JP6768738B2 - Training data generators, methods, and programs - Google Patents

Description

The present invention relates to devices, methods, and programs that automatically generate learning data used for machine learning of document proofing engines.

In recent years, by using computer-based machine learning to create a multi-layered neural network that imitates the human brain neural circuit, the computer itself captures the potential features contained in the data and realizes more accurate and efficient judgment. Learning is widely used in the field of pattern recognition such as image recognition.

Through such deep learning, computers have become able to make decisions that are closer to humans. For example, when humans identify other people or animals, they may recognize facial parts such as eyes, ears, and mouth after seeing the entire physique of the object by making full use of sight and hearing. On the contrary, it goes through a hierarchical process such as recognizing the whole from the parts.

With conventional computers, recognition through such a process is difficult, and identification is performed by a program that calculates the characteristics of an object, but according to deep learning, even a computer can be recognized as a human recognition process. By following the same process, it became possible to recognize the human face and animals in the photograph.

However, in order to enable such recognition, deep learning requires learning with a large amount of learning data. For example, in order to determine whether the image is a gorilla image or a giraffe image by image recognition using deep learning, at least about 100,000 gorilla images and giraffe images are similarly displayed in advance. It is necessary to provide at least about 100,000 images as learning data to a computer for learning.

Ryo Takahashi, "What is a search algorithm that improves the" detection accuracy of particles "that is difficult even with LSTM and Residual Learning?", Www.atmarkit.co.jp/ait/articles/1611/11/news016.html (Recruit Technologies) ( Searched on April 13, 2018) Yuta Hitomi, Hideaki Tamori, Naoaki Okazaki and Kentaro Inui, ‘Proofread Sentence Generation as Multi-Task Learning with Edit Operation Prediction’, In Proceedings of the 8th International Joint Conference on Natural Language Processing.

Deep learning is considered to be applied not only to image recognition but also to document proofreading.

In order to apply deep learning to document proofreading, it is necessary to train the document proofreading engine, which is software for document proofreading, by deep learning.

In order for the deep-learned document proofreading engine to be able to determine document errors with high accuracy, not only correct sentences but also sentences containing errors (hereinafter referred to as "error sentences") are used as training data. Also need. In order to achieve sufficient calibration accuracy, as in the case of image recognition, an extremely large amount (for example, hundreds of thousands) of training data is required.

However, although it is easy to collect learning data of correct sentences, it is not easy to collect learning data of incorrect sentences. This is because, in general, a published document is a document consisting of correct sentences in principle, and a document containing an incorrect sentence is not normally published.

In order to obtain a document containing an error sentence, for example, a document in the process of being created such as an article is collected from an organization that creates many documents, such as a newspaper company or a publisher, and an error contained in the document in the process of being created. One idea would be to get a sentence.

However, with this method, it is unclear how many error sentences can be obtained from the document being created, and it will not greatly contribute to the creation of a large amount of learning data including the error sentences.

Therefore, in order to proofread a document using deep learning, it is not enough to create learning data based on the document being created, and an error sentence is intentionally and spontaneously generated by some method. There is a need.

That said, it is not realistic for humans to generate erroneous sentences. This is because human beings are unfamiliar with the generation of incorrect sentences because they originally think to generate correct sentences. Therefore, even if human beings generate erroneous sentences, a great deal of manpower will be required to generate a sufficient number of erroneous sentences.

In addition, when human beings generate error sentences, there is a concern that the error method will be biased due to individual habits. Even if a sufficient number of error sentences can be generated, if the error tendencies are similar, it will not be effective learning data.

The present invention has been made in view of such circumstances, and a device, a method, and a device capable of automatically generating learning data for machine learning of a document calibration engine by a computer instead of manually. The purpose is to provide a program.

In order to achieve the above object, the present invention takes the following measures.

That is, the invention of claim 1 is an apparatus for generating learning data used for machine learning of a document calibration engine, and includes a processor. The processor cuts out an arbitrary character string of an arbitrary length in order to generate the training data from the document stored in the storage device, and includes the cut out character string in the character string. A process of duplicating an arbitrary one character, a process of deleting an arbitrary one character included in the character string, a process of replacing an arbitrary one character included in the character string with another character, a process of being included in the character string. The training data is generated from the cut out character string by executing at least one of a process of rearranging an arbitrary two characters and a process of adding an arbitrary one character included in the character string. Perform the steps to be performed.

According to the second aspect of the present invention, in the learning data generation device according to the first aspect, the processor further executes a step of outputting the generated learning data to the document calibration engine.

According to the invention of claim 3, when the same learning data is generated by the step of generating the learning data in the learning data generation device according to claim 2, the processor is one of the same learning data. The step of outputting only one to the document calibration engine is executed.

According to the invention of claim 4, when the learning data generation device according to any one of claims 1 to 3 performs a process of deleting the arbitrary one character, the processor causes the arbitrary one character. The step of adding an arbitrary character to the end of the character string in which is deleted to obtain the training data is executed.

The invention of claim 5 further includes the storage device in the learning data generation device according to any one of claims 1 to 4.

The invention of claim 6 is a method of generating training data used for machine learning of a document calibration engine, in order for a processor to generate the training data from a document stored in a storage device . A process of cutting out an arbitrary character string of an arbitrary length, a process of duplicating an arbitrary one character included in the character string with respect to the cut out character string, and an arbitrary one character included in the character string. The process of deleting, the process of replacing any one character included in the character string with another character, the process of rearranging any two characters included in the character string, and the process of rearranging any one character included in the character string. By carrying out at least one of the processes to be added, the step of generating the training data from the cut out character string is executed.

The invention of claim 7 is a program for generating training data used for machine learning of a document calibration engine , and is optional for generating the training data from a document stored in a storage device. A function to cut out an arbitrary character string of the length of, a process of duplicating any one character included in the character string with respect to the cut out character string, and deleting any one character included in the character string. Processing, processing for replacing any one character included in the character string with another character, processing for rearranging any two characters included in the character string, and addition of any one character included in the character string. By performing at least one of the processes, the computer is made to realize the function of generating the training data from the cut out character string.

According to the present invention, devices, methods, and programs capable of automatically generating learning data for machine learning of a document calibration engine, particularly learning data including error sentences, by a computer rather than manually. Can be realized.

It is a block diagram which shows the structural example of the learning data generation apparatus to which the learning data generation method which concerns on embodiment of this invention is applied. It is a flowchart which shows the operation example of the learning data generation apparatus to which the learning data generation method which concerns on embodiment of this invention is applied.

Hereinafter, a learning data generation device to which the learning data generation method according to the embodiment of the present invention is applied will be described with reference to the drawings.

FIG. 1 is a block diagram showing a configuration example of a learning data generation device to which the learning data generation method according to the embodiment of the present invention is applied.

The learning data generation device 10 is a device for generating learning data for machine learning for document proofreading.

Machine learning includes deep learning. The learning data includes learning data including correct sentences and learning data including error sentences. The learning data generation device 10 is particularly suitable for generating learning data including an error sentence. In this specification, the typographical error corresponds to a sentence including typographical errors, omissions, and the like.

The electronic circuit of the learning data generation device 10 includes a CPU 12, a recording medium reading unit 14, a display 15, a key input unit 16, a communication unit 17, a storage device 18, and a memory 20 connected to each other by a bus 11.

The key input unit 16 is a portion that receives operation input from the user, such as a keyboard and a mouse.

The CPU 12 is a computer, controls the operation of each part of the circuit according to the character string cutting program 21, the learning data generation program 22, and the learning data output program 23 stored in the memory 20, and is input from the key input unit 16. Executes the process according to the operation input.

The character string cutting program 21, the learning data generation program 22, and the learning data output program 23 may be stored in the memory 20 in advance, or may be stored in advance from an external recording medium 13 such as a memory card via the recording medium reading unit 14. It may be read and stored in the memory 20.

The character string cutting program 21, the learning data generation program 22, and the learning data output program 23 cannot be rewritten by input from the key input unit 16 by the user.

In addition to such user-rewritable information, the memory 20 is an area in which information corresponding to the input made by the key input unit 16 is sequentially input as an area for storing user-rewritable data. A writable data area 24 is secured.

The communication unit 17 is connected to a communication network such as the Internet, receives a document from an external document server or the like (not shown) via this communication network, and stores the received document in the storage device 18. Alternatively, the learning data output by the learning data output program 23 is transmitted to, for example, a document calibration engine (not shown) to which deep learning is applied via a communication network.

The storage device 18 is composed of, for example, an SSD (Solid State Drive), an HDD (Hard Disk Drive), or the like, and stores a document received by the communication unit 17 from an external document server or the like (not shown) via a communication network. There is. In FIG. 1, the storage device 18 is provided inside the learning data generation device 10, but the storage device 18 may be provided outside the learning data generation device 10.

In the learning data generation device 10 configured in this way, the CPU 12 controls the operation of each part of the circuit according to the instructions described in the character string cutting program 21, the learning data generation program 22, and the learning data output program 23, and the software and the software. By operating in cooperation with the hardware, as explained below, an arbitrary character string is cut out from the document, and training data is automatically generated based on the cut out character string, and the generated training data is generated. To the document calibration engine.

The character string cutting program 21 cuts out an arbitrary character string of an arbitrary length from the document stored in the storage device 18. For example, suppose that the document stored in the storage device 18 is "It is easy to purchase a product online by the Internet, but there are many troubles in which the wrong product is shipped." When cutting out an arbitrary character string having a length of 10 characters as an arbitrary length from this document, it is possible to cut out a character string having a length of 10 characters, for example, "It is easy but wrong". ..

The learning data generation program 22 deletes (1) replace, (2) sort, (3) add, (4) duplicate, and (5) delete the character string cut out by the character string cutout program 21. At least one of the five processes as described above is performed. The specific processing contents of each of the above processes (1) to (5) are shown below together with the processing results for the above-mentioned character string "easy but incorrect".

(1) Replacement: Replace any one character in the cut out character string with another character of the same character type (Hiragana / Katakana / Kanji). By this process, the character string "easy but wrong" can be converted into, for example, "easy but wrong".

(2) Sorting: Sorting two consecutive characters of the same character type in the cut out character string. By this process, the character string "easy but wrong" can be converted into, for example, "easy but wrong".

(3) Addition: Add any one character of the same character type immediately before or after any one character in the cut out character string. By this process, the character string "easy but wrong" can be converted into, for example, "easy but wrong".

(4) Duplicate: Any one character in the cut out character string is duplicated. By this process, the character string "easy but wrong" can be converted into, for example, "easy but wrong".

(5) Delete: Delete any one character in the cut out character string. By this process, the character string "easy but wrong" can be converted into, for example, "easy but wrong".

The learning data generation program 22 determines the character string after conversion by each of the above processes (1) to (5) as learning data.

The character string after conversion by each of the above processes (1) to (5) (for example, (1) "easy to do, wrong", (2) "easy to do, but wrong", (3) " "Easy but wrong", (4) "Easy but wrong", and (5) "Easy but wrong") appear in standard Japanese documents. It is a character string that contains erroneous sentences.

In this way, in each of the above processes (1) to (5), an error sentence is generated from the original sentence by performing the minimum conversion on the correct original sentence, so that the error sentence is a correct sentence. Is similar to. In fact, in most cases, the mistakes that humans make when creating a document are only one place in a sentence. In addition, it is generally more difficult for a computer to make a correct / incorrect judgment for a sentence containing only one error than for a sentence containing many errors.

From this point of view, the error sentence used as the learning data for the deep learning of the document proofreading engine is more preferable as it is similar to the correct sentence. Therefore, the learning data generated by the learning data generation program 22 as described above becomes high-quality learning data suitable for improving the performance of the document proofreading engine.

In addition, when the learning data generation program 22 performs the deletion process of an arbitrary one character in (5) deletion, the sentence after the process is smaller than the number of character strings after the process by the other process. In order to make the length of the character string after the deletion processed uniform with other processing, any one character may be added to the end of the deleted character string. For example, in the above example, by implementing (5), the 10-character character string "easy but incorrect" becomes a 9-character character string "easy but incorrect". Therefore, in order to adjust the number of characters, it is generated by adding an arbitrary character to the end of the character string "It is easy but wrong" after deletion. For example, "It is easy but wrong" Such a sentence may be used as learning data.

Further, in the above example, it is shown that five error sentences are generated from one correct character string by each of the processes (1) to (5), but the above example is included in the correct character string. It is just an example made for one character. In the case of a 10-character character string as in the above example, if each of the above processes (1) to (5) is performed for all 10 characters, a maximum of 50 characters can be obtained from one correct character string. An erroneous statement will be generated. In this way, the learning data generation device 10 generates a large amount of error data from one correct character string.

Considering the large number of types of characters used in Japanese, the possibility is low enough, but as a result of performing each of the above processes (1) to (5), a sentence without errors was obtained. It can be done. However, the learning data generated by the learning data generation program 22 does not necessarily have to include an erroneous sentence, and may not necessarily include an erroneous sentence. Therefore, the learning data generation program 22 determines the sentences obtained by each of the above processes (1) to (5) as learning data regardless of whether the sentences are correct or include error sentences. good.

The learning data output program 23 outputs the learning data generated by the learning data generation program 22 to the communication unit 17, and causes the communication unit 17 to transmit the learning data to a document calibration engine (not shown) via the communication network.

When the learning data output program 23 contains the same learning data in the learning data generated by the learning data generation program 22, only one of the same learning data is documented from the communication unit 17. It may be sent to the calibration engine.

Next, the operation of the learning data generation device to which the learning data generation method according to the embodiment of the present invention configured as described above is applied will be described with reference to the flowchart shown in FIG.

A document is received by the communication unit 17 from an external document server or the like (not shown) via a communication network such as the Internet (S1).

For example, suppose the document received was "The snow that has been falling since yesterday made today the best skiing day and many skiers visited."

The document received by the communication unit 17 is sent from the communication unit 17 to the storage device 18 and stored in the storage device 18 (S2).

An arbitrary character string of an arbitrary length is cut out from the document stored in the storage device 18 by the character string cutting program 21 (S3).

For example, from the document stored in the storage device 18, "Because of the snow that has continued since yesterday, today was the best skiing day and many skiers visited.", A character string of 11 characters as an arbitrary length. When cutting out, for example, a character string having a length of 11 characters such as "Today's best skiing weather" can be cut out.

The character string cut out by the character string cutting program 21 is subjected to (1) replacement, (2) sorting, (3) addition, (4) duplication, and (5) deletion by the learning data generation program 22. Learning data is generated by performing at least one of the five processes as described above (S4).

By performing each of the above processes (1) to (5) on the character string “Today's best skiing weather” cut out in step S3, for example, the following error sentence is generated.

(1) Replacement: "The best skiing weather today"
(2) Sorting: "The best skiing weather right now"
(3) Addition: "Today is the highest ski day"
(4) Duplicate: "Today is the best skiky day"
(5) Deleted: "Today's best ski sum"
As described above, the error sentence generated by each of the above processes (1) to (5) is similar to the correct sentence because only the minimum conversion process is applied to the original sentence. For a computer, it is generally more difficult to judge the correctness of a sentence containing only one error than a sentence containing many errors. Therefore, the error sentence generated as described above is a document proofreading engine. It is a good quality learning data suitable for improving the performance of the computer.

Further, in the above example, it is shown that five error sentences are generated from one correct character string by each of the processes (1) to (5), but the above example is included in the correct character string. It is just an example made for one character. In the case of an 11-character character string as in the above example, if each of the above processes (1) to (5) is performed for all 11 characters, a maximum of 55 can be obtained from one correct character string. An error statement is generated. As described above, according to the learning data program 22, a large amount of learning data is efficiently generated from one correct character string.

The learning data generated by the learning data generation program 22 in this way is output to the communication unit 17 by the learning data output program 23, and is directed from the communication unit 17 to a document calibration engine (not shown) via the communication network. It is transmitted (S5).

In the document proofreading engine, deep learning is performed using the transmitted learning data (S6).

As described above, according to the learning data generation device to which the learning data generation method according to the present embodiment is applied, high-quality learning for machine learning such as deep learning of a document calibration engine by the above-mentioned action. Data can be generated automatically and efficiently by a computer.

Therefore, it is possible to greatly contribute to improving the accuracy of the document proofreading engine using deep learning.

The best mode for carrying out the present invention has been described above with reference to the accompanying drawings, but the present invention is not limited to such a configuration. Within the scope of the claimed technical idea, those skilled in the art can come up with various modifications and modifications, and these modifications and modifications are also the technical scope of the present invention. It is understood that it belongs to.

The learning data generator of the present invention can be used to generate learning data for machine learning of a document calibration engine, particularly deep learning.

10 ... Learning data generator,
11 ... Bus,
12 ... CPU,
13 ... External recording medium,
14 ... Recording medium reader,
15. Display,
16 ... Key input section,
17 ... Communication Department,
18 ... Storage device,
20 ... Memory,
21 ... Character string cutting program,
22 ... Learning data generation program,
23 ... Learning data output program,
24 ... Writable data area.

Claims (7)

A device that generates learning data used for machine learning of document proofing engines.
Equipped with a processor
The processor
A process of cutting out an arbitrary character string of an arbitrary length from a document stored in the storage device in order to generate the learning data, and
A process of duplicating an arbitrary one character included in the character string with respect to the cut out character string, a process of deleting an arbitrary one character included in the character string, and an arbitrary one character included in the character string. Is replaced with another character, any two characters included in the character string are rearranged, and any one character included in the character string is added. A training data generation device that executes a step of generating the training data from the cut out character string. The learning data generation device according to claim 1, wherein the processor further executes a step of outputting the generated learning data to the document calibration engine. If the same training data is generated by the step of generating the training data, the processor executes the step of outputting only one of the same training data to the document calibration engine. 2. The learning data generator according to 2. When the process of deleting the arbitrary one character is executed, the processor executes a step of adding the arbitrary one character to the end of the character string in which the arbitrary one character is deleted to obtain the training data. The learning data generation device according to any one of claims 1 to 3. The learning data generation device according to any one of claims 1 to 4, further comprising the storage device. A method of generating training data used for machine learning of document proofing engines.
The processor
A process of cutting out an arbitrary character string of an arbitrary length from a document stored in the storage device in order to generate the learning data, and
A process of duplicating an arbitrary one character included in the character string with respect to the cut out character string, a process of deleting an arbitrary one character included in the character string, and an arbitrary one character included in the character string. Is replaced with another character, any two characters included in the character string are rearranged, and any one character included in the character string is added. A training data generation method for executing the step of generating the training data from the cut out character string. A program for generating training data used for machine learning of document proofing engines.
A function of cutting out an arbitrary character string of an arbitrary length in order to generate the learning data from a document stored in a storage device.
A process of duplicating an arbitrary one character included in the character string with respect to the cut out character string, a process of deleting an arbitrary one character included in the character string, and an arbitrary one character included in the character string. At least one of a process of replacing the character with another character, a process of rearranging any two characters included in the character string, and a process of adding an arbitrary one character included in the character string. A function to generate the training data from the cut out character string,
A program to realize on a computer.