JP2015228170A - Machine translation method, machine translation program and machine translation apparatus - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a machine translation method, a machine translation program and a machine translation apparatus capable of improving the accuracy in translation of a document component acquired by dividing a sentence as a translation object.SOLUTION: A sentence as a translation object is divided into plural document components. An analysis candidate which is applied to the translation of a document component to be translated is selected from one or more analysis candidates based on the similarity among the analysis candidates of the document components and a corpus of the analyzed document components to perform the translation.

Description

  The present invention relates to a machine translation technique for translating a first language sentence into a second language sentence.

  2. Description of the Related Art Conventionally, machine translation for translating a sentence in a first language to be translated into a sentence in a second language for translation purposes different from the first language using a bilingual dictionary of the first language and the second language is known. (For example, see Patent Documents 1 to 5.)

  There are a variety of documents that require machine translation. Among them, there is a great demand for translation of industrial documents such as operation manuals, laws, contracts, and patent specifications.

  In general, industrial documents are characterized by containing many long sentences. In machine translation, if the sentence to be translated becomes longer, the ambiguity of interpretation increases, which causes problems such as an increase in processing time and an increase in analysis errors, and the translation quality tends to be lowered. Therefore, there is a great need for a technique for performing high-accuracy (high-quality) translation of long sentences.

  For example, in order to translate claims (claims) of English patent specifications with many long sentences, paying attention to the specific style of the claims and dividing each input claim (claim) into components such as constituent requirements A technique for obtaining an appropriate translation result is disclosed (see, for example, Patent Document 6).

  That is, first, an English claim inputted by collation with a pattern (predetermined rule) for dividing a sentence-specific sentence style stored in advance is divided. Next, each divided part is analyzed, and a hierarchical structure (dependency relationship) is created from the modification relationship. Then, reverse the hierarchical structure and correct the order of Japanese words. Finally, each hierarchy is translated and the results are output. As a result, the translation quality is improved because the ambiguity of the interpretation can be reduced as much as the translation unit is shortened as compared with the case where the entire sentence is translated as it is.

JP-A-8-278970 JP 2008-233955 A Special table 2010-535377 JP 2009-230561 A JP 2012-79081 A Japanese Patent Laid-Open No. 9-293075

FIG. 1 is a diagram for explaining the problems of the prior art.
In FIG. 1, an input sentence 100 “A device comprising text storing means for new data and data ranking part for the stored data” is a sentence in the first language to be translated. A pattern 200 ““ part P0 ”comprising“ part P1 ”and“ part P2 ”.” As a predetermined rule for dividing a long sentence is stored in advance.

  The conventional machine translation device first matches the input sentence 100 with the pattern 200, and the input sentence 100 is "A device", "comprising", "text storing means for new data", "and", "data ranking part for the Divided into “stored data” and “.”. “Part P0” of the pattern 200 corresponds to the document part 110 “A device”. “Part P1” of the pattern 200 corresponds to the document part 120 “text storing means for new data”. “Part P2” of the pattern 200 corresponds to the document part 130 “data ranking part for the stored data”.

  Next, the conventional machine translation apparatus analyzes each divided part. As an analysis result of the document part 120 “text storing means for new data”, the dependency from the phrase 121 “text” to the phrase 122 “storing”, the dependency from the phrase 122 “storing” to the phrase 123 “means”, and the phrase The dependency from 124 “for new data” to the phrase 123 “means” is obtained. Similarly, as an analysis result of the document part 130 “data ranking part for the stored data”, the phrase 134 “for the stored data” depends on the phrase 133 “part”, and the phrase 133 “part” to the phrase 132 “ranking”. And the dependency from the phrase 132 “ranking” to the phrase 131 “data”.

  Then, the phrases 121 to 124 are translated using the bilingual dictionary, and "translation part 210" new data text storage means "is combined by combining" for new data "," text "," storage ", and" means "in the order of dependency. obtain. Similarly, phrases 131 to 134 are translated using a bilingual dictionary and combined in the order of dependency to obtain a translation component 230 “data for sorting stored data components”. The document part 110 “A device” obtains a translation part 250 “device” as a translation result. These translations are done independently.

  Then, the conventional machine translation apparatus combines these translation results with the translation results of the standard parts (“comprising” and “and”) in the pattern 200 (translation parts 220 “to” and translation parts 240 “includes”). To do.

  As described above, the translation of the document part 110 “A device”, the translation of the document part 120 “text storing means for new data”, and the translation of the document part 130 “data ranking part for the stored data” are independent. . In this example, the translation result of the document part 130 “data ranking part for the stored data” that should be translated as “data sorting part for stored data” is incorrect.

  That is, in the conventional machine translation, each document part is independently translated one by one in the translation of each document part obtained by dividing the sentence to be translated. For this reason, there has been a problem that the ambiguity cannot be resolved in consideration of the structure of the document parts in the entire sentence to be translated.

  In one aspect, the present invention provides a machine translation method, a machine translation program, and a machine translation apparatus capable of improving the translation accuracy of document parts in translation of each document part obtained by dividing a sentence to be translated. The purpose is to provide.

  In one proposal, the machine translation method is executed by a computer of a machine translation device that translates a translation target sentence described in a first language into a translation target sentence described in a second language. Based on a predetermined rule, a document part is created by dividing a sentence to be translated. One or more analysis candidates are created for each of the created document parts. When a parallel structure is recognized between the created document parts, a combination of the created analysis candidates is created for the analysis candidates created for each document part constituting the parallel structure. A parallel similarity value indicating the similarity between the analysis candidates constituting the created combination is calculated. A corpus value indicating the appearance degree of the created analysis candidate is calculated based on the analyzed corpus that is a result of analysis in the past. Based on the calculated parallel similarity value and the calculated corpus value, an analysis candidate to be applied when translating the document part to be translated is selected from the created one or more analysis candidates.

  According to the embodiment, translation accuracy can be improved in translation of each document part obtained by dividing a sentence to be translated.

It is a figure explaining the problem of a prior art. It is a figure explaining the characteristic of this Embodiment. It is a figure which shows the functional block of the machine translation apparatus which concerns on 1st Embodiment. It is a flowchart which shows the flow of the machine translation process which the computer of a machine translation apparatus performs. It is a figure which shows the example of a pattern. It is a figure explaining how to create an analyzed corpus. It is a figure which shows the example of an analyzed corpus database. It is a figure explaining how to obtain a corpus value. It is a figure explaining the conditions which calculate a parallel similarity value. It is a figure which shows the example of calculation of a parallel similarity value. It is a figure which shows the example of calculation of a comprehensive value. It is a figure which shows the example of the translation result translated correctly. It is a figure which shows the functional block of the machine translation apparatus which concerns on 2nd Embodiment. It is a figure which shows the example when a parallel structure does not exist between document parts. It is a figure which shows an example in case a parallel structure exists between document parts. It is a figure which shows the functional block of the machine translation apparatus which concerns on 3rd Embodiment. It is a block diagram of information processing apparatus.

Hereinafter, embodiments for carrying out the present invention will be described in detail with reference to the drawings.
FIG. 2 is a diagram for explaining the features of the present embodiment.

The present embodiment will be described using the example shown in FIG.
Since the process up to generating the document part 120 is the same as that described with reference to FIG. 1, the description thereof is omitted.

  In FIG. 2, in the present embodiment, a plurality of analysis candidates 120A (P1-a) and 120B (P1-b) are acquired for a document part 120 “text storing means for new data”. Similar to the example shown in FIG. 1, the analysis candidate 120 </ b> A includes a dependency from the phrase “text” to the phrase “storing”, a dependency from the phrase “storing” to the phrase “means”, and the phrase “for new data”. It is a dependency on the phrase “means”. The analysis candidate 120B is different from the analysis candidate 120A, the dependency from the phrase “for new data” to the phrase “means”, the dependency from the phrase “means” to the phrase “storing”, and the phrase “storing” It is a dependency on "text". Similarly, a plurality of analysis candidates 130A (P2-a) and 130B (P2-b) are acquired for the document part 130 “data ranking part for the stored data”.

  Next, based on the pattern for dividing the document to be translated, the divided document parts 120 and the document parts 130 are in a parallel relationship or arranged with “and” and “or” constituting the pattern interposed therebetween. Judgment can be made by sentence. The sentence parts (120, 130) connected in parallel in the sentences constituting one sentence, ie, “and”, “or”, etc., have a similar structure for the parallel requirement. There are many cases. Focusing on the structural features of this document, a “parallel similarity value” is given to the parallel combination of analysis candidates in accordance with the similarity of the document structure (analysis candidate) between the document parts constituting the parallel relationship. In the present embodiment, a plurality of analysis candidates 120A and 120B are acquired for the document part 120, and a plurality of analysis candidates 130A and 130B are acquired for the document part 130. A “parallel similarity value” is given according to the combination of the analysis candidate of the document part 120 (120A and 120B), the combination of the analysis candidate of the document part 130 (130A and 130B), and the commonality of the analysis candidate of each combination.

  In addition, ambiguity may not be sufficiently resolved only by parallel similarity values. Therefore, a “corpus value” is given to each analysis candidate based on the frequency in the analyzed corpus of the target field that has been classified in advance according to the sentence type (field of legal documents, papers, operation manuals, etc.). Here, the corpus value is a value based on the frequency at which candidates appear in the corpus, and how to obtain the corpus value will be described later.

  Then, based on the parallel similarity value and the corpus value, (P1-a) and (P2-b), which are parallel combinations having the highest total value, are selected. As a result, with respect to the document part 130 “data ranking part for the stored data”, the parallel similarity value between the document part 120 and the document part 120 constituting the parallel document is high, and the corpus value in the target field is high (it can be determined to be correct). As a result, the translation component 330 “data sorting unit for stored data” is obtained.

FIG. 3 is a diagram illustrating functional blocks of the machine translation apparatus according to the first embodiment.
In FIG. 3, the machine translation apparatus 500A translates a translation target sentence described in the first language into a translation target sentence described in the second language. The machine translation apparatus 500A includes an input sentence division unit 501, a plurality of analysis candidate creation units 502, a parallel combination creation unit 503, a parallel similarity value assignment unit 504, a frequency acquisition unit 505, a corpus value assignment unit 506, and a total value calculation unit 508. Prepare. Furthermore, the machine translation apparatus 500A includes a candidate selection unit 509, a component translation unit 510, and a combination unit 511.

  Based on the pattern 200 (predetermined rule), the input sentence dividing unit 501 divides the translation target sentence that is the input sentence 100 to create a document part. For example, as described with reference to FIG. 1, the translation target sentence that is the input sentence 100 and the pattern 200 are collated, and a document part corresponding to the part constituting the pattern is created.

  The multiple analysis candidate creation unit 502 creates one or more analysis candidates for each of the document parts created by the input sentence division unit 501. For example, a plurality of analysis candidates 120A and 120B are created from the document part 120 in the above example, and a plurality of analysis candidates 130A and 130B are created from the document part 130.

  When it is determined that the document part created by the input sentence splitting unit 501 has a parallel structure, the parallel combination creation unit 503 determines each of the plurality of document parts constituting the parallel structure created by the multiple analysis candidate creation unit 502. A combination of a plurality of analysis candidates (120A, 120B / 130A, 13B) is created. For example, the analysis candidate 120A and the analysis candidate 130A, the analysis candidate 120A and the analysis candidate 130B, the analysis candidate 120B and the analysis candidate 130A, and the analysis candidate 120B and the analysis candidate 130B are created.

  The parallel similarity value assigning unit 504 calculates a parallel similarity value indicating the similarity between the analysis candidates constituting the combination created by the parallel combination creating unit 503. Specific calculation of the parallel similarity value will be described later.

  The frequency acquisition unit 505 acquires frequency data from the analyzed corpus database (DB) 507 for each of the plurality of analysis candidates generated by the multiple analysis candidate creation unit 502. Here, the analyzed corpus is a result of analysis and translation in the past, and is a corpus classified according to sentence types (fields of legal documents, papers, operation manuals, etc.).

  The corpus value assigning unit 506 calculates a corpus value indicating the appearance degree of the analysis candidates created by the multiple analysis candidate creation unit 502 based on the frequency data acquired for each analysis candidate by the frequency acquisition unit 505. Specific calculation of the corpus value will be described later.

  The total value calculation unit 508 calculates a total value based on the parallel similarity value calculated by the parallel similarity value giving unit 504 and the corpus value calculated by the corpus value giving unit 506. The calculation of the total value will be described later.

  Based on the total value calculated by the total value calculation unit 508, the candidate selection unit 509 analyzes candidates to be applied when translating the document parts to be translated from the plurality of analysis candidates created by the multiple analysis candidate creation unit 502. Select.

  The component translation unit 510 translates the document component based on the analysis candidate selected by the candidate selection unit 509.

  Then, the combining unit 511 combines the translated document parts translated by the part translating unit 510 and outputs them as output sentences.

  FIG. 4 is a flowchart showing the flow of machine translation processing executed by the computer of the machine translation apparatus. FIG. 5 is a diagram illustrating an example of a pattern. FIG. 6 is a diagram for explaining how to create an analyzed corpus. FIG. 7 is a diagram illustrating an example of an analyzed corpus database. FIG. 8 is a diagram for explaining how to obtain the corpus value. FIG. 9 is a diagram illustrating conditions for calculating parallel similarity values. FIG. 10 is a diagram illustrating an example of calculating parallel similarity values. FIG. 11 is a diagram illustrating an example of calculating the total value. FIG. 12 is a diagram illustrating an example of a translation result that is correctly translated.

  An example in which a patent document written in English is translated into a patent document written in Japanese will be described with reference to FIGS.

  First, in step S401, the input sentence dividing unit 501 of the machine translation device 500A sequentially matches the input sentence 100 with a pattern as shown in FIG. For example, the input sentence 100 is “A device comprising text storing means for new data and data ranking part for the stored data”. In the example shown in FIG. 5, the input sentence 100 matches the patterns “part P0” comprising “part P1” and “part P2”. Then, the input sentence dividing unit 501 creates a document part corresponding to each part constituting this pattern. That is, the input sentence dividing unit 501 performs “A device” corresponding to “part P0”, “text storing means for new data” corresponding to “part P1”, and “data ranking part for” corresponding to “part P2”. the stored data ". Note that “present” and “none” shown at the right end of the pattern illustrated in FIG. 5 indicate the presence or absence of a parallel structure in the created document part.

  In step S402, the multiple analysis candidate creation unit 502 has at least one or more document parts arranged with “and” or “or” constituting a pattern between each of the document parts created in step S401. Create analysis candidates. For example, analysis candidates 120A and 120B as described above with reference to FIG. 2 are created from the document part “text storing means for new data”. From the document part “data ranking part for the stored data”, the analysis candidates 130A and 130B as described above with reference to FIG. 2 are created.

  In step S <b> 403, the frequency acquisition unit 505 acquires frequency data from the analyzed corpus DB 507 for each of the plurality of analysis candidates generated by the multiple analysis candidate creation unit 502. Then, the corpus value assigning unit 506 calculates a corpus value indicating the appearance degree of the analysis candidate created in step S402 based on the frequency data acquired for each of the plurality of analysis candidates, and assigns it to the analysis candidate.

The analyzed corpus DB 507 is created as follows.
Analyzed corpus is a database in which the results of past translations are accumulated, classified according to sentence type (fields such as legal documents, papers, operation manuals, etc.), and the source document and the translated document correspond. Attached and registered. The registration data is registered including the analysis results at the time of translation. For example, the case of a document part “text storing means for new data” as shown in FIG.

  As an analysis result, four phrases “text”, “storing”, “means”, and “for new data” are obtained. Here, the phrases “text” and “means” are nouns, and the phrase “for new data” is a noun phrase. However, the phrase “storing” can be an adjective or a verb. Depending on whether the phrase “storing” is an adjective or a verb, an analysis result as shown in FIG. 6B is created or an analysis result as shown in FIG. 6C is created. The analysis results as shown in FIGS. 6B and 6C are determined to be ambiguous.

  When the document part “communication means for new data” as shown in FIG. 6D is analyzed, three phrases “communication”, “means”, and “for new data” are obtained. Here, the phrases “communication” and “means” are nouns, and the phrase “for new data” is a noun phrase. It is determined that the analysis result as shown in FIG. 6D is not ambiguous.

Along with the above analysis results, phrases such as words and phrases that are finally involved are the main parts.
Only the FIG. 6E in which the ambiguous analysis result is excluded from the analysis results as shown in FIGS. 6B, 6C, and 6E is the analyzed corpus.

  In this way, as shown in FIG. 7, an analyzed corpus DB 507 in which an unambiguous analysis result is stored is created.

The corpus value is calculated as follows.
First, the input sentence 100 “A device comprising text storing means for new data and data ranking part for the stored data” is divided based on the corresponding pattern as described above. The divided input sentence 100 includes a part P1 “text storing means for new data” and a part P2 “data ranking part for the stored data” ((A) in FIG. 8).

  Two analysis candidates (P1-a) and (P1-b) as shown in FIGS. 8B and 8C are obtained from the component P1. From the component P2, two analysis candidates (P2-a) and (P2-b) as shown in (D) and (E) of FIG. 8 are obtained.

  The analyzed corpus DB is referred to for these analysis candidates (P1-a) (P1-b) (P2-a) (P2-b). For example, for the analysis candidate (P1-a), the main part frequency and the main part phrase frequency of the analysis candidate are acquired as the main part frequency “538” and the phrase frequency “19”, respectively. The corpus value is obtained by calculating constant α × primary frequency + constant β × phrase frequency. Here, α and β are determined empirically. For example, α = 1 and β = 4. In this case, the corpus value of the analysis candidate (P1-a) is “614”. Similarly, the corpus value of the analysis candidate (P1-b) is “7”. The corpus value of the analysis candidate (P2-a) is “23”. The corpus value of the analysis candidate (P2-b) is “925”.

  Next, in step S404 in FIG. 4, the parallel combination creation unit 503 creates a combination of analysis candidates created in step S402 for the parallel structure between document parts created in step S401. For example, a combination of (P1-a) shown in (B) of FIG. 8 and (P2-a) shown in (D) of FIG. 8, or (P1-a) and (E) of FIG. A combination of (P2-b) is created.

  In step S405, the combinations created in step S404 are extracted in order.

  In step S406, the parallel similarity value assigning unit 504 calculates a parallel similarity value indicating the similarity between the analysis candidates constituting the combination created in step S404, and assigns it between the analysis candidates.

  The parallel similarity value is calculated as a ratio of the number of matching predetermined conditions to the total number of phrases constituting each analysis result. The number of matches of the predetermined condition is the number of pairs in which the part of speech and the dependency direction match between phrases appearing at the same position in each analysis result. For example, the example shown in FIG. 9A includes a dependency from the noun “text” to the verb “storing” (left side) and a dependency from the noun “ranking” to the noun “data”. (Right side). In the example shown in FIG. 9A, neither the part of speech nor the dependency direction matches. The example shown in FIG. 9B includes a dependency from the verbal noun “storing” to the noun “means” (left) and a dependency from the noun “part” to the verb “ranking” (right) ). In the example shown in FIG. 9B, neither the part of speech nor the dependency direction matches. The example shown in FIG. 9C is a dependency from the noun phrase “for new data” to the noun “means” (left side), and from the noun phrase “for the stored data” to the noun “part”. And (right). In the example shown in FIG. 9C, the part of speech and the dependency direction are the same.

  Therefore, in the case of the combination of analysis candidates (P1-a) and (P2-a), the total number of phrases constituting each analysis candidate is 6, as shown in FIG. The number of matches is 1. Therefore, when the parallel similarity value is expressed as a percentage, (1/6) × 100 = 17 (%). Similarly, in the case of a combination of analysis candidates (P1-a) and (P2-b), the total number of phrases constituting each analysis candidate is 6, as shown in FIG. The number of direction matches is 3. Therefore, when the parallel similarity value is expressed as a percentage, (3/6) × 100 = 50 (%). In the case of the combination of analysis candidates (P1-b) and (P2-a), as shown in FIG. 10C, the total number of phrases constituting each analysis candidate is 6, and the direction of dependency is the same. The number is 3. Therefore, when the parallel similarity value is expressed as a percentage, (3/6) × 100 = 50 (%). In the case of a combination of analysis candidates (P1-b) and (P2-b), as shown in FIG. 10 (D), the total number of phrases constituting each analysis candidate is 6, and the direction of dependency is consistent. The number is 1. Therefore, when the parallel similarity value is expressed as a percentage, (1/6) × 100 = 17 (%).

  In step S407, the total value calculation unit 508 calculates a total value based on the parallel similarity value calculated in step S406 and the corpus value calculated in step S403.

  The total value is obtained, for example, by calculating (constant γ) × (parallel similarity value) + (constant δ) × (total of each corpus value). Here, γ and δ are determined empirically, for example, γ = 1 and δ = 0.1. For example, in the case of a combination of analysis candidates (P1-a) and (P2-a), the parallel similarity value of analysis candidates (P1-a) and (P2-a) is 17, as shown in FIG. The corpus value of the analysis candidate (P1-a) is 614, and the corpus value of the analysis candidate (P2-a) is 23. Therefore, the total value calculated based on the above formula is 81. Similarly, in the case of the combination of the analysis candidates (P1-a) and (P2-b), the parallel similarity value is 50, the corpus value of the analysis candidate (P1-a) is 7, and the analysis candidate (P2-b) The corpus value is 925. Therefore, the total value calculated based on the above formula is 204. In the case of the combination of the analysis candidates (P1-b) and (P2-a), the parallel similarity value is 50, the corpus value of the analysis candidate (P1-b) is 7, and the corpus value of the analysis candidate (P2-a) is 23. Therefore, the total value calculated based on the above formula is 53. In the case of the combination of the analysis candidates (P1-b) and (P2-b), the parallel similarity value is 17, the corpus value of the analysis candidate (P1-b) is 7, and the corpus value of the analysis candidate (P2-b) is 925. Therefore, the total value calculated based on the above formula is 110.

  In step S408, it is determined whether the total value has been calculated up to the last combination. When the total value is calculated up to the last combination (step S408: YES), in step S409, the candidate selection unit 509 translates from the plurality of analysis candidates created in step S402 based on the total value calculated in step S407. Select an analysis candidate to be applied to the translation of the target document part. In the example illustrated in FIG. 11, the combination of analysis candidates (P1-a) and (P2-b) that has the maximum value among the calculated total values is selected.

  In step S410, the component translation unit 510 translates the document component based on the analysis candidate selected in step S409.

  In step S411, the combining unit 511 combines the translated document parts translated in step S410 and outputs them as an output sentence as shown in FIG.

  As described above, according to the first embodiment, in the translation of each document part obtained by dividing the sentence to be translated, the parallel similarity value between the document parts in parallel relation and the sentence type are converted. Based on the corpus value in the corresponding field, an analysis candidate for each document part is selected. This selection eliminates the ambiguity in consideration of the structure of the document parts in the entire sentence to be translated, and improves the translation accuracy.

  For example, 30% or more of patent documents are composed of long sentences of 40 morphemes or more. In such a long sentence of 40 morphemes or more, a parallel structure appears in 90% or more of input sentences. Therefore, use of the information of the parallel structure is effective.

  FIG. 13 is a diagram illustrating functional blocks of the machine translation apparatus according to the second embodiment. FIG. 14 is a diagram illustrating an example in which there is no parallel structure between document parts. FIG. 15 is a diagram illustrating an example in which a parallel structure exists between document parts.

  In FIG. 13, the machine translation apparatus 500B translates a translation target sentence described in the first language into a translation target sentence described in the second language. The machine translation apparatus 500B includes an input sentence dividing unit 501, a multiple analysis candidate creating unit 502, a parallel structure presence / absence determining unit 1501, and a first processing control unit 1502. Furthermore, the machine translation apparatus 500B includes a parallel combination creation unit 503, a parallel similarity value assignment unit 504, a frequency acquisition unit 505, a corpus value assignment unit 506, an overall value calculation unit 508, a candidate selection unit 509, a component translation unit 510, and a combination unit 511.

  Except for the parallel structure presence / absence determining unit 1501 and the first processing control unit 1502, the description is omitted because it is the same as the machine translation apparatus 500A described with reference to FIG.

  The parallel structure presence / absence determination unit 1501 determines whether or not a parallel structure exists between the analysis candidates created by the multiple analysis candidate creation unit 502. For example, as illustrated in FIG. 14, when an input sentence 100 such as “A device comprising text storing means for new data” is collated with the pattern ““ part P0 ”comprising“ part P1 ””, There is no document part having a parallel structure with “P0” or “part P1”. On the other hand, as shown in FIG. 15, an input sentence 100 such as “A device comprising text storing means for new data and data ranking part for the stored data” is represented by the pattern “part P0” comprising “part. When collated with “P1” and “part P2”, “part P1” and “part P2” are document parts having a parallel structure.

  Then, when the parallel structure presence / absence determining unit 1501 determines that a parallel structure exists, the first processing control unit 1502 controls the parallel combination creating unit 503 and the parallel similarity value adding unit 504 to perform processing.

  Thereby, the machine translation apparatus 500B selects an appropriate analysis candidate from a plurality of analysis candidates for the document parts constituting the parallel structure described with reference to FIG. 4 only for the input sentence having the parallel structure. Processing is executed. Therefore, it becomes possible to more efficiently translate a sentence to be translated having a parallel structure that requires processing like the present plan according to an appropriate analysis candidate.

FIG. 16 is a diagram illustrating functional blocks of the machine translation apparatus according to the third embodiment.
In FIG. 16, the machine translation device 500C translates a translation target sentence described in the first language into a translation target sentence described in the second language. The machine translation apparatus 500C includes an input sentence division unit 501, a plurality of analysis candidate creation units 502, a parallel combination creation unit 503, a parallel similarity value assignment unit 504, a parallel similarity value threshold determination unit 1601, a second processing control unit 1602, and a frequency acquisition. A unit 505, a corpus value assigning unit 506, a total value calculating unit 508, a candidate selecting unit 509, a component translating unit 510, and a combining unit 511.

  Except for the parallel similarity threshold determination unit 1601 and the second processing control unit 1602, the description is omitted because it is the same as the machine translation device 500A described with reference to FIG.

  The parallel similarity value threshold determination unit 1601 determines whether or not the parallel similarity value calculated by the parallel similarity value giving unit 504 is equal to or greater than a predetermined threshold value.

  The second processing control unit 1602 controls the total value calculation unit 508 to calculate the total value when the parallel similarity value threshold determination unit 1601 determines that the parallel similarity value is equal to or greater than the threshold value.

  As a result, only the input candidate having a high parallel similarity, that is, the analysis candidate having a high similarity of the analysis candidates between the document parts constituting the parallel structure is selected as the analysis candidate selection target. Processing can be reduced.

FIG. 17 is a configuration diagram of the information processing apparatus.
The machine translation apparatuses 500A, 500B, and 500C in FIG. 1 can be realized using an information processing apparatus (computer) as shown in FIG. 17, for example. 17 includes a CPU (Central Processing Unit) 1701, a memory 1702, an input device 1703, an output device 1704, an external recording device 1705, a medium driving device 1706, and a network connection device 1707. These are connected to each other by a bus 1708.

  The memory 1702 is, for example, a semiconductor memory such as a ROM (Read Only Memory), a RAM (Random Access Memory), or a flash memory, and stores programs and data used for machine translation processing executed by the machine translation devices 500A, 500B, and 500C. Store. For example, the CPU 1701 performs the above-described machine translation process by executing a program using the memory 1702.

  The input device 1703 is, for example, a keyboard, a pointing device, or the like, and is used for inputting instructions and information from an operator. The output device 1704 is, for example, a display device, a printer, a speaker, and the like, and is used to output an inquiry to the operator and a processing result.

  The external recording device 1705 is, for example, a magnetic disk device, an optical disk device, a magneto-optical disk device, a tape device, or the like. The external recording device 1705 includes a hard disk drive. The information processing apparatus can store programs and data in the external recording apparatus 1705 and load them into the memory 1702 for use.

  The medium driving device 1706 drives a portable recording medium 1709 and accesses the recorded contents. The portable recording medium 1709 is a memory device, a flexible disk, an optical disk, a magneto-optical disk, or the like. The portable recording medium 1709 includes a compact disk read only memory (CD-ROM), a digital versatile disk (DVD), a universal serial bus (USB) memory, and the like. An operator can store programs and data in the portable recording medium 1709 and load them into the memory 1702 for use.

  As described above, computer-readable recording media for storing programs and data used for machine translation processing include physical (non-temporary) such as the memory 1702, the external recording device 1705, and the portable recording medium 1709. Recording medium).

  The network connection device 1707 is a communication interface that is connected to the communication network 1710 and performs data conversion accompanying communication. The information processing apparatus can receive a program and data from an external apparatus via the network connection apparatus 1707, and can use them by loading them into the memory 1702.

  Although the disclosed embodiment and its advantages have been described in detail, those skilled in the art can make various modifications, additions and omissions without departing from the scope of the present invention as explicitly described in the claims. I will.

The following notes are further disclosed with respect to the embodiment described with reference to the drawings.
(Appendix 1)
A machine translation method executed by a computer of a machine translation device that translates a translation target sentence described in a first language into a translation target sentence described in a second language,
Based on a predetermined rule, create a document part by dividing the sentence to be translated,
Creating one or more analysis candidates for each of the created document parts;
When a parallel structure is recognized between the created document parts, a combination of analysis candidates is created for the analysis candidates created for each document part constituting the parallel structure,
Calculate a parallel similarity value indicating the similarity between the analysis candidates constituting the created combination,
Based on the analyzed corpus that is the result of analysis in the past, a corpus value indicating the appearance degree of the created analysis candidate is calculated,
Based on the calculated parallel similarity value and the calculated corpus value, select an analysis candidate to be applied when translating a document part to be translated from the created one or more analysis candidates.
A machine translation method characterized by the above.
(Appendix 2)
Determining whether a parallel structure exists between the created analysis candidates,
If it is determined that the parallel structure exists, control to create a combination of the analysis candidates,
The machine translation method according to Supplementary Note 1, wherein:
(Appendix 3)
Determining whether the calculated parallel similarity value is equal to or greater than a predetermined threshold;
Control to calculate a corpus value for the analysis candidate determined that the parallel similarity value is equal to or greater than the threshold,
The machine translation method according to Supplementary Note 1, wherein:
(Appendix 4)
In a computer of a machine translation device that translates a translation target sentence described in a first language into a translation target sentence described in a second language,
Based on a predetermined rule, create a document part by dividing the sentence to be translated,
Creating one or more analysis candidates for each of the created document parts;
When a parallel structure is recognized between the created document parts, a combination of analysis candidates is created for the analysis candidates created for each document part constituting the parallel structure,
Calculate a parallel similarity value indicating the similarity between the analysis candidates constituting the created combination,
Based on the analyzed corpus that is the result of analysis in the past, a corpus value indicating the appearance degree of the created analysis candidate is calculated,
Based on the calculated parallel similarity value and the calculated corpus value, select an analysis candidate to be applied when translating a document part to be translated from the created one or more analysis candidates.
A machine translation program for executing a process.
(Appendix 5)
Determining whether a parallel structure exists between the created analysis candidates,
If it is determined that the parallel structure exists, control to create a combination of the analysis candidates,
The machine translation program according to attachment 4, wherein the machine translation program is executed.
(Appendix 6)
Determining whether the calculated parallel similarity value is equal to or greater than a predetermined threshold;
Control to calculate a corpus value for the analysis candidate determined that the parallel similarity value is equal to or greater than the threshold,
The machine translation program according to attachment 4, wherein the machine translation program is executed.
(Appendix 7)
In a machine translation device that translates a translation target sentence described in a first language into a translation target sentence described in a second language,
Based on a predetermined rule, an input sentence dividing unit that divides the translation target sentence to create a document part;
A plurality of analysis candidate creation units for creating one or more analysis candidates for each of the document parts created by the input sentence dividing unit;
A parallel combination creation unit that creates a combination of analysis candidates for the analysis candidates created for each of the document parts constituting the parallel structure when a parallel structure is recognized between the document parts created by the input sentence dividing unit When,
A parallel similarity value assigning unit that calculates a parallel similarity value indicating the similarity between the analysis candidates constituting the combination created by the parallel combination creation unit;
A corpus value giving unit that calculates a corpus value indicating the appearance degree of the analysis candidate created by the plurality of analysis candidate creation units, based on an analyzed corpus that is a result of analysis in the past;
Based on the parallel similarity value calculated by the parallel similarity value giving unit and the corpus value calculated by the corpus value giving unit, the document to be translated from one or more analysis candidates created by the multiple analysis candidate creation unit A candidate selection unit for selecting analysis candidates to be applied when translating parts;
A machine translation device comprising:
(Appendix 8)
A parallel structure presence / absence determining unit that determines whether or not a parallel structure exists between analysis candidates created by the plurality of analysis candidate creating units;
A first process control unit that controls the parallel combination creation unit to create a combination of analysis candidates when the parallel structure presence / absence determination unit determines that a parallel structure exists;
The machine translation apparatus according to appendix 7, characterized by comprising:
(Appendix 9)
A parallel similarity value threshold determination unit that determines whether or not the parallel similarity value calculated by the parallel similarity value giving unit is equal to or greater than a predetermined threshold;
A second processing control unit that performs control so as to calculate a corpus value for an analysis candidate for which a parallel similarity value is determined to be greater than or equal to the threshold value by the parallel similarity value threshold determination unit;
The machine translation apparatus according to appendix 7, characterized by comprising:

100 Input sentence 110, 120, 130 Document part 120A, 120B, 130A, 130B Analysis candidate 121, 122, 123, 124, 131, 132, 133, 134 Phrase 200 Pattern 210, 220, 230, 240, 250 Translation part 500A, 500B, 500C Machine translation apparatus 501 Input sentence division unit 502 Multiple analysis candidate creation unit 503 Parallel combination creation unit 504 Parallel similarity value assignment unit 505 Frequency acquisition unit 506 Corpus value assignment unit 507 Analyzed corpus database (DB)
508 Total value calculation unit 509 Candidate selection unit 510 Component translation unit 511 Connection unit 1501 Parallel structure presence / absence determination unit 1502 First processing control unit 1601 Parallel similarity value threshold determination unit 1602 Second processing control unit 1701 CPU (Central Processing Unit)
1702 Memory 1703 Input device 1704 Output device 1705 External recording device 1706 Medium drive device 1707 Network connection device 1708 Bus 1709 Portable recording medium 1710 Communication network

Claims (5)

A machine translation method executed by a computer of a machine translation device that translates a translation target sentence described in a first language into a translation target sentence described in a second language,
Based on a predetermined rule, create a document part by dividing the sentence to be translated,
Creating one or more analysis candidates for each of the created document parts;
When a parallel structure is recognized between the created document parts, a combination of analysis candidates is created for the analysis candidates created for each document part constituting the parallel structure,
Calculate a parallel similarity value indicating the similarity between the analysis candidates constituting the created combination,
Based on the analyzed corpus that is the result of analysis in the past, a corpus value indicating the appearance degree of the created analysis candidate is calculated,
Based on the calculated parallel similarity value and the calculated corpus value, select an analysis candidate to be applied when translating a document part to be translated from the created one or more analysis candidates.
A machine translation method characterized by the above. Determining whether a parallel structure exists between the created analysis candidates,
If it is determined that the parallel structure exists, control to create a combination of the analysis candidates,
The machine translation method according to claim 1. Determining whether the calculated parallel similarity value is equal to or greater than a predetermined threshold;
Control to calculate a corpus value for the analysis candidate determined that the parallel similarity value is equal to or greater than the threshold,
The machine translation method according to claim 1. In a computer of a machine translation device that translates a translation target sentence described in a first language into a translation target sentence described in a second language,
Based on a predetermined rule, create a document part by dividing the sentence to be translated,
Creating one or more analysis candidates for each of the created document parts;
When a parallel structure is recognized between the created document parts, a combination of analysis candidates is created for the analysis candidates created for each document part constituting the parallel structure,
Calculate a parallel similarity value indicating the similarity between the analysis candidates constituting the created combination,
Based on the analyzed corpus that is the result of analysis in the past, a corpus value indicating the appearance degree of the created analysis candidate is calculated,
Based on the calculated parallel similarity value and the calculated corpus value, select an analysis candidate to be applied when translating a document part to be translated from the created one or more analysis candidates.
A machine translation program for executing a process. In a machine translation device that translates a translation target sentence described in a first language into a translation target sentence described in a second language,
Based on a predetermined rule, an input sentence dividing unit that divides the translation target sentence to create a document part;
A plurality of analysis candidate creation units for creating one or more analysis candidates for each of the document parts created by the input sentence dividing unit;
A parallel combination creation unit that creates a combination of analysis candidates for the analysis candidates created for each of the document parts constituting the parallel structure when a parallel structure is recognized between the document parts created by the input sentence dividing unit When,
A parallel similarity value assigning unit that calculates a parallel similarity value indicating the similarity between the analysis candidates constituting the combination created by the parallel combination creation unit;
A corpus value giving unit that calculates a corpus value indicating the appearance degree of the analysis candidate created by the plurality of analysis candidate creation units, based on an analyzed corpus that is a result of analysis in the past;
Based on the parallel similarity value calculated by the parallel similarity value giving unit and the corpus value calculated by the corpus value giving unit, the document to be translated from one or more analysis candidates created by the multiple analysis candidate creation unit A candidate selection unit for selecting analysis candidates to be applied when translating parts;
A machine translation device comprising:

Images