JP2003308319A - Device for selecting translation, translator, program for selecting translation, and translation program - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To precisely select a translation, and to precisely carry out mechanical translation, in an idiomatic expression in which selection of a proper translation or the translation is difficult in the prior art, without collecting a large volume of parallel translation example data. <P>SOLUTION: A method for outputting the translation corresponding to a translation-objective word in an input text, based on the similarity of a character string between the input text and the parallel translation example data, or/and a method for generating a learning model based on the parallel translation example data, and for outputting an translation option corresponding to one providing the maximum likelihood after the learning model of the highest precision for a learning data is applied to an input composition is/are used independently or combinedly. <P>COPYRIGHT: (C)2004,JPO

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a translation word selection device, a translation device, and programs for translating a text input in a language into another language.

[0002]

2. Description of the Related Art In machine translation, parallel translation data is stored that pairs original texts such as sentences, phrases, sections, or words described in a certain language with translated texts obtained by translating the original texts into another language. Sometimes a database is used. Recently, in particular, not only words, but also databases of examples of sentences and phrases including words (hereinafter referred to as “translation corpus”) have been used. Currently,
A wide variety of bilingual corpora using newspapers, dictionaries, etc. as language resources have been made available on the Internet and available for use.

In machine translation, the translation word selection is considered as one of the important technical elements. However, when the parallel translation example corpus is used, the number and types of examples increase simply as the amount of parallel translation data increases. It is conceivable that instead of using only a single bilingual corpus, collecting as many kinds of bilingual corpus or bilingual data as possible and performing machine translation using them. In this case, the bilingual corpus collected based on the input text in the source language to be translated is referred to, and the bilingual text corresponding to the source text that matches or is most similar to the input text is output as the translation result. The example-based translation word selection method is considered to be the simplest method. In addition to this, a learning-based translation word selection method in which learning data created based on a bilingual corpus is applied to a learning model and a simple statistically probable translation word is output is also considered.

[0004]

Since the example-based translation word selection method refers to a wide variety of parallel translation corpus, it can be presumed that the translation accuracy is improved accordingly. However, regarding a source language word having polysemy, in which there are many types of corresponding translation words, the above method outputs a correct translation word if there is no example in the bilingual corpus that is the same as or similar to the input text. However, there is a problem that it is not flexible. On the other hand, in the learning-based translation word selection method, in order to preferentially output translation words used in statistically high frequency examples, While the accuracy is improved, the accuracy of the translation is decreased for the translated words with low occurrence frequency.

Such problems often occur when the input text contains "idioms" in which a word is combined with other phrases to give a unique expression. As an example, consider the case where the original text contains the ambiguous word “buy” in Japanese. The meaning of “buy” differs between the expressions “buy a book” and “buy a feeling of disgust”. , The English translation corresponding to "buy" differs accordingly. In this case, in the Japanese-English parallel corpus, there are many examples in which the same translated word (buy) as the translated word "buy" in the case of "buying things" is used, and it is considered that the frequency of occurrence of the translated word is high. On the other hand, in an idiomatic expression such as "buy an antipathy", the English translation of "buy" (an
It is considered that there are few examples of "buy" including the English translation word because the tipathy) is special.

Further, in any of the translation word selection methods,
In order to realize highly accurate translation, it is necessary to collect a large number of bilingual corpora, but since there are various variations in natural language, the method of simply collecting a large number of bilingual corpora increases the computational load. However, in reality, it is considered impossible to implement accurate machine translation in a short time by such a method.

In view of the above-mentioned problems, the present invention makes it possible to accurately and properly perform translation word selection and translation in machine translation without imposing an excessive load on the device. The main purpose is that.

[0008]

SUMMARY OF THE INVENTION The present invention basically comprises an original language example consisting of text in a first language, a word contained therein, a translation of the word in a second language and information about the translation. It is input in the first language using the bilingual translation example data storage unit that stores bilingual translation example data that is a pair of language example data and a target language example composed of text translated from the source language example in the second language. The translated word described in the second language corresponding to the translation target word that is the word to be translated included in the input text is selected. The parallel translation example data storage unit used here corresponds to the so-called parallel translation corpus described above, but the number is not limited as long as one or more is used. However, by using a plurality of parallel translation example data storage units, it is possible to increase the number of examples and improve the accuracy of translation word selection. In addition, the parallel translation example data storage unit can be provided as a component of a translation word selection device or a translation device described below, or can be provided in another device that can communicate with the translation word selection device or the like.

According to the present invention, as shown in the schematic configuration diagram of FIG. 1, the input acceptance unit 1 for accepting an input of an input text and its acceptance are accepted as the basic configuration of the first translation word selection device A1. An example extracting unit 2 for extracting at least one or more source language example data containing a word corresponding to the translation target word in the input text from the parallel translation example data storage unit C, and the extracted source language example data and the input text. A similarity detection unit 3 that detects the similarity between the input text and the source language example based on a similarity evaluation of the detected source language example and the similarity that outputs at least the source language example data that has the highest similarity. An evaluation unit 4 and a translation output unit 5 that outputs a translation corresponding to the translation target word in the target language example included in the parallel translation example data corresponding to the output source language example data. And it is characterized in that it is.

With this configuration, it is possible to output a translation word having the highest similarity to the source language example in which the translation target word included in the input text is used. Therefore, when selecting a translated word of an infrequently appearing phrase such as an idiom used in the source language, an appropriate translated word is used without using too much parallel translation example data and without imposing a heavy load on computer processing. It becomes possible to make a selection.

In particular, as a mode in which the similarity detection unit 3 can perform suitable similarity detection, the input text and the source language example included in the extracted source language example data are compared character by character to obtain them. Similarity calculated using at least one of the ratio of the matching character string between the input text and the source language example, or the number of divisions indicating how many parts the matching part is divided and matched based on the difference An example is one in which degree is calculated as similarity.

In order to facilitate the subsequent processing of the source language examples extracted by the example extracting unit 2, the example extracting unit 2 adds sentence endings to the source language examples included in the extracted source language example data. It suffices to perform processing and output the processed source language example. In this case, the similarity detection unit 3 calculates the difference between the input text and the processed source language example in character units. Based on this, at least one of the ratio of the matched character string to the character string of the processed source language example or the number of divisions indicating how many parts the matched portion is divided and matched is calculated as the similarity. It is desirable to configure it as follows.

Further, in the translated word output unit 5, it is assumed that there is a plurality of source language example data having a maximum degree of similarity as a result of being calculated by the similarity detection unit 3 and output by the similarity evaluation unit 4. It In this case, as a result of the operation, it is presumed to be most suitable by outputting a translation corresponding to the translation target word in the parallel translation example data including the character string matching the input text or the source language example having the largest number of divisions. The translated word can be output.

Further, in order to simplify the processing after receiving the input text, it is preferable that the input receiving unit 1 automatically extracts the translation target word by morphological analysis of the input text. The “morphological analysis” refers to an analysis process such as dividing an input text into words and assigning a part of speech to each word, and a predetermined analysis algorithm and analysis dictionary data are used.

Further, when the parallel translation example data includes a source language headword generated based on a word included in the source language example, the example extraction unit 2 at least applies the source language headword to the translation target word. By extracting the source language example data including "" from the parallel translation example data storage unit C, the extraction processing of the source language example data from the parallel translation example data storage unit C can be speeded up.

Furthermore, when the parallel translation example data has a source language headword generated based on a word included in the source language example and a target language headword generated based on a corresponding translated word, The example extracting unit 2 extracts at least the source language example data including the source language headword corresponding to the translation target word, and the translated word output unit 5 includes the source language example data output by the similarity evaluation unit 4 and the example. By outputting the target language headword corresponding to the source language headword extracted by the extraction unit 2, the processing up to the translated word output can be further speeded up.

Further, according to the present invention, as shown in a schematic configuration diagram of FIG. 2, as a basic configuration of the second translation word selection device A2, an input acceptance section 11 for accepting an input of an input text and a parallel translation example data storage section are stored. Learning corresponding to the translation target word in the input text accepted by the input accepting unit by using the learning data created based on the words included in the source language example and the parallel translation example data corresponding to the source language example. A learning model generation unit 12 that generates a model, and applies the generated learning model to a translation target word in an input text, calculates a certainty factor for all translation word candidates of the translation target word, and orders them in order of certainty factor. A learning model application unit 13 that outputs a candidate and a translation word candidate that has the highest certainty factor among the output translation word candidates are selected and output as a translation word corresponding to the translation target word. It is characterized by having a word output section 14. Here, the “learning data” means the words input in the first language created based on the bilingual translation example, the corresponding correct translated words to be output in the second language, and the attributes associated with them. Refers to information such as features. The “learning model” is a functional model that includes parameters estimated using the learning data and is generated by a machine learning method. Further, the ordering of the certainty factors does not matter whether it is a descending order or an ascending order.

According to such a configuration, when a certain amount of learning data is created or collected, an appropriate learning model generated based on the learning data is applied to the target language to be translated, and then the confidence factor is calculated. The highest translation word candidate, that is, the translation word that can be inferred to be the most appropriate, can be output. Therefore, in such a translation word selection device A2, when selecting a translation word, a learning model is generated for each word (word) to be translated, and the translation word is selected by an appropriate model according to each word (word). You will be able to.

In particular, in the learning model generation unit 12, for each translation target word in the input text accepted by the input acceptance unit 11, the bilingual translation example data corresponding to the source language example including it is extracted from the bilingual translation example data storage unit C, If the learning model is configured to be generated based on the extracted parallel translation example data, it is possible to quickly and accurately perform the translated word output process.

In order to increase the accuracy of the translated word to be output, the learning model generating unit 12 uses the learning data to generate a learning model for each learning data, and the input receiving unit 11 receives the learning model. A learning model having the highest accuracy in the learning data is selected for each translation target word in the input text, and the learning model applying unit 13 selects the learning model having the highest accuracy selected in the learning model generating unit 12 in the input text. Should be applied to the translation target word of. The number of learning data to be used may be one or plural.

Also, in this translated word selection device A2,
By automatically extracting the translation target words from the input text by morphological analysis in the input reception unit 11, it is possible to simplify the processing after the reception of the input text. Similarly, when the bilingual example data includes a source language headword generated based on a word included in the source language example, the learning model generation unit 12 causes the learning model generation unit 12 to at least correspond to the source language headword. By extracting the source language example data including "" from the parallel translation example data storage unit C, the extraction processing of the source language example data from the parallel translation example data storage unit C can be speeded up.

The translation word selection device of the present invention also has the above-mentioned 2
It is also possible to dramatically improve the accuracy of the translated word as a mode in which the translated word selection devices A1 and A2 of different types are combined. That is, according to the present invention, as shown in the schematic configuration diagram of FIG. 3, as a basic configuration of the third translated word selection device A3, an input acceptance unit 31 that accepts an input of an input text and an input text accepted by the input acceptance unit 1 are provided. An example extraction unit 32 that extracts at least one or more source language example data including a word corresponding to the translation target word from the parallel translation example data storage unit C, and an example of the source language extracted by the input text and the example extraction unit. A similarity detection unit 33 that detects the similarity between the input text and the source language example based on the data, and the similarity detection unit 3
Included in the source language example stored in the parallel translation example data storage unit 34 and the similarity evaluation unit 34 that compares and evaluates the similarity of the source language example detected in step S3 and outputs at least the source language example data having the highest similarity. Learning model generation for generating a learning model corresponding to the translation target word in the input text accepted by the input accepting unit 31 using the learning data created based on the word and the parallel translation example data corresponding to the source language example The unit 35 and the learning model generated by the learning model generation unit 35 are applied to the translation target words in the input text, the confidence factors are calculated for all the translation word candidates of the translation target words, and the translation word candidates are ordered in the confidence level. Corresponds to the translation target word in the target language example included in the bilingual example data corresponding to the source language example data output by the learning model application unit 36 and the similarity evaluation unit 34 That translation, or, learning model applying section 3
A translation word output unit 37 that selects the most suitable translation word from the translation word candidates output in step 6, that is, the translation word or the translation word candidate with the highest certainty factor and outputs it as the translation word corresponding to the translation target word. It is characterized by.

That is, based on the input text received by the input receiving unit 31 and the parallel translation example data stored in the parallel translation example data storage unit C, the example extraction unit 32 corresponding to the first translation word selection device A1 and the similarity detection. Either the translated word processed by the unit 33 and the similarity evaluation unit 34, or the translated word candidate processed by the learning model generation unit 35 and the learning model application unit 36 corresponding to the second translated word selection device A2,
The translated word is output by the translated word output unit 37. The parallel translation example data storage units C used by the first translation word selection device A1 corresponding part and the second translation word selection device A2 corresponding part may be the same or different.

In this case, it is preferable to adopt either of the following two modes.

That is, first, the first translation word selection device A1
The corresponding part and the corresponding part of the second translation word selection device A2 are operated in parallel, and the translation word output part 37 outputs the parallel translation example data for which the similarity evaluation part 34 has obtained a similarity equal to or higher than a predetermined threshold value. In some cases, the translated word corresponding to the translation target word obtained as a result is output, and when there is no output of the bilingual translation example data for which the similarity equal to or greater than a predetermined threshold value is output, the learning model application unit 36 outputs the result output. It is possible to adopt a mode of outputting a translated word corresponding to the translated target word. In this way, the translated words can be output quickly by parallel processing.

On the other hand, when the corresponding portion of the first translation word selection device A1 is operated first, and the parallel translation example data for which the similarity evaluation unit 34 has obtained a similarity of a predetermined threshold value or more is not output, the second translation word is output. It is also possible to adopt a mode in which the translation model output unit 37 is operated after operating the learning model generation unit 35 and the learning model application unit 36 that are the selection device relevant portions A2. In this way, the similarity evaluation unit 34
If there is parallel translation example data in which the similarity equal to or more than the threshold value is obtained, it is not necessary to operate the second translation word selection device corresponding portion A2, so that the load on the computer processing is reduced and the second translation word selection device A2 When operating the corresponding part, additional parallel translation example data can be additionally collected or selected, for example, by using a different parallel translation example data storage unit C.

In any of the above modes, the parallel translation example data storage unit used by the example extraction unit 32 and the parallel translation example data storage unit C used by the learning model generation unit 35 are based on different language resources. If they are created differently, it is possible to make the number and types of bilingual translation examples more diverse and improve the accuracy of the translated word finally output.

Further, according to the present invention, it is possible to construct a suitable translation device by using any of the translation word selection devices A1, A2 and A3 as described above. That is, the translation device, in addition to the configurations of the translation word selection devices A1, A2, and A3, targets corresponding to the input text based on the translation word output by the translation word output unit in any of them and the parallel translation example data including the translation word. It further comprises a translated sentence output unit for generating and outputting a text. By doing so, not only the translation word corresponding to the translation target word in the input text is selected, but also the target text translated in the second language is generated and output based on the input text in the first language. It becomes possible.

[0029]

BEST MODE FOR CARRYING OUT THE INVENTION An embodiment of the present invention will be described below with reference to FIGS.

This embodiment, whose schematic configuration is shown in FIG.
It is the translated word selection device A3 having the above-described third basic configuration. That is, the translated word selection device A having the first basic configuration
1, a part corresponding to the translation word selection device A2 having the second basic configuration, and a part common to these. Although the parallel translation example data storage unit C is included in the translation word selection device A3, it may be collected from the parallel translation example data storage unit C provided in another device connected by a communication line as needed. Is also possible. In the present embodiment, a case where Japanese is applied as the first language (source language) and English is applied as the second language (target language) will be described.

First, the parallel translation example data storage section C will be described. The bilingual example data storage unit C includes an example composed of Japanese text (hereinafter, “Japanese example”), a word included in the Japanese example, a translated word in English of the word (hereinafter, English translated word), and the English translated word. Japanese-English bilingual example data is stored which is a pair of Japanese example data including various information regarding the above and English example data including an English example including text translated into English corresponding to the Japanese example. It is a database. In addition, the Japanese-English bilingual example data further includes a Japanese headword that can be a translation target word for each Japanese example, and in some cases, an English headword that can be a correct translation corresponding to the Japanese headword is included. May be included. Examples of such Japanese-English bilingual example data include databases created in consideration of the frequency of occurrence based on articles such as newspapers and magazines, Japanese-English bilingual electronic dictionary databases, and other databases that can be used online. The stored data can be used.

FIG. 5 shows a part of an example of Japanese-English parallel translation example data. In this example, three Japanese examples including the Japanese word "nonsense" and corresponding English examples are paired. In this case, "don't care" applies to Japanese headwords and "feel constraine" to English headwords.
"d", "constraint", "refrain", etc. are applicable. However, for English headwords corresponding to Japanese headwords only, or for both Japanese headwords and English headwords, if there are already set, you can use it.
If not, it must be set manually or automatically set by computer processing.

Next, the function of the translation word selection device A3 will be described. This translation word selection device A3 is a first translation word when the internal and external devices of a normal computer such as a CPU and a memory operate according to a predetermined program stored in a storage device such as an HDD of a general-purpose computer or a dedicated computer. The example extraction unit 32, the similarity detection unit 33, and the similarity evaluation unit 34 that function as the selection device A1.
The learning model generation unit 35 and the learning model application unit 36 that have the function of the second translation word selection device A2, and the functions of the input reception unit 31 and the translation word output unit 37 that have the same functions as these are exhibited.

The input receiving section 31 receives input of text data (input text) created in Japanese. The input receiving unit 31 includes an input text processing unit 311. The input text processing unit 311 performs a morphological analysis on the input text and automatically extracts a translation target word from the input text. It should be noted that the translation target word can be designated at the time of inputting the input text, but in this case, the input text processing unit 311 only performs morphological analysis.

The example extraction unit 32 extracts the Japanese example data including the translation target word obtained by the input reception unit 31 in the parallel translation example data storage unit C. At this time, if the parallel translation example data storage unit C contains a Japanese headword, the corresponding translation target word is searched for and extracted by referring to it. The example extracting unit 32 includes a Japanese example processing unit 321 which is a source language example processing unit. The Japanese example processing unit 321 performs sentence end processing on the Japanese example data extracted from the bilingual example data storage unit C. For example, by performing sentence end processing for Japanese example data among the Japanese-English bilingual example data shown in FIG.
"Refrain from my mother", "Refrain from my mother", and "Refrain from my donation" mean "Refrain from my mother", "Refrain from my mother",
"Please refrain from donation."

The similarity detecting section 33 compares the input text accepted by the input accepting section 31 with the Japanese example data extracted by the example extracting section 32, and detects the similarity between them. Specifically, the similarity, which is the ratio of matching between the input text calculated by the similarity calculation unit 331 included in the similarity detection unit 33 and the Japanese example data, is detected as the similarity. That is, the degree of similarity is obtained as the ratio of the matched character strings by comparing the input text and the Japanese example data character by character by dynamic programming. More specifically, the similarity is calculated by the following equation using, for example, the UNIX diff command.

[0037]

[Formula 1]

It is calculated by It should be noted that the Japanese example data used is the one that has been subjected to sentence ending processing by the Japanese example processing unit 321.

The similarity evaluation unit 34 compares and evaluates the similarity detected by the similarity detection unit 33 for each Japanese example data compared with the input text, that is, the similarity obtained by the above equation, and the highest similarity. The Japanese example data for which the degree r has been obtained or the Japanese-English bilingual example data including the Japanese example data is output. At this time, when there are a plurality of Japanese example data for which the maximum similarity r is obtained, the Japanese example data including the longest Japanese example is output as having the highest similarity. However, it is limited to the case where the part that matches the input text is longer than the length of the Japanese headword.

The learning model generator 35 uses the learning data to generate a learning model corresponding to each translation target word in the input text accepted by the input acceptor 31. The learning data is created based on the words included in the Japanese example stored in the bilingual example data storage section C and the English example data corresponding to the Japanese example, and the words input in Japanese. , Corresponding correct words to be output in English, and information such as attributes and features associated with them. Further, in the present embodiment, as a learning model, for example, S
VM (Support Vector Machine), ME (Maximum Entr)
Opy), DL (Decision List), and other known machine learning models are applied. Then, by applying these learning models to each translation target word, the probability that each correct translation word is generated is obtained. that time,
It is necessary to give a feature to each learning model, but in the present embodiment, as a feature, morpheme information that is information obtained from the learning data, characters n-gram, information about a Japanese example that is the maximum match, content Four types of information, that is, information on the appearance frequency of a word and its translation candidate are used. The learning model generation unit 35 includes a learning model selection unit 351. The learning model selection unit 351 performs a cross validation using learning data for each learning model and selects a learning model having the highest accuracy.

The learning model application unit 36 applies the learning model generated by the learning model generation unit 35, specifically, the learning model selected by the learning model selection unit 351, to the translation target word in the input text, thereby The certainty factor is calculated for all the translated word candidates of the translation target word, and the translated word candidates are output by ordering in the certainty factor order. Basically, the certainty factor is the score of the probability distribution of the event (a, b) that becomes the class a (εA) in the context b (εB), where B is the set of contexts and A is the set of classification classes. It is obtained as p (a, b). Note that when such a probability distribution cannot be obtained depending on the type of learning model, for example, when SVM is applied, the probability value is set to 1 for the optimum class and 0 for the other classes for convenience. There is.

The translated word output unit 37 outputs a translated word corresponding to the translation target word in the input text, and is a translated word obtained from either the root of the translated word selection apparatus A1 or the root of the translated word selection apparatus A2, that is, The translation included in the Japanese-English bilingual example data corresponding to the Japanese example data that has the highest similarity in the similarity evaluating unit 34, or the learning model applying unit 3
Among the translation word candidates output in 6, the translation word candidate having the highest certainty factor (score) is selected and output. Specifically, in the present embodiment, a threshold is set for the degree of similarity obtained by the similarity calculator 331 in the similarity detector 33,
When the Japanese example data output by the similarity evaluation unit 34 is equal to or more than the threshold value, the translated word corresponding to the Japanese example data is output. In this embodiment, the threshold value is 1. On the other hand, if there is no Japanese example data above the threshold,
From the translation word candidates output by the learning model application unit 36, the one with the highest degree of certainty is output. When the input acceptance unit 31 accepts the input text, the route of the translation word selection device A1 and the route of the translation word selection device A2 may be operated simultaneously, or the route of the translation word selection device A1 may be operated first. The route of the translated word selection device A2 may be operated only when there is no Japanese example data equal to or more than the threshold value.

Hereinafter, an example of the usage of the translation word selection device A3 of the present embodiment is shown in FIGS. 6 and 7, and the translation word selection device A3.
This will be described with reference to the flowchart showing the operation procedure. In the following explanation, the inseparator of the present invention (participant name, CRL-NYU) word disambiguation contest 2nd SE
NSEVAL {hereinafter, "SENSEVAL2", 200
The translated word selection device A3 of the present embodiment is applied to the Japanese translation task held for one year (SENSEVAL-2 Organization Committee)}, and in this contest, the translated word selection device A3 before improvement is used. However, it has an extremely high reputation.

As a premise, Japanese-English parallel translation example data (320
Japanese headwords, about 20 examples per headword) are given in advance before the contest.
Same as the AL2 Japanese translation task. For 40 words selected from these (20 nouns, 20 verbs), 30 occurrences of test data are used, and the total number of Japanese words to be translated is 1200 words. Also,
It is allowed to use the parallel translation dictionary obtained from a language resource other than the given Japanese-English parallel translation example data and the Japanese-English parallel translation example data based on various newspaper articles in the contest wagon. Further, in order to fairly evaluate the correctness of the finally output translated word, the accuracy of the translated word is evaluated based on the predetermined input text, the translation target word, and the correct translated word.

In order to simplify the explanation, a mode in which the route of the translated word selection device A1 is started and the route is shifted to the translated word selection device A2 when the translated word is not output from the route will be described. As described above, both routes may proceed at the same time. First, when the input acceptance unit 31 accepts an input of an input text (for example, a Japanese text including the expression of "single role" which is an idiomatic expression) (FIG. 6; step S1), the input text processing unit 311 receives the input text. A word to be translated (for example, <buy>) is extracted by morphological analysis of (step S2). Next, the example extraction unit 32 searches the parallel translation example data storage unit C based on the extracted translation target word (<buy>), and extracts Japanese example data including the translation target word (step S3). , Japanese example processing unit 32
Sentence end processing is performed for each Japanese example included in the Japanese example data extracted by 1 (step S4). Next, with respect to each of the Japanese examples subjected to the sentence ending process and the input text, the similarity calculation unit 331 in the similarity detection unit 33.
Calculates the degree of similarity r based on Equation 1 (step S
5). Then, the number of Japanese examples in which the degree of similarity r is maximum is checked (step S6). If the number is 1 (step S6).
6; Y), the similarity evaluation unit 34 outputs Japanese example data including the Japanese example (step S7). on the other hand,
If the number of Japanese examples with the maximum similarity r is 1 or more in step S6 (step S6; N), the Japanese example data including the Japanese example with the longest similar character string is selected (step S6a). , The Japanese example data is output as the one having the highest similarity (step S
7). Here, in this case, the Japanese example with the highest similarity r includes the expression corresponding to the input text (“buy a role”), and the English example corresponding to the Japanese example in this Japanese-English bilingual example data. In addition, it is assumed that the English translation word (<to offer to help>) corresponding to the translation target word is included. Then, the similarity of the output Japanese example data is compared with a predetermined threshold value (for example, 1) (step S8), and if the similarity is equal to or higher than the threshold value (1) (step S8; Y), the translated word output unit 37 outputs the English translation word (for example, <offer>) corresponding to the translation target word (<buy>) (step S9). It should be noted that the English expression corresponding to the Japanese idiomatic expression "single role buy" is "to off
er to help, and in this case, if the correct English translation word for the translation target word “buy” is given as “offer”, the English translation word output in step S9 becomes the correct answer.

On the other hand, in step S8, the threshold value (1)
If there is no Japanese example data above (step S
8; N), that is, when there is no Japanese example data for the same or similar example as the Japanese example including the word to be translated in the input text, the process moves to the root of the translated word translation device A2 {S6 (N )}. In this case, the learning model generation unit 3
5, first, based on the translation target word in the input text received by the input receiving unit 31, the translation word selection device A1
The Japanese-English parallel translation example data storage unit C different from the one used in the above route is searched, and the Japanese example data including the corresponding word is extracted (FIG. 7, step S11). Then, learning data is applied to each Japanese example included in each extracted Japanese example data to generate a learning model (based on one of SVM, DL, and ME) (step S12). Further, the learning model selection unit 351 selects a learning model having the highest accuracy after performing cross validation using the learning data for each of the generated learning models (step S13). The learning model selected here is applied to the translation target word in the input text in the learning model application unit 36, and the confidence factor p is calculated for all the translation word candidates corresponding thereto (step S14), and the confidence factor p is calculated.
The translated word candidates are output in order, for example, in descending order (step S15). Finally, the translation word output unit 3 selects the translation word candidate having the highest confidence p from the output translation word candidates.
It outputs by 7 (step S16). If the output translation candidate matches the correct English translation given in advance, the English translation is correct.

For reference, FIG. 8 shows SENSEVAL2.
The results obtained by the translation word selection devices A1 and A2 in the contest are shown in a list. The result is a word (noun 20, verb 20) that is the translation target word given in the contest.
The accuracy rate of the English translation word output for each is shown as accuracy. As a result of applying the translation word selection device A1 to 100 out of the given 1200 translation target words,
The accuracy was 91.0%. Further, as a result of applying the translation word selection device A2 to 1100 translation target words, the accuracy is 6
It was 0.9%. For comparison, the comprehensive results (A1 + A2) by the translation word selection devices A1 and A2 are also shown in the list. From this result, the translation word selection device A1
For translation target words that were not accurate in
It can be said that it is appropriate to apply the translation word selection device A3 of this embodiment, which is to apply the translation word selection device A2. That is, the route to which the translation word selection device A1 based on the similarity of character strings is applied to an example in which the number of learning data is generally small such as including an idiomatic expression, in other words, an example in which the number of Japanese-English parallel translation example data is small. However, by applying the route of the translation word selection device A2 that obtains a certainty factor by applying the learning data and the learning model when the accuracy is poor in the above route, not only the expression that is usually used It can be said that it is possible to perform highly accurate translation selection, including idiomatic expressions.

The present invention is not limited to the embodiments described above. For example, the translation device is configured by using the translation word selection devices A1 and A2 independently or by providing a translation text output unit that generates and outputs a target text corresponding to the input text based on the translation word output by the translation word output unit. It is also possible. Further, other than that, the specific configuration of each part is not limited to the above-described embodiment, and various modifications can be made without departing from the spirit of the present invention.

[0049]

As described above in detail, according to the present invention, it is possible to improve the accuracy without putting a large amount of manpower and overloading the computer, that is, without collecting a large amount of parallel translation example data. It is possible to perform high translation selection as well as machine translation. In particular, the method based on the similarity of character strings and the method of applying the learning data and the learning model may be used separately, or may be used together to complement each other. It is possible to perform highly accurate translation selection and machine translation even for expressions such as and conventional expressions that occur infrequently.

[Brief description of drawings]

FIG. 1 is a schematic functional configuration diagram of a translation word selection device corresponding to a first aspect of the present invention.

FIG. 2 is a schematic functional configuration diagram of a translation word selection device corresponding to a second aspect of the present invention.

FIG. 3 is a schematic functional configuration diagram of a translation word selection device corresponding to a third aspect of the present invention.

FIG. 4 is a schematic functional configuration diagram of a translation word selection device according to an embodiment of the present invention.

FIG. 5 is a diagram showing an example of Japanese-English parallel translation example data used in the embodiment.

FIG. 6 is a schematic flowchart showing an operation procedure of the embodiment.

FIG. 7 is a schematic flowchart showing an operation procedure of the same embodiment.

FIG. 8 is a diagram showing a list of translation word selection results in a SENSEVAL2 contest to which the present invention has been applied.

[Explanation of symbols]

A1, A2, A3 ... Translated word selection device C ... Parallel translation example data storage unit 1, 11, 21, 31 ... Input acceptance section 2, 32 ... Example extraction unit 3, 33 ... Similarity detection unit 4, 34 ... Similarity evaluation unit 5, 14, 37 ... Translated word output section 12, 35 ... Learning model generation unit 13, 36 ... Learning model application unit 311 ... Input text processing unit 321 ... Source language example processing unit (Japanese example processing unit) 331 ... Similarity calculation unit 351 ... Learning model selection unit

   ─────────────────────────────────────────────────── ─── Continued front page    (72) Inventor Maki Murata             4-2-1 Kanaikitamachi, Koganei City, Tokyo Independent             Communications Research Laboratory (72) Inventor Hitoshi Isahara             4-2-1 Kanaikitamachi, Koganei City, Tokyo Independent             Communications Research Laboratory F term (reference) 5B091 AA05 CA02 CA22 CC01 EA02

Claims (32)

[Claims] 1. Source language example data including a source language example composed of text in a first language and a word included therein, a translated word of the word in a second language, and information related to the translated word, and second to second source language examples. It is a word to be translated that is included in the input text input in the first language by using the parallel translation example data storage unit that stores the parallel translation example data that is paired with the target language example including the text translated in the language. Selects a translated word described in a second language corresponding to a translation target word, and corresponds to the translation target word in the input text accepted by the input acceptance unit, which accepts the input of the input text. At least one or more source language example data including a word is extracted by the example extraction unit that extracts from the parallel translation example data storage unit, and the input text and the example extraction unit. Based on the language example data, the similarity detection unit that detects the similarity between the input text and the source language example and the similarity between the source language example detected by the similarity detection unit are comparatively evaluated, and at least the highest similarity is found. A similarity evaluation section that outputs source language example data, and a translated word that outputs a translated word corresponding to the translation target word in the target language example included in the bilingual example data corresponding to the source language example data output by the similarity evaluation section A translation word selection device comprising: an output unit. 2. The similarity detection unit matches the input text with the source language example based on the difference obtained by comparing the input text and the source language example included in the extracted source language example data on a character-by-character basis. There is a similarity calculation unit that calculates the similarity calculated using at least one of the proportion of the character string or the number of divisions where the matched portion is divided and matches. Claim 1
Described word selection device. 3. The example extraction unit includes a source language example processing unit that performs sentence end processing on the source language example included in the extracted source language example data and outputs a processed source language example.
In the similarity detection unit, the similarity calculation unit, based on the calculation result of the difference obtained by comparing the input text and the processed source language example on a character-by-character basis, 3. The translation word selection device according to claim 2, wherein at least one of the ratio to the character string or the number of divisions at which the matched portion is divided and matched is calculated as the similarity. 4. The similarity degree is calculated when the translated word output section has a plurality of source language example data having a maximum degree of similarity as a result of calculation by the similarity degree calculation section of the similarity detection section and evaluation by the similarity evaluation section. The translated word selection according to claim 3, wherein the translated word corresponding to the translation target word in the bilingual translation example data including the source language example in which the ratio of the character string matching the input text or the division number is the maximum is output as a result of the operation in the operation unit. apparatus. 5. The translation word selection device according to claim 1, 2, 3 or 4, wherein the input receiving unit has an input text processing unit for automatically extracting a translation target word by morphological analysis of the input text. 6. The parallel translation example data includes a source language headword generated based on a word included in the source language example, and the example extraction unit includes at least the source language headword corresponding to the translation target word. 4. The source language example data including the word is extracted from the bilingual example data storage unit.
4. The translation word selection device according to 4 or 5. 7. The parallel translation example data includes a source language headword generated based on a word included in the source language example and a target language headword generated based on a corresponding translated word. The extraction unit extracts at least source language example data including a source language headword corresponding to the translation target word, and a translated word output unit is included in the source language example data output by the similarity evaluation unit. The translation word selection device according to claim 1, 2, 3, 4, or 5, which outputs a target language headword corresponding to the source language headword extracted by the example extraction unit. 8. Source language example data including a source language example including text in a first language and a word included therein, a translated word of the word in a second language, and information about the translated word, and second to second source language examples. It is a word to be translated that is included in the input text input in the first language by using the parallel translation example data storage unit that stores the parallel translation example data that is paired with the target language example including the text translated in the language. Selects a translated word described in a second language corresponding to a translation target word, and includes an input receiving unit that receives an input of the input text, and a word included in the source language example stored in the translation example data storage unit. Using the learning data created based on the parallel translation example data corresponding to the source language example, the learning model corresponding to the translation target word in the input text accepted by the input accepting unit is used. Applying the learning model generated by the learning model generation unit and the learning model generated by the learning model generation unit to the translation target word in the input text, computing the certainty factor for all translation word candidates of the translation target word, and in the certainty factor order. From the learning model application unit that outputs translation word candidates in order, and the translation word candidate that has the highest certainty factor among the translation word candidates output by the learning model application unit, the translation word that is output as the translation word corresponding to the translation target word A translation word selection device comprising: an output unit. 9. A learning model generation unit extracts, from the parallel translation example data storage unit, parallel translation example data corresponding to a source language example including each translation target word in the input text received by the input receiving unit. 9. The translation word selection device according to claim 8, wherein a learning model is generated based on the extracted parallel translation example data. 10. A learning model generation unit generates a learning model corresponding to each learning data, and the learning data has the highest accuracy for each translation target word in the input text accepted by the input acceptance unit. The learning model selecting unit further includes a learning model selecting unit that selects a learning model that becomes, and the learning model applying unit applies the learning model selected by the learning model selecting unit to a translation target word in the input text. 8. A translation word selection device according to item 8 or 9. 11. The translation word selection device according to claim 8, 9 or 10, wherein the input reception unit has an input text processing unit for automatically extracting a translation target word from the input text by morphological analysis. 12. The parallel translation example data includes a source language headword generated based on a word included in the source language example, and the learning model generation unit at least the source language heading corresponding to the translation target word. 9. Source language example data including words is extracted from the bilingual example data storage unit.
The translated word selection device according to 9, 10 or 11. 13. Source language example data including a source language example consisting of text in a first language and a word included therein, a translated word of the word in a second language, and information about the translated word, and second to second source language examples. It is a word to be translated that is included in the input text input in the first language by using the parallel translation example data storage unit that stores the parallel translation example data that is paired with the target language example including the text translated in the language. Selects a translated word described in a second language corresponding to a translation target word, and corresponds to the translation target word in the input text accepted by the input acceptance unit, which accepts the input of the input text. The at least one or more source language example data including the words to be extracted from the parallel translation example data storage unit by the input text and the example extraction unit. Based on the source language example data, the similarity detection unit that detects the similarity between the input text and the source language example and the source language example detected by the similarity detection unit are compared and evaluated, and the highest similarity is found. At least a similarity evaluation unit that outputs source language example data, and learning data created based on the words included in the source language example stored in the source example data storage and the translation example data corresponding to the source language example. Apply the learning model generated by the learning model generation unit that generates a learning model corresponding to the translation target word in the input text received by the input reception unit and the learning model generated by the learning model generation unit to the translation target word in the input text Then, the learning model application unit that calculates the certainty factors for all the translated word candidates of the translation target word and outputs the translated word candidates in order of certainty factor, and the original language output by the similarity evaluation unit Corresponding to the translation target word by selecting the most suitable translation word corresponding to the translation target word in the target language example included in the bilingual translation example data corresponding to the example data or the translation candidate output by the learning model application unit And a translated word output unit that outputs the translated word as a translated word. 14. A translation target word obtained as a result of being output by the similarity evaluation unit when the translation evaluation unit outputs the parallel translation example data for which the similarity evaluation unit has obtained a similarity equal to or greater than a predetermined threshold value. Corresponding to the translation target word obtained as the result output by the learning model application unit when the parallelism example data for which the similarity evaluation unit obtains a similarity equal to or higher than a predetermined threshold is not output. 14. The translation word selection device according to claim 13, wherein the translation word is output. 15. When there is no output of parallel translation example data for which a similarity equal to or higher than a predetermined threshold is obtained by the similarity evaluation unit,
The translation word selection device according to claim 13, wherein the learning model generation unit, the learning model application unit, and the translation output unit are operated. 16. The parallel translation example data storage unit used by the example extraction unit and the parallel translation example data storage unit used by the learning model generation unit are different parallel translation example data storage units created based on different language resources. Claims 13 and 14
Alternatively, the translation word selection device according to item 15. 17. Source language example data including a source language example consisting of a text in a first language and a word included therein, a translated word of the word in a second language, and information about the translated word, and second to second source language examples. A translation in a second language based on an input text input in a first language by using a translation example data storage unit that stores a translation example data paired with a target language example composed of text translated in a language A target text that is a sentence is output, and an input reception unit that receives an input of the input text, and at least one or more source texts including a word corresponding to each translation target word in the input text received by the input reception unit. Based on the example extraction unit that extracts the language example data from the parallel translation example data storage unit, and the source text example data extracted by the input text and the example extraction unit. When the similarity detection unit that detects the similarity between the input text and the source language example and the source language example detected by the similarity detection unit are compared and evaluated, at least the source language example data that has the highest similarity is evaluated. A similarity evaluation unit that outputs, and a translation output unit that outputs a translation corresponding to the translation target word in the target language example included in the parallel translation example data corresponding to the source language example data output by the similarity evaluation unit, A translation device comprising: a translation output unit that generates and outputs a target text corresponding to an input text based on a translation output by a translation output unit and parallel translation example data including the translation. 18. Source language example data including a source language example consisting of text in a first language and a word contained therein, a translated word of the word in a second language, and information about the translated word, and second to second source language examples. It is a word to be translated that is included in the input text input in the first language by using the parallel translation example data storage unit that stores the parallel translation example data that is paired with the target language example including the text translated in the language. Selects a translated word described in a second language corresponding to a translation target word, and includes an input receiving unit that receives an input of the input text, and a word included in the source language example stored in the translation example data storage unit. And learning corresponding to the translation target word in the input text accepted by the input accepting unit, using the learning data created based on the parallel translation example data corresponding to the source language example. The learning model generator that generates Dell and the learning model generated by the learning model generator are applied to the translation target words in the input text, and the confidence factor is calculated for all translation word candidates of the translation target word, and the confidence factors are calculated in order of the confidence factor. From the learning model application unit that outputs translation word candidates in order, and the translation word candidate that has the highest certainty factor among the translation word candidates output by the learning model application unit, the translation word that is output as the translation word corresponding to the translation target word It is characterized by comprising an output unit and a translated sentence output unit for generating and outputting a target text corresponding to the input text based on the translated word output by the translated word output unit and the parallel translation example data including the translated word. Translation device. 19. A source language example data including a source language example consisting of text in a first language and a word included therein, a translated word of the word in a second language, and information about the translated word, and second to second source language examples. It is a word to be translated that is included in the input text input in the first language by using the parallel translation example data storage unit that stores the parallel translation example data that is paired with the target language example including the text translated in the language. Selects a translated word described in a second language corresponding to a translation target word, and corresponds to the translation target word in the input text accepted by the input acceptance unit, which accepts the input of the input text. The at least one or more source language example data including the words to be extracted from the parallel translation example data storage unit by the input text and the example extraction unit. Based on the source language example data, the similarity detection unit that detects the similarity between the input text and the source language example and the source language example detected by the similarity detection unit are compared and evaluated, and the highest similarity is found. At least a similarity evaluation unit that outputs source language example data, and learning data created based on the words included in the source language example stored in the source example data storage and the translation example data corresponding to the source language example. Apply the learning model generated by the learning model generation unit that generates a learning model corresponding to the translation target word in the input text received by the input reception unit and the learning model generated by the learning model generation unit to the translation target word in the input text Then, the learning model application unit that calculates the certainty factors for all the translated word candidates of the translation target word and outputs the translated word candidates in order of certainty factor, and the original language output by the similarity evaluation unit A translation word corresponding to the translation target word in the target language example included in the bilingual translation example data corresponding to the word example data, or a translation word candidate output from the learning model application unit, is selected as the translation target word. A translation output unit that outputs the corresponding translation, and a translation output unit that generates and outputs a target text corresponding to the input text based on the translation output by the translation output unit and the parallel translation example data including the translation A translation device characterized by: 20. Source language example data including a source language example including text in a first language and a word included therein, a translated word of the word in a second language, and information related to the translated word, and second to second source language examples. A computer is operated using the parallel translation example data storage unit that stores the parallel translation example data paired with the target language example composed of the text translated in the language, and the translation included in the input text input in the first language is performed. A program for selecting a translated word written in a second language corresponding to a translation target word that is a power word, the input receiving step of receiving an input of the input text, and the translation target word in the received input text. At least one or more source language example data including the corresponding word,
An example extraction step of extracting from the parallel translation example data storage unit; a similarity detection step of detecting a similarity between the input text and the source language example based on the input text and the extracted source language example data; A similarity evaluation step of comparing and evaluating the similarities of the source language examples and outputting at least the source language example data having the highest similarity, and a target language example included in the bilingual example data corresponding to the output source language example data. A translation word output step of outputting a translation word corresponding to the translation target word in the translation word selection program. 21. The similarity detection step matches the input text and the source language example based on a difference obtained by comparing the input text and the source language example included in the extracted source language example data on a character-by-character basis. A similarity calculation step of calculating the similarity calculated using at least one of the ratio of the character string or the number of divisions at which the matched portion is divided to match The translation word selection program according to claim 21. 22. The example extraction step includes a source language example processing step of subjecting the source language example included in the source language example data extracted in the example extraction step to end-of-sentence processing and outputting a processed source language example. In the calculation step, based on the calculation result of the difference obtained by comparing the input text and the processed source language example on a character-by-character basis, the ratio of the matched character string to the character string of the processed source language example, or the matched portion. 22. The translation word selection program according to claim 21, wherein at least one of the numbers of divisions indicating the number of divisions in which is matched is calculated as the similarity. 23. In the translated word output step, the result output in the similarity output step and evaluated in the similarity evaluation step,
In the case where there are a plurality of source language example data having the highest degree of similarity, the translation in the parallel translation example data including the character string matching the input text or the source language example having the largest division number as a result of the operation in the difference operation step. 23. The translation word selection device according to claim 22, which outputs a translation word corresponding to the target word. 24. Source language example data including a source language example consisting of text in a first language and a word contained therein, a translated word of the word in a second language, and information about the translated word, and second to second source language examples. A computer is operated using the parallel translation example data storage unit that stores the parallel translation example data paired with the target language example composed of the text translated in the language, and the translation included in the input text input in the first language is performed. A program for selecting a translated word written in a second language corresponding to a translation target word that is a power word, and an input receiving step for receiving the input of the input text, and an example for the source language stored in the parallel translation example data storage unit. Using the learning data created based on the parallel translation example data corresponding to the source language example and the words included in A learning model generation step of generating a learning model corresponding to the translation target word in the list, applying the generated learning model to the translation target word in the input text, and determining the certainty factor for all translation word candidates of the translation target word. A learning model applying step of calculating and ordering the candidate words in order of certainty, and outputting the translated word candidate having the highest certainty factor from the output translated word candidates and outputting it as the translated word corresponding to the translation target word. And a translated word output step of: 25. In the learning model generating step, for each translation target word in the input text accepted in the input accepting step, bilingual example data corresponding to a source language example including it is extracted from the bilingual example data storage unit, 25. The translation word selection program according to claim 24, which generates a learning model based on the extracted parallel translation example data. 26. In the learning model generating step, a learning model is generated corresponding to each learning data, and the learning data has the highest accuracy for each of the translation target words in the input text accepted in the input accepting step. 26. The translated word according to claim 24, further comprising a learning model selecting step of selecting a model, wherein the learning model applying step applies the learning model selected in the learning model selecting step to a translation target word in the input text. Choice program. 27. Source language example data including a source language example consisting of text in a first language and a word included therein, a translated word of the word in a second language, and information related to the translated word; and a source language example second. A computer is operated using the parallel translation example data storage unit that stores the parallel translation example data paired with the target language example composed of the text translated in the language, and the translation included in the input text input in the first language is performed. A program for selecting a translated word written in a second language corresponding to a translation target word that is a power word, the input receiving step of receiving an input of the input text, and the translation target word in the received input text. At least one or more source language example data including the corresponding word,
An example extraction step of extracting from the parallel translation example data storage unit; a similarity detection step of detecting a similarity between the input text and the source language example based on the input text and the extracted source language example data; A similarity evaluation step of comparing and evaluating the similarities of the source language examples and outputting at least the source language example data having the highest similarity, and the words included in the source language examples stored in the bilingual example data storage and the corresponding source A learning model generating step of generating a learning model corresponding to a translation target word in the input text accepted by the input accepting unit, using the learning data created based on the parallel translation example data corresponding to the language example; The learning model is applied to the translation target word in the input text, and the certainty factor is calculated for all translation candidate words of the translation target word, A learning model applying step for outputting candidate translations ordered in order of degree, and a translated word corresponding to the translation target word in the target language example included in the bilingual example data corresponding to the source language example data output in the similarity evaluation step, or And a translation word output step of selecting an optimum translation word candidate output from the learning model application step and outputting the translation word as a translation word corresponding to the translation target word. 28. In the translated word output step, when there is output of parallel translation example data for which a similarity equal to or more than a predetermined threshold is obtained in the similarity evaluation step, a translation target word obtained as a result of the output in the similarity evaluation step. Corresponding to the translation target word obtained as the result output by the learning model application unit when there is no output of the parallel translation example data in which the similarity equal to or more than a predetermined threshold is obtained in the similarity evaluation step. 28. The translated word selection program according to claim 27, which outputs translated words. 29. When there is no output of parallel translation example data for which a similarity equal to or greater than a predetermined threshold is obtained in the similarity evaluation step, the learning model generation step, the learning model application step, and the translated word output step are performed. 29. The translation word selection device according to claim 28. 30. Source language example data including a source language example including text in a first language and a word included therein, a translated word of the word in a second language, and information about the translated word, and second to second source language examples. Based on the input text input in the first language, the computer is operated by using the parallel translation example data storage unit that stores the parallel translation example data paired with the target language example consisting of the text translated in the language. A target text, which is a translation in two languages, is output, and an input receiving step of receiving the input of the input text, and at least one or more including a word corresponding to each translation target word in the received input text. An example extraction step of extracting source language example data from the bilingual example data storage, the input text and the extracted source language Based on the example data, a similarity detection step of detecting the similarity between the input text and the source language example, and the comparative evaluation of the similarity of the detected source language example, at least the source language example data having the highest similarity. A similarity evaluation step of outputting, a translated word output step of outputting a translated word corresponding to the translation target word in the target language example included in the parallel translation example data corresponding to the output source language example data, the output translated word, and A translation program output step of generating and outputting a target text corresponding to an input text based on parallel translation example data including the translation word. 31. Source language example data including a source language example including text in a first language and a word included therein, a translated word of the word in a second language, and information about the translated word, and second to second source language examples. Based on the input text input in the first language, the computer is operated by using the parallel translation example data storage unit that stores the parallel translation example data paired with the target language example consisting of the text translated in the language. A target text that is a translation in two languages is output, and an input receiving step of receiving the input of the input text, a word included in the source language example stored in the bilingual example data storage unit, and the source language example. Using the learning data created based on the bilingual example data corresponding to, it corresponds to the translation target word in the input text accepted by the input acceptance unit. And a learning model generation step of generating a learning model, applying the generated learning model to the translation target word in the input text, calculating the certainty factor for all translation word candidates of the translation target word, and ordering in the certainty factor A learning model applying step of outputting a translation word candidate; a translation word output step of selecting a translation word candidate having the highest certainty factor from the output translation word candidates and outputting the translation word candidate as a translation word corresponding to the translation target word; And a translated sentence output step of generating and outputting a target text corresponding to the input text based on the translated word and the parallel translation example data including the translated word. 32. Source language example data including a source language example including text in a first language and a word included therein, a translated word of the word in a second language, and information related to the translated word; Based on the input text input in the first language, the computer is operated by using the parallel translation example data storage unit that stores the parallel translation example data paired with the target language example consisting of the text translated in the language. A target text, which is a translation in two languages, is output, and an input receiving step of receiving the input of the input text, and at least one or more including a word corresponding to each translation target word in the received input text. An example extraction step of extracting source language example data from the bilingual example data storage, the input text and the extracted source language Based on the example data, a similarity detection step of detecting the similarity between the input text and the source language example, and the comparative evaluation of the similarity of the detected source language example, at least the source language example data having the highest similarity. Input acceptance using the similarity evaluation step to output and the learning data created based on the words included in the source language example stored in the source translation example data storage and the parallel translation example data corresponding to the source language example. A learning model generation step of generating a learning model corresponding to a translation target word in the input text accepted by the section, applying the generated learning model to the translation target word in the input text, and selecting a translation candidate of the translation target word. A learning model application step that computes the certainty factors for all and outputs candidate words in order of certainty factor, and a source language that is output in the similarity evaluation step Corresponding to the translation target word by selecting the optimum one from the translation words corresponding to the translation target words in the target language example included in the bilingual translation example data corresponding to the example data, or the translation word candidates output in the learning model applying step And a translated sentence output step of generating and outputting a target text corresponding to the input text based on the output translated word and the parallel translation example data including the translated word. Translation program.