JP6979294B2 - Calibration support device, calibration support method and calibration support program - Google Patents

Description

The present invention relates to a calibration support device, a calibration support method, and a calibration support program.

With the development of artificial intelligence (AI) in recent years, natural language processing using distributed expressions has attracted attention, and for example, a proofreading support device applying natural language processing has been developed.

Non-Patent Document 1 discloses a technique in which one part of a sentence is blank and a word candidate entering the blank part is predicted by using a distributed expression before and after the blank.

"Conext2vec: Learning Generic Context Embedding with Bidilectional LSTM" Proceedings of the 20th SIGNLL Conference on ComputerLangu

The present inventors have found the possibility that the technique of Non-Patent Document 1 can be applied to a calibration support device, and have conducted various studies. However, in the technology of Non-Patent Document 1, since multiple paraphrase candidates are basically predicted, for example, "east exit" and "south exit" are paraphrase candidates of "no" in the sentence "extending from the west exit of Miyazaki station". Candidates that are inappropriate for calibration, such as being included, are also included. Therefore, if the technique of Non-Patent Document 1 is applied to the calibration support device as it is, there is a problem that the accuracy of calibration is not sufficient and the Japanese becomes unnatural.

The present invention has been made in view of the present situation, and is a calibration support device and a calibration support method capable of selecting an appropriate candidate from the prediction results of calibration candidates using distributed representation and supporting accurate calibration. And to provide a calibration support program.

The calibration support device of the present invention is
A word-separated sentence generator that divides the proofreading target sentence into processing units and generates a word-separated sentence,
Of the processing units that make up the word-separated text, the proofreading target determination unit that determines the processing unit that matches the heading in the proofreading history corpus as the proofreading target.
A calibration candidate prediction unit that predicts calibration candidates using a vector possessed by at least one processing unit or processing unit group before and after the calibration target.
Among the calibration candidates, an appropriate candidate determination unit that determines a calibration candidate that matches the calibration result corresponding to the heading in the calibration history corpus as an appropriate candidate.
It is characterized by having.

Further, the calibration support method of the present invention is:
It is a proofreading support method performed by a computer.
A step to generate a word-separated sentence by dividing the calibrated sentence into processing units and generating a word-separated sentence,
A proofreading target determination step for determining a processing unit that matches a heading in the proofreading history corpus as a proofreading target among the processing units constituting the divided sentence.
A calibration candidate prediction step for predicting a calibration candidate using a vector possessed by at least one processing unit or processing unit group before and after the calibration target.
Among the calibration candidates, an appropriate candidate determination step for determining a calibration candidate that matches the calibration result corresponding to the heading in the calibration history corpus as an appropriate candidate.
It is characterized by having.

Further, the calibration support program of the present invention is
On the computer
A step to generate a word-separated sentence by dividing the calibrated sentence into processing units and generating a word-separated sentence,
A proofreading target determination step for determining a processing unit that matches a heading in the proofreading history corpus as a proofreading target among the processing units constituting the divided sentence.
A calibration candidate prediction step for predicting a calibration candidate using a vector possessed by at least one processing unit or processing unit group before and after the calibration target.
Among the calibration candidates, an appropriate candidate determination step for determining a calibration candidate that matches the calibration result corresponding to the heading in the calibration history corpus as an appropriate candidate.
Is characterized by executing.

According to the present invention, even if the prediction result of the calibration candidate using the distributed representation includes a candidate inappropriate for calibration, a more appropriate candidate can be selected, and more accurate calibration becomes possible. ..

It is a block diagram which shows an example of the structure of the calibration support apparatus of 1st Embodiment. It is a flowchart which shows an example of the calibration support method of 1st Embodiment. It is a block diagram which shows an example of the structure of the calibration history corpus generation part of 1st Embodiment. It is a block diagram which shows an example of the structure of the vector trained model generation part of 1st Embodiment. It is a block diagram which shows an example of the structure of the calibration support apparatus of 2nd Embodiment. It is a flowchart which shows an example of the calibration support method of 2nd Embodiment.

Hereinafter, embodiments of the calibration support device and the calibration support method of the present invention will be described in detail with reference to the drawings. The calibration support device and the calibration support method of the present invention are not limited to the embodiments shown below.

1. 1. 1st Embodiment << Calibration support device >>
FIG. 1 is a block diagram showing an example of the configuration of the calibration support device of the present embodiment. In FIG. 1, 1 is a sentence input unit, 2 is a word-separated sentence generation unit, 3 is a calibration target determination unit, 4 is a calibration candidate prediction unit, 5 is an appropriate candidate determination unit, 6 is an output unit, 8 is a calibration history corpus, and 9 Is a dictionary for morphological analysis, and 10 is a vector-learned model.

<Sentence input unit 1>
First, the proofreading target sentence is input to the sentence input unit 1. The input method is not particularly limited, and examples thereof include keyboard input and handwritten input. The sentence input unit 1 outputs the proofreading target sentence to the word-separated sentence generation unit 2.

<Divided sentence generation unit 2>
The word-separated sentence generation unit 2 divides the proofreading target sentence into processing units and generates a word-separated sentence. For example, a morphological analysis of a calibrated target sentence is performed using a morphological analysis dictionary 9 generated in advance, and the calibrated target sentence is divided into a morpheme and a proper noun. It is preferable that the proper noun is not further divided as long as it is included in the morphological analysis dictionary 9. For example, a person's first and last name is divided into "last name" and "first name" in a general morphological analysis, and the result of "last name / first name" is obtained (hereinafter, "/" indicates a delimiter of a processing unit). However, if the morphological analysis dictionary 9 contains, for example, the surname of a celebrity, the surname is determined to be a proper noun, and the surname is not divided into a "last name" and a "first name", but is used as a processing unit. In this embodiment, the processing unit is, in principle, a morpheme or a proper noun, but a morpheme group in which several consecutive morphemes are combined (for example, "elongation" in which two morphemes before and after "elongation /" are combined) is processed. It may be a unit.

The word-separated sentence generation unit 2 divides the proofreading sentence into processing units, and further arranges a sentence beginning symbol such as <bos> at the beginning of the sentence and a sentence ending symbol such as <eos> at the end of the sentence to generate a divided sentence. .. The word-separated sentence generation unit 2 outputs the generated word-separated sentence to the proofreading target determination unit 3.

The word-separated sentence may be generated by using a method other than morphological analysis. For example, as in the method of "SentencePiece", it may be divided into processing units from the viewpoint of probability or the ease of subsequent processing. Also, for example, if the proofreading target sentence is a language that uses spaces, it is divided into processing units by separating it with spaces, and if the proofreading target sentence is a language that does not use spaces, it is divided into processing units for each character. Etc., the processing unit may be appropriately determined depending on the language to be calibrated and the purpose of proofreading.

<Calibration target determination unit 3>
The proofreading target determination unit 3 compares each of the processing units constituting the divided sentence with the heading of the proofreading history corpus 8 generated in advance, and sets the processing unit matching the heading of the proofreading history corpus 8 as the proofreading target. judge. The order of comparison of processing units is not particularly limited. Here, the calibration history corpus 8 is a database in which past calibration histories are accumulated. For example, as shown in Table 1, the processing unit before calibration is used as a heading, and the calibration result and calibration attributes (insertion, deletion, replacement) are used. ) And the number of calibrations (number of appearances in the past) are recorded. Therefore, the proofreading target determination unit 3 determines the processing unit that has been the proofreading target in the past as the proofreading target among the processing units constituting the divided sentence. The details of the calibration history corpus 8 will be described later.

The calibration target determination unit 3 may compare all the processing units with the headings of the calibration history corpus 8, and as shown in the second embodiment, only the processing units estimated to be error locations by the error location automatic detection unit. May be compared with the heading of the calibration history corpus 8.

The calibration target determination unit 3 outputs the determined calibration target to the calibration candidate prediction unit 4. When there are a plurality of calibration targets, the calibration targets may be output one by one, or a plurality of calibration targets may be output at one time. When outputting a plurality of calibration targets at once, a calibration target storage unit is provided between the calibration target determination unit 3 and the calibration candidate prediction unit 4, and the plurality of calibration targets output from the calibration target determination unit 3 are output. , It may be temporarily held in the calibration target storage unit and output to the calibration candidate prediction unit 4 one by one.

<Proofreading candidate prediction unit 4>
When one of the calibration targets is a blank, the calibration candidate prediction unit 4 sets the processing unit for filling the blank, that is, the calibration candidate at least one before or after the calibration target (blank), preferably both processing units or processing units. Prediction is made using a vector (distributed expression) possessed by a group, for example, a character vector, a word vector, a sentence vector, or the like. The vector of the processing unit can be obtained from, for example, the vector trained model 10 generated in advance. Further, the vector possessed by the processing unit group can be calculated using the vector possessed by the processing unit acquired from the vector trained model 10. Here, the vector-learned model 10 is a machine-learned model of a vector (distributed expression) such as a word vector from a sentence after proofreading that has been proofread in the past in each processing unit and accumulated as a trained model. .. The details of the vector trained model 10 will be described later.

Examples of the prediction method include the following methods. In the calculation of similarity, for example, cosine similarity can be used.
(1) From the vector-learned model 10, each word vector possessed by a plurality of morphemes before and after the blank is acquired, and the average vector of these is calculated. A morpheme having a word vector having a high degree of similarity to the calculated average vector is acquired from the vector-learned model 10, and the morpheme is predicted as a proofreading candidate.
(2) From the vector-learned model 10, each word vector contained in the morpheme group before and after the blank is acquired, and the sentence vector before and after the blank is calculated by using, for example, "context2vc". A morpheme having a word vector having a high degree of similarity to the calculated sentence vector is acquired from the vector-learned model 10, and the morpheme is predicted as a proofreading candidate.

When the processing unit at the beginning or end of the proofreading target is the proofreading target, the vector of the sentence beginning symbol <bos> and the sentence ending symbol <eos> may be used, or only after or before the proofreading target. The vector of the processing unit or the processing unit group may be used.

Further, when predicting a calibration candidate, the calibration candidate prediction unit 4 may predict the calibration candidate by replacing at least one of the other calibration targets with one of the appropriate candidates. In particular, when it is replaced with the optimum candidate, the accuracy of prediction of the calibration candidate is improved, which is preferable.

The calibration candidate prediction unit 4 outputs the predicted calibration candidate to the appropriate candidate determination unit 5. When there are a plurality of calibration candidates, the calibration candidates may be output one by one, or a plurality of calibration candidates may be output at one time. When outputting a plurality of calibration candidates at once, a calibration candidate storage unit is provided between the calibration candidate prediction unit 4 and the appropriate candidate determination unit 5, and a plurality of calibration candidates output from the calibration candidate prediction unit 4 are output. , It may be temporarily held in the calibration candidate storage unit and output to the appropriate candidate determination unit 5 one by one. Further, the calibration candidate prediction unit 4 may output the similarity of the calibration candidates together with the calibration candidates.

<Appropriate candidate determination unit 5>
Among the calibration candidates, the appropriate candidate determination unit 5 determines the calibration candidate that matches the calibration result corresponding to the heading (heading that matches the calibration target) in the calibration history corpus 8 as an appropriate candidate. The appropriate candidate determination unit 5 may determine appropriate candidates for all calibration targets, or may determine appropriate candidates only for calibration targets whose calibration candidates do not include the calibration target itself.

Further, the appropriate candidate determination unit 5 may determine one of the appropriate candidates as the optimum candidate. When there is only one appropriate candidate, the appropriate candidate may be determined as the optimum candidate. The method of determining the optimum candidate when there are a plurality of appropriate candidates is not particularly limited. For example, the similarity of the calibration candidates obtained from the calibration candidate prediction unit 4 and the calibration candidates obtained from the calibration history corpus 8 (matching the calibration candidates). The method of determining the optimum candidate in consideration of the number of calibrations of the calibration result), the part of speech of the calibration target blanked, and the method of determining the optimum candidate by using the width priority search as shown in the second embodiment. Be done.

The appropriate candidate determination unit 5 outputs the determined appropriate candidate to the output unit 6.

<Output unit 6>
The output unit 6 outputs an appropriate candidate in association with the proofreading target sentence. The method of associating with the calibrated sentence is not particularly limited, and examples thereof include the following methods.
(1) Output the proofreading target sentence, the proofreading target, and the appropriate candidate in association with each other.
(2) Output the proofreading target sentence, the proofreading target, the appropriate candidate, and the appropriate degree thereof in association with each other.
(3) Output the proofreading target sentence, the proofreading target, and the optimum candidate in association with each other.
(4) Output a proofread sentence in which the proofreading target is replaced with the optimum candidate. At this time, if the optimum candidate is "<del> ... </ del>", the processing unit is deleted. For example, in the example of Table 1, if the optimum candidate is "<del> soon </ del>", "soon" is deleted. Further, when the number of morphemes constituting the optimum candidate is larger than the number of morphemes constituting the calibration target, the insertion is performed before or after the processing unit. For example, in the example of Table 1, when the calibration target is "2 cases" and the optimum candidate is "2nd case", the "eye" is inserted after the "2 cases".

If there is no calibration target, or if there is a calibration target but there is no appropriate candidate, only the calibration target sentence may be output, or the calibration target or the fact that there is no appropriate candidate may be output together with the calibration target sentence. good.

The output method is not particularly limited, and examples thereof include display on a display and printout.

≪Proofreading support method≫
FIG. 2 is a flowchart showing an example of the calibration support method of the present embodiment.

<Step 1 (S1)>
When the proofreading target sentence is input to the sentence input unit 1, the divided writing sentence generation unit 2 divides the proofreading target sentence into processing units and generates a divided writing sentence. For example, a morphological analysis of a calibrated target sentence is performed using a morphological analysis dictionary 9 generated in advance, and the calibrated target sentence is divided into a morpheme and a proper noun. Further, the word-separated sentence generation unit 2 generates a word-separated sentence by arranging a sentence beginning symbol such as <bos> at the beginning of the sentence and a sentence ending symbol such as <eos> at the end of the sentence.

<Steps 2 and 3 (S2, S3)>
The proofreading target determination unit 3 compares each of the processing units constituting the divided sentence with the heading of the proofreading history corpus 8 generated in advance. The calibration target determination unit 3 may compare all the processing units with the heading of the calibration history corpus 8, and as shown in the second embodiment, the error location automatic detection unit that estimates the error location by machine learning makes an error. Only the processing units presumed to be locations may be compared with the headings of the calibration history corpus 8. The order of comparison of processing units is not particularly limited. When the processing unit does not match the heading of the calibration history corpus 8, the calibration target determination unit 3 compares the next processing unit with the heading of the calibration history corpus 8. On the other hand, if the processing unit matches the heading of the calibration history corpus 8, the processing unit is determined as the calibration target, and the process proceeds to step 4.

<Step 4 (S4)>
When one of the calibration targets is a blank, the calibration candidate prediction unit 4 sets the processing unit for filling the blank, that is, the calibration candidate at least one before or after the calibration target (blank), preferably both processing units or processing units. Prediction is made using a vector (distributed expression) possessed by a group, for example, a character vector, a word vector, a sentence vector, or the like. The method of obtaining the vector possessed by the processing unit or the processing unit group and the specific example of the prediction method using these are as described in the column of "<Calibration candidate prediction unit 4>".

<Steps 5 and 6 (S5 and S6)>
The appropriate candidate determination unit 5 compares each of all the calibration candidates with the calibration result corresponding to the heading (heading matching the calibration target) in the calibration history corpus 8. If the calibration candidate does not match the calibration result, the appropriate candidate determination unit 5 compares the next calibration candidate with the calibration result of the calibration history corpus 8. On the other hand, if the calibration candidate matches the calibration result, the calibration candidate is determined to be an appropriate candidate. Then, when the determination of all the calibration candidates is completed, the process returns to step 2. The appropriate candidate determination unit 5 may determine one of the appropriate candidates as the optimum candidate. In that case, the calibration target is replaced with the optimum candidate, and when the process returns to step 2, the calibration candidate is predicted in step 4. It is preferable because the accuracy of the above is improved.

The appropriate candidate determination unit 5 may compare all the calibration targets with the calibration results of the calibration history corpus 8, or compare only the calibration targets whose calibration targets do not include the calibration target itself with the calibration results of the calibration history corpus 8. May I. In that case, the appropriate candidate determination unit 5 first determines whether or not the calibration target itself is included in the calibration candidate, and if the calibration target itself is included, returns to step 2.

<Step 7 (S7)>
After repeating steps 2 to 6 until all the processing units are processed, the output unit 6 outputs an appropriate candidate in association with the proofreading target sentence.

In the method shown in FIG. 2, each time the calibration target is determined in step 2, the process proceeds from step 3 to step 6, the necessity of calibration of the calibration target is determined, and the process returns to step 2 to determine the next processing unit. However, a method of repeating steps 4 to 6 until all the processing units are processed may be used. Further, when the processing speed is prioritized, the necessity of calibration of all the calibration targets may be determined at the same time without repeating steps 2 to 6.

≪Proofreading history corpus generation≫
FIG. 3 is a block diagram showing an example of the configuration of the calibration history corpus generation unit that generates the calibration history corpus 8 used in the present embodiment. In FIG. 3, 13 is a proofreading database, 14 is a sentence pair acquisition unit, 15 is a sentence pair division word-separated sentence generation unit, and 16 is a proofreading history acquisition unit.

<calibrated database 13>
The proofread database 13 stores the proofread pre-proofreading sentences 12 and the proofreading sentences 11 that have been proofread in the past. The sentences stored in the proofread database 13 are preferably sentences in the same field as or related to the proofreading target sentences. For example, when the proofreading target sentence is a newspaper article, it is preferable that the sentence stored in the proofreading database 13 is a newspaper article.

<Sentence pair acquisition unit 14>
The sentence pair acquisition unit 14 acquires sentence pairs before and after proofreading from the proofread database 13. Specifically, the sentence pair acquisition unit 14 divides each of the pre-proofreading sentence 12 and the post-proofreading sentence 11 stored in the proofreading database 13 into sentences. The method of dividing into sentences is, for example, a method of separating with a punctuation mark when the proofreading target is Japanese and a period when the proofreading target is English, but the method is not limited to these, and the language of the proofreading target and the purpose of the proofreading. It may be selected as appropriate. Then, these sentences are compared to obtain sentence pairs before and after proofreading. The method of obtaining a sentence pair is not particularly limited, but for example, a word vector is obtained from the vector trained model 10, and the similarity between the sentence before proofreading and the sentence after proofreading is calculated using those word vectors. A method of obtaining a sentence pair based on the calculated similarity can be mentioned.

When manually acquiring a pair of sentences before and after proofreading, the sentence pair acquisition unit 14 may not be provided.

<Sentence vs. word-separated sentence generation unit 15>
The sentence before and after proofreading is divided into the sentence before proofreading and the sentence after proofreading into processing units to generate a word-separated sentence. The method of generating a divided sentence is as described in "<Divided sentence generation unit 2>".

<Proofreading history acquisition unit 16>
Use the word-separated sentences to compare the sentence pairs before and after proofreading, and acquire the proofreading history. The method of acquiring the calibration history is not particularly limited, and examples thereof include a method of automatically calculating and acquiring changes before and after calibration using an edit graph.

The calibration history is acquired as a record in which the calibration result and the calibration attributes (insertion, deletion, replacement) are associated with each processing unit before calibration. Here, when the calibration attribute is replacement, the replaced processing unit is the calibration result. When the calibration attribute is deleted, the calibration result is obtained by sandwiching the deleted processing unit with, for example, the deletion symbols "<del>" and "</ del>". If the calibration attribute is insert, the calibration result will include the processing unit before calibration.

When the calibration history is manually acquired, the calibration history acquisition unit 16 may not be provided.

<Proofreading history corpus 8>
The calibration history acquired by the calibration history acquisition unit 16 is stored in the calibration history corpus 8. As described above, for example, as shown in Table 1, a record in which the processing unit before calibration is used as a heading and the calibration result is associated with calibration attributes (insertion, deletion, replacement) and calibration count (past appearance count). Saved as. In the example shown in Table 1, one heading is associated with one calibration result, calibration attribute and combination of calibration times, but of course, one heading is associated with multiple calibration results, calibration attributes and calibrations. A combination of times may be associated. It is preferable that the calibration history corpus 8 is sequentially updated when new calibration history data is obtained from the viewpoint of calibration accuracy.

≪Generation of vector trained model 10≫
FIG. 4 is a block diagram showing an example of the configuration of the vector-learned model generation unit that generates the vector-learned model 10 used in the present embodiment. In FIG. 4, reference numeral 17 is a vector calculation unit.

The vector calculation unit 17 acquires the proofread sentence 11 from the proofread database 13 and divides it into sentences. The method of dividing into sentences is, for example, a method of separating with a punctuation mark when the proofreading target is Japanese and a period when the proofreading target is English, but the method is not limited to these, and the language of the proofreading target and the purpose of the proofreading. It may be selected as appropriate.

The vector calculation unit 17 divides the divided sentence into processing units and generates a divided sentence. The method of generating a divided sentence is as described in "<Divided sentence generation unit 2>". Here, in order to perform proofreading in which the processing unit of the proofreading target sentence is deleted, for example, the deleted processing unit sandwiched between the deletion symbols "<del>" and "</ del>" is also used as the processing unit. The method can be mentioned. For that purpose, it is preferable that the vector calculation unit 17 acquires a pair of sentences before and after proofreading from the proofreading database 13 to acquire a proofreading history, and determines a processing unit based on the acquired proofreading history. The method of acquiring the calibration history is as described in the column of "<< Generation of calibration history corpus >>".

Further, in order to perform proofreading by inserting before or after the processing unit of the proofreading target sentence, or proofreading by replacing or deleting a plurality of consecutive processing units, for example, "n-gram" (combining adjacent n processing units). There is also a method of using the processing unit including the ones that have been used. Specifically, if the sentence obtained by dividing the sentence 11 after proofreading is "According to the prefectural police, it will be the second case in the prefecture." In addition, a combination of multiple morphemes before and after is also included in the processing unit. However, in "n-gram", as n becomes larger, the number of processing units increases, so it is preferable to determine an appropriate value for n by computer resources.
1-gram: "<bos> / prefectural police / ni / by / to /, / prefecture / de / ha / 2 cases / eyes / no / grasp / and / become /. / <eos>"
2-gram: "<bos> To the prefectural police / by / by / and /, within the prefecture / within the prefecture / in / in 2 cases / 2nd case / grasp / grasp / become / become / ./. <Eos>"
3-gram: "<bos> To the prefectural police / According to / According to /, within the prefecture /, within the prefecture / within the prefecture / in 2 cases / is the 2nd case / 2nd case / grasping the eyes / And / become / become / become. / Become. <Eos>"

The vector calculation unit 17 statistically determines the frequency of appearance of processing units, which processing units and processing units are arranged close to each other, and in what context the processing units are used, from the generated word-separated text. Machine learning is performed to obtain a vector (distributed expression) possessed by a processing unit such as a word vector. It is preferable to obtain vectors for various symbols such as the beginning symbol, the ending symbol, and the deletion symbol. For example, "word2vec", "GloVe" and the like can be used to obtain a vector.

≪Specific example≫
Hereinafter, the present embodiment will be further described with reference to specific examples.

<Specific example 1>
Explains an example of replacing one part of the calibration target sentence, specifically, an example of replacing the "system" of the calibration target sentence "The prefectural police set up an investigation headquarters of 100 people at the Kumagaya police station on the 8th." do. The records shown in Table 2 are stored in the calibration history corpus 8 used in this example.

(1) Step 1
The word-separated sentence generation unit 2 performs morphological analysis on the calibrated target sentence output from the sentence input unit 1 using the morphological analysis dictionary 9, and divides it into morphemes, which are processing units. Further, the sentence beginning symbol <bos> and the sentence ending symbol <eos> are arranged at the beginning of the sentence to generate the following divided sentence.
"<Bos> / Prefectural Police / Ha / 8th /, / Kumagaya Police Station / Ni / 100 people / System / / Investigation Headquarters / Established /. / <Eos>"

(2) Steps 2 and 3
The proofreading target determination unit 3 compares each of the morphemes constituting the above-mentioned word-separated sentence with the heading of the proofreading history corpus 8 in order from the beginning of the sentence, and since the "system" matches the heading of the proofreading history corpus 8, it is a proofreading target. And proceed to step 4.

(3) Step 4
When the calibration target "system" is blank, the calibration candidate prediction unit 4 predicts a morpheme that fills the blank as a calibration candidate. Specifically, the morpheme group before and after the calibration target "system""<bos> / prefectural police / on the 8th /, / Kumagaya police station / ni / 100 people""/ investigation headquarters / established /. / <Eos> The word vector of each morpheme included in "" is acquired from the vector-learned model 10, and each sentence vector is calculated by using "context2vc". Then, a morpheme having a word vector having a high degree of similarity to the calculated sentence vector is acquired from the vector-learned model 10, and the morpheme is predicted as a proofreading candidate. The results are shown in Table 3.

(4) Steps 5 and 6
The appropriate candidate determination unit 5 compares each of the calibration candidates with the calibration results shown in Table 2. As a result, since only the "position" matches the calibration result, this is determined as the optimum candidate, the calibration target "system" is replaced with the optimum candidate "position", and the process returns to step 2.

(5) Step 2
The proofreading target determination unit 3 compares the morphemes after the "position" of the word-separated sentence in which the "system" of the above-mentioned word-separated sentence is replaced with the "position" in order with the heading of the calibration history corpus 8. Since there is nothing that matches the heading of the calibration history corpus 8, the process proceeds to step 7.

(6) Step 7
The output unit 6 displays on the display a calibrated sentence "The prefectural police set up an investigation headquarters with 100 people at the Kumagaya police station on the 8th."

<Specific example 2>
An example of deleting one part of the proofreading sentence, specifically, an example of deleting "soon" in the proofreading sentence "A wooden building was soon visible after crossing the Kamitsugawa No. 2 Dam." .. The records shown in Table 4 are stored in the calibration history corpus 8 used in this example.

(1) Step 1
In the same manner as in Specific Example 1, the following word-separated sentence is generated.
"<Bos> / Kamitsugawa No. 2 Dam / Beyond / Exceed / To /, / Soon / Wooden / No / Building / Can be seen / / Kita /. / <Eos>"

(2) Steps 2 and 3
In the same manner as in Specific Example 1, "soon" is determined as the calibration target, and the process proceeds to step 4.

(3) Step 4
Similar to the first embodiment, the calibration candidate prediction unit 4 predicts a morpheme that fills the blank as a calibration candidate when the calibration target “soon” is blank. The results are shown in Table 5.

(4) Steps 5 and 6
The appropriate candidate determination unit 5 compares each of the calibration candidates with the calibration results shown in Table 4. As a result, only "<del> soon </ del>" matches the calibration result, so this is judged as the optimum candidate, and the calibration target "soon" is changed to the optimum candidate "<del> soon </ del>". Replace and return to step 2.

(5) Step 2
The proofreading target determination unit 3 calibrates the morphemes after "<del> soon </ del>" in the word-separated sentence in which "soon" in the above-mentioned word-separated sentence is replaced with "<del> soon </ del>". Contrast with the heading of the history corpus 8. Since there is nothing that matches the heading of the calibration history corpus 8, the process proceeds to step 7.

(6) Step 7
The output unit 6 replaced the proofreading target "soon" with the optimum candidate "<del> soon </ del>", that is, the proofread sentence "Soon" was deleted. I can see the building. "Is displayed on the display.

<Specific example 3>
An example of inserting in one place of the proofreading target sentence, specifically, an example of inserting an "eye" after "2 cases" of the proofreading target sentence "According to the prefectural police, two cases are grasped in the prefecture." do. The records shown in Table 6 are stored in the calibration history corpus 8 used in this example.

(1) Step 1
In the same manner as in Specific Example 1, the following word-separated sentence is generated.
"<Bos> / Prefectural Police / Ni / By / To /, / Prefectural / De / is / 2 cases / / Grasp / To / Become /. / <eos>"

(2) Steps 2 and 3
In the same manner as in Specific Example 1, "2 examples" are determined as calibration targets, and the process proceeds to step 4.

(3) Step 4
Similar to the specific example 1, the calibration candidate prediction unit 4 predicts a morpheme that fills the blank as a calibration candidate when the calibration target “2 examples” is blank. The results are shown in Table 7.

(4) Steps 5 and 6
The appropriate candidate determination unit 5 compares each of the calibration candidates with the calibration results shown in Table 6. As a result, since only the "second case" matches the calibration result, this is determined as the optimum candidate, the calibration target "two cases" is replaced with the optimum candidate "second case", and the process returns to step 2.

(5) Step 2
The proofreading target determination unit 3 compares the morphemes after the "second example" of the divided word sentence in which the "two examples" of the above divided sentence are replaced with the "second example" in order with the heading of the calibration history corpus 8. Since there is nothing that matches the heading of the calibration history corpus 8, the process proceeds to step 7.

(6) Step 7
The output unit 6 replaced the proofreading target "2 cases" with the optimum candidate "2nd case", that is, the proofread sentence "eyes" inserted after "2 cases", "according to the prefectural police, the second case in the prefecture". Will be grasped. ”Is displayed on the display.

<Specific example 4>
An example of proofreading one part of the sentence to be proofread but not proofreading, specifically, proofreading the "prefectural police" in the proofreading sentence "The prefectural police set up a 100-person investigation headquarters at the Kumagaya police station on the 8th." An example in which the target is determined but the appropriate candidate is not determined and the calibration is not performed will be described. The records shown in Table 8 are stored in the calibration history corpus 8 used in this example.

(1) Step 1
In the same manner as in Specific Example 1, the following word-separated sentence is generated.
"<Bos> / Prefectural Police / Ha / 8th /, / Kumagaya Police Station / Ni / 100 people / Position / No / Investigation Headquarters / Established /. / <Eos>"

(2) Steps 2 and 3
In the same manner as in Specific Example 1, the "prefectural police" is determined as the calibration target, and the process proceeds to step 4.

(3) Step 4
Similar to the first embodiment, the calibration candidate prediction unit 4 predicts a morpheme that fills the blank as a calibration candidate when the calibration target “prefectural police” is blank. The results are shown in Table 9.

(4) Step 5
The appropriate candidate determination unit 5 determines whether or not the calibration candidate does not include the calibration target “prefectural police”. In this example, since the calibration candidates shown in Table 9 include "prefectural police", the process returns to step 2.

(5) Step 2
The proofreading target determination unit 3 compares the morphemes behind the "prefectural police" in the above-mentioned word-separated sentence with the headings of the proofreading history corpus 8 in order. Since there is nothing that matches the heading of the calibration history corpus 8, the process proceeds to step 9.

(6) Step 9
The output unit 6 displays the calibration target sentence "The prefectural police set up an investigation headquarters with 100 personnel at the Kumagaya police station on the 8th."

<Specific example 5>
An example of proofreading multiple parts of the proofreading sentence, specifically, the proofreading sentence "The section between Tokyo (Shinagawa) and Nagoya will open in 2027, and the distance from Nagoya to Osaka will be extended to 45 years." The section between Tokyo (Shinagawa) and Nagoya will open in 2027, and the distance from Nagoya to Osaka will be extended to 45 years. " The records shown in Table 10 are stored in the calibration history corpus 8 used in this example.

(1) Step 1
In the same manner as in Specific Example 1, the following word-separated sentence is generated.
"<Bos> / Tokyo / (/ Shinagawa /) / and / Nagoya / no / ma / ha / 2027 / ni / opening /, / Nagoya / to / Osaka / to / ha / 45 years / ni / mo / growth / Schedule / Da /. / <eos>"

(2) Steps 2 and 3
In the same manner as in Specific Example 1, "," (a reading point) is determined as a proofreading target, and the process proceeds to step 4.

(3) Step 4
Similar to the first embodiment, the calibration candidate prediction unit 4 predicts a morpheme that fills the blank as a calibration candidate when the calibration target “,” is blank. The results are shown in Table 11.

(4) Steps 5 and 6
The appropriate candidate determination unit 5 compares each of the calibration candidates with the calibration results shown in Table 10. As a result, since only "shi," matches the calibration result, this is determined as the optimum candidate, the calibration target "," is replaced with the optimum candidate "shi,", and the process returns to step 2.

(5) Steps 2 and 3
The proofreading target determination unit 3 compares the morphemes after the "shi," in the word-separated sentence in which "," in the above-mentioned word-separated sentence is replaced with "shi," in order with the heading of the proofreading history corpus 8, and "mo". Is determined as a calibration target, and the process proceeds to step 4.

(6) Step 4
Similar to the first embodiment, the calibration candidate prediction unit 4 predicts a morpheme that fills the blank as a calibration candidate when the calibration target “mo” is blank. The results are shown in Table 12.

(7) Steps 5 and 6
The appropriate candidate determination unit 5 compares each of the calibration candidates with the calibration results shown in Table 10. As a result, only "<del> also </ del>" matches the calibration result, so this is determined as the optimum candidate, and the calibration target "mo" is changed to the optimum candidate "<del> also </ del>". Replace and return to step 2.

(8) Steps 2 and 3
In the proofreading target determination unit 3, the word-separated sentence "<del> is also </ del>" in which "," is replaced with "," and "mo" is replaced with "<del> and </ del>". The morphemes after "del>" are sequentially compared with the headings of the calibration history corpus 8, and "elongation" is determined as the calibration target, and the process proceeds to step 4.

(9) Step 4
Similar to the first embodiment, the calibration candidate prediction unit 4 predicts a morpheme that fills the blank as a calibration candidate when the calibration target “stretches” is blank. The results are shown in Table 13.

(10) Steps 5 and 6
The appropriate candidate determination unit 5 compares each of the calibration candidates with the calibration results shown in Table 10. As a result, only "extension" matches the calibration result, so this is determined as the optimum candidate, the calibration target "elongation" is replaced with the optimum candidate "extension", and the process returns to step 2.

(11) Step 2
In the proofreading target determination unit 3, the word-separated sentence in which "," is replaced with "," and "mo" is replaced with "<del> also </ del>" is replaced with "extended". The morphemes after "extend" in are sequentially compared with the headings of the calibration history corpus 8. Since there is nothing that matches the heading of the calibration history corpus 8, the process proceeds to step 7.

(12) Step 7
The output unit 6 displays a proofread sentence "The section between Tokyo (Shinagawa) and Nagoya will open in 2027, and the distance from Nagoya to Osaka will be extended to 45 years."

<Specific example 6>
An example of generating the calibration history corpus 8 used in Specific Examples 1 to 3 will be described.

[Specific Example 6-1 (Calibration history corpus 8 used in Specific Example 1)]
The sentence pair acquisition unit 14 divides each of the pre-proofreading sentence 12 and the post-proofreading sentence 11 stored in the proofreading database 13 by kuten and divides them into sentences. Then, a word vector is acquired from the vector-learned model 10, the similarity between the sentence before proofreading and the sentence after proofreading is calculated using those word vectors, and based on the calculated similarity, before and after the following proofreading. Get a pair of sentences.
Sentence before proofreading: "One vehicle system every 40 minutes on weekdays, two vehicles two vehicles every 30 minutes on weekends and holidays."
Sentence after proofreading: "One head and one vehicle are prepared every 40 minutes on weekdays, and two heads and two vehicles are prepared every 30 minutes on weekends and holidays."

The sentence-to-separate word-separated sentence generation unit 15 performs morphological analysis on each of the pre-calibration sentence and the post-calibration sentence using the morphological analysis dictionary 9, divides the sentence into morphological elements, which are processing units, and generates a sentence-to-separate word-separated sentence. do. The proofreading history acquisition unit 16 compares the divided sentences of the sentences before and after the proofreading of the sentence-to-divided sentence by using the edit graph, and acquires the proofreading history that the "system" is replaced with the "position" once. Then, as shown in Table 2, the processing unit "system" before calibration is used as a heading, and the record is stored in the calibration history corpus 8 as a record in which the calibration result "position", the calibration attribute "insertion", and the number of calibrations are associated with each other. .. The number of calibrations is updated every time the calibration history that the "system" is replaced with the "position" is acquired.

[Specific Example 6-2 (Calibration history corpus 8 used in Specific Example 2)]
In the same manner as in Specific Example 6-1 the following pair of sentences before and after proofreading is acquired.
Pre-proofreading sentence: "The fire was extinguished after about four and a half hours, but he was burned to the hospital and died shortly thereafter."
Post-proofreading sentence: "The fire was extinguished after about four and a half hours, but the whole body was burned and taken to the hospital, and <del> soon </ del>died."

The sentence-to-partition word-separated sentence generation unit 15 generates a sentence-to-partition word-separated sentence in the same manner as in the specific example 6-1. The calibration history acquisition unit 16 displays the calibration history that "soon" was replaced with "<del> soon </ del>" once, that is, "soon" was deleted, in the same manner as in Specific Example 6-1. Obtained and stored in the calibration history corpus 8 as a record shown in Table 4.

[Specific Example 6-3 (Calibration history corpus 8 used in Specific Example 3)]
In the same manner as in Specific Example 6-1 the following pair of sentences before and after proofreading is acquired.
Sentence before proofreading: "According to the prefectural police, there are two cases in the prefecture."
Post-proofreading sentence: "According to the prefectural police, this is the second case in the prefecture."

The sentence-to-partition word-separated sentence generation unit 15 generates a sentence-to-partition word-separated sentence in the same manner as in the specific example 6-1. The calibration history acquisition unit 16 calibrates that "2 cases" are replaced once with "2nd case", that is, "eyes" are inserted after "2 cases" in the same manner as in Specific Example 6-1. The history is acquired and stored in the calibration history corpus 8 as a record shown in Table 6.

2. 2. 2nd Embodiment << Calibration support device >>
FIG. 5 is a block diagram showing an example of the configuration of the calibration support device of the present embodiment. In FIG. 5, the same components as those in the first embodiment are designated by the same reference numerals, and are the same as those in the first embodiment unless otherwise specified. In FIG. 5, reference numeral 7 is an error location automatic detection unit.

<Error location automatic detection unit 7>
The error location automatic detection unit 7 estimates an error location (a location that is grammatically incorrect and should be corrected) by machine learning for each of the processing units constituting the divided sentence.

The method of inferring the error location is not particularly limited, and for example, "Liu, Zhuoran, and Yang Liu." Exploiting Unlabeled Data for Natural Grammatic Error Detection. Known "grammatical error detection" methods such as those described in "arXiv preprint arXiv: 1611.08987 (2016)" can be used.

Specifically, for example, it is composed of the following three-layer neural network, and when a sentence is input, labels 0 to 4 (0: do not change, 1: replace, 2: delete, 3: (before the word). There is a method of outputting an array of input sentence lengths with () inserted).
Input layer: Vector of input sentence (sentence that may contain errors) Intermediate layer: Vector mapped to lower dimension of input sentence Output layer: Difference between correct sentence (sentence manually calibrated) and input sentence Vector of what is expressed as in the example below

More specifically, when the word-separated sentence "<bos> / criminal / ha / fence / is / rides / crosses / from / invades / / and / sees / is /. / <Eos>" is input, The vector "0,0,0,1,0,1,1,2,3,0,0,1,1,0,0" is output. This output means that "fence", "ride", "beyond", "mi", and "reru" can be replaced, "from" can be deleted, and "intrusion" can be inserted before that.

An error from the output may be calculated from the vector of the output layer using a loss function based on the Softmax function, and the optimum parameters of the neural network may be predicted from the error by the error back propagation method. Further, the type of label is not particularly limited, and for example, two types of labels of 0 and 1 (indicating whether or not it is an error location) may be used.

The error location automatic detection unit 7 outputs the processing unit estimated to be an error location to the calibration target determination unit 3 together with the label, preferably. When there are a plurality of processing units presumed to be an error location, the processing targets may be output one by one, or a plurality of processing units may be output at one time. Further, when a plurality of processing units are output at one time, a processing unit storage unit is provided between the calibration target determination unit 3 and the plurality of processing units output from the error location automatic detection unit 7 to be processed once. It may be held in the unit storage unit and output to the calibration target determination unit 3 one by one.

<Calibration target determination unit 3>
The calibration target determination unit 3 determines the calibration target by comparing only the processing unit estimated by the error location automatic detection unit 7 as the error location with the heading of the calibration history corpus 8. Therefore, there is a possibility that the processing speed will be improved as compared with the case where the calibration target is determined for all the processing units constituting the divided sentence.

Further, the calibration target determination unit 3 may use the label output by the error location automatic detection unit 7 to concatenate continuous processing units and treat them as one processing unit (n-gram). For example, in the above example, the labels "1" (replacement) output by the error location automatic detection unit 7 are continuous for "ride" and "beyond", and "mi" and "reru", respectively. One word (2-gram) "overcoming" and "seen" in which both are connected may be treated as a processing unit.

<Appropriate candidate determination unit 5>
The appropriate candidate determination unit 5 may use the label output by the error location automatic detection unit 7 as the optimum candidate determination material. For example, in the above example, when there are a plurality of appropriate candidates for "kara", the label "2" (deleted) output by the error location automatic detection unit 7 is added, and "<del> to </ del>" is added. It may be determined as the optimum candidate.

≪Proofreading support method≫
FIG. 6 is a flowchart showing an example of the calibration support method of the present embodiment.

<Step 11 (S11)>
The word-separated sentence generation unit 2 generates a word-separated sentence in the same manner as in step 1 of the first embodiment.

<Steps 12 to 14 (S12 to S14)>
Except for the point that the error location automatic detection unit 7 determines the calibration target only for the processing unit estimated to be the error location, and the point that the calibration target is determined for all the processing units estimated to be the error location before proceeding to the next step. 1 The calibration target determination unit 3 determines the calibration target in the same manner as in steps 2 and 3 of the embodiment.

<Steps 15 to 19 (S15 to S19)>
In this example, the optimum candidate was determined by incorporating a probabilistic model that takes context into consideration. Hereinafter, the determination of the optimum candidate by the combination of breadth-first search such as beam search and calibration history corpus will be described. Breadth-first beam search is classified as knowledgeable search. ..

Hereinafter, the calibration targets are the first calibration target, the second calibration target ... Nth calibration target (N is an integer of 2 or more) in order from the beginning of the sentence. Further, the appropriate candidates for one calibration target are the first appropriate candidate, the second appropriate candidate ... the M appropriate candidate (M is an integer of 1 or more).

First, an appropriate candidate for the first calibration target is determined by the following procedure.

[Step 15]
The calibration candidate prediction unit 4 predicts the calibration candidate to be calibrated with a predetermined search width in the same manner as in step 4 of the first embodiment. At that time, for example, the similarity obtained by "Conextext2Vec", the log-likelihood obtained from the similarity, and the like are obtained as scores.

[Steps 16 and 17]
The appropriate candidate determination unit 5 determines an appropriate candidate for the first calibration target in the same manner as in steps 5 and 6 of the first embodiment.

Next, in the following procedure, the appropriate candidate for the second proofreading target is determined for the word-separated sentence in which the first proofreading target is replaced with the first appropriate candidate.

[Step 18]
The appropriate candidate determination unit 5 replaces the first calibration target with the first appropriate candidate, and returns to step 15.

[Step 15]
The calibration candidate prediction unit 4 predicts the calibration candidate to be calibrated in the second calibration target with a predetermined search width and obtains a score in the same manner as in step 4 of the first embodiment.

[Steps 16 and 17]
The appropriate candidate determination unit 5 determines an appropriate candidate for the second calibration target in the same manner as in steps 5 and 6 of the first embodiment.

Next, in the following procedure, the appropriate candidate for the third proofreading target is determined for the word-separated sentence in which the first proofreading target is replaced with the first appropriate candidate and the second proofreading target is replaced with the first appropriate candidate.

[Step 18]
The appropriate candidate determination unit 5 replaces the first calibration target with the first appropriate candidate, replaces the second calibration target with the first appropriate candidate, and returns to step 15.

[Step 15]
The calibration candidate prediction unit 4 predicts the calibration candidate to be calibrated in the third calibration target with a predetermined search width and obtains a score in the same manner as in step 4 of the first embodiment.

[Steps 16 and 17]
The appropriate candidate determination unit 5 determines an appropriate candidate for the third calibration target in the same manner as in steps 5 and 6 of the first embodiment.

After that, steps 15 to 18 are repeated in the same manner, and for the nth calibration target (n is an integer of 2 to N), the first calibration target to the n-1 calibration target are each appropriate candidate (first appropriate candidate to first). The appropriate candidate is determined by replacing it with one of M appropriate candidates), and all combinations of appropriate candidates for each of the first calibration target to the Nth calibration target are obtained.

[Step 19]
The appropriate candidate determination unit 5 totals the scores of the appropriate candidates constituting the combination of the appropriate candidates for each of the first calibration target to the Nth calibration target, obtains the total score for each combination, and constitutes the combination with the smallest total score. Each appropriate candidate is judged to be the optimum candidate.

<Step 20 (S20)>
The output unit 6 outputs the optimum candidate in association with the calibration target sentence, for example, outputting a calibrated sentence in which the calibration target is replaced with the optimum candidate. Further, the output unit 6 may output a combination of appropriate candidates for each of the first calibration target to the Nth calibration target in the order of total scores.

≪Specific example≫
Hereinafter, the present embodiment will be further described with reference to specific examples.

<Specific example 7>
An example of proofreading multiple parts of the proofreading target sentence, specifically, proofreading the proofreading target sentence "The criminal seems to have invaded after overcoming the fence." To "The criminal seems to have invaded after overcoming the fence." An example will be described. The records shown in Table 14 are stored in the calibration history corpus 8 used in this example.

(1) Step 11
The following word-separated sentence is generated in the same manner as in step 1 of the specific example 1.
"<Bos> / Criminal / Ha / Fence / Ride / Cross / From / Invade / Intrude / To / See / Can /. / <eos>"

(2) Step 12
The error location automatic detection unit 7 estimates the error location of the divided sentence by machine learning, and the estimation result "0,0,0,1,0,1,1,2,3,0,0,1,1,0" , 0 "is output. That is, the error location automatic detection unit 7 presumes that there is a possibility that "fence", "ride", "over", "mi", and "being" are replaced, "from" is deleted, and "intrusion" is inserted before that. do.

(3) Steps 13 and 14
Among the morphemes constituting the above-mentioned word-separated sentence, the proofreading target determination unit 3 is a morpheme presumed to be an error part by the error part automatic detection unit 7, that is, "fence", "ride", "beyond", "from", "intrusion", and "intrusion". Each of "mi" and "reru" is compared with the heading of the calibration history corpus 8 in order from the beginning of the sentence. At this time, for "ride" and "over", the label "1" (replacement) output by the error location automatic detection unit 7 is continuous, so one word (2-gram) "overcoming" in which both are connected is continuous. Process as "te". "Mi" and "reru" are also treated as "seen" in the same way. As a result, "fence", "overcoming", "from", and "intrusion" match the headings of the calibration history corpus 8, so these are the first calibration target, the second calibration target, the third calibration target, and the fourth calibration, respectively. It is determined as a target, and the process proceeds to step 15.

(4) Step 15
Similar to step 4 of the specific example 1, the calibration candidate prediction unit 4 predicts a morpheme that fills the blank as a calibration candidate when the first calibration target “fence” is blank. However, the search width was set to 5, and the log-likelihood obtained from the similarity obtained by "Conextext2Vec" was obtained as a score. The results are shown in Table 15.

(5) Steps 16 and 17
The appropriate candidate determination unit 5 compares each of the calibration candidates with the calibration results shown in Table 14. As a result, since the "fence" and "moat" and the "fence" which is the calibration target itself match the calibration result, these are determined as the first appropriate candidate, the second appropriate candidate, and the third appropriate candidate, respectively.

(6) Step 18
The appropriate candidate determination unit 5 replaces the first calibration target “fence” with the first appropriate candidate “fence”, and returns to step 15.

(7) Step 15
In the same manner as in "(4) Step 15", the calibration candidate prediction unit 4 predicts a morpheme that fills the blank as a calibration candidate when the second calibration target "overcoming" is blank. The results are shown in Table 16.

(8) Steps 16 and 17
The appropriate candidate determination unit 5 compares each of the calibration candidates with the calibration results shown in Table 14. As a result, "overcoming", "overcoming", and "jumping over" match the calibration results, and these are determined as the first appropriate candidate, the second appropriate candidate, and the third appropriate candidate, respectively.

(9) Step 18
The appropriate candidate determination unit 5 replaces the first calibration target “fence” with the first appropriate candidate “fence”, replaces the second calibration target “overcoming” with the first appropriate candidate “overcoming”, and returns to step 15.

(10) Step 15
In the same manner as in "(4) Step 15", the calibration candidate prediction unit 4 predicts a morpheme that fills the blank as a calibration candidate when the third calibration target "kara" is blank. The results are shown in Table 17.

(11) Steps 16 and 17
The appropriate candidate determination unit 5 compares each of the calibration candidates with the calibration results shown in Table 14. As a result, "to" and "from <del> to </ del>" match the calibration results, so these are determined as the first appropriate candidate and the second appropriate candidate, respectively. Further, the appropriate candidate determination unit 5 determines the second appropriate candidate “<del> to </ del>” as the optimum candidate in consideration of the label “2” (deleted) output by the error location automatic detection unit 7.

(12) Step 18
The appropriate candidate determination unit 5 replaces the first calibration target "fence" with the first appropriate candidate "fence", replaces the second calibration target "overcoming" with the first appropriate candidate "overcoming", and replaces the third calibration target "from". Is replaced with the optimum candidate "<del> to </ del>", and the process returns to step 15.

(13) Step 15
In the same manner as in "(4) Step 15", the calibration candidate prediction unit 4 predicts a morpheme that fills the blank as a calibration candidate when the fourth calibration target "intrusion" is blank. The results are shown in Table 18.

(14) Steps 16 and 17
The appropriate candidate determination unit 5 compares each of the calibration candidates with the calibration results shown in Table 14. As a result, ", invasion" and "illegal invasion" match the calibration results, and these are determined as the first appropriate candidate and the second appropriate candidate, respectively.

(15) Step 18
The appropriate candidate determination unit 5 replaces the first calibration target "fence" with the first appropriate candidate "fence", replaces the second calibration target "overcoming" with the second appropriate candidate "over", and returns to step 15.

(16) Steps 15-18
Hereafter, all combinations of appropriate candidates are obtained in the same manner. The results are shown in Table 19.

(17) Step 19
The appropriate candidate determination unit 5 obtains the total score for each combination of appropriate candidates, and determines each appropriate candidate constituting the combination having the smallest total score as the optimum candidate. In this example, as shown in Table 19, the first calibration target "fence" is replaced with the first appropriate candidate "fence", the second calibration target "overcoming" is replaced with the first appropriate candidate "overcoming", and the third The total score of the combination in which the calibration target "kara" is replaced with the optimum candidate "<del> to </ del>" and the fourth calibration target "kara" is replaced with the first appropriate candidate ", intrusion" is 1.21 + 1.59 + 1. 62 + 1.58 = 6.00, which is the minimum. Therefore, the first appropriate candidate "fence" for the first calibration target "fence", the first appropriate candidate "overcome" for the second calibration target "overcome", and the optimum candidate for the third calibration target "kara". It is determined that "from <del> to </ del>" and the first appropriate candidate ", intrusion" for the fourth calibration target "from" are the optimum candidates.

(18) Step 20
The output unit 6 displays on the display a proofread sentence "The criminal seems to have climbed over the fence and invaded." With each proofreading target replaced with the optimum candidate. Further, the output unit 6 displays other combinations of appropriate candidates on the display in ascending order of the total score.

1: Sentence input unit 2: Word-separated sentence generation unit 3: Proofreading target determination unit 4: Proofreading candidate prediction unit 5: Appropriate candidate determination unit, 6: Output unit, 7: Error location automatic detection unit, 8: Proofreading History corpus, 9: Morphological analysis dictionary, 10: Vector trained model, 11: Post-proofreading sentence, 12: Pre-proofreading sentence, 13: Pre-proofreading database, 14: Sentence pair acquisition unit, 15: Sentence-paired word-separated sentence generation unit , 16: Calibration history acquisition unit, 17: Vector calculation unit

Claims (10)

A word-separated sentence generator that divides the proofreading target sentence into processing units and generates a word-separated sentence,
Of the processing units that make up the word-separated text, the proofreading target determination unit that determines the processing unit that matches the heading in the proofreading history corpus as the proofreading target.
A calibration candidate prediction unit that predicts calibration candidates using a vector possessed by at least one processing unit or processing unit group before and after the calibration target.
Among the calibration candidates, an appropriate candidate determination unit that determines a calibration candidate that matches the calibration result corresponding to the heading in the calibration history corpus as an appropriate candidate.
A calibration support device characterized by having. Further, it has a vector-learned model, and the calibration candidate prediction unit uses a vector possessed by a processing unit acquired from the vector-learned model or a vector possessed by a processing unit group calculated using the vector possessed by the processing unit. The calibration support device according to claim 1, wherein the calibration candidate is predicted. Further, it has an error location automatic detection unit that estimates an error location by machine learning, and the calibration target determination unit has only a processing unit that is estimated to be an error location by the error location automatic detection unit among the processing units. The calibration support device according to claim 1 or 2, wherein the calibration target is determined. The calibration support device according to any one of claims 1 to 3, wherein the appropriate candidate determination unit determines one of the appropriate candidates as the optimum candidate. The calibration support device according to claim 4, wherein the appropriate candidate determination unit determines an optimum candidate by using breadth-first search. When predicting a calibration candidate for one calibration target, the calibration candidate prediction unit replaces at least one other calibration target with one of the appropriate candidates for the other calibration target to predict the calibration candidate. The calibration support device according to any one of claims 1 to 5, wherein the calibration support device is characterized by the above. The calibration support device according to any one of claims 1 to 6, wherein the appropriate candidate determination unit determines an appropriate candidate when the calibration candidate does not include the calibration target. Further, it has a calibration history corpus generation unit that generates the calibration history corpus.
The calibration history corpus generator
A sentence pair acquisition unit that acquires pairs of sentences before and after proofreading from a proofread database that stores sentences before and after proofreading.
A sentence-to-word-separated sentence generation unit that divides each of the pre-proofreading sentence and the post-proofreading sentence into processing units to generate a word-separated sentence.
A proofreading history acquisition unit that acquires proofreading history by comparing sentence pairs before and after proofreading using the above-mentioned word-separated sentences.
The calibration support device according to any one of claims 1 to 7, wherein the calibration support device is characterized by the above. It is a proofreading support method performed by a computer.
A step to generate a word-separated sentence by dividing the calibrated sentence into processing units and generating a word-separated sentence,
A proofreading target determination step for determining a processing unit that matches a heading in the proofreading history corpus as a proofreading target among the processing units constituting the divided sentence.
A calibration candidate prediction step for predicting a calibration candidate using a vector possessed by at least one processing unit or processing unit group before and after the calibration target.
Among the calibration candidates, an appropriate candidate determination step for determining a calibration candidate that matches the calibration result corresponding to the heading in the calibration history corpus as an appropriate candidate.
A calibration support method characterized by having. On the computer
A step to generate a word-separated sentence by dividing the calibrated sentence into processing units and generating a word-separated sentence,
A proofreading target determination step for determining a processing unit that matches a heading in the proofreading history corpus as a proofreading target among the processing units constituting the divided sentence.
A calibration candidate prediction step for predicting a calibration candidate using a vector possessed by at least one processing unit or processing unit group before and after the calibration target.
Among the calibration candidates, an appropriate candidate determination step for determining a calibration candidate that matches the calibration result corresponding to the heading in the calibration history corpus as an appropriate candidate.
A proofreading support program characterized by running.

Images