JP2009075795A - Machine translation apparatus, machine translation method, and program - Google Patents

Abstract

A machine translation device that realizes machine translation in consideration of syntax information without learning is provided.
A translation target text data storage unit for storing a translation target text in a source language, a translation model information storage unit for storing a translation model, and statistically machine-translating the translation target text using the translation model. A machine translation unit 13; a post-translation text data storage unit 14 that stores post-translation text; a syntax analysis unit 15 that parses the translation target text to obtain tree structure information; a tree structure information storage unit 16 that stores tree structure information; Judgment unit 18 for judging whether the translated text can be realized by converting the leaf of the tree structure indicated by the tree structure information from the original language to the target language and replacing the nodes of the tree structure; A selection unit 19 for selecting the post-translation text determined by the determination unit 18 and an output unit 20 for outputting the selection result.
[Selection] Figure 1

Description

  The present invention relates to a machine translation apparatus that performs statistical machine translation.

  In recent years, statistical translation (SMT), particularly phrase-based statistical translation (PBSMT), has begun to be widely used as machine translation (for example, Non-Patent Document 1). One of the biggest problems in PBSMT is the rearrangement of phrases (especially opposite). The reason is that the rearrangement model in PBSMT is a model that depends on the number of words ahead (after) to be moved when rearranging.

In order to solve this problem, many attempts have been made to introduce syntax information into statistical translation. These attempts are roughly classified into three types depending on which one of the source language and target language syntax information to be used. The first uses syntactic information of both the translation source language and the translation destination language, and is called tree-to-tree translation. The second method uses syntax information of only the translation destination language, and is called string-to-tree translation. The third is called tree-to-string translation, and uses syntax information of only the source language. Like these methods, the performance of statistical translation can be improved by using syntax information.
Daniel Marcu, William Wong, “A phrase-based, joint property model for statistical machine translation”, Proc. EMNLP-2002, pp. 133-139, 2002

  However, by introducing syntax information into the model, the number of parameters to learn increases. This is particularly true in tree-to-tree translation. The increase in the number of parameters means that the quality and quantity requirements for the parallel corpus, which is training data for statistical translation, will increase. In terms of quality, it will further spur PBSMT, which has a severe data sparseness problem in the first place. In terms of quality, the parallel language corpus, which is the learning data, must reflect the structure of the source language, but it can be translated into several different syntaxes with the same meaning. . This means that the way the structure is reflected is not necessarily consistent, which also adds to the data sparseness issue.

  Generally speaking, the introduction of syntactic information for word-based and phrase-based statistical translation increases the ability of the model to express, but has the problem of causing serious data sparseness problems in terms of learning model parameters. .

  The present invention has been made to solve the above problems, and in order to avoid this data sparseness problem, a machine translation apparatus in which a tree structure constraint model, which is a syntax information model that does not require parameter learning, is introduced. The purpose is to provide.

  In order to achieve the above object, a machine translation apparatus according to the present invention includes a translation target text data storage unit that stores translation target text data that is text data of a source language to be translated, and translation from the source language to the target language. A translation model information storage unit that stores translation model information used in the machine, a machine translation unit that statistically translates the text data to be translated using the translation model information, and a text to be translated by the machine translation unit Information indicating a tree structure of the text data to be translated by parsing the text data to be translated, and a text data storage unit for the text to be translated that accumulates the text data after translation which is text data of a target language obtained by machine translation of the data A syntax analysis unit for obtaining tree structure information, a tree structure information storage unit for storing the tree structure information, and the post-translation The post-translation text data stored by the text data storage unit is a tree structure indicated by the tree structure information stored by the tree structure information storage unit, and the tree structure obtained from the translation target text data corresponding to the post-translation text data A determination unit that determines whether or not it can be realized by converting the source language of the leaf of the tree into the target language, replacing the nodes of the tree structure, converting the source language of the leaf of the tree structure into the target language, and the tree structure A selection unit that selects the post-translation text data determined by the determination unit to be realized by exchanging the nodes, and an output unit that outputs a selection result by the selection unit.

  With this configuration, a decision using tree structure information is made, and post-translation text data is selected according to the decision result, thereby introducing restrictions on syntax information without performing parameter learning on syntax information. Machine translation can be realized. Therefore, it is possible to prevent a problem of data sparseness related to learning data. In addition, since learning about syntax information is not required, the processing load related to generation of translation model information is lighter than when learning about syntax information is performed, and the data amount of translation model information may be reduced. it can. Also, by making a decision using the tree structure information, machine translation with the tree structure constraint model introduced can be realized, and the accuracy of the translation result can be improved.

In the machine translation device according to the present invention, the machine translation unit performs a word-based statistical machine translation, and the determination unit stores the post-translation text data accumulated by the post-translation text data accumulation unit, The tree structure information is stored in the tree structure information stored by the tree structure information storage unit, and the word unit from the source language of the leaf of the tree structure obtained from the text data to be translated corresponding to the translated text data to the target language. It may be determined whether it can be realized by conversion and replacement of nodes of the tree structure.
With such a configuration, it is possible to appropriately introduce restrictions on tree structure information during word-based statistical machine translation.

In the machine translation device according to the present invention, the machine translation unit performs phrase-based statistical machine translation, and the determination unit stores the post-translation text data accumulated by the post-translation text data accumulation unit, It is a tree structure indicated by the tree structure information accumulated by the tree structure information accumulation unit, and the phrase unit from the source language of the leaf of the tree structure obtained from the text data to be translated corresponding to the translated text data to the target language It may be determined whether or not the conversion can be realized by replacing the node of the tree structure without dividing the phrase that is a unit translated by the phrase-based statistical machine translation and the conversion including the conversion.
With such a configuration, it is possible to appropriately introduce restrictions on tree structure information in the phrase-based statistical machine translation.

  In the machine translation apparatus according to the present invention, the machine translation unit performs left-to-right machine translation at a translation destination, and the post-translation text data includes text data being translated. Yes, each subtree in the tree structure indicated by the tree structure information stored by the tree structure information storage unit is untranslated indicating that the subtree includes only untranslated leaves or consists only of untranslated subtrees. Translated to indicate that the subtree contains only translated leaves, or consists only of translated subtrees, and the subtree contains only translated and untranslated leaves, The determination unit further comprises a classification unit that classifies only a translated subtree or classifies that the translation includes only one translated subtree. When the subtree including two or more subtrees under translation appears using the classification result by the classification unit, the translated text data is converted from the source language of the leaf of the tree structure to the target language. , The machine translation unit determines that the post-translation after the addition of the text when adding a new post-translation target language text to the post-translation text data. You may add so that text data may be selected by the said selection part.

  With such a configuration, it becomes possible to perform machine translation in which restrictions relating to tree structure information are appropriately introduced during left-to-right machine translation at the translation destination.

  In the machine translation device according to the present invention, the machine translation unit generates a plurality of post-translation text data corresponding to the text data to be translated, and the selection unit generates a plurality of post-translation text data, One or two or more post-translation text data determined by the determination unit to be realized by converting the leaf of the tree structure from the source language to the target language and replacing the nodes of the tree structure may be selected.

  With such a configuration, it is possible to select an appropriate one that satisfies the constraints of the tree structure information from a plurality of post-translation text data that has been translated in advance, and the selected one can be used as a final translation result. .

  According to the machine translation apparatus and the like according to the present invention, it is possible to realize machine translation in which restrictions on syntax information are introduced without performing parameter learning on syntax information.

  Hereinafter, a machine translation apparatus according to the present invention will be described using embodiments. In the following embodiments, components and steps denoted by the same reference numerals are the same or equivalent, and repetitive description may be omitted.

(Embodiment 1)
A machine translation apparatus according to Embodiment 1 of the present invention will be described with reference to the drawings. The machine translation apparatus according to the present embodiment introduces a restriction using a tree structure in statistical machine translation.

  FIG. 1 is a block diagram showing a configuration of a machine translation apparatus 1 according to this embodiment. The machine translation apparatus 1 according to the present embodiment includes a translation target text data storage unit 11, a translation model information storage unit 12, a machine translation unit 13, a post-translation text data storage unit 14, a syntax analysis unit 15, a tree The structure information storage unit 16, the classification unit 17, the determination unit 18, the selection unit 19, and the output unit 20 are provided.

  The translation target text data storage unit 11 stores translation target text data that is text data of the source language to be translated. This translation target text data is machine-translated by a machine translation unit 13 described later. Accordingly, the translation target text data storage unit 11 stores translation target text data as data to be machine translated. The text data to be translated may be, for example, text data of one sentence to be translated, or text data of a plurality of sentences having a series of coordinates (for example, text data such as business letters and books). There may be.

  The process in which information is stored in the translation target text data storage unit 11 does not matter. For example, the translation target text data may be stored in the translation target text data storage unit 11 via a recording medium, and the translation target text data transmitted via a communication line or the like is stored in the translation target text data storage unit. 11 may be stored, or the text data to be translated input via the input device may be stored in the text data storage unit 11 to be translated. The storage in the translation target text data storage unit 11 may be temporary storage in the RAM or the like of translation target text data read from an external storage device or the like, or may be long-term storage. The translation target text data storage unit 11 can be realized by a predetermined recording medium (for example, a semiconductor memory, a magnetic disk, an optical disk, etc.).

  The translation model information storage unit 12 stores translation model information used for translation from the source language to the target language. Here, the translation model information is information having a translation unit of the source language included in the bilingual corpus, a translation unit of the target language, and a probability associated with each other. The translation unit of the target language is a translation unit of the target language having a translation relationship with the translation unit of the source language associated with the translation unit of the target language, and is a translation unit included in the bilingual corpus. The probability is a probability relating to the source language translation unit and the target language translation unit with which the probability is associated. The translation unit is, for example, a word, a morpheme, a phrase, or the like. The phrase is not a linguistic phrase but a word string composed of several words. The fact that the source language A and the target language B have a parallel translation relationship means that the source language A translated into the target language becomes B or vice versa. The probabilities regarding the source language translation unit and the target language translation unit include, for example, the probability of the target language translation unit when the source language translation unit is given, and the target language translation unit. Such as the probability of the source language translation unit. A method for generating translation model information is already known, and a description thereof will be omitted. The translation model information is generated using a bilingual corpus.

  The process in which information is stored in the translation model information storage unit 12 does not matter. For example, the translation model information may be stored in the translation model information storage unit 12 via a recording medium, and the translation model information transmitted via a communication line or the like is stored in the translation model information storage unit 12. Alternatively, the translation model information input via the input device may be stored in the translation model information storage unit 12. The translation model information storage unit 12 may store the translation model information read from an external storage device or the like temporarily in a RAM or the like, or may be a long-term storage. The translation model information storage unit 12 can be realized by a predetermined recording medium (for example, a semiconductor memory, a magnetic disk, an optical disk, etc.).

  The machine translation unit 13 statistically machine translates the translation target text data read from the translation target text data storage unit 11 using the translation model information stored in the translation model information storage unit 12. This statistical machine translation method is already known and will not be described in detail. The machine translation unit 13 may perform, for example, a word-based statistical machine translation or a phrase-based statistical machine translation. In the case of word-based statistical machine translation, the source language character string is replaced with the target language character string in units of words (or possibly morphemes similar to words). Become. On the other hand, in the case of the phrase-based statistical machine translation, the replacement of the source language character string to the target language character string is performed in phrase units. Note that even in the case of phrase-based statistical machine translation, replacement of a source language character string with a target language character string in units of words may be performed. Depending on whether the machine translation unit 13 performs word-based statistical machine translation or phrase-based statistical machine translation, appropriate translation model information is accumulated in the translation model information storage unit 12. It is preferable to use the translation model information.

  In addition, the machine translation unit 13 may or may not perform a target side left-to-right machine translation at the translation destination. In the present embodiment, the former case will be mainly described. When the machine translation unit 13 performs a left-to-right machine translation at the translation destination, when adding a new translated target language text to the translated text data, The post-translation text data is added to be selected by the selection unit 19. The selection by the selection unit 19 will be described later. Further, when the machine translation unit 13 performs left-to-right machine translation at the translation destination, the post-translation text data may include text data being translated. In left-to-right machine translation at the translation destination, words that are translated one by one at the translation destination (target language) are added, so even text data that is in the middle of translation This is because the text data after translation can be obtained. Here, the text data being translated is text data that does not completely correspond to the text data to be translated. In other words, the text data being translated is partly translated text data of the translation target text data. The machine translation unit 13 also supports a plurality of post-translation text data corresponding to the text data to be translated (for example, corresponding to a sentence in one source language) when left-to-right machine translation is not performed at the translation destination. A plurality of target language sentences) may be generated.

  The machine translation unit 13 may perform machine translation using information other than translation model information. For example, machine translation may be performed using language model information together with translation model information. As language model information, for example, an N-gram language model may be used. The language model information is stored in a recording medium (not shown), and the machine translation unit 13 may use the language model information by reading the language model information from the recording medium (not shown).

  The translated text data storage unit 14 stores the translated text data in a predetermined recording medium. The post-translation text data is text data in a target language in which the machine translation unit 13 machine-translates the text data to be translated. Therefore, the translation target text data and the translated text data have a parallel translation relationship. Further, the recording medium in which the post-translation text data storage unit 14 stores the post-translation text data is, for example, a semiconductor memory, an optical disk, a magnetic disk, etc., and the post-translation text data storage unit 14 may have, Alternatively, it may exist outside the post-translation text data storage unit 14 (may be outside the machine translation apparatus 1). Further, this recording medium may or may not store the post-translation text data temporarily. Moreover, the process of the machine translation part 13 and the process of the post-translation text data storage part 14 may be made integrally. For example, when the machine translation unit 13 generates post-translation text data, the post-translation text data may be stored in a recording medium such as a memory, and the storage may be performed by the post-translation text data storage unit 14. .

  The syntax analysis unit 15 parses the translation target text data read from the translation target text data storage unit 11 to obtain tree structure information that is information indicating the tree structure of the translation target text data. This tree structure information does not need to include a label (for example, part of speech, subject, predicate, etc.) obtained by general syntax analysis. The tree structure information is generally configured for each sentence of the text data to be translated, but this need not be the case. In other words, the tree structure information is information indicating a structure in which words, morphemes, and the like included in a sentence are structured with labels such as a subject, a predicate, a noun phrase, a verb phrase, and the like (as described above, the label itself) May not be included in the tree structure information). This tree structure information is generally known as a syntax tree, and a detailed description thereof is omitted. FIG. 2 is a diagram illustrating an example of tree structure information of “This is a pen.”. As shown in FIG. 2, “a” and “pen” constitute a noun phrase and are grouped together in one node. Also, “is” and “a pen” are grouped together in one node to form a verb phrase. Further, since “This”, “is a pen”, and “.” Constitute a sentence, they are grouped together in one node. In the case of a syntax tree, labels such as “noun phrase (NP)” and “verb phrase (VP)” are generally attached to each node in FIG. 2, but as described above, Since the machine translation device 1 according to the embodiment does not use the label, the tree structure information may not include information on the label. Note that in the tree structure as shown in FIG. 2, the one in the lowest hierarchy (one that does not have a child below it) may be called a leaf. In FIG. 2, “This”, “is”, and the like are leaves.

  Moreover, although the tree structure information has been described with reference to FIG. 2, the data structure of the tree structure information does not matter. For example, when the above-mentioned “This is a pen.” Is parsed by a syntax analyzer (parser), (S1 (S (NP (DT This)) (VP (AUX is) (NP (DT a) (NN pen) )) (..))). From this, by removing a label (for example, S indicating a sentence, NP indicating a noun phrase, VP indicating a verb phrase, etc.), it is a bracketed sentence ((This) ((is) ((A) (pen))) (.)) Is obtained. This shows the same structure as the tree structure information of FIG. Therefore, the data structure of the tree structure information may be a data structure of such a bracketed sentence. Needless to say, the data structure is not limited to the data structure as long as it can represent the configuration of the tree structure indicated by the tree structure information.

  The tree structure information storage unit 16 stores the tree structure information in a predetermined recording medium. This recording medium is, for example, a semiconductor memory, an optical disk, a magnetic disk, or the like, and may be included in the tree structure information storage unit 16 or outside the tree structure information storage unit 16 (even outside the machine translation device 1). May be present). Further, this recording medium may or may not temporarily store the tree structure information. Further, the processing of the syntax analysis unit 15 and the processing of the tree structure information storage unit 16 may be performed integrally. For example, when the syntax analysis unit 15 generates the tree structure information, the tree structure information may be stored in a recording medium such as a memory, and the storage may be performed by the tree structure information storage unit 16.

  The classification unit 17 classifies each subtree in the tree structure indicated by the tree structure information accumulated by the tree structure information accumulation unit 16. This classification is performed for determination by the determination unit 18 described later. More specifically, the classification unit 17 classifies each subtree in the tree structure into untranslated, translated, and being translated. Untranslated means that the subtree to be classified includes only untranslated leaves or consists of only untranslated subtrees. “Translated” indicates that the subtree to be classified includes only translated leaves or consists only of translated subtrees. “Translated” indicates that the subtree to be classified includes only translated and untranslated leaves, or includes only untranslated and translated subtrees, or includes only one subtree being translated. The tree structure information to be classified by the classification unit 17 is tree structure information corresponding to a sentence that has been left-to-right machine translated by the machine translation unit 13. For example, the classification unit 17 may not classify a subtree that cannot be classified (for example, one that includes two or more subtrees being translated), or may classify the subtree into NG. Further, information on the result of classification by the classification unit 17 may be temporarily stored in a recording medium (not shown). For example, classification results (information indicating untranslated, being translated, and translated) may be accumulated in association with each subtree, or information identifying each subtree is accumulated in association with the classification result. May be.

  The determination unit 18 determines whether the post-translation text data stored by the post-translation text data storage unit 14 can be realized by a predetermined process. The predetermined processing is a tree structure indicated by the tree structure information accumulated by the tree structure information accumulating unit 16, and is obtained from the original language of the leaf of the tree structure obtained from the text data to be translated corresponding to the post-translation text data. It is conversion to language and replacement of nodes of the tree structure. Whether it can be realized is whether post-translation text data can be obtained by such processing. In the predetermined process, the order of conversion from the original language of the leaf of the tree structure to the target language and the replacement of the nodes of the tree structure are not limited. In addition, this determination is a determination as to whether or not the translated text data satisfies the restriction on the tree structure indicated by the tree structure information. Therefore, when the determination unit 18 determines that the translated text data can be realized by converting the leaf of the tree structure from the original language to the target language and replacing the nodes of the tree structure, the translated text data Is determined to satisfy the restriction on the tree structure indicated by the tree structure information (syntax restriction on the tree structure). This determination is made, for example, by converting the tree-structured leaf source language into the target language and replacing the tree-structured nodes as much as possible, and in the generated text data after translation You may judge by whether it is included. If included, the determination unit 18 determines that the post-translation text data can be realized by a predetermined process. Needless to say, this determination may be made according to other equivalent determination conditions as long as the results are the same.

  When the machine translation unit 13 performs word-based statistical machine translation, the determination unit 18 uses the post-translation text data accumulated by the post-translation text data accumulation unit 14 and the tree structure accumulated by the tree structure information accumulation unit 16. The tree structure indicated by the information, from the source language of the leaf of the tree structure obtained from the text data to be translated corresponding to the translated text data (that is, in a parallel translation relationship with the translated text data) to the target language It may be determined whether it can be realized by conversion in units of words and replacement of nodes of the tree structure.

  When the machine translation unit 13 performs the phrase-based statistical machine translation, the determination unit 18 accumulates the post-translation text data accumulated by the post-translation text data accumulation unit 14 by the tree structure information accumulation unit 16. A tree structure indicated by the tree structure information, including conversion of a phrase unit from the source language of the leaf of the tree structure obtained from the text data to be translated corresponding to the post-translation text data, and a phrase-based It may be determined whether or not the phrase that is a unit translated by statistical machine translation can be realized by exchanging nodes of the tree structure without dividing the phrase. “Conversion including phrase unit conversion” means that conversion other than phrase unit may be included. The conversion other than the phrase unit is, for example, conversion in word units.

When the machine translation unit 13 performs left-to-right machine translation and classification by the classification unit 17 is performed as in the present embodiment, the determination unit 18 performs the classification result by the classification unit 17. When a subtree including two or more subtrees under translation appears, the translated text data is converted from the source language of the tree-structured leaf to the target language and the nodes of the tree-structure are replaced. It may be determined that it cannot be realized. This is because by making such a determination, the same determination as described above is performed. The reason will be described later. In the present embodiment, a case will be described in which the determination unit 18 makes a determination using the classification result.
The determination unit 18 may pass the determination result to the selection unit 19 or may temporarily store it in a recording medium (not shown).

  The selection unit 19 selects the post-translation text data determined by the determination unit 18 that can be realized by converting the source language of the tree-structured leaf into the target language and replacing the nodes of the tree structure. That is, the post-translation text data determined to satisfy the restriction on the tree structure indicated by the tree structure information is selected. This selection may be a selection after a series of translations by the machine translation unit 13 (for example, translation of one sentence when translation is performed in sentence units), and the machine translation unit 13 performs machine translation. It may be the selection of the subsequent translated word, morpheme, phrase, etc. in the middle of going. In the former case, for example, since the selection is performed after the fact, the machine translation unit 13 generates a plurality of post-translation text data having a translation relation with the translation target text data. Then, the selection unit 19 determines that the determination unit 18 determines that the plurality of post-translation text data can be realized by converting the source language of the leaf of the tree structure to the target language and replacing the nodes of the tree structure. Select two or more post-translation text data. The selected post-translation text data may be one piece or two or more pieces. In the latter case, that is, when the machine translation unit 13 is performing machine translation, when selecting a translation unit (word, morpheme, phrase, etc.) of the next target language, for example, the machine translation unit When 13 translates the text data to be translated, selection is made for each word, morpheme, and phrase, so that only one post-translation text data that has a translation relationship with the text data to be translated is generated.

  The output unit 20 outputs the selection result by the selection unit 19. Here, the selection result may be post-translation text data selected by the selection unit 19 or information indicating which post-translation text data the selection unit 19 has selected. In the latter case, for example, in the recording medium in which the post-translation text data storage unit 14 stores the post-translation text data, even if the flag or the like given to the post-translation text data selected by the selection unit 19 is the selection result. The information for identifying the selected post-translation text data may be the selection result. In addition, when the machine translation unit 13 performs left-to-right machine translation, multiple selections are made before a sentence is translated. Certain post-translation text data can be information indicating the result of the multiple selections.

  Here, the output may be, for example, display on a display device (for example, a CRT or a liquid crystal display), transmission via a communication line to a predetermined device, printing by a printer, or audio output by a speaker. Alternatively, it may be stored in a recording medium or delivered to another component. The output unit 20 may or may not include an output device (for example, a display device or a printer). The output unit 20 may be realized by hardware, or may be realized by software such as a driver that drives these devices.

  The translation target text data storage unit 11, the translation model information storage unit 12, the post-translation text data storage unit 14 stores the post-translation text data, and the tree structure information storage unit 16 stores the tree structure information. Any two or more storage units or recording media with the recording medium to be recorded may be realized by the same recording medium, or may be realized by separate recording media. In the former case, for example, a region storing translation target text data is the translation target text data storage unit 11, and a region storing translation model information is the translation model information storage unit 12.

  Next, the operation of the machine translation apparatus 1 according to this embodiment will be described using the flowchart of FIG. The flowchart of FIG. 3 shows the operation of the machine translation apparatus 1 when the machine translation unit 13 performs left-to-right machine translation. Also, the flowchart of FIG. 3 is for explaining the process of machine-translating a sentence. Therefore, when a plurality of sentences are continuously machine-translated, the series of processes shown in the flowchart of FIG. 3 may be repeated for the number of sentences.

  (Step S101) The syntax analysis unit 15 reads the translation target text data from the translation target text data storage unit 11, and parses the translation target text data to generate tree structure information. The translation target text data read by the syntax analysis unit 15 is translation target text data that is subsequently machine translated by the machine translation unit 13.

(Step S102) The tree structure information storage unit 16 stores the tree structure information generated by the syntax analysis unit 15 in a predetermined recording medium.
(Step S103) The machine translation unit 13 sets the counter i to 1.

  (Step S104) The machine translation unit 13 machine translates the i-th translation unit (for example, a word, a morpheme, a phrase, etc.) at the translation destination. That is, in order to extend the post-translation text data in the translation destination language (target language) from the beginning of the sentence to the end of the sentence, the i-th translation unit is translated. It is assumed that a plurality of candidates are translated when the i-th translation unit is translated. The number of candidates is not limited. For example, the number of candidates to be translated by the machine translation unit 13 may be determined in advance, or a translation result having a likelihood equal to or greater than a predetermined value may be adopted as a candidate.

  (Step S105) The post-translation text data storage unit 14 stores a plurality of candidates translated by the machine translation unit 13 in association with the translated word text data that has been translated and stored so far. In this accumulation, it is preferable to accumulate a plurality of candidates in descending order of likelihood (that is, descending order of likelihood). This is because the likelihood increases as the value of a counter j described later decreases.

(Step S106) The determination unit 18 sets the counter j to 1.
(Step S107) The classification unit 17 classifies the tree structure indicated by the tree structure information with respect to the j-th translation unit among the i-th translation units translated by the machine translation unit 13. That is, each sub-tree of the tree structure is classified as untranslated, being translated, or translated. In addition, when there exists a subtree that cannot be classified into any of them, the classification unit 17 may classify the subtree into NG.

  (Step S108) The determination unit 18 uses the classification result of the classification unit 17 to convert the j-th translation unit translated by the machine translation unit 13 from the source language of the tree-structured leaf to the target language. It is judged whether it can be realized by conversion to and replacement of nodes of the tree structure.

  (Step S109) As a result of the determination by the determination unit 18, the selection unit 19 converts the tree-structured leaf source language into the target language by adding the jth translation unit to the post-translation text data so far. If it is determined that the tree structure node can be replaced, the j-th translation unit is selected. Therefore, the jth translation unit is added to the end of the post-translation text data. Then, the process proceeds to step S110. On the other hand, as a result of the determination by the determination unit 18, the j-th translation unit added to the post-translation text data up to that point is the conversion of the tree-structured leaf from the source language to the target language and the tree-structured node. If it is determined that the change cannot be realized by the replacement, the selection unit 19 does not select the j-th translation unit and proceeds to step S113.

  (Step S110) The machine translation unit 13 determines whether the sentence has been translated to the end of the sentence. For example, in the case of left-to-right machine translation, a bit may be set in a source language word corresponding to a translated text. In that case, if all the words in the source language sentence being translated have bits set, it may be determined that the sentence has been translated to the end of the sentence. If the sentence has been translated to the end of the sentence, the process proceeds to step S111. If not, the process proceeds to step S112.

  (Step S <b> 111) The output unit 20 reads out the translated text data created by the selection unit 19 repeating the selection from the recording medium accumulated in the translated text data storage unit 14, and outputs it. Then, the series of processes for machine translation of the text data to be translated ends.

(Step S112) The machine translation unit 13 increments the counter i by 1. Then, the process returns to step S104.
(Step S113) The determination unit 18 increments the counter j by 1. Then, the process returns to step S107.

  Here, a determination method in word-based statistical machine translation, a determination method in phrase-based statistical machine translation, and a left-to-right machine translation method at the translation destination will be briefly described.

[Method of judgment in word-based statistical machine translation]
First, as the simplest case, consider a case in which all words in the text data to be translated and the text data after translation have a one-to-one correspondence. As described above, the “word” here includes units similar to words such as morphemes. When the word s _i in the translation target text data is translated into S _i at the translation destination, the translation target text data s ₁ , s ₂ ,..., S _N are word sets S ₁ , S ₂ ,. Translated as a rearranged _N word order. In this case, the number of possible combinations of translated text data is N! (N factorial). The purpose of the tree structure constraint model using the tree structure information according to this embodiment is N! The restriction is to reduce the search space of the street. The tree structure constraint model is based on the assumption that translation is possible according to the following two rules, and does not require parameter training.

Rule 1: If the word s _i of the text data to be translated has a relationship such as a dependency relationship with s _j , the word S _i of the post-translation text data also has a relationship with S _j .
Rule 2: Arcs of tree structure information representing the relationship between words do not intersect.

The determination as to whether or not the above-mentioned rule is satisfied is as follows: “Left of the tree structure (the tree structure indicated by the tree structure information) obtained from the translation target text data corresponding to the translated text data. This is the same as the determination of “whether it can be realized by conversion from the source language (s _i ) to the target language (S _i ) and switching the nodes of the tree structure”. The determination unit 18 makes this determination.

  A case where this determination is performed on the tree structure of FIG. When the tree structure in FIG. 4 is indicated by a sentence in parentheses, ((ab) (cd)) is obtained. From this tree structure, the translated text data is converted from the source language of the leaf of the tree structure obtained from the text data to be translated corresponding to the translated text data to the target language, and the tree structure is determined based on the above determination. The post-translation text data that can be determined to be realized by the replacement of the nodes is limited to one of the following. [ABCD], [ABDC], [BACD], [BADC], [CDAB], [CDBA], [DCAB], [DCBA]. Here, the words A, B, C, and D in the post-translation text data are parallel translations of the words a, b, c, and d in the translation target text data, respectively.

  For example, consider [ACBD] as post-translation text data. It can be seen from the parenthesized sentences of the text data to be translated that a and b have a relationship, and c and d have a relationship. On the other hand, when rule 1 is applied, A and B and C and D are also related. However, if this relationship is applied to the word string [ACBD], arcs indicating the respective relationships intersect, and rule 2 cannot be satisfied. Therefore, it can be seen that the post-translation text data [ACBD] is inappropriate.

  Here, a case will be described in which the determination by the determination unit 18 is applied to the translated text data. When the leaf of the tree structure ((ab) (cd)) obtained from the text data to be translated corresponding to the post-translation text data [ACBD] is converted into the target language, ((AB) (CD)) is obtained. Next, it is determined whether or not the translated text data [ACBD] can be realized by replacing these nodes. In this case, in ((AB) (CD)), AB is a pair (corresponding to one node), and it is never separated by replacing the node, but the translated text data [ACBD ], A and B are separated. Therefore, the translated text data [ACBD] is converted from the original language of the leaf of the tree structure obtained from the text data to be translated corresponding to the translated text data to the target language, and the nodes of the tree structure are replaced. It will be judged that it cannot be realized. Such a determination is performed by the determination unit 18. Here, for example, the determination unit 18 configures all post-translation text data that can be reached by exchanging nodes from ((AB) (CD)). This determination may be performed by determining whether or not post-translation text data to be determined is included, or may be determined by other methods.

  The following example will be described. Assume that the translated text data is [DCAB]. Then, similarly to the above description, when the leaf of the tree structure ((ab) (cd)) obtained from the text data to be translated corresponding to the post-translation text data [DCAB] is converted into the target language, ((AB) ( CD)). Next, it is determined whether or not the translated text data [DCAB] can be realized by replacing these nodes. In this case, [DCAB] is obtained by exchanging C and D (replacement of node 2 in FIG. 4), and (AB) and (DC) (replacement of node 3 in FIG. 4). . Therefore, the determination unit 18 determines that the translated text data [DCAB] satisfies the restriction of the tree structure constraint model.

  An example of an actual sentence is a Japanese translation of an English sentence (He (eats (large bread) quickly).)) (He eats (big bread) quickly). Despite the big differences, you can see that the translation is done while satisfying the above rules. Therefore, it can be inferred that this tree structure constraint model is appropriate.

When this tree structure constraint model is not used, there are 4 possible combinations of post-translation text data for ((ab) (cd))! = 24. On the other hand, when this model is introduced, the number is reduced to 8. In the case of a binary tree structure composed of N words, the combination when this model is introduced is ^2N-1 . This is because the number of nodes in this binary tree is N-1, and there are two selection branches for each node, with or without replacement. This number of combinations is N when this model is not introduced! Compared to, it is very small. Actually, it is about 1 / 7,000 when N = 10, and 1/2 × 10 ¹² when N = 20. More generally, the number of combinations at the time of introduction of this model when not a binary tree is Π _{i = 1} ⁿ (B _i !). Here, n represents the number of nodes included in the tree, and B _i represents the number of branches of the i-th node.

The following expression (1) is an expression representing statistical translation when this model is not used, and P (f | e) and P (e) represent a translation model and a language model, respectively.

On the other hand, when the proposed model according to the present embodiment is used, a new term P (e | T) is added and is expressed by the following equation.

  Here, P (e | T) is a tree structure constraint model, and T represents a tree structure of text data to be translated. The value of P (e | T) is 1 if e satisfies the model constraints and 0 otherwise. When e satisfies the constraints of the model, the translated text data is converted from the source language of the leaf of the tree structure obtained from the text data to be translated corresponding to the translated text data to the target language, and the tree This is a case where it can be realized by exchanging the nodes of the structure.

[Method of judgment in phrase-based statistical machine translation]
Next, the judgment method in the word-based statistical machine translation is extended so that it can be applied to the phrase-based model. Usually, the word alignment is n to m (including 0 to m and n to 0). However, in the phrase-based model, the phrase-to-phrase alignment is always one-to-one even if the number of words contained in each phrase is different. For this reason, the above-mentioned rule for one-to-one words can be applied to the phrase-to-phrase correspondence as it is. It is assumed that the phrase ph _i includes the word s _n and the phrase ph _j includes the word s _m . Here, if the word _{s n} and the word _{s m} has a relation, also defined as having a relationship phrase ph _i and phrase ph _j. This allows the above rules to be expanded into phrases as follows:

Rule 1: If the source language phrase ph _i has a relationship such as a dependency relationship with ph _j , the translated language phrase PH _i also has a relationship with PH _j .
Rule 2: Arcs representing the relationship between phrases do not intersect.

Here, PH _n represents a parallel translation phrase of the translation source language phrase ph _n . The translation source language phrase is a phrase included in the translation target text data, and the translation destination language phrase is a phrase included in the translated text data.

  The determination as to whether or not the above-mentioned rule is satisfied is as follows: “Left of the tree structure (the tree structure indicated by the tree structure information) obtained from the translation target text data corresponding to the translated text data. This is the same as the determination of “whether it can be realized by converting the phrase unit from the original language of the phrase into the target language and replacing the nodes of the tree structure (provided that the phrase is not divided)”. The determination unit 18 makes this determination.

  When the extended rule is applied to this phrase, as in the case of one-to-one word correspondence, for each node of the tree structure representing the parenthesized sentence of the source language, the subtree (or the word that is a leaf) immediately below it ) The translated language sentence can be obtained by changing the order of each other. However, there are restrictions on the nodes that can be replaced, and a node that includes only a part of a phrase cannot be replaced. For example, as shown in FIG. 5, consider a case where bcd forms a phrase ph in a tree structure ((abc) ((de) (fg))). In this case, since node 1 includes bc which is a part of phrase ph, it cannot be replaced. Similarly, the nodes 2 and 4 cannot be switched. On the other hand, the node 3 does not include a phrase, and the node 5 includes the entire phrase, and thus can be replaced. For example, when the node 2 is replaced, the phrase ph is divided into two parts in the translation target language, which is contrary to the phrase-to-phrase correspondence in the phrase-based model is 1: 1. . As a result, only [APHEFG], [APHEGF], [GFEPHA], and [FGEPHA] are allowed as parallel translations of this tree structure. Here, PH represents a parallel translation phrase of ph.

  Here, a specific example in the case where the above-described determination is applied to translated text data will be described. Assume that the translated text data is [DEFGABC]. When the leaf of the tree structure ((abc) ((de) (fg))) obtained from the text data to be translated corresponding to the post-translation text data [DEFGABC] is converted into the target language, ((ABC) ((DE ) (FG))). Next, it is possible to determine whether or not the translated text data [DEFGABC] can be realized by replacing these nodes. This is because the node 5 in FIG. However, in the post-translation text data [DEFGABC], the phrase PH is divided into D and BC. Therefore, the determination unit 18 converts the post-translation text data [DEFGABC] from the source language of the leaf of the tree structure obtained from the text data to be translated corresponding to the post-translation text data to the target language, It is determined that it cannot be realized by replacing the nodes of the tree structure (however, the phrase is not divided). Therefore, the post-translation text data [DEFGABC] is not selected by the selection unit 19.

  On the other hand, if the post-translation text data is [GFEDCBA], the determination unit 18 can obtain the post-translation text data by replacing the node of ((ABC) ((DE) (FG))). It is determined that the phrase PH is not divided. Therefore, the post-translation text data [GFEDCBA] can be selected by the selection unit 19.

[Left-to-right machine translation method at the translation destination]
Next, an algorithm for introducing the above-described tree structure constraint model into a left-to-right machine translation (decoder) at the translation destination will be described. In this machine translation, post-translation text data (translation target language sentence) is generated in order from left to right (that is, from the beginning to the end of the sentence). When a new translation unit for extending the post-translation text data to the right is generated, a bit is set for the word of the translation target text data corresponding to the translation unit. When the bits for all the words in the text data to be translated are set, it is determined that the translated text data has reached the end of the sentence. In order to incorporate the tree structure constraint model into this flow, each time a target language translation unit added to the post-translation text data is newly translated, it is checked whether the tree structure constraint model is satisfied. It is necessary to select one that satisfies the constraints of the structural constraint model.

  Prior to the description of this check algorithm, the tree structure subtree of the text data to be translated is classified into four types: “untranslated”, “translated”, “under translation”, and “NG”.

If a subtree consists only of words that are leaves, and all the words are untranslated (no bits are set), the subtree is untranslated.
If a subtree consists only of untranslated subtrees, that subtree is also untranslated.

If a subtree consists only of words that are leaves and all the words have been translated (bits are set), the subtree has been translated.
If a subtree consists only of translated subtrees, that subtree is also translated.

If a subtree consists only of words that are leaves and contains both untranslated words and translated words, the subtree is being translated.
If a subtree contains both translated and untranslated subtrees, the subtree is being translated.
If a subtree contains only one subtree being translated, that subtree is being translated.

If a subtree contains more than one translating subtree, the subtree is NG. This is because if left-to-right machine translation at the translation destination satisfies the above rules, there will be one boundary where untranslated and translated in the tree structure, and there will be two or more subtrees being translated This is because it cannot exist.
If a subtree contains an NG subtree, that subtree is also NG.

  In this embodiment, the translation unit candidates are selected so that an NG subtree does not occur. Therefore, the classification unit 17 has three types of classifications “untranslated”, “under translation”, and “translated”. However, as described above, when the classification unit 17 performs four types of classification, “untranslated”, “under translation”, “translated”, and “NG”, and the classification unit 17 performs NG classification. The determination unit 18 may determine that the translated text data cannot be realized by converting the tree-structured leaf source language into the target language and replacing the tree-structured nodes. Even if such a determination is made, in the case of the above-described explanation, that is, “when a subtree including two or more subtrees under translation appears using the classification result, the translated text data is converted into a tree-structured leaf. This is substantially the same as the case where the determination unit 18 determines that it cannot be realized by conversion from the original language to the target language and switching the nodes of the tree structure.

During machine translation of the translation destination Left to Right, if a new NG subtree is generated when a new target language translation unit is added to the post-translation text data, the post-translation after the translation unit is added Text data cannot satisfy the tree structure constraint model. Here, the reason will be described. It is assumed that the subtree on the translation target text data side is composed of word strings [x ₁ , x ₂ ,..., X _n ]. At this time, the bilingual translation of this subtree is obtained by rearranging the word order of the bilingual word set {X ₁ , X ₂ ,..., X _N }. This is because rule 2 cannot be satisfied if another word interrupts in the middle. For this reason, the word following the last translated word in the sub-tree being translated must be an untranslated word in this sub-tree. This means that the next word to be translated must be chosen from the untranslated words in the translating subtree. For example, in a tree structure ((ab) ((cd) (ef))), when a and b are translated words and c, d, e and f are untranslated words, they are translated next. The word that should be should be one of c, d, e, f. In this way, if two or more subtrees are included during translation, this condition cannot be satisfied.

Next, let us consider a case where the translation phrase PH of the translation source language phrase ph follows the translation sentence hypothesis (during generation). Assume that the post-translation text data at that time is (S ₁ , S ₂ ,..., S _i ), which satisfies the tree structure constraint model. In order to prevent the post-translation text data followed by the phrase PH from generating two or more subtrees being translated (that is, to satisfy the tree structure constraint model), one of the following conditions must be satisfied: . Here, T represents the smallest sub-tree under translation including any of (S ₁ , S ₂ ,..., S _i ) in the tree structure.

(Condition 1) T is still being translated even after the phrase PH is followed, and no other subtree being translated is generated.
(Condition 2) T becomes a translated subtree by following the phrase PH.

  It is necessary and sufficient for the above condition 1 that the phrase ph is included in the untranslated portion of T. In addition, it is necessary and sufficient for Condition 2 that the phrase ph includes all untranslated portions of T. From this, it can be determined by the following procedure whether the existing target text data added with the new target language translation unit satisfies the tree structure constraint model.

  (1) For existing post-translation text data, the minimum sub-tree under translation included in the corresponding tree structure is stored.

  (2) When a new phrase PH is added to the existing post-translation text data, the translation source language phrase ph, which is a translation of the phrase, and the untranslated subtree not stored in (1) above Compare with the translated part. If ph is included in the untranslated part of the smallest translating subtree, or if ph contains the untranslated part of the smallest translating subtree, the new post-translation text data is followed by PH. Generates post-translation text data and correspondingly updates the smallest sub-translation subtree. Otherwise, the new phrase PH is discarded and the same determination is made for the next phrase. Therefore, the determination unit 18 may determine whether the restriction of the tree structure constraint model is satisfied by making such a determination.

Next, the operation of the machine translation apparatus 1 according to the present embodiment will be described using a specific example.
In this specific example, it is assumed that the text data to be translated is “abcdefghi”. Here, abc... Are source language words and morphemes. It is assumed that the translation target text data “abcdefghi” is stored in the translation target text data storage unit 11.

  It is assumed that the user operates the machine translation apparatus 1 to start the machine translation process for the text data to be translated “abcdefghi”. Then, the syntax analysis unit 15 reads out the translation target text data “abcdefghhi” from the translation target text data storage unit 11, parses it, and generates tree structure information “(((ab) (cd)) ((efg) (hi ))) ”And pass it to the tree structure information storage unit 16 (step S101). Then, the tree structure information storage unit 16 stores the parenthesized sentence “(((ab) (cd)) ((efg) (hi)))” as a tree structure information on a recording medium (not shown) ( Step S102). This tree structure information is illustrated in FIG. Further, as shown in FIG. 6, each node is given a node ID (for example, N001). It is assumed that the node IDs are assigned in order such as N001 and N002 from the left of the hierarchy close to the leaf.

  Next, the machine translation unit 13 reads the translation target text data from the translation target text data storage unit 11 and translates the first translation unit (steps S103 and S104). It is assumed that a plurality of translation result candidates are created by the machine translation. The post-translation text data storage unit 14 stores the plurality of translation result candidates on a recording medium (not shown) (step S105). The post-translation text data accumulating unit 14 accumulates the translation result candidates in descending order of the likelihood (probability value) of the machine translation. Therefore, the translation result candidates after accumulation are in descending order of likelihood. It is assumed that the first candidate is a translated phrase “FE” of the phrase “ef”.

  Here, when classification or determination is made regarding a candidate of a certain translation result, the translated flag corresponding to the translated translation unit including the candidate in the translation target text data storage unit 11 as described above. Is set to “1”. FIG. 7 is a diagram showing an example of the correspondence between the translated flag set in this way and the text data to be translated. In FIG. 7, the source language words and morphemes corresponding to the translated flag “1” have been translated, and the source language words and morphemes corresponding to the translated flag “0” have not been translated. The translated flag may be set by, for example, the machine translation unit 13, the determination unit 18, or other components.

  Next, the determination unit 18 instructs the classification unit 17 to perform classification related to the first candidate. Then, the classification unit 17 performs classification for each node using the tree structure information accumulated by the tree structure information accumulation unit 16 and the relationship between the translation target text data and the translated flag shown in FIG. For example, the node with the node ID “N001” is classified as untranslated because both a and b are not translated (the translated flag is “0”). On the other hand, the node with the node ID “N003” is classified during translation because e and f have been translated but g has not been translated. The result of the classification is as shown in FIG. 8 (steps S106 and S107). 8 is stored in a recording medium (not shown) by the classification unit 17.

  The determination unit 18 refers to the information indicating the classification result shown in FIG. 8 accumulated by the classification unit 17 and determines whether there is a subtree (node) including two or more subtrees (nodes) being translated. Here, although the nodes with node IDs “N003”, “N006”, and “N007” are being translated, each of the nodes has only one node being translated. It is determined that there is no subtree including the subtree being translated (step S108). Therefore, the selection unit 19 selects the first translation result candidate (step S109). Then, in the recording medium in which the post-translation text data storage unit 14 stores the translation result candidates, the selection unit 19 determines the translation result candidate “FE” as post-translation text data. It is assumed that all translation result candidates other than the selected translation result, that is, the second and subsequent translation result candidates are all discarded.

  The machine translation unit 13 refers to the information indicating the relationship between the translation target text data and the translated flag shown in FIG. 7, and determines that the translated flags are not all “1”, so that it is not yet the end of the sentence. The next translation unit, that is, the second translation unit is translated (steps S110, S112, and S104).

  As a result of the translation, the first candidate is “C” corresponding to “c”, the second candidate is “G” corresponding to “g”, and the third candidate corresponds to “h”. It is assumed that “H” and fourth and subsequent candidates are also obtained. These translation result candidates are accumulated by the post-translation text data accumulating unit 14 as in the case of machine translation of the first translation unit (step S105).

  Thereafter, as in the case of the first translation unit, the classification related to the first translation result candidate is performed by the classification unit 17 (steps S106 and S107), and two or more are determined by the determination unit 18 based on the classification result. It is determined whether there is a subtree including the subtree being translated (step S108). In this case, since the nodes with node IDs “N002”, “N003”, “N005”, and “N006” are classified during translation, the determination unit 18 includes a subtree including two or more subtrees (N005 and N006) being translated. It is determined that “N007” exists (this subtree “N007” may be classified as NG). Therefore, the selection of “C” is not performed (step S109), and the classification and determination for the second candidate are performed (step S113).

Similar to the above description, classification relating to the second translation result candidate is performed by the classification unit 17 (step S107), and based on the classification result, a subtree including two or more subtrees being translated is determined by the determination unit 18. It is determined whether it exists (step S108). The classification result in this case is that “translated” corresponding to the node ID “N003” in the classification result of FIG. 8 is only “translated”, so the determination unit 18 selects two or more subtrees under translation. Determine that there is no subtree to contain. Accordingly, the “C” is selected (step S109), and the translation result candidate “G” is added to the existing post-translation text data. As a result, the post-translation text data at this point is “FEG”. Thereafter, similarly to the above description, the next translation unit is translated (steps S110, S112, S104). In this way, by sequentially selecting the translation unit candidates to be added to the post-translation text data, the post-translation text data will be extended, and finally the post-translation text data up to the end of the sentence will be Will be created. The output unit 20 outputs the final post-translation text data (step S111). For example, the output unit 20 displays the translated text data on the display. By viewing the display on the display, the user can know the translation result of the text data to be translated “abcdefghi”.
Next, an experimental example of the machine translation apparatus 1 according to this embodiment will be described.

  First, the evaluation scale in the experiment will be described. In this experimental example, four evaluation scales WER, PER, BLEU, and NIST were used. Here, the effectiveness of the machine translation apparatus 1 according to the present embodiment for each evaluation scale will be briefly considered.

WER: This scale can take into account global word order permutations. Therefore, the machine translation apparatus 1 according to the present embodiment is expected to work effectively for this scale.
PER: This measure basically cannot take into account word order. Therefore, it is expected that the proposed method according to the present embodiment is not effective for this scale.

BLEU: Since this scale focuses on ngram, it is possible to consider the replacement of the middle distance word order. For example, when the reference translation translation (w ₁ , w ₂ ,..., W _n ) is a translation result (w ₁ , w ₂ ,..., W _j−1 , X, w _{j + 1} ,..., W _n ). Both WER and BLEU show high values. However, BLEU shows the same high value for the translation results (w _{j + 1} ,..., W _n , X, w ₁ , w ₂ ,..., W _j−1 ), whereas the WER value is 0. Therefore, the machine translation apparatus 1 according to this embodiment is effective for BLEU, but is not expected to be as high as WER.

  NIST: This scale also pays attention to ngram as in BLEU. However, since the weight for the low-order ngram is larger than BLEU, the effectiveness of the machine translation apparatus 1 according to the present embodiment is expected to be lower than that of BLEU.

[English-Japanese News Translation Experiment]
First, in order to evaluate the performance of the machine translation apparatus 1 according to the present embodiment, an English-Japanese news translation experiment was performed. In order to use the method according to the present embodiment, first, tree structure information (sentences in parentheses) of translation target text data is required. For this purpose, the syntax analysis unit 15 parses the translation target text data (syntax analysis). )There is a need to. Performance degradation due to a purging error at this time is expected. The purpose of this experiment is to evaluate the performance of the machine translation apparatus 1 according to the present embodiment and to evaluate performance degradation due to a parsing error. In order to evaluate performance degradation due to a parsing error, the method according to the present embodiment uses a result of parsing automatically (that is, by the parsing unit 15) and a correct parse result (provided manually). Evaluation was made for two cases.

  As the experimental corpus, Yomiuri Shimbun, Reuters and Wall Street Journal were used as training corpora. Each data size (number of sentences) is 145K, 57K, and 14K. Also, 1,787 sentences from the Wall Street Journal were used as development sets, and 1,787 sentences were also used as evaluation sets. The Wall Street Journal text used in the experiment is included in the Pentree Bank Corpus and is given a manual parse tree. Details of these corpora are shown in the table of FIG.

  As translation model information in this experimental example, a phrase-based translation model was used, and GIZA ++ was used for training. Regarding this GIZA ++, refer to the following document 1. The machine translation unit 13 performs machine translation using a language model. SRI language model tool kit was used for training the language model. Refer to the following document 2 for this. The language model is a word trigram, and smoothing was performed by Kneser-Nee discounting (see the following document 3). For optimization of decoding parameters, minimum error training (refer to the following document 4) is used, and optimization is performed for BLEU. Further, as the syntax analysis unit 15 that extracts the tree structure information from the text data to be translated, a Charniak parser (see the following document 5) is used.

  Reference 1: F.R. J. et al. Och, H.C. Ney, “A Systemical Comparison of Variant Statistical Alignment Models”, Computational Linguistics, No. 5; 1, Vol. 29, pp. 19-51, 2003.

  Reference 2: A. Stockke, “SRILM-An Extensible Language Model Tool”, Proc. ICSLP '02, 2002. (Http://www.speech.sri.com/projects/srilm/)

  Reference 3: R.A. Kneser, H .; Ney, “Improved backing-off for gram language model”, Proceedings of the IEEE International Conference of Acoustics, Speech, and Signaling. Vol. 1, pp. 181-184, 1995.

  Reference 4: F.R. J. et al. Och, “Minimum error rate training for statistical machine training”, Proc. ACL, 2003.

  Reference 5: E.M. Charniak, “A Maximum-Entropy-Inspired Parser”, Proc. NAACL-2000, pp. 132-139, 2000.

  In the experiment, comparison was made under three conditions. “Base-line” is a case where the method according to the present embodiment is not used (when simple phrase-based statistical machine translation is performed), and “Chariniak” is a case where a Charniak parser is used to extract tree structure information. “Oracle” is a case where a tree structure of a pen tree bank (that is, a manual tree structure) is used. As the decoder, a Pharaoh (refer to the following document 6) compatible decoder CleopATRa originally developed by the inventors was used for the tree structure constraint model according to the present embodiment. All parameters at this time are the same, and those optimized under Base-line conditions were used. FIG. 10 shows the evaluation results under each condition. Under the condition using the Charniak parser (Chariniak), the WER was improved by about 4%, and the BLEU was improved by about 0.6. The amount of improvement for each evaluation standard is consistent with the above-mentioned prediction, followed by WER and BLEU. There is no significant difference between the result of Oracle using the correct tree structure and Chariniak, and it can be said that the accuracy of the Charniak parser is sufficient for the machine translation apparatus 1 according to the present embodiment. Here, the tree structure for the evaluation sentence set is 60% the same for Chariniak and Oracle.

  Reference 6: P.M. Koehn, “PHARAOH: A beam search decoder for phase-basic static machine translation models”, Proc. AMTA, 2004. (Http://www.isi.edu/publications/licensed-sw/farah/)

[English-Japanese Patent Translation Experiment]
Next, we conducted translation experiments on patents that have different domains from the previous experiment. Detailed information on the patent translation experiment corpus is shown in FIG. The model training, parameter optimization method, and decoding are the same as those used in the news translation experiment.

  FIG. 12 shows the experimental results. In FIG. 12, “Proposed” is the result of the machine translation apparatus 1 according to the present embodiment. As with the news translation experiment, the improvement to WER is the largest with 4.9%, followed by BLEU with 1.5. From this experimental result, it was confirmed that the machine translation apparatus 1 according to the present embodiment is also effective for different domains.

[English-Chinese translation experiment]
We conducted an English-Chinese translation experiment as the last different language pair. The corpus used in the experiment was the one used in the SSMT2007 UK / China Limited Track, and its details are shown in FIG. The conditions for model training, etc. are the same as for the English-Japanese experiment, but the evaluation set is used as it is for parameter optimization only, and the parameters are closed (that is, the evaluation data is I did n’t have another development).

  FIG. 14 shows the experimental results. The number of Chinese references in this experiment is 4 (1 for Japanese reference). The unit of evaluation is a character (kanji). As with the UK-Japan experiment, the improvement to WER is the largest at 4.9%, followed by BLEU at 1.9. From this experimental result, it was confirmed that the machine translation apparatus 1 according to the present embodiment is also effective for different language pairs.

  As described above, according to the machine translation device 1 according to the present embodiment, parameter structure-related parameter learning is performed by generating tree-structure information of translation target text data and introducing restrictions on the tree-structure information. And machine translation using syntax information can be realized. Therefore, it is possible to prevent a problem of data sparseness related to learning data. Also, the machine translation method can be directly incorporated into the left-to-right machine translation at the translation destination, and in this case, it acts as a new restriction on the rearrangement of words. According to the above experimental example, when the phrase-based statistical machine translation is adopted in the machine translation apparatus 1 according to the present embodiment, an improvement of 1.9% for BLEU and 4.9% for WER is achieved in an English-Chinese translation experiment. It was confirmed that it works effectively for global word reordering with a performance improvement of 4.9% in WER.

  In the machine translation apparatus 1 according to the present embodiment, a case where left-to-right machine translation is mainly performed at the translation destination has been described. However, as described above, the machine translation unit 13 performs a plurality of post-translation texts. Data may be generated in advance, and a plurality of post-translation text data generated that match the tree structure constraint conditions may be selected. In that case, since it is not necessary to perform classification by the classification unit 17, the machine translation device may not include the classification unit 17. FIG. 15 shows a configuration of the machine translation apparatus 2 that does not include the classification unit 17 and that the selection unit 19 selects from the plurality of post-translation text data generated by the machine translation unit 13 that matches the constraint condition of the tree structure. FIG. Each component included in the machine translation apparatus 2 is the same as that described above. However, the determination unit 18 determines that it can be realized by converting a plurality of post-translation text data generated by the machine translation unit 13 from the source language of the leaf of the tree structure to the target language and replacing the nodes of the tree structure. The selection unit 19 selects one or more post-translation text data.

  FIG. 16 is a flowchart showing the operation of the machine translation apparatus 2 shown in FIG. In the flowchart of FIG. 16, the processes in steps S101 and S102 are the same as those in the flowchart of FIG.

  (Step S201) The machine translation unit 13 machine translates the text data to be translated. At the time of this machine translation, the machine translation unit 13 generates a plurality of post-translation text data having a parallel translation relationship with the translation target text data. The generated post-translation text data is, for example, a sentence.

(Step S202) The post-translation text data storage unit 14 stores a plurality of post-translation text data generated by the machine translation unit 13 in a predetermined recording medium.
(Step S203) The determination unit 18 sets the counter i to 1.

  (Step S204) The determination unit 18 refers to the tree structure information stored in the tree structure information storage unit 16 and uses the i-th post-translation text data stored in the post-translation text data storage unit 14 as the original tree structure leaf. It is determined whether or not it can be realized by converting the language into the target language and switching the nodes of the tree structure. When the machine translation unit 13 performs word-based statistical machine translation, the determination unit 18 obtains the i-th post-translation text data from the translation target text data corresponding to the post-translation text data. It may be determined whether or not it can be realized by conversion of a word unit from the original language of the leaf of the tree structure to the target language and replacement of the nodes of the tree structure. When the machine translation unit 13 performs the phrase-based statistical machine translation, the determination unit 18 obtains the i-th post-translation text data from the translation target text data corresponding to the post-translation text data. Conversion including the conversion of the phrase unit from the source language to the target language of the leaf of the tree structure, and replacement of the node of the tree structure without dividing the phrase that is the unit translated by the phrase-based statistical machine translation It may be determined whether it can be realized.

  (Step S205) If the determination unit 18 determines that the selection can be realized, the selection unit 19 selects the i-th post-translation text data, and proceeds to step S206. On the other hand, if the determination unit 18 determines that it cannot be realized, the i-th post-translation text data is not selected, and the process proceeds to step S207.

(Step S206) The output unit 20 outputs the i-th post-translation text data selected by the selection unit 19. As described above, the output by the output unit 20 may be output of information indicating the selection result (for example, information for identifying the selected post-translation text data). Then, a series of machine translation processes ends.
(Step S207) The determination unit 18 increments the counter i by 1. Then, the process returns to step S204.

  Note that the flowchart of FIG. 16 is also for explaining the process of machine-translating a sentence, like the flowchart of FIG. Therefore, when a plurality of sentences are continuously machine-translated, the series of processes shown in the flowchart of FIG. 16 may be repeated for the number of sentences.

  In the flowchart of FIG. 16, the case has been described in which the determination regarding other post-translation text data is not performed when the selection unit 19 performs the selection. Therefore, in the flowchart of FIG. 16, only one post-translation text data is selected. On the other hand, even if a selection is made by the selection unit 19, if there is post-translation text data that can be selected and selected for subsequent post-translation text data, the post-translation text data is selected. You may make it do. In that case, two or more post-translation text data can be selected.

  In the flowchart of FIG. 16, the selection unit 19 determines the post-translation text data that can be realized by the conversion from the source language of the leaf of the tree structure to the target language and the replacement of the nodes of the tree structure by the determination unit 18. Among them, it is possible to select a predetermined number or ratio of post-translation text data in descending order of the likelihood calculated at the time of machine translation, or the likelihood is determined in advance. Alternatively, post-translation text data that is higher than the threshold value may be selected. As described above, the selection by the selection unit 19 includes at least post-translation text data determined by the determination unit 18 to be realized by converting the leaf of the tree structure from the source language to the target language and replacing the nodes of the tree structure. If so, there is some degree of arbitraryness in other respects.

  In the flowchart of FIG. 16, the case where any post-translation text data is selected from the plurality of post-translation text data translated by the machine translation unit 13 is described. For example, all post-translation text data is determined. If there is no post-translation text data that can be selected even after performing the above, it is determined that an error has occurred, and the series of processing may be terminated without outputting the post-translation text data. In that case, an error message may be output or not.

  In addition, the specific method for determining whether the translated text data meets the constraints of the tree structure is the above-mentioned “Method of Judgment in Word-Based Statistical Machine Translation” or “Phrase-based Statistical Machine Translation”. This is the same as the specific example in the “determination method”, and the description thereof is omitted.

  In the above-described embodiment, the case where the machine translation apparatus is a stand-alone has been described. However, the machine translation apparatus may be a stand-alone apparatus or a server apparatus in a server / client system. In the latter case, the output unit or the like outputs information via a communication line, for example.

  In the above embodiment, each process or each function may be realized by centralized processing by a single device or a single system, or may be distributedly processed by a plurality of devices or a plurality of systems. It may be realized by doing.

  In the above embodiment, when two or more components included in the machine translation apparatus include a communication device or an input device, the two or more components may have a physically single device. Alternatively, it may have a separate device.

  In the above embodiment, each component may be configured by dedicated hardware, or a component that can be realized by software may be realized by executing a program. For example, each component can be realized by a program execution unit such as a CPU reading and executing a software program recorded on a recording medium such as a hard disk or a semiconductor memory. The software that implements the machine translation apparatus in the above embodiment is the following program. In other words, this program uses a translation model information stored in a translation model information storage unit in which translation model information used in translation from the source language to the target language is stored. Machine translation unit that statistically translates translation target text data stored in a translation target text data storage unit in which translation target text data is stored, and the machine translation unit converts the translation target text data This is information indicating a tree structure of the text data to be translated by parsing the text data to be translated and a text data storage section for storing the text data after translation which is text data in the target language that has been machine translated. A syntax analysis unit for obtaining tree structure information; a tree structure information storage unit for storing the tree structure information; The post-translation text data stored by the post-translation text data storage unit is a tree structure indicated by the tree structure information stored by the tree structure information storage unit, and is obtained from the translation target text data corresponding to the post-translation text data Conversion from the source language of the tree-structured leaf to the target language, a determination unit that determines whether the tree-structured node can be replaced, conversion from the source language of the tree-structured leaf to the target language, This is to function as a selection unit that selects the post-translation text data determined by the determination unit to be realized by replacing nodes in a tree structure, and an output unit that outputs a selection result by the selection unit.

  In the program, the functions realized by the program do not include functions that can be realized only by hardware. For example, a function that can be realized only by hardware such as a modem or an interface card in an output unit that outputs information is not included in at least the function realized by the program.

  Further, this program may be executed by being downloaded from a server or the like, and a program recorded on a predetermined recording medium (for example, an optical disk such as a CD-ROM, a magnetic disk, a semiconductor memory, or the like) is read out. May be executed by

  Further, the computer that executes this program may be singular or plural. That is, centralized processing may be performed, or distributed processing may be performed.

  FIG. 17 is a schematic diagram showing an example of the external appearance of a computer that executes the program and realizes the machine translation apparatus according to the embodiment. The above-described embodiment is realized by computer hardware and a computer program executed on the computer hardware.

  In FIG. 17, the computer system 100 includes a computer 101 including a CD-ROM (Compact Disk Read Only Memory) drive 105 and an FD (Flexible Disk) drive 106, a keyboard 102, a mouse 103, and a monitor 104.

  FIG. 18 is a diagram illustrating a computer system. 18, in addition to the CD-ROM drive 105 and the FD drive 106, a computer 101 includes a CPU (Central Processing Unit) 111, a ROM (Read Only Memory) 112 for storing a program such as a bootup program, A CPU (Random Access Memory) 113 that is connected to the CPU 111 and temporarily stores application program instructions and provides a temporary storage space, a hard disk 114 that stores application programs, system programs, and data, a CPU 111 and a ROM 112. Etc. to each other. The computer 101 may include a network card (not shown) that provides connection to the LAN.

  A program that causes the computer system 100 to execute the functions of the machine translation apparatus according to the above-described embodiment is stored in the CD-ROM 121 or FD 122, inserted into the CD-ROM drive 105 or FD drive 106, and transferred to the hard disk 114. May be. Instead, the program may be transmitted to the computer 101 via a network (not shown) and stored in the hard disk 114. The program is loaded into the RAM 113 at the time of execution. The program may be loaded directly from the CD-ROM 121, the FD 122, or the network.

  The program does not necessarily include an operating system (OS) or a third-party program that causes the computer 101 to execute the functions of the machine translation apparatus according to the above-described embodiment. The program may include only a part of an instruction that calls an appropriate function (module) in a controlled manner and obtains a desired result. How the computer system 100 operates is well known and will not be described in detail.

  Further, the present invention is not limited to the above-described embodiment, and various modifications are possible, and it goes without saying that these are also included in the scope of the present invention.

  As described above, according to the machine translation device or the like according to the present invention, it is possible to introduce restrictions on tree structure information into machine translation without performing learning by syntax information, and there is an effect that a problem of data sparseness does not occur. It is useful as a machine translation device that performs machine translation.

1 is a block diagram showing the configuration of a machine translation apparatus according to Embodiment 1 of the present invention. The figure which shows an example of the tree structure in the embodiment Flowchart showing the operation of the machine translation apparatus according to the embodiment The figure which shows an example of the tree structure in the embodiment The figure which shows an example of the tree structure in the embodiment The figure which shows an example of the tree structure in the embodiment The figure which shows an example of a response | compatibility with the translation object text data and translated flag in the embodiment The figure which shows an example of the classification result by the classification | category part in the embodiment The figure for demonstrating the English-Japanese corpus in the experiment example of the embodiment The figure which shows an example of the experimental result in the same embodiment The figure for demonstrating the English-Japanese corpus in the experiment example of the embodiment The figure which shows an example of the experimental result in the same embodiment The figure for demonstrating the English-Chinese corpus in the experiment example of the embodiment The figure which shows an example of the experimental result in the same embodiment The block diagram which shows the other structure of the machine translation apparatus in the embodiment Flowchart showing the operation of the machine translation device in the same embodiment Schematic diagram showing an example of the appearance of the computer system in the embodiment The figure which shows an example of a structure of the computer system in the embodiment

Explanation of symbols

DESCRIPTION OF SYMBOLS 1, 2 Machine translation apparatus 11 Translation object text data storage part 12 Translation model information storage part 13 Machine translation part 14 Post-translation text data storage part 15 Syntax analysis part 16 Tree structure information storage part 17 Classification part 18 Judgment part 19 Selection part 20 Output section

Claims (7)

A translation target text data storage unit that stores translation target text data that is text data of a source language to be translated;
A translation model information storage unit for storing translation model information used in translation from the source language to the target language;
Using the translation model information, a machine translation unit that statistically machine translates the text data to be translated,
A post-translation text data storage unit that stores post-translation text data that is text data of a target language obtained by machine-translating the text data to be translated by the machine translation unit;
A syntax analysis unit that obtains tree structure information that is information indicating a tree structure of the translation target text data by parsing the translation target text data;
A tree structure information storage unit for storing the tree structure information;
The post-translation text data stored by the post-translation text data storage unit is a tree structure indicated by the tree structure information stored by the tree structure information storage unit, and is obtained from the translation target text data corresponding to the post-translation text data. A determination unit that determines whether or not it can be realized by converting the leaf of the tree structure from the source language to the target language and replacing the nodes of the tree structure;
A selection unit that selects the post-translation text data determined by the determination unit to be realized by conversion from a source language of a leaf of the tree structure to a target language, and replacement of nodes of the tree structure;
An output unit that outputs a selection result by the selection unit. The machine translation unit performs statistical machine translation based on words,
The determination unit is a tree structure indicated by the tree structure information accumulated by the tree structure information accumulation unit, the post-translation text data accumulated by the translated text data accumulation unit, and a translation target corresponding to the translated text data The machine translation device according to claim 1, wherein it is determined whether or not it can be realized by conversion of a word unit from a source language of a leaf of a tree structure obtained from text data to a target language and replacement of nodes of the tree structure. The machine translation unit performs a phrase-based statistical machine translation,
The determination unit is a tree structure indicated by the tree structure information accumulated by the tree structure information accumulation unit, the post-translation text data accumulated by the translated text data accumulation unit, and a translation target corresponding to the translated text data The conversion that includes the conversion of the phrase unit from the source language of the tree-structured leaf obtained from the text data to the target language, and the phrase that is the unit translated by the phrase-based statistical machine translation, without dividing The machine translation apparatus according to claim 1, wherein it is determined whether or not it can be realized by exchanging nodes of a tree structure. The machine translation unit performs machine translation of left to right at the translation destination,
The post-translation text data includes text data being translated,
Each subtree in the tree structure indicated by the tree structure information stored by the tree structure information storage unit is untranslated indicating that the subtree includes only untranslated leaves or consists only of untranslated subtrees; Translated to indicate that the subtree contains only translated leaves or consists only of translated subtrees, and the subtree contains only translated and untranslated leaves, or untranslated and translated A classifying unit that classifies only that sub-tree or is in translation indicating that only one sub-tree is being translated,
When the subtree including two or more subtrees under translation appears using the classification result of the classification unit, the determination unit converts the translated text data from the source language of the tree-structured leaf to the target language. Judging that it can not be realized by conversion to and replacement of nodes of the tree structure,
The machine translation unit adds the post-translation text data after the addition of the text so that the post-translation text data is selected by the selection unit when adding a new post-translation target language text to the post-translation text data. The machine translation device according to any one of claims 1 to 3. The machine translation unit generates a plurality of post-translation text data corresponding to the text data to be translated,
1 or 2 or more which the said judgment part judged that the said selection part was realizable by conversion from the original language of the leaf of the said tree structure to the target language, and replacement | exchange of the node of the said tree structure from several translated text data The machine translation apparatus according to claim 1, wherein the post-translation text data is selected. Translation target text data storage unit storing translation target text data which is text data of the source language to be translated, and translation model information storage unit storing translation model information used in translation from the source language to the target language A machine translation method processed using a machine translation unit, a post-translation text data storage unit, a syntax analysis unit, a tree structure information storage unit, a determination unit, a selection unit, and an output unit. ,
A machine translation step in which the machine translation unit statistically machine translates the text data to be translated using the translation model information;
The post-translation text data storage unit stores post-translation text data that is post-translation text data that is text data of a target language obtained by machine translating the text data to be translated in the machine translation step;
A syntax analysis step in which the syntax analysis unit obtains tree structure information which is information indicating a tree structure of the translation target text data by parsing the translation target text data;
A tree structure information storage step in which the tree structure information storage unit stores the tree structure information;
The determination unit is a tree structure indicated by the tree structure information stored in the tree structure information storage step, the post-translation text data stored in the post-translation text data storage step, and a translation target corresponding to the post-translation text data A determination step for determining whether the tree structure leaf obtained from the text data can be realized by conversion from the source language to the target language and replacing the nodes of the tree structure;
A selection step of selecting the post-translation text data determined in the determination step that the selection unit can be realized by conversion from the source language of the leaf of the tree structure to a target language and replacement of nodes of the tree structure;
An output step in which the output unit outputs a selection result in the selection step. Computer
Translation target text data, which is text data in the source language to be translated, using translation model information stored in a translation model information storage unit in which translation model information used in translation from the source language to the target language is stored A machine translation unit that statistically machine translates the text data to be translated stored in the text data storage unit to be translated,
A post-translation text data storage unit that stores post-translation text data that is text data in a target language obtained by machine-translating the text data to be translated by the machine translation unit;
A syntax analysis unit that obtains tree structure information that is information indicating a tree structure of the translation target text data by parsing the translation target text data;
A tree structure information storage unit for storing the tree structure information;
The post-translation text data stored by the post-translation text data storage unit is a tree structure indicated by the tree structure information stored by the tree structure information storage unit, and is obtained from the translation target text data corresponding to the post-translation text data. A determination unit that determines whether or not it can be realized by converting the leaf of the tree structure from the source language to the target language and replacing the nodes of the tree structure;
A selection unit that selects the post-translation text data determined by the determination unit to be realized by conversion from a source language of a leaf of the tree structure to a target language, and replacement of nodes of the tree structure;
The program for functioning as an output part which outputs the selection result by the said selection part.

Images