JP2008225744A - Machine translation device and program - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To more highly precisely achieve machine translation than a conventional manner. <P>SOLUTION: This machine translation device is configured to divide an accepted sentence into two or more words and phrases, and to acquire generalization information corresponding to one or more divided words and phrases, and to calculate the degree of similarity of a sentence using the acquired generalization information and one or more words and phrases among the two or more divided words and phrases and a sentence in each source language owned by the two or more sets of translation sentence sets, and to output one or more sentence in the source language similar to the accepted sentence, and to accept a permission instruction corresponding to one source language among one or more output sentences in the source language, and to acquire a sentence in target language corresponding to the source language corresponding to the permission instruction, and to acquire words and phrases in the target language making a pair with the words and phrases in the source language corresponding to the acquired generalization information, and to replace the words and phrases corresponding to the generalization information in the sentence in the acquired target language with the acquired words and phrases in the target language, and to configure the sentence in the target language, and to output the configured sentence in the target language. Thus, it is possible to highly precisely achieve machine translation. <P>COPYRIGHT: (C)2008,JPO&INPIT

Description

  The present invention relates to a machine translation apparatus that performs machine translation with high accuracy.

  Conventionally, there has been a technique for allowing a system user to confirm whether a translation result is good or bad by translating the translation result again in the reverse direction and presenting the result (see Non-Patent Document 1).

  Moreover, there is one method for realizing translation based on analogy, and in this method, a semantic dictionary and an edit distance are used for searching for similar parallel translation examples (see Non-Patent Document 2).

In addition, as a method for calculating the similarity between two sentences, there is a process by scoring that uses both the edit distance and tf / idf (see Non-Patent Document 3).
Uchimoto, K., Hayashida, N., Ishida, T., and Isahara, H., "Automatic rating of machine translatability," Proc. Of MT Summit X, pp. 235-242 (2005). Sumita, E., "Example-based machine translation using DP-matching between word sequences," Proc. Of ACL 2001 Workshop on Data-Driven Machine Translation, pp. 9-16 (2001). Watanabe, T. and Sumita, E., "Example-based decoding for statistical machine translation," Proc. Of MT Summit IX, pp. 410-417 (2003).

  However, the conventional method has a problem that it is difficult to guarantee the performance in a small unit such as a sentence or an utterance although it may be useful in an average tendency using a large number of data.

  The machine translation device of the first invention includes a phrase generalization information storage unit capable of storing one or more phrase generalization information having a phrase and generalization information that is generalized information of the phrase, a phrase in the source language A phrase set storage unit storing at least one phrase set having a pair of a target language phrase and a target sentence, and a translation sentence set having a source language sentence and a target language sentence as a pair. A translated sentence set storage unit that stores two or more sets of translated sentence sets having generalization information corresponding to some words and phrases of the source language sentence and the target language sentence; a sentence receiving unit that receives a sentence; A phrase division unit that divides the sentence into two or more words; a generalization information acquisition unit that acquires generalization information corresponding to one or more words divided by the phrase division unit from the phrase generalization information storage unit; The generalization information acquired by the generalization information acquisition unit and the phrase division unit Similarity calculation for calculating the similarity between a sentence using one or more of two or more divided phrases and sentences in each source language included in two or more translation sentence sets in the translation sentence set storage unit A similar sentence output unit that outputs one or more source language sentences similar to the sentence received by the sentence reception unit using the similarity calculated by the similarity calculation unit, and the similar sentence output unit Of the one or more source language sentences that have been output, the permission instruction reception unit that receives an input of a permission instruction that is an instruction indicating permission for a sentence in one source language, and the permission instruction received by the permission instruction reception unit A target language sentence acquisition unit that acquires a target language sentence corresponding to a source language sentence, and a target language phrase that is paired with a source language phrase corresponding to the generalization information acquired by the generalization information acquisition unit, A target language phrase acquisition unit acquired from the phrase set storage unit; The target language constituting the sentence of the target language to be output by replacing the phrase corresponding to the generalization information in the target language sentence acquired by the target language sentence acquisition unit with the target language phrase acquired by the target language phrase acquisition unit A machine translation apparatus including a sentence composing unit and an output unit that outputs a sentence in a target language configured by the target language sentence composing unit.

  With this configuration, highly accurate machine translation can be performed.

  The machine translation apparatus according to the second aspect of the invention also includes a phrase set storage unit that stores one or more sets of phrases each having a pair of a source language phrase and a target language phrase, a source language sentence, and a target language. Two or more sets of translation sentences having generalized information corresponding to a part of phrases of the source language sentence and the target language sentence included in the translation sentence set. A translated sentence set storage unit, a sentence receiving unit that receives a sentence, a phrase dividing unit that divides the sentence into two or more words and acquires generalization information corresponding to the words, and the phrase dividing unit Among the generalization information corresponding to two or more words and phrases, generalization information to be replaced with a phrase is determined and acquired, and the generalization information acquired by the generalization information acquisition section and the phrase division section Sentences using one or more words out of two or more words divided by , Using a similarity calculation unit that calculates a similarity with sentences in each source language included in two or more sets of translation sentences in the translation sentence set storage unit, and the similarity calculated by the similarity calculation unit, A similar sentence output unit that outputs one or more source language sentences similar to the sentence received by the sentence reception unit, and one source language sentence among the one or more source language sentences output by the similar sentence output unit A permission instruction receiving unit that receives an input of a permission instruction that is an instruction indicating permission to the target, and a target language sentence acquisition that acquires a sentence in a target language corresponding to a sentence in the source language corresponding to the permission instruction received by the permission instruction receiving unit A target language phrase acquisition unit that acquires from the phrase set storage unit a target language phrase that is paired with a source language phrase corresponding to the generalization information acquired by the generalization information acquisition unit, and the target language In the sentence of the target language acquired by the sentence acquisition unit A target language sentence constructing unit that configures a target language sentence to be output by replacing a phrase corresponding to generalization information with a target language phrase acquired by the target language phrase acquiring unit, and the target language sentence configuring unit The machine translation apparatus includes an output unit that outputs a sentence in the target language.

  With this configuration, it is not necessary to prepare word generalization information, and highly accurate machine translation can be performed.

  The machine translation device according to the third aspect of the invention relates to either the first or second invention, a voice receiving unit that receives voice, and voice recognition processing of the voice received by the voice receiving unit. The sentence receiving unit is a machine translation device that acquires a sentence formed by the sentence forming unit.

  With this configuration, it is possible to perform machine translation with high accuracy on a sentence input by voice.

  Further, in the machine translation device according to the fourth aspect of the present invention, in the first to third aspects of the invention, the similarity calculation unit is the number of generalization information acquired by the generalization information acquisition unit. Number of generalization information possessed by a source language sentence or a target language sentence included in two or more translation sentence sets of the first sentence generalization information acquisition means for acquiring the number of generalization information and the translation sentence set storage unit A second generalization information number acquisition means for acquiring the second generalization information number, and a similarity calculation target for selecting a source language sentence having a second generalization information number that matches the first generalization information number A sentence selection unit, a sentence using the generalization information acquired by the generalization information acquisition unit and one or more words out of two or more words divided by the word division unit, and the similarity calculation target sentence selection unit A machine translation apparatus comprising similarity calculation means for calculating similarity with one or more source language sentences selected by That.

  With this configuration, high-precision machine translation can be performed at high speed.

  The machine translation device according to the fifth aspect of the invention receives the sentence accepting unit when the permission instruction accepting unit does not accept the input of the permission instruction with respect to either the first or second invention. A re-input prompting unit that performs output for prompting re-input or editing of a sentence is further provided, and the sentence receiving unit is a machine translation device that receives a sentence again.

  With this configuration, machine translation with higher accuracy can be performed more reliably.

  Further, the machine translation apparatus of the sixth aspect of the invention provides the second invention with an output for prompting re-input of the voice when the permission instruction receiving unit does not accept the input of the permission instruction, and re-input The voice receiving unit is a machine translation device that receives voice again.

  With this configuration, it is possible to perform machine translation with high accuracy on a sentence that is input by voice.

  In the machine translation device of the seventh invention, the generalization information acquisition unit acquires generalization information for only a noun or adjective phrase in any one of the first to sixth inventions. Machine translation device.

  With this configuration, machine translation can be performed efficiently and with high accuracy.

  The machine translation apparatus according to the present invention can perform machine translation with high accuracy.

  Hereinafter, embodiments of a machine translation apparatus and the like will be described with reference to the drawings. In addition, since the component which attached | subjected the same code | symbol in embodiment performs the same operation | movement, description may be abbreviate | omitted again.

(Embodiment 1)
In the present embodiment, for example, the following machine translation apparatus will be described. In other words, this machine translation device accepts a sentence, divides it into words (morphemes, etc.), generalizes noun content words, etc., and uses similar parallel translation examples to obtain sentences containing information that generalizes similar expressions (noun content words). Etc.), the noun content word of the searched similar expression is replaced with the noun content word of the received sentence, and the replaced sentence is output. Then, when the machine translation device receives an OK instruction from the user, the machine translation device acquires a target language sentence corresponding to the similar expression, and uses the noun of the received sentence as a noun content word term of the target language sentence. Rewrite the target content word into a translated word to compose a sentence and output it.

  FIG. 1 is a block diagram of a machine translation apparatus according to the present embodiment.

  The machine translation apparatus includes a phrase generalization information storage unit 101, a phrase set storage unit 102, a translated sentence set storage unit 103, a sentence reception unit 104, a phrase division unit 105, a generalization information acquisition unit 106, a similarity calculation unit 107, a similarity A sentence output unit 108, a permission instruction accepting unit 109, a target language sentence acquiring unit 110, a target language phrase acquiring unit 111, a target language sentence constructing unit 112, an output unit 113, and a re-input prompting unit 114 are provided.

  The similarity calculation unit 107 includes a first generalization information number acquisition unit 1071, a second generalization information number acquisition unit 1072, a similarity calculation target sentence selection unit 1073, and a similarity calculation unit 1074.

  The phrase generalization information storage unit 101 can store one or more pieces of phrase generalization information. The phrase generalization information includes a phrase and generalization information. The generalization information is information obtained by generalizing the corresponding phrase. The unit of the phrase is preferably a morpheme, but may be a word or a phrase. The generalization information is, for example, “place name” for the phrase “Tokyo”. The generalization information is, for example, “person name” for the phrase “Natsume Soseki”. The generalization information is, for example, part of speech or segmentation information of part of speech (such as “place name” or “person name”). The phrase generalization information storage unit 101 is preferably a nonvolatile recording medium, but can also be realized by a volatile recording medium.

  The phrase set storage unit 102 stores one or more sets of phrase sets. The phrase set is information having a pair of a source language phrase and a target language phrase. That is, the phrase set storage unit 102 is a so-called term dictionary. The phrase set is, for example, “Tokyo, Tokyo”. The phrase set storage unit 102 is preferably a nonvolatile recording medium, but can also be realized by a volatile recording medium.

  The translated sentence set storage unit 103 stores two or more sets of translated sentence sets. The translation sentence set is information having a pair of a source language sentence and a target language sentence. Also, some of the phrases constituting the source language sentence and the target language sentence included in the translation sentence set usually have generalization information associated with each other. In other words, when the source language sentence is “I want to go to Kyoto”, for example, the sentence in the translation set is “Kyoto [place name] / to / go / toi / no / is / ga”. Further, for example, the sentence in the target language included in the translation sentence set is “I'd like to go to / Kyoto [place name]”. It is preferable that a sentence included in the translation sentence set includes information on a delimiter (here, “/”). Needless to say, the delimiter information is not limited to “/” but may be a space or “,”. The translated sentence set storage unit 103 is preferably a non-volatile recording medium, but can also be realized by a volatile recording medium.

  The sentence receiving unit 104 receives a sentence. In addition, the sentence receiving unit 104 may receive a sentence twice or more. The case where the sentence receiving unit 104 receives a sentence for the second time or later is a case where the permission instruction receiving unit 109 has not received a permission instruction. The case where the permission instruction is not received includes a case where a non-permission instruction is received, or a case where the instruction is not received for a predetermined time or more. The sentence input means may be anything such as a numeric keypad, keyboard, mouse or menu screen. The sentence receiving unit 104 can be realized by a device driver for input means such as a numeric keypad or a keyboard, control software for a menu screen, and the like.

  The phrase dividing unit 105 divides a sentence into two or more words. The phrase dividing unit 105 performs morphological analysis on the sentence to obtain two or more morphemes. Note that the phrase dividing unit 105 may divide the sentence into phrases by processing other than morphological analysis. The word division unit 105 can be realized by, for example, a morphological analyzer (ChaSen) URL <see http://chasen.naist.jp/hiki/ChaSen/>. Note that the morphological analyzer also has various processing methods, and there is no particular difference in the algorithm and processing results. The phrase dividing unit 105 can be usually realized by an MPU, a memory, or the like. The processing procedure of the phrase dividing unit 105 is usually realized by software, and the software is recorded on a recording medium such as a ROM. However, it may be realized by hardware (dedicated circuit).

  The generalization information acquisition unit 106 acquires generalization information corresponding to one or more words divided by the word division unit 105 from the word generalization information storage unit 101 and arranges it on the memory. The generalized information acquisition unit 106 can be usually realized by an MPU, a memory, or the like. The processing procedure of the generalized information acquisition unit 106 is usually realized by software, and the software is recorded on a recording medium such as a ROM. However, it may be realized by hardware (dedicated circuit).

  The similarity calculation unit 107 calculates the similarity between the first sentence and the second sentence, and arranges it on the memory. The first sentence is a sentence using the generalization information acquired by the generalization information acquisition unit 106 and one or more words out of two or more words divided by the word division unit 105. The first sentence is, for example, a sentence composed of two or more words divided by the word dividing unit 105 (for example, “Tokyo / To / Go / Tai / Nan / Is / ga”) as generalization information in the sentence. A sentence (for example, “[place name] / to / go / tai / n / is / ga”) in which a corresponding word (for example, “Tokyo”) is replaced with generalized information (eg, [place name]). The second sentence is a sentence in each source language included in two or more translation sentence sets in the translation sentence set storage unit 103. A sentence in the source language included in the translation sentence set usually also includes generalization information therein. Moreover, the algorithm which calculates the similarity degree of two sentences is not ask | required. The first example of an algorithm for calculating the similarity between two sentences is based on tf / idf. The second example is a method for calculating the similarity based on the edit distance. A third example is a method of calculating the similarity by using both the edit distance and tf / idf. Furthermore, the fourth example is a method in which the ratio of words that are simply matched is used as the similarity.

First, the first example will be described. The similarity (P _{tf / idf} (I _k , J ₀ )) based on tf / idf in the first example is calculated by the following formula 1.

In Equation 1, _{J 0} is the source language _{sentence, J 0, i} is the i-th morpheme of _{J 0, df (J 0,} i) document frequency of morpheme _{J 0, i (document frequency)} , N is the translation sentence This is the number of examples in the set storage unit 103 (the number of translation sets). The document frequency is, for example, the appearance frequency in the translated sentence set storage unit 103, the number of home pages in which morphemes J0 _{, i} appear on the Web, and the like.
If there is a target morpheme in J _k , the term frequency is 1 and if it is not 0. This score is normalized by the length (| J ₀ |) of the sentence J _{0 in} the source language.

Next, a second example will be described. The second example is a method of calculating the similarity by using the edit distance together. The edit distance (dis (J _k , J ₀ )) is expressed by Equation 2 below.

In Equation 2, k is a numerical value equal to or less than the number of sentences in the source language in the translated sentence set storage unit 103 for which the edit distance is calculated. I (J _k , J ₀ ), D (J _k , J ₀ ), and S (J _k , J ₀ ) are the numbers of insertion, deletion, and replacement errors, respectively. Note that the method for calculating the number of insertions, omissions, and replacement errors is a well-known technique, and thus detailed description thereof is omitted here.

Next, a third example will be described. The third example is a method of calculating the similarity (score) using the edit distance and tf / idf together. The third example is calculated using Equation 3 below.

In Formula 3, when dis (J _k , J ₀ ) is normalized, it may be performed on the sentence length | J ₀ | of the source language, but here, | J _k + J so as not to be negative. ₀ | is used. The value of the weight α is appropriately selected and stored so that “0 <α <1”.

Note that the similarity calculation unit 107 holds in advance formula information for calculating the similarity, such as Formula 1 to Formula 3, reads the formula information, and reads the necessary information (I (J _k , J ₀ ), etc.) are obtained and substituted into the mathematical formula to obtain the similarity, and the similarity is placed in the memory.

  Further, the similarity calculation unit 107 scores each sentence in the source language in the translated sentence set storage unit 103 by using, for example, tf / idf described above, and the similarity is calculated only for the top N sentences. May be calculated.

  The similarity calculation unit 107 can usually be realized by an MPU, a memory, or the like. The processing procedure of the similarity calculation unit 107 is usually realized by software, and the software is recorded on a recording medium such as a ROM. However, it may be realized by hardware (dedicated circuit).

  The first generalization information number acquisition unit 1071 acquires the first generalization information number that is the number of generalization information acquired by the generalization information acquisition unit 106 and arranges it on the memory. It can be said that the number of first generalization information is the number of words / phrases having corresponding generalization information among two or more words / phrases of the sentence received by the sentence reception unit 104. The first generalized information number acquisition unit 1071 can be usually realized by an MPU, a memory, or the like. The processing procedure of the first generalized information number acquiring unit 1071 is usually realized by software, and the software is recorded in a recording medium such as a ROM. However, it may be realized by hardware (dedicated circuit).

  The second generalization information number acquisition unit 1072 acquires the second generalization information number and arranges it on the memory. The number of second generalization information is the number of generalization information included in each source language sentence or each target language sentence included in two or more translation sentence sets in the translation sentence set storage unit 103. Usually, generalized information is included in the information of the source language sentence or the target language included in the translated sentence set, and the second generalized information number acquisition unit 1072 examines the generalized information, Get the number of second generalization information. The second generalized information number acquisition unit 1072 can be usually realized by an MPU, a memory, or the like. The processing procedure of the second generalized information number acquiring unit 1072 is usually realized by software, and the software is recorded on a recording medium such as a ROM. However, it may be realized by hardware (dedicated circuit).

  The similarity calculation target sentence selecting unit 1073 selects a sentence in the source language having the number of second generalization information that matches the number of first generalization information from the translation sentence set storage unit 103 and places it on the memory. The similarity calculation target sentence selection unit 1073 can be usually realized by an MPU, a memory, or the like. The processing procedure of the similarity calculation target sentence selecting means 1073 is usually realized by software, and the software is recorded on a recording medium such as a ROM. However, it may be realized by hardware (dedicated circuit).

  The similarity calculation unit 1074 selects a sentence that uses the generalization information acquired by the generalization information acquisition unit 106 and one or more words out of two or more words divided by the word division unit 105, and a similarity calculation target sentence selection The similarity with one or more source language sentences selected by the means 1073 is calculated and placed on the memory. The similarity calculation unit 1074 can be realized typically as an MPU, a memory, or the like. The processing procedure of the similarity calculation means 1074 is usually realized by software, and the software is recorded on a recording medium such as a ROM. However, it may be realized by hardware (dedicated circuit).

  The similar sentence output unit 108 outputs one or more source language sentences similar to the sentence received by the sentence receiving unit 104 using the similarity calculated by the similarity calculating unit 107. The similar sentence output unit 108 normally outputs a source language sentence corresponding to the maximum similarity. However, the similar sentence output unit 108 may output a source language sentence having the highest degree of similarity, or may output all source language sentences having a similarity degree equal to or higher than a predetermined value. Here, the output is a concept including display on a display and transmission to an external device (display device). The similar sentence output unit 108 may be considered as including or not including an output device such as a display. The similar sentence output unit 108 can be implemented by output device driver software, or output device driver software and an output device.

  The permission instruction receiving unit 109 receives an input of a permission instruction which is an instruction indicating permission for one source language sentence among one or more source language sentences output by the similar sentence output unit 108. The permission instruction input means may be anything such as a numeric keypad, a keyboard, a mouse, or a menu screen. The permission instruction accepting unit 109 can be realized by a device driver of an input unit such as a numeric keypad or a keyboard, menu screen control software, or the like.

  The target language sentence acquisition unit 110 acquires a source language sentence corresponding to the permission instruction received by the permission instruction reception unit 109, and sends a target language sentence corresponding to the source language sentence from the translated sentence set storage unit 103. Obtain and place on memory. The target language sentence acquisition unit 110 can usually be realized by an MPU, a memory, or the like. The processing procedure of the target language sentence acquisition unit 110 is usually realized by software, and the software is recorded in a recording medium such as a ROM. However, it may be realized by hardware (dedicated circuit).

  The target language phrase acquisition unit 111 sets a phrase in the target language that is paired with a source language phrase (a phrase in the accepted source language sentence) corresponding to the generalization information acquired by the generalization information acquisition unit 106. Obtained from the set storage unit 102. The target language phrase acquisition unit 111 can usually be realized by an MPU, a memory, or the like. The processing procedure of the target language phrase acquisition unit 111 is usually realized by software, and the software is recorded on a recording medium such as a ROM. However, it may be realized by hardware (dedicated circuit).

  The target language sentence constructing unit 112 replaces the phrase corresponding to the generalization information in the sentence of the target language acquired by the target language sentence acquiring unit 110 with the phrase of the target language acquired by the target language phrase acquiring unit 111 and outputs the phrase. Construct a sentence in the target language. The target language sentence constructing unit 112 can usually be realized by an MPU, a memory, or the like. The processing procedure of the target language sentence constructing unit 112 is usually realized by software, and the software is recorded in a recording medium such as a ROM. However, it may be realized by hardware (dedicated circuit).

  The output unit 113 outputs a sentence in the target language configured by the target language sentence constructing unit 112. Here, the output is a concept including display on a display, transmission to an external device (display device), and the like. The output unit 113 may be considered as including or not including an output device such as a display. The output unit may be realized by output device driver software, or output device driver software and an output device.

  The re-input prompting unit 114 performs output for prompting re-input or editing of the sentence received by the sentence receiving unit 104. It is assumed that the user re-inputs or edits a sentence by using an input unit based on an output prompting the user to re-input or edit the sentence. Then, the sentence receiving unit 104 receives a sentence re-input or edited by the user. Further, the content of the output prompting re-input or editing is not limited. The re-input prompting unit 114 can usually be realized by an MPU, a memory, or the like. The processing procedure of the re-input prompting unit 114 is usually realized by software, and the software is recorded on a recording medium such as a ROM. However, it may be realized by hardware (dedicated circuit).

  Next, the operation of the machine translation apparatus will be described with reference to the flowcharts of FIGS.

  (Step S201) The sentence receiving unit 104 determines whether a sentence has been received. If a sentence is accepted, the process goes to step S202, and if no sentence is accepted, the process returns to step S201.

  (Step S202) The phrase dividing unit 105 divides the sentence received in step S201 into phrases. For example, the phrase dividing unit 105 divides the sentence received in step S201 into a plurality of morphemes. Here, it is assumed that the phrase dividing unit 105 obtains part-of-speech information for each phrase and information about whether it is a noun content word. Note that the noun content words are, for example, nouns and adjectives. The noun content word may contain other parts of speech, but preferably no verbs. For example, a known morphological analyzer can not only divide a sentence into morphemes but also acquire part-of-speech information and the like, and thus detailed description thereof is omitted.

  (Step S203) The generalization information acquisition unit 106 substitutes 1 for a counter i.

  (Step S204) The generalization information acquisition unit 106 determines whether or not the i-th word is present in the word obtained in step S202. If the i-th word / phrase exists, the process goes to step S205, and if the i-th word / phrase does not exist, the process goes to step S210.

  (Step S205) The generalization information acquisition unit 106 determines whether or not the i-th word is a noun content word. If the i-th phrase is a noun content word, go to step S206, and if the i-th phrase is not a noun content word, go to step S209.

  (Step S206) The generalization information acquisition unit 106 acquires generalization information corresponding to the word from the word generalization information storage unit 101 using the i-th word as a key.

  (Step S207) The generalization information acquisition unit 106 determines whether generalization information has been acquired in step S206. If generalization information can be acquired, it will go to step S208, and if generalization information cannot be acquired, it will go to step S209.

  (Step S208) The generalization information acquisition unit 106 inserts the generalization information acquired in step S206 at a location corresponding to the i-th word / phrase. At this time, the i-th word is separately stored in the memory.

  (Step S209) The generalization information acquisition unit 106 increments the counter i by 1. The process returns to step S204.

  (Step S210) The similarity calculation unit 107 selects a source language sentence satisfying a predetermined condition from one or more source language sentences held in each translation set of the translation sentence set storage unit 103. Here, the predetermined condition is, for example, a sentence in which the above-described score of tf / idf is the top N (N is a natural number of 2 or more). The process of selecting the source language sentence is also referred to as preliminary selection.

  (Step S211) The similarity calculation unit 107 substitutes 1 for a counter i.

  (Step S212) The similarity calculation unit 107 determines whether or not the i-th sentence is present among the sentences selected in step S210. If the i-th sentence exists, go to step S213, and if the i-th sentence does not exist, go to step S215.

  (Step S213) The similarity calculation unit 107 calculates the similarity between the i-th sentence and the source language sentence (including generalization information) obtained through the processing up to step S208. The i-th sentence is usually a sentence including generalization information. The similarity calculation process will be described with reference to the flowchart of FIG.

  (Step S214) The similarity calculation unit 107 increments the counter i by 1. Go to step S212.

  (Step S215) The similar sentence output unit 108 acquires a sentence using the similarity calculated by the similarity calculation unit 107. For example, the similar sentence output unit 108 acquires a sentence having the highest similarity. The similar sentence output unit 108 may acquire the top three sentences having the highest similarity.

  (Step S216) The similar sentence output unit 108 outputs the sentence acquired in step S215.

  (Step S217) The permission instruction reception unit 109 determines whether or not an input of a permission instruction that is an instruction indicating permission for one source language sentence among the one or more source language sentences output in Step S216 is received. to decide. If an input of a permission instruction is accepted, the process goes to step S218. If an input of a permission instruction is not accepted, the process goes to step S219.

  (Step S218) The target language sentence acquisition unit 110 or the like performs processing for outputting a translated sentence. The process ends. The translated sentence output process will be described with reference to the flowchart of FIG.

  (Step S219) The re-input prompting unit 114 performs output for prompting re-input or editing of the sentence received by the sentence receiving unit 104. Return to step S201

  In the flowchart of FIG. 3, the translation set preliminary selection process in step S210 may be omitted. Further, the determination in step S205 may not be performed.

  Next, the similarity calculation process in step S213 will be described with reference to the flowchart of FIG.

  (Step S301) The first generalization information number acquisition unit 1071 of the similarity calculation unit 107 acquires the first generalization information number that is the number of generalization information acquired by the generalization information acquisition unit 106, and stores it in the memory. Deploy.

  (Step S302) The second generalization information number acquisition unit 1072 acquires the number of second generalization information, which is the number of generalization information included in the i-th source language sentence, and arranges it on the memory.

  (Step S303) The similarity calculation target sentence selection unit 1073 determines whether or not the number of first generalized information acquired in step S301 matches the number of second generalized information acquired in step S302. If they match, the process goes to step S304, and if they do not match, the process returns to the upper process.

  (Step S304) The similarity calculation means 1074 calculates the tf / idf score of the i-th source language sentence (usually including generalization information) and the input source language sentence (including generalization information). Obtained by the above-described calculation formula and method.

  (Step S305) The similarity calculation means 1074 calculates the edit distance between the i-th source language sentence (usually including generalization information) and the input source language sentence (including generalization information) as described above. Obtained by formula and method.

  (Step S306) The similarity calculation unit 1074 reads information on an expression for calculating the similarity (for example, Expression 3), and uses the tf / idf score acquired in Step S304 and the edit distance acquired in Step S305. The similarity (score) is calculated.

  (Step S307) The similarity calculation means 1074 associates the similarity (score) calculated in step S306 with the i-th source language sentence (the sentence selected in the sentence set storage unit 103). And temporarily store it in memory. The similarity calculating unit 1074 associates the similarity (score) calculated in step S306 with a sentence in the target language corresponding to the i-th source language sentence (a sentence in the translated sentence set storage unit 103). It may be temporarily stored in the memory. Return to upper process.

  Needless to say, the method of calculating similarity in the flowchart of FIG. 3 is an example.

  Next, the translated sentence output process of step S218 will be described with reference to the flowchart of FIG.

  (Step S401) The target language sentence acquisition unit 110 acquires a source language sentence corresponding to the permission instruction received by the permission instruction reception unit 109, and converts the target language sentence corresponding to the source language sentence into a translated sentence set. Obtained from the storage unit 103 and placed on the memory.

  (Step S402) The target language phrase acquisition unit 111 assigns 1 to the counter i.

  (Step S403) The target language phrase acquisition unit 111 determines whether or not there is a phrase to be replaced with the i-th generalization information in the phrases (morphemes, words, phrases, etc.) constituting the sentence acquired in Step S401. Judging. If there is a phrase to be replaced with the i-th generalization information, the process goes to step S404, and if there is no phrase to be replaced with the i-th generalization information, the process goes to step S408.

  (Step S404) The target language phrase acquisition unit 111 is a phrase corresponding to the phrase to be replaced with the i-th generalization information, acquires the phrase in the sentence of the source language that has received the input, and places it on the memory. .

  (Step S405) The target language phrase acquisition unit 111 acquires the target language phrase corresponding to the phrase in the source language sentence acquired in step S404 from the phrase set storage unit 102, and arranges it on the memory.

  (Step S406) The target language sentence constructing unit 112 replaces the phrase corresponding to the generalization information in the target language sentence acquired in step S401 with the target language phrase acquired in step S405.

  (Step S407) The target language phrase acquisition unit 111 increments the counter i by one.

  (Step S408) The output unit 113 outputs the sentence in the target language configured by the target language sentence constructing unit 112 in step S406. Return to upper process.

  Hereinafter, a specific operation of the machine translation apparatus in the present embodiment will be described. The machine translation device here is a machine translation device that performs Japanese-English translation.

  The phrase generalization information storage unit 101 holds the phrase generalization information management table shown in FIG. The phrase generalization information management table stores one or more phrase generalization information. The phrase generalization information includes “phrase” and “generalization information”.

  The phrase set storage unit 102 holds a phrase set management table shown in FIG. The phrase set management table stores one or more phrase sets. The phrase set has “source language” and “target language”. The “source language” is a phrase of the source language (here, Japanese). The “target language” is a phrase of the target language (here, English).

  The translated sentence set storage unit 103 holds a translated sentence set management table shown in FIG. The translation sentence set management table stores one or more translation sentence sets. The translation sentence set includes “source language sentence” and “target language sentence”. The “source language sentence” is a sentence in the source language (here, Japanese). The “target language sentence” is a sentence of a target language (here, English). Generalized information is associated with some words in each sentence in the translated sentence set (such as “place name” for “Tokyo”).

  In such a situation, it is assumed that the user inputs a colloquial sentence such as “I want to go to Tokyo” using a keyboard or the like.

  Next, the sentence reception unit 104 receives the sentence “I want to go to Tokyo”. The phrase dividing unit 105 divides the received sentence “I want to go to Tokyo” into morphemes, and obtains “Tokyo / Ni / Go / Tai / Nan / Is / Ga”. The morpheme separator “/” may be another character code (for example, a space or “,”). Here, the phrase dividing unit 105 also outputs part-of-speech information for each morpheme. That is, “Tokyo” is output as “noun”.

  Next, the generalization information acquisition unit 106 reads the part of speech “noun” and “adjective” corresponding to the noun content words stored in advance, and acquires the first word “Tokyo” corresponding to “noun”. The phrase “Tokyo” is a noun content word.

  Then, the generalization information acquisition unit 106 searches the word generalization information management table in FIG. 5 using “Tokyo” as a key, acquires generalization information [location name] corresponding to the word, and arranges it on the memory. .

  Then, the generalization information acquisition unit 106 inserts the acquired generalization information [place name] into the location corresponding to the first word “Tokyo” and reads “[place name] / ni / go / tai / n / Get ".

  And since there are no noun content words (here, nouns or adjectives) other than “Tokyo” in “Tokyo / ni / go / tai / n / is / ga”, the generalized information acquisition unit 106 Then, the process of acquiring generalization information is terminated. Note that the generalization information acquisition unit 106 separately stores the phrase “Tokyo” corresponding to [place name] on the memory.

  Next, the similarity calculation unit 107 calculates the tf / idf between the source language sentence “[place name] / ni / go / tai / n / d / ga” and each sentence of the translation sentence set management table of FIG. Calculate the score. FIG. 8 shows, for example, sentences in the top three languages (where N is “3”) having the highest tf / idf score in the translated sentence set management table. Note that the tf / idf score calculation method is well-known and is omitted here. FIG. 8 shows a set of pre-selected source language sentences.

  Next, the similarity calculation unit 107 calculates the similarity to the sentence “[place name] / ni / go / tai / n / d / ga” for each of the three sentences in FIG. To do.

  First, the first generalization information number acquisition unit 1071 of the similarity calculation unit 107 includes first generalization information that is the number of generalization information of the sentence “[place name] / ni / go / tai / n / is / ga”. The number “1” is acquired and placed on the memory.

  Then, the second generalization information number acquisition unit 1072 acquires the second generalization information number “1” which is the number of generalization information included in the first source language sentence of FIG. 8 and arranges it on the memory. .

  The similarity calculation target sentence selection unit 1073 determines that the acquired first generalization information number “1” matches the acquired second generalization information number “1”, and the similarity calculation unit 1074 determines that 1 The tf / idf score of the sentence [[place name] / ni / go / tai / no / is / ga] and the sentence [[place name] / ni / go / tai / n / da / ga] in the second source language, Obtained by the above-described calculation formula and method. Here, the score of tf / idf that has already been acquired may be read.

  Then, the similarity calculation means 1074 inputs the sentence “[place name] / ni / go //” of the source language in which the first original sentence “[place name] / ni / go / tai / no / is / ga” is input. The edit distance of “Tai / Nan / Isuda / Ga” is acquired by the above-described calculation formula and method.

  Next, the similarity calculation unit 1074 reads information on an expression for calculating the similarity (for example, Expression 3), and calculates the similarity (score) using the acquired tf / idf score and the acquired editing distance. To do. Here, it is assumed that the similarity calculation unit 1074 can calculate, for example, “0.9” as the similarity of the sentence “[place name] / ni / go / tai ////” in the source language. (Is this value strange?)

  Next, the second generalization information number acquisition unit 1072 acquires the second generalization information number “2”, which is the number of generalization information included in the second source language sentence in FIG. To do. Then, the similarity calculation target sentence selection unit 1073 determines that the acquired first generalization information number “1” and the acquired second generalization information number “2” do not match, and processes the third sentence Migrate to

  Then, the second generalization information number acquisition unit 1072 acquires the second generalization information number “1” that is the number of generalization information included in the third source language sentence of FIG. 8 and arranges it on the memory. .

  The similarity calculation target sentence selection unit 1073 determines that the acquired first generalization information number “1” matches the acquired second generalization information number “1”. The tf / idf score of the sentence [[place name] / to / coming / tai / no / is / ne] and the sentence [[place name] / ni / go / tai / n / de / ga] Obtained by the above-described calculation formula and method. Here, the score of tf / idf that has already been acquired may be read.

  Then, the similarity calculation means 1074 inputs the source language sentence “[place name] / ni / go //” where the first source language sentence “[place name] / to / coming / want / no / is / ne” is input. The edit distance of “Tai / Nan / Isuda / Ga” is acquired by the above-described calculation formula and method.

  Next, the similarity calculation unit 1074 reads information on an expression for calculating the similarity (for example, Expression 3), and calculates the similarity (score) using the acquired tf / idf score and the acquired editing distance. To do. Here, it is assumed that the similarity calculation unit 1074 can calculate, for example, “0.55” as the similarity of the sentence “[place name] / to / coming / want / no / is / ne” in the source language.

  Then, the similar sentence output unit 108 uses the similarity (“0.9” or “0.55”) calculated by the similarity calculation unit 107 to set “0.9” having the highest similarity. The corresponding source language sentence “[place name] / ni / go / tai / no / is / ga” is acquired and placed in the memory.

  Next, the similar sentence output unit 108 changes the generalization information [place name] of the acquired sentence “[place name] / ni / go / tai / no / de / ga” to the original phrase “Tokyo”, Get “Tokyo / Ni / Go / Tai / No / Is / Ga”. The original phrase “Tokyo” is arranged on the memory when the generalization information is acquired, and the phrase “Tokyo” on the memory is read out.

  Then, the similar sentence output unit 108 outputs the sentence “I want to go to Tokyo” on the display. Here, as shown in FIG. 9, the similar sentence output unit 108 displays not only the sentence “I want to go to Tokyo” but also a button (“YES” button) that prompts the user to input a permission instruction and a non-permission instruction. A screen having a “NO” button is also output.

  Next, since the sentence “I want to go to Tokyo” that is close to the sentence “I want to go to Tokyo” is displayed, the user clicks the “YES” button with an input means such as a mouse. Is pressed.

  Next, when the “YES” button is pressed, the permission instruction receiving unit 109 receives an input of a permission instruction.

  Then, the target language sentence acquisition unit 110 acquires the source language sentence “I want to go to Tokyo” corresponding to the permission instruction received by the permission instruction reception unit 109. Then, the target language sentence acquisition unit 110 reads the sentence “I'd like to go to / Kyoto” in the target language corresponding to the sentence “I want to go to Tokyo” in the source language (the sentence with “ID = 2”). [Place name] ”is acquired from the translated sentence set management table of FIG. 7 and placed on the memory.

  Next, the target language phrase acquisition unit 111 acquires the phrase “Tokyo” in the sentence of the source language that is the input corresponding to the phrase that is replaced with the first generalization information [place name]. Note that the phrase “Tokyo” is arranged on the memory. Then, the target language phrase acquisition unit 111 acquires the phrase “Tokyo” in the target language corresponding to the phrase “Tokyo” in the acquired source language sentence from the phrase set management table of FIG. 6 and places it on the memory. To do.

  And the target language sentence structure part 112 acquired the phrase (Kyoto [place name]) corresponding to the generalization information in the sentence "I'd like to go to / Kyoto [place name]" of the acquired target language. Replace with the phrase “Tokyo” in the target language to obtain “I'd like to go to / Tokyo”. Then, the target language sentence constructing unit 112 deletes the delimiter “/” and obtains the sentence “I'd like to go to Tokyo”.

  Then, the output unit 113 outputs the sentence “I'd like to go to Tokyo” in the target language configured by the target language sentence composing unit 112. With the above processing, the machine translation process is completed.

  In this specific example, in the display of FIG. 9, for example, when the user presses the “NO” button or does not input for a certain period of time, as shown in FIG. Present a screen that prompts the user to re-enter the sentence. In this case, the sentence previously input by the user is displayed in an editable form.

  When the user edits the sentence and presses the “OK” button, the above translation processing is performed again. The process of the re-input prompting unit 114 may be performed twice or more.

  As described above, according to the present embodiment, by selecting a sentence similar to the input sentence, presenting it to the user, and performing translation using the translated sentence corresponding to the sentence selected by the user, Machine translation can be performed with high accuracy.

  Note that according to the present embodiment, generalization information is acquired using the phrase generalization information stored in the phrase generalization information storage unit. However, when the phrase dividing unit divides a sentence into two or more phrases, generalization information (usually part of speech) corresponding to each phrase may be acquired. In such a case, the phrase generalization information storage unit becomes unnecessary. In such a case, the machine translation device is configured to store a phrase set storage unit that stores one or more sets of phrases having a pair of a source language phrase and a target language phrase, and a source language sentence and a target language sentence. A translation sentence set that includes two or more sets of translation sentence sets having generalization information corresponding to some words and phrases of the source language sentence and the target language sentence of the translation sentence set A sentence set storage unit; a sentence receiving unit that receives a sentence; a phrase dividing unit that divides the sentence into two or more words and acquires generalized information (usually parts of speech) corresponding to the words; and the phrase dividing unit Among generalization information corresponding to two or more words divided by the generalization information (for example, nouns and adjectives) to be replaced with words, and obtaining the generalization information acquisition unit, and the generalization information acquisition unit The acquired generalization information and the phrase dividing unit divided A similarity calculation unit that calculates a similarity between a sentence using one or more of the above phrases and a sentence in each source language included in two or more sets of translation sentences in the translation sentence set storage unit; Using the similarity calculated by the similarity calculator, a similar sentence output unit that outputs one or more source language sentences similar to the sentence received by the sentence receiver, and the similar sentence output unit outputs 1 Among the above source language sentences, a permission instruction receiving unit that receives an input of a permission instruction that is an instruction indicating permission for one source language sentence, and a source language corresponding to the permission instruction received by the permission instruction receiving unit A target language sentence acquisition unit that acquires a sentence in a target language corresponding to a sentence, and a phrase in the target language that is paired with a source language phrase that corresponds to the generalization information acquired by the generalization information acquisition unit. A target language phrase acquisition unit acquired from a storage unit, and the target language In the target language sentence acquired by the sentence acquisition unit, the phrase corresponding to the generalization information is replaced with the target language phrase acquired by the target language phrase acquisition unit, and the target language sentence configuration that configures the output target language sentence And a machine translation apparatus comprising an output unit that outputs a sentence in a target language configured by the target language sentence constructing unit.

  Further, according to the present embodiment, there are many algorithms for calculating the similarity between two sentences, and there is no limitation. This also applies to other embodiments in this specification.

  Further, according to the present embodiment, the source language is Japanese and the target language is English. However, it goes without saying that the source language and the target language are not limited. This also applies to other embodiments in this specification.

  Furthermore, the processing in the present embodiment may be realized by software. Then, this software may be distributed by software download or the like. Further, this software may be recorded and distributed on a recording medium such as a CD-ROM. This also applies to other embodiments in this specification. The software that realizes the machine translation apparatus according to the present embodiment is the following program. In other words, the program includes a computer that receives a sentence receiving unit that receives a sentence, a phrase dividing unit that divides the sentence into two or more words, and generalization information corresponding to one or more words that the word dividing unit divides. A generalization information acquisition unit acquired from one or more word generalization information stored in a storage medium, generalization information acquired by the generalization information acquisition unit, and two or more words divided by the word division unit A similarity calculation unit for calculating a similarity between a sentence using one or more of the phrases and a sentence in each source language included in two or more translation sentence sets stored in a storage medium; and the similarity calculation A similar sentence output unit that outputs one or more source language sentences similar to the sentence received by the sentence receiving unit using the similarity calculated by the unit, and one or more source languages output by the similar sentence output unit Is an instruction indicating permission for a sentence in one source language. A permission instruction receiving unit that receives an input of a permission instruction, a target language sentence acquiring unit that acquires a sentence in a target language corresponding to a sentence in a source language corresponding to the permission instruction received by the permission instruction receiving unit, and the generalization information A target language phrase acquisition unit that acquires, from one or more word / phrase sets stored in a storage medium, a target language phrase that is paired with a source language phrase corresponding to the generalization information acquired by the acquisition unit; In the target language sentence acquired by the sentence acquisition unit, the phrase corresponding to the generalization information is replaced with the target language phrase acquired by the target language phrase acquisition unit, and the target language sentence configuration that configures the output target language sentence And a program for causing the target language sentence configuration unit to function as an output unit that outputs a sentence in the target language.

  In addition, the program includes a sentence receiving unit that receives a sentence, a phrase dividing unit that divides the sentence into two or more words and acquires generalization information corresponding to the words, and the word dividing unit divides the computer Among the generalization information corresponding to two or more words and phrases, generalization information to be replaced with a phrase is determined and acquired, and the generalization information acquired by the generalization information acquisition section and the phrase division section Similarity that calculates the similarity between sentences that use one or more of two or more words divided by, and sentences of each source language that two or more translation sentences stored in the storage medium have Using the similarity calculated by the similarity calculator, the similar sentence output unit that outputs one or more source language sentences similar to the sentence received by the sentence receiver, and the similar sentence output unit One source language sentence out of one or more source language sentences output by A permission instruction receiving unit that receives an input of a permission instruction that is an instruction indicating permission to the target, and a target language sentence acquisition that acquires a sentence in a target language corresponding to a sentence in a source language corresponding to the permission instruction received by the permission instruction receiving unit Language and a target language phrase that is acquired from one or more phrase sets stored in a storage medium, and a target language phrase that is paired with a source language phrase corresponding to the generalization information acquired by the generalization information acquisition unit In the target language sentence acquired by the acquisition unit and the target language sentence acquisition unit, the phrase corresponding to the generalization information is replaced with the target language phrase acquired by the target language phrase acquisition unit, and the target language sentence to be output And a program for functioning as an output unit that outputs a sentence in the target language configured by the target language sentence configuration unit.

  Further, in the program, the similarity calculation unit includes a first generalization information number acquisition unit that acquires a first generalization information number that is the number of generalization information acquired by the generalization information acquisition unit, and the translation Second generalization information number acquisition means for acquiring the number of second generalization information, which is the number of generalization information included in a source language sentence or a target language sentence included in two or more sets of translated sentence sets in a sentence set storage unit; , A similarity calculation target sentence selection unit that selects a sentence in a source language having a second generalization information number that matches the first generalization information number, the generalization information acquired by the generalization information acquisition unit, and the phrase Similarity calculation for calculating the similarity between a sentence using one or more of two or more words divided by the dividing unit and one or more source language sentences selected by the similarity calculation target sentence selecting means. It is preferable that the program functions as a means.

  Further, in the above program, when the permission instruction receiving unit does not receive the input of the permission instruction, the computer is further provided as a re-input prompting unit that performs an output for prompting re-input or editing of the sentence received by the sentence receiving unit. It is preferable that the function is made to function, and the sentence receiving unit functions to accept a sentence again.

    In the above program, it is preferable that the generalization information acquisition unit functions so as to acquire generalization information for only a noun or adjective phrase.

(Embodiment 2)
In the present embodiment, the sentence input is a speech input, and after performing speech recognition processing, machine translation is performed by the same algorithm as in the first embodiment.

  FIG. 11 is a block diagram of the machine translation apparatus in the present embodiment.

  The machine translation apparatus includes a phrase generalization information storage unit 101, a phrase set storage unit 102, a translated sentence set storage unit 103, a speech reception unit 1101, a sentence composition unit 1102, a sentence reception unit 1104, a phrase division unit 105, and a generalization information acquisition. Unit 106, similarity calculation unit 107, similar sentence output unit 108, permission instruction reception unit 109, target language sentence acquisition unit 110, target language phrase acquisition unit 111, target language sentence composition unit 112, output unit 113, re-input prompting unit 114.

  The voice reception unit 1101 receives voice. The voice reception unit 1101 may receive voice again. The voice input means is, for example, a microphone, but may be read from a storage medium. The voice input means may be anything such as a microphone 305 or a reading means from a recording medium. The voice reception unit 1101 can be realized by a device driver or the like of input means such as the microphone 305.

  The sentence constructing unit 1102 performs speech recognition processing on the speech accepted by the speech accepting unit 1101, and composes a sentence. The algorithm for the speech recognition process is not limited. Since the voice recognition process is a known technique, detailed description thereof is omitted. The sentence constructing unit 1102 can usually be realized by an MPU, a memory, or the like. The processing procedure of the sentence composition unit 1102 is usually realized by software, and the software is recorded in a recording medium such as a ROM. However, it may be realized by hardware (dedicated circuit).

  The sentence reception unit 1104 receives a sentence. This acceptance is acquisition of the sentence which the sentence structure part 1102 comprised. The sentence reception unit 1104 serves to pass the sentence formed by the sentence composition unit 1102 to the phrase dividing unit 105. The sentence receiving unit 1104 is, for example, executing software that implements the phrase dividing unit 105 or calling a function that implements the phrase dividing unit 105. The sentence reception unit 1104 can be realized by an MPU, a memory, software, and the like.

  Next, the operation of the machine translation apparatus will be described. This machine translation device is different from the machine translation device of the first embodiment in that a sentence received from a user is received by voice. That is, the speech receiving unit 1101 of the machine translation apparatus receives a sentence by speech, and the sentence constructing unit 1102 performs a speech recognition process, acquires a character string that constitutes a sentence to be translated, and places it on a memory. Then, the sentence receiving unit 1104 passes the sentence obtained by the sentence constructing unit 1102 to the phrase dividing unit 105. The subsequent translation processing is the same as that of the machine translation apparatus of the first embodiment.

  Further, the re-input prompting unit 114 normally prompts the user to re-input a sentence by voice. Here, “prompt the user to re-enter a sentence” means, for example, that a sentence “Please speak to the microphone again” is displayed on the display or output to the speaker.

  As described above, according to the present embodiment, a sentence uttered by a user can be machine translated with extremely high accuracy.

  The software that realizes the machine translation apparatus according to the present embodiment is the following program. In other words, the program includes a computer that receives a sentence receiving unit that receives a sentence, a phrase dividing unit that divides the sentence into two or more words, and generalization information corresponding to one or more words that the word dividing unit divides. A generalization information acquisition unit acquired from one or more word generalization information stored in a storage medium, generalization information acquired by the generalization information acquisition unit, and two or more words divided by the word division unit A similarity calculation unit for calculating a similarity between a sentence using one or more of the phrases and a sentence in each source language included in two or more translation sentence sets stored in a storage medium; and the similarity calculation A similar sentence output unit that outputs one or more source language sentences similar to the sentence received by the sentence receiving unit using the similarity calculated by the unit, and one or more source languages output by the similar sentence output unit Is an instruction indicating permission for a sentence in one source language. A permission instruction receiving unit that receives an input of a permission instruction, a target language sentence acquiring unit that acquires a sentence in a target language corresponding to a sentence in a source language corresponding to the permission instruction received by the permission instruction receiving unit, and the generalization information A target language phrase acquisition unit that acquires, from one or more word / phrase sets stored in a storage medium, a target language phrase that is paired with a source language phrase corresponding to the generalization information acquired by the acquisition unit; In the target language sentence acquired by the sentence acquisition unit, the phrase corresponding to the generalization information is replaced with the target language phrase acquired by the target language phrase acquisition unit, and the target language sentence configuration that configures the output target language sentence And a program for causing the target language sentence configuration unit to function as an output unit that outputs a sentence in the target language.

  In addition, the program includes a sentence receiving unit that receives a sentence, a word dividing unit that divides the sentence into two or more words and acquires generalized information corresponding to the words, and the word dividing unit divides the computer Among the generalization information corresponding to two or more words and phrases, generalization information to be replaced with a phrase is determined and acquired, and the generalization information acquired by the generalization information acquisition section and the phrase division section Similarity that calculates the similarity between sentences that use one or more of two or more words divided by, and sentences of each source language that two or more translation sentences stored in the storage medium have Using the similarity calculated by the similarity calculator, the similar sentence output unit that outputs one or more source language sentences similar to the sentence received by the sentence receiver, and the similar sentence output unit Of one or more source language sentences output by, one source language sentence A permission instruction receiving unit that receives an input of a permission instruction that is an instruction indicating permission to the target, and a target language sentence acquisition that acquires a sentence in a target language corresponding to a sentence in a source language corresponding to the permission instruction received by the permission instruction receiving unit Language and a target language phrase that is acquired from one or more phrase sets stored in a storage medium, and a target language phrase that is paired with a source language phrase corresponding to the generalization information acquired by the generalization information acquisition unit In the target language sentence acquired by the acquisition unit and the target language sentence acquisition unit, the phrase corresponding to the generalization information is replaced with the target language phrase acquired by the target language phrase acquisition unit, and the target language sentence to be output And a program for functioning as an output unit that outputs a sentence in the target language configured by the target language sentence configuration unit.

  Further, in the above program, the computer further causes a voice reception unit that receives a voice and a voice recognition process on the voice received by the voice reception unit to further function as a sentence configuration unit that constitutes a sentence. It is preferable to function so as to acquire a sentence formed by the sentence constructing unit.

(Evaluation experiment)
Hereinafter, an evaluation experiment using the machine translation apparatus in the present embodiment will be described.

  The Japanese speech recognition performed by the sentence composition part 1102 is ATRASR (Ito Gen, Yutaka Toyo, Takatoshi Satoshi, Satoshi Nakamura, "Summary and Evaluation Report of ATRASR, Integrated Speech Recognition Environment," 2004 Autumn Meeting of the Acoustical Society of Japan Lecture Proceedings I, 1-P-30, pp. 221-222 (2004). The acoustic model is MDL-SSS (Takashi Tsuji, Shigeki Matsuda, Masayoshi Fujimoto, Herbordt, W., Toshiharu Horiuchi, Satoshi Nakamura, "Evaluation of Japanese speech recognition in ATR-Japanese acoustic model-," Acoustical Society of Japan 2006 The Spring Research Conference Proceedings, 1-P-21, pp. 185-186 (2006).) Language model is multi-class composite bigram (Hirofumi Yamamoto, Genichiro Kikui, "Evaluation of speech recognition in ATR-corpus and We used the language model, "The Acoustical Society of Japan 2006 Spring Meeting, 1-P-22, pp. 187-188 (2006)."

Three types of test sets as shown in FIG. 12 were used. In FIG. 12, BTEC (Basic Travel Expression Corpus) (“Takezawa, T., Sumita, E., Sugaya, F., Yamamoto, H., and Yamamoto, S.,“ Toward a broad-coverage bilingual corpus for speech translation of travel conversations in the real world, "Proc. of International Conference on Language Resources and Evaluation, pp. 147-152 (2002)." and "Kikui, G., Sumita, E., Takezawa, T., and Yamamoto, S., "Creating corpora for speech- to-speech translation," Proc. of 8 th European Conference on Speech Communication and Technology, Vol. 1, pp. 381-384 (2003). "reference), travel conversation basic representation This is the speech of the corpus.

  MAD (Machine-Aided Dialogs) ("Takezawa, T. and Kikui, G.," A comparative study on human communication behaviors and linguistic characteristics for speech-to-speech translation, "Proc. Of International Conference on Language Resources and Evaluation, pp. 1589-1592, (2004) ”is a task-execution-type dialogue conducted by a Japanese speaker and an English speaker via a speech translation system. A typist speaks instead of a speech recognition system. These are collected in the form of transcription and input to a machine translation system.

  FED (Field Experiment Data) ("Genichiro Kikui, Toshiyuki Takezawa, Masahide Mizushima, Seiichi Yamamoto, Hiroshi Sasaki, Tsune Kawai, Satoshi Nakamura," Monitoring Experiment in Spoken Dialogue Translation System in Real Environments, "Japan Society of Acoustics 2005 Fall Research Presentation (See Proceedings of the Conference, 1-7-10, pp. 11-12 (2006).) Is the data selected from the monitoring experiment conducted at Kansai International Airport. The average utterance length is about the same for BTEC and FED, and MAD is long. Perplexity increases in the order of BTEC, MAD, and FED.

  Since it is assumed that the user will use it, the setting is made so that the result is displayed immediately after the talk. Specifically, when processing time is expressed in real time factor (RTF), “RTF = 1” is set. FIG. 13 shows information on the result of Japanese speech recognition.

  Moreover, the generalization information mentioned above prepared two types of generalization information (condition A and condition B). Both types of generalization information are generalized parts of speech. “Condition A” is a generalization of “person name” and “place name” only. “Condition B” is a generalization of “common nouns”, “formal nouns”, “sa variant nouns”, “adjective nouns”, “sa deformed nouns”, “adjectives”, “adjective verbs”, “numerical numbers”, and “adverbs”. That is, the noun content words of “condition B” are “common noun”, “formal noun”, “sa variant noun”, “adjective noun”, “sa deformed noun”, “adjective”, “adjective verb”, “numerical number”, and “adverb”.

  The procedure for generalizing and matching parts of speech is as follows, as described above. First, the surface representation (for example, “Tokyo”) to be generalized is deleted. Separately, the information (for example, “Tokyo”) is left. Second, the parallel translation example (translated sentence set storage unit 103) is searched. Third, if the number of objects to be generalized in the parallel translation pair is different, it is deleted from the candidates. Fourth, the original surface representation is added to the candidate including the generalized one.

As a large-scale corpus, the travel conversation basic expression corpus BTEC was used. The outline is shown in FIG. BTEC1, BTEC2, BTEC3, and BTEC4 were used, which were divided into subsets according to the time of collection and which language was used as the starting point. BTEC1, BTEC2, BTEC3 are expressions of scenes where Japanese people mainly go to Europe and America. BTEC4 is an expression of a scene where a traveler from the United States or Australia comes to Japan. The total number of utterance expressions in the bilingual format to be searched is about 490,000. Trilingual parallel between Japanese and English, but experiments were conducted in the direction of Japanese and English. The number N _{r of} preliminary selections was set to 30.

  Experiments were performed using the three types of test sets shown in FIG. 12, with the value of the search weight parameter α set to 0.1 increments from 0 to 1. The accuracy rate (Top 1 Accuracy) that perfectly matched the first candidate using the speech recognition result as input, and the accuracy rate (Top 30 Accuracy) that includes the exact match up to 30th, correct morpheme sequence The results obtained for the accuracy rate (Oracle Accuracy) with respect to condition A (Condition A) and condition B (Condition B) are shown in the figure together with the utterance accuracy rate (Correct Utterance). FIG. 15 shows the results of BTEC, FIG. 16 shows the results of MAD, and FIG. 17 shows the results of FED.

  The BTEC experiment is closed. The MAD and FED experiments are open results, although typical expressions such as greetings may be included. As for BTEC, the generalization condition A is better than the generalization condition B, but for other MADs and FEDs, the generalization condition B is better than the generalization condition A. Only in the case of generalization condition A in BTEC, when the candidates up to 30th are given by inputting the speech recognition result, the accuracy rate is improved compared to the case of only 1st place, but other test sets and conditions are 1 There is not much difference between only the place and the cumulative value up to the 30th place.

  Next, manual evaluation results are shown.

  The results shown in FIGS. 15, 16, and 17 are obtained by calculating the ratio that the obtained similar use example completely matches the original language expression as a morpheme string. In fact, even if the expression is slightly different, it may be considered that the same semantic content is transmitted. Therefore, I manually evaluated the content of FED, which is a test set of real dialogue voice. Only the first candidate selected as a similar example was classified into the following ABCD rank. (A) indicates that “the morpheme string is completely matched”. (B) indicates that “the expressions are different but have the same meaning”. (C) indicates "contains some useful information". (D) indicates “meaning is different and useless”.

  In addition, what was not obtained similar example. It was classified into the following two. (OK) indicates that “the speech recognition result is correct in units of utterances”. (NG) indicates that “the speech recognition result is incorrect”.

  FIG. 18 shows a manual evaluation result for the generalization condition A, and FIG. 19 shows a manual evaluation result for the generalization condition B.

In the real conversation voice test set FED, the rate of outputting similar examples in the case of the generalization condition A is about 75%, but about one-third of those that are useless are included. In the case of the generalization condition B, the ratio of outputting a similar example is about 60%, but there are few useless ones. An example in which rank B in generalization condition A is improved to rank B in generalization condition B is shown below.
[Input Voice] There is a train or bus, which is better?
[Generalization condition A] Print paper is glossy or silky. Which is better?
[Generalization condition B] Which is better for trains and buses?

Moreover, although the example for similar use was not searched, the example of what is classified into OK that the recognition result is correct is shown.
[Example of OK] ¥ 230.

  Based on the above, we examined a method for searching linguistic expressions similar to speech recognition results using a large-scale corpus, and conducted an evaluation experiment using multiple test sets of spoken speech of travel conversation sentences and real conversation speech. Showed the effect.

  In each of the above embodiments, each process (each function) may be realized by centralized processing by a single device (system), or by distributed processing by a plurality of devices. May be.

  FIG. 20 shows the external appearance of a computer that executes the programs described in this specification to realize the machine translation apparatus according to the various embodiments described above. The above-described embodiments can be realized by computer hardware and a computer program executed thereon. FIG. 20 is an overview of the computer system 340, and FIG. 21 is a block diagram of the computer system 340.

  20, the computer system 340 includes a computer 341 including a FD (Flexible Disk) drive and a CD-ROM (Compact Disk Read Only Memory) drive, a keyboard 342, a mouse 343, a monitor 344, and a microphone 345. .

  In FIG. 21, in addition to the FD drive 3411 and the CD-ROM drive 3412, a computer 341 includes a CPU (Central Processing Unit) 3413, a bus 3414 connected to the CD-ROM drive 3412 and the FD drive 3411, and a boot-up program. ROM (Read-Only Memory) 3415 for storing programs such as a RAM, and a RAM (Random Access Memory) 3416 connected to the CPU 3413 for temporarily storing application program instructions and providing a temporary storage space , An application program, a system program, and a hard disk 3417 for storing data. Although not shown here, the computer 341 may further include a network card that provides connection to the LAN.

  A program that causes the computer system 340 to execute the functions of the machine translation apparatus according to the above-described embodiment is stored in the CD-ROM 3501 or the FD 3502, inserted into the CD-ROM drive 3412 or the FD drive 3411, and further stored in the hard disk 3417. May be forwarded. Alternatively, the program may be transmitted to the computer 341 via a network (not shown) and stored in the hard disk 3417. The program is loaded into the RAM 3416 at the time of execution. The program may be loaded directly from the CD-ROM 3501, the FD 3502, or the network.

  The program does not necessarily include an operating system (OS) or a third-party program that causes the computer 341 to execute the functions of the machine translation apparatus according to the above-described embodiment. The program only needs to include an instruction portion that calls an appropriate function (module) in a controlled manner and obtains a desired result. How the computer system 340 operates is well known and will not be described in detail.

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

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

  As described above, the machine translation apparatus according to the present invention has an effect that machine translation can be performed with extremely high accuracy, and is useful as a machine translation apparatus or the like.

Block diagram of machine translation apparatus in embodiment 1 Flow chart for explaining the operation of the machine translation apparatus A flowchart for explaining the operation of the similarity calculation process Flowchart for explaining the operation of the translated sentence output process Figure showing synonym generalization information management table Diagram showing synonym phrase management table Figure showing the translation set management table The figure which shows the example of the sentence of the selected source language The figure which shows the example of the same screen output The figure which shows the example of the same screen output Block diagram of machine translation apparatus according to Embodiment 2 Diagram showing the test set The figure which shows the information of the same Japanese voice recognition result The figure which shows the outline of the travel conversation basic expression corpus The figure which shows the result of the same BTEC The figure which shows the result of the same MAD The figure which shows the result of the same FED The figure which shows the manual evaluation result with respect to the generalization condition A The figure which shows the manual evaluation result with respect to the generalization condition B External view of a computer that implements the machine translation device Block diagram of a computer system that implements the machine translation device

Explanation of symbols

DESCRIPTION OF SYMBOLS 101 Phrase generalization information storage part 102 Phrase set storage part 103 Translation sentence set storage part 104, 1104 Sentence reception part 105 Phrase division part 106 Generalization information acquisition part 107 Similarity calculation part 108 Similar sentence output part 109 Permit instruction | indication reception part 110 Target language sentence acquisition unit 111 Target language phrase acquisition unit 112 Target language sentence composition unit 113 Output unit 114 Re-input prompting unit 1071 First generalization information number acquisition unit 1072 Second generalization information number acquisition unit 1073 Similarity calculation target sentence selection Means 1074 Similarity Calculation Means 1101 Voice Accepting Unit 1102 Sentence Constructing Unit

Claims (10)

A phrase generalization information storage unit capable of storing one or more phrase generalization information having generalization information that is generalized information of the phrase and the phrase;
A phrase set storage unit that stores at least one phrase set having a pair of a source language phrase and a target language phrase;
A translation set having a pair of a source language sentence and a target language sentence, and having generalization information corresponding to some words and phrases of the source language sentence and the target language sentence of the translation sentence set A translated sentence set storage unit storing two or more sentence sets;
A sentence reception unit for receiving sentences;
A phrase dividing unit that divides the sentence into two or more phrases;
A generalization information acquisition unit that acquires generalization information corresponding to one or more words divided by the word division unit from the word generalization information storage unit;
A sentence using one or more words out of two or more words divided by the generalization information and the word dividing part acquired by the generalization information acquisition part, and two or more sets of translation sentences in the translation sentence set storage part A similarity calculation unit for calculating the similarity with each source language sentence included in the set;
A similarity sentence output unit that outputs one or more source language sentences similar to the sentence received by the sentence reception unit using the similarity calculated by the similarity calculation unit;
Among the one or more source language sentences output by the similar sentence output unit, a permission instruction receiving unit that receives an input of a permission instruction that is an instruction indicating permission for one source language sentence;
A target language sentence acquisition unit that acquires a target language sentence corresponding to a source language sentence corresponding to the permission instruction received by the permission instruction reception unit;
A target language phrase acquisition unit that acquires from the phrase set storage unit a target language phrase that is paired with a source language phrase corresponding to the generalization information acquired by the generalization information acquisition unit;
The purpose of constructing a sentence of the target language to be output by replacing the phrase corresponding to the generalization information in the sentence of the target language acquired by the target language sentence acquiring unit with the phrase of the target language acquired by the target language phrase acquiring unit A language sentence composition part;
A machine translation apparatus comprising: an output unit that outputs a target language sentence configured by the target language sentence constructing unit. A phrase set storage unit that stores at least one phrase set having a pair of a source language phrase and a target language phrase;
A translation set having a pair of a source language sentence and a target language sentence, and having generalization information corresponding to some words and phrases of the source language sentence and the target language sentence of the translation sentence set A translated sentence set storage unit storing two or more sentence sets;
A sentence reception unit for receiving sentences;
A phrase dividing unit that divides the sentence into two or more words and acquires generalization information corresponding to the words;
A generalization information acquisition unit for determining and acquiring generalization information to be replaced with a phrase from among the generalization information corresponding to two or more words divided by the phrase division unit;
A sentence using one or more words out of two or more words divided by the generalization information and the word dividing part acquired by the generalization information acquisition part, and two or more sets of translation sentences in the translation sentence set storage part A similarity calculation unit for calculating the similarity with each source language sentence included in the set;
A similar sentence output unit that outputs one or more source language sentences similar to the sentence received by the sentence receiving unit using the similarity calculated by the similarity calculating unit;
Among the one or more source language sentences output by the similar sentence output unit, a permission instruction receiving unit that receives an input of a permission instruction that is an instruction indicating permission for one source language sentence;
A target language sentence acquisition unit that acquires a target language sentence corresponding to a source language sentence corresponding to the permission instruction received by the permission instruction reception unit;
A target language phrase acquisition unit that acquires from the phrase set storage unit a target language phrase that is paired with a source language phrase corresponding to the generalization information acquired by the generalization information acquisition unit;
The purpose of constructing a sentence of the target language to be output by replacing the phrase corresponding to the generalization information in the sentence of the target language acquired by the target language sentence acquiring unit with the phrase of the target language acquired by the target language phrase acquiring unit A language sentence composition part;
A machine translation apparatus comprising: an output unit that outputs a target language sentence configured by the target language sentence constructing unit. A voice reception unit for receiving voice;
Voice recognition processing is performed on the voice received by the voice reception unit, and a sentence composing unit that constitutes a sentence is provided.
The machine translation device according to claim 1, wherein the sentence receiving unit acquires a sentence formed by the sentence composing unit. The similarity calculation unit includes:
First generalization information number acquisition means for acquiring a first generalization information number that is the number of generalization information acquired by the generalization information acquisition unit;
Acquisition of the number of second generalization information for acquiring the number of second generalization information that is the number of generalization information that the sentence of the source language or the sentence of the target language possessed by two or more sets of translation sentences in the translation sentence set storage unit Means,
A similarity calculation target sentence selection means for selecting a sentence in a source language having a second generalization information number that matches the first generalization information number;
Sentence using generalization information acquired by the generalization information acquisition unit and one or more words out of two or more words divided by the word division unit, and one or more selected by the similarity calculation target sentence selection unit The machine translation apparatus according to claim 1, further comprising a similarity calculation unit that calculates a similarity with a source language sentence. When the permission instruction accepting unit does not accept the input of the permission instruction,
A re-input prompting unit for performing an output for prompting re-input or editing of the sentence received by the sentence receiving unit;
The machine translation device according to claim 1, wherein the sentence receiving unit receives the sentence again. When the permission instruction accepting unit does not accept the input of the permission instruction,
An output for prompting re-input of voice is provided, and a re-input prompting unit is further provided.
The machine translation apparatus according to claim 2, wherein the voice receiving unit receives voice again. The generalized information acquisition unit
The machine translation apparatus according to claim 1, wherein generalization information is acquired only for a noun or adjective phrase. Computer
A sentence reception unit for receiving sentences;
A phrase dividing unit that divides the sentence into two or more phrases;
A generalization information acquisition unit that acquires generalization information corresponding to one or more words divided by the word division unit from one or more word generalization information stored in a storage medium;
A sentence using generalization information acquired by the generalization information acquisition unit and one or more words out of two or more words divided by the word division unit, and two or more sets of translated sentences stored in a storage medium A similarity calculation unit for calculating the similarity with each source language sentence included in the set;
A similar sentence output unit that outputs one or more source language sentences similar to the sentence received by the sentence receiving unit using the similarity calculated by the similarity calculating unit;
Among the one or more source language sentences output by the similar sentence output unit, a permission instruction receiving unit that receives an input of a permission instruction that is an instruction indicating permission for one source language sentence;
A target language sentence acquisition unit that acquires a target language sentence corresponding to a source language sentence corresponding to the permission instruction received by the permission instruction reception unit;
A target language phrase acquisition unit that acquires a target language phrase that is paired with a source language phrase corresponding to the generalization information acquired by the generalization information acquisition unit, from one or more phrase sets stored in a storage medium; ,
The purpose of constructing a sentence of the target language to be output by replacing the phrase corresponding to the generalization information in the sentence of the target language acquired by the target language sentence acquiring unit with the phrase of the target language acquired by the target language phrase acquiring unit A language sentence composition part;
The program for functioning as an output part which outputs the sentence of the target language which the said target language sentence structure part comprised. Computer
A sentence reception unit for receiving sentences;
A phrase dividing unit that divides the sentence into two or more words and acquires generalization information corresponding to the words;
A generalization information acquisition unit for determining and acquiring generalization information to be replaced with a phrase from among the generalization information corresponding to two or more words divided by the phrase division unit;
A sentence using generalization information acquired by the generalization information acquisition unit and one or more words out of two or more words divided by the word division unit, and two or more sets of translated sentences stored in a storage medium A similarity calculation unit for calculating the similarity with each source language sentence included in the set;
A similar sentence output unit that outputs one or more source language sentences similar to the sentence received by the sentence receiving unit using the similarity calculated by the similarity calculating unit;
Among the one or more source language sentences output by the similar sentence output unit, a permission instruction receiving unit that receives an input of a permission instruction that is an instruction indicating permission for one source language sentence;
A target language sentence acquisition unit that acquires a target language sentence corresponding to a source language sentence corresponding to the permission instruction received by the permission instruction reception unit;
A target language phrase acquisition unit that acquires a target language phrase that is paired with a source language phrase corresponding to the generalization information acquired by the generalization information acquisition unit, from one or more phrase sets stored in a storage medium; ,
The purpose of constructing a sentence of the target language to be output by replacing the phrase corresponding to the generalization information in the sentence of the target language acquired by the target language sentence acquiring unit with the phrase of the target language acquired by the target language phrase acquiring unit A language sentence composition part;
The program for functioning as an output part which outputs the sentence of the target language which the said target language sentence structure part comprised. Computer
A voice reception unit for receiving voice;
Voice recognition processing is performed on the voice received by the voice reception unit, and the voice reception unit further functions as a sentence configuration unit that forms a sentence.
The program according to claim 8 or 9, wherein the sentence receiving unit functions to acquire a sentence formed by the sentence composing unit.