Advertisement text generation system
By fine-tuning existing learning models to enhance the relationship between advertising text appeal information and communication reasons, the system improves the quality of advertising text generation from 61.6 to 73.3 points, addressing the low-quality issue in conventional methods.
Patent Information
- Application Number
- JP2025037094
- Authority / Receiving Office
- JP · JP
- Patent Type
- Applications
- Current Assignee / Owner
- Priority Date
- 2024-05-13
- Filing Date
- 2025-03-10
- Publication Date
- 2025-11-26
Smart Images
Figure 2025172689000001_ABST
Abstract
Description
[Technical Field]
[0001] The present invention relates to an advertising text generation system that generates advertising text using an existing learning model. [Background technology]
[0002] Conventionally, a device has been proposed that generates advertising text (e.g., a catchphrase for a live streamer) using an existing learning model (e.g., GPT (Generative Pre-trained Transformer)) (see Patent Literature 1). This conventional device receives text data converted from video data of the streamer's live stream and information about the streamer as input, and outputs a catchphrase and introduction for the streamer. [Prior art documents] [Patent documents]
[0003] [Patent Document 1] Patent No. 7316598 Summary of the Invention [Problem to be solved by the invention]
[0004] However, as with conventional devices, simply generating advertising text using an existing learning model (such as "GPT") had the problem of low quality advertising text (for example, if a copywriter creates 15 catchy slogans, the average evaluation score is 69.3 points, whereas if 15 catchy slogans are generated simply using an existing learning model, the average evaluation score is 61.6 points).
[0005] The present invention has been made in consideration of the above-mentioned problems, and aims to provide a catchphrase generation system that can improve the quality of advertising text generated using existing learning models. [Means for solving the problem]
[0006] The advertising text generation system of the present invention includes a first system that generates advertising text appeal information indicating what is desired to be communicated about an advertising text target based on advertising text target information related to the advertising text target for which advertising text is to be generated; a second system that evaluates the advertising text appeal information generated by the first system and outputs an evaluation score for the advertising text appeal information; a third system that generates advertising text indicating how to communicate what is desired to be communicated in the advertising text appeal information based on the advertising text appeal information; and a fourth system that evaluates the advertising text generated by the third system and outputs an evaluation score for the advertising text, wherein the third system applies the following to an existing learning model generated by machine learning using predetermined learning data: The system is equipped with an additional learning unit that performs additional learning through fine tuning of the relationship between predetermined advertising text appeal information, advertising text indicating how to communicate what is intended to be communicated in the advertising text appeal information, and the reason why the intended communication in the advertising text appeal information leads to the intended communication; an input unit that receives advertising text appeal information selected from the advertising text appeal information generated by the first system based on the evaluation points output by the second system; and an estimation output unit that, based on the relationship additionally learned by the additional learning unit, uses the advertising text appeal information input from the input unit as input and estimates and outputs advertising text indicating how to communicate what is intended to be communicated in the advertising text appeal information and the reason why the intended communication in the advertising text appeal information leads to the intended communication.
[0007] According to this configuration, a first system generates advertising text appeal information (e.g., "Your health depends on whether you stay hydrated") indicating what the advertising text target (e.g., "soft drinks") wants to convey from advertising text target information (e.g., "Make people aware of the symptoms of dehydration"), and a second system outputs an evaluation score (e.g., "85 points") for the advertising text appeal information. A third system generates advertising text (e.g., "Drink to live") indicating how to communicate what the advertising text appeal information wants to convey from the advertising text appeal information, and a fourth system outputs an evaluation score (e.g., "90 points") for the advertising text.
[0008] In this case, the third system uses an existing learning model that has been additionally trained through fine tuning to learn the relationship between the advertising text appeal information and the advertising text and reasons (the reasons that lead from what you want to communicate (WHAT) to how to communicate it (HOW)), and when advertising text appeal information (advertising text appeal information generated by the first system that is selected based on the evaluation points output by the second system) is input, the advertising text and reasons (the reasons that lead from what you want to communicate (WHAT) to how to communicate it (HOW)) are estimated and output.
[0009] If advertising text is generated from advertising text appeal information simply using an existing learning model (such as "GPT"), the quality of the generated advertising text will be low (for example, if 15 advertising texts are generated, the average evaluation score will be 61.6 points). However, if advertising text and reasons are generated from advertising text appeal information using an existing learning model that has been additionally trained, as described above, to not only the relationship between advertising text appeal information and advertising text, but also the relationship between the reasons (the reasons leading from what you want to communicate (WHAT) to how you communicate it (HOW)), the quality of the generated advertising text will be high (for example, if 15 advertising texts are generated, the average evaluation score will be 73.3 points). In this way, the quality of advertising text generated using existing learning models can be improved.
[0010] In addition, in the advertising text generation system of the present invention, the fourth system may include a second additional learning unit that uses fine tuning to additionally learn the relationship between a predetermined advertising text, an evaluation score for the advertising text, and the reason for the evaluation score for the advertising text using an existing learning model generated by machine learning using predetermined learning data; a second input unit that receives advertising text selected by a user from the advertising text generated by the third system; and a second estimation output unit that uses the advertising text input from the second input unit as input based on the additional learning relationship in the second additional learning unit and estimates and outputs the evaluation score for the advertising text and the reason for the evaluation score for the advertising text.
[0011] According to this configuration, the fourth system uses an existing learning model that has been additionally trained through fine tuning to learn the relationship between the advertising text and the evaluation score and reason (reason for that evaluation score) of the advertising text, and when advertising text (advertising text selected by the user from the advertising text generated by the third system) is input, the evaluation score and reason (reason for that evaluation score) of the advertising text is estimated and output.
[0012] If an evaluation score for advertising text is generated from advertising text simply using an existing learning model (such as "GPT"), the accuracy of the evaluation score for the generated advertising text will be low, whereas if an evaluation score and reason for advertising text are generated from advertising text using an existing learning model that has been additionally trained, as described above, to learn not only the relationship between the advertising text and the evaluation score for the advertising text but also the reason (the reason for setting that evaluation score), the accuracy of the evaluation score for the generated advertising text will be higher. In this way, the accuracy of the evaluation score for advertising text generated using an existing learning model can be improved.
[0013] Furthermore, in the advertising text generation system of the present invention, the first system may include a third additional learning unit that uses fine tuning to additionally learn, in an existing learning model generated by machine learning using predetermined learning data, the relationship between advertising text object information about a predetermined advertising text object, advertising text appeal information indicating what is desired to be communicated about the advertising text object, and the reason for selecting the advertising text appeal information; a third input unit that receives input of advertising text object information about the advertising text object for which advertising text is to be generated; and a third estimation output unit that uses, as input, the advertising text object information about the advertising text object input from the third input unit, based on the additional learning made by the third additional learning unit, and estimates and outputs the advertising text appeal information indicating what is desired to be communicated about the advertising text object and the reason for selecting the advertising text appeal information.
[0014] According to this configuration, the first system uses an existing learning model that has been additionally trained through fine tuning to learn the relationship between the advertising text target information and the advertising text appeal information and reasons (reasons for choosing that advertising text appeal information), and when advertising text target information is input, the advertising text appeal information and reasons (reasons for choosing that advertising text appeal information) are estimated and output.
[0015] If advertising text appeal information is generated from advertising text target information simply using an existing learning model (such as "GPT"), the quality of the generated advertising text appeal information will be low. However, if, as described above, an existing learning model that has been additionally trained to not only learn the relationship between the advertising text target information and the advertising text appeal information, but also the relationship between the advertising text target information and the reason (the reason for selecting that advertising text appeal information), generates advertising text appeal information and the reason from the advertising text target information, the quality of the generated advertising text appeal information will be high. In this way, the quality of advertising text appeal information generated using an existing learning model can be improved.
[0016] In addition, in the advertising text generation system of the present invention, the second system may include a fourth additional learning unit that uses fine tuning to additionally learn the relationship between predetermined advertising text appeal information, the evaluation score of the advertising text appeal information, and the reason for the evaluation score, in an existing learning model generated by machine learning using predetermined learning data; a fourth input unit that receives advertising text appeal information selected by a user from the advertising text appeal information generated by the first system; and a fourth estimation output unit that uses the advertising text appeal information input from the fourth input unit as input, based on the additional learning relationship caused by the fourth additional learning unit, and estimates and outputs the evaluation score of the advertising text appeal information and the reason for the evaluation score.
[0017] According to this configuration, the second system uses an existing learning model that has been additionally trained through fine tuning to learn the relationship between the advertising text appeal information and the evaluation score and reason (reason for that evaluation score) of the advertising text appeal information, and when advertising text appeal information (advertising text appeal information selected by the user from the advertising text appeal information generated by the first system) is input, the second system estimates and outputs the evaluation score and reason (reason for that evaluation score) of the advertising text appeal information.
[0018] If an evaluation score of advertising text appeal information is generated from advertising text appeal information simply using an existing learning model (such as "GPT"), the accuracy of the evaluation score of the generated advertising text appeal information will be low. However, if an evaluation score and reason for the advertising text appeal information are generated from advertising text appeal information using an existing learning model that has additionally learned not only the relationship between the advertising text appeal information and the evaluation score of the advertising text appeal information but also the reason (the reason for the evaluation score) as described above, the accuracy of the evaluation score of the generated advertising text appeal information will be higher. In this way, the accuracy of the evaluation score of advertising text appeal information generated using an existing learning model can be improved.
[0019] The advertising text generation system of the present invention includes an additional learning unit that additionally learns, through fine tuning, the relationship between predetermined advertising text appeal information indicating what is desired to be communicated about the target of the advertising text, advertising text indicating how to communicate what is desired to be communicated in the advertising text appeal information, and the reason why the desired communication method is reached from the desired communication method in the advertising text appeal information in an existing learning model generated by machine learning using predetermined learning data; an input unit that receives advertising text appeal information generated based on advertising text target information related to the advertising text target for which advertising text is to be generated; and an estimation output unit that, based on the relationship additionally learned by the additional learning unit, uses the advertising text appeal information input from the input unit as input and estimates and outputs advertising text indicating how to communicate what is desired to be communicated in the advertising text appeal information and the reason why the desired communication method is reached from the desired communication method in the advertising text appeal information.
[0020] According to this configuration, an existing learning model is used that has been additionally trained through fine tuning to learn the relationship between advertising text appeal information, advertising text, and reasons (reasons for going from what you want to communicate (WHAT) to how to communicate it (HOW)), and when advertising text appeal information is input, advertising text and reasons (reasons for going from what you want to communicate (WHAT) to how to communicate it (HOW)) are estimated and output.
[0021] If advertising text is generated from advertising text appeal information simply using an existing learning model (such as "GPT"), the quality of the generated advertising text will be low (for example, if 15 advertising texts are generated, the average evaluation score will be 61.6 points). However, if advertising text and reasons are generated from advertising text appeal information using an existing learning model that has been additionally trained, as described above, to not only the relationship between advertising text appeal information and advertising text, but also the relationship between the reasons (the reasons leading from what you want to communicate (WHAT) to how you communicate it (HOW)), the quality of the generated advertising text will be high (for example, if 15 advertising texts are generated, the average evaluation score will be 73.3 points). In this way, the quality of advertising text generated using existing learning models can be improved.
[0022] The advertising text generation system of the present invention may also include a second additional learning unit that uses fine tuning to additionally learn the relationship between a predetermined advertising text, an evaluation score for the advertising text, and the reason for the evaluation score for the advertising text, in an existing learning model generated by machine learning using predetermined learning data; a second input unit that receives advertising text selected by a user from the advertising texts output from the estimation output unit; and a second estimation output unit that uses the advertising text input from the second input unit as input based on the additional learning relationship in the second additional learning unit, and estimates and outputs the evaluation score for the advertising text and the reason for the evaluation score for the advertising text.
[0023] According to this configuration, an existing learning model is used that has been additionally trained through fine tuning to learn the relationship between the advertising text and the evaluation score and reason (reason for that evaluation score) of the advertising text, and when advertising text (advertising text selected by the user from the advertising text output from the estimation output unit) is input, the evaluation score and reason (reason for that evaluation score) of the advertising text is estimated and output.
[0024] If an evaluation score for advertising text is generated from advertising text simply using an existing learning model (such as "GPT"), the accuracy of the evaluation score for the generated advertising text will be low, whereas if an evaluation score and reason for advertising text are generated from advertising text using an existing learning model that has been additionally trained, as described above, to learn not only the relationship between the advertising text and the evaluation score for the advertising text but also the reason (the reason for setting that evaluation score), the accuracy of the evaluation score for the generated advertising text will be higher. In this way, the accuracy of the evaluation score for advertising text generated using an existing learning model can be improved.
[0025] The program of the present invention is a program executed in an advertising text generation system, and the program causes the advertising text generation system to execute the following processes: a process of additionally learning, through fine tuning, a relationship between predetermined advertising text appeal information indicating what is desired to be communicated about the target of the advertising text, advertising text indicating how to communicate what is desired to be communicated in the advertising text appeal information, and the reason why the desired communication in the advertising text appeal information leads to the desired communication; a process of inputting advertising text appeal information generated based on advertising text target information related to the advertising text target for which advertising text is to be generated; and a process of using the advertising text appeal information input from the input unit as input, based on the relationship additionally learned by the additional learning unit, to estimate and output advertising text indicating how to communicate what is desired to be communicated in the advertising text appeal information and the reason why the desired communication in the advertising text appeal information leads to the desired communication.
[0026] Like the system described above, this program uses an existing learning model that has been fine-tuned to learn the relationship between advertising text appeal information and the advertising text and reasons (the reasons that lead from what you want to communicate (WHAT) to how to communicate it (HOW)).When advertising text appeal information is input, the advertising text and reasons (the reasons that lead from what you want to communicate (WHAT) to how to communicate it (HOW)) are estimated and output.
[0027] If advertising text is generated from advertising text appeal information simply using an existing learning model (such as "GPT"), the quality of the generated advertising text will be low (for example, if 15 advertising texts are generated, the average evaluation score will be 61.6 points). However, if advertising text and reasons are generated from advertising text appeal information using an existing learning model that has been additionally trained, as described above, to not only the relationship between advertising text appeal information and advertising text, but also the relationship between the reasons (the reasons leading from what you want to communicate (WHAT) to how you communicate it (HOW)), the quality of the generated advertising text will be high (for example, if 15 advertising texts are generated, the average evaluation score will be 73.3 points). In this way, the quality of advertising text generated using existing learning models can be improved.
[0028] The method of the present invention is a method executed in an advertising text generation system, and the method includes: additionally learning, through fine tuning, a relationship between predetermined advertising text appeal information indicating what is desired to be communicated about the target of the advertising text, advertising text indicating how to communicate what is desired to be communicated in the advertising text appeal information, and a reason why the desired communication method is reached from what is desired to be communicated in the advertising text appeal information, in an existing learning model generated by machine learning using predetermined learning data; receiving advertising text appeal information generated based on advertising text target information related to the advertising text target for which advertising text is to be generated; and using the advertising text appeal information input from the input unit as input based on the relationship additionally learned by the additional learning unit, estimating and outputting advertising text indicating how to communicate what is desired to be communicated in the advertising text appeal information and a reason why the desired communication method is reached from what is desired to be communicated in the advertising text appeal information, using the advertising text appeal information input from the input unit as input.
[0029] With this method, as with the above system, an existing learning model is used that has been fine-tuned to learn the relationship between the advertising text appeal information and the advertising text and reasons (the reasons for going from what you want to communicate (WHAT) to how to communicate it (HOW)), and when advertising text appeal information is input, the advertising text and reasons (the reasons for going from what you want to communicate (WHAT) to how to communicate it (HOW)) are estimated and output.
[0030] If advertising text is generated from advertising text appeal information simply using an existing learning model (such as "GPT"), the quality of the generated advertising text will be low (for example, if 15 advertising texts are generated, the average evaluation score will be 61.6 points). However, if advertising text and reasons are generated from advertising text appeal information using an existing learning model that has been additionally trained, as described above, to not only the relationship between advertising text appeal information and advertising text, but also the relationship between the reasons (the reasons leading from what you want to communicate (WHAT) to how you communicate it (HOW)), the quality of the generated advertising text will be high (for example, if 15 advertising texts are generated, the average evaluation score will be 73.3 points). In this way, the quality of advertising text generated using existing learning models can be improved. [Effects of the Invention]
[0031] According to the present invention, it is possible to improve the quality of advertising text generated using an existing learning model. [Brief explanation of the drawings]
[0032] [Figure 1] 1 is a block diagram showing a configuration of an advertising text generation system according to an embodiment of the present invention. [Figure 2] FIG. 2 is a diagram showing an example (data sample) of training data for additional learning (fine tuning) performed in the first system of the advertising text generation system. [Figure 3] FIG. 10 is a diagram showing an example (data sample) of training data for additional learning (fine tuning) performed in the second system of the advertising text generation system. [Figure 4] FIG. 10 is a diagram showing an example (data sample) of training data for additional learning (fine tuning) performed in the third system of the advertising text generation system. [Figure 5] FIG. 10 is a diagram showing an example (data sample) of training data for additional learning (fine tuning) performed in the fourth system of the advertising text generation system. [Figure 6] FIG. 2 is a sequence diagram illustrating the operation of the advertising text generation system according to the embodiment of the present invention. [Figure 7] FIG. 2 is a sequence diagram illustrating the operation of the advertising text generation system according to the embodiment of the present invention. [Figure 8] FIG. 2 is a sequence diagram illustrating the operation of the advertising text generation system according to the embodiment of the present invention. [Figure 9] FIG. 2 is a diagram showing an example of input data (advertising text target information) and output data (advertising text appeal information and reasons) of the first system in the embodiment of the present invention. [Figure 10] FIG. 10 is a diagram showing an example of input data (advertising text appeal information) and output data (evaluation points and reasons) of the second system in the embodiment of the present invention. [Figure 11(a)] FIG. 10 is a diagram showing examples of input data (advertising text appeal information) and output data (advertising text and reason) of the third system in the embodiment of the present invention. [Figure 11(b)] FIG. 10 is a diagram showing examples of input data (advertising text appeal information) and output data (advertising text and reason) of the third system in the embodiment of the present invention. [Figure 12] FIG. 10 is a diagram showing an example of input data (advertising text) and output data (evaluation points and reasons) of the fourth system in the embodiment of the present invention. [Figure 13] FIG. 10 is a block diagram showing the configuration of an advertising text generation system according to another embodiment. DETAILED DESCRIPTION OF THE INVENTION
[0033] An advertising text generation system according to an embodiment of the present invention will be described below with reference to the accompanying drawings. In this embodiment, a business support system used in a cloud service or the like that supports the creation of advertising text will be exemplified. The advertising text generation system according to this embodiment has a function for generating advertising text using an existing learning model. These functions are realized by a program stored in the memory area of the advertising text generation system.
[0034] The configuration of an advertising text generation system according to an embodiment of the present invention will be described with reference to the drawings. In this embodiment, a case where a catchy slogan is generated using an existing learning model (i.e., the advertising text is a catchy slogan) will be described. FIG. 1 is a block diagram showing the configuration of the advertising text generation system according to this embodiment. As shown in FIG. 1, advertising text generation system 1 is connected to user device 2 via a network N such as the Internet. Advertising text generation system 1 is configured, for example, by a cloud server device. User device 2 is configured, for example, by a business computer device, and includes an input unit 3 such as a keyboard and a mouse, and a display unit 4 such as a display.
[0035] The advertising text generation system 1 includes a first system 10, a second system 20, a third system 30, and a fourth system 40. The first system 10 has a function of generating catchphrase appeal information indicating what the advertiser wants to communicate about a catchphrase target based on catchphrase target information related to the target for which the catchphrase is to be generated. The second system 20 has a function of evaluating the catchphrase appeal information generated by the first system 10 and outputting an evaluation score for the catchphrase appeal information. The third system 30 has a function of generating a catchphrase indicating how to communicate what the advertiser wants to communicate with the catchphrase appeal information based on the catchphrase appeal information. The fourth system 40 has a function of evaluating the catchphrase generated by the third system 30 and outputting an evaluation score for the catchphrase. Here, the catchphrase corresponds to the advertising text of the present invention, and the catchphrase target corresponds to the advertising text target of the present invention. The catchphrase target information corresponds to the advertising text target information of the present invention, and the catchphrase appeal information corresponds to the advertising text appeal information of the present invention.
[0036] The first system 10 includes an additional learning unit 11, a memory unit 12, an input unit 13, and an estimation output unit 14. The additional learning unit 11 of the first system 10 has a function of additionally learning the relationship between catchphrase target information, catchphrase appeal information, and the reason (reason for selecting that catchphrase appeal information) through fine tuning in an existing learning model. The catchphrase target information is information about a predetermined catchphrase target, and the catchphrase appeal information is information indicating what you want to communicate about the catchphrase target (WHAT). The existing learning model is a learning model (e.g., "GPT") generated by machine learning using predetermined learning data. Any method, such as deep learning using a neural network, is used for machine learning. Furthermore, known methods can be used for additional learning (fine tuning). The additional learning unit 11 of the first system 10 corresponds to the third additional learning unit of the present invention.
[0037] FIG. 2 shows an example of training data (data sample) for additional training (fine tuning) performed by the first system 10. When additional training of an existing learning model is performed by the first system 10, for example, as shown in FIG. 2, information on the product name or service name (e.g., "Smart Sound Pillow") and its supplemental information (e.g., "Smart Sound Pillow is...") are used as the "catchphrase target information" of the input data. In addition, for example, "Providing comfortable sleep with relaxing sounds" is used as the "catchphrase appeal information (WHAT)" of the output data, and for example, "For users who want to improve the quality of their sleep" is used as the "reason (reason for the WHAT)" of the output data. Such training data can be prepared in advance (e.g., prepared in advance by an experienced copywriter).
[0038] The storage unit 12 of the first system 10 is configured with a large-capacity memory or the like, and stores data and programs necessary for performing the processing of the first system 10 (processing for generating catchphrase appeal information based on catchphrase target information). In addition, the storage unit 12 stores, for example, training data for additional learning (fine tuning) performed by the first system 10.
[0039] The input unit 13 of the first system 10 receives catchphrase target information (information relating to a catchphrase target for which a catchphrase is to be generated) input (user input) via the input unit 3 of the user device 2. The input unit 13 of the first system 10 corresponds to the third input unit of the present invention.
[0040] The estimation output unit 14 of the first system 10 receives the target information for the catchphrase (e.g., "Relaxing sleep aids") input from the input unit 13 of the first system 10 based on the relationships additionally learned by the additional learning unit 11 of the first system 10, and estimates and outputs the appealing information for the catchphrase (e.g., "Provides ultimate relaxation") and the reason for it (e.g., "High-quality materials and ergonomic design provide maximum comfort") (see FIG. 9). For example, in the case of a neural network that has undergone the above additional learning (fine tuning), estimation is performed by inputting the target information for the catchphrase into the input layer and outputting the appealing information for the catchphrase and the reason for it from the output layer. The estimation output unit 14 of the first system 10 corresponds to the third estimation output unit of the present invention.
[0041] The second system 20 includes an additional learning unit 21, a memory unit 22, an input unit 23, and an estimation output unit 24. The additional learning unit 21 of the second system 20 has a function of additionally learning, by fine tuning, the relationship between predetermined catchphrase appeal information, the evaluation score of that catchphrase appeal information, and the reason (the reason for that evaluation score) for that catchphrase appeal information. The existing learning model is a learning model (such as "GPT") generated by machine learning using predetermined learning data. Any method, such as deep learning using a neural network, is used for the machine learning. Furthermore, a known method can be used for the additional learning (fine tuning). The additional learning unit 21 of the second system 20 corresponds to a fourth additional learning unit of the present invention.
[0042] FIG. 3 shows an example of training data (data sample) for additional learning (fine tuning) performed by the second system 20. When additional learning of an existing learning model is performed by the second system 20, for example, as shown in FIG. 3, "Low-calorie and satisfying" is used as the "catchphrase appeal information" of the input data. Furthermore, "88 points" is used as the "evaluation score" of the output data, and "the health aspects of the product are clearly stated" is used as the "reason (reason for the evaluation score)" of the output data. Such training data can be prepared in advance (for example, prepared in advance by an experienced copywriter).
[0043] The storage unit 22 of the second system 20 is configured with a large-capacity memory or the like, and stores data and programs necessary for the processing of the second system 20 (processing for evaluating catchphrase appeal information and outputting evaluation scores for the catchphrase appeal information). In addition, the storage unit 22 stores, for example, training data for additional learning (fine tuning) performed by the second system 20.
[0044] The input unit 23 of the second system 20 receives catch phrase appeal information selected (selected by user input) by the input unit 3 of the user device 2 from the catch phrase appeal information generated by the first system 10. The input unit 23 of the second system 20 corresponds to the fourth input unit of the present invention.
[0045] The estimation output unit 24 of the second system 20 receives, as input, the catchphrase appeal information (e.g., "We offer the ultimate relaxation") input from the input unit 23 of the second system 20, based on the relationship additionally learned by the additional learning unit 21 of the second system 20, and estimates and outputs the evaluation score of the catchphrase appeal information (e.g., "75 points") and the reason for the evaluation (e.g., "This catchphrase emphasizes the appeal of high-quality materials and design, and effectively conveys the importance of relaxation") (see FIG. 10). For example, in the case of a neural network that has undergone the above additional learning (fine tuning), estimation is performed by inputting the catchphrase appeal information to the input layer and outputting the evaluation score of the catchphrase appeal information and the reason for the evaluation from the output layer. The estimation output unit 24 of the second system 20 corresponds to the fourth estimation output unit of the present invention.
[0046] The third system 30 includes an additional learning unit 31, a memory unit 32, an input unit 33, and an estimation output unit 34. The additional learning unit 31 of the third system 30 has a function of additionally learning, by fine tuning, the relationship between catchphrase appeal information, catchphrases, and their reasons (the reason for leading from what the catchphrase appeal information wants to communicate (WHAT) to how to communicate it (HOW)) in an existing learning model. A catchphrase is information indicating how to communicate what the catchphrase appeal information wants to communicate. The existing learning model is a learning model (e.g., "GPT") generated by machine learning using predetermined learning data. Any method, such as deep learning using a neural network, is used for machine learning. Furthermore, known methods can be used for additional learning (fine tuning). The additional learning unit 31 of the third system 30 corresponds to the additional learning unit of the present invention.
[0047] FIG. 4 shows an example of training data (data sample) for additional learning (fine-tuning) performed by the third system 30. When additional learning of an existing learning model is performed by the third system 30, for example, as shown in FIG. 4, the "catchphrase appeal information (WHAT)" of the input data may be, for example, "Changing the future with innovative technology." The "catchphrase (HOW)" of the output data may be, for example, "Turning yesterday's common sense inside out," and the "reason (the reason for going from what the catchphrase appeal information wants to convey (WHAT) to how it is conveyed (HOW))" of the output data may be, for example, "It boldly expresses innovation and appeals to the power to overturn existing concepts." Such training data may be prepared in advance (for example, prepared in advance by an experienced copywriter).
[0048] The storage unit 32 of the third system 30 is configured with a large-capacity memory or the like, and stores data and programs necessary for the processing of the third system 30 (processing for generating a catchy slogan based on catchy slogan appeal information). The storage unit 32 also stores, for example, training data for additional learning (fine tuning) performed by the third system 30.
[0049] The input unit 33 of the third system 30 receives, from the catch phrase appeal information generated by the first system 10, catch phrase appeal information selected (selected by user input) by the input unit 3 of the user device 2 based on the evaluation points output by the second system 20. The input unit 33 of the third system 30 corresponds to the input unit of the present invention.
[0050] The estimation output unit 34 of the third system 30 receives, as input, the catchphrase appeal information (e.g., "Reset your daytime fatigue" or "Supports a good night's sleep") input from the input unit 33 of the third system 30 based on the additional learning performed by the additional learning unit 31 of the third system 30. The estimation output unit 34 estimates and outputs a catchphrase (e.g., "Feel refreshed and ready to wake up the next morning" or "To wake up refreshed") and its reason (e.g., "This catchphrase simply conveys that you can completely reset your daytime fatigue by the next morning" or "It explains the importance of good sleep while using expressions that evoke a relaxing, comfortable environment") based on the input of the catchphrase appeal information (e.g., "Reset your daytime fatigue" or "Supports a good night's sleep") from the input unit 33 of the third system 30 (see FIGS. 11(a) and 11(b)). For example, in the case of a neural network that has undergone the above additional learning (fine-tuning), estimation is performed by inputting the catchphrase target information into the input layer and outputting the catchphrase appeal information and its reason from the output layer. The estimation output unit 34 of the third system 30 corresponds to the estimation output unit of the present invention.
[0051] The fourth system 40 includes an additional learning unit 41, a memory unit 42, an input unit 43, and an estimation output unit 44. The additional learning unit 41 of the fourth system 40 has a function of additionally learning the relationship between a predetermined catchphrase, the evaluation score of that catchphrase, and the reason (the reason for that evaluation score) by fine tuning in an existing learning model. The existing learning model is a learning model (such as "GPT") generated by machine learning using predetermined learning data. Any method such as deep learning using a neural network is used for the machine learning. Furthermore, a known method can be used for the additional learning (fine tuning). The additional learning unit 41 of the fourth system 40 corresponds to the second additional learning unit of the present invention.
[0052] FIG. 5 shows an example of training data (data sample) for additional training (fine tuning) performed by the fourth system 40. When additional training of an existing learning model is performed by the fourth system 40, for example, as shown in FIG. 5, a "catchphrase" such as "It's up to you to build a new future" is used as the "catchphrase" of the input data. In addition, a "58 points" is used as the "evaluation score" of the output data, and a "reason (reason for the evaluation score)" of the output data is used such as "It has a bright image, but lacks specificity." Such training data can be prepared in advance (for example, prepared in advance by an experienced copywriter).
[0053] The storage unit 42 of the fourth system 40 is configured with a large-capacity memory or the like, and stores data and programs necessary for the processing of the fourth system 40 (processing for evaluating catchphrases and outputting evaluation scores for those catchphrases). The storage unit 42 also stores, for example, training data for additional learning (fine tuning) performed by the fourth system 40.
[0054] The input unit 43 of the fourth system 40 receives a catchphrase selected (selected by user input) by the input unit 3 of the user device 2 from among the catchphrases generated by the third system 30. The input unit 43 of the fourth system 40 corresponds to the second input unit of the present invention.
[0055] The estimation output unit 44 of the fourth system 40 receives a catchphrase (e.g., "Feel refreshed and ready the next morning") input from the input unit 43 of the fourth system 40 based on the relationship additionally learned by the additional learning unit 41 of the fourth system 40, and estimates and outputs an evaluation score (e.g., "85 points") of the catchphrase appeal information and the reason for the evaluation (e.g., "This catchphrase succinctly expresses that good quality sleep makes waking up the next day comfortable, and is effective in attracting the user's attention") (see FIG. 12). For example, in the case of a neural network that has undergone the above additional learning (fine tuning), estimation is performed by inputting the catchphrase to the input layer and outputting the evaluation score of the catchphrase and the reason for the evaluation from the output layer. The estimation output unit 44 of the fourth system 40 corresponds to the second estimation output unit of the present invention.
[0056] The operation of the advertising text generation system 1 configured as above will be described with reference to the sequence diagrams of FIGS.
[0057] As shown in Fig. 6, in the advertising text generation system 1 of this embodiment, first, in the first system 10, using, for example, the training data shown in Fig. 2, the relationship between the catch phrase target information, the catch phrase appeal information, and the reason (the reason for selecting that catch phrase appeal information) is additionally learned by fine tuning to the existing learning model (S1). Also, in the second system 20, using, for example, the training data shown in Fig. 3, the relationship between a predetermined catch phrase appeal information, the evaluation score of that catch phrase appeal information, and the reason (the reason for selecting that evaluation score) is additionally learned by fine tuning to the existing learning model (S2).
[0058] Furthermore, in the third system 30, using training data such as that shown in Fig. 4, the relationship between the catch phrase appeal information, the catch phrase, and the reason for it (the reason for going from what the catch phrase appeal information wants to convey (WHAT) to how to convey it (HOW)) is additionally learned by fine tuning to the existing learning model (S3). Also, in the fourth system 40, using training data such as that shown in Fig. 5, the relationship between a predetermined catch phrase, the evaluation score of that catch phrase, and the reason for it (the reason for setting that evaluation score) is additionally learned by fine tuning to the existing learning model (S4).
[0059] When generating a catchphrase using the advertising text generation system 1 of this embodiment, catchphrase target information is input via the input unit 3 of the user device 2 (S10), and the input catchphrase target information is transmitted from the user device 2 to the first system 10 (S11). As shown in Fig. 9, the first system 10 uses the catchphrase target information transmitted from the user device 2 (e.g., "Relaxing sleep aids") as input and estimates and outputs catchphrase appeal information (e.g., "Provides ultimate relaxation") and the reason for the appeal (e.g., "High-quality materials and ergonomic design provide maximum comfort") (S12). The estimated and output catchphrase appeal information and the reason for the appeal are transmitted from the first system 10 to the user device 2 (S13), and displayed on the display unit 4 of the user device 2 (S14).
[0060] Next, as shown in Fig. 7, when one or more pieces of catchphrase appeal information are selected by the input unit 3 of the user device 2 from one or more pieces of catchphrase appeal information displayed on the display unit 4 of the user device 2 (S15), the selected catchphrase appeal information is transmitted from the user device 2 to the second system 20 (S16). The second system 20, for example, as shown in Fig. 10, inputs the catchphrase appeal information transmitted from the user device 2 (e.g., "We offer the ultimate relaxation") and estimates and outputs an evaluation score (e.g., "75 points") and the reason for the evaluation (e.g., "This catchphrase emphasizes the appeal of high-quality materials and design, and effectively communicates the importance of relaxation") for the catchphrase appeal information (S17). The estimated and output evaluation score and the reason for the catchphrase appeal information are transmitted from the second system 20 to the user device 2 (S18) and displayed on the display unit 4 of the user device 2 (S19).
[0061] Next, one or more pieces of catchphrase appeal information are selected by the input unit 3 of the user device 2 from among one or more pieces of catchphrase appeal information displayed on the display unit 4 of the user device 2 based on the respective evaluation points (S20). The selected catchphrase appeal information is transmitted from the user device 2 to the third system 30 (S21). The third system 30, for example, as shown in Figures 11(a) and 11(b), uses the catchphrase appeal information transmitted from the user device 2 (e.g., "Reset your daytime fatigue" or "Supports a good night's sleep") as input, and estimates and outputs a catchphrase (e.g., "Feel refreshed and ready to go the next morning" or "To start the day feeling refreshed") and the reason for the catchphrase (e.g., "This catchphrase is intended to simply convey that you can completely reset your daytime fatigue by the next morning" or "It explains the importance of a good night's sleep while using expressions that evoke a relaxing and comfortable environment") (S22). The estimated and output catchphrase and the reason for it are transmitted from the third system 30 to the user device 2 (S23) and displayed on the display unit 4 of the user device 2 (S24).
[0062] 8, when one or more catch phrases are selected by the input unit 3 of the user device 2 from one or more catch phrases displayed on the display unit 4 of the user device 2 (S25), the selected catch phrases are transmitted from the user device 2 to the fourth system 40 (S26). The fourth system 40, for example, as shown in FIG. 12, inputs the catch phrase transmitted from the user device 2 (e.g., "Wake up refreshed and ready the next morning") and estimates and outputs an evaluation score (e.g., "85 points") for the catch phrase appeal information and the reason for the evaluation score (e.g., "This catch phrase succinctly expresses that good quality sleep makes waking up comfortable the next day, and is effective in attracting the user's attention") (S27). The estimated and output evaluation score for the catch phrase and the reason for the evaluation score are transmitted from the fourth system 40 to the user device 2 (S28) and displayed on the display unit 4 of the user device 2 (S29).
[0063] According to the advertising text generation system 1 of this embodiment, the first system 10 generates catchphrase appeal information (e.g., "Your health depends on whether you stay hydrated") indicating what the advertiser wants to communicate about the catchphrase target (e.g., "soft drinks") from catchphrase target information (e.g., "Make dehydration a familiar symptom"), and the second system 20 outputs an evaluation score (e.g., "85 points") for the catchphrase appeal information. The third system 30 generates a catchphrase (e.g., "Drink to live") indicating how to communicate what the advertiser wants to communicate with the catchphrase appeal information from the catchphrase appeal information, and the fourth system 40 outputs an evaluation score (e.g., "90 points") for the catchphrase.
[0064] In the advertising text generation system 1 of this embodiment, the first system 10 uses an existing learning model that has been additionally trained through fine tuning to learn the relationship between the catch phrase target information and the catch phrase appeal information and reasons (reasons for choosing that catch phrase appeal information), and when catch phrase target information is input, the catch phrase appeal information and reasons (reasons for choosing that catch phrase appeal information) are estimated and output.
[0065] If catch phrase appeal information is generated from catch phrase target information simply using an existing learning model (such as "GPT"), the quality of the generated catch phrase appeal information will be low. However, if catch phrase appeal information and the reason are generated from catch phrase target information using an existing learning model that has been additionally trained, as described above, to learn not only the relationship between the catch phrase target information and the catch phrase appeal information but also the relationship with the reason (the reason for selecting that catch phrase appeal information), the quality of the generated catch phrase appeal information will be high. In this way, the quality of catch phrase appeal information generated using an existing learning model can be improved.
[0066] Furthermore, in the advertising text generation system 1 of this embodiment, the second system 20 uses an existing learning model in which the relationship between the catch phrase appeal information and the evaluation score and reason (reason for the evaluation score) of the catch phrase appeal information is additionally learned by fine tuning, and when catch phrase appeal information (the catch phrase appeal information selected by the user from the catch phrase appeal information generated by the first system 10) is input, the evaluation score and reason (reason for the evaluation score) of the catch phrase appeal information is estimated and output.
[0067] If an evaluation score for catch phrase appeal information is generated from catch phrase appeal information simply using an existing learning model (such as "GPT"), the accuracy of the evaluation score for the generated catch phrase appeal information will be low. However, if an evaluation score and reason for catch phrase appeal information are generated from catch phrase appeal information using an existing learning model that has additionally learned not only the relationship between the catch phrase appeal information and the evaluation score for the catch phrase appeal information but also the relationship with the reason (the reason for the evaluation score) as described above, the accuracy of the evaluation score for the generated catch phrase appeal information will be higher. In this way, the accuracy of the evaluation score for catch phrase appeal information generated using an existing learning model can be improved.
[0068] Furthermore, in the advertising text generation system 1 of this embodiment, the third system 30 uses an existing learning model in which the relationship between the catch phrase appeal information and the catch phrase and the reason (the reason for going from what you want to communicate (WHAT) to how to communicate it (HOW)) has been additionally learned by fine tuning, and when catch phrase appeal information (the catch phrase appeal information generated by the first system 10 that has been selected based on the evaluation points output by the second system 20) is input, the catch phrase and the reason (the reason for going from what you want to communicate (WHAT) to how to communicate it (HOW)) are estimated and output.
[0069] If catchphrases are generated from catchphrase appeal information simply using an existing learning model (such as "GPT"), the quality of the generated catchphrases will be low (for example, if 15 catchphrases are generated, the average evaluation score will be 61.6 points). However, if catchphrases and reasons are generated from catchphrase appeal information using an existing learning model that has been additionally trained, as described above, to not only the relationship between catchphrase appeal information and catchphrases, but also the relationship between the reasons (the reasons leading from what you want to communicate (WHAT) to how to communicate it (HOW)), the quality of the generated catchphrases will be high (for example, if 15 catchphrases are generated, the average evaluation score will be 73.3 points). In this way, it is possible to improve the quality of catchphrases generated using existing learning models.
[0070] Furthermore, in the advertising text generation system 1 of this embodiment, the fourth system 40 uses an existing learning model in which the relationship between the catch phrase and the evaluation score and reason (reason for that evaluation score) of the catch phrase has been additionally learned by fine tuning, and when a catch phrase (a catch phrase selected by the user from among the catch phrases generated by the third system 30) is input, the evaluation score and reason (reason for that evaluation score) of the catch phrase are estimated and output.
[0071] If an existing learning model (such as "GPT") is simply used to generate a catchphrase evaluation score from a catchphrase, the accuracy of the generated catchphrase evaluation score will be low. However, if an existing learning model that has been additionally trained as described above to not only learn the relationship between the catchphrase and the catchphrase evaluation score but also the relationship with the reason (the reason for the evaluation score) is used to generate a catchphrase evaluation score and reason from the catchphrase, the accuracy of the generated catchphrase evaluation score will be higher. In this way, the accuracy of the evaluation score of a catchphrase generated using an existing learning model can be improved.
[0072] Although the embodiments of the present invention have been described above by way of example, the scope of the present invention is not limited to these, and can be modified and changed according to the purpose within the scope of the claims.
[0073] In the above embodiment, an example was described in which the advertising text generation system 1 sends "catchphrase appeal information and the reasons therefor," "evaluation score of the catchphrase appeal information and the reasons therefor," "catchphrase and the reasons therefor," and "evaluation score of the catchphrase and the reasons therefor" to the user device 2, and these are displayed on the user device 2, but the scope of the present invention is not limited to these.
[0074] For example, as shown in FIG. 13 , the advertising text generation system 1 may include a transmission control unit 50 that controls information transmitted from the catch phrase generation system 1 to the user device 2. The transmission control unit 50 has a function of controlling the exclusion of the "reason" from the information transmitted from the advertising text generation system 1 to the user device 2. In this case, the advertising text generation system 1 estimates and outputs "catch phrase appeal information and its reason," "evaluation score of the catch phrase appeal information and its reason," "catch phrase and its reason," and "evaluation score of the catch phrase and its reason," but the advertising text generation system 1 transmits to the user device 2 only "catch phrase appeal information," "evaluation score of the catch phrase appeal information," "catch phrase," and "evaluation score of the catch phrase," which are displayed on the user device 2. In other words, the "reason" is not transmitted from the advertising text generation system 1 to the user device 2, and the "reason" is not displayed on the user device 2.
[0075] Furthermore, in the above embodiment, an example has been described in which the advertising text is a catchphrase, but the scope of the present invention is not limited to this. Advertising text can also include text other than a catchphrase, as long as it is text used in advertising (including PR). For example, advertising text can include body copy, shoulder copy, naming, narration and storyboards for commercials (radio commercials, television commercials, etc.), company slogans, company statements, commercial song lyrics, web banner copy, in-store POP copy, website and landing page text, listing ads, poster ad copy, newspaper and magazine ad copy, social media ads (banners, timeline ads, etc.), email newsletter and direct mail (DM) copy, press releases, and OOH (Out Of Home) ad copy.
[0076] For example, if the advertising text is body copy, shoulder copy, narration or storyboard for a commercial (radio commercial, TV commercial, etc.), lyrics for a commercial song, copy for a web banner, copy for a point-of-purchase advertisement, wording for a website or landing page, or listing advertisement, the product name or service name, etc., is input as the advertising text target information. If the advertising text is a name, the concept of the product or service, etc., is input as the advertising text target information. If the advertising text is a company slogan, the company's philosophy, vision, goals and values, etc., are input as the advertising text target information. If the advertising text is a company statement, background information such as the company's history, growth strategy, and social role is input as the advertising text target information. [Industrial Applicability]
[0077] As described above, the advertising text generation system of the present invention has the effect of being able to improve the quality of advertising text generated using existing learning models, and is useful as a business support system used in cloud services that support advertising text creation work. [Explanation of symbols]
[0078] 1. Advertising text generation system 2. User Device 3 Input section 4 Display section 10. System 1 11 Additional Learning Section (Third Additional Learning Section) 12 Storage section 13 Input section (third input section) 14 Estimation output unit (third estimation output unit) 20 Second System 21 Additional Learning Section (4th Additional Learning Section) 22 Memory section 23 Input section (fourth input section) 24 Estimation output unit (fourth estimation output unit) 30 Third System 31 Additional Learning Section 32 Storage section 33 Input section 34 Estimation output section 40 Fourth System 41 Additional Learning Section (Second Additional Learning Section) 42 Storage section 43 Input section (second input section) 44 Estimation output unit (second estimation output unit) 50 Transmission control section N Network
Claims
1. a first system that generates advertising text appeal information indicating what is to be communicated about an advertising text target based on advertising text target information related to the advertising text target for which advertising text is to be generated; a second system that evaluates the advertising text appeal information generated by the first system and outputs an evaluation score of the advertising text appeal information; a third system that generates advertising text based on the advertising text appeal information, the advertising text indicating how to convey what is intended to be conveyed by the advertising text appeal information; a fourth system that evaluates the advertising text generated by the third system and outputs an evaluation score for the advertising text; Equipped with The third system is an additional learning unit that additionally learns, by fine tuning, the relationship between predetermined advertising text appeal information, advertising text indicating a way of conveying what is intended to be conveyed by the advertising text appeal information, and the reason why what is intended to be conveyed by the advertising text appeal information leads to the way of conveying the information, in an existing learning model generated by machine learning using predetermined learning data; an input unit to which advertising text appeal information selected from the advertising text appeal information generated by the first system based on the evaluation points output by the second system is input; an estimation output unit that uses the advertising text appeal information input from the input unit as an input based on the relationship additionally learned by the additional learning unit, estimates and outputs advertising text indicating a way of communicating what is intended to be communicated in the advertising text appeal information, and a reason why what is intended to be communicated in the advertising text appeal information leads to the way of communicating; An advertising text generation system comprising:
2. The fourth system is a second additional learning unit that additionally learns, by fine tuning, a relationship between a predetermined advertising text, an evaluation score for the advertising text, and a reason for the evaluation score for the advertising text in an existing learning model generated by machine learning using predetermined learning data; a second input unit to which advertising text selected by a user from the advertising texts generated by the third system is input; a second estimation output unit that estimates and outputs an evaluation score for the advertising text and a reason for the evaluation score, based on the relationship additionally learned by the second additional learning unit, using the advertising text input from the second input unit as an input; The advertising text generation system of claim 1 , comprising:
3. The first system is a third additional learning unit that additionally learns, by fine tuning, the relationship between advertising text target information related to a predetermined advertising text target, advertising text appeal information indicating what is to be communicated about the advertising text target, and the reason for selecting the advertising text appeal information, in an existing learning model generated by machine learning using predetermined learning data; a third input unit to which advertising text target information relating to an advertising text target for which advertising text is to be generated is input; a third estimation output unit that receives advertising text target information related to the advertising text target input from the third input unit based on the relationship additionally learned by the third additional learning unit, estimates and outputs advertising text appeal information indicating what is to be communicated about the advertising text target and a reason for selecting the advertising text appeal information; and The advertising text generation system of claim 1 , comprising:
4. The second system is a fourth additional learning unit that additionally learns, by fine tuning, the relationship between predetermined advertising text appeal information, the evaluation score of the advertising text appeal information, and the reason for the evaluation score in an existing learning model generated by machine learning using predetermined learning data; a fourth input unit to which advertising text appeal information selected by a user from the advertising text appeal information generated by the first system is input; a fourth estimation output unit that estimates and outputs an evaluation score of the advertising text appeal information and a reason for the evaluation score based on the relationship additionally learned by the fourth additional learning unit, using the advertising text appeal information input from the fourth input unit as an input; The advertising text generation system according to claim 3 , comprising:
5. an additional learning unit that additionally learns, by fine tuning, the relationship between predetermined advertising text appeal information indicating what is intended to be communicated about the subject of the advertising text, advertising text indicating how to communicate what is intended to be communicated in the advertising text appeal information, and the reason why what is intended to be communicated in the advertising text appeal information leads to the way of communication, in an existing learning model generated by machine learning using predetermined learning data; an input unit to which advertising text appeal information generated based on advertising text target information related to an advertising text target for which advertising text is to be generated is input; an estimation output unit that uses the advertising text appeal information input from the input unit as an input based on the relationship additionally learned by the additional learning unit, estimates and outputs advertising text indicating a way of communicating what is intended to be communicated in the advertising text appeal information, and a reason why what is intended to be communicated in the advertising text appeal information leads to the way of communicating; An advertising text generation system comprising:
6. a second additional learning unit that additionally learns, by fine tuning, a relationship between a predetermined advertising text, an evaluation score for the advertising text, and a reason for the evaluation score for the advertising text in an existing learning model generated by machine learning using predetermined learning data; a second input unit to which advertising text selected by a user from the advertising texts output from the estimation output unit is input; a second estimation output unit that estimates and outputs an evaluation score for the advertising text and a reason for the evaluation score, based on the relationship additionally learned by the second additional learning unit, using the advertising text input from the second input unit as an input; The advertising text generation system of claim 5 , comprising:
7. A program executed in an advertising text generation system, The program is configured to: A process of additionally learning, by fine tuning, the relationship between predetermined advertising text appeal information indicating what is intended to be communicated about the subject of the advertising text, advertising text indicating how to communicate what is intended to be communicated in the advertising text appeal information, and the reason why what is intended to be communicated in the advertising text appeal information leads to the said way of communication, in an existing learning model generated by machine learning using predetermined learning data; A process of inputting advertising text appeal information generated based on advertising text target information related to an advertising text target for which advertising text is to be generated; A process of using the advertising text appeal information input from the input unit as an input based on the relationship additionally learned by the additional learning unit, estimating and outputting advertising text indicating a way of communicating what is intended to be communicated in the advertising text appeal information, and a reason why the way of communicating is reached from what is intended to be communicated in the advertising text appeal information; A program that executes.
8. 1. A method performed in an advertising text generation system, comprising: The method comprises: An existing learning model generated by machine learning using predetermined learning data is additionally trained by fine tuning the relationship between predetermined advertising text appeal information indicating what is intended to be communicated about the subject of the advertising text, advertising text indicating how to communicate what is intended to be communicated in the advertising text appeal information, and the reason why what is intended to be communicated in the advertising text appeal information leads to the said way of communication; Advertising text appeal information generated based on advertising text target information related to an advertising text target for which advertising text is to be generated is input; Based on the relationship additionally learned by the additional learning unit, the advertising text appeal information input from the input unit is used as an input, and an advertising text indicating a way of communicating what is intended to be communicated in the advertising text appeal information and a reason why the way of communicating is derived from what is intended to be communicated in the advertising text appeal information are inferred and output; A method comprising:
Citation Information
Patent Citations
server
JP7316598B1