JP2016057815A - Sentence rewrite processing device, learning device, method, and program - Google Patents

Abstract

PROBLEM TO BE SOLVED: To impart consistently character-likeness as a whole sentence.SOLUTION: An operation on morphemes is classified for each morpheme included in an input sentence by a rewrite place estimation unit 226. Determination is made for each of morphemes which are classified into replacement or deletion and which are auxiliary verbs by an auxiliary verb replacement unit 230 as to whether or not to delete the morpheme, the morpheme determined to be deleted is deleted from the input sentence, and for those classified into replacement, an auxiliary verb is inserted into each of positions in the input sentence from which the morphemes have been deleted. A sentence-ending particle is inserted in to each of positions in the input sentence from which the morphemes have been deleted by a sentence-ending particle insertion unit 234. For each of morphemes which are classified into deletion and which are postpositional particles other than the sentence-ending particle, the morphemes are deleted from the input sentence by a postpositional particle replacement unit 236, and for each of morphemes which are classified into replacement and which are postpositional particles other than the sentence-ending particle, the morphemes in the input sentence are replaced with predetermined postpositional particles.SELECTED DRAWING: Figure 5

Description

  The present invention relates to a sentence rewriting processing apparatus, a learning apparatus, a method, and a program, and more particularly, to a sentence rewriting processing apparatus, a learning apparatus, a method, and a program for rewriting a sentence having a certain character character.

  As a conventional technique, there is a character addition by rewriting only for a function word sequence (end of sentence expression) at the end of a sentence. In the prior art, sentence end expressions that are used in a biased manner are extracted for each attribute value of the author using text data to which the attribute of the author is assigned and used for characterizing the utterance (Non-patent Document 1).

Chiaki Miyazaki, Toru Hirano, Ryuichiro Higashinaka, Toshiro Makino, Yoshihiro Matsuo, “Conversion of sentence ending expression to give utterance character”, Artificial Intelligence Society meeting (SIG-SLUD-68), pp. 41-46, 2013.

  However, in the method of Non-Patent Document 1, since only the sentence end expression is a target for rewriting, there may be a mismatch between the language expression appearing in the sentence and the language expression used at the sentence end. For example, if the rewritten sentence is “It's raining today but I don't care”, there is a mismatch in the linguistic expressions used in the sentence “I” and “Yo” at the end of the sentence.

  The present invention has been made to solve the above-described problems, and an object of the present invention is to provide a sentence rewriting processing apparatus, method, and program that can be rewritten into a sentence having a consistent character character as a whole sentence. To do.

  Further, the present invention is made to solve the above problems, and a learning apparatus, method, and method capable of learning a parameter or a model to be given to a sentence having a consistent character character as a whole sentence, And to provide a program.

  In order to achieve the above object, a sentence rewriting processing device according to a first invention is a sentence rewriting processing device that rewrites an input sentence that has been subjected to morphological analysis into a sentence having character-like character representing a specific person image, To classify each morpheme included in the morpheme-analyzed input sentence as a replacement, deletion, or no operation for the morpheme feature value and the previously learned operation on the morpheme for imparting character character. Based on the rewrite location estimation model, the operation for the morpheme is classified as replacement, deletion, or no operation, a rewrite location estimation unit, and the rewrite location estimation unit is classified as replacement or deletion, and an auxiliary verb For each morpheme, write the morpheme to give the morpheme notation and pre-learned character character The morpheme that is determined to be deleted based on the morpheme notation and the rewrite parameter. An auxiliary verb replacement unit that replaces the morpheme determined to be deleted or replaced with a predetermined morpheme or morpheme string, and the position of the input sentence where the morpheme is deleted or replaced by the auxiliary verb replacement unit For each, a final particle insertion unit that inserts a predetermined final particle following the position of the input sentence based on the notation of the morpheme deleted or replaced at the position and the rewriting parameter, and For each morpheme that is classified as deleted by the rewrite location estimator and is a particle other than the final particle, the rewrite parameter Therefore, the morpheme is deleted from the input sentence, classified into replacement by the rewrite location estimation unit, and for each morpheme that is a particle other than a final particle, based on the rewrite parameter, the input sentence And a particle replacement unit that replaces the morpheme with a predetermined particle.

  The sentence rewrite processing method according to the second invention is a morphological analysis input sentence including a rewrite location estimation unit, an auxiliary verb replacement unit, a final particle insertion unit, and a particle replacement unit. A sentence rewriting processing method in a sentence rewriting processing apparatus that rewrites a sentence having character-likeness, wherein the rewriting location estimation unit includes, for each morpheme included in the input sentence after morpheme analysis, a feature quantity of the morpheme, Replacing and deleting the operation for the morpheme based on the rewritten location estimation model for classifying the operation for the morpheme for imparting character likeness as replacement, deletion, or no operation in advance. Or the auxiliary verb replacement unit is classified as replacement or deletion by the rewrite location estimation unit, and the morpheme is an auxiliary verb. Whether the morpheme is deleted, replaced, or not manipulated based on the morpheme notation and a rewriting parameter that is learned in advance and represents a rule for rewriting the morpheme to give the character character And deleting the morpheme determined to be deleted or replacing the morpheme determined to be deleted with a predetermined morpheme or morpheme sequence based on the morpheme notation and the rewriting parameter. The final particle insertion unit is based on the notation of the morpheme deleted or replaced at the position and the rewriting parameter for each position of the input sentence from which the morpheme has been deleted or replaced by the auxiliary verb replacement unit. Then, after the position of the input sentence, a predetermined final particle is inserted, and the particle replacement unit For each morpheme that is a particle other than a final particle, the morpheme is deleted from the input sentence based on the rewrite parameter, and is classified as a replacement by the rewrite location estimator. For each morpheme that is a particle other than a final particle, the morpheme of the input sentence is replaced with a predetermined particle based on the rewriting parameter.

  According to the first and second aspects of the present invention, the rewrite location estimation unit applies a morpheme feature amount and a pre-learned morpheme for imparting character likeness to each morpheme included in the input sentence that has been subjected to morpheme analysis. Based on the rewrite location estimation model for classifying the operation as replacement, deletion, or no operation, the operation for the morpheme is classified as replacement, deletion, or no operation, and the rewrite location estimation unit by the auxiliary verb replacement unit For each of the morphemes that are classified as replacements or deletions, and that are auxiliary verbs, based on the morpheme notation and the rewriting parameters representing the rules for rewriting the morpheme in order to impart character likeness, learned in advance, The morpheme is deleted or replaced with a predetermined morpheme or morpheme string, and the final particle insertion unit is replaced with the auxiliary verb replacement unit. For each position of the input sentence from which the morpheme has been deleted or replaced, a predetermined end following the position of the input sentence based on the notation of the morpheme deleted or replaced at that position and the rewriting parameters. A particle is inserted, and the particle replacement unit is classified as deleted by the rewrite location estimation unit, and for each morpheme that is a particle other than the final particle, the morpheme is deleted from the input sentence based on the rewrite parameter, and the rewrite location For each morpheme that is classified as a replacement by the estimation unit and is a particle other than the final particle, the morpheme of the input sentence is replaced with a predetermined particle based on the rewrite parameter.

  As described above, according to the first and second inventions, for each morpheme included in the morpheme-analyzed input sentence, the operations on the morpheme are classified as replacement, deletion, or no operation based on the rewritten location estimation model. Then, according to the classification result and the rewriting parameter, the morpheme whose part of speech is an auxiliary verb is deleted or replaced with a predetermined morpheme or morpheme sequence, and the morpheme which becomes a final particle is inserted. By deleting the morpheme that becomes or replacing it with another particle, the entire sentence can be rewritten into a sentence having a consistent character character.

  A learning device according to a third aspect of the present invention is a learning device that learns a rewriting parameter that represents a rule for rewriting a morpheme of a sentence in order to give a character a character representing a specific person image, and has the character character Each of the characters included in the character-free sentence in the pair, and each of the characters included in the character-present sentence in the pair. A replacement character string alignment unit that obtains the correspondence relationship between and a decision to determine whether or not to delete a specific auxiliary verb acquired based on the correspondence relationship obtained for each of the pairs by the replacement character string alignment unit Is the corpus composed of pairs of the first rewrite parameter, a sentence with a certain character character and a sentence without it The obtained second rewriting parameter for determining whether or not to replace a specific auxiliary verb, a usable end result acquired from a corpus composed of a pair of a sentence having a character-like character and a sentence having no specific character A third rewriting parameter for determining a particle, a second rewriting parameter for determining whether or not to use a specific particle acquired from a corpus composed of a pair of a sentence having a character-like character and a sentence having no character At least one of four rewriting parameters and a fifth rewriting parameter for determining whether or not to omit a particle acquired based on the correspondence acquired for each of the pairs by the replacement character string alignment unit. And a rewrite parameter acquisition unit that acquires one rewrite parameter.

  A learning method according to a fourth aspect of the present invention includes a replacement character string alignment unit and a rewrite parameter acquisition unit, and a rewrite representing a rule for rewriting a morpheme of a sentence in order to impart character likeness to a sentence to a sentence A learning method in a learning device for learning a parameter, wherein the replacement character string alignment unit is configured for each of a pair of a character-less sentence not having character character and a character-present sentence having character character. A correspondence relationship between each of the characters included in the characterless sentence in the pair and each of the characters included in the character presence sentence in the pair is acquired, and the rewrite parameter acquiring unit is configured to perform the replacement by the replacement character string alignment unit. Whether to delete a specific auxiliary verb acquired based on the correspondence acquired for each pair A first rewriting parameter for determining whether or not to replace a specific auxiliary verb obtained from a corpus composed of a pair of a sentence having a character-like character and a sentence having no character 2 rewriting parameters, a third rewriting parameter for determining a usable final particle obtained from a corpus composed of a pair of a sentence having a character-like character and a sentence having no character-character, having a character-like character A fourth rewrite parameter for determining whether or not to use a specific particle, obtained from a corpus composed of a sentence and a sentence pair, and each of the pairs by the replacement character string alignment unit Less of the fifth rewriting parameters acquired based on the acquired correspondence relationship and used to determine whether to omit the particle Also acquires one rewrite parameters.

  According to the third and fourth inventions, the replacement character string alignment unit includes, for each pair of a character-less sentence having no character character and a character-present sentence having character character, included in the character-less sentence in the pair. The correspondence between each of the characters to be included and each of the characters included in the sentence with characters in the pair is acquired, and the specific auxiliary verb acquired based on the correspondence acquired for each of the pairs is deleted by the rewrite parameter acquisition unit A first rewrite parameter for determining whether to replace or not, to determine whether or not to replace a specific auxiliary verb acquired from a corpus composed of a pair of a sentence having a character-like character and a sentence having no character The second rewrite parameter for, obtained from a corpus consisting of a pair of a sentence with a certain character character and a sentence without A third rewriting parameter for determining a usable final particle, whether or not to use a specific particle obtained from a corpus composed of a pair of a sentence having a character-like character and a sentence having no character At least one rewriting parameter among the fourth rewriting parameter for determining and the fifth rewriting parameter for determining whether or not to omit the particle acquired based on the correspondence acquired for each of the pairs. get.

  Thus, according to the third and fourth inventions, the characters included in the characterless sentence for each pair of the characterless sentence that does not represent character character and the character present sentence that represents character character. And a correspondence relationship between each of the characters included in the sentence with character and the first rewriting parameter, the second rewriting parameter, the third rewriting parameter, the fourth rewriting parameter, and the fifth rewriting. By acquiring at least one rewrite parameter among the parameters, it is possible to determine a parameter for giving the sentence a consistent character character as the whole sentence.

  A learning device according to a fifth aspect of the present invention learns a rewrite location estimation model for classifying operations on a sentence morpheme for giving a character like a specific person image to a sentence as replacement, deletion, or no operation. Each of the characters included in the character-free sentence in the pair, for each of a pair of a character-less sentence not having character character and a character-present sentence having character character, A replacement character string alignment unit that acquires a correspondence relationship with each of the characters included in the character presence sentence in the pair, and for each of the pairs, based on the correspondence relationship acquired by the replacement character string alignment unit, in the pair For each of the morphemes contained in the character-less sentence, it has been rewritten to the character-present sentence in the pair A rewrite location estimation model creation unit that learns the rewrite location estimation model based on each notation of the morpheme to which the label is assigned, as a manipulation for the morpheme And.

  A learning method according to a sixth aspect of the invention includes an operation on a sentence morpheme for adding a character character representing a specific person image to a sentence, including a replacement character string alignment unit and a rewritten location estimation model creation unit. A learning method in a learning device that learns a rewrite location estimation model for classifying as replacement, deletion, or no operation, wherein the replacement string alignment unit includes the character-less sentence and the character that do not have the character character For each pair of characters with a character that has a character, obtain a correspondence relationship between each of the characters included in the character-less sentence in the pair and each of the characters included in the character-present sentence in the pair, The rewrite location estimation model creation unit creates a correspondence relationship acquired by the replacement character string alignment unit for each of the pairs. Therefore, for each morpheme included in the character-free sentence in the pair, a label indicating replacement, deletion, or no operation is given as an operation on the morpheme when the character-in-the-character sentence in the pair is rewritten. The rewriting location estimation model is learned based on the notation of each morpheme with the label.

  According to the fifth and sixth inventions, the replacement character string alignment unit includes, for each pair of a character-less sentence having no character character and a character-present sentence having character character, included in the character-less sentence in the pair. The correspondence between each of the characters to be included and each of the characters included in the sentence with characters in the pair, and based on the correspondence obtained by the replacement character string alignment unit for each of the pairs by the rewrite location estimation model creation unit Then, for each morpheme included in the characterless sentence in the pair, a label indicating substitution, deletion, or no operation was given as an operation on the morpheme when it was rewritten to a charactery sentence in the pair, and the label was given Based on the notation of each morpheme, the rewritten location estimation model is learned.

  Thus, according to the fifth and sixth inventions, for each pair of a character-less sentence not expressing character character and a character-present sentence expressing character character, the characters included in the character-less sentence For each character of the character and each character included in the sentence with the character, and based on the correspondence obtained for each of the pairs, the operation for the morpheme is classified as replacement, deletion, or no operation By learning the rewrite location estimation model, it is possible to learn a model for imparting character consistency to the sentence as a whole sentence.

  Moreover, the program of this invention is a program for functioning a computer as each part which comprises said sentence rewriting processing apparatus and learning apparatus.

  As described above, according to the sentence rewriting processing device, method, and program of the present invention, based on the rewritten location estimation model, for each morpheme included in the morpheme-analyzed input sentence, the operation on the morpheme is replaced, Delete or classify as no operation, delete morpheme whose part of speech is an auxiliary verb according to rewrite parameters, or replace with a predetermined morpheme or morpheme sequence, insert morpheme as final particle, part of speech is other than final particle By deleting a morpheme that is a particle of or substituting it with a specific particle, the sentence as a whole can be rewritten into a sentence having a consistent character character.

  Further, according to the learning apparatus, method, and program of the present invention, it is possible to learn a parameter or model to be given to a sentence having a character character consistent as a whole sentence.

It is a block diagram which shows the functional structure of the learning apparatus which concerns on the 1st Embodiment of this invention. It is a block diagram which shows the functional structure of the rewriting parameter determination part in the learning apparatus which concerns on the 1st Embodiment of this invention. It is a figure which shows the example of the pair to which part of speech has been given. It is a figure which shows the example of a rewriting parameter set. It is a block diagram which shows the functional structure of the sentence rewriting processing apparatus which concerns on the 1st Embodiment of this invention. It is a figure which shows the example which provided the presumed label. It is a figure which shows the example of the processing result of deletion of an auxiliary verb by an auxiliary verb replacement part. It is a figure which shows the example of the processing result of the replacement of an auxiliary verb by an auxiliary verb replacement part. It is a figure which shows the example of the process result of a final particle insertion part. It is a figure which shows the example of the process result of a particle replacement part. It is a flowchart figure which shows the learning process routine in the learning apparatus which concerns on the 1st Embodiment of this invention. It is a flowchart figure which shows the determination process routine of the rewriting parameter in the learning apparatus which concerns on the 1st Embodiment of this invention. It is a flowchart figure which shows the learning process routine of the rewriting location estimation model in the learning apparatus which concerns on the 1st Embodiment of this invention. It is a flowchart figure which shows the sentence rewriting process routine which concerns on the 1st Embodiment of this invention. It is a block diagram which shows the functional structure of the learning apparatus which concerns on the 2nd Embodiment of this invention. It is a block diagram which shows the functional structure of the sentence rewriting processing apparatus which concerns on the 2nd Embodiment of this invention. It is a flowchart figure which shows the learning process routine in the learning apparatus which concerns on the 2nd Embodiment of this invention. It is a flowchart figure which shows the determination process routine of the rewriting parameter in the learning apparatus which concerns on the 2nd Embodiment of this invention. It is a flowchart figure which shows the sentence rewriting process routine based on the 2nd Embodiment of this invention.

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

<Principle of Embodiment of the Present Invention>
First, the principle of the embodiment of the present invention will be described. In the present embodiment, in order to give a consistent character character as a whole sentence, rewriting is performed not only on the sentence end expression but also on function words appearing in the sentence. In the function word, rewriting is performed by focusing on “particles” and “auxiliary verbs” in which character characteristics are easy to appear.

  In preparation for rewriting, a sentence with no character character (hereinafter referred to as a character-less sentence) and a sentence with character character (hereinafter referred to as a character-with-sentence) are used for each character. Information (rewriting parameters) required for rewriting “participant particle, final particle”) “auxiliary verb (mas, is, da)” is acquired in advance. When rewriting an unknown input sentence, a morpheme to be rewritten in the input sentence is determined by a machine learning technique, and rewriting according to the rewriting parameters for the target character is performed.

  For this reason, in the present embodiment, a particular character character is given to the sentence by rewriting the “particle” and “auxiliary verb” of the sentence written in Japanese. Specifically, information necessary for rewriting "particles" and "auxiliary verbs" is acquired from a corpus that consists of sentences with specific character-like sentences and sentences that do not have them, and newly input according to the obtained information Rewrite the sentence. Here, the sentence written in Japanese may be any textual Japanese data such as a speech recognition result of an utterance or a text chat. Further, the “character” here refers to a person image of a speaker. In previous studies in the field of linguistics, examples of human figures include age, gender, occupation, hierarchy, age, appearance, appearance, and personality.

<Configuration of Learning Device According to First Embodiment of the Present Invention>
Next, the configuration of the learning device according to the first embodiment of the present invention will be described. As shown in FIG. 1, a learning device 100 according to an embodiment of the present invention is a computer that includes a CPU, a RAM, and a ROM that stores a program for executing a learning processing routine described later and various data. Can be configured. Functionally, the learning device 100 includes an input unit 10, a calculation unit 20, and an output unit 90 as shown in FIG.

  The input unit 10 accepts each of a pair of a character-free sentence and a character-with-text rewritten so as to manually emphasize a certain character like the character-less sentence. For example, a character-less sentence “There is a delicious sushi restaurant near the station” and a female character's character sentence “There is a delicious sushi restaurant near the station” are accepted. In the first embodiment, it is assumed that a specific character to be processed is determined in advance, and the character ID of the character is stored in a memory (not shown). The character ID is used to identify a rewrite parameter set stored in a rewrite parameter set storage unit 44 described later.

  The calculation unit 20 includes a sentence pair storage unit 22, a rewrite parameter determination unit 30, a rewrite location estimation model creation unit 60, and a rewrite location estimation model storage unit 62.

  The sentence pair storage unit 22 stores a character-less sentence and a character-added sentence pair received by the input unit 10. In the following description, a pair of a sentence without a character and a sentence with a character will be simply referred to as a “pair”.

  The rewriting parameter determination unit 30 determines a rewriting parameter that defines a rule for converting the morpheme notation for rewriting from a character-less sentence to a predetermined character-with-character sentence, and rewriting is performed for each character ID. A rewrite parameter set that is a set of parameters is output to the output unit 90. The rewrite parameter determination unit 30 includes a replacement character string alignment unit 32 and a rewrite parameter acquisition unit 33, as shown in FIG.

  First, the replacement character string alignment unit 32 uses dynamic programming for each of the pairs stored in the sentence pair storage unit 22 and is included in each of the characters included in the character-less sentence and the character-included sentence. The character unit is associated with each character, and the character correspondence in the pair is acquired. Here, as rewrite locations between the pairs, consecutive replacement locations, deletion locations, or insertion locations are collectively connected as one rewrite location. Next, for each pair, a morpheme analysis is performed on each of the character-free sentence and the character-present sentence included in the pair, and the left end is positioned within the range of the rewritten portion for each rewritten portion of the pair. The part of speech of the morpheme to be given is assigned as the part of speech of the rewritten portion, and a part-of-speech assigned pair as shown in FIG. In the example of the part-of-speech added pair in FIG. 3, consecutive rewrite locations such as “ri”, “mas”, “ru”, and “yo” are connected to form one rewrite location. In FIG. 3, “Ri” of “Risuma” and “Ru” of “Ruyo” are regarded as part of the morpheme “A” and “A”, and the left end of the morpheme is included in the range of the rewritten part. Because there is no part of speech, no part of speech is given.

  The rewrite parameter acquisition unit 33 includes an auxiliary verb deletion determination unit 34, an auxiliary verb replacement determination unit 36, a final particle selection unit 38, a particle replacement determination unit 40, a particle omission omission determination unit 42, and a rewrite parameter set storage unit 44. It has.

The auxiliary verb deletion determination unit 34 determines a rule as to whether or not to delete the auxiliary verbs “is” and “mas” in the rewriting of the sentence. Specifically, for each pair with part-of-speech added, the number of characters that have been deleted by rewriting to the character-presented sentence and the character-presented sentence appearing in the no-character sentence included in the pair. The number of times that the character-less sentence is deleted is compared with the sum D _{1 of} the number of times that the character-less sentence is deleted and the total number D ₂ that is not deleted for each of the character-less sentences. If statistically was significantly greater than the sum D ₂ of the number of times the sum D ₁ of the number of times has not been deleted, which is, "we", to determine the "masu" as the "delete", otherwise. Determines “is” and “mas” as “do not delete”, and stores the determined result in the rewrite parameter set storage unit 44 as the first rewrite parameter. In the first embodiment, a chi-square test is used as a statistical test.

The auxiliary verb replacement determination unit 36 determines a rule as to whether or not to replace the auxiliary verb “DA” in the sentence rewriting. Specifically, for each pair with part of speech, the number of occurrences of “Da”, “Mas”, and “Da” appears in each of the characterless sentences included in the pair, and each characterless sentence is counted. to get the total sum E ₁ of the number of times that appear, on the other hand, "it is" to each of the characters there statements that are included in the pair, and count the number of times that the "cell" and "I" appears, appears in each of the characters there sentence to get the sum total E ₂ of the number of times that, compared with the sum total E ₂ of the number of times that appear in the number of times the sum E ₁ and Characterization There sentence of which appears in the character without a statement, the sum E ₂ of the number of times that appear in the character there sentence , if statistically was significantly greater than the sum E ₁ of the number of times that appear in the character without a sentence, it is determined as the replacement to the "I (do it)", in other cases, but "it's (do ) "Is decided not to be replaced, and the result of the decision is As recombinant parameter, and stores the rewritten parameter set storage unit 44. In the present embodiment, chi-square test is used as a statistical test.

The final particle selector 38 determines a final particle rule that is easily used in sentence rewriting. Specifically, for each pair of parts of speech that has been given a part-of-speech, a character-less sentence of the pair is determined for each predetermined final particle (such as na / yo / no / wa / ne) that is likely to be used by a specific character. counting the number of times that the final particle appears, it obtains the sum G ₁ of the number of times the final particle to each of the characters without statement appears, while counting the number of times the final particle appears in character there sentence of the pair Then, the sum G ₂ of the number of times that the final particle appears in each of the sentences with characters is obtained, and the sum G _{1 of} the number of times of appearance in the sentences without characters is compared with the sum G _{2 of} the number of appearances of the sentences with characters. the sum G ₂ times appearing in character there statement, if statistically was significantly greater than the sum G ₁ times appearing in character without sentence, it determines that easily the final particles are used to determine The result is the third rewrite parameter As data, together with the value representing the statistical significance, and stores the rewritten parameter set storage unit 44. In the first embodiment, a chi-square test is used as a statistical test.

The particle replacement determination unit 40 determines a rule as to whether or not to use “te” in sentence rewriting. Specifically, for each pair of parts of speech given, count the number of times “te” appears in the characterless sentence in the pair, and sum the number of times “te” appears in each characterless sentence. H ₁ is acquired, and on the other hand, the number of times “te” appears in the sentence with character included in the pair is counted, and the total number H ₂ of the number of times “te” appears in each sentence with character is obtained. , compared with the sum H ₂ of the number of times that appear in the number of times the sum H ₁ and Characterization There sentence of which appears in the character without a statement, the sum H ₂ of the number of times that appear in the character there statement, the number of times that appear in the character without sentence If the sum statistically was significantly greater compared to H _1, it decides to use the "I", in other cases, decides not to use the "I", the determined result, the fourth As the rewrite parameter of It is stored in the set storage unit 44. In the first embodiment, a chi-square test is used as a statistical test.

The particle omission omission determination unit 42 determines a rule for whether or not omission of particles other than the final particle is omissible in sentence rewriting. Specifically, for each pair of parts of speech, the number of times that each particle other than the final particle that appears in the character-less sentence in the pair has been deleted by rewriting to the character-indicated sentence in the pair is calculated. Count and obtain the sum I _{1 of} the number of times particles other than final particles were deleted by rewriting in each characterless sentence, while each particle other than final particles appearing in the characterless sentence in the pair , Count the number of times that the pair was not deleted by rewriting to the character with sentence, and obtain the sum I _{2 of} the number of times the particles other than the final particle were not deleted by rewriting in each of the character-less sentences comparing the number of times the sum total I ₂ times that were not removed as I _1, and the sum I ₁ of the number of deleted, compared to the total sum I ₂ times that were not removed If it is statistically significant, the particle “can be omitted” is determined. Otherwise, the particle “cannot be omitted” is determined, and the determined result is used as the fifth rewriting parameter. Store in the set storage unit 44. In the first embodiment, a chi-square test is used as a statistical test.

  The rewrite parameter set storage unit 44 includes a rewrite parameter received from the auxiliary verb deletion determination unit 34, the auxiliary verb replacement determination unit 36, the final particle selection unit 38, the particle replacement determination unit 40, and the particle omission omission determination unit 42, and a memory (illustrated). Data that is a set of the character ID stored in (omitted) is stored as a rewrite parameter set of the target character. FIG. 4 shows an example of the rewrite parameter set. In FIG. 4, the auxiliary verbs “is” and “mas” selected by the auxiliary verb deletion determination unit 34 are “deleted” or “deleted” in order from the top for each character ID positioned as an identifier when stored in the rewrite parameter set storage unit 44. Information of “do not”, information of “replace” or “do not” to the auxiliary verb “da” selected by the auxiliary verb replacement determination unit 36, and the final particle selected as the final particle that is easily used by the final particle selection unit 38 Example of storing information, “use” or “do not” information of “te” selected by the particle replacement determination unit 40, and “no omission” and “no omission” of the particles determined by the particle omission omission determination unit 42 In each case, the bold part represents the item selected or determined.

  Based on the correspondence acquired by the replacement character string alignment unit 32, the rewritten location estimation model creation unit 60 assigns each teacher label to each of the replaced morpheme, the deleted morpheme, and the morpheme to which no operation has been added. SUB (-B / I), DEL (-B / I), and O (-B represents the first morpheme at the rewritten location, -I represents a morpheme other than the first), a teacher label, and various A rewritten part to model a sequence of substitution, deletion, and no operation by a machine learning method using feature quantities, and classify operations on morphemes for imparting character character as substitution, deletion, or no operation The estimated model is learned, stored in the rewritten location estimated model storage unit 62, and output to the output unit 90. Specifically, using the conditional random field (CRF) for machine learning in the first embodiment, the morpheme to which the teacher label is assigned and the teacher label as the feature quantity for the teacher label are given. The notation, the part of speech, and the character type of each of the morphemes that are located within the two preceding morphemes and the morphemes that are located within the two following the morphemes given the teacher label are used. In the first embodiment, the character types are kanji (label “a”), hiragana (label “b”), katakana (label “c”), alphabet (label “d”), and numbers (label “e”). ), When a plurality of character types such as kanji and hiragana are mixed in one morpheme, the character type is expressed by a combination of labels such as “ab”. Note that the character type is not an essential feature amount in learning, and may not be used.

<Configuration of sentence rewrite processing device according to first embodiment of the present invention>
Next, the configuration of the sentence rewrite processing device according to the first embodiment of the present invention will be described. As shown in FIG. 5, the sentence rewrite processing device 200 according to the first embodiment of the present invention includes a CPU, a RAM, and a ROM that stores a program and various data for executing a sentence rewrite processing routine to be described later. And a computer including Functionally, the sentence rewriting processing device 200 includes an input unit 210, a calculation unit 220, and an output unit 290 as shown in FIG.

  The input unit 210 accepts a characterless sentence.

  The calculation unit 220 includes a morphological analysis unit 222, a rewrite location estimation model storage unit 224, a rewrite location estimation unit 226, a rewrite parameter set storage unit 228, an auxiliary verb replacement unit 230, a final particle insertion unit 234, and a particle replacement. Part 236.

  The morpheme analysis unit 222 performs morpheme analysis on the characterless sentence received by the input unit 210, and acquires each morpheme of the characterless sentence and each part of speech of the morpheme.

The rewrite location estimation model storage unit 224 stores the same rewrite location estimation model as the rewrite location estimation model stored in the rewrite location estimation model storage unit 62 of the learning device 100.

  The rewrite location estimation unit 226 includes a feature quantity similar to the morpheme feature quantity acquired by the rewrite location estimation model creation unit 60 of the learning device 100 in the morpheme-analyzed characterless sentence acquired by the morpheme analysis unit 222. Each morpheme is extracted and replaced for each morpheme of the characterless sentence based on the feature amount of each morpheme included in the characterless sentence and the rewrite location estimation model stored in the rewrite location estimation model storage unit 224. Estimating whether a morpheme to be deleted (SUB (-B / I)), a morpheme to be deleted (DEL (-B / I)), and a morpheme to which an operation should not be added (O) A label is given to make the input sentence with a rewritten part label. In the first embodiment, as in the learning device 100, Conditional Random Field (CRF) is used for estimation. An example of input / output of the rewrite location estimation unit 226 is shown in FIG.

  The rewrite parameter set storage unit 228 stores the same rewrite parameter set as the rewrite parameter set stored in the rewrite parameter set storage unit 44 of the learning apparatus 100.

  The auxiliary verb replacement unit 230 deletes or separates the auxiliary verbs “is” and “mas” included in the input sentence with the rewritten portion label based on the rewriting parameter set of the specific character stored in the rewriting parameter set storage unit 228. Is replaced with a morpheme or morpheme string. Specifically, for each morpheme whose notation is “is” and “mas” included in the input sentence with the rewrite location label, the estimated label of the morpheme estimated by the rewrite location estimation unit 226 means replacement. If the DEL signifies SUB or deletion, and the rewrite parameter set for a specific character specifies that “is” or “masu” be deleted, as shown in FIG. delete. When deleting “mas”, it is necessary to replace the verb on the left of “mas” with the final form.

  And, the morpheme whose notation of the input sentence with the rewrite location label is “is” and “mas”, and the estimated label of the morpheme estimated by the rewrite location estimation unit 226 means SUB meaning replacement or deletion And the location of the morpheme based on each part of speech of the morpheme immediately before and after the morpheme (for example, the part of speech of the previous morpheme is a noun and the part of speech of the morpheme immediately after is a connective particle) In the grammar where “da” or “da” is required, or when it is specified to replace “da” in the rewrite parameter set of a specific character As shown in FIG. 8, “is” or “mass” in the morpheme portion is replaced with “da” or “da”.

  The final particle insertion unit 234 converts the rewritten parameter set storage unit 228 into a rewritten parameter-labeled input sentence that has been processed based on the rewrite parameter set of a specific character and the processing result of the auxiliary verb replacement unit 230. Insert a final particle. Specifically, the morpheme (notation is “da” or “da”) replaced in the auxiliary verb replacement unit 230 with respect to the morpheme whose notation of the input sentence with the rewritten location label is “is” and “mas”. For each morpheme), when a final particle that can be used is defined in the rewrite parameter set, as shown in FIG. 9, it can be used immediately after “da” or “dan” of the morpheme part. Inserts the final particle with the highest frequency of occurrence in the sentence with characters among the final particles. In addition, for each morpheme in which the notation of the input sentence with the rewrite location label is “is” and “mas” and deleted by the auxiliary verb replacement unit 230 (morpheme with the notation “φ”), rewriting is performed. When a usable final particle is defined in the parameter set, “φ” in the morpheme part is replaced with a final particle having the highest frequency of appearance in the character presence sentence among usable final particles. If the input statement with the rewritten part label is “Da” and “Mas” and the rewriting operation is not performed, the final particle that can be used immediately after “Da” and “Mas” The final particle with the highest frequency of occurrence in characters with characters is inserted.

  The particle replacement unit 236 is a processed result obtained as a result of processing by the auxiliary verb replacement unit 230 and the final particle insertion unit 234 based on the rewrite parameter set of a specific character stored in the rewrite parameter set storage unit 228. The particle included in the input sentence with the rewritten location label is deleted or replaced, and output to the output unit 290 as an output result. Specifically, for each morpheme whose part-of-speech included in the processed input sentence with the rewritten location label is a particle other than the final particle, DEL which means deletion is attached as an estimated label of the morpheme, and When the particle parameter “can be omitted” is defined in the rewriting parameter set of a specific character, the morpheme that is a particle other than the final particle is deleted. In addition, for each morpheme whose part of speech included in the input sentence with the rewrite location label is a particle other than the final particle, a SUB meaning substitution is attached as an estimated label of the morpheme, and the rewrite parameters of a specific character When it is defined that “te” is used for the set, as shown in FIG. 10, the notation of the morpheme that is a particle other than the final particle is replaced with “te”.

<Operation of the learning apparatus according to the first embodiment of the present invention>
Next, the operation of the learning device 100 according to the first embodiment of the present invention will be described. First, in the input unit 10, each of a pair of a character-less sentence and a character-present sentence is stored in the received sentence pair storage unit 22. Then, when each of the pair of the character-free sentence and the character-present sentence is read from the sentence pair storage unit 22, the learning device 100 executes a learning processing routine shown in FIG.

  First, in step S100, based on each of the read pairs, a rewriting parameter that defines a rule for converting a morpheme notation for rewriting from a character-less sentence to a predetermined character-with-character sentence is determined. Then, the data is output to the output unit 90 as a rewrite parameter set for a specific character.

  Next, in step S102, for each pair of parts of speech that has been acquired in step S100, the replaced morpheme, the deleted morpheme, Each morpheme that has not been subjected to an operation is assigned a respective teacher label, and a series of replacement, deletion, and no operation is modeled using a machine learning technique, and a rewrite location estimation model is used as a rewrite location estimation model. While storing in the memory | storage part 62, it outputs to the output part 90, and complete | finishes a learning process routine.

  Step S100 will be described in detail in the rewrite parameter determination processing routine in FIG.

  In step S200 of FIG. 12, for each read pair, using dynamic programming, character units in each of the characters included in the character-free sentence and each character included in the character-added sentence in the pair. And the rewriting part which is the correspondence of the character in the said pair is acquired. When acquiring rewrite locations between the pairs, consecutive replacement locations, deletion locations, or insertion locations are connected together as a single rewrite location.

  Next, in step S202, morphological analysis is performed on the character-less sentence and the character-present sentence in the pair for each read pair, and the part of speech is assigned to each rewritten location in the pair acquired in step S200. As described in the operation of the replacement character string alignment unit 32, the part-of-speech assignment condition is such that morphological analysis is performed on each of the character-free sentence and the character-present sentence included in the pair, , The part of speech of the morpheme whose left end is located within the range of the rewritten part is given as the part of speech of the rewritten part.

  Next, in step S204, based on each of the part-of-speech added pairs acquired in step S202, a rule as to whether or not to delete the auxiliary verbs “is” and “mas” in rewriting a character with a character of a specific character. And the determined result is stored in the rewrite parameter set storage unit 44 as a rewrite parameter.

  Next, in step S206, on the basis of a corpus composed of a pair of a sentence having a character-like character and a sentence having no character, the auxiliary verb “Da” is rewritten in the rewriting of a character with a character. The rule whether or not to replace the auxiliary verbs “is” and “mas” deleted in step S204 is determined, and the determined result is stored in the rewrite parameter set storage unit 44 as a rewrite parameter.

  Next, in step S208, a final particle rule that is likely to be used in rewriting a sentence with a character of a specific character is determined based on a corpus composed of a pair of a sentence with a specific character and a sentence with no specific character. The determined result is stored in the rewrite parameter set storage unit 44 as a rewrite parameter.

  Next, in step S210, based on a corpus composed of a sentence having a character-like character and a sentence having no character-like character, the word “te” is rewritten to a predetermined character-with-character sentence. The rule of whether or not to use is determined, and the determined result is stored in the rewrite parameter set storage unit 44 as a rewrite parameter.

  Next, in step S212, based on each part-of-speech given pair acquired in step S202, a rule is determined as to whether or not a particle can be omitted when rewriting a character with a character. The result is stored in the rewrite parameter set storage unit 44 as a rewrite parameter.

  Next, in step S214, a rewrite parameter set for a specific character is set by setting each of the rewrite parameters acquired in steps S204 to S212 and a character ID for a specific character stored in a memory (not shown). Is stored in the rewrite parameter set storage unit 44.

  Next, in step S216, the rewrite parameter set for the specific character acquired in step S214 is output to the output unit 90, and the rewrite parameter determination processing routine ends.

  Step S102 will be described in detail in the learning process routine for the rewritten location estimation model in FIG.

  In step S300 of FIG. 13, for each pair of parts of speech given in step S100, the replaced morpheme is deleted based on the notation of the character without sentence and the character with sentence in the pair. Each teacher label is assigned to each morpheme and each morpheme that has not been manipulated.

  Next, in step S302, a feature amount is extracted for each morpheme of the character-less sentence in the pair assigned the label in step S300. A specific example of the feature amount has been described in the description of the operation of the rewritten location estimation model creation unit 60.

  Next, in step S304, based on the teacher label for each morpheme included in the character-less sentence in each part-of-speech given pair acquired in step S300 and the feature amount for each morpheme acquired in step S302, A rewrite location estimation model for estimating the rewrite location is learned and stored in the rewrite location estimation model storage unit 62.

  Next, in step S306, the rewrite location estimation model acquired in step S304 is output to the output unit 90, and the rewrite location estimation model learning processing routine ends.

<Operation of sentence rewriting processing device according to first embodiment of the present invention>
Next, the operation of the sentence rewrite processing device 200 according to the first embodiment of the present invention will be described. First, in the input unit 210, the rewrite location estimation model learned in the learning device 100 is received and stored in the rewrite location estimation model storage unit 224. Next, the input unit 210 receives the rewrite parameter set determined by the learning device 100 and stores it in the rewrite parameter set storage unit 228. Then, when the characterless sentence is received in the input unit 210, the sentence rewriting processing device 200 executes a sentence rewriting processing routine shown in FIG.

  First, in step S400, the rewrite location estimation model stored in the rewrite location estimation model storage unit 224 is read.

  Next, in step S <b> 402, a rewrite parameter set for a specific character stored in the rewrite parameter set storage unit 228 is read. The specific character is a character specified by the character ID.

  Next, in step S404, a morphological analysis is performed on the characterless sentence received by the input unit 210.

  Next, in step S406, for each morpheme included in the morpheme-analyzed sentence obtained in step S404, a feature quantity similar to that in step S302 is extracted, and the feature quantity and the rewrite obtained in step S400 are extracted. Based on the location estimation model, an estimated label is assigned and an input sentence with a rewritten location label is acquired.

  Next, in step S408, on the basis of the rewrite parameter acquired in step S402, whether or not to delete each of the morphemes that are represented by “is” and “mas” for the input sentence with the rewrite location label acquired in step S406. Determine and delete the morpheme determined to be deleted.

  Next, in step S410, based on the rewrite parameter acquired in step S402, the input sentence with the rewritten portion label processed in step S408 is replaced with “DA” and “DA” (that is, deleted in step S408). If it is determined to replace the morpheme, it is replaced with the morpheme having the notation “da” or “da”.

  Next, in step S412, on the basis of the rewrite parameter acquired in step S402, it is determined whether to insert a final particle for the input sentence with the rewritten portion label processed in steps S408 to S410, and it is determined to insert it. In this case, an appropriate final particle is inserted at an appropriate predetermined position.

  Next, in step S414, based on the rewrite parameters acquired in step S402, it is determined whether or not to delete or replace particles other than final particles for the input sentence with the rewritten portion label processed in steps S408 to S412. When it is determined to be deleted or replaced, the particles other than the final particle are deleted, or the particle is replaced with “te”.

  Next, in step S416, the input sentence with the rewritten portion label processed in steps S408 to S414 is output from the output unit 290 as a rewrite result from the character-less sentence received in the input unit 210 to the specific character-present sentence. To complete the sentence rewriting routine.

  As described above, according to the learning device according to the first embodiment of the present invention, it is possible to learn a parameter or a model to be given to a sentence having a character character consistent as a whole sentence.

  Further, according to the sentence rewriting processing device according to the first exemplary embodiment of the present invention, the operation for the morpheme is replaced or deleted for each morpheme included in the input sentence after the morpheme analysis based on the rewritten location estimation model. Or, classify as no operation and delete the morpheme whose part of speech is an auxiliary verb according to the rewrite parameter, or replace it with a predetermined morpheme or morpheme sequence, insert the morpheme as final particle, and the part of speech is other than final particle By deleting a morpheme that becomes a particle and replacing it with another particle, the sentence as a whole can be rewritten into a sentence having a consistent character character.

  Moreover, it becomes possible to eliminate the mismatch between the linguistic expression appearing in the sentence and the linguistic expression used at the end of the sentence, which could not be prevented by the method of Non-Patent Document 1 (for example, “Today is raining I do n’t care ”→“ It ’s raining today but I do n’t care ”). As a result, it is possible to give consistent character characteristics as a whole sentence to the utterances of a computer (dialog system) that has a conversation with a human.

  In addition, when the learning device and the sentence rewriting processing device according to the present embodiment are applied to a system that interacts with people (dialogue system), it becomes possible to give character to the utterance of the system, making the dialogue system more human-like. It can be easy to exist.

  In addition, when the learning device and sentence rewriting processing device according to the present embodiment are applied to word-of-mouth and Q & A summaries, the character characteristics that appear in the contents of posts written by multiple people with different dialects, genders, and ages are unified. If it becomes possible to generate a more natural summary sentence that does not notice that it is composed of sentences written by multiple people, and can provide consistent character characteristics as a whole sentence, character characteristics It is possible to further enhance the effect of unifying.

  Note that the present invention is not limited to the above-described embodiment, and various modifications and applications are possible without departing from the gist of the present invention.

  For example, in the first embodiment, the case where the chi-square test is used as a statistical test has been described. However, the present invention is not limited to this, and other ratios such as a binomial test and a log likelihood ratio test are used. An assay method may be used.

In addition, the number of times in the first embodiment, the auxiliary verb "It's (do it)" in the case to determine whether or not to replace the, which appears in the number of times the sum E ₁ and Characterization There sentence of which appears in the character without sentence It has been described for the case of comparing the sum E ₂ of, but is not limited thereto. For example, when rewriting from a character-free sentence to a character-with-sentence, the total number F _{1 of} the number of times “I” and “Masu” have been deleted, and the “I” and “Masu” morphemes are “ When the sum F _{2 of} the number of times of replacement with “Da” is compared, and the sum F _{1 of} the number of times of deletion is statistically significantly larger than the sum of the number of times of replacement F ₂ , You may decide to replace it with “Da”.

  In the first embodiment, the case where CRF is used for machine learning has been described. However, the present invention is not limited to this. For example, any machine learning method may be used as long as it can model a sequence.

  In the first embodiment, a case has been described in which feature amounts are extracted for all five morphemes including each of the two morphemes before and after the morpheme to which a label is assigned. It is not limited to. For example, two consecutive morphemes among the five morphemes including two morphemes before and after the morpheme to which the label is attached may be used as feature quantity extraction targets.

  In the first embodiment, the case where the conditional random field (CRF) is used for the estimation of the rewritten portion has been described. However, the present invention is not limited to this. For example, any method may be used as long as it is a sequence modeling method.

  In the first embodiment, a case has been described in which the final particle to be inserted is a final particle having the largest statistical significance value, but the present invention is not limited to this. For example, a final particle to be inserted at random may be selected from available final particles, or a rule for selecting a final particle to be manually inserted in advance is specified, and insertion is performed based on the rule. A final particle may be selected.

  In the first embodiment, the case where all of the first to fifth rewrite parameters are used as the rewrite parameters has been described. However, the present invention is not limited to this. For example, at least one of the first to fifth rewrite parameters may be used.

  Next, a learning device and a sentence rewriting processing device according to the second embodiment will be described.

  In the second embodiment, the learning apparatus determines a rewriting parameter set for each of a plurality of characters, and the sentence rewriting processing apparatus specifies a character to be given by a character ID in the first embodiment. And different. In addition, about the structure and effect | action similar to the learning apparatus 100 and the sentence rewriting processing apparatus 200 which concern on 1st Embodiment, the same code | symbol is attached | subjected and description is abbreviate | omitted.

<Configuration of Learning Device According to Second Embodiment of the Present Invention>
Next, the configuration of the learning device according to the second embodiment of the present invention will be described. As shown in FIG. 15, a learning device 300 according to an embodiment of the present invention is a computer that includes a CPU, a RAM, and a ROM that stores a program for executing a learning processing routine described later and various data. Can be configured. Functionally, the learning apparatus 300 includes an input unit 310, a calculation unit 320, and an output unit 90 as shown in FIG.

  The input unit 310 accepts each of a pair of a character-without sentence and a character-with-text rewritten so as to manually emphasize the character-like character with respect to the character-less sentence. In the second embodiment, since there are a plurality of target characters, it is assumed that a character ID representing a character of the character presence sentence is assigned to the character presence sentence.

  The calculation unit 320 includes a sentence pair storage unit 322, a rewrite parameter determination unit 330, a rewrite location estimation model creation unit 60, and a rewrite location estimation model storage unit 62.

  The sentence pair storage unit 322 stores each of a character-less sentence and a character-present sentence pair for a plurality of characters received by the input unit 10.

  The rewriting parameter determination unit 330 determines a rewriting parameter that defines a rewriting rule from a sentence without a character to a sentence with a character for each character for each character, and a rewriting parameter set that is a set of rewriting parameters for each character ID. Is output to the output unit 90. Specifically, first, each of the sentence pairs for a plurality of characters stored in the sentence pair storage unit 322 is determined based on the character ID assigned to the character-present sentence included in the sentence pair. Classification is made for each character having the same character ID. Then, similar to the rewrite parameter determination unit 30 of the learning device 100 according to the first embodiment, the processing is performed on the sentence pair set for each classified character, and the character-less sentence for each character is changed to the specific character. A rewriting parameter that defines a rewriting rule for a character-with-character is determined, and a rewriting parameter set that is a set of rewriting parameters for each character ID is output to the output unit 90.

<Configuration of sentence rewrite processing device according to second embodiment of the present invention>
Next, the configuration of the sentence rewrite processing device according to the second embodiment of the present invention will be described. As shown in FIG. 16, a sentence rewrite processing device 400 according to the second embodiment of the present invention includes a CPU, a RAM, and a ROM that stores a program and various data for executing a sentence rewrite processing routine to be described later. And a computer including Functionally, the sentence rewriting processing device 400 includes an input unit 410, a calculation unit 420, and an output unit 290 as shown in FIG.

  The input unit 410 receives the character ID of the character to be given and the no character sentence.

  The calculation unit 420 includes a morphological analysis unit 222, a rewrite location estimation model storage unit 224, a rewrite location estimation unit 226, a rewrite parameter set storage unit 428, an auxiliary verb replacement unit 430, a final particle insertion unit 434, and a particle replacement. Part 436.

  The rewrite parameter set storage unit 428 stores the same rewrite parameter set as the rewrite parameter set for each character ID determined by the learning device 300.

  Hereinafter, in the processing in the auxiliary verb replacement unit 430, the final particle insertion unit 434, and the particle replacement unit 436, the rewrite parameter set having the same character ID as the character ID of the character to be given, acquired in the input unit 410 Since the processing is the same as that of the first embodiment except that is used, detailed description thereof is omitted.

<Operation of the learning apparatus according to the second embodiment of the present invention>
Next, the operation of the learning device 300 according to the second embodiment of the present invention will be described. First, in the input unit 310, each of a pair of a character-less sentence and a character-present sentence for a plurality of characters is stored in the received sentence pair storage unit 322. Then, when each of the pairs of the character-less sentence and the character-present sentence for the plurality of characters assigned with the plurality of character IDs is read from the sentence pair storage unit 322, the learning apparatus 300 performs a learning processing routine shown in FIG. Execute.

  First, in step S500, based on each of a pair of a character-less sentence and a character-present sentence of a specific character for a plurality of read characters, a predetermined specific character is determined from a character-less sentence for each character ID. A rewriting parameter that defines rules for rewriting a character with a character is determined, and is output to the output unit 90 as a rewriting parameter set for a specific character.

  Step S500 will be described in detail in the rewrite parameter determination processing routine in FIG.

  In step S600 of FIG. 18, for each read pair, pairs having the same character ID are collected and classified into pairs for each character ID based on the character ID assigned to the character presence sentence of the pair. From the next step S204 to step S602, it is assumed that processing is performed on the pair set classified as the same character based on the character ID.

  In step S602, each rewrite parameter acquired in steps S204 to S212 and the character ID of the specific character to be processed are set and stored in the rewrite parameter set storage unit 44 as a specific character rewrite parameter set. To do.

  Next, in step S604, it is determined whether or not the processing from step S204 to step S602 has been completed for all characters assigned to the pair read from the sentence pair storage unit 322. If the processing from step S204 to step S602 has been completed for all characters assigned to the pair read from the sentence pair storage unit 322, the process proceeds to step S216, and the pair read from the sentence pair storage unit 322 If the processing from step S204 to step S602 has not been completed for all the assigned characters, the character to be processed is changed, and the processing from step S204 to step S602 is repeated.

<Operation of sentence rewriting processing device according to second embodiment of the present invention>
Next, the operation of the sentence rewrite processing device 400 according to the second embodiment of the present invention will be described. First, in the input unit 410, the rewrite location estimation model learned in the learning device 300 is received and stored in the rewrite location estimation model storage unit 224. Next, the input unit 410 accepts the rewrite parameter set for each character determined by the learning device 300 and stores it in the rewrite parameter set storage unit 428. Then, when the input unit 410 receives the character ID of the character to be given and the character-less sentence, the sentence rewriting processing device 400 executes a sentence rewriting processing routine shown in FIG.

  Next, in step S700, a rewrite parameter set having the same character ID as the character ID of the character to be given and received in the input unit 410, which is stored in the rewrite parameter set storage unit 228, is read. Steps S404 to S416 are the same processing except that the rewrite parameter set read in step S700 is used, and thus detailed description thereof is omitted.

  In addition, about the other structure and effect | action of the class classification apparatus which concern on 2nd Embodiment, since it is the same as that of 1st Embodiment, description is abbreviate | omitted.

  As described above, according to the learning device according to the second embodiment of the present invention, it is possible to learn a parameter or a model to be given to a sentence having a character character consistent as a whole sentence.

  Further, according to the sentence rewriting processing device according to the second exemplary embodiment of the present invention, the operation for the morpheme is replaced or deleted for each morpheme included in the input sentence after the morpheme analysis based on the rewritten location estimation model. Classify as no operation, delete morpheme whose part of speech becomes auxiliary verb according to rewrite parameter, or replace with morpheme or morpheme sequence determined in advance, insert morpheme as final particle, part of speech other than final particle By deleting a morpheme that becomes a particle and replacing it with a specific particle, the sentence can be rewritten into a sentence with a consistent character character.

  Note that the present invention is not limited to the above-described embodiment, and various modifications and applications are possible without departing from the gist of the present invention.

  For example, in the second embodiment, in the learning of the rewrite location estimation model, a case has been described in which the sentence pairs for all characters are comprehensively processed and one rewrite location estimation model is learned. It is not limited to. For example, a rewrite location estimation model may be learned for each character. In this case, in the sentence rewriting processing device, a rewrite location estimation model to be used in the rewrite location estimation unit is determined based on the label given to the characterless sentence to be rewritten. Further, the conversion from the character-free sentence to the character-added sentence can be executed only by the rewrite parameter set even if the rewrite location estimation model does not exist. In this case, it is possible to input from the input unit 210 shown in FIG. 5 an input sentence with a rewritten part label added manually to the character-less sentence, and to perform the processing after the auxiliary verb replacement part 230.

  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 or provided via a network. It is also possible to do.

DESCRIPTION OF SYMBOLS 10 Input part 20 Computation part 22 Sentence pair memory | storage part 30 Rewriting parameter determination part 32 Replacement character string alignment part 33 Rewriting parameter acquisition part 34 Auxiliary verb deletion determination part 36 Auxiliary verb replacement determination part 38 Final particle selection part 40 Adjective replacement determination part 42 Availability determination unit 44 Rewrite parameter set storage unit 60 Rewrite site estimation model creation unit 62 Rewrite site estimation model storage unit 90 Output unit 100 Learning device 200 Sentence rewrite processing device 210 Input unit 220 Operation unit 222 Morphological analysis unit 224 Rewrite site estimation model storage Unit 226 rewrite location estimation unit 228 rewrite parameter set storage unit 230 auxiliary verb replacement unit 234 final particle insertion unit 236 particle replacement unit 290 output unit 300 learning device 310 input unit 320 calculation unit 322 sentence pair storage unit 330 rewrite parameter determination Part 400 sentence rewriting processing apparatus 410 input unit 420 calculating unit 428 rewrites the parameter set storage unit 430 auxiliary verb substitution unit 434 final particle insertion unit 436 particle replacement unit

Claims (8)

A sentence rewriting processing device that rewrites an input sentence that has been subjected to morpheme analysis into a sentence having character-like characteristics representing a specific person image, and for each morpheme included in the input sentence that has been analyzed for morpheme, Replacing and deleting the operation for the morpheme based on the rewritten location estimation model for classifying the operation for the morpheme for imparting character likeness as replacement, deletion, or no operation in advance. Or a rewrite location estimation unit that classifies as no operation,
Rewrite that represents the rules for rewriting the morpheme to give the character-likeness and the notation of the morpheme and the character learned in advance for each of the morphemes that are classified as replacement or deletion by the rewrite location estimation unit and are auxiliary verbs Determining whether to delete or replace or not operate the morpheme based on the parameter, deleting the morpheme determined to be deleted based on the morpheme notation and the rewriting parameter, or An auxiliary verb replacement unit for replacing the morpheme determined to be replaced with a predetermined morpheme or morpheme sequence;
For each position of the input sentence from which the morpheme has been deleted or replaced by the auxiliary verb replacement unit, the position of the input sentence is determined based on the notation of the morpheme deleted or replaced at the position and the rewrite parameter. Subsequently, a final particle insertion part for inserting a predetermined final particle;
For each morpheme that is classified as a deletion by the rewrite location estimation unit and is a particle other than a final particle, the morpheme is deleted from the input sentence based on the rewrite parameter, and replaced by the rewrite location estimation unit For each morpheme that is classified and is a particle other than a final particle, a particle replacement unit that replaces the morpheme of the input sentence with a predetermined particle based on the rewriting parameter;
A sentence rewrite processing device including A learning device that learns a rewriting parameter that represents a rule for rewriting a morpheme of a sentence in order to give a character a character representing a specific person image, the characterless sentence not having the character character and the character character Replacement character string alignment for obtaining a correspondence relationship between each character included in the character-free sentence in the pair and each character included in the character-present sentence in the pair for each pair with a character-with-sentence And
A first rewriting parameter for determining whether or not to delete a specific auxiliary verb acquired based on the correspondence acquired for each of the pairs by the replacement character string alignment unit, a sentence having a specific character character A second rewrite parameter for determining whether or not to replace a specific auxiliary verb, obtained from a corpus composed of a pair of and a sentence not having, a pair of a sentence having a certain character character and a sentence not having A third rewrite parameter for determining a usable final particle obtained from a corpus constituted by a specific corpus obtained from a corpus consisting of a sentence having a character-like character and a sentence having no character A fourth rewriting parameter for determining whether or not to use a particle, and each of the pairs by the replacement character string alignment unit It was obtained based on the acquired correspondence relationship with the rewriting parameter acquisition unit that acquires at least one of rewriting the parameters of the fifth rewrite parameter for determining whether to omit particle,
Including a learning device. A learning device that learns a rewrite location estimation model for classifying operations on a sentence morpheme for imparting character-likeness representing a specific person image to a sentence as replacement, deletion, or no operation, the character-likeness For each of a pair of a character-less sentence having no character and a character-present sentence having character character, each of the characters included in the character-less sentence in the pair and a character included in the character-present sentence in the pair A replacement character string alignment unit for acquiring a correspondence relationship with each of
For each of the pairs, based on the correspondence acquired by the replacement character string alignment unit, for each of the morphemes included in the character-less sentence in the pair, As operations for morphemes, a replacement, deletion, or a label indicating no operation is given, and based on each notation of the morpheme with the label, a rewrite location estimation model creation unit that learns the rewrite location estimation model,
Including a learning device. A sentence rewriting processing device that rewrites an input sentence that has undergone a morphological analysis into a sentence having a character character representing a specific person image, including a rewriting location estimation part, an auxiliary verb replacement part, a final particle insertion part, and a particle replacement part In the sentence rewriting processing method,
For each morpheme included in the input sentence that has undergone the morpheme analysis, the rewrite location estimation unit replaces and deletes the feature amount of the morpheme and an operation on the morpheme that has been learned in advance to give the character likeness Or, based on the rewritten location estimation model for classifying without operation, classifying the operation for the morpheme as replacement, deletion, or no operation,
The auxiliary verb replacement unit is classified into replacement or deletion by the rewrite location estimation unit, and for each morpheme that is an auxiliary verb, the morpheme is given in order to give the character-likeness and the learned morpheme. Based on a rewriting parameter that represents a rule for rewriting the morpheme, whether to delete, replace, or not operate the morpheme, and determined to delete based on the morpheme notation and the rewrite parameter Deleting or replacing a morpheme with a predetermined morpheme or a morpheme sequence,
The final particle insertion unit, for each position of the input sentence where the morpheme has been deleted or replaced by the auxiliary verb replacement unit, based on the morpheme notation deleted or replaced at the position, and the rewriting parameters, Following the position of the input sentence, insert a predetermined final particle,
The particle replacement unit deletes the morpheme from the input sentence based on the rewrite parameter for each morpheme that is classified as a deletion by the rewrite location estimation unit and is a particle other than a final particle, and the rewriting A sentence rewriting method for replacing the morpheme of the input sentence with a predetermined particle based on the rewriting parameter for each morpheme that is classified as a replacement by the location estimation unit and is a particle other than a final particle. A learning method in a learning device that learns a rewriting parameter representing a rule for rewriting a morpheme of a sentence in order to give the sentence character-likeness representing a specific person image, including a replacement character string alignment unit and a rewriting parameter acquisition unit Because
The replacement character string alignment unit, for each of a pair of a character-less sentence not having the character character and a character-like sentence having the character character, each of the characters included in the character-less sentence in the pair, Obtaining a correspondence relationship with each of the characters included in the character presence sentence in the pair;
The rewrite parameter acquisition unit is a first rewrite parameter for determining whether or not to delete a specific auxiliary verb acquired based on the correspondence relationship acquired for each of the pairs by the replacement character string alignment unit, A second rewrite parameter for determining whether or not to replace a specific auxiliary verb, obtained from a corpus composed of a sentence having a specific character character and a sentence having no specific character, a sentence having a specific character character A third rewrite parameter for determining usable final particles, obtained from a corpus that consists of pairs of sentences with and without, composed of pairs of sentences with a particular character character and sentences without A fourth rewriting parameter obtained from the corpus for determining whether or not to use a specific particle, and the replacement character string Imento part by acquired on the basis of the correspondence relation, each acquired for the pair, learning how to get at least one rewriting parameters of the fifth rewrite parameter for determining whether to omit particle. The operations on the sentence morpheme for adding the character character representing the specific person image to the sentence, including the replacement character string alignment unit and the rewritten location estimation model creating unit, are classified as replacement, deletion, or no operation. A learning method in a learning device for learning a rewrite location estimation model for
The replacement character string alignment unit, for each of a pair of a character-less sentence not having the character character and a character-like sentence having the character character, each of the characters included in the character-less sentence in the pair, Obtaining a correspondence relationship with each of the characters included in the character presence sentence in the pair;
The rewritten location estimation model creation unit, for each of the pairs, based on the correspondence acquired by the replacement character string alignment unit, for each morpheme included in the no-character sentence in the pair, A label indicating substitution, deletion, or no operation is assigned as an operation on the morpheme when it is rewritten into a sentence, and the rewritten location estimation model is learned based on the notation of each morpheme with the label. How to learn.   The program for functioning a computer as each part which comprises the sentence rewriting processing apparatus of Claim 1.   The program for functioning a computer as each part which comprises the learning apparatus of Claim 2 or 3.

Images