Text creation support device and text creation support method

Abstract

It helps generate customer-attracting text about products or services. [Solution] A text creation support device 30 comprises a storage device that stores text output from a trained model 31 that takes customer and product attributes as input and outputs text encouraging customers to purchase the product, and conversion information which is information about each customer's subsequent purchase history of the product based on the text, and a computing device that performs an evaluation information generation process that generates information representing the evaluation of the output text based on the conversion information.

Description

Technical Field

[0001] The present invention relates to a text creation support device and a text creation support method.

Background Art

[0002] With the technological development of Large Language Models (LLMs), attempts have been made to utilize LLMs for attracting customers. As a technology for this purpose, for example, Patent Document 1 discloses an information processing apparatus comprising: a transaction information acquisition unit that acquires transaction information including entries describing transactions at a store; a text separation unit that separates transaction text from the transaction information acquired by the transaction information acquisition unit; an attribute extraction unit that inputs the transaction text separated by the text separation unit into a plurality of natural language processing models to respectively extract a plurality of attributes corresponding to the transaction text; and an output unit that outputs the plurality of attributes extracted by the attribute extraction unit.

Prior Art Documents

Patent Documents

[0003]

Patent Document 1

Summary of the Invention

Problems to be Solved by the Invention

[0004] In attracting customers, it is important not only to analyze customer purchasing behavior using LLMs as in Patent Document 1, but also to provide information that can actually attract customers (for example, advertisements that lead to customers purchasing products). In this regard, it is conceivable to generate such information using an LLM similarly to Patent Document 1. However, since customers and products differ for each business operator, appropriate outputs cannot always be obtained with an LLM.

[0005] The present invention has been made in view of these circumstances, and its purpose is to provide a text creation support device and a text creation support method that can help generate customer-attracting text relating to a product or service. [Means for solving the problem]

[0006] One aspect of the present invention that solves the aforementioned problems is a text creation support device comprising a storage device that stores, respectively, text output from a trained model that takes customer and product attributes as input and outputs text encouraging the customer to purchase the product, and conversion information which is information about each customer's subsequent behavioral history regarding the purchase of the product based on the text, and a computing device that performs an evaluation information generation process that generates information representing the evaluation of the output text based on the conversion information. [Effects of the Invention]

[0007] According to the present invention, it is possible to help generate customer-attracting text related to a product or service.

[0008] Other configurations and effects will be clarified by the following description of the embodiments. [Brief explanation of the drawing]

[0009] [Figure 1] This figure shows an example of the configuration of the transaction support system 1 according to this embodiment. [Figure 2] This figure shows an example of the hardware provided by each information processing device in the transaction support system 1. [Figure 3] This figure shows an example of the functional components of the text creation support device 30. [Figure 4] This figure shows an example of a member table of 400. [Figure 5] This figure shows an example of product table 500. [Figure 6] This figure shows an example of the Hobbies and Preferences Table 600. [Figure 7] This figure shows an example of the recommended products table 700. [Figure 8] This figure shows an example of the recommended text table 800. [Figure 9] This figure shows an example of conversion table 900. [Figure 10] This figure shows an example of LLM management table 1000. [Figure 11] This figure shows an example of prompt template 1100. [Figure 12] This is a flowchart illustrating the overview of the processes performed in Transaction Support System 1. [Figure 13] This is a flowchart illustrating an example of the recommendation text provision process s1. [Figure 14] This is a flowchart illustrating the details of the information merging process s100. [Figure 15] This is a flowchart explaining the details of the recommendation text generation process s200. [Figure 16] This is a flowchart that explains the details of the model adjustment process s2. [Figure 17] This is a flowchart explaining the details of the tuning process s300. [Modes for carrying out the invention]

[0010] Hereinafter, embodiments of the present invention will be described with reference to the drawings.

[0011] Figure 1 is a diagram showing an example of the configuration of a transaction support system 1 according to the present embodiment. The transaction support system 1 includes a transaction support device 10 that is operated by a predetermined operator and provides predetermined products or services, one or a plurality of customer terminals 20 used by customers who purchase the aforementioned products or services, and a text creation support device 30 that supports the creation of texts (hereinafter referred to as recommendation texts) for prompting each customer to purchase products or services, which the transaction support device 10 provides to the customer terminals 20. In the following description, the customers are assumed to be members registered by the operator associated with the transaction support device 10, but the present invention is not intended to be limited to this configuration.

[0012] The respective information processing devices of the customer terminal 20, the transaction support device 10, and the text creation support device 30 are connected via a wired or wireless communication network 5 such as the Internet, a LAN (Local Area Network), a WAN (Wide Area Network), or a dedicated line.

[0013] The transaction support device 10 is, for example, an EC (Electronic Commerce) site. The transaction support device 10 displays information on products or services sold by the operator on the screen of the customer terminal 20, and accepts input of a purchase or a purchase candidate for the product or service (for example, adding to a shopping cart) from the member.

[0014] Furthermore, the transaction support device 10 transmits a notification (an advertisement such as an email) including a recommendation text related to a product or service created by the text creation support device 30 to the customer terminal 20. The customer terminal 20 displays the recommendation text received from the transaction support device 10 on the screen, so that the member can check the recommendation text. The member uses this recommendation text as a basis for deciding whether to purchase the product or service. In the following description, the recommendation texts are assumed to be generated for products, but the same applies to services.

[0015] Figure 2 shows an example of the hardware provided by each information processing device in the transaction support system 1. Each information processing device includes a computing device 91 such as a CPU (Central Processing Unit) or GPU (Graphics Processing Unit), a main memory 92 such as RAM (Random Access Memory) or ROM (Read Only Memory), an auxiliary storage device 95 such as an HDD (Hard Disk Drive) or SSD (Solid State Drive), an input device 97 consisting of a keyboard, mouse, touch panel, etc., an output device 98 for screen display consisting of a monitor (display), etc., an input interface 93 for sending and receiving data with the input device 97, an output interface 96 for sending and receiving data with the output device 98, and a communication device 94 consisting of a NIC (Network Interface Card), wireless communication module, USB (Universal Serial Interface) module, or serial communication module, etc.

[0016] Figure 3 shows an example of the functional components of the text creation support device 30. First, the text creation support device 30 stores data for the trained model 31, member table 400, product table 500, hobby / preference table 600, recommended product table 700, recommended text table 800, conversion table 900, LLM management table 1000, and prompt template 1100.

[0017] The trained model 31 is a mathematical model (for example, a Large Language Model (LLM)) that takes text prompts as input and outputs text corresponding to the input prompts. The text creation support device 30 stores each version of the trained model 31, which is managed by the LLM management table 1000 described later.

[0018] The trained model 31 is a neural network composed of various parameters and functions, and has an input layer into which input data is input, one or more intermediate layers (hidden layers) that extract and output features from the input data, and an output layer that outputs output data from the features. For example, the trained model 31 is a model that takes the attributes of a member or product as input and outputs text (recommendation text) that encourages the member to purchase that product. The trained model 31 can be constructed using, for example, BERT (Bidirectional Encoder Representations from Transformers), XLNet, or GPT (Generative Pre-trained Transformer). Note that the trained model 31 may be stored in an information processing device other than the text creation support device 30, and the text creation support device 30 may call the trained model 31.

[0019] Member table 400 is a table that stores attribute information for each member.

[0020] Product table 500 is a table that stores attribute information for each product.

[0021] The Hobbies and Preferences Table 600 is a table that stores information about each member's hobbies and preferences, in particular, from among the attribute information of each member.

[0022] The Recommended Products Table 700 is a table that stores information on products that are considered to be particularly likely to be purchased by each member, based on their attribute information (hereinafter referred to as "recommended products").

[0023] The recommendation table 800 is a table that stores attribute information for recommendation texts.

[0024] Conversion Table 900 is a table that stores conversion information that formed the basis for generating recommendation messages. Specifically, Conversion Table 900 contains information about each member's subsequent purchase history of products based on the recommendation messages.

[0025] The LLM management table 1000 is a table that manages the trained models 31 for each version.

[0026] The prompt template 1100 is a prompt template that will be input to the trained model 31.

[0027] (Member Table) Figure 4 shows an example of the member table 400. The member table 400 contains the following data: item number 401, each member's ID (member ID 402), each member's last name 403, each member's first name 404, each member's gender 405, and each member's age 406. The member table 400 is obtained, for example, from the transaction support device 10 and updated as needed. Note that the member attributes shown here are just examples and may be given attributes of the member, or information obtained by the business operator through member behavior analysis.

[0028] (Product Table) Figure 5 shows an example of a product table 500. The product table 500 contains the following data: item number 501, product ID (product ID 502), category ID (category ID 503), product name 504, target gender 505, size 506, and URL 507 of the website (transaction support device 10) page where information about each product is posted. The product table 500 is obtained, for example, from the transaction support device 10 and updated as needed.

[0029] (Hobbies and Preferences Table) Figure 6 shows an example of a hobby / preference table 600. The hobby / preference table 600 contains the following data: item number 601, ID of each hobby / preference (hobby / preference ID 602), member ID 603 of the member associated with each hobby / preference, and content 604 of the hobby / preference corresponding to the ID. The hobby / preference table 600 is obtained, for example, from the transaction support device 10 and updated as needed.

[0030] (Recommended Products Table) Figure 7 shows an example of the recommended products table 700. The recommended products table 700 contains the following data: item number 701, each member's ID (member ID 702), each member's recommended product ID (product ID 703), and data update date 704. The recommended products table 700 is set up automatically or by the administrator, for example, based on the member's usage history of the transaction support device 10 (identified in the conversion table 900). For example, when a certain product is purchased or reserved by a member, the text creation support device 30 sets that product or other products belonging to the same category as that product as recommended products for that member.

[0031] (Recommended Text Table) Figure 8 shows an example of the recommendation text table 800. The recommendation text table 800 stores information on multiple texts (the text of the recommendation text). Specifically, the recommendation text table 800 contains the following data: item number 801, recommendation text ID (recommendation text ID 802), member ID associated with the recommendation text (member ID 803), product ID 804 associated with the product, the content of the recommendation text 805, LLMID 806 which is the ID of the LLM (trained model 31) that generated the recommendation text, conversion ID 807 which is the ID of the conversion on which the recommendation text is based, and data update date 808. The content of the recommendation text 805 in the recommendation text table 800 is based, for example, on a recommendation text obtained from the transaction support device 10 or generated by the text creation support device 30.

[0032] (Conversion Table) Figure 9 shows an example of a conversion table 900. The conversion table 900 contains the following data for each conversion: item number 901 for each conversion, ID for each conversion (conversion ID 902), ID of the member corresponding to each conversion (member ID 903), ID of the product corresponding to each conversion (product ID 904), a flag indicating whether or not the member purchased the product (purchase status 905), the date and time the member purchased the product (purchase date and time 906), a flag indicating whether or not the member accessed the product page on the transaction support device 10 (product site visit 907), a flag indicating whether or not the notification of the recommendation text for the product sent to the member was opened on the customer terminal 20 (whether the member accessed the recommendation text) (email delivery status 908), and a conversion score 909 (evaluation value of the recommendation text corresponding to that conversion).

[0033] The conversion table 900 may be obtained, for example, from the transaction support device 10, or it may be entered by the administrator.

[0034] (LLM Management Table) Figure 10 shows an example of the LLM management table 1000. The LLM management table 1000 contains the following data: item number 1001, ID of each version of the LLM (LLMID 1002), storage location of each version of the trained model 31 in the text creation support device 30 (LLM storage path 1003), and data update date 1004.

[0035] (Prompt template) Figure 11 shows an example of a prompt template 1100. This template 1100 includes an instruction section 1101 that instructs the system to generate recommendation texts based on the attributes of the member and the products, a product attribute specification section 1102 in which the attributes of the products are set, and a customer attribute specification section 1103 in which the attributes of the member are set.

[0036] Next, as shown in Figure 2, the text creation support device 30 includes the following functional units: a recommended sentence generation unit 301, a recommended sentence transmission unit 302, a DB access unit 303, an information acquisition unit 304, an information merging unit 305, an evaluation information generation unit 306, and an adjustment unit 307.

[0037] The recommendation text generation unit 301 generates recommendation texts for each member using the trained model 31.

[0038] The recommendation message transmission unit 302 sends the recommendation message generated by the recommendation message generation unit 301 to the transaction support device 10. The transaction support device 10 then attaches the received recommendation message to the corresponding product information and sends it to the corresponding member's customer terminal 20.

[0039] The DB access unit 303 accesses each table.

[0040] The information acquisition unit 304 acquires predetermined data from each table. For example, the information acquisition unit 304 acquires and stores the text (recommendation text) output from the trained model 31 and the conversion information (conversion table 900), which is information about the purchase history of each member who accessed the recommendation text (for example, when the recommendation text was displayed on the screen of the customer terminal 20).

[0041] The information merging unit 305 combines the data acquired by the information acquisition unit 304 to generate merged data. The merged data is information that combines the attributes of each member, the attributes of the products, and the conversion information, with each member's attributes and product attributes being associated with each other.

[0042] The evaluation information generation unit 306 generates information representing the evaluation of the text (recommendation text) output from the trained model 31, based on the conversion information (conversion table 900). In this embodiment, the evaluation information generation unit 306 calculates an evaluation value (score) that represents the evaluation of the output text.

[0043] The adjustment unit 307 adjusts the parameters that make up the trained model 31 based on the information (score) generated by the evaluation information generation unit 306 and the text output from the trained model 31. With the parameters of the trained model 31 adjusted by the adjustment unit 307, the recommendation sentence generation unit 301 becomes able to generate more appropriate recommendation sentences.

[0044] The functional units of the text creation support device 30 described above are realized by the hardware of the text creation support device 30, or by the arithmetic unit 91 of the text creation support device 30 reading and executing programs stored in the main memory 92 or auxiliary memory 95. These programs are stored in storage devices such as secondary storage devices, non-volatile semiconductor memory, hard disk drives, SSDs, or recording media readable by each device, such as IC cards, SD cards, or DVDs. Each device may be realized in whole or in part using virtual information processing resources provided using virtualization technology or process space isolation technology, such as a virtual server provided by a cloud system. Furthermore, all or part of the functions provided by these devices may be realized by services provided by the cloud system via an API (Application Programming Interface), for example.

[0045] Next, we will explain the processes performed in the transaction support system 1.

[0046] Figure 12 is a flowchart illustrating the overview of the processes performed by the trading support system 1. The trading support system 1 executes a recommendation statement provision process s1, which generates recommendation statements and provides them to the trading support device 10.

[0047] Furthermore, the transaction support system 1 executes a model tuning process s2 that adjusts (tunes) the trained model 31 based on the member's subsequent behavioral history based on the generated recommendation text, enabling the generation of more appropriate recommendation texts. Note that the recommendation text provision process s1 and the model tuning process s2 are independent processes and may be executed repeatedly.

[0048] The details of each process are explained below.

[0049] <Recommendation text provision process s1> Figure 13 is a flowchart illustrating an example of the recommendation text provision process s1. The recommendation text provision process s1 is started, for example, at a predetermined timing (e.g., a predetermined time, a predetermined time interval), or when the administrator makes a predetermined input to the text creation support device 30.

[0050] The text creation support device 30 obtains the member table 400, the product table 500, the hobby / preference table 600, the recommended product table 700, and the recommended text table 800, respectively (s11).

[0051] The text creation support device 30 uses the information obtained in s11 as input values ​​to execute the information merging process s100 and generate merged information (s12). Specifically, the text creation support device 30 obtains merged information for each member by specifying each member registered in the member table 400 and calling the information merging process s100 for each member.

[0052] The text creation support device 30 obtains the prompt template 1100 (s13).

[0053] The text creation support device 30 executes a recommendation sentence generation process s200 for each member that was processed in s12, using the combined information obtained in s11 and the prompt template 1100 obtained in s13 as input values ​​to generate a new recommendation sentence (s14).

[0054] The text creation support device 30 sends each recommendation sentence output by the recommendation sentence generation process s200 to the transaction support device 10, associating it with the corresponding member and product (s15).

[0055] Subsequently, the transaction support device 10 generates information (promotional information) encouraging the purchase of the corresponding product based on the received recommendation message, and displays the generated promotional information on the screen of the corresponding member's customer terminal 20.

[0056] <Information merging process> Figure 14 is a flowchart illustrating the details of the information merging process s100. The information merging process s100 is executed for the specified member (hereinafter referred to as "the member").

[0057] The text creation support device 30 combines information from the member table 400 and the hobby / preference table 600 using the member ID as a key, and generates combined information that associates the member's attribute information with their hobbies and preferences (s101).

[0058] The text creation support device 30 refers to the recommended products table 700 and checks whether there are any recommended products for the member in question (s102).

[0059] If there are recommended products (s102:YES), the text creation support device 30 executes the process in s103. If there are no recommended products (s102:NO), the information merging process s100 terminates.

[0060] In s103, the text creation support device 30 uses the member ID and product category ID as keys to combine information from the product table 500 and the recommended product table 700 to identify products for which recommendation text should be generated for the member (hereinafter referred to as "target products"), and combines this information with the combined information generated in s101. For example, the text creation support device 30 refers to the product table 500, identifies all products in the same category as the recommended product identified in s102, and obtains the record for each identified product. The text creation support device 30 then combines the information from each obtained record with the combined information generated in s101 (record information from the member table 400 and the hobby / preference table 600).

[0061] The text creation support device 30 then refers to the recommendation text table 800 and determines whether a recommendation text corresponding to each of the member's target products is registered (s104).

[0062] If recommendation texts corresponding to each of the member's target products are registered (s104:YES), the text creation support device 30 executes the process s105 for those target products. If no recommendation texts corresponding to each of the member's target products are registered (s104:NO), the information merging process s100 for the member is terminated.

[0063] In s105, the text creation support device 30 identifies the conversion information for each target product corresponding to the member in question using the information in the recommendation text table 800, and combines it with the combined information generated in s103. For example, the text creation support device 30 retrieves the information of a record in the recommendation text table 800 in which the member ID 803 is set to the member ID and the product ID of the target product for the member is set to the product ID 804, and combines the retrieved information with the combined information. With this, the information merging process s100 for the member in question is completed.

[0064] <Recommended text generation process> Figure 15 is a flowchart illustrating the details of the recommendation text generation process s200.

[0065] The text creation support device 30 obtains the input value, which is the prompt template 1100, and the combined information obtained in s12 (s201).

[0066] The text creation support device 30 inputs a prompt with combined information set in the prompt template 1100 to the trained model 31, causing it to output recommendation texts for each recommended product for the member, and then retrieves each outputted recommendation text (s202).

[0067] Specifically, the text creation support device 30 sets the product information from the combined information in the product attribute specification section 1102 of the prompt template 1100, and sets the member attribute information from the combined information in the customer attribute specification section 1103. The text creation support device 30 inputs the prompt with this information into the trained model 31 to obtain a recommendation sentence.

[0068] The text creation support device 30 adds the recommended text obtained in s202 to the recommended text table 800 (s203). Specifically, the text creation support device 30 generates a new record in the recommended text table 800 and sets the recommended text obtained in s202 to the content of the recommended text 805 of the generated record. The text creation support device 30 also sets the member ID 803 to the member's ID, the product ID corresponding to the member to the product ID 804 to the product ID of the member, the LLMID of the LLM (trained model 31) that generated the recommended text to the LLMID 806 to the member's conversion ID 807 to the member's conversion ID and the current date to the update date 808.

[0069] <Model adjustment process s2> Figure 16 is a flowchart illustrating the details of the model adjustment process s2. The model adjustment process s2 is executed, for example, at predetermined timings (e.g., predetermined time intervals, predetermined times).

[0070] The text creation support device 30 obtains the member table 400, the product table 500, the hobby / preference table 600, the recommended product table 700, and the recommended text table 800, respectively (s21).

[0071] The text creation support device 30, similar to s12, executes information merging process s100 using each piece of information acquired in s11 as input values, and obtains the merged information (s22).

[0072] Furthermore, the text creation support device 30 obtains the conversion table 900 (s23).

[0073] The text creation support device 30 combines the information from the conversion table 900 and the information from the recommendation text table 800 using the conversion ID as a key, and generates information that associates the member's conversion with the recommendation text (s24).

[0074] The text creation support device 30 calls a tuning process s300 to update the trained model 31 using the information acquired or generated by s24 as input values ​​(s26).

[0075] The text creation support device 30 stores the LLM updated by the tuning process s300 and registers information related to that LLM in the LLM management table 1000 (s27). Specifically, the text creation support device 30 registers the path of the LLM updated in s26 in the LLM storage path 1003 of the new record in the LLM management table 1000, and registers the current date in s26 in the update date 1004.

[0076] <Tuning process s300> Figure 17 is a flowchart illustrating the details of the tuning process s300.

[0077] The text creation support device 30 obtains conversion information for each member from the conversion table 900 (s301).

[0078] The text creation support device 30 generates information representing each member's evaluation of the recommendation text based on the information acquired in s301 (s302).

[0079] In this embodiment, the text creation support device 30 calculates an evaluation value indicating how much the recommendation text contributed to the member's purchase of the product. For example, the text creation support device 30 calculates each evaluation value using a method such that the value when a member who accessed the recommendation text subsequently purchased the product is higher than the value when the member who accessed the recommendation text did not subsequently purchase the product.

[0080] Specifically, the text creation support device 30 may calculate a score for each item in the conversion table 900 and sum up the calculated scores. For example, the text creation support device 30 may set a high predetermined score if the product is purchased, a low predetermined score if it is not purchased, an intermediate predetermined score if the product is not purchased but is considered for purchase, and a higher score if the product is purchased earlier. Alternatively, for example, the text creation support device 30 may calculate the score by inputting a prompt to the trained model 31 instructing it to calculate a score corresponding to the content of each item.

[0081] The text creation support device 30 creates a dataset for optimizing the recommended sentences output by the trained model 31 based on the score calculated in s302 (s303).

[0082] For example, the text creation support device 30 identifies recommended sentences that correspond to combinations of member attributes, hobbies and preferences, and product attributes, and whose scores are above a predetermined value (e.g., 0.8 or higher) (i.e., preferred recommended sentences). The text creation support device 30 stores these combinations and recommended sentences as a dataset related to preferred recommended sentences. Also, for example, the text creation support device 30 identifies recommended sentences that correspond to combinations of member attributes, hobbies and preferences, and product attributes, and whose scores are below a predetermined value (e.g., 0.3 or lower) (i.e., undesirable recommended sentences). The text creation support device 30 stores these combinations and recommended sentences as a dataset related to undesirable recommended sentences. The dataset may include score information.

[0083] Note that the method of creating datasets described here is just one example. For example, the text creation support device 30 may set combinations that consider only some of the member's attributes, or it may exclude product attributes from the combination elements. Also, for example, the text creation support device 30 may include input data (prompts) for the trained model 31 corresponding to the outputted recommendation sentences in each dataset.

[0084] Then, the text creation support device 30 optimizes the parameters in the trained model 31 based on the dataset created in s303 (s304).

[0085] The optimization method for the trained model 31 is not particularly limited, but for example, the text creation support device 30 fine-tunes the trained model 31 using DPO (Direct Preference Optimization). In this case, the text creation support device 30 trains the trained model 31 based on the created dataset.

[0086] Furthermore, for example, the text creation support device 30 may employ a reinforcement learning algorithm such as RHLF (Reinforcement Learning from Human Feedback). In this case, the text creation support device 30 creates a reward model using a dataset and trains the trained model 31 to maximize the reward from this reward model (PPO: Proximal Policy Optimization).

[0087] By performing this type of tuning, for example, the trained model 31 learns specifically from recommendation sentences that scored highly (for example, recommendation sentences that customers actually read and then purchased), and learns to generate new recommendation sentences that have the characteristics of those recommendation sentences.

[0088] As described above, the text creation support device 30 of this embodiment acquires recommendation sentences output from the trained model 31 and conversion table information, which is information on the subsequent purchase history of each customer based on those recommendation sentences. Based on the conversion information, it generates information representing the evaluation of the output recommendation sentences.

[0089] In other words, the text creation support device 30 of this embodiment evaluates the recommendation sentences output from the trained model 31 based on customer conversion information and generates information representing the evaluation of those recommendation sentences.

[0090] This evaluation information can be used to help provide products or services using recommendation text. Thus, the text creation support device 30 of this embodiment can help generate text that can attract customers regarding products or services.

[0091] Furthermore, the text creation support device 30 of this embodiment adjusts the parameters that constitute the trained model 31 based on the evaluation information generated above and the recommended sentences output above.

[0092] By updating the trained model 31 based on the evaluation results of the recommended sentences, the trained model can be tuned to generate more suitable recommended sentences.

[0093] Specifically, the text creation support device 30 of this embodiment calculates an evaluation value (score) that represents the evaluation of the recommended sentences, and adjusts the parameters that constitute the trained model 31 based on a dataset that associates the calculated scores with the recommended sentences.

[0094] This allows for a trained model 31 that accurately reflects the quality of the recommended sentences.

[0095] Furthermore, the text creation support device 30 of this embodiment calculates an evaluation value for each recommendation sentence based on conversion information such that the value when a customer subsequently purchases a product based on the recommendation sentence is higher than the value when a customer subsequently does not purchase a product based on the recommendation sentence.The text creation support device 30 then adjusts the parameters that constitute the trained model 31 based on a dataset that associates desirable recommendation sentences with evaluation values ​​above a predetermined value and the attributes of the customer or product related to those recommendation sentences, and a dataset that associates undesirable recommendation sentences with evaluation values ​​below a predetermined value and the attributes of the customer or product related to those recommendation sentences.

[0096] In this way, by tuning the trained model 31 using both favorable and unfavorable recommendation statements in relation to customer or product attributes, it is possible to train a trained model 31 that can create recommendation statements that accurately reflect customer or product attributes.

[0097] The above description of embodiments is for the purpose of facilitating understanding of the present invention and does not limit it. The present invention can be modified and improved without departing from its spirit, and equivalents thereof are included.

[0098] For example, some of the hardware components of each device in this embodiment may be provided in other devices.

[0099] Furthermore, each program of each device may be provided in other devices, a program may consist of multiple programs, or multiple programs may be integrated into a single program.

[0100] The method for generating evaluation information (calculating evaluation values ​​and scores) described in this embodiment is just one example; any evaluation method that indicates how much the recommendation text contributed to the customer's purchase of a product or service can be adopted.

[0101] Furthermore, in this embodiment, the method described for a member to access the recommendation text was that the transaction support device 10 sends a notification (email, etc.) to the customer terminal 20, but other methods may also be adopted. For example, when a member's customer terminal 20 accesses a product page on the transaction support device 10, the transaction support device 10 may automatically display the recommendation text for that product for that member, thereby allowing the member to access the recommendation text. [Explanation of symbols]

[0102] 30 Text creation support device, 31 Trained model, 301 Recommended sentence generation unit, 305 Information merging unit, 306 Evaluation information generation unit, 307 Adjustment unit

Claims

1. A storage device that stores text output from a trained model that takes customer and product attributes as input and outputs text encouraging the customer to purchase the product, and conversion information which is information about each customer's subsequent purchase history of the product based on the text, and A computing device that performs an evaluation information generation process to generate information representing the evaluation of the output text based on the aforementioned conversion information. A text creation support device equipped with the following features.

2. The aforementioned computing device is Based on the generated information and the output text, a tuning process is further performed to adjust the parameters that constitute the trained model. The text creation support device according to claim 1.

3. The aforementioned computing device is In the evaluation information generation process, an evaluation value representing the evaluation of the output text is calculated based on the conversion information. In the tuning process, the parameters constituting the trained model are adjusted based on the data that associates the calculated evaluation value with the output text. The text creation support device according to claim 2.

4. The storage device acquires multiple texts output from the trained model and each conversion information related to each text, The aforementioned computing device is In the evaluation information generation process, based on the conversion information, an evaluation value is calculated for each text such that the value when the customer subsequently purchases the product based on the text is higher than the value when the customer subsequently does not purchase the product based on the text. In the tuning process, the parameters constituting the trained model are adjusted based on data that associates texts with an evaluation value greater than or equal to a predetermined value with customer or product attributes related to the texts, and data that associates texts with an evaluation value less than or equal to a predetermined value with customer or product attributes related to the texts. The text creation support device according to claim 2.

5. A text creation support method using an information processing device comprising a storage device and a computing device, wherein the computing device is configured to store text output from a trained model that takes customer and product attributes as input and outputs text encouraging the customer to purchase the product, and conversion information which is information about each customer's subsequent purchase history of the product based on the text, the computing device is configured to store text output from a trained model that takes customer and product attributes as input and outputs text encouraging the customer to purchase the product, and conversion information which is information about each customer's subsequent purchase history of the product based on the text, and the computing device is configured to store text output from a trained model that takes customer and product attributes as input and outputs text encouraging the customer to purchase the product, and conversion information which is information about each customer's subsequent purchase history of the product based on the text, and the computing device is configured to store text output from a trained model that takes customer and product attributes as input and outputs text encouraging the customer to purchase the product, and the computing device is configured to store text output from a trained model that takes customer and product attributes as input and outputs text encouraging the customer to purchase the product, and the computing device is configured to store text output from a trained model that takes customer and product attributes as input and outputs text encouraging the customer to purchase the product, and conversion information which is information about each customer Based on the aforementioned conversion information, an evaluation information generation process is executed to generate information representing the evaluation of the output text. Text creation support methods.

6. The aforementioned computing device Based on the generated information and the output text, a tuning process is further performed to adjust the parameters that constitute the trained model. The text creation support method according to claim 5.

7. The aforementioned computing device In the evaluation information generation process, an evaluation value representing the evaluation of the output text is calculated based on the conversion information. In the tuning process, the parameters constituting the trained model are adjusted based on the data that associates the calculated evaluation value with the output text. The text creation support method according to claim 6.

8. The storage device acquires a plurality of texts output from the trained model and each conversion information related to each of the texts, The aforementioned computing device In the evaluation information generation process, based on the conversion information, an evaluation value is calculated for each text such that the value when the customer subsequently purchases the product based on the text is higher than the value when the customer subsequently does not purchase the product based on the text. In the tuning process, the parameters constituting the trained model are adjusted based on data that associates texts with an evaluation value greater than or equal to a predetermined value with customer or product attributes related to the texts, and data that associates texts with an evaluation value less than or equal to a predetermined value with customer or product attributes related to the texts. The text creation support method according to claim 6.

Images