Article creation support device

Abstract

Provided is a sentence creation support device that creates more appropriate sentences for each of the action target individuals selected from among the prediction targets based on the characteristics of the action target individuals. 【Solution means】The sentence creation support device 1 includes a model creation module 11 that creates a model showing the relationship between the first explanatory variable and the objective variable based on the first explanatory variable and the objective variable of each of a plurality of learning targets, a scoring module 12 that inputs the second explanatory variable of each of the plurality of prediction targets into the model and calculates the score of each of the plurality of prediction targets, a selection module 13 that selects a plurality of action targets from among the plurality of prediction targets based on the score, a first prompt creation module 14 that creates a first prompt for instructing the generation AI 3 to estimate the characteristics of each of the plurality of action targets by referring to the data related to each of the plurality of action targets, and a second prompt creation module 15 that creates a second prompt for instructing the generation AI to generate a sentence considering the characteristics.

Description

Technical Field

[0001] The present invention relates to a text creation support device that supports text creation using generative AI in order to more effectively execute actions based on scores calculated using the results of machine learning.

Background Art

[0002] A system has been put into practical use and improved in which machine learning is performed to create a model from a combination of a first explanatory variable indicating an attribute or feature of a learning target and an objective variable that is a known result regarding the learning target, and a result regarding the prediction target is predicted by inputting a second explanatory variable indicating an attribute or feature of the prediction target into this model.

Prior Art Documents

Patent Documents

[0003]

Patent Document 1

Patent Document 2

Summary of the Invention

Problems to be Solved by the Invention

[0004] As an example, in the field of marketing, it is possible to predict the purchases of another customer who is the prediction target using the attributes and purchase history data of a plurality of customers who are the learning targets. By taking actions such as sending a guidance text to customers with a high purchase probability among the prediction targets, high effects can be obtained at a lower cost.

[0005] However, even if the prediction accuracy has been improved, the conventional text of the guidance text has been uniform, or there has been only a method of creating the text for each target group segmented to a certain extent.

[0006] One aspect of the present invention relates to creating more appropriate contexts individually based on the individual characteristics of action targets selected from among targets to be predicted using the results of machine learning, and enabling effective execution of the actions.

Means for Solving the Problems

[0007] A text creation support device according to one aspect of the present invention includes a model creation module that creates a model showing the relationship between the first explanatory variable and the objective variable based on the first explanatory variable and the objective variable of each of a plurality of learning targets; a scoring module that inputs the second explanatory variable of each of a plurality of prediction targets into the model and calculates the score of each of the plurality of prediction targets; a selection module that selects a plurality of action targets from among the plurality of prediction targets based on the score; a first prompt creation module that creates a first prompt for instructing a first generative AI to estimate the characteristics of each of the plurality of action targets by referring to the data related to each of the plurality of action targets; a second prompt creation module that creates a second prompt for instructing a second generative AI to generate a text considering the characteristics; and includes.

Brief Description of the Drawings

[0008]

Figure 1

Figure 2

Figure 3A

Figure 3B

Figure 3C

Figure 3D

Figure 3E

Figure 3F

Figure 4A

Figure 4B

Figure 4C

Figure 4D

Figure 4E

Figure 4F

Figure 4G

Figure 4H

Figure 5A

Figure 5B

Figure 5C

Figure 5D

Figure 5E

Figure 5F

Figure 5G

Figure 6A

Figure 6B

Figure 6C

Figure 6D

Figure 6E

Figure 6F

Figure 6G

Figure 7

Figure 8A

Figure 8B

Figure 8C

Best Mode for Carrying Out the Invention

[0009] Hereinafter, embodiments of the present invention will be described in detail with reference to the drawings. Each embodiment described below shows an example of the present invention and does not limit the content of the present invention. Also, not all of the configurations and operations described in each embodiment are essential as the configurations and operations of the present invention. Note that the same reference numerals are assigned to the same components, and redundant explanations are omitted.

[0010] <1. First Embodiment> <1-1. Configuration> FIG. 1 shows a text creation support device 1 according to the first embodiment and an external device connected to the text creation support device 1. The text creation support device 1 is a computer system including a CPU, a memory, etc. (not shown). The text creation support device 1 may be composed of a single computer or may be composed of a plurality of computers connected by a network. The text creation support device 1 is connected to external devices such as a database 2 and a generation AI 3.

[0011] The database 2 stores data for learning and prediction targets and other data. The database 2 is not limited to one in which data is stored in a single storage, and may be one in which data is distributed and stored in a plurality of storages. The text creation support device 1 acquires various data from the database 2 and performs model creation, scoring, selection of an action target, and creation of a prompt.

[0012] The generative AI 3 includes a large language model (LLM). The large language model is a language model constructed using a large amount of text data and deep learning techniques, and processes tasks such as text generation, translation, answering questions, text summarization, sentiment analysis, etc. according to the prompt sent from the text creation support device 1. It is desirable for the large language model to have an attention mechanism. The attention mechanism is a mechanism for extracting important parts from the input, and greatly contributes to improving the processing speed and accuracy of the large language model.

[0013] The generative AI 3 may further include image generation AI. The image generation AI is a system that generates an image from a prompt described in text. For example, a diffusion model is used. Note that the "image" in the following description may also be a video.

[0014] The text creation support device 1 sends a prompt to the generative AI 3 and obtains the output from the generative AI 3.

[0015] <1-2. Functions and Operations> Figure 2 shows the functions and operations of the text creation support device 1 according to the first embodiment. The text creation support device 1 includes a model creation module 11, a scoring module 12, a selection module 13, a first prompt creation module 14, and a second prompt creation module 15. These modules are realized by loading a program into the memory included in the text creation support device 1 and being executed by the CPU.

[0016] The model creation module 11 obtains the first explanatory variable and the objective variable of each of a plurality of learning targets from the database 2, and creates a model showing the relationship between the first explanatory variable and the objective variable. A specific example of model creation will be described later with reference to FIGS. 3A and 3B.

[0017] The scoring module 12 acquires second explanatory variables for each of the multiple prediction targets from the database 2, inputs the second explanatory variables into the model, and calculates a score for each of the multiple prediction targets. Specific examples of scoring will be described later with reference to Figures 3C to 3E.

[0018] The selection module 13 selects a plurality of behavior targets from the plurality of predicted targets based on the scores. A specific example of behavior target selection will be described with reference to FIG. 3F.

[0019] The first prompt creation module 14 acquires data on each of the plurality of behavior targets from the database 2 and creates a first prompt that instructs the generation AI 3 to estimate the characteristics of each of the plurality of behavior targets. A specific example of creating the first prompt will be described later with reference to Figures 4A to 4F.

[0020] The second prompt generation module 15 acquires the characteristics of each of the multiple behavioral targets estimated by the generation AI 3 and generates a second prompt that instructs the generation AI 3 to generate a sentence taking these characteristics into consideration. A specific example of generating a second prompt will be described later with reference to Figures 4G and 4H.

[0021] In this application, the generation AI3 that estimates the characteristics of the target of action in response to the first prompt may be referred to as the first generation AI, and the generation AI3 that generates a sentence in response to the second prompt may be referred to as the second generation AI. The first and second generation AIs may be the same generation AI or different generation AIs.

[0022] <1-3. First concrete example> <1-3-1. Model Creation> 3A shows an example of training data used for model creation in the first specific example of the first embodiment. The training data includes various information about each of a large number of customers identified by their customer IDs, such as gender, address, membership type, and whether or not they have made a purchase.

[0023] FIG. 3B shows an example in which the learning target data in FIG. 3A is digitized for model creation. For model creation, for example, gender is converted to 0 or 1, address is converted to a value from 1 to 47 according to the prefecture, membership type is converted to a value from 1 to 4 according to the rank, and these are used as the first explanatory variables. Also, for example, the presence or absence of purchase is converted to 0 or 1, and this is used as the target variable. The model creation module 11 creates a function showing the relationship between a plurality of the first explanatory variables and the target variable as a model.

[0024] <1-3-2. Scoring> FIG. 3C shows an example of prediction target data used for scoring in the first specific example of the first embodiment. The prediction targets are composed of customers different from the customers constituting the learning targets. The prediction target data includes information such as gender, address, and membership type for each of a number of customers identified by customer ID. However, the prediction target data does not include data on the presence or absence of purchase. The presence or absence of purchase of the prediction target is unknown information. For items other than the presence or absence of purchase, the items included in the prediction target data are common with the items included in the learning target data.

[0025] FIG. 3D shows an example in which the prediction target data in FIG. 3C is digitized for scoring. Similar to the learning target data, gender, address, membership type, etc. are each converted to a numerical value, and these are used as the second explanatory variables.

[0026] FIG. 3E shows an example of scores calculated in the first specific example of the first embodiment. By inputting the second explanatory variables of each of a plurality of prediction targets into the model, scores for each of the plurality of prediction targets are calculated. This score is a value corresponding to the target variable in the learning target data, for example, a value indicating the probability that each of the prediction targets purchases a product.

[0027] <1-3-3. Selection of Action Targets> FIG. 3F shows the theoretical effect obtained by selecting the action target in the first specific example of the first embodiment. Assume that the prediction target is, for example, 460,000 customers, and if a guidance text is sent to all 460,000 customers by direct mail, 60,000 customers will purchase the product. Since it is extremely costly to send the guidance text to all 460,000 customers, for example, when sending the guidance text to 230,000, which is half of them, if 230,000 are randomly selected (when not analyzed), it is considered that 30,000 will purchase the product. That is, the number of purchasers is proportional to the number of guidance text sent. On the other hand, if 230,000 are selected in descending order of score (when using the analysis result), the number of customers who purchase the product is expected to increase, for example, to 51,000. The ratio of the number of action targets to the number of prediction targets is not limited to half, and even if it is other ratios (excluding 0 and 1), higher results can be obtained when using the analysis result than when not analyzed.

[0028] <1-3-4. First Prompt Creation> FIG. 4A shows an example of action target data used for creating the first prompt in the first specific example of the first embodiment. Only the data of customer ID "bbbb", which is one of the multiple action targets selected based on the score from among the prediction targets, is shown in FIG. 4A.

[0029] The action target data may include items included in the prediction target data. Also, the action target data may include items not included in the prediction target data. For example, as the purchase data of customer ID "bbbb", it includes the date, amount, and product type of the product with donation purchased by customer ID "bbbb".

[0030] Figure 4B shows an example in which external data is added to the action target data shown in Figure 4A. The external data shown in Figure 4B is event data including events for each month corresponding to the date included in the action target data shown in Figure 4A, the business sentiment index CI and DI, and the Nikkei average stock price N225. The business sentiment index CI is an index indicating the direction of business expansion or contraction. The business sentiment index DI is an index representing the troughs and peaks of the economy with 50 as the boundary. The event data is extracted from the data stored in database 2 separately from the learning target data and the prediction target data according to the date included in the action target data.

[0031] Figure 4C shows an example of the first prompt generated in the first specific example of the first embodiment. The first prompt includes an instruction sentence, auxiliary information, purchase data, and event data. The instruction sentence includes a sentence that requests the generation AI 3 to generate an inference about the characteristics such as the age and personality of the action target. The instruction sentence may be stored in advance as a template. The purchase data and event data are the purchase data and event data shown in Figure 4B.

[0032] The auxiliary information may include information stored in advance as a template. The auxiliary information may also include information derived from the action target data. For example, if the correlation between the purchase data of customer ID "bbbb" and the trend of stock prices by industry is obtained, and there is a high positive correlation between the deposit amount of customer ID "bbbb" and the stock price of the construction industry, the presumption that the workplace of customer ID "bbbb" is in the construction industry may be included in the auxiliary information. Also, if the correlation between the purchase data of customer ID "bbbb" and the trend of the market value of various assets is obtained, and there is a high positive correlation between the deposit amount of customer ID "bbbb" and the evaluation amount of real estate investment trusts, the presumption that the assets held by customer ID "bbbb" are mainly real estate may be included in the auxiliary information.

[0033] Figure 4D shows an example of the response from the generation AI 3 to the first prompt shown in Figure 4C. The generation AI 3 outputs the result of inferring the characteristics such as the age and personality of the action target. By making an inference taking into account the external data, the inference accuracy can be improved.

[0034] Figure 4E shows a further example of the first prompt. The first prompt includes an instruction sentence that requests the generative AI 3 to generate speculations on the advantages and disadvantages as characteristics of the target of action.

[0035] Figure 4F shows an example of a response from the generative AI 3 to the first prompt shown in Figure 4E. The generative AI 3 outputs the result of speculating on the characteristics including the advantages and disadvantages of the target of action.

[0036] <1-3-5. Creation of the second prompt> Figure 4G shows an example of the second prompt generated in the first specific example of the first embodiment. The second prompt includes an instruction sentence that requests to write a thank-you letter for the purchase of the product so that the purchase of the product can be continued in the future, taking into account the speculated advantages and disadvantages. The instruction sentence also requests not to mention the workplace and assets, which are uncertain information. The instruction sentence may be stored in advance as a template. When the first prompt and the second prompt are sent to the same generative AI 3, it is assumed that the speculation result of the characteristics generated in response to the first prompt is carried over inside the generative AI 3. When the first prompt and the second prompt are sent to different generative AI 3s, it is necessary to include the speculation result of the characteristics generated in response to the first prompt in the second prompt.

[0037] Figure 4H shows an example of a response from the generative AI 3 to the second prompt shown in Figure 4G. The generative AI 3 outputs the text of the letter to the target of action. By generating the text taking into account the speculated advantages and disadvantages, the customer who receives the letter may feel that "they understand me", enhancing the goodwill towards the sender of the letter and increasing the likelihood of leading to the purchase of the product.

[0038] The document creation support device 1 may further request the generation AI 3 to generate an image to be attached to the letter. As an example, when the image generation AI constituting the generation AI 3 operates in response to an English prompt, the document creation support device 1 outputs a prompt such as "Please describe an image that would be effective to attach to a letter in English" to the large language model constituting the generation AI 3. Then, the image generation AI is made to generate an image based on the response from the large language model. The document creation support device 1 generates a prompt for the large language model so as to generate an html document of the letter linked to the image file from the file name of the image and the text of the letter in FIG. 4H, whereby the html document linked to the image file can be obtained from the large language model.

[0039] <1-3-6. Additional Prompt Examples> A prompt including the following command sentences may be further added. The command sentences may be stored in advance as templates.

[0040] (1) Before the instruction to estimate features, an instruction to list what viewpoints there are is provided. For example, before the first prompt shown in FIG. 4C, a prompt including a command sentence such as "List the important viewpoints to consider when estimating a persona in marketing." is given to the generation AI 3. Then, the command sentence included in the first prompt shown in FIG. 4C is changed to something like "Based on the 'listed viewpoints', infer Mr. A's age, personality, family composition, and lifestyle." and given to the generation AI 3. By making the estimation according to the listed viewpoints, it is possible to reduce omissions in consideration and enhance the sophistication of thinking.

[0041] (2) Make the prompt structure such that the examination of features is repeated several times to obtain the best answer. For example, after the first prompt shown in FIG. 4C, give the generation AI3 a prompt including an instruction sentence such as "Based on the listed viewpoints, create two more personas for Mr. A." Then, give the generation AI3 a prompt including an instruction sentence such as "Compare and consider or improve the above personas and present the most effective persona." In this way, another aspect of the action target can be found, and the accuracy can be further improved.

[0042] (3) Add a procedure to refine the generated text in light of the original data. For example, after the second prompt shown in FIG. 4G, give the generation AI3 a prompt including an instruction sentence such as "Refer to the auxiliary information, the member's purchase data, and the event data of the current month, verify if there are any inconveniences or points worthy of note, and refine this letter." By going back to the original data that was the basis of the speculation for consideration, the accuracy can be further improved.

[0043] <1-4. Second specific example> In the second specific example, different examples of prompt creation are shown. The model creation, scoring, and selection of the action target are the same as in the first specific example.

[0044] <1-4-1. First prompt creation> FIG. 5A shows an example of action target data used for the first prompt creation in the second specific example of the first embodiment. Only the data of customer ID "cccc", who is one of the multiple action targets selected based on the score from among the prediction targets, is shown in FIG. 5A. The action target data includes, for example, the attribute information and purchase history of customer ID "cccc" of a company engaged in the manufacture and sale of confectionery and nutritional supplements. In the second specific example, it is not necessary to import external data as shown in FIG. 4B.

[0045] Figure 5B shows an example of the first prompt generated in the second specific example of the first embodiment. The first prompt includes an instruction sentence, attribute information, and a purchase history. The instruction sentence includes a sentence that requests the generation AI 3 to generate an inference about characteristics such as the persona and personality of the action target. The attribute information and the purchase history are the attribute information and the purchase history shown in Figure 5A.

[0046] Figure 5C shows an example of the response from the generation AI 3 to the first prompt shown in Figure 5B. The generation AI 3 outputs the result of inferring characteristics such as the persona and personality of the action target.

[0047] Figure 5D shows a further example of the first prompt. The first prompt includes an instruction sentence that requests the generation AI 3 to generate an inference about the motivation to purchase and the preference for products as characteristics of the action target.

[0048] Figure 5E shows an example of the response from the generation AI 3 to the first prompt shown in Figure 5D. The generation AI 3 outputs the result of inferring characteristics including the motivation to purchase and the preference for products of the action target.

[0049] <1-4-2. Creation of the second prompt> Figure 5F shows an example of the second prompt generated in the second specific example of the first embodiment. The second prompt includes an instruction sentence that requests writing a thank-you letter for product purchase so as to increase the purchase desire of the action target based on the inferred motivation and preference. Also, the instruction sentence requests not to mention the attribute information which is personal information.

[0050] Figure 5G shows an example of the response from the generation AI 3 to the second prompt shown in Figure 5F. The generation AI 3 outputs the text of the letter to the action target. By generating the text based on the inferred motivation and preference, the customer who receives the letter is likely to recall feelings such as "Come to think of it, that was delicious" when purchasing in the past, increasing the likelihood of a repeat purchase.

[0051] The document creation support device 1 may further request the generation AI 3 to generate an image to be attached to the letter.

[0052] <1-5. Third specific example> In the third specific example, different examples of prompt creation are shown. The model creation, scoring, and selection of the action target are the same as in the first specific example.

[0053] <1-5-1. First prompt creation> FIG. 6A shows an example of action target data used for the first prompt creation in the third specific example of the first embodiment. Only the data of customer ID "dddd", who is one of the multiple action targets selected based on the score from among the prediction targets, is shown in FIG. 6A. The action target data shall include, for example, the shopping records of customer ID "dddd" in a grocery store, as well as the data of the household composition. In the third specific example, it is not necessary to import external data as shown in FIG. 4B.

[0054] FIG. 6B shows an example of the first prompt generated in the third specific example of the first embodiment. The first prompt includes an instruction sentence, a shopping record, and a specification of the output format. The instruction sentence includes a sentence that requests the generation AI 3 to generate an inference of the dishes for each date in the household of the action target. The shopping record is the shopping record shown in FIG. 6A.

[0055] FIG. 6C shows an example of the response from the generation AI 3 to the first prompt shown in FIG. 6B. The generation AI 3 outputs the inferred dishes and a list of the ingredients used extracted from the shopping record.

[0056] FIG. 6D shows a further example of the first prompt. The first prompt includes an instruction sentence that requests the generation AI 3 to propose an ingredient to be added to the inferred dishes.

[0057] FIG. 6E shows an example of a response from the generative AI 3 to the first prompt shown in FIG. 6D. The generative AI 3 outputs a presumed dish, ingredients extracted from the shopping record, and proposed ingredients.

[0058] <1-5-2. Creation of the second prompt> FIG. 6F shows an example of the second prompt generated in the third specific example of the first embodiment. The second prompt includes an instruction sentence that, based on the proposed ingredients, requests writing a recommended menu for the next shopping of the action target and a guidance text appealing the ingredients.

[0059] FIG. 6G shows an example of a response from the generative AI 3 to the second prompt shown in FIG. 6F. The generative AI 3 outputs the text of the guidance text for the action target. By generating a text based on the presumed dish and the proposed ingredients, the customer who receives the letter may think "I can make such a dish with the ingredients that I and my family like" or "It seems more delicious if I add one more item", increasing the likelihood of leading to further purchases.

[0060] The document creation support device 1 may further request the generative AI 3 to generate an image to be attached to the letter.

[0061] <1-6. Other examples> Although the cases where the learning target and the prediction target are customers in the first to third specific examples have been described, the present application is not limited thereto. The learning target and the prediction target may be, for example, medical examinees. By using "whether or not suffering from a specific disease" as the target variable for the learning target medical examinees and scoring the prediction target medical examinees, it is possible to predict a person who is likely to suffer from a specific disease in the future and set them as the action target. By having the generative AI 3 estimate the characteristics of the action target and generate a guidance text for preventive vaccination based on the characteristics, it is possible to obtain a more effective guidance text and help prevent diseases.

[0062] The learning target and prediction target can be employees. By using "whether promoted or not" as the target variable for the employee learning target and scoring the employees to be predicted, the action target can be selected according to the likelihood of promotion. Let the generative AI3 estimate the characteristics of the employees to be the action target, and let the generative AI3 generate advice and training guidance texts based on those characteristics, so that advice and training guidance texts suitable for those employees can be obtained.

[0063] When the learning target and prediction target are people, the characteristics to be estimated include the following. (1) Psychological motivation (purpose, motivation, needs, worries, background, feelings) (2) Preferences (hobbies, inclinations, likes and dislikes, favorite activities, styles, trends) (3) Thoughts and concepts (values, outlook on life, social outlook, worldview, family outlook, philosophy of life) (4) Personality traits (strengths, weaknesses, personality characteristics, behavior patterns, lifestyle) (5) Demographics (age, gender, occupation, income, education level, digital skills) (6) Cultural and regional characteristics (regionality, language, cultural values)

[0064] The learning target and prediction target are not limited to people and can also be things, such as products. By using "whether the product is a hit or not" as the target variable for the product learning target and scoring the products to be predicted, products that are likely to be hits in the future can be predicted and used as the action target. Let the generative AI3 estimate the characteristics of the products to be the action target, and let the generative AI3 generate promotional texts based on those characteristics, so that promotional texts with high promotional effects for those products can be obtained.

[0065] The learning target and the prediction target may be, for example, mechanical devices. By using "whether or not it has failed" as the target variable for the mechanical device to be learned and performing scoring on the mechanical device to be predicted, it is possible to predict mechanical devices that are likely to fail in the future and use them as the action targets. By having the generative AI3 estimate the characteristics of the mechanical device that becomes the action target and having the generative AI3 generate a maintenance guidance text for the owner of the mechanical device based on those characteristics, a more persuasive guidance text can be obtained.

[0066] When the learning target and the prediction target are things, the characteristics to be estimated are as follows. (1) Cost performance, durability, design, size, functionality, convenience (2) Customer satisfaction, brand reliability, history, safety (3) Exclusivity, rarity, eco-friendliness

[0067] <1-7. Effect> According to the first embodiment, the text creation support device 1 A model creation module 11 that creates a model showing the relationship between the first explanatory variable and the target variable based on the first explanatory variable and the target variable of each of a plurality of learning targets, A scoring module 12 that inputs the second explanatory variable of each of a plurality of prediction targets into the model and calculates the score of each of the plurality of prediction targets, A selection module 13 that selects a plurality of action targets from among the plurality of prediction targets based on the scores, A first prompt creation module 14 that creates a first prompt for instructing the generative AI3 to estimate the characteristics of each of the plurality of action targets by referring to the data related to each of the plurality of action targets, A second prompt creation module 15 that creates a second prompt for instructing the generative AI3 to generate a text considering the characteristics, including.

[0068] According to this, by selecting an action target based on a score, not only can the probability of achieving the goal be improved, but also an appealing sentence can be generated by considering each feature of the action target. Further, instead of generating a sentence directly from the data regarding each action target, by dividing it into a step of estimating features and a step of creating a sentence considering the features, it becomes clear under what kind of estimation the sentence was generated, and individual verification is also possible. Furthermore, since what kind of features should be estimated (advantages and disadvantages, purchase motivations and product preferences, dishes and proposed ingredients, etc.) can be specified in the first prompt, it becomes possible to direct the sentence to be generated. In this way, the features of the action target (person or thing) can be grasped in detail and multifariously, and a sentence suitable for each individual action target can be generated.

[0069] <2. Second Embodiment> <2-1. Functions and Operations> FIG. 7 shows the functions and operations of the sentence creation support device 1e according to the second embodiment. The sentence creation support device 1e further includes a third prompt creation module 16 and a variable creation module 17 in addition to the various modules included in the sentence creation support device 1 described with reference to FIG. 2. These modules are realized by a program being loaded into the memory included in the sentence creation support device 1e and executed by the CPU.

[0070] The third prompt creation module 16 acquires data regarding each of a plurality of learning targets and prediction targets from the database 2, and creates a third prompt for instructing the generation AI 3 to analyze the data of each of the plurality of learning targets and prediction targets. The data of each of the plurality of learning targets and prediction targets includes text data. Although a system for classifying text data has been conventionally known, the information contained in the text data could not be fully utilized only by classification. The sentence creation support device 1e can perform scoring with high accuracy by causing the generation AI 3 to analyze the text data. A specific example of creating the third prompt will be described later with reference to FIGS. 8A to 8C.

[0071] In the present application, the generative AI 3 that performs analysis in response to the third prompt may be referred to as the third generative AI. The third generative AI may be the same generative AI as either or both of the first and second generative AIs, or may be a separate generative AI.

[0072] The variable creation module 17 acquires the analysis results of the data regarding each of the learning target and the prediction target from the generative AI 3, and creates the first and second explanatory variables. The created explanatory variables are stored in the database 2. The first explanatory variable is used for model creation, the second explanatory variable is used for scoring, and the subsequent processing is the same as that in the first embodiment.

[0073] <2-2. Specific example> <2-2-1. Creation of the third prompt> FIG. 8A shows an example of the data regarding each of the learning target and the prediction target used for creating the third prompt in the specific example of the second embodiment. In this specific example, the learning target and the prediction target are stores. Among the plurality of stores included in FIG. 8A, those with unknown store names are less likely to be prediction targets, but can be learning targets. For the stores that are learning targets, by preparing the target variable separately, it becomes possible to create a model by the model creation module 11. The target variable is, for example, the presence or absence of trouble occurring at the store.

[0074] FIG. 8A shows the records of inquiries from customers regarding each store. The time zone is the time zone when the inquiry by phone was received, the business type is the business type of the store, and the classification is the classification of the inquiry content. The data of the time zone, business type, and classification may be automatically assigned or may be manually assigned. The text conversion of the inquiry voice may also be automatically performed or may be manually performed.

[0075] FIG. 8B shows an example of a third prompt generated in a specific example of the second embodiment. The third prompt includes an instruction sentence, analysis conditions, an input sentence, and a specification of an output format. The instruction sentence includes a sentence that requests the generation AI 3 to generate a sentiment analysis of a customer based on the text data of the inquiry voice. The instruction sentence may be stored in advance as a template. The input sentence is the text data of the inquiry voice shown in FIG. 8A.

[0076] FIG. 8C collectively shows examples of responses from the generation AI 3 to the third prompt shown in FIG. 8B for a plurality of stores. The generation AI 3 outputs positive, negative, and neutral scores.

[0077] <2-2-2. Variable Creation> The variable creation module 17 acquires the scores shown in FIG. 8C and creates first and second explanatory variables. The first and second explanatory variables may use the scores shown in FIG. 8C as they are, or may be further processed. As an example of a processed explanatory variable, a value obtained by subtracting the negative score from the positive score may be used. As another example of a processed explanatory variable, a principal component score obtained by performing principal component analysis using a value obtained by quantifying the time zone, business type, classification, etc. shown in FIG. 8A and the scores shown in FIG. 8C may be used.

[0078] By using the first and second explanatory variables obtained in this way, for example, the possibility of trouble occurring with customers in the store to be predicted can be scored, and before such trouble occurs, a warning message can be sent to the action target with a high score. To create this warning message, the generation AI 3 can be made to estimate the characteristics of the store to be the action target, and the generation AI 3 can be made to generate a sentence based on those characteristics, thereby obtaining an effective warning message for that store.

[0079] Although an example has been described above in which text data is given to the generative AI 3 for sentiment analysis, the present application is not limited to this. The generative AI 3 may be made to extract keywords from the text data, and for each of the extracted keywords, the presence or absence of the keyword may be used as an explanatory variable (0 or 1).

[0080] Although an example has been described above in which the text data given to the generative AI 3 is the text-converted data of an inquiry voice, the present application is not limited to this. For example, the text data may be a questionnaire entry sentence.

[0081] The text creation support device 1e may further request the generative AI 3 to generate an image to be attached to the attention-grabbing sentence.

[0082] <2-3. Effect> According to the second embodiment, the text creation support device 1e a third prompt creation module 16 that creates a third prompt for instructing a third generative AI to output an analysis result by referring to text data regarding each of a plurality of learning targets and a plurality of prediction targets; a variable creation module 17 that creates a first explanatory variable and a second explanatory variable based on the analysis result; further includes.

[0083] According to this, by analyzing text data to create an explanatory variable, the accuracy of the score can be improved. Even when the text data is insufficient, the large language model can probabilistically make the most likely interpretation and supplement the information, thus avoiding a decrease in the accuracy of the score.

Explanation of Signs

[0084] 1, 1e... Text creation support device, 2... Database, 3... Generative AI, 11... Model creation module, 12... Scoring module, 13... Selection module, 14... First prompt creation module, 15... Second prompt creation module, 16... Third prompt creation module, 17... Variable creation module

Claims

1. A model creation module that creates a model showing the relationship between the first explanatory variable and the objective variable based on the first explanatory variable and the objective variable of each of a plurality of learning targets; A scoring module that inputs the second explanatory variable of each of a plurality of prediction targets into the model and calculates the score of each of the plurality of prediction targets; A selection module that selects a plurality of action targets among the plurality of prediction targets based on the score; A first prompt creation module that creates a first prompt for instructing a first generative AI to estimate the characteristics of each of the plurality of action targets by referring to the data related to each of the plurality of action targets; A second prompt creation module that creates a second prompt for instructing a second generative AI to generate a sentence considering the characteristics; A sentence creation support device comprising:

2. The sentence creation support device according to Claim 1, A third prompt creation module that creates a third prompt for instructing a third generative AI to output an analysis result by referring to text data related to each of the plurality of learning targets and the plurality of prediction targets; A variable creation module that creates the first explanatory variable and the second explanatory variable based on the analysis result; The sentence creation support device further comprising:

Images