JP6896552B2 - Explanatory text evaluation method, explanatory text evaluation device and explanatory text evaluation program - Google Patents

Description

The present application relates to an explanatory text evaluation method, an explanatory text evaluation device, and an explanatory text evaluation program.

Advertisements on product packages, TVs, newspapers, magazines, etc. usually contain explanatory texts indicating how to use them, their effects, and the like. In order for consumers to properly understand quality, effectiveness, safety, etc., and to purchase and use products, various laws and regulations such as the Pharmaceutical and Medical Device Law (Pharmaceutical Machinery Law) and the Freebie Labeling Law are required for explanations. Various regulations and guidelines have been formulated.

Expressions that describe the features of a product must not include content that goes beyond the indications and effects that consumers would normally understand under the regulations of various laws and regulations. For this reason, the description of the product is usually checked by an expert or the like.

In order to select an appropriate word as an explanatory sentence, it is conceivable to use a technique or the like (see, for example, Patent Document 1) for predicting a word that can be an unknown word from the words before and after the unknown word in the document.

U.S. Pat. No. 9037464

However, it is difficult for the person in charge of planning and developing the product to evaluate whether or not the description is appropriate in accordance with regulations such as various laws and regulations. Is being done.

On the other hand, it takes time to educate skilled workers, and today, the number of inquiries from the persons in charge of each department is increasing, and the situation is that they are busy responding. Further, when the explanation is evaluated based on the experience of the expert, the judgment may be different depending on the expert, and there is a problem that it is difficult to obtain a certain evaluation.

Therefore, in one aspect, the present invention aims to quickly obtain a stable evaluation of the description of a product.

According to one form, in response to the reception of the evaluation request of the explanatory text attached to the product, the monocultural processing step of monoculturing the explanatory text and the content of the simple text for each simple sentence are at least the explanatory representation of the product. A simple sentence unit evaluation step that evaluates whether or not the content complies with the laws and regulations related to the above using a model obtained by machine learning in advance, and a sentence about the suitability as a description of the product are associated with each evaluation result. By referring to the definition table, the sentence corresponding to the evaluation result by the simple sentence unit evaluation step is acquired, and the result display that enables the sentence to be displayed according to the selection of the simple sentence on the screen for displaying the explanation. Result display of data creation A data creation step and a description evaluation method performed by a computer are provided.

Further, as a means for solving the above-mentioned problems, an explanatory text evaluation device and an explanatory text evaluation program can be used.

A stable evaluation can be quickly obtained for the explanation.

It is a figure which shows the network configuration example of the explanatory text evaluation system in 1st Example. It is a figure which shows the hardware configuration in the explanatory text evaluation system in 1st Example. It is a flowchart for demonstrating the model creation verification process in 1st Example. It is a flowchart for demonstrating the evaluation process of the explanatory text in the explanatory text evaluation system in 1st Example. It is a figure which shows the functional structure example in the appeal sentence evaluation system in 2nd Example. It is a figure which shows the data structure example of the definition table in 2nd Example. It is a figure which shows the data structure example of the teacher data in 2nd Example. It is a flowchart for demonstrating the knowledge base creation process in 2nd Example. It is a flowchart for demonstrating the discriminant model creation process in 2nd Example. It is a flowchart for demonstrating the appeal sentence evaluation process in 2nd Example. It is a figure which shows the initial screen example displayed on the client terminal in 2nd Example. It is a figure which shows the screen example at the time of receiving the result display data in 2nd Example. It is a figure which shows the display example of the detail of the evaluation result in 2nd Example. It is a figure which shows the other screen example at the time of receiving the result display data in 2nd Example. It is a figure which shows the other display example of the detail of the evaluation result in 2nd Example. It is a figure which shows the data structure example of the definition table in 3rd Example. It is a figure which shows the data structure example of the teacher data in 3rd Example. FIG. 18 is a diagram showing a display example of details of the evaluation result in the third embodiment.

Hereinafter, embodiments for carrying out the present invention will be described with reference to the drawings, but the present invention is not limited to the following embodiments and does not deviate from the scope of the present invention. Can be modified and substituted in various ways.

[First Example]
First, the explanation evaluation system 1000 that evaluates whether the explanation about the products, services (services), etc. to be sold complies with the laws and regulations of the product field and each country or region (hereinafter referred to as the country). The outline of is explained.

FIG. 1 is a diagram showing a network configuration example of the explanatory text evaluation system in the first embodiment. In FIG. 1, the explanatory text evaluation system 1000 has a server device 100 and a plurality of client terminals 3, each of which can be connected via the network 2.

When the server device 100 receives the evaluation request 4r of the explanation from the client terminal 3, the explanation included in the evaluation request 4r is based on the product field, the country, and the evaluation viewpoint specified in the selection information 4s of the evaluation request 4r. Evaluate 4q. Based on the evaluation result, the result display data 5r to which the evaluation explanatory information 5p is added is provided to the client terminal 3. The explanatory text 4q is text data, voice data, or the like.

When the server device 100 receives the evaluation request 4r, the server device 100 evaluates the explanatory text 4q using the model created in advance based on the evaluation viewpoint specified in the selection information 4s of the evaluation request 4r, and the evaluation explanation is based on the evaluation result. The result display data 5r to which the information 5p is added is output.

The evaluation explanation information 5p links a code for displaying a risky sentence and a word in an identifiable manner, and links each explanation to the sentence and the word so that the explanation is displayed when the risky sentence and the word are selected. Includes data such as code and description to be made. In addition to the description, alternatives may be included.

The client terminal 3 transmits an evaluation request 4r including an explanatory text 4q to be evaluated specified by the user to the server device 100 in response to the operation of the user, and receives and displays the result display data 5r from the server device 100.

The server device 100 that provides the evaluation explanatory information 5p may target the explanatory text 4q of the product, service (service), or the like. Examples are cosmetics, food, real estate / finance, etc. In order to correspond to the description 4q of various products, as illustrated in FIG. 1, the server device 100 includes a cosmetic product field SB1, a food product field SB2, a real estate / financial product field SB3, and the like. Each of them is provided with a discriminant model group. Discrimination model groups may be provided for each country. Here, an outline of the functional configuration when a discrimination model group is provided for each country in each product field is shown.

For example, the server device 100 has a model creation verification unit 1150, an explanatory text evaluation unit 1180, an explanatory text DB 1137, and a cosmetic product field SB1. Etc. (collectively, it may be referred to as a plurality of discriminant model groups HMG).

Discrimination model group HMG1 corresponded to Japanese laws and regulations, with the discriminant model SB1a based on the evaluation aspect _{a 1,} a discrimination model SB1b like based on the evaluation aspect _{b 1.} Discrimination model group HMG2 corresponded to Korean laws and regulations, discriminant model SB2a based on the evaluation aspect _{a 2,} a discrimination model SB2b like based on the evaluation aspect _{b 2.}

Discrimination model group HMG3 corresponded to Chinese laws and regulations, discriminant model SB3a based on the evaluation aspect a _3, a discrimination model SB3b like based on the evaluation aspect b _3. Discrimination model group HMG4 corresponded to US laws and regulations, discriminant model SB4a based on the evaluation aspect a _4, a discriminant model SB4b like based on the evaluation aspect b _4.

Discrimination model group HMG5 corresponded to European law, discriminant model SB5a based on the evaluation aspect _{a 5,} has a discriminant model SB5b like based on the evaluation aspect _{b 5.} The discrimination models SB1a, ..., SB5b and the like are simply referred to as discrimination models SB unless any of them is specified.

Each of these discriminant model SBs is a model created by the model creation verification unit 1150 by the model creation verification unit 1150 and whose appropriateness has been verified by using the explanatory text DB1137 that has accumulated many explanatory texts collected for discriminant model creation. Is.

The explanatory text DB1137 is a database that manages the explanatory texts collected and accumulated for each country in each product field. The collected and accumulated explanatory text corresponds to corpus data, and the explanatory text DB1137 includes corpus data for each product field and country.

The model creation verification unit 1150 and the explanatory text evaluation unit 1180 are realized by processing by executing a program corresponding to the CPU 111 (FIG. 2) described later, and these processes will be described later.

In FIG. 1, the network configuration in which the client terminal 3 and the server device 100 are connected via the network 2 has been described, but the user may evaluate the explanatory text 4q in a stand-alone device having the function of the server device 100. .. In addition, individual server devices may be provided with the same functional configuration for each product field. Alternatively, it may be an individual server device for each country.

Here, the laws and regulations related to each product field are as follows.
・ Cosmetics, Pharmaceutical Machinery Law, etc. ・ Foods (general foods, foods for specified health use, nutritionally functional foods, foods with functional claims, health foods)
The prize labeling law, the Specified Commercial Transactions Law, the JAS law, etc., the real estate / financial prize labeling law, etc.

In addition, regarding cosmetics, the laws and regulations of each major country are shown.
・ Korea Cosmetics Law, Advertising Guidelines, etc. ・ China Cosmetics Satellite Supervision Ordinance, People's Republic of China Consumer Rights Protection Law, etc. ・ United States
Federal Food, Drug, and Cosmetic Act, Federal Trade Commission Act, Fair Packaging and Labeling Act, etc.-Europe
Cosmetic Regulation, Unfair commercial practices directive, etc.

In the following description, a case where a discrimination model is created for at least one product field in one server device 100 based on the laws and regulations of a plurality of countries will be described.

In the first embodiment, even a user who is not sufficiently familiar with various domestic and foreign laws and regulations can evaluate the created explanatory text 4q.

FIG. 2 is a diagram showing a hardware configuration in the explanatory text evaluation system according to the first embodiment. In FIG. 2, the server device 100 is an information processing device controlled by a computer, and is a CPU (Central Processing Unit) 111, a main storage device 112, an auxiliary storage device 113, an input device 114, and a display device 115. The communication I / F (interface) 117 and the drive device 118 are provided and connected to the bus B1.

The CPU 111 corresponds to a processor that controls the server device 100 according to a program stored in the main storage device 112. A RAM (Random Access Memory), a ROM (Read Only Memory), or the like is used in the main storage device 112, and is obtained by a program executed by the CPU 111, data required for processing by the CPU 111, and processing by the CPU 111. Store or temporarily store the data, etc.

An HDD (Hard Disk Drive) or the like is used in the auxiliary storage device 113, and data such as a program for executing various processes is stored in the auxiliary storage device 113. A part of the program stored in the auxiliary storage device 113 is loaded into the main storage device 112 and executed by the CPU 111 to realize various processes. The storage unit 130 corresponds to the main storage device 112 and / or the auxiliary storage device 113.

The input device 114 has a mouse, a keyboard, and the like, and is used by the user to input various information necessary for processing by the server device 100. The display device 115 displays various information required under the control of the CPU 111. The input device 114 and the display device 115 may be a user interface using an integrated touch panel or the like. The communication I / F 117 communicates through a network such as wired or wireless. Communication by communication I / F 117 is not limited to wireless or wired.

The program that realizes the processing performed by the server device 100 is acquired, for example, by downloading from the external server device that provides the program via the network 2.

The drive device 118 interfaces the storage medium 119 (for example, a CD-ROM or the like) set in the drive device 118 with the server device 100. The program that realizes the processing according to this embodiment may be provided to the server device 100 by a storage medium 119 such as a CD-ROM (Compact Disc Read-Only Memory).

In this case, the program stored in the storage medium 119 is installed in the server device 100 by setting the storage medium 119 in which the program for realizing various processes according to the present embodiment described later is stored in the drive device 118. You may. The installed program can be executed by the server device 100.

The storage medium 119 for storing the program is not limited to the CD-ROM, and is one or more non-transitory, tangible (tangible) having a structure as data that can be read by a computer. ) Any medium. The computer-readable storage medium may be a DVD (Digital Versatile Disk) disk, a portable recording medium such as a USB memory, or a semiconductor memory such as a flash memory, in addition to the CD-ROM.

The client terminal 3 is an information processing device controlled by a computer, and is a CPU (Central Processing Unit) 11, a main storage device 12, an auxiliary storage device 13, an input device 14, a display device 15, and a communication I. It has a / F (interface) 17 and a drive device 18, and is connected to the bus B2.

The CPU 11 corresponds to a processor that controls the client terminal 3 according to a program stored in the main storage device 12. A RAM (Random Access Memory), a ROM (Read Only Memory), or the like is used in the main storage device 12, and is obtained by a program executed by the CPU 11, data required for processing by the CPU 11, and processing by the CPU 11. Store or temporarily store the data, etc.

An HDD (Hard Disk Drive) or the like is used in the auxiliary storage device 13, and data such as a program for executing various processes is stored in the auxiliary storage device 13. Various processes are realized by loading a part of the program stored in the auxiliary storage device 13 into the main storage device 12 and executing the program in the CPU 11. The storage unit 30 corresponds to the main storage device 12 and / or the auxiliary storage device 13.

The input device 14 has a mouse, a keyboard, and the like, and is used by the user to input various information necessary for processing by the client terminal 3. The display device 15 displays various necessary information under the control of the CPU 11. The input device 14 and the display device 15 may be a user interface using an integrated touch panel or the like. The communication I / F17 communicates through a network such as wired or wireless. Communication by communication I / F17 is not limited to wireless or wired.

The processing performed by the client terminal 3 may be realized by connecting to the server device 100 via a Web browser.

The drive device 18 interfaces the storage medium 19 (for example, a CD-ROM or the like) set in the drive device 18 with the client terminal 3.

As another method for realizing the processing performed by the client terminal 3, a program for realizing various processes according to the present embodiment, which will be described later, is stored in the storage medium 19, and the program stored in the storage medium 19 is stored. , May be installed on the client terminal 3 via the drive device 18. The installed program can be executed by the client terminal 3.

In this case, the storage medium 19 for storing the program is not limited to the CD-ROM, and is one or more non-transitory having a structure as data that can be read by a computer. It can be a tangible medium. The computer-readable storage medium may be a DVD (Digital Versatile Disk) disk, a portable recording medium such as a USB memory, or a semiconductor memory such as a flash memory, in addition to the CD-ROM.

The client terminal 3 may be a laptop, a tablet terminal, or the like. In that case, the storage medium 19 is an SD (Secure Digital) memory card or the like, and the drive device 18 interfaces the storage medium 19 set in the drive device 18 with the terminal 3.

The model creation verification process by the model creation verification unit 1150 will be described. FIG. 3 is a flowchart for explaining the model creation verification process in the first embodiment. In FIG. 3, a plurality of explanatory texts (corpus data) are selected for each combination of product fields, countries, and evaluation viewpoints from the explanatory text DB 1137, and the following processing is performed for each combination. To select the corpus data from the description DB 1137, a screen for selecting a product field and a country may be displayed on the display device 115, and a matching or related type of corpus data may be extracted based on the user's specification.

First, the model creation / verification unit 1150 performs collection / structuring processing (step S110). In the collection / structuring process, the model creation / verification unit 1150 collects explanatory texts for creating a discrimination model, samples the collected explanatory texts, and structures them in association with appropriateness labels. In the corpus data extracted from the explanatory text DB1137, the appropriateness of the sampled explanatory text is determined, and the determination result corresponds to the appropriateness label. The suitability label indicates either inappropriate, cautionary, or proper. The descriptive text associated with the aptitude label is the teacher data 138.

Judgment of appropriateness is made by the judgment of a skilled person with expertise based on various laws and regulations related to the product field, its operation policy, and a collection of questions and answers.

Next, the model creation / verification unit 1150 performs a feature extraction process (step S120). In the feature extraction process, the model creation / verification unit 1150 performs morpheme decomposition and then performs natural language processing to create an algorithm for extracting sentence features. Then, the model creation / verification unit 1150 generates a sentence vector space by the created algorithm.

Then, the model creation / verification unit 1150 performs a discrimination model creation process by machine learning (step S130). In the discrimination model creation process, the model creation / verification unit 1150 creates a discrimination model SBi based on the teacher data 138 among the explanations accumulated in the explanation DB 1137 for the sentence features.

In machine learning, as an example, a multi-support vector machine (multi-SVM) or the like is used to learn the boundary surface 3f indicating the classification of aptitude labels in the sentence vector space. In the process of learning the boundary surface 3f, the model creation verification unit 1150 performs the accuracy verification process (step S140). In the accuracy verification process, the model creation verification unit 1150 can repeat the process of calculating the accuracy of the model while changing various coefficients included in the model by using the explanatory text without the appropriateness label, and can classify with the highest accuracy. Generate a model.

In addition to the teacher data 138, oversampling is performed as appropriate. In oversampling, in the sentence vector space, a virtual point (hereinafter, virtual) is provided so that the document feature point exists in the vicinity defined by a predetermined distance between the points indicating the sentence feature associated with the appropriateness label. Point 3v) is to be added. By adding such a virtual point 3v, the accuracy of the model can be maintained even when the collected teacher data 138 has a bias of the appropriateness label. As an example of such processing, SMOTE (Synthetic Minority Over-sampling Technique) or the like may be used.

By repeating this model creation verification process as a whole, a discrimination model for each combination of product fields, countries, and evaluation viewpoints is created. That is, one or a plurality of discrimination models SBi are created.

FIG. 4 is a flowchart for explaining the evaluation process of the explanatory text in the explanatory text evaluation system in the first embodiment. In FIG. 4, the client terminal 3 accepts selection of the product field, country, and evaluation viewpoint from the screen for causing the server device 100 displayed on the display device 15 to evaluate the explanatory text 4q (step S211), and further. Read the evaluation description specified by the user (step S212).

The client terminal 3 creates an evaluation request 4r including the selection information 4s that specifies the product field, the country, and the evaluation viewpoint received in step S211 and the explanatory text read in step S212, and transmits the evaluation request 4r to the server device 100. (Step S213). The evaluation request 4r is transmitted to the server device 100.

When the server device 100 receives the evaluation request 4r from the client terminal 3, the explanation evaluation unit 1180 identifies the product field and one or more evaluation viewpoints from the selection information 4s of the received evaluation request 4r (step S321).

Then, the explanatory text evaluation unit 1180 selects the discriminant model group HMG based on the product field and the country specified in the selection information 4s (step S322). Further, the explanatory text evaluation unit 1180 selects the discrimination model SB for each evaluation viewpoint based on one or more evaluation viewpoints specified in the selection information 4s (step S323).

Next, the explanatory text evaluation unit 1180 analyzes the explanatory text 4q included in the evaluation request 4r, determines the presence or absence of risk for each discrimination model SB selected in step S323 (step S234), and evaluates the explanatory text information 5p. The result display data 5r including the above is created and transmitted to the client terminal 3 (step S235). The explanatory text evaluation process by the explanatory text evaluation unit 1180 in the server device 100 ends.

When the terminal 3 receives the result display data 5r, it acquires the evaluation explanatory information 5p from the result display data 5r, displays it on the display device 15 (step S214), and ends the process at the terminal 3.

The following is a case where a discrimination model for evaluating an appeal sentence explaining cosmetics is created based on the US law in the first embodiment described above, and the appeal sentence to be evaluated is evaluated using the created discrimination model. 2 Examples will be used, and the specific processing contents will be described. In the second embodiment, the appeal sentence 4p (FIG. 5) is an example of the explanatory sentence 4q of the first embodiment.

[Second Example]
FIG. 5 is a diagram showing a functional configuration example in the appeal sentence evaluation system in the second embodiment. In the appeal sentence evaluation system 1002 shown in FIG. 5, the server device 100 mainly has a knowledge base creation unit 60, a discrimination model creation unit 70, and an appeal sentence evaluation unit 80. The knowledge base creation unit 60, the discrimination model creation unit 70, and the appeal sentence evaluation unit 80 are realized by processing performed by executing a program corresponding to the CPU 111.

Here, the knowledge base creation unit 60 and the discrimination model creation unit 70 perform preprocessing and correspond to the model creation verification unit 1150 of the first embodiment. Further, the appeal sentence evaluation unit 80 corresponds to the explanatory sentence evaluation unit 1180 of the first embodiment.

Further, the storage unit 130 stores the explanatory text DB 1137, the knowledge base 137, the teacher data 138, the discrimination model 139, the definition table 140, the evaluation request 4r, the result display data 5r, and the like. The knowledge base 137 and teacher data 138 in the second embodiment correspond to corpus data classified into the product field "cosmetics" and the country "US" in the explanatory text DB1137 of the first embodiment.

The knowledge base creation unit 60 extracts the corresponding cosmetic appeal from the explanatory text DB1137 based on the product field "cosmetics" and the country "US" specified in the selection information 4s of the evaluation request 4r, and extracts the extracted cosmetics. The appeal sentence is divided into simple sentences, and the distributed expression (vector expression) of the word (or sentence) for each division is analyzed. The extracted cosmetic appeal text corresponds to corpus data. The analysis result is stored in the storage unit 130 as a knowledge base 137. The knowledge base creation unit 60 includes an appeal sentence collection unit 62, a monoculture processing unit 64, and a distributed expression analysis unit 66.

The appeal sentence collection unit 62 collects the appeal sentences and causes the sentences of the area (beauty) to be targeted to be selected according to the operation of a skilled person who has specialized knowledge about the appeal sentences. The selection of beauty texts can be omitted if sufficient at the collection stage.

The monoculture processing unit 64 divides the texts in the target area, that is, the texts of beauty, which are selected in advance in the collected appeal texts, into simple text units. In the case of Japanese, the monoculture processing unit 64 also recognizes word breaks for documents.

The distributed expression analysis unit 66 analyzes the distributed expression of words separated by the monoculture processing unit 64. As an example of the distributed expression analysis unit 66, Doc2Vec (registered trademark) or Word2Vec (registered trademark), which is an open source code of Google (registered trademark), may be used. Here, machine learning without teacher data is performed, and the analysis result of the distributed representation is output to the storage unit 130 as a knowledge base 137.

The discrimination model creation unit 70 creates a discrimination model 139 for discriminating whether the appeal text complies with the law and guidelines and has appropriate content as cosmetics from the viewpoint of evaluation.

The evaluation viewpoint is not limited to one. There may be a plurality of evaluation viewpoints, and each evaluation viewpoint may be divided into more detailed levels.

In this embodiment, the suitability (suitability) of cosmetics as an appeal is shown as a risk value.

The discrimination model creation unit 70 further includes a monoculture processing unit 72, a risk value acquisition unit 74, and a machine learning unit 76. The monoculture processing unit 72 monocultures the appeal sentences sampled for learning from the collected appeal sentences (hereinafter, the appeal sentences for learning).

When the risk value acquisition unit 74 receives a simple sentence from the monoculture processing unit 72, the risk value acquisition unit 74 acquires the risk value from the expert by displaying an evaluation screen or the like on the expert client terminal 3. Experts set the suitability of cosmetics as an appeal as a risk value. The risk value follows the preset definition table 140. A simple sentence is associated with the risk value and is stored in the storage unit 130 as teacher data 138.

The machine learning unit 76 creates a discrimination model 139 so that the teacher data 138 obtained by the risk value acquisition unit 74 can be classified more accurately on the knowledge base 137 obtained by the knowledge base creation unit 60. The machine learning unit 76 uses a multi-SVM or the like. When the teacher data is biased, the machine learning unit 76 uses SMOTE or the like to perform oversampling to add a virtual point 3v.

The discrimination model 139 is a model for discriminating whether or not a word or sentence in an appeal sentence is suitable as an appeal for cosmetics. The discriminant model 139 corresponds to one of the discriminant model group HMG4 for the United States of the first embodiment. For convenience of explanation, one discrimination model 139 will be described, but a plurality of discrimination models may be provided.

The appeal sentence evaluation unit 80 evaluates the appeal sentence 4p included in the evaluation request 4r according to the evaluation request 4r transmitted from the client terminal 3 via the network 2 and received by the communication I / F 117. The evaluation request 4r is stored in the storage unit 130. The appeal sentence evaluation unit 80 includes a single culture processing unit 82, a simple sentence unit evaluation unit 84, and a result display data creation unit 86.

The monoculture processing unit 82 monocultures the appeal sentence 4p in response to the input of the appeal sentence 4p included in the evaluation request 4r, and inputs the appeal sentence 4p to the single sentence unit evaluation unit 84.

The simple sentence unit evaluation unit 84 evaluates the appeal sentence 4p in single sentence units using the discrimination model 139, and determines whether or not the simple sentence is suitable as an appeal for cosmetics. The evaluation result is notified to the evaluation result transmission unit 85. The evaluation result shows the evaluated simple sentence and the risk value defined in the definition table 140.

The result display data creation unit 86 transmits the result display data 5r to which the evaluation explanation information 5p is added to the client terminal 3 of the request source of the evaluation request 4r via the network 2 based on the evaluation result by the simple sentence unit evaluation unit 84. ..

The client terminal 3 accesses the server device 100 with a Web browser (or application) in response to a user operation, causes the server device 100 to evaluate the appeal text 4p, and receives the result display data 5r from the server device 100. .. Based on the result display data 5r, the problematic part in the appeal sentence 4p is specified, and the evaluation explanation showing the degree of risk is displayed on the display device 15.

When the screen G90 for evaluation is displayed on the display device 15 on the client terminal 3, the user sets a desired appeal sentence 4p and causes the server device 100 to evaluate.

In response to the reception of the result display data 5r from the server device 100, the part of the review in the appeal sentence 4p is specified on the screen G90 based on the evaluation explanation information 5p, and the reason for the review is explained in a sentence. Therefore, the user can easily determine how much the word (or sentence) to be reviewed should be revised.

FIG. 6 is a diagram showing a data configuration example of the definition table in the second embodiment. In FIG. 6, the definition table 140 is a table in which a risk value and an explanation are associated with each evaluation viewpoint, and has items such as an evaluation viewpoint, a risk value, and an explanation.

The evaluation viewpoint shows, for example, the suitability of cosmetics as an appeal. Words corresponding to the evaluation viewpoint specified in advance by the expert are specified to have a high risk within the evaluation viewpoint. The risk value indicates the high risk. There may be multiple evaluation viewpoints.

In the explanation, explanations for each of the risk values are set. In the result display data 5r transmitted to the client terminal 3, the evaluation explanatory information 5p includes the explanatory text corresponding to the risk value determined by the simple sentence unit evaluation unit 84.

In this data structure example, the risk value is defined from "1" to "3". If the risk value is "3", it means that the risk value is "NG because it is out of the category of cosmetics". If the risk value is "2", it means that "there is a risk of being taken if it is out of the category of cosmetics". When the risk value is "1", it indicates that "it is an expression in the category of cosmetics and there is no problem".

By giving an explanation based on the risk value from such an evaluation viewpoint on the screen G90, it is possible to provide a stable evaluation even to a user in a department who does not have sufficient expertise in the appeal statement.

FIG. 7 is a diagram showing a data structure example of the teacher data in the second embodiment. In FIG. 7, the teacher data 138 is a table showing the evaluation results for each appeal sentence obtained by the risk value acquisition unit 74, and has items such as a product name, an evaluation result, and an appeal sentence.

The product name indicates the name of the target product in which the appeal statement is described. The product name may be omitted. The value of the evaluation result is the risk value of the definition table 140 as illustrated in FIG. The appeal sentence indicates the evaluated sentence.

In the data structure example of FIG. 7, an appeal sentence in English is mainly shown, but Japanese or another language may be used. As an example, the appeal statement evaluated for cosmetics with the product name "AAA" is "AAA completely erases acne scars rejuvenating dermal collagen.", And the evaluation result shows a risk value of "3", that is, , Clearly out of the category of cosmetics, and has been shown to be NG.

FIG. 7 shows an example when there is one evaluation viewpoint, but when two or more evaluation viewpoints are selected, the risk value for each evaluation viewpoint is shown. Moreover, although the appeal sentence is shown in a sentence for easy understanding, the coordinates in the sentence vector space may be shown.

By accumulating more teacher data 138, the accuracy of the evaluation result by the appeal sentence evaluation unit 80 can be improved. The teacher data 138 is not limited to the appeal sentence evaluated by an expert who has specialized knowledge at the time of collecting the appeal sentence.

The evaluation result provided to the client terminal 3 may be stored in the storage unit 130 as a log, and may be added to the teacher data 138 when evaluated again by a skilled person having specialized knowledge.

Next, the processing by each processing unit 60, 70, and 80 will be described with a flowchart. First, the knowledge base creation process by the knowledge base creation unit 60 will be described. FIG. 8 is a flowchart for explaining the knowledge base creation process in the second embodiment.

In FIG. 8, in the knowledge base creation unit 60, the appeal sentence collection unit 62 collects a plurality of appeal sentences 4c (step S611). In the collection of appeal sentences 4c, as an example, in the case of English sentences, about 90,000 cases are collected. The product name may be acquired for each appeal sentence 4c.

In addition, the appeal sentence collection unit 62 classifies and selects the appeal sentence 4c (step S612). The classification and selection of the appeal sentence 4c may be omitted if it has already been sufficiently done.

Next, the monoculture processing unit 64 monocultures the sentence of the appeal sentence 4c (step S613). Divide the appeal sentence into simple sentences for each product. If the appeal sentence 4c is in Japanese, word break recognition is also performed.

Then, the distributed expression analysis unit 66 performs an analysis process of the distributed expression for each simple sentence (step S614). In the analysis process of the distributed expression, the distributed expression (vector expression) of the word (or sentence) is analyzed by machine learning without teacher data, and the analysis result is accumulated in the knowledge base 137.

Next, the discriminant model creation process by the discriminant model creation unit 70 will be described. FIG. 9 is a flowchart for explaining the discrimination model creation process in the second embodiment.

In FIG. 9, the discrimination model creation unit 70 selects an appeal sentence (hereinafter, learning appeal sentence 4t) for learning (step S631). A learning appeal sentence 4t may be selected from the collected appeal sentences 4c by an expert, or may be randomly extracted. The selected learning appeal sentence 4t is stored in the storage unit 130.

The monoculture processing unit 72 monocultures the sentence of the learning appeal sentence 4t (step S632), and the risk value acquisition unit 74 acquires the risk value for each simple sentence from the expert (step S633). The risk value acquisition unit 74 may display the risk value setting screen on the expert client terminal 3 and acquire the risk value for each simple sentence from the expert. The risk value setting screen may be any screen in which the risk value can be set for a simple sentence. The set risk values are stored in the teacher data 138 in association with each simple sentence.

The machine learning unit 76 uses the teacher data 138 to count the number of simple sentences for each risk value (step S634), and calculates the ratio of the single sentences for each risk value to the total number of single sentences contained in the teacher data 138 (step S634). S635).

The machine learning unit 76 evaluates the degree of bias in the number of sentence samples for each risk value (step S636). When the ratio of the number of sentences for each risk value exceeds a predetermined threshold value (YES in step S636), the machine learning unit 76 determines that the teacher data 138 of the corresponding risk value is biased, and oversamples (YES). Step S637) and step S637. On the other hand, when the ratio of the number of sentences for each risk value does not exceed a predetermined threshold value (NO in step S636), the machine learning unit 76 proceeds to step S637.

Instead of oversampling in step S637, if it is a premise that statistical significance can be guaranteed for teacher data 138, do not use a part of simple sentences with a high risk value in teacher data 138. May be good (downsampling).

When the risk value is set for each simple sentence for all of the learning appeal sentences 4t and the teacher data 138 with little variation can be prepared, the machine learning unit 76 performs machine learning of the knowledge base 137 using the teacher data 138. (Step S638).

Therefore, the discrimination model 139 for determining whether or not the product is suitable as an appeal for cosmetics is output to the storage unit 130 (step S639).

Next, the appeal sentence evaluation process by the appeal sentence evaluation unit 80 will be described. FIG. 10 is a flowchart for explaining the appeal sentence evaluation process in the second embodiment. In FIG. 10, first, on the client terminal 3, the appeal sentence 4p that the user wants to evaluate is input on the screen G90 (step S641), and the evaluation button 91a (FIG. 11) is pressed (step S643). The evaluation request 4r including the input appeal sentence 4p is transmitted to the server device 100.

When the server device 100 receives the evaluation request 4r, the appeal sentence evaluation unit 80 is notified, and in the appeal sentence evaluation unit 80, the monoculture processing unit 82 identifies the language from the country specified in the selection information 4s and evaluates it. The appeal sentence 4p included in the request 4r is monocultured (step S651), and one simple sentence is input to the simple sentence unit evaluation unit 84 (step S652).

The simple sentence unit evaluation unit 84 selects the discrimination model 139 from the product field specified in the selection information 4s, and uses the discrimination model 139 for discriminating the suitability as an appeal of cosmetics, and the input single sentence is a cosmetic product. The risk value corresponding to the suitability of the appeal is determined (step S653), and whether or not there is a risk is determined (step S654). That is, it may be determined from the definition table 140 of FIG. 6 whether or not the risk value is “2” or more.

When the risk value of the evaluation result is "2" or more, the result display data creation unit 86 acquires the explanatory text associated with the risk value using the definition table 140 (step S655).

When the risk value is "2" or more, the result display data creation unit 86 searches for a simple sentence similar to the input simple sentence using the knowledge base 137 and acquires it as an alternative plan (step S656).

After that, the result display data creation unit 86 creates the result display data 5r based on the evaluation result by using the explanatory text, the alternative plan, etc. obtained for the simple sentence having the risk value “2” or more (step S657). For simple sentences with a risk value of "2" or more in the appeal sentence 4p, the display can be distinguished by changing the character color, highlighting, underlining, etc., and alternatives can be displayed according to the specification of the simple sentence. The data to be used is added to the result display data 5r.

Upon completion of the processing of the result display data creation unit 86, the monoculture processing unit 82 determines whether or not there is the next unprocessed simple sentence (step S658). When the next unprocessed simple sentence exists (YES in step S658), the monoculture processing unit 82 returns to step S652, inputs the next unprocessed simple sentence to the single sentence unit evaluation unit 84, and performs the same as described above. The process is repeated.

On the other hand, when the next unprocessed simple sentence does not exist (NO in step S658), it is determined that the evaluation of the appeal sentence 4p has been completed, and the appeal sentence evaluation unit 80 has the result display data 5r stored in the storage unit 130. Is transmitted to the client terminal 3 that is the source of the evaluation request 4r (step S659).

On the client terminal 3, risky sentences and words are identifiablely displayed on the screen G90 with respect to the appeal sentence 4p input in step S641, and when those sentences or words are selected, the risk description is displayed. Is displayed. The description may include alternatives.

FIG. 11 is a diagram showing an example of an initial screen displayed on the client terminal in the second embodiment. The screen G90 as shown in FIG. 11 is displayed on the display device 15 of the client terminal 3. In FIG. 11, the screen G90 initially displayed has only the risk check area 9a. The risk check area 9a has an appeal text input area 7a, an evaluation button 7b, a reset button 7c, and the like.

The appeal sentence input area 7a is an area for the user to input the appeal sentence 4p, and in the initial state, a message or the like prompting the user to input the appeal sentence 4p may be displayed. The evaluation button 7b is a button for evaluating whether the appeal sentence 4p input in the appeal sentence input area 7a is appropriate as an appeal for cosmetics. In response to the pressing of the evaluation button 7b, the evaluation request 4r is transmitted to the server device 100. The reset button 7c is a button for returning the appeal text input area 7a to the initial state.

When the user inputs the appeal sentence 4p in the appeal sentence input area 7a of the screen G90 and presses the evaluation button 7b, the evaluation request 4r is transmitted from the client terminal 3 to the server device 100. When the result display data 5r is received from the server device 100, for example, the screen G90 undergoes a screen transition as shown in FIG.

FIG. 12 is a diagram showing a screen example at the time of receiving the result display data in the second embodiment. The screen G90 shown in FIG. 12 has an evaluation result region 9b in addition to the risk check region 9a. Since the configuration of the risk check area 9a is as described with reference to FIG. 11, the description thereof will be omitted.

The evaluation result area 9b has a storage format selection area 9c and an evaluation result display area 8a. The storage format selection area 9c is an area for selecting a data format on the client terminal 3 and storing the evaluation result. The evaluation result display area 8a has a display area 8m in which the evaluation result for the appeal sentence 4p input to the appeal sentence input area 7a of the risk check area 9a can be changed based on the result display data 5r received from the server device 100. ..

The display area 8 m is an area for displaying the risk value and the like and displaying the evaluation result in a changeable manner by reviewing by the user, such as ID, judgment sentence, risk value, number of risk words, similar sentences, and unknown words. Has an item.

The ID is an identification number assigned to each sentence by dividing the appeal sentence 4p input in the appeal sentence input area 7a of the risk check area 9a into sentence units. The judgment sentence indicates a delimited sentence. The risk value indicates the degree of suitability of cosmetics as a beginner to the text. The evaluation result indicates which of the three risk values is, and is displayed so that it can be corrected by a user such as an expert.

The number of risk words indicates the number of words that are judged to be at risk in the sentence. The similar sentence indicates whether or not there is a sentence similar to the appeal sentence 4p in the knowledge base 137. The number of unknown words indicates the number of words for which it cannot be determined whether or not it is appropriate for appealing cosmetics.

In this example, the appeal sentence 4p input by the user is only one sentence 7d "AAA completely erases acne scars rejuvenating dermal collagen." Therefore, in the evaluation result display area 8a, the ID "1" and the judgment sentence "AAA completely erase" are used. It is shown that there were erases acne scars rejuvenating ... ”, a risk value of“ 3 ”, a risk word number of“ 2 ”, a similar sentence“ yes ”, and an unknown word number of“ 0 ”. When the user presses the selection button 8p next to the ID "1", the details of the evaluation result are displayed, for example, as shown in FIG.

FIG. 13 is a diagram showing a display example of details of the evaluation result in the second embodiment. In FIG. 13, in the evaluation result area 9b of the screen G90, in addition to the storage format selection area 9c and the evaluation result display area 8a shown in FIG. 12, the judgment sentence display area 8b, the risk word display area 8c, and the unknown word display are displayed. Detailed information such as area 8d, similar sentence display area 8e, and remarks display area 8f is displayed.

The determination sentence display area 8b is an area for displaying the entire sentence 7d of the ID "1". In the risk word display area 8c, the risk word detected in the sentence 7d and the risk value thereof are shown. The unknown word display area 8d indicates a word that could not be determined in the sentence 7d. The similar sentence display area 8e shows a sentence similar to the sentence 7d searched from the knowledge base 137, its risk value, and the degree of similarity (%). Remarks The display area 8f displays explanations such as the basis of the risk value and countermeasures for the sentence 7d.

In this example, the sentence 7d "AAA completely erases acne scars rejuvenating dermal collagen." Is displayed in the judgment sentence display area 8b, and the "acne" and "dermal" parts 9r judged to be at risk are easily formed. It is displayed so that it can be identified. The part 9r may be shown in a color different from the color of other words or highlighted.

In the risk word display area 8c, it is shown that the risk value “2” is determined for each of the “acne” and “dermal” of the portion 9r that is discriminatively displayed in the judgment sentence display area 8b. The unknown word display area 8d indicates that it was not detected in this sentence 7d due to a blank or "-" or the like.

In the similar sentence display area 8e, three similar sentences are shown, and the risk value and the degree of similarity of each are shown. It is shown that the similar sentence “AAA erases △△△….” Had a risk value of “1” and a similarity of “77%”. In addition, it is shown that the similar sentence "○○○○ moderately ○○○○ ...." had a risk value of "1" and a similarity of "75%". Furthermore, it is shown that the similar sentence “AAA reduces □□□□….” Had a risk value of “1” and a similarity of “75%”.

Remarks In the display area 8f, an explanatory text such as "out of the range of cosmetics" is displayed. This description may be modified by a user such as an experienced expert.

In FIGS. 12 and 13, an example of the appeal sentence 4p having only one sentence 7d has been described, but the appeal sentence 4p may be composed of a plurality of sentences. When the appeal sentence 4p contains a plurality of sentences, the evaluation result is shown for each sentence.

FIG. 14 is a diagram showing another screen example when the result display data in the second embodiment is received. In FIG. 14, the configuration of the screen G90 is the same as that in FIG. 12, so the description thereof will be omitted. FIG. 14 shows a case where the user inputs an appeal sentence 4p composed of a plurality of sentences 7e, 7f, 7g, 7h, and 7i in the risk check area 9a of the screen G90.

Specifically, the appeal sentence 4p is "BBB cream has been repackaged and is now available in a newly designed pack." In sentence 7e and "○○○○○○○○○○○." In sentence 7f. , Sentence 7g "This product is 100% protection against uv.", Sentence 7h "△△△△ ....", and Sentence 7i "□□□□□□□ ..." ..

The server device 100 divides the appeal sentence 4p into each sentence. For example, this appeal sentence 4p is divided into 5 sentences. The server device 100 assigns an ID to each of the five divided sentences, and the single sentence unit evaluation unit 84 obtains an evaluation result.

In the evaluation result area 9b, each of the five sentences obtained by dividing the appeal sentence 4p is identified by ID "1", ID "2", ID "3", ID "4", and ID "5", and the evaluation of each sentence is performed. The results are shown. In this example, the evaluation result of the sentence 7e with the ID “1” indicates a risk value “1”, a risk word number “0”, a similar sentence “none”, an unknown word number “0”, and the like. The evaluation result of the sentence 7f of the ID "2" indicates a risk value "2", a risk word number "1", a similar sentence "none", an unknown word number "0", and the like.

Further, the evaluation result of the sentence 7g of the ID "3" indicates a risk value "3", a risk word number "1", a similar sentence "yes", an unknown word number "0", and the like. The evaluation result of the sentence 7h of the ID "4" indicates a risk value "3", a risk word number "1", a similar sentence "yes", an unknown word number "0", and the like. The evaluation result of the sentence 7i of the ID "5" indicates a risk value "1", a risk word number "0", a similar sentence "none", an unknown word number "0", and the like.

Here, it is desirable to display the level of risk in an identifiable manner. As an example, the sentences 7g and 7h of ID "3" and ID "4" indicating the highest risk value "3" are shown in a reddish color, and the ID "2" indicating the next highest risk value "2" is shown. The sentences 7f of "" are shown in a yellowish color, and the sentences 7e and 7i of ID "1" and ID "5" having the smallest risk value (determined to have no risk) may be shown in a blueish color.

By moving the cursor closer to the "risk value" in the evaluation result display area 8a in the evaluation result area 9b of the screen G90 of FIG. 14, the risk value legend 9d can be displayed. An explanation is given for each value of the risk value indicating the level of risk in the legend 9d. For example, the value "1" is "an expression of the cosmetics category and there is no problem", the value "2" is "there is a risk of being taken as being out of the cosmetics category", and the value "3" is " It clearly deviates from the category of cosmetics, and explanations such as "NG" are displayed.

When the user presses the selection button 8p of the ID "3" on the screen G90, the details of the evaluation result as shown in FIG. 15 are displayed.

FIG. 15 is a diagram showing another display example of the details of the evaluation result in the second embodiment. In FIG. 15, the evaluation result area 9b of the screen G90 shows the details of the evaluation result of the sentence 7g of the ID “3” selected by the user.

In this case, the sentence 7g "This product is 100% protection against uv." Is displayed in the judgment sentence display area 8b, and the part 9t of "100% protection against uv" judged to be at risk can be easily displayed. It is displayed so that it can be identified. The part 9t may be shown in a color different from the color of other words or highlighted.

In the risk word display area 8c, it is shown that the “100% protection against uv” of the portion 9r identifiablely displayed in the judgment sentence display area 8b is determined to be the risk value “3”. The unknown word display area 8d indicates that it was not detected in this sentence 7d due to a blank or "-" or the like.

In the similar sentence display area 8e, three similar sentences are shown, and the risk value and the degree of similarity of each are shown. The similar sentence "The product is environmentally-friendly and cruelty-free." Is shown to have a risk value of "1" and a similarity of "36%". In addition, the similar sentence "The product is said to help *** **." Is shown to have a risk value of "1" and a similarity of "34%". Furthermore, the similar sentence "This product retails in 100g pack." Is shown to have a risk value of "1" and a similarity of "30%".

Remarks In the display area 8f, an explanatory text such as "out of the category of cosmetics" is displayed. This description may be modified by a user such as an experienced expert.

In this way, in the second embodiment, the risky expression in the appeal sentence 4p is automatically discriminated, and the reason for the discrimination, the alternative plan, etc. are displayed to provide a stable evaluation to any user. Can be done. In addition, high-risk words and sentences can be easily changed to preferable expressions. Further, for the expert, the time required for inquiries regarding the appeal text from various departments can be reduced, and the burden can be reduced.

[Third Example]
Next, the case of evaluating the description of food based on Japanese laws and regulations in the above-mentioned first embodiment will be described as a third embodiment with a definition table, teacher data, and an example of evaluation results. In the third embodiment, the details of the functional configuration are the same as those in the second embodiment, and thus are omitted.

FIG. 16 is a diagram showing a data configuration example of the definition table in the third embodiment. In FIG. 16, the definition table 240 is a table in which a risk value and an explanation are associated with each evaluation viewpoint, as in the second embodiment, and has items such as an evaluation viewpoint, a risk value, and an explanation.

The evaluation viewpoint shows, for example, the suitability as an explanation of food, as in the first embodiment. For words that correspond to the evaluation items specified in advance by the expert, the high risk within the evaluation items is specified. The risk value indicates the high risk. There may be multiple evaluation viewpoints.

In the explanation, explanations for each of the risk values are set. In the result display data 5r transmitted to the client terminal 3, the evaluation explanatory information 5p includes the explanatory text corresponding to the risk value determined by the simple sentence unit evaluation unit 84.

In this data structure example, the risk value is defined from "1" to "3". When the risk value is "3", it indicates that "the description of the food corresponds to a misidentification label and is NG". When the risk value is "2", it indicates that "the description of the food has a risk of being taken if it corresponds to a misidentification label". When the risk value is "1", it indicates that "the description of the food is an appropriate expression and there is no problem".

As in the second embodiment, the explanation based on the risk value from the evaluation viewpoint is given on the screen G90, so that it is stable even for users in the department who do not have sufficient expertise in the explanation of food. An evaluation can be provided.

FIG. 17 is a diagram showing a data structure example of teacher data in the third embodiment. In FIG. 17, the teacher data 238 is a table showing the evaluation results for each explanatory text obtained by the same processing unit as the risk value acquisition unit 74 of the second embodiment, and is a product name, evaluation result, explanatory text, etc. It has the item of.

The product name indicates the name of the target product in which the description is described. The product name may be omitted. The value of the evaluation result is the risk value of the definition table 240 as illustrated in FIG. The descriptive text indicates the evaluated text.

In the data structure example of FIG. 17, Japanese explanations are mainly shown, but other languages may be used. As an example, the description evaluated for the product with the product name "Aiueo" is "Dramatically reduce body fat with a drink once a day without exercise or dietary restrictions !!" The evaluation result shows a risk value of "3", that is, it is clearly shown that the description of the food corresponds to the misidentification label and is NG.

FIG. 17 shows an example when there is one evaluation viewpoint, but when two or more evaluation viewpoints are selected, the risk value for each evaluation viewpoint is shown. In addition, although the explanatory text is shown in sentences for ease of understanding, the coordinates in the text vector space may be shown.

FIG. 18 is a diagram showing a display example of details of the evaluation result in the third embodiment. In FIG. 18, as in the second embodiment, the evaluation result area 9b of the screen G90 shows the details of the evaluation result of the sentence 7k included in the explanatory text 4q.

In this case, in the judgment sentence display area 8b, the sentence 7k "3 tablets a day improves immunity! Excellent effect on prevention of lifestyle-related diseases and cancer !!" The part 9w of "excellent effect on cancer prevention" is displayed in an easily distinguishable manner. The part 9w may be shown in a color different from the color of other words or highlighted.

In the risk word display area 8c, it is shown that the “excellent effect on cancer prevention” of the portion 9w identifiablely displayed in the judgment sentence display area 8b is determined to be the risk value “3”. The unknown word display area 8d indicates that it was not detected in this sentence 7k due to a blank or "-" or the like.

In the similar sentence display area 8e, one similar sentence is shown, and the risk value and the degree of similarity are shown. The similar sentence "You can supplement the nutritional components that tend to be deficient in your daily diet" is shown to have a risk value of "1" and a similarity of "29%".

Remarks In the display area 8f, an explanatory text such as "corresponding to a misidentified display of food" is displayed. This description may be modified by a user such as an experienced expert.

In this way, in the third embodiment, the risky expression in the explanatory text 4q is automatically discriminated, and the reason for the discriminant, the alternative plan, etc. are displayed to provide a stable evaluation to any user. Can be done. In addition, high-risk words and sentences can be easily changed to preferable expressions. Further, for a skilled person, the time required for inquiries about explanatory texts from various departments can be reduced, and the burden can be reduced.

2 Network 3 Client terminal 4r Evaluation request 4p Appeal text 4q Description 4s Selection information 5r Result display data 5p Evaluation explanation information 60 Knowledge base creation department 62 Appeal text collection department 64 Monocultural processing department 66 Distributed expression analysis department 70 Discrimination model creation department 72 Monoculture processing unit 74 Risk value acquisition unit 76 Machine learning unit 80 Appeal sentence evaluation unit 82 Monoculture processing unit 84 Single sentence unit evaluation unit 86 Result display data creation unit 100 Server device 11, 111 CPU
12, 112 Main storage device 13, 113 Auxiliary storage device 14, 114 Input device 15, 115 Display device 17, 117 Communication I / F
18, 118 Drive device 19, 119 Storage medium 30, 130 Storage unit 137 Knowledge base 138 Teacher data 139 Discrimination model G90 Screen 1137 Description DB
1150 Model Creation Verification Department 1180 Description Evaluation Department

Claims (10)

In response to the reception of the evaluation request of the description attached to the product, the monoculture processing step to monoculture the description and
A simple sentence unit evaluation step for evaluating whether or not the content of the simple sentence is at least in accordance with the laws and regulations related to the explanatory expression of the product for each simple sentence using a model obtained by machine learning in advance.
By referring to the definition table in which sentences related to the suitability as a description of the product are associated with each evaluation result, the sentence corresponding to the evaluation result by the single sentence unit evaluation step is acquired, and the description is displayed on the screen. In the result display data creation step of creating the result display data that makes it possible to display the sentence according to the selection of the simple sentence,
A description evaluation method performed by a computer. The result display data creation step further
With reference to the knowledge base that accumulates the analysis results of distributed expressions for the explanations collected in advance, a sentence similar to the simple sentence of the explanation of the evaluation request is acquired as an alternative, and the explanation is displayed on the screen. The explanatory text evaluation method according to claim 1, wherein, in addition to the text, result display data that enables display of alternatives is created according to the selection of the simple text. The simple sentence unit evaluation step is
The explanatory text evaluation method according to claim 1 or 2, wherein the appeal text of the evaluation request is evaluated in units of simple sentences. The evaluation request specifies at least one of the product fields, countries, and evaluation viewpoints to which the product belongs.
The explanatory text evaluation method according to any one of claims 1 to 3, further comprising a model selection step for selecting the model corresponding to the designation of the evaluation request. In the model selection step, when two or more evaluation viewpoints are specified, the model is selected for each evaluation viewpoint.
The explanatory text evaluation method according to claim 4, wherein the simple sentence unit evaluation step evaluates the content of the simple sentence for each selected model in accordance with the above-mentioned laws and regulations. A plurality of explanations related to the plurality of products collected for creating the model are collected, sampled from the collected explanations, the appropriateness of the sampled explanations is judged in accordance with the above-mentioned laws and regulations, and the result of the judgment is obtained. Collection and structuring steps to associate and structure,
A feature extraction step of analyzing the sampled explanatory text to extract sentence features to generate a sentence vector space, and
A model creation step for creating the model by classifying the sentence features by machine learning using the explanatory text for which the appropriateness has been determined as teacher data.
The explanatory text evaluation method according to any one of claims 1 to 5, further comprising. In the model creation step, the teacher data is classified using a multi-support vector machine, and if the number of teacher data for each classification is biased, the text vector space is located near a predetermined distance. The explanatory text evaluation method according to claim 6, wherein oversampling is performed by adding points so that document feature points exist. Any of claims 6 to 7, wherein the collection / structuring step, the feature extraction step, and the model creation step are repeated for each combination of the product field, country, and evaluation viewpoint to which the product belongs. The explanation evaluation method described in one item. In response to the reception of the evaluation request of the explanation attached to the product, the monoculture processing department that monocultures the explanation, and
A simple sentence unit evaluation unit that evaluates whether or not the content of the simple sentence is at least in accordance with the laws and regulations related to the explanatory expression of the product for each simple sentence using a model obtained by machine learning in advance.
By referring to the definition table in which sentences related to the suitability as a description of the product are associated with each evaluation result, the sentence corresponding to the evaluation result by the single sentence unit evaluation unit is acquired, and the description is displayed on the screen. In the result display data creation unit that creates the result display data that makes it possible to display the sentence according to the selection of the simple sentence,
Explanatory text evaluation device with. Computer,
A monoculture processing means that monocultures the description in response to the reception of the evaluation request of the description attached to the product.
A simple sentence unit evaluation means for evaluating whether or not the content of the simple sentence is at least in accordance with the laws and regulations related to the explanatory expression of the product for each simple sentence using a model obtained by machine learning in advance.
By referring to the definition table in which sentences related to the suitability as a description of the product are associated with each evaluation result, the sentence corresponding to the evaluation result by the simple sentence unit evaluation means is acquired, and the description is displayed on the screen. Result display data creation means for creating result display data that enables the sentence to be displayed according to the selection of the simple sentence.
An explanatory text evaluation program that functions as.

Images