JP2016045751A - Machine translation device, machine translation method, machine translation program, and recording medium - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a machine translation device that selects a field dictionary used for translation of a translation target sentence with high speed and precision, a machine translation method, a machine translation program, and a recording medium.SOLUTION: A machine translation device includes: a field correspondence information storage unit for storing field correspondence information showing correspondence between a translation condition and a field; a dictionary storage unit for storing a plurality of field dictionaries preserving translation in accordance with the field; and a dictionary determination unit for determining the field dictionary used for translation of the translation target sentence from the plurality of field dictionaries based on the correspondence between the translation condition that coincides with a predetermined unit description in the translation target sentence and the field.SELECTED DRAWING: Figure 1

Description

  The present invention relates to a machine translation device, a machine translation method, a machine translation program, and a recording medium that estimate the field of a translation target sentence.

  When translating a sentence written in a specific natural language into another natural language, machine translation, in which a computer is used to translate mechanically, is widely used instead of manual translation.

  One method for improving the translation accuracy of machine translation is to use a field dictionary. The field dictionary is a dictionary in which dictionary data for controlling a translation word is stored when a sentence to be translated is described in a specific field. For example, when translating English “court” into Japanese, multiple translations such as “Court” and “Court” are applicable. Can't determine if is appropriate. However, it is appropriate to translate “court” if the sentence to be translated is related to “law” and “coat” if it is related to “sports”. A field dictionary is a dictionary in which dictionary data for controlling translations according to fields is stored. By using the field dictionary, it becomes possible to appropriately select the translation of the sentence to be translated according to the field.

  Regarding the method of selecting a field dictionary to be used, if the field of the sentence to be translated is known in advance, a method of manually setting the field and selecting the field dictionary to be used is simple. On the other hand, if the field of the text to be translated is not known in advance or if there are many text to be translated and it takes time to manually set the field, it would be convenient if the field could be estimated automatically during machine translation. is there. However, if the field is mistakenly estimated and a field dictionary other than the field of the original text to be translated is selected, translation into the wrong translated word is performed all at once, resulting in a significant decrease in translation accuracy.

As a typical method for solving such a problem, there is a method for estimating a field of a translation target sentence by using a vector operation. For example, Patent Document 1 describes a method for calculating the degree of approximation between a multidimensional feature vector having a frequency of appearance of each word of a translation target sentence as a one-dimensional element and a representative feature vector in each field. In general, cosine similarity is often used for calculating the degree of approximation of two vectors. The cosine similarity is a method in which the similarity is higher as the angle between two vectors is smaller. The cosine of the angle θ formed by the vector x and the vector y can be obtained by dividing the inner product of x and y by the length (norm) of x and y as shown in Equation (1).

  Moreover, in patent document 2, the appearance frequency information in each field | area dictionary of the word contained in a translation object sentence is used.

JP 2007-080221 A JP 2001-005812 A

  However, in the method of Patent Document 1, since the degree of approximation of a vector having the dimension of the number of words that is an element of the feature vector is calculated for each field, the calculation amount is very large and the translation speed is slow. For example, when the cosine similarity is used as the degree of approximation, it is necessary to multiply the number of words that are the elements of the feature vector by the number of fields. Furthermore, if the number of fields increases or the number of words that are elements of feature vectors increases, the amount of calculation increases. In addition, when the number of words in the translation target sentence is small, the estimation accuracy tends to decrease.

  In the method of Patent Literature 2, since the appearance frequency of words included in the translation target sentence is obtained for each field dictionary, the accuracy of estimation decreases when many words having different translation words are included in different field dictionaries. Cheap. Further, when the number of words in the translation target sentence is small, the estimation accuracy is likely to be further reduced.

  An object of the present invention is to provide a machine translation apparatus, a machine translation method, a machine translation program, and a recording medium that perform selection of a field dictionary used for translation of a translation target sentence at high speed and high accuracy.

  In order to solve the above-described problem, the machine translation apparatus of the present invention includes a field correspondence information storage unit that stores field correspondence information indicating a correspondence relationship between a translation condition and a field, and a plurality of words corresponding to the field. A dictionary storage unit for storing the field dictionary, and a translation of the sentence to be translated from a plurality of the field dictionaries based on the correspondence relationship between the field and the translation condition that matches a predetermined unit description in the sentence to be translated And a dictionary determining unit for determining the field dictionary to be used.

  In addition, the machine translation device of the present invention includes a field correspondence information storage unit that stores field correspondence information indicating a correspondence relationship between a translated word and a field, and a dictionary that stores a plurality of field dictionaries that hold the translated words corresponding to the field. A translation rule storage unit for storing the translation rule including a storage unit, a translation condition and the translation corresponding to the translation condition, and the translation corresponding to the translation condition matching a predetermined unit description in the translation target sentence And a dictionary determination unit that determines the field dictionary to be used for translating the sentence to be translated from a plurality of the field dictionaries based on the correspondence relationship between the field and the field.

  Further, the machine translation method of the present invention is based on translation conditions that match a predetermined unit description in the sentence to be translated, and field correspondence information that indicates a correspondence relationship between the translation conditions and the fields, and is translated according to the field. The field dictionary to be used for translation of the translation target sentence is determined from a plurality of field dictionaries holding

  The machine translation method of the present invention includes a translation rule including a translation condition and a translation corresponding to the translation condition, and the translation and field corresponding to the translation condition matching a predetermined unit description in the translation target sentence. The field dictionary to be used for translation of the translation target sentence is determined from a plurality of field dictionaries holding the translated words corresponding to the field based on field correspondence information indicating the correspondence relationship between the field and the field.

  In addition, the machine translation program of the present invention includes a field correspondence information storage function for storing field correspondence information indicating a correspondence relationship between translation conditions and fields, and a plurality of field dictionaries holding translated words corresponding to the fields. Based on the correspondence relationship between the dictionary storage function for storing and the translation condition that matches a predetermined unit description in the translation target sentence and the field, the translation target sentence is used from a plurality of the field dictionaries. A dictionary determination function for determining a field dictionary is realized.

  In addition, the machine translation program of the present invention includes a field correspondence information storage function for storing field correspondence information indicating a correspondence relationship between a translated word and a field, and a plurality of field dictionaries for holding the translation words corresponding to the field. Corresponding to the translation rule storage function for storing the translation rule including the dictionary storage function to store, the translation rule including the translation condition and the translation corresponding to the translation condition, and the translation condition matching the predetermined unit description in the translation target sentence Based on the correspondence relationship between the translated word and the field, a dictionary determination function for determining the field dictionary to be used for translation of the translation target sentence from a plurality of field dictionaries is realized.

  With the machine translation device, machine translation method, machine translation program, and recording medium of the present invention, it becomes possible to select a field dictionary used for translation of a translation target sentence at high speed and with high accuracy.

It is a figure which shows the structural example of the machine translation apparatus of embodiment of this invention. It is a figure which shows the operation example of embodiment of this invention. It is a figure which shows the other structural example of the machine translation apparatus of embodiment of this invention. It is a figure which shows the other operation example of embodiment of this invention. It is a figure which shows the example of data of a dictionary memory | storage part. It is a figure which shows the example of data of a translation rule. It is a figure which shows the example of data of field corresponding | compatible information. It is a figure which shows the structural example of the machine translation apparatus of 2nd embodiment of this invention. It is a figure which shows the operation example of 2nd embodiment of this invention. It is a figure which shows the example of creation of a field classification tree. It is a figure which shows the example of data of the field corresponding | compatible information of 4th embodiment of this invention. It is a figure which shows the other structural example of the machine translation apparatus of embodiment of this invention. It is a figure which shows the other structural example of the machine translation apparatus of embodiment of this invention.

  FIG. 1 shows a configuration example of a machine translation apparatus according to an embodiment of the present invention.

  The machine translation device 1 includes a dictionary storage unit 11, a field correspondence information storage unit 12, and a dictionary determination unit 13. The dictionary storage unit 11 is a part that stores a plurality of field dictionaries. The field dictionary is a dictionary in which dictionary data for controlling a translation word is stored when a sentence to be translated is described in a specific field. The field correspondence information storage unit 12 is a part that stores field correspondence information indicating a correspondence relationship between a translation rule condition part (translation condition) and a field. The dictionary determination unit 13 compares the translation conditions in the translation target sentence 2 and the field correspondence information storage unit 12, and based on the matching translation conditions and field correspondence information in the translation target sentence 2, a plurality of fields in the dictionary storage unit 11. This is a part for determining a field dictionary to be used for translation of the translation target sentence 2 from the dictionary.

  By configuring the machine translation apparatus 1 as described above, it becomes possible to select a field dictionary used for translation of a translation target sentence at high speed and with high accuracy.

  Next, FIG. 2 shows an example of the operation of the dictionary determination unit 13 of the machine translation apparatus 1 in the configuration of FIG.

  First, the translation conditions 2 and the translation conditions stored in the field correspondence information storage unit 12 are compared (S101). Next, a field dictionary to be used for translation of the translation target sentence 2 is determined based on the matching translation conditions in the translation target sentence 2 and the field correspondence information in the field correspondence information storage unit 12 (S102).

  By operating the machine translation apparatus 1 as described above, it is possible to select a field dictionary used for translation of a translation target sentence at high speed and with high accuracy.

  FIG. 3 shows another configuration example of the machine translation apparatus according to the embodiment of the present invention.

  In FIG. 3, a translation rule storage unit 13 is added to FIG. The translation rule storage unit 13 is a part for storing a translation rule including a translation rule condition (translation condition) and a translation corresponding to the translation condition. The dictionary storage unit 11 is the same as described in FIG. The field correspondence information storage unit 12 is a part that stores field correspondence information indicating a correspondence relationship between a translation word and a field, instead of the above-described translation condition and field correspondence. The dictionary determination unit 13 compares the translation conditions of the translation target sentence 2 and the translation rule storage unit 13, and based on the translation word corresponding to the translation condition matched in the translation target sentence 2 and the field correspondence information, This is a part for determining a field dictionary to be used for translation of the translation target sentence 2 from the field dictionary.

  By configuring the machine translation apparatus 1 as described above, it becomes possible to select a field dictionary used for translation of a translation target sentence at high speed and with high accuracy.

  Next, an example of the operation of the dictionary determination unit 13 of the machine translation apparatus 1 in the configuration of FIG. 3 is shown in FIG.

  First, the translation condition 2 and the translation conditions stored in the translation rule storage unit 13 are compared (S201). Next, a field dictionary to be used for translation of the translation target sentence 2 is determined based on the translation corresponding to the translation condition that matches in the translation target sentence 2 and the field correspondence information in the field correspondence information storage unit 12 (S202).

  By operating the machine translation apparatus 1 as described above, it is possible to select a field dictionary used for translation of a translation target sentence at high speed and with high accuracy.

  Next, the first embodiment of the present invention will be described using a specific example.

  First, the configuration example is the same as in FIG.

  An example of data stored in the dictionary storage unit 11 is shown in FIG. In FIG. 5, two, two, and three dictionary data are held in each of the “law”, “tennis”, and “basket” field dictionaries. Using this field dictionary, for example, “court” in English means “court” or “court” or “tennis” field dictionary when “law” field dictionary is selected. "Coat" will be translated.

  As for the storage format of the field dictionary, in FIG. 5, the dictionary data (original word translation word pair) is stored for each field. However, a field dictionary in a format in which field information is given in units of dictionary data may be used. I do not care.

  Next, in order to explain the field correspondence information storage unit 12, first, translation rules will be described. FIG. 6 shows an example of translation rule data. The translation rule is composed of a condition part (translation condition) and an action part, and means that the action part operates when there is a part that matches the condition part in the sentence to be translated. In the case of FIG. 6, two translation rules are held, and each condition part is “adjacent (federal, court)” and “co-occurrence (cager, traveling)”. In this embodiment, “adjacent (federal, court)” is “caged” and “court” is “cager” when “federal” and “court” are adjacent in the translation target document. This means that both “traveling” are present in the translation target document. Since various methods can be considered for the determination criterion of “co-occurrence”, it is also possible to apply the method. For example, when the number of characters or the number of words between the words specified in the condition part is within a predetermined number, or when the specified words are simultaneously present in the translation target sentence at a predetermined number or more, “co-occurrence” It is also possible to judge. In the translation rule of FIG. 6, when “adjacent (federal, court)” is applied, “federal” in that part is translated as “federal”, “court” is translated as “court”, and “co-occurrence (cager, When traveling) is applied, “traveling” is translated as “traveling”.

  An example of the field correspondence information storage unit 12 is shown in FIG. In this example, data that associates the field with the condition part (translation condition) of the translation rule is stored, and the translation condition of “adjacent (federal, court)” is “law”, but “co-occurrence (cager, traveling) ) "Is associated with" basket ". In FIG. 7, one field is associated with one translation condition, but a plurality of fields may be associated with one translation condition.

  Next, the operation of the dictionary determination unit 13 will be described. As shown in FIG. 2, the dictionary determination unit 13 first compares the translation target sentence 2 with the translation conditions stored in the field correspondence information storage unit 12 (S101). Next, a field dictionary to be used for translation of the translation target sentence 2 is determined based on the matching translation conditions in the translation target sentence 2 and the field correspondence information in the field correspondence information storage unit 12 (S102). As shown in FIG. 3, the translation rule storage unit 13 is added to FIG. 1, and the translation conditions 2 and the translation conditions stored in the translation rule storage unit 13 are compared in S101 of FIG. The same result is obtained.

  For the sake of simplicity, an example will be given in which only the translation condition of “adjacent (federal, court)” in the field correspondence information in FIG. At this time, referring to the field correspondence information in FIG. 7, the “law” field is associated with the translation rule of “adjacent (federal, court)”. Therefore, it can be estimated that the field of the translation target sentence is “law”. As described above, the field can be estimated by examining the field associated with the matching translation condition in the translation target sentence 2 with the field correspondence information in the field correspondence information storage unit 12. The field dictionary of the field estimated in this way is determined as the field dictionary used for translation of the translation target sentence 2.

  As a method for determining a dictionary when a plurality of translation rules are matched, the field corresponding information storage unit 12 checks the field associated with the matched translation condition, and calculates the fitness for each field from the number of matching translation conditions matched. However, a method in which a field having a high fitness is used as an estimation field can be considered. Any function may be used for the calculation of the fitness as long as the fitness increases as the number of matches increases based on the number of matches of the matched translation conditions.

As an example, a case will be described in which a linear function related to the number of matches of each translation condition is used for calculation of the fitness. The number of translation conditions registered in the field correspondence information storage unit 12 is a, and the number of fields is b. The weight coefficient between the i-th translation condition (1 ≦ i ≦ a) and the j-th field (1 ≦ j ≦ b) is m (i, j) (m (i, j) ≧ 0) in advance. It is determined and stored in the field correspondence information storage unit 12 or the like. The weighting factor between the translation condition and the field that are not associated in the field correspondence information storage unit 12 is assumed to be zero. At this time, if the number of matches of the i-th translation condition is x _i , the fitness can be calculated by a linear function relating to the number of matches of each translation condition, for example, Equation (2).

  When the function of Equation (2) is used for calculating the fitness, the number of multiplications is the product of the number of translation conditions registered in the field correspondence information storage unit 12 and the number of fields (a × b times). . However, since the weighting coefficient between the unrelated translation condition and the field is 0, the field associated with the matched translation condition is checked in the field correspondence information storage unit 12, and multiplication (weight) for the unrelated field is performed. (Multiplication with a coefficient of 0) can not be performed. In this way, the number of multiplications can be reduced. For example, when there is one field associated with each translation condition as in the field correspondence information of FIG. 7, the number of multiplications is the number of matching translation conditions (a maximum). In the method using the cosine similarity of the feature vector of the translation target sentence and the field dictionary, the number of words (c) that are the elements of the feature vector is multiplied by the number of fields (b) (c × b times). . In the method according to the present embodiment, the number of multiplications can be reduced by making the number of field condition information translation conditions smaller than the number of words that are elements of feature vectors. In addition, if the number of fields associated with each translation condition of the field correspondence information is reduced, the number of multiplications can be further reduced.

  Furthermore, in the formula (2), it is possible to set all the weighting factors in the field associated with each translation condition to 1. At this time, the fitness is the total number of matching of the translation conditions associated with each field. In this way, the number of matching of the translation conditions is counted and totaled for each field, and the number of multiplications is 0. Therefore, the calculation amount is significantly larger than the method of calculating the cosine similarity of the feature vector. Can be reduced.

  For example, in the case of the field correspondence information in FIG. 7, the weighting factors of the “law” field of the translation condition “federal, court” and the “basket” field of the translation condition “cager, traveling” are set to 1. The other weighting factors are set to 0. If the number of matches for the translation condition of “adjacent (federal, court)” is 10 and the number of matches for the translation condition of “cager, traveling” is 3, the conformity in the “law” field. Is calculated as 10, and the fitness of the “basket” field is calculated as 3. As a result, a “law” having a higher fitness is estimated as the field of the translation target text. Also, if the weight factor for the “law” field of “adjacent (court)” is 1, and the weight factor for the “basket” field of “cager, traveling” is 5, the fitness of the “law” field Is 10, and the fitness of the “basket” field is 15. As a result, a “law” having a higher fitness is estimated as the field of the translation target text.

  The number of times that the translation condition is met is the number of times that it is determined that the translation condition is met. In the case of the present embodiment, the number of matches in the case of a translation rule with “adjacent” as a condition is the number of times that the specified word is present adjacent to each other. The number of matches in the case of the translation condition with “co-occurrence” as a condition is the total number of appearances of the specified word in the translation target document.

  Further, it is possible to apply a plurality of fields at the same time and use a plurality of field dictionaries at the same time for translation, instead of narrowing the estimated field of the translation target sentence to one.

  As described above, in the first embodiment of the present invention, since the field associated with the matched translation condition is examined and counted by the field correspondence information, it is compared with the method of calculating the cosine similarity of the feature vector. The amount of calculation can be reduced, and the time for field estimation can be shortened. Moreover, even if the number of words in the translation target sentence is small, the field can be estimated if there is a translation condition that satisfies the condition. Therefore, it becomes possible to select the field dictionary used for translation of the translation target sentence at high speed and with high accuracy.

  Note that there may be a case where the translation of the determined field dictionary is different from the translation of the translation rule. In such a case, as a method of processing the competitive relationship, there is a method of preferentially translating the translation rule translation word for a translation condition with a high number of matches, or performing field estimation again without the conflicting rule. Conceivable. However, conflict resolution is not the core technology of the present invention, and it is not necessary to determine the resolution means, and therefore the method is not finalized.

  Next, a second embodiment of the present invention will be described.

  A configuration example of the second embodiment of the present invention is shown in FIG. A field dictionary management unit 15 and a translation rule storage unit 13 are added to FIG. The translation rule storage unit 13 is a part that stores translation rules. The field dictionary management unit 15 is a part that generates field correspondence information from the translation rules in the translation rule storage unit 13 and the field dictionary in the dictionary storage unit 11 and stores the field correspondence information in the field correspondence information storage unit 12.

  Next, an example of the operation of the field dictionary management unit 15 will be described with reference to FIG.

  First, a field classification tree is created from the field dictionary in the dictionary storage unit 11 (S301). The field classification tree is a condition tree for estimating a field from dictionary data (original word translation word pair). An example of creating a field classification tree is shown in FIG. The dictionary data held in the field unit is converted into a format in which the field is associated with the dictionary data.

  Next, dictionary data is extracted one by one from the field classification tree, and it is confirmed whether the dictionary data is in the action part of the translation rule storage part 14 (S302 to 303). In some cases, the translation condition and the field are linked and registered in the field correspondence information storage unit (S304 to S306). This is repeated until there is no unconfirmed translation rule (S307). When there is no unconfirmed translation rule, one new dictionary data is extracted (S302). The steps of S302 to S307 are repeated until there is no unconfirmed dictionary data (S308).

  In the case of the field classification tree of FIG. 10, the dictionary data is extracted from the field classification tree until dictionary data existing in the action part of the translation rule in the translation rule storage unit 14 of FIG. 6 is found. Since “court → courtroom, court” exists in the translation rule storage unit 14, the translation rule “adjacent (federal, court)” and the field “law” are connected and registered in the field correspondence information storage unit 12. Similarly, since “traveling → travelling” exists in the translation rule storage unit 14, the translation rules “cager, traveling” and “basket” are connected and registered in the field correspondence information storage unit 12. In this way, the field correspondence information of FIG. 7 can be obtained.

  Here, the generation of the field correspondence information storage unit 12 in the field dictionary management unit 15 may be performed at the time when the translation target document 2 is input, but is generated in advance before the translation target document 2 is input. By holding the data, the time required for field estimation can be shortened.

  As described above, the field dictionary management unit 15 can generate field correspondence information in which a translation condition and a field are associated from the translation rule in the translation rule storage unit 14 and the field dictionary in the dictionary storage unit 11. By using the field correspondence information generated by the field dictionary management unit 15, it is possible to perform field estimation by the dictionary determination unit 13 described in the first embodiment. As a result, as in the first embodiment, the field associated with the matching translation condition is examined and counted in the field correspondence information, so that the amount of calculation can be reduced compared with the method of performing the vector operation, and the time for field estimation is reduced. Can be shortened. Moreover, even if the number of words in the translation target sentence is small, the field can be estimated if there is a matching translation condition. Therefore, also in the second embodiment, it becomes possible to select a field dictionary used for translation of a translation target sentence at high speed and with high accuracy.

  Next, a third embodiment of the present invention will be described.

  In the first and second embodiments, the data (for example, FIG. 7) in which the translation condition is associated with the field is used as the field correspondence information. However, in the third embodiment, the dictionary data (original language word pair) and the field are Use the associated data.

  The field correspondence information of the third embodiment is obtained by converting the field dictionary in the dictionary storage unit 11 into a format for associating the field with dictionary data in the same manner as the field classification tree generation method described in the second embodiment. Can be generated. For example, a field classification tree as shown in FIG. 10 can be used as field correspondence information. Furthermore, it is also possible to reduce the dictionary data of the field classification tree to only those existing in the action part of the translation rule in the translation rule storage unit 14.

  Next, the field estimation method in the third embodiment will be described. In the field estimation method of the third embodiment, the number of translations of the dictionary data (source language translation pair) translated in the action part of the translation rule is used instead of the number of matching of the translation conditions used in the first and second embodiments. use. Field estimation can be performed in the same manner as in the first and second embodiments, except that the number of dictionary data translations is used instead of the number of translation condition matches. The number of translations of dictionary data can be obtained by summing up the number of matches of translation conditions corresponding to certain dictionary data. If there is only one translation condition corresponding to a certain dictionary data, the number of translations of the dictionary data and the number of matches of the translation conditions corresponding to the dictionary data are the same.

The degree of matching exemplified in Expression (2) can also be calculated using dictionary data instead of translation conditions. That is, the number of dictionary data registered in the field correspondence information storage unit 12 is a, the number of fields is b, the i-th dictionary data (1 ≦ i ≦ a) and the j-th field (1 ≦ j ≦ b). The degree of fitness can be calculated with m (i, j) (m (i, j) ≧ 0) as the weighting coefficient between and the number of translations of the i-th dictionary data as x _i .

  For example, it is assumed that the conditional part of the translation rule “adjacent (federal, court)” of the translation rules in FIG. The action part of this translation rule is “translation (federal → federal)” and “translation (court → court)”. When the dictionary data of this action part is confirmed by the field classification tree of FIG. 10, “court → court” corresponds to “law”. In Equation (2), when the fitness is calculated by setting the weighting coefficient between the dictionary data having a correspondence relationship and the field as 1, for example, the fitness of the “law” field is 1, and the fitness of the other fields is 0. . As a result, the “law” field can be estimated as the field of the translation target sentence 2.

  As described above, in the third embodiment, the number of translations of dictionary data is used in place of the number of translation condition matches in the first and second embodiments, so that the first and second embodiments Similarly, field estimation can be performed. Therefore, also in the third embodiment, it becomes possible to select a field dictionary used for translation of a translation target sentence at high speed and with high accuracy.

  Next, a fourth embodiment of the present invention will be described.

  In the fourth embodiment, as the field correspondence information, the firing condition data of the neural network that outputs the fitness of each field when the number of times of translation of the dictionary data translated by the action part of the translation rule is input is used.

  The field dictionary management unit 15 generates firing condition data from the field dictionary in the dictionary storage unit 11. First, as in the second embodiment, a field classification tree is created from a field dictionary. Then, using the created field classification tree, a neural network is constructed in which the connection between the dictionary data (original word translation word pair) and the field is a neuron connection. An example of a neural network obtained from the field classification tree in FIG. 10 is shown in FIG. When the number of dictionary data translations is input, firing is performed according to the determined firing conditions, and the degree of conformity of each field is output. m (i, j) is a weighting coefficient between the i-th dictionary data and the j-th field. When the sum of the number of times of translation of dictionary data multiplied by the weighting coefficient exceeds a predetermined threshold, it is considered to have fired. When fired, the sum of the number of translations of the dictionary data multiplied by the weighting coefficient is output as the fitness. Here, for example, when the threshold value of the ignition condition is set to 0, the fitness is calculated by the expression (2) of the third embodiment. Firing condition data indicating the relationship between the number of translations of dictionary data and the fitness of each field of the neural network obtained in this way is held as the field correspondence information storage unit 12.

  The dictionary determination unit 13 performs field estimation using the firing condition data stored in the field correspondence information storage unit 12. A neural network is constructed from the firing condition data, the number of translations of the dictionary data applied to the translation target sentence 2 is input to the neural network, the degree of adaptation to each field is calculated, and the field of the translation target sentence 2 is estimated be able to. For example, it is assumed that “adjacent (federal, court)” is applied among the translation rules of FIG. The action part of this translation rule is “translation (federal → federal)” and “translation (court → court)”. When the dictionary data of this action part is input to the neural network of FIG. 11, “law” is estimated as a field because “court → court” is related to “law”.

  As described above, in the fourth embodiment, the calculation of the fitness similar to the linear function (equation (2)) regarding the number of translations in the third embodiment can be performed using a neural network. . In the description so far, the method for deriving the fitness of each field from the number of times the dictionary data is translated has been described. However, by replacing the dictionary data with the translation conditions, the translation can be performed as in the first and second embodiments. It is also possible to calculate the fitness of each field from the number of times the condition is met.

  As described above, in the fourth embodiment, as in the first, second, and third embodiments, the field can be estimated from the number of matching translation conditions, and the translation target sentence can be translated. The field dictionary to be used can be selected at high speed and with high accuracy.

  Although the machine translation device that estimates the field of the text to be translated and determines the field dictionary has been described above, it is possible to add a translation unit 16 and perform translation according to the determined field dictionary as shown in FIG. It is. Furthermore, as shown in FIG. It is also possible to add the communication unit 17 and input the translation target sentence 2 on the network and return the translation result 3 via communication. Furthermore, each part of the machine translation apparatus can be used by being distributed on the network.

  Further, although “adjacent” and “co-occurrence” of a predetermined word have been described as examples of translation conditions, not only words but “adjacent” or “co-occurrence” of sentences may be used as translation conditions. Here, the aforementioned words and sentences are also referred to as unit descriptions.

  The machine translation device, machine translation method, machine translation program, and recording medium of the present invention can be applied not only to natural language translation but also to kanji kana conversion and automatic application of field tags to sentences. .

  Each of the above-described embodiments is a preferred embodiment of the present invention, and various modifications can be made without departing from the scope of the present invention.

  A part or all of the above-described embodiment can be described as in the following supplementary notes, but is not limited thereto.

(Appendix 1)
A field correspondence information storage unit for storing field correspondence information indicating correspondence between translation conditions and fields;
A dictionary storage unit that stores a plurality of field dictionaries that hold translations according to the field;
Dictionary determination for determining the field dictionary to be used for translating the sentence to be translated from a plurality of the field dictionaries based on the correspondence relationship between the field and the translation condition that matches a predetermined unit description in the sentence to be translated And a machine translation device.

(Appendix 2)
The dictionary determination unit further calculates the degree of adaptation for each field based on the number of matches of the translation condition in the translation target sentence and the correspondence relationship with the field, and the field having a high degree of fitness The machine translation device according to appendix 1, wherein the field dictionary corresponding to is determined as the field dictionary used for translation of the translation target sentence.

(Appendix 3)
The machine translation device further includes:
A translation rule storage unit for storing a translation rule including the translation condition and the translation corresponding to the translation condition;
A field dictionary management unit that generates the field correspondence information from the translation rule and the field dictionary;
The dictionary determination unit further calculates, using the field correspondence information, the field corresponding to the translation condition that matches the predetermined unit description in the translation target sentence, and calculates the fitness.
The machine translation device according to appendix 2, wherein the degree of conformity increases as the number of matches of the translation condition corresponding to the field increases.

(Appendix 4)
The machine translation device according to appendix 3, wherein the fitness is further obtained by calculating a linear function related to the number of times the translation condition is matched.

(Appendix 5)
The fitness is further calculated by multiplying a weighting factor between the translation condition and the field by the number of matches of the translation condition for each field,
The weighting factor is a real value greater than or equal to 0, the weighting factor between the unrelated translation condition and the field is 0, and the weighting factor between the associated translation condition and the field The machine translation apparatus according to appendix 4, wherein is greater than 0.

(Appendix 6)
The machine translation device according to appendix 5, wherein a weighting factor between the translation condition and the field associated with each other is 1.

(Appendix 7)
The field dictionary management unit further sets the firing condition, which is the neural network firing condition as the calculation method of the fitness, as the field correspondence information,
The machine translation device according to any one of appendix 4 to appendix 6, wherein the dictionary determination unit further calculates the fitness by inputting the number of matches of the translation condition into a neural network of the firing condition.

(Appendix 8)
A field correspondence information storage unit that stores field correspondence information indicating correspondence between translated words and fields;
A dictionary storage unit that stores a plurality of field dictionaries that hold the translated words according to the field;
A translation rule storage unit that stores the translation rule including a translation condition and the translation corresponding to the translation condition;
The field dictionary used for translating the sentence to be translated from a plurality of the field dictionaries based on the correspondence relationship between the translation word corresponding to the translation condition and the field corresponding to the translation unit description in the sentence to be translated A machine translation apparatus comprising: a dictionary determination unit that determines

(Appendix 9)
The dictionary determination unit further determines the fitness for each field based on the correspondence relationship between the field and the number of translations of the translated words translated by the translation rule in which the translation condition is matched in the sentence to be translated. 9. The machine translation device according to appendix 8, wherein the field dictionary corresponding to the field having a high degree of fitness is calculated as the field dictionary to be used for translation of the translation target sentence.

(Appendix 10)
The machine translation device further includes:
A field dictionary management unit for generating the field correspondence information from the field dictionary,
The dictionary determining unit further uses the field correspondence information to check the field corresponding to the translated word translated by the translation rule whose translation condition matches in the translation target sentence, and calculates the fitness. ,
The machine translation apparatus according to appendix 9, wherein the fitness level increases as the number of translations of the translation corresponding to the field increases.

(Appendix 11)
The machine translation apparatus according to appendix 10, wherein the fitness is further obtained by calculating a linear function related to the number of translations of the translated word.

(Appendix 12)
The fitness is further calculated by multiplying, for each field, a product of a weighting factor between the translated word and the field and the number of translations of the translated word.
The weighting factor is a real value greater than or equal to 0, the weighting factor between the uncorresponding translation and the field is 0, and the weighting factor between the associated translation and the field is 0. The machine translation device according to attachment 11, wherein the machine translation device is larger.

(Appendix 13)
The machine translation apparatus according to appendix 12, wherein the weighting factor between the translated word and the field associated with each other is 1.

(Appendix 14)
The field dictionary management unit further sets the firing condition, which is the neural network firing condition as the calculation method of the fitness, as the field correspondence information,
The machine translation device according to appendix 11 to appendix 13, wherein the dictionary determination unit further calculates the fitness by inputting the number of translations of the translation into the neural network of the firing condition.

(Appendix 15)
The machine translation device further includes:
15. The machine translation device according to any one of supplementary notes 1 to 14, further comprising: a translation unit that performs translation using the field dictionary used for translation of the translation target sentence determined by the dictionary determination unit.

(Appendix 16)
The machine translation device further includes:
The machine translation apparatus according to claim 15, further comprising a data communication unit that receives the translation target sentence via a network and returns a translation result.

(Appendix 17)
The translation from a plurality of field dictionaries holding translation words corresponding to the field based on the field condition information indicating a translation condition that matches a predetermined unit description in the sentence to be translated and the translation condition and the field A machine translation method comprising: determining the field dictionary to be used for translation of a target sentence.

(Appendix 18)
The machine translation method further includes:
Based on the number of matches of the translation condition in the translation target sentence and the correspondence relationship with the field, the degree of matching for each field is calculated, and the field dictionary corresponding to the field with the high degree of matching is The machine translation method according to supplementary note 17, wherein the field dictionary is determined as the field dictionary used for translation of the translation target sentence.

(Appendix 19)
The machine translation method further includes:
From the translation rule including the translation condition and the translation corresponding to the translation condition, and the field dictionary, the field correspondence information in which the translation condition is associated with the field is generated, and the field correspondence information is used. , Examining the field corresponding to the translation condition that matches the predetermined unit description in the translation target sentence, and calculating the fitness, the fitness is the number of matches of the translation condition corresponding to the field The machine translation method according to appendix 18, wherein the machine translation method increases as the value increases.

(Appendix 20)
The machine translation method according to appendix 19, wherein the fitness is further obtained by calculating a linear function related to the number of matches of the translation condition.

(Appendix 21)
The fitness is further calculated by multiplying a weighting factor between the translation condition and the field by the number of matches of the translation condition for each field,
The weighting factor is a real value greater than or equal to 0, the weighting factor between the unrelated translation condition and the field is 0, and the weighting factor between the associated translation condition and the field The machine translation method according to appendix 20, wherein is greater than zero.

(Appendix 22)
The machine translation method according to appendix 21, wherein the weighting factor between the translation conditions associated with the field is 1.

(Appendix 23)
The machine translation method further includes:
Using the firing condition as the neural network firing condition as the neural network firing condition as the field correspondence information, and inputting the number of matches of the translation condition into the neural network of the firing condition to calculate the fitness. 23. The machine translation method according to supplementary note 20 to supplementary note 22, which is characterized.

(Appendix 24)
A translation rule including a translation condition and a translation corresponding to the translation condition, and field correspondence information indicating a correspondence relationship between the translation corresponding to the translation condition and the field corresponding to the predetermined unit description in the translation target sentence Based on the plurality of field dictionaries holding the translated words corresponding to the field, the field dictionary to be used for translation of the translation target sentence is determined.

(Appendix 25)
The machine translation method further includes:
Based on the number of translations of the translated words translated by the translation rule that matches the translation condition in the translation target sentence and the correspondence relationship with the field, the degree of fitness for each field is calculated, and the degree of matching is large. 25. The machine translation method according to appendix 24, wherein the field dictionary corresponding to the field is determined as the field dictionary used for translating the sentence to be translated.

(Appendix 26)
The machine translation method further includes:
The field correspondence information is generated from the field dictionary, and the field correspondence information is used to check the field corresponding to the translated word translated by the translation rule whose translation condition matches in the translation target sentence. 26. The machine translation method according to claim 25, wherein a degree of conformity is calculated, and the degree of conformity increases as the number of translations of the translation corresponding to the field increases.

(Appendix 27)
The machine translation method according to appendix 26, wherein the goodness of fit is further obtained by calculating a linear function related to the number of translations of the translated word.

(Appendix 28)
The fitness is further calculated by multiplying, for each field, a product of a weighting factor between the translated word and the field and the number of translations of the translated word.
The weighting factor is a real value greater than or equal to 0, the weighting factor between the uncorresponding translation and the field is 0, and the weighting factor between the associated translation and the field is 0. 28. The machine translation method according to appendix 27, which is larger.

(Appendix 29)
29. The machine translation method according to appendix 28, wherein the weighting factor between the translated word and the field associated with each other is 1.

(Appendix 30)
The machine translation method further includes:
Generating the firing condition as the field correspondence information using the method for calculating the fitness as a firing condition of a neural network, and calculating the fitness by inputting the number of translations of the translation into the neural network of the firing condition. 30. The machine translation method according to appendix 27 to appendix 29, wherein:

(Appendix 31)
The machine translation method further includes:
The machine translation method according to appendix 17 to appendix 30, wherein translation is performed using the field dictionary used for translating the determined translation target sentence.

(Appendix 32)
The machine translation method further includes:
32. The machine translation method according to appendix 31, wherein the translation target sentence is received via a network and a translation result is returned.

(Appendix 33)
On the computer,
A field correspondence information storage function for storing field correspondence information indicating correspondence between translation conditions and fields;
A dictionary storage function for storing a plurality of field dictionaries that hold translations according to the field;
Dictionary determination for determining the field dictionary to be used for translating the sentence to be translated from a plurality of the field dictionaries based on the correspondence relationship between the field and the translation condition that matches a predetermined unit description in the sentence to be translated Machine translation program for realizing functions and.

(Appendix 34)
The dictionary determination function further calculates the degree of adaptation for each field based on the number of matches of the translation condition in the sentence to be translated and the correspondence relationship with the field, and the field having a high degree of fitness 34. The machine translation program according to appendix 33, wherein the field dictionary corresponding to is determined as the field dictionary used for translation of the translation target sentence.

(Appendix 35)
The machine translation program is further stored on a computer.
A translation rule storage function for storing a translation rule including the translation condition and the translation corresponding to the translation condition;
A field dictionary management function for generating the field correspondence information from the translation conditions and the field dictionary;
Realized,
The dictionary determining function further calculates the fitness by examining the field corresponding to the translation condition that matches the predetermined unit description in the translation target sentence using the field correspondence information,
The machine translation program according to appendix 34, wherein the fitness level increases as the number of matches of the translation conditions corresponding to the field increases.

(Appendix 36)
36. The machine translation program according to appendix 35, wherein the fitness is further obtained by calculating a linear function related to the number of matches of the translation condition.

(Appendix 37)
The fitness is further calculated by multiplying a weighting factor between the translation condition and the field by the number of matches of the translation condition for each field,
The weighting factor is a real value greater than or equal to 0, the weighting factor between the unrelated translation condition and the field is 0, and the weighting factor between the associated translation condition and the field 37. The machine translation program according to appendix 36, wherein is greater than zero.

(Appendix 38)
38. The machine translation program according to appendix 37, wherein the weighting factor between the associated translation condition and the field is 1.

(Appendix 39)
The field dictionary management function further sets the firing condition with the method for calculating the fitness as the firing condition of the neural network as the field correspondence information,
The machine translation program according to appendix 36 to appendix 38, wherein the dictionary determination function further inputs the number of matches of the translation condition into a neural network of the firing condition and calculates the fitness.

(Appendix 40)
On the computer,
A field correspondence information storage function for storing field correspondence information indicating the correspondence between translated words and fields;
A dictionary storage function for storing a plurality of field dictionaries that hold the translated words according to the field;
A translation rule storage function for storing the translation rule including a translation condition and the translation corresponding to the translation condition;
The field dictionary used for translating the sentence to be translated from a plurality of the field dictionaries based on the correspondence relationship between the translation word corresponding to the translation condition and the field corresponding to the translation unit description in the sentence to be translated A machine translation program for realizing a dictionary determination function for determining

(Appendix 41)
The dictionary determination function is further configured to determine a degree of fitness for each field based on the correspondence between the number of translations of the translated word translated by the translation rule with the translation condition matched in the translation target sentence and the field. 41. The machine translation program according to appendix 40, wherein the field dictionary corresponding to the field having a high degree of fitness is calculated and used as the field dictionary used for translation of the translation target sentence.

(Appendix 42)
The machine translation program is further stored on a computer.
Realizing a field dictionary management function for generating the field correspondence information from the field dictionary,
The dictionary determination function further calculates, using the field correspondence information, the field corresponding to the translated word translated by the translation rule whose translation condition matches in the translation target sentence, and calculates the fitness. ,
The machine translation program according to appendix 41, wherein the fitness level increases as the number of translations of the translation corresponding to the field increases.

(Appendix 43)
The machine translation program according to appendix 42, wherein the fitness is further obtained by calculating a linear function related to the number of translations of the translated word.

(Appendix 44)
The fitness is further calculated by multiplying, for each field, a product of a weighting factor between the translated word and the field and the number of translations of the translated word.
The weighting factor is a real value greater than or equal to 0, the weighting factor between the uncorresponding translation and the field is 0, and the weighting factor between the associated translation and the field is 0. 44. The machine translation program according to appendix 43, which is larger.

(Appendix 45)
45. The machine translation program according to appendix 44, wherein the weighting factor between the translated word and the field associated with each other is 1.

(Appendix 46)
The field dictionary management function further generates, as the field correspondence information, the firing condition in which the calculation method of the fitness is a firing condition of a neural network,
46. The machine translation program according to appendix 43 to appendix 45, wherein the dictionary determination function further calculates the fitness by inputting the number of translations of the translation into the neural network of the firing condition.

(Appendix 47)
The machine translation program is further stored on a computer.
47. The machine translation program according to appendix 33 to appendix 46 for realizing a translation function for performing translation using the field dictionary used for translation of the translation target sentence determined by the dictionary determination unit.

(Appendix 48)
The machine translation program is further stored on a computer.
48. The machine translation program according to appendix 47 for realizing a data communication unit that receives the translation target sentence via a network and returns a translation result.

(Appendix 49)
A computer-readable recording medium on which the machine translation program according to appendix 33 to appendix 48 is recorded.

DESCRIPTION OF SYMBOLS 1 Machine translation apparatus 11 Dictionary memory | storage part 12 Field correspondence information storage part 13 Dictionary determination part 14 Translation rule memory | storage part 15 Field dictionary management part 16 Translation part 17 Communication part 2 Translation object sentence 3 Translation result 3 Translation result

Claims (10)

A field correspondence information storage unit for storing field correspondence information indicating correspondence between translation conditions and fields;
A dictionary storage unit that stores a plurality of field dictionaries that hold translations according to the field;
Dictionary determination for determining the field dictionary to be used for translating the sentence to be translated from a plurality of the field dictionaries based on the correspondence relationship between the field and the translation condition that matches a predetermined unit description in the sentence to be translated And a machine translation device. The dictionary determination unit further calculates the degree of adaptation for each field based on the number of matches of the translation condition in the translation target sentence and the correspondence relationship with the field, and the field having a high degree of fitness The machine translation device according to claim 1, wherein the field dictionary corresponding to is determined as the field dictionary used for translation of the translation target sentence. The machine translation device further includes:
A translation rule storage unit for storing a translation rule including the translation condition and the translation corresponding to the translation condition;
A field dictionary management unit that generates the field correspondence information from the translation rule and the field dictionary;
The dictionary determination unit further calculates, using the field correspondence information, the field corresponding to the translation condition that matches the predetermined unit description in the translation target sentence, and calculates the fitness.
The machine translation device according to claim 2, wherein the degree of conformity increases as the number of matches of the translation rule corresponding to the field increases. The machine translation device according to claim 3, wherein the fitness is further obtained by calculating a linear function related to the number of matches of the translation condition. A field correspondence information storage unit that stores field correspondence information indicating correspondence between translated words and fields;
A dictionary storage unit that stores a plurality of field dictionaries that hold the translated words according to the field;
A translation rule storage unit that stores the translation rule including a translation condition and the translation corresponding to the translation condition;
The field dictionary used for translating the sentence to be translated from a plurality of the field dictionaries based on the correspondence relationship between the translation word corresponding to the translation condition and the field corresponding to the translation unit description in the sentence to be translated A machine translation apparatus comprising: a dictionary determination unit that determines The translation from a plurality of field dictionaries holding translation words corresponding to the field based on the field condition information indicating a translation condition that matches a predetermined unit description in the sentence to be translated and the translation condition and the field A machine translation method comprising: determining the field dictionary to be used for translation of a target sentence. A translation rule including a translation condition and a translation corresponding to the translation condition, and field correspondence information indicating a correspondence relationship between the translation corresponding to the translation condition and the field corresponding to the predetermined unit description in the translation target sentence Based on the plurality of field dictionaries holding the translated words corresponding to the field, the field dictionary to be used for translation of the translation target sentence is determined. On the computer,
A field correspondence information storage function for storing field correspondence information indicating correspondence between translation conditions and fields;
A dictionary storage function for storing a plurality of field dictionaries that hold translations according to the field;
Dictionary determination for determining the field dictionary to be used for translating the sentence to be translated from a plurality of the field dictionaries based on the correspondence relationship between the field and the translation condition that matches a predetermined unit description in the sentence to be translated Machine translation program for realizing functions and. On the computer,
A field correspondence information storage function for storing field correspondence information indicating the correspondence between translated words and fields;
A dictionary storage function for storing a plurality of field dictionaries that hold the translated words according to the field;
A translation rule storage function for storing the translation rule including a translation condition and the translation corresponding to the translation condition;
The field dictionary used for translating the sentence to be translated from a plurality of the field dictionaries based on the correspondence relationship between the translation word corresponding to the translation condition and the field corresponding to the translation unit description in the sentence to be translated A machine translation program for realizing a dictionary determination function for determining   A computer-readable recording medium on which the machine translation program according to claim 8 or 9 is recorded.

Images