JP6096138B2 - Word order replacement apparatus, method and program, and translation quality evaluation apparatus and program - Google Patents

Description

  The present invention relates to a word order replacement apparatus, method and program, and a translation quality evaluation apparatus and program, and more particularly to a word order replacement apparatus, method and method for a sentence written in a language subsequent to a main part such as Japanese. The present invention relates to a program, a translation quality evaluation apparatus, and a program.

  In recent years, in machine translation research, statistical machine translation (SMT: Statistical Machine Translation) that automatically generates a translation dictionary by statistically analyzing a large number of translated documents and creates a translator using the translation dictionary Is sweeping the world. However, between languages with completely different word orders, such as Japanese-English translation and English-Japanese translation, the difference in word order becomes a problem, and it may be difficult to perform high-quality translation using SMT.

  In addition, with regard to automatic translation evaluation devices used to improve translators in SMT, BLEU and other evaluation methods that have been used as standard are the human evaluation points for Japanese-English and English-Japanese translations. Is known to have a low correlation (see, for example, Non-Patent Document 1).

  Thus, an automatic evaluation method has been proposed in which SMT between languages having different word orders is focused on the word order (see, for example, Patent Document 1). This method is a method called RIBES (Rank-based Intuitive Bilingual Evaluation Score) in Non-Patent Document 1. RIBES evaluates the closeness of the global word order between an ideal translation (reference translation) created by a human and a translation by SMT, using a rank correlation coefficient. In RIBES, English-Japanese translation and Japanese-English translation show a high correlation with human evaluation points.

JP 2012-079107 A

Isao Goto, Ka Po Chow, Bin Lu, Eiichiro Sumita, Benjamin K. Tsou, "Overview of the Patent Machine Translation Task at the NTCIR-10 Workshop", Proceedings of the 10th NTCIR Conference, pp.260-286, 2010.

  However, it is known that languages following the main part such as Japanese have a relatively high degree of freedom in word order compared to other languages, and if there is only one reference translation, the translation quality is correct. Evaluation may not be possible. For example, when the reference translation of "He read the book." Is "He read the book.", The translation "He read the book." The translation, “He read.” Is disappointing. Therefore, there is a translated sentence that cannot be highly evaluated by the evaluation by RIBES, while the translated sentence should be highly evaluated by human evaluation.

  The present invention has been made in view of the above circumstances, and an object thereof is to provide a word order changing device, method, and program capable of generating a plurality of sentences having the same meaning but different word orders. It is another object of the present invention to provide a translation quality evaluation apparatus and method that can appropriately evaluate the quality of translation into a language having a high degree of freedom in word order.

  In order to achieve the above object, the word order permutation apparatus according to the present invention includes a specific phrase ending with a predetermined particle in a dependency tree indicating a dependency relationship between sentence phrases, a phrase that modifies the specific phrase, The sentence order is changed based on a replacement part that replaces each of a plurality of partial trees that modify the same verb, and a dependency tree in which the replacement part is replaced by the replacement part. And a generation unit that generates a word order replacement sentence.

  According to the word order exchanging device according to the present invention, the replacement unit includes a specific phrase that ends with a predetermined particle and a phrase that modifies the specific phrase in a dependency tree that indicates a dependency relationship between the phrases of the sentence. Each of the subtrees that modify the same verb is replaced with a tree. And a production | generation part produces | generates the word order replacement | exchange sentence which replaced the word order of the sentence based on the dependency tree by which the subtree was replaced by the replacement part.

  As described above, the replacement of the word order is limited to the replacement of a partial tree including a specific phrase that ends with a predetermined particle and a phrase that modifies the specific phrase, and that modifies the same verb. Therefore, it is possible to eliminate a word order in which the meaning is changed or the expression is unnatural, and to generate a plurality of sentences having the same meaning and different word order as the original sentence.

  The predetermined particles may be “ha”, “ga”, “ha”, “ni”, “de”, and “to”. Phrases ending with these particles are particles that can be reordered when modifying the same verb.

  Further, the word order exchanging device according to the present invention is arranged in front of the other subtree including a verb different from the verb to be modified by the one subtree including the predetermined particle other than the particle “ha” by the replacement unit. A dependency tree in which the one subtree is replaced, or a dependency tree in which the other subtree including the predetermined particle other than the particle “O” is replaced before the one subtree including the adjective On the basis of the above, it may be configured to include a restricting unit that restricts the generating unit from generating the word order replacement sentence. Thereby, the sentence of the unnatural word order can be excluded more appropriately.

In order to achieve the above object, the translation quality evaluation apparatus according to the present invention modifies a specific phrase ending with a predetermined particle in a dependency tree indicating a dependency relation between sentence phrases, and modifies the specific phrase. A subtree comprising a phrase to be replaced, a replacement unit that replaces each of a plurality of subtrees that modify the same verb, and a dependency tree in which the subtree is replaced by the replacement unit. A generating unit that generates a word order replacement sentence in which the word order is changed, and the other subtree including a verb different from the verb modified by the one subtree including the predetermined particle other than the particle “ha” by the replacement unit. A dependency tree in which the one subtree is replaced before, or a relationship tree in which the other subtree including the predetermined particle other than the particle “O” is replaced before the one subtree including the adjective Based on the shrub Word order and limiting section by serial generating unit limits so as not to generate the word order replacement sentence, the translated sentence and machine translation the original text, ideal translation a reference translation is of the original, and without changing the meaning of the reference translation A calculation unit that calculates the similarity with each of the word-order replacement reference translations in which the words are replaced, and an evaluation unit that evaluates the translation quality of the translation based on the similarity calculated by the calculation unit. ing.

  According to the translation quality evaluation apparatus according to the present invention, the calculation unit translates the original sentence by machine translation, the reference translated sentence that is an ideal translation of the original sentence, and the word order replacement that changes the word order without changing the meaning of the reference translated sentence The similarity with each reference translation is calculated. Then, the evaluation unit evaluates the translation quality of the translation based on the similarity calculated by the calculation unit. Thereby, the quality of translation into a language with a high degree of freedom of word order can be appropriately evaluated compared with the case where it evaluates with one reference translation.

In addition, the word order replacement reference translation sentence may be a word order replacement sentence for the reference translation sentence generated by the generation unit under the restriction of the restriction unit . Accordingly, the translation can be evaluated using a plurality of reference translations having the same meaning and different word orders without requiring manual labor.

  Further, the word order replacement method according to the present invention is a word order replacement method in a word order replacement device including a replacement unit and a generation unit, wherein the replacement unit indicates a dependency relationship between sentence clauses. A subtree consisting of a specific phrase ending with a predetermined particle and a phrase that modifies the specific phrase, replacing each of a plurality of subtrees that modify the same verb, and the generation unit includes: Generating a word order replacement sentence in which the word order of the sentence is changed based on a dependency tree in which the subtree has been changed by the replacement section.

  Moreover, the word order replacement program according to the present invention is a program for causing a computer to function as each unit constituting the word order replacement device.

  Further, the translation quality evaluation program according to the present invention is a program for causing a computer to function as each part constituting the above translation quality evaluation apparatus.

  As described above, according to the word order replacement device, method, and program of the present invention, the word order replacement is a subtree consisting of a specific phrase that ends with a predetermined particle and a phrase that modifies the specific phrase. Limit the replacement of subtrees that modify the same verb. Therefore, it is possible to eliminate the word order in which the meaning is changed or the expression is unnatural and to generate a plurality of sentences having the same meaning but different word orders.

  Moreover, according to the translation quality evaluation apparatus and program concerning this invention, the translation quality of a translation is evaluated also using the word order replacement | exchange reference translation sentence which replaced the word order, without changing the meaning of a reference translation sentence. Therefore, an effect is obtained that the quality of translation into a language having a high degree of freedom in word order can be appropriately evaluated as compared with the case of evaluating with one reference translation.

It is a functional block diagram which shows the translation quality evaluation system containing the word order replacement apparatus and translation quality evaluation apparatus which concern on this Embodiment. It is a figure which shows an example of a dependency tree. It is a figure which shows an example of a dependency tree. It is a flowchart which shows an example of the word order replacement | exchange process in this Embodiment. It is a figure which shows an example of a dependency tree. It is a figure which shows an example of the dependency tree represented by the connection between nodes. It is a flowchart which shows an example of a commutative particle list creation process. It is a figure which shows an example of the recursive program which outputs a permutation. It is a flowchart which shows an example of a permutation output process. It is a flowchart which shows an example of another predicative particle restriction | limiting process. It is a flowchart which shows an example of the translation quality evaluation process in this Embodiment. It is a figure which shows an example of the dependency tree represented by the connection between nodes. It is a figure which shows an example of the dependency tree represented by the connection between nodes. It is a figure which shows an example of the dependency tree represented by the connection between nodes.

  Hereinafter, embodiments of the present invention will be described in detail with reference to the drawings.

<Principle of this embodiment>
In this embodiment, the word order of the reference translation sentence represented by a dependency tree indicating the dependency relation between clauses is switched to generate a word order replacement reference translation sentence. In order to generate a Japanese-style head final (subsequent main part) word order sentence, it is sufficient to follow the dependency tree with a post-order.

  The post-order is a term that represents the order in which all nodes of a tree structure are traversed in computer science. The post-order is used alongside pre-order and in-order, and has been used as a standard for a long time. In the pre-order, the parent node is processed before the child node of the tree structure is processed (for example, output). In the in-order, the parent node is processed while the child node is being processed. In post-order, the parent node is processed after all the child nodes are processed.

  The sentence generated by post-order from the dependency tree satisfies the Japanese word order condition that "the expression to be modified comes after the expression to be modified", but in the output sentence, It includes sentences with unnatural word order, such as sentences that are semantically misleading and difficult to understand.

  Therefore, in the present embodiment, an unnatural word order sentence is avoided, and a word order replacement reference translation sentence that is another word order sentence that is valid as Japanese is generated in the same meaning as the original reference translation sentence. Then, the reference translation set including the word order replacement reference translation is used for evaluating the quality of the machine translation. This makes it possible to give a high evaluation to a machine translation sentence that has been obtained only with a low evaluation so far, using a sentence having the same meaning but having a different word order from the original reference translation sentence.

  In the present embodiment, in order to prevent generation of an unnatural word order sentence, the word order is rearranged by “replacement of substitutive particle subtrees that modify the same verb”. In addition, the word order rearrangement inappropriate for use as a reference translation sentence is restricted by “another prescriptive particle restriction”.

  First, a commutative particle is a specific particle attached to the end of a phrase. This is a particle whose meaning does not change before and after the replacement even if it is modified with other phrases that end with a particular particle. Specifically, the particles “ha”, “ga”, “wa”, “ni”, “de”, and “to”. For example, a sentence such as “Taro bought a map at a bookstore” may be divided into phrases (parts separated by “/”) such as “Taro / at a bookstore / buy a map /”. it can. “Taro ga”, “in bookstore”, and “map” are all phrases that end with commutative particles (hereinafter referred to as “commutative particle clauses”), and qualify the same verb “Bought” Therefore, the order of these commutative particle clauses can be changed. For example, it is possible to change phrases such as “Map / at bookstore / Taro / Bought”.

  Here, the reason why only a specific particle is used as a commutative particle is that the allowed word order varies depending on the use of the particle. For example, when the particles “from” and “to” are recognized as commutative particles, the order of the sentence “walked from Tokyo to Shinagawa” is changed, and a sentence “walked from Tokyo to Shinagawa” is generated. . The spoken language may be “I walked from Tokyo to Shinagawa”, but it is a slightly unnatural word order when it is used as a reference translation for evaluating the quality of machine translation. Therefore, in order not to allow such a change of word order, only specific particles are recognized as commutative particles.

  Next, the “commutative particle subtree” will be described. For example, consider the case where the original reference translation is “Taro / Blanker / Neighborhood / Bookstore / Large / Map / Bought”. If you just replace the commutative particle clauses, the sentence “Map / Hanaya / Neighborhood / Taro is / Large / Bookstore / Bought” will also be generated. Will change the meaning. In order to rearrange the word order without changing the meaning, it is necessary to move together the clauses that modify the commutative particle clauses. In this example sentence, if the phrase that modifies this commutative particle clause is also included for the commutative particle clause “bookstore”, it becomes “at / next to the bookstore”. In addition, if the commutative particle clause “map is” and a clause that modifies the commutative particle clause is included, “large / map” is obtained. When rearranging in this unit, it becomes “Large / Map / Taro / Hanaya / Neighborhood / Bookstore / Buy /” and the meaning is the same as the original reference translation. A part consisting of a commutative particle clause and a clause that modifies the commutative particle clause, that is, a subtree rooted at a node corresponding to the commutative particle clause is a “commutative particle subtree”. .

  Thus, by exchanging the word order on a permutable particle subtree basis, it is possible to prevent generation of an unnatural word order sentence. However, when there are a plurality of predicates in a sentence, as described above, an unnatural sentence order may be generated by exchanging a commutative particle subtree that modifies the same verb.

For example,
Hanako / operated / at the flower shop / next door / at the bookstore / Taro / buyed the map /
In subtitles, the replacement particle sub-tree “Hanako / Managed / Hanaya / Neighborhood / Bookstore” and “Map” are modified to modify the verb “Bought”.
I bought a map / Hanako / operated / florist / next to / bookstore / Taro / bought.

Becomes an unnatural sentence order. The unnatural reason is that there is another verb “running” in front of the verb “buy” that is modified by “map” or “taro”, and “map” is in front of this verb. It is because it can be read like “I run a map”.

Therefore, in the present embodiment, a restriction of “other prescriptive particle restriction” is provided to prevent the generation of an unnatural word order sentence as described above. Specifically, the “other prescriptive particle restrictions” are the following restrictions.
The only commutative particle subtree that can appear before another verb that is not the original verb is the commutative particle subtree that ends with the commutative particle “ha”.
• The only commutative particle subtree that can appear before an adjective is the commutative particle subtree that ends with the commutative particle "O".

<Configuration of the present embodiment>
FIG. 1 is a block diagram showing a translation quality evaluation system 40 including a word order replacement device 10 and a translation quality evaluation device 20 according to an embodiment of the present invention. The word order changing device 10 can be configured by a computer including a CPU, a RAM, and a ROM that stores a program for executing a word order changing process described later. Moreover, the translation quality evaluation apparatus 20 can be comprised with a computer provided with CPU, RAM, and ROM which memorize | stored the program for performing the translation quality evaluation process mentioned later.

  The word order changing device 10 receives a dependency tree indicating a dependency relationship between phrases of the reference translation, which is an ideal translation of the original for evaluating the translation quality of the translation obtained by machine translation of the original. A known dependency analyzer can be used to obtain a dependency tree from the reference translation.

Here, many dependency analyzers have problems with processing of quotation marks and parentheses. For example,
“Hou went to America.”
Is analyzed by a standard Japanese dependency analyzer, the clauses are divided as shown in FIG. 2 and the following, for example.
“Hiro / went to America /” / said.

For this reason, if you try to switch between “Haga” and “To America” by phrase,
He was told to America that "Hiro went."
It becomes the word order like. Therefore, in this embodiment, as shown in FIG. 3 and below, the clause of the sentence inside the parentheses is re-cut into the clause of the parenthesis, and the clause of the opening parenthesis is modified to modify the clause of the closing parenthesis. It is assumed that a dependency tree is accepted as an input.
“/ Hiro / went to America / /” /

  As shown in FIG. 1, functionally, the computer that constitutes the word order changing device 10 performs a replacement part 11 that replaces a replaceable particle subtree that modifies the same verb, and a pre-adjunct particle restriction determination. It can be expressed by a configuration including a restriction unit 12 and a generation unit 13 that generates a word order replacement reference translation based on a dependency tree in which a commutative particle subtree is replaced. Hereinafter, each part is explained in full detail.

  The replacement unit 11 receives the dependency tree of the input reference translation sentence, and specifies a commutative particle clause that modifies the same verb for each verb included in the reference translation sentence. The replacement unit 11 outputs all permutations of a plurality of commutative particle clauses when there are a plurality of commutative particle clauses that modify the same verb. Each permutation corresponds to each pattern of replacement of a commutative particle subtree consisting of commutative particle clauses and clauses that modify commutative particle clauses.

  The restriction unit 12 excludes a permutation corresponding to the replacement of a subjunction particle subtree restricted by the above-mentioned prejunctive particle restriction from the permutations output by the replacement unit 11, and outputs the remaining permutations.

  The generation unit 13 generates each of the word order replacement reference translations based on each dependency tree in which the commutative particle subtree is replaced according to each of the permutations output from the restriction unit 12. The generation unit 13 combines the original reference translation text and the generated word order replacement reference translation text into a reference translation set 30 and stores it in a predetermined storage area.

  As shown in FIG. 1, the computer constituting the translation quality evaluation apparatus 20 functionally calculates a similarity between the evaluation target translation and each of the reference translations, and an evaluation target based on the similarity. It can be expressed by a configuration including an evaluation unit 22 that evaluates the quality of the translation. Hereinafter, each part is explained in full detail.

  The calculation unit 21 receives an evaluation target translated sentence that is an evaluation target of the quality of machine translation as an input. Further, the reference translation set 30 stored in a predetermined storage area is read out. Then, the calculation unit 21 calculates the degree of similarity between the evaluation target translation and each of the plurality of reference translations (the original reference translation and the word order replacement reference translation) included in the reference translation set 30. The similarity can be calculated using a known method such as notation matching.

  The evaluation unit 22 obtains and outputs an evaluation result for the evaluation target translation based on the similarity to each of the plurality of reference translations calculated by the calculation unit 21. The evaluation result can be an evaluation based on the highest similarity among the similarities for each of the plurality of reference translations, an evaluation based on the average of the similarities, and the like.

<Operation of the present embodiment>
Next, the operation of the word order changing device 10 of the present embodiment will be described. When the dependency translation dependency tree is input to the word order replacement device 10, the word order replacement device 10 performs the word order replacement processing shown in FIG.

  In step S10, the replacement unit 11 receives a dependency tree of the input reference translation. Here, for example, a dependency tree as shown in FIG. 5 is accepted. In the example of FIG. 5, each morpheme constituting the sentence corresponds to each line, and each line includes information such as morpheme notation, part of speech, and utilization form. In addition, a phrase information line indicating information on the phrase is provided before the line corresponding to the first morpheme of each phrase. In the example of FIG. 5, a line starting with “*” represents a clause information line, and the clause information line includes a clause id for identifying the clause and a clause id of the clause that the clause modifies ( Hereinafter, information of “modification destination clause id”) is included.

  Further, FIG. 6 shows an example in which the dependency tree of FIG. 5 is represented by each node as a node (square frame in FIG. 6), and the dependency relationship between the clauses is expressed by connection between nodes. In the example of FIG. 6, the nodes are connected by an arrow from the child node corresponding to the parent-child relationship to the parent node. Also, the number added above or to the left of each node is the phrase id of the phrase corresponding to that node.

  Next, in step S20, the replacement unit 11 executes a commutative particle list creation process. The commutative particle list creation process is a process of creating a commutative particle list by specifying a commutative particle clause that modifies the same verb for each verb included in the reference translation sentence. Here, an example of commutative particle list creation processing when the dependency tree shown in FIG. 5 is input will be described with reference to FIG.

  In step S21, the replacement unit 11 sets the variable id corresponding to the clause id and the variable link corresponding to the clause id of the modification destination to be empty. In addition, the commutative particle list A [link] is set to be empty. The commutative particle list A [link] is a list that stores the phrase ids of commutative particle phrases that modify the verb whose phrase id is link. For example, if commutative particle clauses with clause id = '0' and '1' modify a verb with clause id = '3', A ['3'] = ['0', '1'] It expresses.

  Next, in step S22, the next line of the dependency tree is read. Here, since this step is executed for the first time in this routine, the first line of the dependency tree is read.

  Next, in step S23, it is determined whether the line read in step S22 is the last line or a phrase information line. If the read line is “EOS”, it can be determined that it is the last line, and if “*” is described at the beginning of the read line, it can be determined that it is a clause information line. it can. If the determination is affirmative, the process proceeds to step S24. If the determination is negative, the process proceeds to step S28. Here, since the first line is “* 0 3D”, it is determined that it is a phrase information line, and the process proceeds to step S24.

  In step S24, the replacement unit 11 determines whether the variable id is empty or whether id is already stored in A [link] corresponding to the variable link. If the determination is affirmative, the process proceeds to step S26. If the determination is negative, the process proceeds to step S25. Here, since the variable id is empty, an affirmative determination is made, and the routine proceeds to step S26.

  In step S26, the replacement unit 11 determines whether or not the line read in step S22 is the last line. If the determination is affirmative, the process proceeds to step S30. If the determination is negative, the process proceeds to step S27. Here, since the read line is not “EOS”, a negative determination is made and the process proceeds to step S27.

  In step S27, the phrase id described in the line read in step S22, that is, the phrase information line, is acquired and set in the variable id. Also, the modification-destination clause id described in the clause information line is acquired and set in the variable link. Here, id = '0', link = '3', and clause id = '0' (hereinafter referred to as “0th clause”. The same applies to clauses with other clause ids) “Taro ga” is processed It becomes a target.

  Returning to step S22, the next line "Taro noun, ..." is read, a negative determination is made in step S23, and the flow proceeds to step S28. In step S28, the morpheme notation and part of speech of the line read in step S22 are referred to and it is determined whether or not the morpheme is a commutative particle. If the determination is affirmative, the process proceeds to step S29. If the determination is negative, the process returns to step S22. Here, since the morpheme “Taro” is a noun and not a commutative particle, the process returns to step S22.

  In step S22, the next line “ga particle, case particle,...” Is read, and in step S23, a negative determination is made, and the flow proceeds to step S28. Here, since the morpheme “ga” is a specific particle predetermined as a commutative particle, an affirmative determination is made and the process proceeds to step S29.

  In step S29, the variable id, that is, the phrase id of the currently processed clause is added to A [link] for the variable link. Here, A ['3'] = ['0']. This indicates that the 0th phrase “Taro ga” is a commutative particle phrase that modifies the third phrase “eat”.

  Returning to step S22, the next line “* 1 3D” is read, an affirmative determination is made in step S23, and the process proceeds to S24. Since the variable id is '0' and A [link] is A ['3'] = ['0'], the process proceeds to step S26, a negative determination is made, and the process proceeds to step S27. In step S27, id = '1' and link = '3' are set, and the first phrase “rotary sushi is” is processed.

  With respect to the next line and the next line, the process returns to step S22 again through steps S22, S23, and S28, and the next line “particle, case particle,...” Is read. Then, after step S23, it is determined as a commutative particle in S28, and id = '1' is added to A ['3'] in step S29. Therefore, A ['3'] becomes A ['3'] = ['0', '1'].

  By repeating the same process and starting from the line “* 2 3D”, id = ‘2’ and link = ’3’ are set, and the second phrase “sushi” is the processing target. Since the commutative particle “O” exists in the second phrase, id = “2” is added to A ['3'], and A ['3'] = ['0', '1', '2 '].

  Next, in the processing of the row “* 3 4D”, id = “3” and link = “4” are set, and the processing of the third phrase “eat” is started. Since there is no commutative particle in the third phrase, it is not added to A ['4'] and remains empty.

  Next, when the line “* 4 6D” is read in step S22, the process passes through S23, and in step S24, id = “3”, so that id is not empty and A [′ 4 ”] is set. Since id = '3' is not stored, the process proceeds to step S25. In step S25, the replacement unit 11 adds ('X' + 'id') to A ['link']. Here, A ['4'] = ['X3']. This indicates that the third clause that modifies the fourth clause is not a commutative particle clause.

  Thus, the commutative particle list A [link] includes information such as 'X3' that is not a commutative particle clause, such as prepositions, adverbs, particles other than commutative particles, and verb conjunctive forms. deep. For example, there may be a clause other than a commutative particle clause between commutative particle clauses that modify the same verb. In this case, A [link] becomes, for example, A [link] = ['4', 'X6', '8']. In this case, if “X” is exchanged across [X8], “X6”, “4”], an unnatural expression may occur. In order to prevent this, information on clauses other than the commutative particle clause that modifies the same verb as the commutative particle clause is also left with a symbol such as 'X'.

  Next, the process proceeds to step S27 through step S26, id = '4' and link = '6' are set, and the fourth phrase becomes the processing target. Since there is a commutative particle “ni” in the fourth phrase, A ['6'] = ['4'].

  Next, in the process starting from the line “* 5 6D”, the fifth clause is a processing target. In the fifth clause, there is a commutative particle “O”, so id = “5” is added to A ['6'], and A ['6'] = ['4', '5']. .

  When the line read in step S22 is the last line (“EOS”), the process proceeds to step S30, and the replacement unit 11 outputs A [link]. Here, A ['3'] = ['0', '1', '2'], A ['6'] = ['4', '5'] is output. Then, the commutative particle list creation process ends, and the process returns to the word order replacement process shown in FIG.

  Next, in step S40 of the word order replacement process shown in FIG. 4, the replacement unit 11 executes the permutation output process. The permutation output process is a process for outputting all permutations of a plurality of commutative particle clauses for each verb that is modified by a plurality of commutative particle clauses. By replacing the commutative particle subtree rooted at the commutative particle clause according to this permutation, replacement of the commutative particle subtree that modifies the same verb is realized.

  Specifically, for example, a permutation of commutative particle clauses can be output using a recursive program as shown in FIG. 8 or a method of creating a permutation of an integer list. Here, as an example, a permutation output process using a method for creating a permutation of an integer list will be described as shown in FIG. The permutation output process shown in FIG. 9 is executed for each verb included in the reference translation. For example, the verbs to be processed can be sequentially searched from the dependency tree root with priority given to the width. All verbs in the reference translation may be processed, but the closer to the root of the dependency tree, the more global replacement occurs, so it is the closest to the root of the dependency tree and is modified to multiple commutative particle clauses. You may limit it to a verb and set it as a processing target. The processes other than step S54 of the permutation output process shown in FIG. 9 are permutation creation processes of the integer list.

  In step S41 of the permutation output process shown in FIG. 9, the replacement unit 11 stores [0] as the initial permutation in permList and sets 1 to the variable k. The variable n is set to the number of ids stored in the commutative particle list A [link] that modifies the verb to be processed. For example, when the verb included in the phrase id = “6” is to be processed, since A [′ 6 ′] = [′ 4 ′, ’5 ′], n = 2. In addition, when information other than a commutative particle clause is included in A [link] like 'X3', it is also included.

  Next, in step S42, the replacement unit 11 determines whether or not k <n. If k <n, the process proceeds to step S43. If k ≧ n, the process proceeds to step S53. In step S43, the replacement unit 11 sets newPermList to be empty, sets 0 to the variable j, and calculates the factorial (kfact) of k. Next, in step S44, the replacement unit 11 determines whether or not j <kfact. If j <kfact, the process proceeds to step S45. If j ≧ kfactk, the process proceeds to step S51.

  In step S45, the replacement unit 11 sets 0 to the variable i. Next, in step S46, the replacement unit 11 determines whether or not i ≦ k. If i ≦ k, the process proceeds to step S47, and if i> k, the process proceeds to step S50. In step S47, the replacement unit 11 creates a permutation pjk in which k is inserted immediately before the i-th element of the j-th permutation of permList. Next, in step S48, the replacement unit 11 adds the created permutation pjk to newPermList.

  Next, in step S49, the replacement unit 11 increments the variable i by 1, and returns to step S46. When a negative determination is made in step S46 and the process proceeds to step S50, the replacement unit 11 increments the variable j by 1, and the process returns to step S44. When a negative determination is made in step S44 and the process proceeds to step S51, the replacement unit 11 replaces the permutation stored in permList with the permutation stored in newPermList. Next, in step S52, the replacement unit 11 increments the variable k by 1, and returns to step S42.

  When a negative determination is made in step S42 and the process proceeds to step S53, the replacement unit 11 outputs the permutation of the integer list stored in permList. Here, [0, 1] and [1, 0] are output. Next, in step S54, the replacement unit 11 associates the output integer of the permutation with the phrase id of the commutative particle list, and outputs the permutation in which the integer of the permutation is replaced with the phrase id. Here, since A ['6'] = ['4', '5'], the 0th clause id is '4' and the first clause id is '5'. Accordingly, ['4', '5'] and ['5', '4'] are output. Then, the permutation output process ends, and the process returns to the word order replacement process shown in FIG.

  Next, in step S60 of the word order replacement process shown in FIG. 4, the restriction unit 12 executes another prescriptive particle restriction process. The pre-subject particle restriction process is a process of eliminating a permutation corresponding to the replacement of a subjunctive sub-partial tree restricted by the above-mentioned pre-subject particle restriction from the output permutations. With reference to FIG. 10, an example of another preferential particle restricting process when the dependency tree shown in FIG. 5 is input will be described. In the example of FIG. 10, in order to eliminate permutations that do not satisfy the “other prescriptive particle restriction”, a phrase that includes a prescription or a phrase that is modified with an expression that includes a prescription is marked in advance. is there.

  In step S <b> 61 of the another prescriptive particle restriction process illustrated in FIG. 10, the restriction unit 12 sets “verb” as the processing target Y. In step S62, the restriction unit 12 sets the variable id corresponding to the clause id and the variable link corresponding to the clause id to be modified to be empty. Also, the mark “Y” [id] is set to be empty. The mark “Y” [id] is “True” when the phrase whose phrase id is “id” includes the part of speech set to Y, and becomes “False” when the phrase is not included. For example, if Y = 'verb' and the sixth clause contains a verb, then 'verb' ['6'] = True.

  Next, in step S63, the next line of the dependency tree is read. Here, since this step is executed for the first time in this routine, the first line of the dependency tree is read.

  Next, in step S64, as in step S23 of the commutative particle list creation process (FIG. 7), it is determined whether the line read in step S63 is the last line or phrase information line. If the determination is affirmative, the process proceeds to step S65. If the determination is negative, the process proceeds to step S71. Here, since the first line is “* 0 3D”, it is determined that it is a phrase information line, and the process proceeds to step S65.

  In step S <b> 65, the limiting unit 12 determines whether the variable id is empty or “True” or “False” is already registered in “Y” [id]. If the determination is affirmative, the process proceeds to step S67. If the determination is negative, the process proceeds to step S66. Here, since the variable id is empty, an affirmative determination is made, and the routine proceeds to step S67.

  In step S <b> 67, the limiting unit 12 determines whether the variable id is empty or “False” has already been registered in “Y” [id]. If the determination is affirmative, the process proceeds to step S69. If the determination is negative, the process proceeds to step S68. Here, since the variable id is empty, an affirmative determination is made, and the routine proceeds to step S69.

  In step S69, the limiting unit 12 determines whether or not the line read in step S63 is the last line. If the determination is affirmative, the process proceeds to step S73. If the determination is negative, the process proceeds to step S70. Here, since the read line is not “EOS”, a negative determination is made and the process proceeds to step S70.

  In step S70, similar to step S27 of the commutative particle list creation process, the phrase id and the qualifying phrase id described in the phrase information line read in step S63 are acquired and set in the variable id and variable link. To do. Here, id = '0', link = '3', and the 0th phrase "Taroga" is the processing target.

  Returning to step S63, the next line "Taro noun, ..." is read, a negative determination is made in step S64, and the flow proceeds to step S71. In step S71, the part of speech of the morpheme read in step S63 is referred to and it is determined whether or not the part of speech of the morpheme is set to 'Y'. If the determination is affirmative, the process proceeds to step S72. If the determination is negative, the process returns to step S63. Here, since the morpheme “Taro” is a noun and not a “verb” set to “Y”, the process returns to step S63. Similarly, for the next line “ga particle,...”, The process returns to step S 63 via steps S 64 and S 71.

  In step S63, the next line “* 1 3D” is read, an affirmative determination is made in step S64, and the process proceeds to S65. Since the variable id is “0” and “Y” [id] is not registered, a negative determination is made and the process proceeds to step S66. In step S66, the restriction unit 12 registers “False” in “Y” [id]. Here, 'verb' ['0'] = True. Then, after steps S67 and S69, the first phrase is set as a processing target in step S70. In the same manner, it is determined that there are no verbs in the first and second clauses.

  Next, returning to step S63, the next line "eat verb, ..." is read. Next, when the process proceeds to step S71 through step S64, since the part of speech of the morpheme read in step S63 is “verb”, an affirmative determination is made and the process proceeds to step S72. In step S <b> 72, the restricting unit 12 registers “True” in “Y” [id]. Here, 'verb' ['3'] = True. Since the next line “ta auxiliary verb,...” Is not a verb, the process returns to step S63 via steps S64 and S71.

  In step S63, the next line “* 4 6D” is read, an affirmative determination is made in step S64, and the process proceeds to step S65. Since the variable id is '3' and 'True' is registered in 'Y' [id], an affirmative determination is made in step S65, a negative determination is made in step S67, and the process proceeds to step S68. In step S <b> 68, the restricting unit 12 registers “True” in “Y” [link]. Here, 'verb' ['4'] = True. This is because there is a verb in the third phrase “eated”, so the subtree “after eaten” that has the fourth phrase “after” modified in the third phrase also contains a verb. Represents that.

  Next, after step S69, the fourth phrase is set as a processing target in step S70. Since “after” and “ni” in the fourth phrase are not verbs, 'verb' ['4'] does not change. Returning to step S63, when the next line “* 5 6D” is read, through steps S64, S65, and S67, the limiting unit 12 registers “True” in “Y” [link] in step S68. . Here, 'verb' ['6'] = True. As a result, the subtree rooted at the sixth phrase “I saw” that is modified by the fourth phrase “after” also before determining whether or not a verb is directly included in “I saw.” Are determined to contain verbs. That is, in the dependency tree shown in FIG. 6, all the clauses on the path from the node corresponding to the third clause “eat” to the root node “saw” are “subtrees rooted in that clause”. Includes a predicate (in this case a verb).

  When the line read in S63 is the last line ("EOS"), the process proceeds to step S73, and the restriction unit 12 outputs 'Y' [id]. As a result of the above processing, “verb” [id] = True is output for a clause including a verb and a clause modified by a clause including a verb, and “verb” [id] = False is output for a clause that is not. The Here, for id = '3', '4', '6', 'verb' [id] = True, for id = '0', '1', '2', '5', 'verb' [ id] = False is output.

  Next, in step S74, the restriction unit 12 determines whether or not the part of speech set as the processing target Y is an 'adjective'. Since the current setting is “verb”, a negative determination is made, the process proceeds to step S75, and the restriction unit 12 sets “adjective” as the processing target Y, and the process returns to step S62. Thereafter, the processing is performed in the same manner as in the case of the 'verb', and it is determined whether the phrase includes an adjective and whether the phrase is modified by a phrase including an adjective or not, and outputs an 'adjective' [id]. Here, as shown in FIG. 5, since the reference translation does not include an adjective, 'adjective' [id] = False is output for all ids.

  Next, in step S77, the restriction unit 12 receives all the permutations output from the permutation unit 11, and for each permutation, a permutation corresponding to the replacement of a commutative particle subtree that satisfies the “other prescriptive particle restriction”. It is determined whether or not. A permutation corresponding to replacement of a subjunction sub-subtree that does not satisfy the “ancillary pre-participant particle restriction” is excluded from the commutative particle list.

  For example, in the example of FIG. 5, when searching for a verb with breadth priority from the root of the dependency tree and paying attention to the sixth phrase “saw”, the commutative particle list A ['6'] = [['4', '5' ], ['5', '4']] are output. The permutation ['4', '5'] is a permutation as in the original reference translation. The permutation ['5', '4'] is a commutative particle part rooted at the 4th phrase, “After Taro eats sushi with carousel,” and the 5th phrase at the root part. This means replacing the tree “Kabuki”.

  Here, 'verb' ['4'] = True, and therefore, the commutative particle subtree rooted at the fourth clause includes a verb. Therefore, if a commutative particle subtree rooted at the fifth clause is placed in front of a commutative particle subtree rooted at the fourth clause, then the commutative particle included in the fifth clause will be “another preface particle”. " In the “adjunct pre-participant restriction”, the pre-additive pre-subject in the case where the pre-subject is a verb is allowed only the commutative particle “ha”, but the commutative particle included in the fifth clause is “a”. And does not meet the pre-participant particle limit. Therefore, the permutation ['5', '4'] is based on the dependency tree in which the commutative particle subtrees are replaced according to this permutation. It is excluded from the commutative particle list A ['6'] before generating the sentence "I saw."

  Further, the commutative particle list includes information such as 'X3' as information on a clause that is not a commutative particle clause in the process of step S25 of the commutative particle list creation processing (FIG. 7). There is a case. Based on this information, the restriction unit 12 straddles a clause other than a commutative particle clause when a clause other than a commutative particle clause is sandwiched between commutative particle clauses that modify the same verb. A permutation corresponding to the replacement is determined, and the permutation is excluded from the commutative particle list.

  The restricting unit 12 outputs the commutative particle list from which the permutation has been removed, ends the pre-junction particle restricting process, and returns to the word order replacing process shown in FIG.

  In step S80 of the word order permutation process shown in FIG. 4, the generation unit 13 is a dependency in which the commutative particle subtree is replaced according to each permutation included in the commutative particle list output from the restriction unit 12 in step S60. Based on each of the trees, each of the word order replacement reference sentences is generated. Then, the generation unit 13 combines the original reference translation sentence and the generated word order replacement reference translation sentence into a reference translation sentence set 30, stores it in a predetermined storage area, and ends the word order replacement process.

  Next, when an evaluation target translation is input to the translation quality evaluation apparatus 20, the translation quality evaluation apparatus 20 executes a translation quality evaluation process shown in FIG.

  In step S90 of the translation quality evaluation process shown in FIG. 11, the calculation unit 21 receives the input evaluation target translation. Further, the calculation unit 21 reads the reference translation set 30 stored in a predetermined storage area. Next, in step S <b> 91, the calculation unit 21 calculates the similarity between the evaluation target translation and each of the plurality of reference translations included in the reference translation set 30.

  Next, in step S92, the evaluation unit 22 obtains and outputs an evaluation result for the evaluation target translation based on the similarity to each of the plurality of reference translations calculated by the calculation unit 21, and performs a translation quality evaluation process. finish.

  As described above, according to the word order exchanging device according to the present embodiment, in the dependency unit of the phrase unit, the commutative particle subtree rooted at the phrase including the commutative particle that modifies the same verb is replaced. Thus, it is possible to generate a plurality of sentences having the same meaning and different word order as the reference translation sentence. It is possible to prevent the generation of an unnatural word order sentence by exchanging permutable particle subtree units.

  In addition, by replacing the commutative particle subtree before the verb different from the verb to be modified by the commutative particle subtree, or by replacing the commutative particle subtree before the adjective, It is possible to further prevent generation of an unnatural word order sentence.

  Further, according to the translation quality evaluation apparatus according to the present embodiment, by using a plurality of reference translations having the same meaning but different word orders, the quality of translation into a language having a high degree of freedom in word order can be appropriately evaluated. Can do. In addition, by using the reference translation generated by the word order replacement device according to the present embodiment, it is possible to increase the reference translation used for evaluation without any manual labor. Many automatic translation quality evaluation tools are configured to accept multiple reference translations, so there is no need to modify these tools, and translation quality of languages with high degree of word order freedom, such as Japanese. It can also handle evaluation.

  Hereinafter, the effects of the present embodiment will be described more specifically with reference to specific examples.

  As shown in FIG. 12, consider a case where a dependency tree indicating a reference translation “Taro ate French cuisine in Paris” is input. As shown in FIG. 12, “I ate” is a parent node, “Taro is”, “In Paris”, and “French cuisine” are child nodes. When a sentence is generated by a post-order from a dependency tree of “Taro ate French cuisine in Paris”, the parent node “eated” is output last. However, since there is an arbitrary order in which the child nodes “Taro ga”, “In Paris”, and “French cuisine” are traced, the following 3! = 6 word orders exist. These six patterns are all valid as Japanese and have the same meaning as the original reference translation.

1. Taro / I ate / French cuisine in Paris.
2. Taro eats / French cuisine / in Paris.
3. In Paris / Taro / French cuisine /
4). In Paris / French cuisine / Taro / ate.
5. I ate French cuisine / Taro / Paris.
6). I ate French cuisine / in Paris / Taro /.

  Each of these is a sentence that can be generated by “replacement of substitutive subpartial trees that modify the same verb (eaten)” in the present embodiment.

  Next, when a dependency tree (FIG. 6) indicating that the reference translation “Taro saw Kabuki after eating sushi with conveyor belt sushi”, also used as a specific example of the above embodiment, is input. think of. In this reference translation, there are two commutative particle clauses that modify the verb “see” at the end of the sentence: “after” and “kabuki wo”, and the commutative particle clause that modifies the verb “eat” is There are three types: “Taro ga”, “In Kaiten Sushi” and “Sushi”. Therefore, the sentence generated by post-order is 2! × 3! = 12 types exist.

1. Taro / at a conveyor belt sushi / eating sushi / after / watching Kabuki.
2. Taro / Sushi / At Kaiten Sushi / Eat / Later / Kabuki / I saw it.
3. At Conveyor Belt Sushi / Taro / Eat Sushi / Later / Kabuki / Saw.
4). At Conveyor Belt Sushi / Sushi / Taro / Eat / Later / Kabuki / Saw.
5. I saw sushi / Taro / at a conveyor belt sushi / after / kabuki.
6). I saw sushi / at conveyor belt sushi / Taro / ate / later / Kabuki / saw.
7). I saw Kabuki / Taro / conveyor sushi / sushi / after / after.
8). I saw Kabuki / Taro / Sushi / Rotating Sushi / Eat / After /.
9. I saw Kabuki / at Kaiten Sushi / Taro / Sushi / Eat / After /.
10. I saw Kabuki / at Kaiten Sushi / Sushi / Taro / Eat / After /.
11. I saw Kabuki, sushi, Taro, sushi, ate, and later.
12 I saw Kabuki / Sushi / Rotating Sushi / Taro / Eat / After /.

  When a sentence is generated by exchanging the order of words in simple phrase units, the meaning is, for example, “Taro / Kabuki / Eat / After / Rotating Sushi / Sushi / I saw”. A sentence that changes or an unnatural sentence may be generated. In the present embodiment, it is possible to eliminate the generation of a sentence whose meaning changes or an unnatural sentence by exchanging in units of substituting particle subtrees.

  Moreover, although said 1-12 are all the head final word order in which a to-be-modified word comes after a modifier, 7-12 is an unnatural sentence. These are because the commutative particle clause “Kabuki” that modifies the verb “see” comes before another verb “eat”. In the present embodiment, such replacement of the commutative particle sub-tree is excluded by “another prescriptive particle restriction”.

  Although it seems that there are a lot of word orders allowed in the present embodiment at first glance, as shown in the example of FIG. There are 5040 ways. Even if you exclude the phrase “I saw” that is determined to be located at the end, the word order of the remaining 6 phrases is 6! There are 720 ways. In the present embodiment, only the six word orders 1 to 6 described above remain, which is only 0.8% with respect to 720 patterns, and it is understood that there are strong restrictions on the allowed word order.

  In addition, for comparison with the present embodiment, consider a case where, when a noun has a plurality of modifiers, the modifiers are replaced to generate a word order replacement sentence. For example, as shown in FIG. 13, when a dependency tree indicating “he bought a white shirt with a large collar” is input, “the shirt is large” and “white” are 2 There are two modifiers. In the sentence generated by the post-order from this dependency tree, there is a sentence "He bought a big shirt with a white collar." In this word order, “white” may be interpreted as modifying the “collar”, but as in this embodiment, when generating a word order permutation sentence by exchanging a commutative particle subtree. , Such a sentence can be prevented from being generated.

  There may also be a problem when the order of sentences having a conjunction at the beginning of a sentence is changed. For example, as shown in FIG. 14, when a dependency tree indicating “and he did not buy the book” is input, “or” is a conjunction indicating addition, but is generated by changing the word order. In the case of the sentence "He has not bought the book again," "also" is an adverb meaning "again." In this embodiment, since a subtree rooted in a phrase including a commutative particle that modifies the same verb is used as a replacement target, such a problematic word order sentence is not generated.

  In addition, the word order switching device of the present invention is not limited to the example of the above embodiment. For example, idiomatic phrases may be registered in advance so that expressions such as idiomatic phrases are not divided into a plurality of clauses, and the order of words that can be used to delimit idiomatic phrases may be restricted. For example, a sentence in which the word order is changed such as “all students become one” and “all students become one” is somewhat unnatural. This is thought to be because the idiom “become” was cut. Therefore, common phrases such as “becomes”, “becomes”, and “falls in” are registered in advance to eliminate the word order in which these expressions are cut and replaced.

  Also, when comparing expressions that modify the same clause, there is a tendency to prefer a word order in which a long modifier is preceded and a short modifier is followed (Reference: Yuko Yamashita, Kimihisa Kondo, “Language Constraints and Long -before-short word order tendency ", IEICE, IEICE Technical Report, TL2011-19, pp.61-65.). Therefore, when the commutative particle subtree is replaced, the lengths of the two may be compared to restrict the shorter commutative particle subtree from coming forward.

  In SMT, the naturalness of a generated sentence is generally evaluated by a monolingual model. Regarding the word order replacement reference translations generated in the present embodiment, the naturalness may be evaluated using a monolingual model, and those whose scores are too low may be deleted from the reference translation set 30.

  Moreover, although the said embodiment demonstrated the case where the word order replacement apparatus 10 and the translation quality evaluation apparatus 20 were comprised by a separate computer, as a translation quality evaluation apparatus 20 which has the function of the word order replacement apparatus 10, one computer is used. You may comprise.

  Although the above-described word order changing device has a computer system therein, the “computer system” includes a homepage providing environment (or display environment) if a WWW system is used.

  Further, in the present specification, the embodiment has been described in which the program is installed in advance. However, the program can be provided by being stored in a computer-readable recording medium.

10 word order replacement device 11 replacement unit 12 restriction unit 13 generation unit 20 translation quality evaluation device 21 calculation unit 22 evaluation unit 30 reference translation set

Claims (6)

A dependency tree indicating a dependency relationship between clauses of a sentence, a subtree consisting of a specific clause ending with a predetermined particle and a clause that modifies the specific clause, and a plurality of portions that modify the same verb A replacement section for replacing each of the trees;
Based on the dependency tree in which the subtree is replaced by the replacement unit, a generation unit that generates a word order replacement sentence in which the word order of the sentence is replaced;
A dependency in which the one subtree is replaced by the replacement unit before the other subtree including a verb different from the verb modified by the one subtree including the predetermined particle other than the particle “ha”. Based on the dependency tree in which the other subtree including the predetermined particle other than the particle “O” is replaced before the one subtree including the tree or the adjective, the word order replacement sentence is generated by the generation unit. A restriction part that restricts not to generate,
Word order replacement device including   2. The word order switching device according to claim 1, wherein the predetermined particles are “ha”, “ga”, “ha”, “ni”, “de”, and “to”. A dependency tree indicating a dependency relationship between clauses of a sentence, a subtree consisting of a specific clause ending with a predetermined particle and a clause that modifies the specific clause, and a plurality of portions that modify the same verb A replacement section for replacing each of the trees;
Based on the dependency tree in which the subtree is replaced by the replacement unit, a generation unit that generates a word order replacement sentence in which the word order of the sentence is replaced;
A dependency in which the one subtree is replaced by the replacement unit before the other subtree including a verb different from the verb modified by the one subtree including the predetermined particle other than the particle “ha”. Based on the dependency tree in which the other subtree including the predetermined particle other than the particle “O” is replaced before the one subtree including the tree or the adjective, the word order replacement sentence is generated by the generation unit. A restriction part that restricts not to generate,
A calculation unit that calculates a similarity between a translation obtained by machine-translating an original sentence, a reference translated sentence that is an ideal translation of the original sentence, and a word-order-replaced reference translated sentence in which the word order is changed without changing the meaning of the reference translated sentence; ,
An evaluation unit that evaluates the translation quality of the translation based on the similarity calculated by the calculation unit,
The translation quality evaluation apparatus which makes the said word order replacement reference translation sentence the word order replacement sentence about the said reference translation sentence produced | generated by the said production | generation part under the restriction | limiting by the said restriction | limiting part . A word order replacement method in a word order replacement device including a replacement unit, a generation unit, and a restriction unit,
The replacement unit is a subtree consisting of a specific clause ending with a predetermined particle and a clause that modifies the specific clause in a dependency tree indicating a dependency relationship between the clauses of the sentence, and the same verb Replacing each of the plurality of subtrees to be modified;
The generation unit generates a word order replacement sentence in which the word order of the sentence is replaced based on a dependency tree in which the subtree is replaced by the replacement unit;
The one subtree before the other subtree including the verb different from the verb modified by the one subtree including the predetermined particle other than the particle “ha” by the restricting unit. Based on the dependency tree in which the other subtree including the predetermined particle other than the particle “O” is replaced before the one of the dependency tree including the adjective or the subtree including the adjective Limiting the generation of the word order replacement sentence according to:
Word order replacement method including   A word order exchanging program for causing a computer to function as each unit constituting the word order exchanging apparatus according to claim 1.   The translation quality evaluation program for functioning a computer as each part which comprises the translation quality evaluation apparatus of Claim 3.

Images