of translated language B phrases in language B. The In the statistical translation technology widely used in the field, the translation of words and the rearrangement of words are statistically modeled from the correspondence between the words of language A and the words of language B estimated from a large number of parallel translations. The language B input sentence is searched for a translation sentence of the language B constituted by translation of a probable phrase and rearrangement of the phrase based on the statistical model.

  In general, it is extremely difficult to comprehensively search for all translation candidates, so the number of translation candidates for each word is limited, and the distance of word sorting is limited within a certain range. By doing so, machine translation processing with a practical calculation amount is realized.

  However, depending on the combination of language A and language B to be translated, the corresponding words may appear in a significantly different order, and the reordering is large enough to accurately translate between such languages. Since translation processing in consideration of distance is required, there is a problem that an increase in calculation amount is unavoidable.

As a technique for dealing with the above problem, there is a technique called “pre-ordering” for rearranging the words of the language A so as to approach the order of the corresponding words of the language B before performing the translation process ( Patent Document 1, Non-Patent Document 1).
Also, in the techniques of Non-Patent Document 2 and Non-Patent Document 3, in Japanese-to-English translation, syntactic analysis is used to estimate the dependency structure of Japanese phrases, and the order of the phrases is changed. You can get close to English word order.

JP 2011-175500 A

Michael Collins, Philipp Koehn, Ivona Kucerova, "Clause Restructuring for Statistical Machine Translation", In Proceedings of the 43rd Annual Meeting on Association for Computational Linguistics, pp. 531-540, 2005 Mamoru Komachi, Yuji Matsumoto, Masaaki Nagata, "Phrase Reordering for Statistical Machine Translation Based on Predicate-Argument Structure", In Proceedings of International Workshop on Spoken Language Translation (IWSLT 2006), 2006 Katsuhito Sudoh, Kevin Duh, Hajime Tsukada, Masaaki Nagata, Xianchao Wu, Takuya Matsuzaki and Jun'ichi Tsujii, "NTT-UT Statistical Machine Translation in NTCIR-9 PatentMT", In Proceedings of NTCIR-9, 2011.

  Non-patent document 1 is intended for translation from German to English, and patent document 1 is intended for translation from English to Japanese. The language of the input sentence (language A) is changed to the corresponding phrase in the translated language (language B). It uses a rule that rearranges them closer to each other. While these technologies can be sorted fairly accurately by using parsing and appropriate rules on the language A side, the necessary rules are different for different languages A and B, so new rules are added. There is a problem that needs to be defined.

  In addition to the dependency structure, Non-Patent Document 2 identifies the subject and object using a technique that estimates the relationship between the predicate and subject and object, called predicate term structure analysis. Rules are used to move the subject and object in order. Non-Patent Document 3 uses a dependency structure and a particle to estimate the subject and object, and uses the same rearrangement rule as Non-Patent Document 2. These methods can be expected to arrange the order of clauses in the order close to English, but the word order in the clauses does not change, so the phrase “I arrived in Tokyo” will only be in the order of words “I arrived in Tokyo” There is a problem that the position of the preposition "at" remains different from the corresponding "arrived at Tokyo" in English.

  In the present invention, there is provided a word order rearrangement device, a translation device, a translation model learning device, a method, and a program, which are made to solve the above-described problems and can accurately rearrange the word order of an input sentence. For the purpose.

  In order to achieve the above object, the word order rearranging device according to the first invention includes, for each phrase of an input sentence, a syntax analysis unit that classifies each element in the phrase into two or more types of elements, and each phrase A rearrangement unit that rearranges the word order of the input sentence by rearranging the order of the two or more types of elements classified by the syntax analysis unit in accordance with a predetermined rearrangement rule. Has been.

  A word order rearrangement method according to a second invention is a word order rearrangement method of a word order rearrangement device including a syntax analysis unit and a rearrangement unit, wherein the syntax analysis unit includes: Each element in the clause is classified into two or more types of elements, and the reordering unit divides into a result of dependency analysis for each clause of the input sentence and a predicate term structure analysis result for each predicate of the input sentence. Based on this, the order of each clause of the input sentence written in Japanese is rearranged, and the order of two or more types of elements classified by the syntax analysis unit is rearranged.

  According to the first and second inventions, the syntax analysis unit classifies the elements in each clause of the input sentence into two or more types of elements, and the rearrangement unit classifies the two or more types of elements classified. Rearrange the order.

  Thus, according to the first invention and the second invention, each element in the clause of the input sentence is classified into two or more types of elements, and the order of the two or more types of classified elements is rearranged. The word order of the input sentence can be rearranged with high accuracy.

  The word order rearrangement device according to a third aspect of the present invention is the word order rearrangement device for rearranging the word order of an input sentence described in Japanese in the order of words similar to a sentence described in a specific language different from Japanese. For each clause of a sentence, a parsing unit that classifies each element in the clause into two or more types of elements, and the order of the clauses of the sentence described in the specific language, the order of the clauses of the sentence described in Japanese The order of the clauses of the input sentence is rearranged in accordance with a predetermined clause rearrangement rule for rearranging the input sentence, and for each of the rearranged input sentences, the sentence clause of the sentence written in Japanese is sorted. According to a predetermined element rearrangement rule for rearranging the order of the two or more types of elements into the order of the two or more types of elements of a sentence clause described in the specific language, the classification is performed by the parsing unit. Said 2 By rearranging the order of the above components like, it is configured to include a, a sort rearrangement unit the word order of the input sentence.

  A word order rearrangement method according to a fourth invention includes a syntax analysis unit and a rearrangement unit, and changes the word order of an input sentence described in Japanese to a sentence described in a specific language different from Japanese. A word order rearrangement method of a word order rearrangement device that rearranges words in close order, wherein the syntax analysis unit classifies each element in the phrase into two or more types of elements for each phrase of the input sentence, The replacement unit is configured to determine the order of the clauses of the input sentence according to a predetermined clause rearrangement rule for rearranging the order of the clauses of the sentence described in Japanese to the order of the clauses of the sentence described in the specific language. For the input sentence thus rearranged, the order of the two or more types of elements of the sentence clause described in Japanese for each of the clauses is changed to the 2 of the sentence clauses described in the specific language. To sort in the order of more than types of elements According Because defined elements collation rules, by rearranging the order of the classified the two or more elements by the parser, rearranges word order of the input sentence.

  According to the third and fourth inventions, the syntax analysis unit classifies each element in each clause of the input sentence written in Japanese into two or more types of elements, and the rearrangement unit determines in advance. The order of clauses of the input sentence is rearranged according to the determined clause rearrangement rule, and the order of two or more types of classified elements is rearranged according to a predetermined element rearrangement rule.

  Thus, according to the third and fourth inventions, each element in each phrase of an input sentence written in Japanese is classified into two or more types of elements, and a predetermined phrase rearrangement rule The order of the phrases in the input sentence is rearranged, and the word order of the input sentence is rearranged with high accuracy by rearranging the order of the two or more types of classified elements according to a predetermined element rearrangement rule. I can do it.

  A translation device according to a fifth aspect of the invention is a translation device for translating an input sentence written in Japanese into a sentence written in a specific language different from Japanese, and for each phrase of the input sentence, A parsing unit for classifying each element of the sentence into two or more elements, and a predetermined order for rearranging the order of sentence clauses described in Japanese into the order of sentence clauses described in the specific language The order of clauses of the input sentence is rearranged according to the clause rearrangement rule, and the order of the two or more types of elements of the sentence clauses described in Japanese for each of the clauses of the rearranged input sentence. The order of the two or more types of elements classified by the parsing unit according to a predetermined element rearrangement rule for rearranging in the order of the two or more types of elements of the sentence clause described in the specific language Sort this The input unit in which the order of the elements is rearranged by the rearrangement unit based on the weight for each of a plurality of types of translation models and the plurality of types of translation models, according to A translation unit that translates a sentence into a sentence written in the specific language.

  A translation method according to a sixth aspect of the present invention includes a syntax analysis unit, a rearrangement unit, and a translation unit, and converts an input sentence written in Japanese into a sentence written in a specific language different from Japanese. A translation method of a translation apparatus for translating, wherein the syntax analysis unit classifies each element in the clause into two or more types of elements for each clause of the input sentence, and the sorting unit is in Japanese. Rearranging the order of the clauses of the input sentence according to a predetermined clause rearrangement rule for rearranging the order of the clauses of the described sentence to the order of the clauses of the sentence described in the specific language, and For each of the input sentences, the order of the two or more types of elements in the sentence clause described in Japanese is changed to the order of the two or more types of elements in the sentence clause described in the specific language. Pre-defined element sorting rules for sorting Then, the word order of the input sentence is rearranged by rearranging the order of the two or more types of elements classified by the syntax analysis unit, and the translation unit includes a plurality of types of translation models and the plurality of types of translations. Based on the weight for each model, the input sentence in which the order of the elements is rearranged by the rearranging unit is translated into a sentence described in the specific language.

  According to the fifth and sixth inventions, the syntax analysis unit classifies each element in each clause of the input sentence into two or more types of elements, and the rearrangement unit determines a predetermined clause rearrangement rule. The order of clauses in the input sentence is rearranged according to the above, and the order of two or more types of classified elements is rearranged according to a predetermined element rearrangement rule. Translate to the sentence described in.

  As described above, according to the fifth and sixth inventions, each element in each clause of the input sentence is classified into two or more types of elements, and the phrase of the input sentence is determined according to a predetermined clause rearrangement rule. The order of the input sentence is rearranged, the order of two or more classified elements is rearranged according to a predetermined element rearrangement rule, and the input sentence is translated into a sentence written in a specific language. Can translate well.

  The translation model learning device according to the seventh aspect of the present invention provides a bilingual data that is a pair of a prepared sentence or phrase written in Japanese and a sentence or phrase written in a specific language different from Japanese. A translation model learning device for learning a plurality of types of translation models for translating a sentence described in Japanese into a sentence described in the specific language and weights for each of the plurality of types of translation models based on the set A learning data language analyzing unit for classifying each element in the phrase into two or more elements for each sentence of the sentence or phrase described in Japanese of each parallel translation data of the set of parallel translation data; Sentences or phrases written in Japanese in each bilingual data in accordance with a predetermined clause rearrangement rule for rearranging the order of clauses of sentences described in the above into the order of clauses of sentences described in the specific language The order of clauses is rearranged, and the sentence or phrase of the sentence or phrase described in the specific language is the order of the two or more types of elements of the clause of the sentence or phrase for each clause. By rearranging the order of the two or more types of elements classified by the learning data language analysis unit according to a predetermined element rearrangement rule for rearranging in the order of the two or more types of elements. Or, a learning data rearrangement unit that rearranges the word order of words and phrases, a phrase or phrase written in Japanese of each parallel translation data that has been rearranged in the element order by the learning data rearrangement unit, and the identification of each parallel translation data Based on a sentence or phrase described in a language, the translation model learning unit that learns the plurality of types of translation models, and the order of elements are rearranged by the learning data rearrangement unit Based on a sentence or phrase described in Japanese of each parallel translation data, a sentence or phrase described in the specific language of each parallel translation data, and a plurality of types of translation models learned by the translation model learning unit A model weight learning unit that learns weights for each of the plurality of types of translation models.

  A translation model learning method according to an eighth invention includes a learning data language analysis unit, a learning data rearrangement unit, a translation model learning unit, and a model weight learning unit, and is prepared in advance and written in Japanese A sentence written in Japanese is converted into a sentence written in the specific language based on a set of parallel translation data that is a pair of a sentence or phrase written in a specific language different from Japanese. A translation model learning method of a translation model learning device for learning a plurality of types of translation models for translation and weights for each of the plurality of types of translation models, wherein the learning data language analysis unit For each sentence of a sentence or phrase written in Japanese of each bilingual data, each element in the phrase is classified into two or more types of elements, and the learning data rearrangement unit Phraseual Rearranging the order of sentences or phrases in Japanese in each bilingual data according to a predetermined phrase rearrangement rule for rearranging the order into the order of sentences in the specific language; With respect to the rearranged sentence or phrase, for each phrase, the two or more elements of the sentence or phrase phrase described in the specific language in the order of the two or more elements of the sentence or phrase phrase The word order of the sentence or phrase is rearranged by rearranging the order of the two or more types of elements classified by the learning data language analysis unit according to a predetermined element rearrangement rule for rearranging in the order of The translation model learning unit includes a sentence or phrase described in Japanese of each parallel translation data in which the order of elements is rearranged by the learning data rearrangement unit, and the identification of each parallel translation data Based on a sentence or phrase described in words, the plurality of types of translation models are learned, and the model weight learning unit is arranged in the Japanese of each parallel translation data in which the order of elements is rearranged by the learning data rearranging unit. The plurality of types of translation based on a sentence or phrase described in words, a sentence or phrase described in the specific language of each parallel translation data, and a plurality of types of translation models learned by the translation model learning unit Learn the weights for each of the models.

  According to the seventh invention and the eighth invention, the learning data language analysis unit classifies each element in each phrase of a sentence or phrase described in Japanese of each parallel translation data into two or more types of elements. The learning data rearrangement unit rearranges the order of the sentences or phrases in Japanese in each bilingual data according to a predetermined phrase rearrangement rule, and classifies according to a predetermined element rearrangement rule. The order of the two or more types of elements that have been arranged is rearranged, and the translation model learning unit describes the sentence or phrase that is written in Japanese for each parallel translation data and the specific language for each parallel translation data. A plurality of types of translation models are learned on the basis of the sentence or phrase that has been made, and a weight for each of the plurality of types of translation models is learned by the model weight learning unit.

  Thus, according to the seventh and eighth inventions, each element in each phrase of a sentence or phrase written in Japanese is classified into two or more types of elements, and a predetermined phrase rearrangement is performed. Rearrange the order of sentences or phrases in Japanese according to the rules, rearrange the word order by rearranging the order of two or more classified elements according to a predetermined element rearrangement rule, Each of a plurality of types of translation models and a plurality of types of translation models based on sentences or phrases written in Japanese with the word order rearranged and sentences or phrases written in a specific language included in the corresponding parallel translation data By learning the weight for, it is possible to learn a translation model that can accurately translate.

  In the first invention, each element in the clause may be classified into a content word and a function word.

  In the first invention, based on the dependency relationship between clauses obtained from the result of dependency analysis on the input sentence, whether or not the order of the two or more types of elements of the clause is rearranged for each clause is determined. The order of two or more elements may be rearranged.

  Further, in the first invention, for each phrase, two or more types of phrase elements based on the configuration of the element of at least one of the phrase source and destination clauses obtained from the result of dependency analysis on the input sentence It may be determined whether or not the order is rearranged, and the order of two or more types of elements may be rearranged.

  The first invention is based on the dependency relationship between clauses obtained from the result of dependency analysis on the input sentence for each clause, and the configuration of elements of at least one of the clause source and destination clauses. Thus, it may be determined whether or not the order of two or more types of elements in the clause is rearranged, and the order of two or more types of elements may be rearranged.

  In the first invention, the order of two or more types of elements of the clause may be rearranged so that the functional word of the clause is arranged at the head of the clause for each clause.

  Moreover, the program of this invention is a program for functioning a computer as each part which comprises said word order rearrangement apparatus, translation apparatus, and translation model learning apparatus.

  As described above, according to the word order rearrangement device, method, and program of the present invention, each element in the clause of the input sentence is classified into two or more types of elements, and the order of the two or more types of classified elements is sorted. By rearranging, the word order of the input sentence can be rearranged with high accuracy.

  Further, according to the translation apparatus, method, and program of the present invention, each element in each clause of the input sentence is classified into two or more types of elements, and the phrase of the input sentence is determined according to a predetermined phrase rearrangement rule. Rearranging the order, rearranging the order of two or more classified elements according to a predetermined element rearrangement rule, and translating the input sentence into a sentence written in a specific language, the input sentence is accurately Can translate.

  Further, according to the translation model learning apparatus, method, and program of the present invention, each element in each phrase of a sentence or phrase described in Japanese is classified into two or more kinds of elements, and a predetermined phrase arrangement is performed. Rearrange the order of the sentences or phrases in Japanese according to the replacement rules, and rearrange the word order by rearranging the order of two or more types of classified elements according to the predetermined element rearrangement rules Based on sentences or phrases written in Japanese with rearranged word order and sentences or phrases written in a specific language included in the corresponding parallel translation data, learn multiple types of translation models and arrange the word order For each of multiple types of translation models based on the sentences or phrases written in Japanese, the sentences or phrases described in a specific language included in the corresponding parallel translation data, and multiple types of translation models. That by learning the weights, it is possible to learn accurately the translation can be carried out the translation model.

It is the schematic which shows the structure of the translation apparatus which concerns on embodiment of this invention. It is the schematic which shows the structure of the translation model learning apparatus which concerns on embodiment of this invention. It is a flowchart which shows the content of the translation model learning process routine in the translation model learning apparatus which concerns on embodiment of this invention. It is a flowchart which shows the content of the translation processing routine in the translation apparatus which concerns on embodiment of this invention. It is a figure which shows the analysis result of KNP. It is a figure which shows the interpretation of the analysis result of KNP. It is a figure which shows the example which rearranged the phrase according to the rearrangement rule of a phrase. It is a figure which shows the example which rearranged the element in a clause. It is a figure which shows the word division | segmentation result of English learning data. It is a figure which shows the rearrangement result of Japanese learning data.

  Hereinafter, embodiments of the present invention will be described in detail with reference to the drawings.

<Outline of the invention>
In the embodiment of the present invention, a sorting rule to be applied to a Japanese parsing result for pre-sorting for translation from Japanese to English is provided. Realize machine translation into In the embodiment of the present invention, not only the order of the clauses is changed based on the rearrangement rule, but also, for example, a function word such as a particle is moved to the head of the clause so as to match the position of the English preposition, Based on the rearrangement rules, the words in the phrase are rearranged so that Japanese can be rearranged in the order of words closer to English than in the prior art. In the embodiment of the present invention, a word having meaning only by the word is `` content word '', a word that is used only in a form attached to the content word, and represents a syntactic or semantic role of the attached content word. “Function word”.

<Configuration of translation device>
A translation apparatus according to an embodiment of the present invention will be described. As shown in FIG. 1, a translation apparatus 100 according to an embodiment of the present invention is a computer that includes a CPU, a RAM, and a ROM that stores a program for executing a translation processing routine described later and various data. Can be configured. The translation apparatus 100 functionally includes an input unit 10, a calculation unit 20, and an output unit 70 as shown in FIG.

  The input unit 10 receives a Japanese input sentence from an input device such as a keyboard. The input sentence is pre-processed by removing tags such as HTML and XML, normalizing the notation, and the like, and is input in a state of being converted into an appropriate input format at the time of input. Note that the input unit 10 may accept input from the outside via a network or the like.

  The calculation unit 20 includes a language analysis unit 30, a rearrangement unit 40, a translation unit 42, and a model storage unit 44.

  The language analysis unit 30 includes a morpheme analysis unit 32, a syntax analysis unit 34, and a predicate term structure analysis unit 36. The language analysis unit 30 performs morphological analysis, syntax analysis, and predicate term structure analysis on the input sentence described in Japanese accepted by the input unit 10.

The morpheme analysis unit 32 uses a known morpheme analyzer (such as JUMAN, MeCab) for the input sentence written in Japanese accepted by the input unit 10 and performs morpheme analysis (word break and part of speech identification). Do.

The syntax analysis unit 34 performs syntax analysis on the sentence described in Japanese that has been morphologically analyzed by the morphological analysis unit 32. In this embodiment, since a dependency structure in Japanese phrase units is used, syntax analysis is performed using a known dependency analyzer (KNP, CaboCha, etc.). These dependency analyzers analyze dependency between clauses, and classify “content words” and “function words” into each word as each element in the clause. Note that “content word” and “function word” are examples of two or more types of elements, and there is a possibility that there are clauses to which only content word classification is given and clauses to which only function word classification is given.

The predicate term structure analysis unit 36 performs a predicate term structure analysis on a Japanese sentence parsed by the syntax analysis unit 34 using a known predicate term structure analyzer (KNP, SynCha, etc.). Here, predicate term parsing refers to a predicate (verb, adjective, action noun), a subject called “ga case”, an object called “wo case”, “ni case” in Japanese. Is to identify a word representing the object called. Through the predicate term structure analysis, the dependency between clauses can determine not only the information of the dependency source and the relationship destination, but also what syntactic role the dependency represents.

  The rearrangement unit 40 uses the Japanese dependency structure and predicate term structure for the sentence described in Japanese obtained by the language analysis unit 30, for the sentence described in Japanese, in English. The clauses and the words in the clauses are rearranged so as to be similar to the word order of the written sentences. Specifically, the clauses are rearranged according to the following clause rearrangement rules for rearranging the order of the clauses of sentences written in Japanese to the order of clauses of sentences written in English: Then, to rearrange the order of "content word" and "function word" in the sentence clause written in Japanese into the order of "content word" and "function word" in the sentence sentence written in English The words in each clause are rearranged according to the predetermined word rearrangement rules. The word rearrangement rule is an example of the element rearrangement rule.

In English, since the predicate is placed immediately after the subject, the predicate clauses are rearranged according to the following clause rearrangement rules (1) to (3) as in Non-Patent Document 2.
(1) The predicate clause is moved immediately after the subject clause (G).
(2) If there is no subject clause, the predicate clause is moved immediately before either the target clause (wo case) or the target clause (d.).
(3) When the predicate clause is a verb-continuous clause, the predicate clause is moved immediately after the related word (modified word).

When none of the subject phrase, the object phrase, and the target phrase is obtained, the following paragraph (4) rearrangement rule is used as an exception.
(4) If none of the subject clause, object clause, and target clause exists, the predicate clause is moved to the second position counted from the end of the sentence.
Since the dependency structure is a hierarchical structure, an operation of moving the predicate clause as a dependency destination according to the above-described clause rearrangement rule is performed in each dependency hierarchy.

Finally, the words (content words and function words) of each clause are rearranged according to the following word rearrangement rule (5) in the clause.
(5) Move the function word to the beginning of the phrase.
Whether or not to perform rearrangement according to the word rearrangement rule in (5) above is determined depending on the dependency relationship between clauses, the configuration of elements in the clause of the clause, or in the clause of the clause whose destination is the clause Judge based on the composition of the elements.

  Specifically, in the dependency relationship between clauses, if the function word in the predicate clause of the whole sentence is an auxiliary verb or final particle, it is not necessary to move the function word to the beginning of the clause (before the verb) If it is a predicate clause of the whole sentence (there is no relation point), it is determined not to perform rearrangement based on the word rearrangement rule of (5) above.

  In addition, in the configuration of the element in the clause of the clause, when the function word of the subject clause is “case particle” “ga” or “ha”, the rearrangement based on the word rearrangement rule of the above (5) is not performed. Judge as follows.

  In addition, in the configuration of the clause element of the clause with the relevant clause as the destination, if the source clause (the clause with the relevant clause as a destination) contains the case particles “ga” and “ha”, When the phrase includes the case particle “”, since the clause is a predicate phrase, it is determined not to perform rearrangement based on the word rearrangement rule of (5) above. Further, it may be determined whether or not to perform rearrangement based on the word rearrangement rule based on the configuration of the elements in the clause of the clause having the clause as a source.

  In addition, the dependency relationship between the above-mentioned clauses, the configuration of the element in the clause of the clause, the configuration of the element in the clause of the clause that is related to the clause, and the configuration of the element in the clause of the clause that is related to the clause It is also possible to determine whether or not the elements in the phrase are to be rearranged by combining rearrangement determination conditions based on each of the above.

  The translation unit 42 translates the Japanese input sentences rearranged by the rearrangement unit 40 using a known machine translator, and is stored in the model storage unit 44 from among a plurality of translation candidate sentences. Based on the plurality of translation models and the weights for each of the translation models, a translation candidate sentence with an optimal translation score is selected and output to the output unit 70. The translation method is a known statistical machine of Non-Patent Document 4 (Phillip Koehn, Franz Josef Och, and Daniel Marcu. 2003. Statistical Phrase-based Translation. In Proc. HLT-NAACL, pages 263-270.). It may be realized by a translation technique, and detailed description is omitted.

  The model storage unit 44 stores the same weights for the translation model and the translation model as the weights for the translation model and the translation model stored in the model storage unit 262 of the translation model learning device 200 described later.

  The output unit 70 restores or adds XML or HTML tags deleted in the pre-input processing to the sentence written in English translated by the translation unit 42, and outputs it to the outside through an output device or a network.

<Configuration of translation model learning device>
Next, the configuration of the translation model learning device according to the embodiment of the present invention will be described. As shown in FIG. 2, the translation model learning device 200 according to the embodiment of the present invention includes a CPU, a RAM, a ROM for storing a program and various data for executing a translation model learning process routine to be described later, It can comprise with the computer which includes. Functionally, the translation model learning apparatus 200 includes an input unit 210, a calculation unit 220, and an output unit 270 as shown in FIG.

  The input unit 210 accepts a parallel learning corpus that is a set of a plurality of sentence pairs and phrase pairs (parallel translation data) that are translated into Japanese and English from an input device such as a keyboard. Note that the input unit 210 may accept input from the outside via a network or the like.

  The arithmetic unit 220 includes a learning parallel corpus 250, a model learning unit 252, and a model storage unit 262.

  The learning parallel corpus 250 stores the learning parallel corpus received by the input unit 210. The learning parallel corpus is a text file in which Japanese and English translations are written at the same number of lines.

  The model learning unit 252 includes a learning data language analysis unit 254, a learning data rearrangement unit 256, a model parallel corpus 257, a translation model learning unit 258, a translation unit 259, and a model weight learning unit 260. ing. The model learning unit 252 learns statistical translation models and model weights using the learning parallel corpus stored in the learning parallel corpus 250 and stores them in the model storage unit 262.

  The learning data language analyzing unit 254 performs Japanese and English language analysis for each Japanese and English parallel data included in the learning parallel corpus. The Japanese language analysis performs the same language analysis (morpheme analysis, syntax analysis, predicate term structure analysis) as the language analysis performed by the language analysis unit 30 of the translation apparatus 100. On the other hand, English language analysis identifies word breaks. Since English is usually divided for each word, it is possible to cut out only the end-of-sentence symbol or the like, or any other word division method generally used in this field may be used.

  The learning data rearrangement unit 256 uses the rearrangement unit 40 of the translation apparatus 100 to translate Japanese sentences or phrases included in the bilingual data for each of a plurality of bilingual data analyzed by the learning data language analysis unit 254. The word order is rearranged by rearranging the clauses and words in the clauses using rearrangement rules similar to the rearrangement rules used. The learning data rearrangement unit 256 combines the rearranged sentence or phrase described in Japanese and the sentence or phrase described in English (separated into words) included in the corresponding parallel translation data. Each of them is stored in the model parallel corpus 257 as a model parallel corpus.

  The model parallel corpus 257 includes a sentence or phrase described in Japanese sorted by the learning data sorting unit 256 and a sentence or phrase written in English included in the corresponding parallel translation data (delimited by words). A parallel translation data set is stored as a parallel corpus for models.

  The translation model learning unit 258 learns a plurality of statistical translation models based on the set of parallel translation data stored in the model parallel corpus 257 and stores it in the model storage unit 262. Statistical translation models include “phrase translation model”, “phrase rearrangement model”, “language model”, and the like that are used in statistical machine translation techniques represented by Non-Patent Document 4, and restrictions on machine translation. As appropriate, the configuration can be selected. As a learning method, Non-Patent Document 4 or the like is widely known, but the present embodiment is not limited to a specific method and can be applied.

  For example, the translation model learning unit 258 uses a statistical phrase based on bilingual data including a sentence or phrase written in Japanese and a sentence or phrase written in English stored in the model parallel corpus 257. Learn each of translation model and phrase rearrangement model. In addition, the language model is learned based on sentences or phrases described in English stored in the model parallel corpus 257.

  The translation unit 259 includes a known machine translator and a plurality of translation models stored in the model storage unit 262 for each of Japanese sentences or phrases stored in the model parallel corpus that is weight learning data. And a plurality of translation candidates using the weights for each of the plurality of translation models. For example, a translation score (for example, a weighted addition value of the scores of each translation model) is calculated using the weight for each translation model stored in the model storage unit 262, and the calculated translation score is calculated from a certain value. Only translation candidates having a larger value are extracted as translation candidates.

  The model weight learning unit 260 includes a plurality of translation candidates for the weight learning data and sentences or phrases (correct answer translations) written in English corresponding to the weight learning data stored in the model parallel corpus 257. Based on this, a translation evaluation scale (for example, BLEU value) for each translation candidate extracted by the translation unit 259 is calculated.

  Then, based on the translation evaluation scales of the translation candidates extracted by the translation unit 259, the model weight learning unit 260 increases the translation candidate that is a better translation for each of the translation models stored in the model storage unit 262. The weight for each translation model is optimized so that the translation score calculated using the weight (for example, the weighted addition value of the scores of each translation model) is high, and the weight for each translation model is assigned to the model storage unit 262. To remember.

  The extraction of translation candidates performed in the translation unit 259 and the learning of weights performed in the model weight learning unit 260 are repeatedly performed until the weights for each of the translation models converge.

  The output unit 270 outputs the translation model and model weight stored in the model storage unit 262.

<Operation of translation model learning device>
Next, the operation of the translation model learning device 200 according to the embodiment of the present invention will be described. First, a learning parallel corpus, which is a set of sentence pairs and phrase pairs (translation data) that are parallel translations of Japanese and English, is input by the input unit 210 and stored in the learning parallel corpus 250. Then, when the CPU executes the program stored in the ROM of the translation model learning device 200, the translation model learning processing routine shown in FIG. 3 is executed.

  First, in step S200, the learning parallel corpus stored in the learning parallel corpus 250 is read.

  Next, in step S202, for each of the Japanese sentences or phrases included in the learning parallel corpus read in step S200, using a known morphological analyzer (such as JUMAN, MeCab), Identification of part of speech).

  Next, in step S204, for each sentence or phrase described in Japanese that has been morphologically analyzed in step S202, the dependency between phrases is analyzed using a known dependency analyzer (KNP, SynCha, etc.). At the same time, a classification of “content word” and “function word” is given to each word in the phrase.

  Next, in step S206, for each sentence or phrase described in Japanese that has been parsed in step S204, a predicate term structure analysis is performed using a known predicate term structure analyzer (KNP, SynCha, etc.). For a predicate clause (verb, adjective, action noun), a subject clause called “ga case”, an object clause called “wo case”, and a target clause called “ni case” are identified.

  Next, in step S207, word breaks are identified for each sentence or phrase described in English included in the parallel learning corpus read in step S200.

  Next, in step S208, for each sentence or phrase described in Japanese that has been subjected to the predicate term structure analysis in step S206, the phrase is rearranged using the phrase rearrangement rule, and then the word rearrangement is performed. Using the rules, the word order is rearranged by rearranging the words in the clause, the rearranged sentence or phrase written in Japanese, and the sentence or phrase written in English included in the corresponding bilingual data ( Together with the one acquired in step S207) and stored in the model parallel corpus 257 as a model parallel corpus.

  Next, in step S212, based on a set of bilingual data composed of sentences or phrases written in Japanese and sentences or phrases written in English, stored in the model parallel corpus 257, statistical phrases Each of the translation model and the phrase rearrangement model is learned, and the language model is learned based on each sentence or phrase described in English stored in the model parallel corpus 257. Then, each learned translation model is stored in the model storage unit 262. In addition, the initial value of the weight for each translation model is stored in the model storage unit 262.

  Next, in step S214, for each sentence or phrase described in Japanese whose word order is rearranged in step S208, a known machine translator, each translation model learned in step S212, and a model storage unit 262 A plurality of translation candidates are created based on the translation score using the weight for each of the translation models stored in.

  In step S216, for each of the plurality of translation candidates created in step S214, based on the sentence or phrase (correct translation) described in English corresponding to the translation candidate stored in the model parallel corpus 257. The translation evaluation scale for each translation candidate is calculated.

  Then, based on the translation evaluation scale of each translation candidate and the translation score of each translation candidate, the weight for each translation model is optimized, the weight for each translation model is learned, and stored in the model storage unit 262.

  Next, in step S218, it is determined whether the weights for each of the translation models learned in step S216 have converged. If it has converged, the process ends. If not converged, the process proceeds to step S214.

<Operation of translation device>
Next, the operation of translation apparatus 100 according to the embodiment of the present invention will be described. First, the input unit 10 inputs a plurality of translation models learned by the translation model learning device 200 and weights for each of the translation models, and stores them in the model storage unit 44. When a sentence written in Japanese is input by the input unit 10, the translation processing routine shown in FIG. 4 is executed by the CPU executing the program stored in the ROM of the translation apparatus 100. .

  First, in step S100, an input sentence written in Japanese is accepted.

  Next, in step S102, a morphological analysis (word delimiter and part-of-speech identification) is performed on the input sentence described in Japanese received in step S100 by using a known morphological analyzer (such as JUMAN, MeCab). Do.

  Next, in step S104, the input sentence subjected to morphological analysis in step S102 is parsed using a known dependency analyzer (KNP, CaboCha, etc.) to analyze the dependency between clauses. The words “content word” and “function word” are assigned to each word in the phrase.

  Next, in step S106, a predicate term structure analyzer (KNP, SynCha, etc.) is used for the input sentence parsed in step S104 to analyze the predicate term structure, Identify clauses, object clauses, and target clauses.

  Next, in step S108, the input sentence that has been subjected to the predicate term structure analysis in step S106 is rearranged using the clause rearrangement rule, and then the word in the clause is used using the word rearrangement rule. Rearrange the word order by rearranging.

  Next, in step S112, the input sentence obtained by rearranging the word order acquired in step S108 is translated using a known machine translator, and stored in the model storage unit 44 from a plurality of translation candidate sentences. A translation candidate sentence having an optimal translation score is selected based on a plurality of translation models and a weight for each of the translation models.

  Next, in step S114, the translation result selected in step S112 is output and the process ends.

<Example of Translation Model Learning Device 200>
Next, an embodiment in which a statistical translation model is learned from a Japanese-English parallel corpus of about 3 million sentences will be described below.

  The learning data language analysis unit 254 of the translation model learning device 200 performs language analysis of Japanese and English. Japanese language analysis performs the same processing as the language analysis unit 30 of the translation apparatus 100. English language analysis uses only the word segmentation program bundled with Moses and performs only word segmentation. The Japanese language analysis result is the same as that of the language analysis unit 30. An example of the English word division result is shown in FIG.

  In the learning data rearrangement unit 256 of the translation model learning device 200, the same sort processing as that of the rearrangement unit 40 of the translation device 100 is performed for the Japanese language of the parallel corpus. The result of this processing is shown in FIG.

  As described above, the parallel corpus of Japanese and English is rewritten into a rearranged set of Japanese word strings and a set of English word strings. Each set of word strings is a text file in which one sentence is stored in one line separated by single-byte spaces for each word as in the example of the result of word separation. In the translation model learning, a phrase program model DB: phrase-table.gz, phrase rearrangement model DB: from a text file representing a set of Japanese and English word strings by a learning program provided by Moses. reordering-table.wbe-msd-bidirectional-fe.gz ".

  Further, “language model DB: ja.5gram.arpa.gz” is created from a text file of a set of English word strings by a known language model learning program SRILM. In this embodiment, a word 5-gram language model is created.

  Further, the weight of each model is learned by a known method (Non-Patent Document 5) called “Minimum Error Rate Training (MERT)” that determines the optimum value of the weight between models, and the model DB Along with this information, write it to the translation program settings file.

<Example of Translation Device 100>
An example in the case where a Japanese sentence “the data storage device 10 is connected to the data collection device 20” is input to a computer terminal on which the translation device 100 is mounted is shown below.

  In this embodiment, Japanese morphological analysis software JUMAN known for Japanese language analysis and syntactic analysis software KNP including known predicate term structure analysis are used. In this embodiment, in order to perform the morphological analysis step, the syntax analysis step, and the predicate term structure analysis step collectively, the Japanese sentence input from the input unit 10 is input from the standard input of the computer terminal to JUMAN, and the output is It is configured to directly input KNP. The KNP analysis result is output as shown in FIG. In FIG. 5, the information that is not used in the present embodiment is partly omitted.

  The symbol “*” at the beginning of a line indicates the beginning of a phrase, and the symbol “+” at the beginning of a line indicates that the word indicated in the following line is a content word. The interpretation of the analysis result is as shown in FIG.

  The rearrangement unit 40 of the translation apparatus 100 rearranges the Japanese sentence clauses according to the clause rearrangement rules. In the present embodiment, the predicate clause of the above Japanese sentence is moved immediately after the subject clause. However, considering that the punctuation mark is a symbol representing the end of the sentence, it is left at the end of the sentence. The result of rearrangement is as shown in FIG. Then, when the function words in the clauses of each clause are rearranged to the head of the clauses except for the particles “ha” and “ga” in the subject clause, the result is as shown in FIG.

  In the case where the phrase “data storage device 10” is further modified by another clause, such as “the data storage device 10 shown in FIG. 1 is connected to the data collection device 20”. Requires hierarchical processing. First, based on the relationship between the phrases “shown in FIG. 1” → “data storage device 10 is”, when the predicate clause is a verb-use phrase, the predicate clause is moved immediately after the word concerned. Reordering rules for clauses, if there is no subject clause, reordering rules for moving the predicate clause immediately before one of the target clause (wo case) or the target clause (d), Then, according to the word rearrangement rule, the data is rearranged as “data storage device 10“ shown in FIG. 1 ”, and then the phrase rearrangement is performed according to the phrase rearrangement rule described above.

  The translation unit 42 of the translation apparatus 100 translates the Japanese sentences rearranged by the rearrangement unit 40 into English. In this embodiment, the statistical translation model (phrase translation model, phrase rearrangement model, language model) learned by the translation model learning unit 258 of the translation model learning device 200 and the weights for each model are used, and known statistical translation software is used. Moses was used. In this embodiment, the output result of “the data storage device 10 is connected to a data collecting device 20.” of Moses is output to the output unit 70.

  It has been found that the machine translation program realized by the method of the above embodiment shows higher translation performance than the machine translation program configured by the prior art and the machine translation program based on the conventional postfix rearrangement method. In an experiment using a statistical model learned with a Japanese-English parallel corpus of about 3 million sentences, the value of the evaluation index BLEU that is most commonly used in the field is composed of conventional techniques that do not perform pre-ordering processing. While the machine translation program was 0.2956, the machine translation program of this embodiment achieved 0.3170.

As described above, according to the translation apparatus according to the embodiment of the present invention, each word in each phrase of the input sentence is classified into “content word” and “function word”, and a predetermined phrase arrangement is performed. The order of clauses in the input sentence is rearranged according to the replacement rule, the word order is rearranged by rearranging the order of the “content word” and “functional word” according to the predetermined word rearrangement rule, and the word order is rearranged. By translating the input sentence into a sentence described in the translation destination language, the input sentence can be translated with high accuracy.

  Further, according to the translation model learning apparatus according to the embodiment of the present invention, each word in each phrase of a Japanese sentence or phrase is classified into “content word” and “function word”, and a predetermined phrase is determined. Rearrange the order of Japanese sentence or phrase clause according to the rearrangement rules, and rearrange the word order by rearranging the order of “content words” and “functional words” according to the predetermined word rearrangement rules Learn multiple translation models based on Japanese sentences or phrases that have been rearranged and rearranged, and sentences or phrases written in the target language included in the corresponding bilingual data, and rearrange the word order Based on a sentence or phrase written in Japanese, a sentence or phrase written in a specific language included in the corresponding parallel translation data, and a plurality of kinds of translation models, weights for each of the plurality of kinds of translation models are given. By learning It is possible to learn a translation model that can be performed with high accuracy translation.

  Also, in the translation from Japanese to English, it is possible to rearrange the Japanese word order to the word order close to English, so the difference in word order in the translation from Japanese to English is very small. Becomes easier.

  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.

  For example, a case has been described in which morphological analysis, syntax analysis, and predicate term structure analysis are performed separately, but the present invention is not limited to this, and morphological analysis and syntax analysis may be performed simultaneously. Further, syntax analysis and predicate term structure analysis may be performed simultaneously.

  Moreover, although the case where word order was rearranged so that Japanese is similar to the word order of English was demonstrated, it is not limited to this, The first language of two languages from which languages differ is similar to the word order of a 2nd language In this way, the word order may be rearranged.

DESCRIPTION OF SYMBOLS 10 Input part 20 Operation part 30 Language analysis part 32 Morphological analysis part 34 Syntax analysis part 36 Predicate term structure analysis part 40 Rearrangement part 42 Translation part 44 Model storage part 70 Output part 100 Translation apparatus 200 Translation model learning apparatus 210 Input part 220 Computing unit 250 Learning parallel corpus 252 Model learning unit 254 Learning data language analysis unit 256 Learning data rearrangement unit 257 Model parallel corpus 258 Translation model learning unit 259 Translation unit 260 Model weight learning unit 262 Model storage unit 270 Output unit

Claims (14)

For each clause of the input sentence, a parsing unit that classifies each element in the clause into two or more types of elements,
A rearrangement unit that rearranges the word order of the input sentence by rearranging the order of the two or more elements classified by the syntax analysis unit according to a predetermined rearrangement rule for each clause.
Sorting device including word order.   The word order rearrangement device according to claim 1, wherein the syntax analysis unit classifies each element in the phrase into a content word and a function word.   The reordering unit reorders the order of the two or more types of elements of the clause for each clause based on a dependency relationship between the clauses obtained from a result of dependency analysis on the input sentence. The word order rearrangement apparatus according to claim 1, wherein the rearrangement apparatus rearranges the order of the two or more types of elements.   The reordering unit, for each of the clauses, based on a configuration of elements of at least one of the clause source and the destination clause of the clause obtained from the result of dependency analysis on the input sentence, the two types of the clauses The word order rearrangement device according to claim 1 or 2, wherein it is determined whether or not the order of the elements is rearranged, and the order of the two or more types of elements is rearranged.   The rearrangement unit, for each clause, a dependency relationship between the clauses obtained from a result of dependency analysis on the input sentence, and a configuration of elements of at least one clause of the clause source and destination The word order rearrangement device according to claim 1 or 2, wherein it is determined whether or not the order of the two or more types of elements of the phrase is to be rearranged based on the order, and the order of the two or more types of elements is rearranged.   The word order rearrangement according to claim 2, wherein the rearrangement unit rearranges the order of the two or more types of elements of the clause so that the functional word of the clause is arranged at the head of the clause for each clause. apparatus. In a word order rearrangement device that rearranges the word order of input sentences written in Japanese into the order of words close to sentences written in a specific language different from Japanese,
For each clause of the input sentence, a syntax analyzer that classifies each element in the clause into two or more types of elements;
Rearranging the order of clauses of the input sentence according to a predetermined clause rearrangement rule for rearranging the order of clauses of sentences written in Japanese to the order of clauses of sentences described in the specific language; For the rearranged input sentence, for each of the phrases, the order of the two or more kinds of elements of the sentence sentence described in Japanese is the two or more kinds of elements of the sentence phrase described in the specific language. A rearrangement unit that rearranges the word order of the input sentence by rearranging the order of the two or more types of elements classified by the syntax analysis unit in accordance with a predetermined element rearrangement rule for rearranging in the order of When,
Sorting device including word order. In a translation device that translates input sentences written in Japanese into sentences written in a specific language different from Japanese,
For each clause of the input sentence, a syntax analyzer that classifies each element in the clause into two or more types of elements;
Rearranging the order of clauses of the input sentence according to a predetermined clause rearrangement rule for rearranging the order of clauses of sentences written in Japanese to the order of clauses of sentences described in the specific language; For the rearranged input sentence, for each of the phrases, the order of the two or more kinds of elements of the sentence sentence described in Japanese is the two or more kinds of elements of the sentence phrase described in the specific language. A rearrangement unit that rearranges the word order of the input sentence by rearranging the order of the two or more types of elements classified by the syntax analysis unit in accordance with a predetermined element rearrangement rule for rearranging in the order of When,
A translation for translating the input sentence in which the order of the elements is rearranged by the rearrangement unit into a sentence described in the specific language based on a plurality of kinds of translation models and weights for each of the plurality of kinds of translation models And
Translation device including A sentence written in Japanese based on a set of bilingual data prepared in advance, which is a pair of a sentence or phrase written in Japanese and a sentence or phrase written in a specific language different from Japanese A translation model learning apparatus for learning a plurality of types of translation models for translating a sentence described in the specific language and a weight for each of the plurality of types of translation models,
A learning data language analysis unit for classifying each element in the phrase into two or more elements for each phrase of the sentence or phrase described in Japanese of each parallel translation data of the set of parallel translation data;
Sentences written in Japanese in each bilingual data according to a predetermined clause rearrangement rule for rearranging the order of clauses of sentences written in Japanese to the order of clauses of sentences described in the specific language Alternatively, the order of the phrase clauses is rearranged, and, for each of the rearranged sentences or phrases, the sentence in which the order of the two or more types of elements of the sentence or phrase clause is described in the specific language or By rearranging the order of the two or more types of elements classified by the learning data language analysis unit according to a predetermined element rearrangement rule for rearranging in the order of the two or more types of elements of the phrase of the phrase , A learning data rearrangement unit for rearranging the word order of the sentence or phrase;
Based on the sentence or phrase described in Japanese of each parallel translation data in which the order of the elements is rearranged by the learning data rearrangement unit, and the sentence or phrase described in the specific language of each parallel translation data, A translation model learning unit for learning multiple types of translation models;
A sentence or phrase described in Japanese of each parallel translation data in which the order of elements is rearranged by the learning data rearrangement unit, a sentence or phrase described in the specific language of each parallel translation data, and the translation model learning A model weight learning unit that learns weights for each of the plurality of types of translation models based on a plurality of types of translation models learned by the unit;
Translation model learning device including A word order rearrangement method of a word order rearrangement device including a parsing unit and a rearrangement unit,
The parsing unit classifies each element in the clause into two or more types of elements for each clause of the input sentence,
The reordering unit determines the order of the clauses of the input sentence written in Japanese based on the result of dependency analysis for each clause of the input sentence and the predicate term structure analysis result for each predicate of the input sentence. The word order rearrangement method of rearranging and rearranging the order of two or more types of elements classified by the syntax analysis unit. A word order rearrangement device that includes a parsing unit and a rearrangement unit, and rearranges the word order of input sentences written in Japanese into the order of words similar to sentences written in a specific language different from Japanese. A replacement method,
The parsing unit classifies each element in the clause into two or more types of elements for each clause of the input sentence,
The rearrangement unit includes a phrase clause of the input sentence according to a predetermined phrase rearrangement rule for rearranging the order of sentence phrases described in Japanese to the order of sentence phrases described in the specific language. For the input sentence thus rearranged, the order of the two or more types of elements of the sentence clause described in Japanese is set for each of the clauses of the sentence clause described in the specific language. By rearranging the order of the two or more types of elements classified by the syntax analysis unit in accordance with a predetermined element rearrangement rule for rearranging in the order of the two or more types of elements, the word order of the input sentence Sort the word order. A translation method for a translation apparatus, comprising: a parsing unit, a sorting unit, and a translation unit, and translating an input sentence written in Japanese into a sentence written in a specific language different from Japanese ,
The parsing unit classifies each element in the clause into two or more types of elements for each clause of the input sentence,
The rearrangement unit includes a phrase clause of the input sentence according to a predetermined phrase rearrangement rule for rearranging the order of sentence phrases described in Japanese to the order of sentence phrases described in the specific language. For the input sentence thus rearranged, the order of the two or more types of elements of the sentence clause described in Japanese is set for each of the clauses of the sentence clause described in the specific language. By rearranging the order of the two or more types of elements classified by the syntax analysis unit in accordance with a predetermined element rearrangement rule for rearranging in the order of the two or more types of elements, the word order of the input sentence Sort
The translation unit is described in the specific language with the input sentence in which the order of the elements is rearranged by the rearrangement unit based on a plurality of types of translation models and a weight for each of the plurality of types of translation models. Translate to sentence Translation method. A sentence or phrase written in Japanese, which includes a learning data language analysis unit, a learning data rearrangement unit, a translation model learning unit, and a model weight learning unit, is different from Japanese A plurality of types of translation models for translating sentences described in Japanese into sentences described in the specific language based on a set of parallel translation data that is a pair with sentences or phrases described in the specific language; A translation model learning method of a translation model learning device for learning a weight for each of the plurality of types of translation models,
The learning data language analysis unit classifies each element in the phrase into two or more elements for each sentence of a sentence or a phrase described in Japanese of each parallel translation data of the set of parallel translation data,
The learning data rearrangement unit is configured to convert each bilingual data in accordance with a predetermined phrase rearrangement rule for rearranging the order of sentences of sentences described in Japanese to the order of sentences of sentences described in the specific language. Rearranging the order of clauses of sentences or phrases described in Japanese, and for the rearranged sentences or phrases, the order of the two or more types of elements of the sentence or phrase clauses for each of the phrases The two or more types classified by the learning data language analysis unit according to a predetermined element rearrangement rule for rearranging in the order of the two or more types of elements in a sentence or phrase phrase described in a specific language Rearrange the word order of the sentence or phrase by rearranging the order of the elements,
The translation model learning unit includes a sentence or phrase described in Japanese of each parallel translation data in which the order of elements is rearranged by the learning data rearrangement unit, and a sentence or phrase described in the specific language of each parallel translation data Learn the multiple types of translation models based on words and phrases,
The model weight learning unit includes a sentence or phrase described in Japanese of each parallel translation data in which the order of elements is rearranged by the learning data rearrangement unit, and a sentence or phrase described in the specific language of each parallel translation data A translation model learning method for learning a weight for each of the plurality of types of translation models based on a word and a plurality of types of translation models learned by the translation model learning unit.   The program for functioning a computer as each part which comprises the word order rearrangement apparatus of any one of Claims 1-7, the translation apparatus of Claim 8, or the translation model learning apparatus of Claim 9. .