JP3778705B2 - Bilingual document matching system - Google Patents

Description

[0001]
BACKGROUND OF THE INVENTION
The present invention relates to a bilingual document association system for associating sentences in a bilingual document composed of an original sentence such as English-Japanese and a translated sentence.
[0002]
[Prior art]
There is a system that uses a bilingual dictionary to associate each sentence in a bilingual document composed of original and translated sentences such as English-Japanese. For example, as literature on such systems:
Bilingual text matching using bilingual dictionaries and statistical information Takehito Utsuro Yuji Matsumoto Vol. 12 No. 5 sep. There was 1995.
[0003]
In this document, a method is described in which correspondence between sentences in a bilingual document is performed by dynamic programming (DP) using a bilingual dictionary. In order to perform such association using dynamic programming, first, the original sentence and the translated sentence are divided into sentences, the form of the sentence is analyzed, and divided into words. Then, independent words are taken out from these words and evaluated by how much the independent words in each sentence correspond using a bilingual dictionary. For example, the following formula is used as an evaluation method.
[0004]
h (x, y) = 2 × fm (x, y) / (fj (x) + fj (y))
here,
x is a sentence in the original sentence (may be multiple sentences)
y is a sentence in the translation (may be multiple sentences)
h (x, y) is the evaluation function
fm (x, y) is the number of independent words that correspond in x and y
fj (x) is the number of independent words in x
fj (y) is the number of independent words in y
It is.
[0005]
By using such an evaluation function, the value of the evaluation function between corresponding sentences becomes large, and the evaluation function between sentences that do not correspond becomes small. This is examined from the beginning of the sentence, and the combination that maximizes the sum of the evaluation functions is taken as the solution of the association problem.
[0006]
[Problems to be solved by the invention]
However, as described above, in order to perform the association using the evaluation function, it takes time to calculate the evaluation value because the bilingual dictionary is drawn or the character string matching process is performed. Furthermore, the method using DP takes time proportional to the number of sentences in the original sentence × the number of sentences in the translated sentence, and there is a problem in the execution speed. Even though most of the sentence correspondence is one-to-one sentence correspondence, including one sentence to two sentences and two sentences to one sentence, it is almost 100% of the probability. This is because the evaluation values of all possible combinations are calculated in consideration of the possibility of becoming three sentences versus one sentence, one sentence versus four sentences, and four sentences versus one sentence.
[0007]
In order to cope with such a problem, a method of not calculating one sentence to three sentences, three sentences to one sentence, etc. can be considered easily. However, the probability values as described above may change depending on the field of the bilingual document, and in some cases, the correspondence between one sentence and three sentences may be the optimal solution. However, such a method has a problem that, when there is a correspondence between one sentence and three sentences, it cannot be found, and in some cases, the precision of association is lowered.
[0008]
From this point of view, it has been desired to realize a bilingual document association system that can speed up the processing without reducing the accuracy of association.
[0009]
[Means for Solving the Problems]
The present invention employs the following configuration in order to solve the above-described problems.
<Constitution>
The present invention is a bilingual document sentence correspondence system for associating a plurality of source sentence sentences included in a source language document with a plurality of target sentence sentences included in a translation document of the source language document. Compare the number of translations, When the number of source language sentences is equal to the number of target language sentences, each initial path is obtained by sequentially setting each corresponding source sentence and each corresponding target sentence in a one-to-one correspondence according to the sentence sequence, When the number of source texts is larger than the number of target sentences, each initial path is set by sequentially setting each corresponding pair corresponding to either one source sentence or each target word sentence according to the sentence sequence. If the number of source language sentences is smaller than the number of translated word sentences, each corresponding set corresponding to either one-to-one or one-to-two according to the sentence sequence is set in order. Each time the initial path is obtained and the number of correspondences between the source sentence and the translated sentence is sequentially increased, combination management means for sequentially obtaining other paths composed of other corresponding pairs, and for each initial path and each other path, Each evaluation value indicating the degree of correspondence between the original sentence and the translated sentence is assigned to the corresponding group. Calculate based on the evaluation elements of the sentence, and sequentially add the calculated evaluation values to obtain the total evaluation value for each path. When calculating the evaluation value calculation means and any of the initial paths Per pass Evaluation value Total The maximum value of the sum is stored as a provisional solution in association with the corresponding initial path, the evaluation value setting evaluation value is held, the evaluation value already calculated for the other paths is taken, and the intermediate sum is calculated. Before the evaluation value of the uncalculated corresponding pair is calculated, the number of the uncalculated corresponding pair is examined, the number and the set evaluation value are integrated, and the integrated value is added to the intermediate sum. If it is determined that the addition result is smaller than the provisional solution, the calculation of the other path is stopped and the evaluation value of the other path is Total When it is determined that the sum is larger than the provisional solution, the provisional solution is updated. Further, when the number of correspondences between the original sentence and the corresponding translation sentence in the corresponding pair reaches the maximum number of correspondence sentences in the other paths, the combination management means corresponds to the maximum correspondence. A calculation result management means that terminates the generation of other paths related to the correspondence pairs that are equal to or greater than the number of attached sentences and uses the path corresponding to the current provisional solution as the solution for the correspondence of the sentences;
It is characterized by providing.
[0010]
This set evaluation value is the maximum evaluation value that the combination can take and Evaluation criteria calculated from the mean and variance of the combination One of them.
[0012]
DETAILED DESCRIPTION OF THE INVENTION
The present invention preferentially calculates from the most likely ones, such as one sentence versus one sentence, one sentence versus two sentences, etc., and uses the provisional solution to narrow down the possibility of correct answers. The amount of calculation is reduced.
[0013]
Hereinafter, embodiments of the present invention will be described in detail using specific examples.
[0014]
<< Specific Example 1 >>
<Constitution>
FIG. 1 is a block diagram showing a specific example 1 of the bilingual document association system of the present invention.
The system shown in the figure includes an original text file 101, a translation text file 102, a text segmentation means 103, a morpheme analysis means 104, a combination management means 105, an evaluation value calculation means 106, a calculation result management means 107, an original text file 108 with a corresponding tag, and a corresponding tag. It consists of a translation file 109 and a bilingual dictionary 110.
[0015]
The original text file 101 is a document file composed of a plurality of original texts, for example, an English text file. The translation file 102 is a document file composed of a plurality of translations, for example, a Japanese document file.
[0016]
The sentence dividing unit 103 has a function of dividing the document file of the original sentence file 101 and the translated sentence file 102 into sentences. For example, it is configured to divide by a period if it is English, or by a punctuation if it is Japanese.
[0017]
The morpheme analyzing unit 104 has a function of performing morphological analysis on the English sentence or the Japanese sentence divided by the sentence dividing unit 103 and dividing the word into words. The sentence dividing unit 103 and the morpheme analyzing unit 104 can use known configurations.
[0018]
The combination management unit 105 obtains a combination of the corresponding pair of the original sentence and the translated sentence, and based on the results of the sentence dividing unit 103 and the morpheme analyzing unit 104, from the number of sentences in the original sentence file 101 and the translated sentence file 102 of the original sentence and the translated sentence, It has a function for obtaining a combination that has a high possibility of an optimal solution for matching an original sentence and a translated sentence.
[0019]
The evaluation value calculation means 106 obtains an evaluation value that becomes higher as the original sentence corresponds to the translated sentence, and calculates the sum of the evaluation values of the combinations that are likely to be obtained by the combination management means 105 as a provisional solution. It has the required function. Specifically, the evaluation value may be calculated based on the evaluation function formula described in the section of the related art, or the evaluation value may be calculated by other means.
[0020]
The calculation result management means 107 Evaluation value calculation means 106 The table (not shown) for storing the DP evaluation values calculated in step 1 is calculated, and the evaluation value of the combination (path) of the original sentence and the translation sentence obtained by the combination management means 105 is calculated sequentially from the first sentence of the combination The evaluation value calculation means 106 is instructed to do so, and the obtained evaluation value is stored in the table. Then, based on the evaluation values in these tables, if it can be determined that the value will not exceed the provisional solution even if it is calculated to the end in any combination, it has a function to stop the evaluation value calculation at that time ing.
[0021]
The corresponding tagged original text file 108 and the corresponding tagged translated text file 109 are obtained by adding tags for indicating correspondence between sentences to the original text file 101 and the translated text file 102, respectively.
[0022]
The bilingual dictionary 110 is a dictionary in which there are a plurality of translated words when an original word for association is drawn. For example, when the original text is English and the translated text is Japanese, it corresponds to an English-Japanese dictionary.
[0023]
The above-described bilingual document mapping system is composed of a microcomputer or the like, and the original text file 101, the translated text file 102, the corresponding tagged text file 108, the corresponding tagged text file 109, and the bilingual dictionary 110 are an external storage device such as a hard disk device. Alternatively, it is provided in a semiconductor memory. The sentence dividing unit 103 to the calculation result managing unit 107 are configured by a program corresponding to each unit, a processor for executing the program, a main storage device, and the like.
[0024]
<Operation>
FIG. 2 is a flowchart showing the operation of this example.
First, the sentence dividing means 103 separates the original text file 101 and the translated text file 102 (step S100). Here, m is the number of sentences in the original sentence, and n is the number of sentences in the translated sentence. Further, in the combination management means 105, 2 is set to a variable i for gradually increasing the corresponding group of sentences.
[0025]
Next, it is checked whether the number of sentences in the original sentence is equal to the number of sentences in the translated sentence, that is, m = n (step S101). Here, if they are equal, the combination of sentences for calculating the evaluation value is only (1, 1) (step S103), and the process proceeds to step S106. On the other hand, if they are not equal in step S101, it is determined whether m> n is satisfied (step S102).
[0026]
In step S102 m <n If so, the corresponding pairs to be calculated are (1, 1) and (2, 1) (step S104), the process proceeds to step S106, and if m> n, the corresponding pairs to be calculated are (1, 1), ( 1 and 2) (step S105), the process proceeds to step S106.
[0027]
In step S <b> 106, the evaluation value calculation unit 106 calculates the evaluation function of the corresponding group and sends the calculation result to the calculation result management unit 107. Thereby, the calculation result management means 107 sets the path with the highest score as a provisional solution (step S107). The path uniquely indicates what route (corresponding pair of the original sentence and the translated sentence) is associated through the first sentence to the last sentence.
[0028]
Next, the calculation result management means 107 sets the corresponding pairs to be calculated as (1, 1), (1, 2), (2, 1) (step S108), and checks whether there is an uncalculated branch (step S109). . A branch is an element of a path. From a certain point in the path (5th sentence in English, 4th sentence in Japanese, etc.), what kind of correspondence the next original and translated text is It is shown.
[0029]
If there is an uncalculated branch in step S109, the evaluation value of that branch is calculated, and this calculated value is stored in the DP table in the calculation result management means 107 (step S110). Next, it is determined by this calculated value whether a solution having a higher evaluation value than the provisional solution described above is obtained (step S111). If a high solution is obtained, the solution is updated as a provisional solution. (Step S112).
[0030]
On the other hand, if it is unclear in step S111 whether a solution higher than the provisional solution can be obtained, it is determined whether the path cannot be an optimal solution (step S113). In step S113, if it is clearly determined that the evaluation value does not become larger than the provisional solution, the node is marked, calculation of the branch ahead is suspended (step S114), and the process returns to step S109. That is, the calculation as the path is stopped. If it is unclear in step S113 that the optimal solution cannot be obtained, the process directly returns to step S109.
[0031]
By repeating the processing of step S109 to step S114, a provisional solution of the paths in the corresponding pairs (1, 1), (1, 2), (2, 1) is obtained, and the provisional solution is obtained during the processing. Paths found not to be greater than the solution are deferred from further branch computations. In step S114, the calculation is not suspended but suspended because it may be calculated before the node when the value of i is incremented in step S115 and later.
[0032]
If there are no more uncalculated branches in step S109, the process proceeds to step S115, and the value of the variable i for gradually increasing the corresponding pairs of sentences is incremented. That is, the value of i is set to 3.
[0033]
Next, it is checked whether the value of i is larger than a predetermined maximum number of correspondence sentences (step S116). The maximum number of correspondence sentences indicates how many sentences correspond to one sentence. If the maximum number of correspondence sentences is 4, the correspondence of one sentence to four sentences and four sentences to one sentence is examined. Means that. In this step S116, if i ≦ maximum number of correspondence sentences, the calculation result management means 107 adds (1, i), (i, 1) to the correspondence pair to be calculated (step S117), and step S109. Return to.
[0034]
On the other hand, if i> the maximum number of associated sentences in step S116, the current provisional solution is set as the optimum solution, and the path of the optimum solution is set as the sentence association solution (step S118), and the association process is terminated.
[0035]
The above processing will be described using a more specific example.
[0036]
FIG. 3 is an explanatory diagram of paths for taking correspondence between files composed of nine English sentences and seven Japanese sentences.
[0037]
In the figure, E1 to E9 represent the first to ninth sentences in English, respectively, and J1 to J7 represent the first to seventh sentences in Japanese, respectively. The numbers with circles represent the order in which the branch evaluation values are calculated. For example, the first branch evaluates the correspondence between the first sentence in English and the first sentence in Japanese. The second branch evaluates the correspondence between the first sentence of the English sentence to the second sentence and the first sentence of the Japanese sentence, ... 27th branch evaluates the correspondence of the ninth sentence of the English sentence and the seventh sentence of the Japanese sentence It means to do.
[0038]
First, the processing of step S100 to step S107 in the flowchart of FIG. 2 will be described.
[0039]
In this example, since there are more English sentences, the combination of (1, 1) and (2, 1) is calculated first. If the corresponding combinations are limited to these two ways, the path that can reach the goal (upper right in the figure) is a narrow range as shown in FIG.
[0040]
The calculation is performed in the order of the numbers given to the branches shown in the figure (it is not always necessary to calculate in this way). When calculation is completed up to 27th, the path with the highest evaluation value sum is checked by DP, and the sum of the evaluation values is calculated. Is a provisional solution. Here, it is assumed that 2-5-9-14-19-23-26 is a provisional solution.
[0041]
Next, the process of step S108-step S114 is demonstrated.
If the combination of (1, 2) is added in step S108, the path in FIG. 3 becomes as follows.
[0042]
FIG. 4 is an explanatory diagram of paths when the combination of (1, 2) is added.
For example, when the path No. 31-12-19-23-26 obtained a higher evaluation value than the provisional solution at the time of calculating No. 31 in the figure, the provisional solution is updated at that time, The path is stored (steps S111 to S112).
[0043]
Next, when calculating the number 36, if it is found that a better solution than the provisional solution cannot be obtained through any path (this method will be described later), the end point of the branch number 36 (goal side) And the evaluation of the branch ahead is suspended (steps S113 to S114). In the example of FIG. 4, it is not necessary to calculate the branches 40, 41, 44, and 45.
[0044]
The determination process in step S113 is performed as follows. If the maximum evaluation value of each branch is 1, the provisional solution is 4.8, and the sum of the branches of 2-6-36 is 1.2, the solution passing through the path cannot be 4.2 or more. That is, since it does not become larger than the provisional solution, there is no point in calculating the evaluation value of the path. Therefore, the calculation as the path is stopped.
[0045]
The processing from step S115 to step S118 is for calculating correspondences up to the maximum number of correspondence sentences. At that time, as in the case where the above corresponding pairs are (1, 1), (1, 2), (2, 1), the calculation of branches that cannot be the optimal solution by the calculation of one sentence versus three sentences. You do n’t have to.
[0046]
<effect>
As described above, according to the first specific example, a solution having a high possibility of correspondence between the original sentence and the translated sentence is set as a provisional solution, and when the evaluation value calculation process is found not to be higher than the provisional solution, Since the calculation as the path is stopped, the calculation amount in the DP can be reduced, and the processing time can be shortened. In addition, since calculation such as one sentence versus three sentences is also taken into account, the accuracy of the document contents or the like does not deteriorate.
[0047]
Furthermore, when a value with a higher sum of evaluation values than the previous provisional solution is found in any combination, the found value is used as a new provisional solution. Therefore, it is possible to obtain the effect of speeding up the processing and improving the accuracy at the same time.
[0048]
In particular, in this specific example, when the path of the optimal solution matches or is close to the path of the provisional solution that was first obtained, the number of evaluation value calculations that can be omitted compared to the prior art increases and the effect is increased. Become prominent.
[0049]
<< Specific Example 2 >>
Specific example 2 sets whether a value that includes evaluation values of a plurality of branches (corresponding pairs of original and translated sentences) at a certain ratio is set as a reference value for each branch, and whether or not the path is an optimal solution. Is determined using this reference value.
[0050]
<Constitution>
FIG. 5 is a configuration diagram of the bilingual document association system of the second specific example.
The illustrated system includes an original text file 101, a translation text file 102, a sentence division means 103, a morpheme analysis means 104, a combination management means 105, an evaluation value calculation means 106, a calculation result management means 107, an original text file 108 with a corresponding tag, and a corresponding tag. The translation file 109, the bilingual dictionary 110, and the reference value calculation means 111 are included. Here, since the original text file 101 to the bilingual dictionary 110 are the same as the configuration of the specific example 1, the description thereof is omitted.
[0051]
The reference value calculation means 111 calculates the average, variance, etc. of each branch, and a certain value such that the evaluation value of a certain percentage of branches does not exceed that value, for example, 95% of the branches are within this reference value. It has a function to derive a certain value. The formula shown with a broken line in FIG. 5 shows an example of how to obtain the reference value. For example, when the branch value has a normal distribution, the average is μ, and the variance is D, the reference value is x such that p = 0.95 in the equation shown.
[0052]
The calculation result management unit 107 is configured to determine a path that cannot be higher than the provisional solution using the reference value obtained by the reference value calculation unit 111. In other words, the evaluation value of the path is calculated assuming that the branch for which the evaluation value is not calculated is at most the reference value.
[0053]
<Operation>
FIG. 6 is a flowchart illustrating the operation of the second specific example.
2 differs from the specific example 1 shown in FIG. 2 in that the reference value calculation means 111 obtains the reference value in step S208 and the calculation result management means 107 is likely to be the optimal solution in step S214. When determining whether or not, the reference value obtained in step S208 is used.
[0054]
Steps S200 to S207 in Specific Example 2 are the same as the processes in Steps S100 to S107 of Specific Example 1. Next, in step S208, an evaluation reference value is calculated from the average and variance of branches based on, for example, the formula in FIG. The subsequent steps S209 to S213 are the same as the steps S108 to S112 of the first specific example.
[0055]
In step S214, when it is determined whether or not there is a possibility of an optimal solution as described above, an evaluation value of a branch that has not yet been calculated is calculated as a reference value at most, and a determination is made based on this. That is. That is, in the specific example 1, the evaluation value of the path is calculated by setting the evaluation value of the branch not yet calculated as 1, that is, the evaluation value as the highest value. On the other hand, in the specific example 2, the evaluation value of the path is calculated on the assumption that most of the branches are within the reference value, for example, 0.7.
[0056]
By using such a reference value, it is not always possible to guarantee that an optimal solution is found, but the number of branches for calculating evaluation values can be further reduced as compared with the first specific example.
[0057]
The reason that it is not always possible to guarantee that the optimum solution is found is as follows. That is, since the reference value is a value that is not the highest value such as 0.7, when the evaluation value of the branch is actually calculated rather than when the evaluation value of the branch that has not been calculated yet is calculated with the reference value. There is a possibility that the evaluation value of the path becomes larger. However, it is considered that such a possibility is hardly possible in actual processing by setting an appropriate reference value.
[0058]
Moreover, since the optimal solution is not necessarily 100% correct for the sentence matching problem, considering that it is necessary to manually check whether the response is correct later, it takes time to determine the optimal solution. It may be more useful to find a reasonable solution in a short time than to find it.
[0059]
Since the processing of subsequent steps S215 to S219 in the specific example 2 is the same as the steps S114 to S118 in the specific example 1, the description thereof is omitted here.
[0060]
The above process will be described using a more specific example. Here, also in the specific example 2, it is assumed that the target file is composed of 9 English sentences and 7 Japanese sentences. In addition, as an explanatory diagram of a path for taking correspondence between these sentences, a description will be given with reference to FIGS.
[0061]
First, the processing of step S200 to step S207 in the flowchart of FIG. 6 will be described.
[0062]
In this example, since there are more English sentences, the combination of (1, 1) and (2, 1) is calculated first. If the corresponding combinations are limited to these two ways, the path that can reach the goal (upper right in the figure) is a narrow range as shown in FIG.
[0063]
The calculation is performed in the order of the numbers given to the branches shown in the figure (it is not always necessary to calculate in this way). When calculation is completed up to 27th, the path with the highest evaluation value sum is checked by DP, and the sum of the evaluation values is calculated. Is a provisional solution. Here, it is assumed that 2-5-9-14-19-23-26 is a provisional solution. The above is the same as the specific example 1.
[0064]
In step S208, a branch reference value is calculated. Here, it is assumed that the reference value becomes 0.7.
[0065]
Next, the process of steps S209 to S215 will be described.
If the combination of (1, 2) is added in step S209, the path of FIG. 3 becomes as shown in FIG.
[0066]
For example, when the path No. 31-12-19-23-26 obtained a higher evaluation value than the provisional solution at the time of calculating No. 31 in the figure, the provisional solution is updated at that time, The path is stored (steps S212 to S213).
[0067]
Next, when calculating the number 36, if it is found that a better solution than the provisional solution cannot be obtained through any path (this method will be described later), the end point of the branch number 36 (goal side) And hold the evaluation of the branch from there. In the example of FIG. 4, it is not necessary to calculate the branches 40, 41, 44, and 45.
[0068]
The above determination method is performed as follows. If the reference value of the evaluation value of each branch is 0.7, the provisional solution is 4.8, and the sum of the branches of 2-6-36 is 2.2, the solution passing through the path must be 4.3 or higher Don't be. That is, since it does not become larger than the provisional solution, there is no point in calculating the evaluation value of the path. Therefore, the calculation as the path is stopped.
[0069]
Here, a comparison with Example 1 is as follows. That is, in the specific example 1, since 1 which is the maximum evaluation value of each branch is used, if the provisional solution is 4.8 and the sum of the branches of 2-6-36 is 2.2, The solution passing through the path is 5.2 at the maximum, and it cannot be determined that it cannot be the optimal solution at this point. On the other hand, in the specific example 2, since the calculation can be stopped at this point, the number of branches for calculating the evaluation value can be reduced as compared with the specific example 1.
[0070]
The processing from step S216 to step S219 is for calculating correspondences up to the maximum number of correspondence sentences. At that time, as in the case where the above corresponding pairs are (1, 1), (1, 2), (2, 1), the calculation of branches that cannot be the optimal solution by the calculation of one sentence versus three sentences. You do n’t have to.
[0071]
<effect>
As described above, according to the specific example 2, when a solution having a high possibility of correspondence between the original sentence and the translated sentence is set as a provisional solution, and it is found that the evaluation value calculation process does not become a value higher than the provisional solution, Since the calculation as the path is stopped, the calculation amount in the DP can be reduced, the processing time can be shortened, and the accuracy is not lowered as in the first specific example.
[0072]
Furthermore, in the specific example 2, a reference value is prepared in advance, and it is determined whether the path cannot be an optimal solution using this reference value. Therefore, the processing time can be shortened.
[0073]
In the specific examples 1 and 2, the case of associating English sentences with Japanese sentences has been shown. However, by changing the bilingual dictionary, it can also be used for associating all languages. Further, the order in which the branch evaluation values are calculated is not limited to the order shown in the figure, and the order may be changed to some extent.
[0074]
Further, in specific examples 1 and 2, even when m> 2n, it can be easily dealt with by changing the correspondence set calculated first. In such a case, one sentence to one sentence, two sentences to one sentence, and three sentences to one sentence are calculated first.
[0075]
In the specific example 2, the reference value is determined in advance by the average or variance of the evaluation values of the branches, but the value of the branch until the provisional solution is first found (the branch value shown in FIG. 3) It may be calculated by the value of the processing up to step S206 in FIG. In this way, the calculation for the reference value can be performed simultaneously with the association processing.
[Brief description of the drawings]
FIG. 1 is a configuration diagram showing a specific example 1 of a bilingual document association system of the present invention;
FIG. 2 is a flowchart showing the operation of specific example 1 of the bilingual document association system of the present invention.
FIG. 3 is an explanatory diagram of paths for taking correspondence between files composed of nine English sentences and seven Japanese sentences in the bilingual document association system of the present invention;
FIG. 4 is an explanatory diagram of paths when a combination of (1, 2) is added in the bilingual document association system of the present invention.
FIG. 5 is a block diagram showing a specific example 2 of the bilingual document association system of the present invention;
FIG. 6 is a flowchart showing the operation of specific example 2 of the bilingual document association system of the present invention.
[Explanation of symbols]
101 Original file
102 Translated file
105 Combination management means
106 Evaluation value calculation means
107 Calculation result management means
111 Reference value calculation means

Claims (2)

A bilingual document sentence association system for associating a plurality of original sentence sentences included in a source language document with a plurality of translated sentence sentences included in a translated word document of the original document,
When the number of the source language sentences and the number of the translated texts are compared, and the number of the source language sentences is equal to the number of the translated texts, the original text sentences and the translated texts are made to correspond one-to-one according to the sentence sequence. Each initial path formed by sequentially setting each corresponding pair is obtained, and when the number of the source language sentences is larger than the number of the translation words, each of the source language sentences and each of the translation words are one-to-one and two-to-one according to the sentence sequence. Each corresponding path corresponding to any one of the above is sequentially set to obtain each initial path, and when the number of the source sentence is smaller than the number of the translated sentences, the original sentence and each translated sentence are arranged in a sentence Each of the correspondence paths corresponding to either one-to-one or one-to-two according to the above is obtained in turn, and each time the number of correspondences between the source sentence and the translated sentence is successively increased, Unions that sequentially seek other paths consisting of pairs And management means,
For each of the initial path and the other path, corresponding to each corresponding set, each evaluation value indicating the correspondence between the original sentence and the translated sentence is calculated based on the evaluation element of the sentence, and each of the calculated Evaluation value calculating means for sequentially adding evaluation values to obtain a total sum of evaluation values for each path ;
Each said hold in correspondence with the initial path to the appropriate maximum value of the total sum of the evaluation value for each path when determined either initial path as an interim solution, and, setting the evaluation value of the evaluation value , The calculated evaluation value for the other path is taken and the intermediate sum is calculated, and before the evaluation value of the uncalculated corresponding group is calculated, the uncalculated corresponding group The number and the set evaluation value are added up, the added value is added to the intermediate sum, and if the addition result is determined to be smaller than the provisional solution, the calculation of the other path is stopped, When the total sum of the evaluation values of the other paths it is determined that the tentative solution is greater than, updates the interim solution, further, the corresponding number of the maximum with the corresponding sets of source language sentences and translated word sentences in the other path When the number of correspondence sentences is reached, the combination management means will not exceed the maximum number of correspondence sentences. The corresponding set together to stop the production of other paths related, parallel texts statement associating system, characterized in that it comprises a calculation result management unit to correspondence solutions sentence corresponding path to the current interim solutions. 2. The bilingual document association system according to claim 1, wherein the set evaluation value is one of a maximum evaluation value that can be taken by the combination and an evaluation reference value calculated from an average and a variance of the combination .

Images