A device for generating fortune-telling texts

Abstract

To provide users with a large amount of up-to-date information reflected in fortune-telling and other results, thereby encouraging them to input further user attributes. [Solution] The fortune-telling text generation device 1 according to the present invention comprises an attribute acquisition unit 112 that acquires one or more user attributes relating to a user, a search word generation unit 117 that generates search words used to refer to academic literature published in an external academic literature database D based on the user attributes, and a fortune-telling text generation unit 118 that inputs instructions to generate fortune-telling text and prompts including the search words into a large-scale language model that refers to academic literature, and generates fortune-telling text relating to the user described above. The generation device 1 of the present invention provides step-by-step protection of personal information by an attribute management unit 111 that manages user attributes in association with their scope of publication.

Description

Technical Field

[0001] The present invention relates to a fortune-telling text generation device.

Background Art

[0002] Based on the user attributes input by the user, the portrait of the user is narrowed down and utilized for personality diagnosis, fortune-telling, and other applications. In such a case, if there are few input user attributes, the portrait of the user cannot be sufficiently narrowed down, and there is a possibility that the application cannot be appropriately implemented.

[0003] However, if an attempt is made to have many user attributes input from the beginning, there is concern that the user may abandon the use of the application. This is because the user hesitates to input user attributes, for example, due to its complexity and concerns about personal information leakage.

[0004] Therefore, there is a need for technical means to attract more user interest and make the user voluntarily input many user attributes.

[0005] Regarding technologies that further attract user interest and encourage users to voluntarily input more user attributes in conventional services and technologies, Patent Document 1 discloses a fortune-telling device characterized by comprising: means for displaying a screen prompting a fortune-telling requester to input personal information; means for displaying a screen prompting the fortune-telling requester to select the content of the fortune-telling requester from two or more pre-prepared standard phrases; means for receiving input of the personal information and input for selecting either of the standard phrases, and reading one or more question sentences corresponding to the selected standard phrase and two or more answer selection options corresponding to the question sentences from a pre-prepared table; means for displaying the read question sentences and answer options on a screen and prompting input for selecting an answer option; and means for receiving input for selecting an answer option and outputting a fortune-telling result based on the received personal information input, the standard phrase selection input and the answer option input. Furthermore, Patent Document 1 shows a method for searching a psychology database that provides certain answers that differ depending on the answer options, based on known psychology, and outputting a mixture of search results from the fortune-telling database and the psychology database.

[0006] The technology described in Patent Document 1 can enable users of fortune-telling to feel satisfied that they have received direct answers to the specific problems they are facing, which is what they were looking for in fortune-telling. [Prior art documents] [Patent Documents]

[0007] [Patent Document 1] Japanese Patent Publication No. 2007-236755 [Overview of the project] [Problems that the invention aims to solve]

[0008] Incidentally, psychologists have discovered a vast amount of knowledge related to psychology, and moreover, such knowledge is constantly being updated. Reflecting this vast amount of knowledge in the psychology database of Patent Document 1 would require considerable effort. Furthermore, even if a psychology database reflecting this vast amount of knowledge were to be built once, adding questions and answer choices based on newly discovered knowledge to such a database and continuously reflecting the latest findings would place a heavy burden on the administrator of the psychology database.

[0009] Therefore, the technology described in Patent Document 1 is limited to providing users with answers that reflect psychological insights to the extent permitted by the economic circumstances of the application. Thus, the technology described in Patent Document 1 has room for further improvement in providing users with fortune-telling results that reflect a large amount of up-to-date knowledge and in encouraging them to input further user attributes.

[0010] The object of the present invention is to provide a technical means that prompts users to input further user attributes by providing them with results of fortune-telling, etc., that reflect a large amount of up-to-date knowledge. In this invention, "fortune-telling, etc." is not limited to those titled "fortune-telling," but refers to all applications that refer to the user's personality, such as fortune-telling, personality tests, psychological tests, and personality assessments. [Means for solving the problem]

[0011] In response to the aforementioned challenges, a portion of academic literature databases containing knowledge from various academic fields has been made public in order to make researchers' findings known to the world. To utilize this knowledge, a large-scale language model has been realized that has the ability to search academic literature databases and generate answers based on the search results. Therefore, at first glance, it seems that by simply instructing such a large-scale language model to search academic literature databases using the parameters used for fortune-telling and to generate fortune-telling text, it would be possible to provide users with fortune-telling results that reflect a large amount of up-to-date knowledge.

[0012] In other words, for example, for a user born on April 1, 2000, by passively partitioning the birth year into the Chinese zodiac to make 2000 the "Year of the Dragon," and by passively partitioning the birth date into the ruling constellation to make April 1, 2000 the "Aries," a prompt is generated stating, "For subjects who are both in the Year of the Dragon and the Aries, please search the academic literature database for information and generate astrology text," and by implementing a system to input this prompt into a large-scale language model, it appears at first glance that the above problem can be solved.

[0013] However, it is known that large-scale language models may sometimes respond as if non-existent search results exist when they cannot obtain search results that directly correspond to a given question. In the research, development, and application of large-scale language models, this phenomenon is called hallucination. Hallucination is thought to be caused by various factors, exemplified by the properties of the training data, training process, or architecture.

[0014] The inventors noticed that when they passively segmented user attributes according to criteria related to fortune-telling and then prompted a large-scale language model to generate fortune-telling texts based on these results, a phenomenon resembling hallucination occurred frequently. The inventors then performed the same processing on multiple types of large-scale language models and, despite referencing the same academic literature database, the generated fortune-telling texts differed in meaning, leading them to believe that hallucination was indeed occurring.

[0015] The frequent occurrence of hallucinations in this process is thought to stem from the fact that, given the general perception that fortune-telling is unscientific, knowledge directly corresponding to the results of passive division regarding fortune-telling is either not stored in academic literature databases, or if it is, only a small amount. Frequent hallucinations mean that fortune-telling texts cannot be said to reflect a large amount of up-to-date knowledge.

[0016] Therefore, in order to solve the above problems, the inventors diligently investigated and explored means of solving the hallucination problem. As a result, they found that the above objectives can be achieved by generating search words for an academic literature database from user attributes, inputting instructions to generate fortune-telling text and prompts containing the search words into a large-scale language model that references academic literature. Thus, the inventors completed the present invention.

[0017] The invention relating to the first feature of the present invention provides a fortune-telling text generation device comprising: an attribute acquisition unit that acquires one or more user attributes relating to a user; a search word generation unit that generates search words used for referencing academic literature published in an external academic literature database based on the user attributes; and a fortune-telling text generation unit that inputs instructions for generating fortune-telling text and prompts including the search words to a large-scale language model that references the academic literature, thereby generating fortune-telling text relating to the user.

[0018] This invention converts user attributes acquired by an attribute acquisition unit into search terms used for referencing in an academic literature database using a search term generation unit. In other words, this invention actively divides user attributes using a search term generation unit.

[0019] As a result, the large-scale language model in the fortune-telling text generation unit of this invention can reference academic literature using search words generated by active segmentation suitable for referencing academic literature, rather than passive segmentation related to fortune-telling that is prone to hallucination. Furthermore, the fortune-telling text generation unit can generate fortune-telling texts that reflect a large amount of up-to-date knowledge from various academic fields, such as psychology or sociology, stored in an academic literature database, using the large-scale language model. As a result, this invention can provide users with persuasive fortune-telling texts that reflect a large amount of up-to-date knowledge and encourage them to input further user attributes. In this invention, "fortune-telling text" is not limited to texts titled "fortune-telling," but includes all kinds of texts that include descriptions of the user's personality, such as fortune-telling results, personality diagnosis results, psychological test results, and personality diagnosis results.

[0020] Therefore, this invention can provide a technical means to encourage users to input further user attributes by providing them with results such as fortune-telling that reflect a large amount of up-to-date knowledge.

[0021] The invention relating to the second feature of the present invention is the invention relating to the first feature, wherein the search word generation unit generates the search words by a process that includes a procedure for performing active segmentation to convert the user attributes into search words for referring to academic literature, and the fortune-telling text generation unit provides a generation device that generates fortune-telling text associated with the results of passively segmenting the user attributes according to criteria related to fortune-telling.

[0022] The present invention generates fortune-telling text that includes a description that correlates the results of passive division according to criteria related to fortune-telling, such as passive division that converts the year of birth into the Chinese zodiac, and passive division that converts the date of birth into a guardian constellation, with knowledge obtained from academic literature, using a fortune-telling text generation unit.

[0023] Therefore, the present invention can provide users with persuasive divination texts that reflect a large amount of up-to-date knowledge without compromising the mystery of divination. As a result, the present invention can further encourage the input of additional user attributes. Thus, the present invention can provide technical means for giving users divination results and the like that reflect a large amount of up-to-date knowledge and encouraging the input of additional user attributes.

[0024] The invention according to the third feature of the present invention is the invention according to the first feature or the second feature, wherein the attribute acquisition unit is configured to acquire user attributes indicating the user's troubles, the search word generation unit is configured to further generate trouble search words based on the troubles, and the divination text generation unit is configured to input a prompt further including the trouble search words to the large language model, and provides a generation device.

[0025] Those seeking divination often look to divination for clues to solve their own troubles. The present invention allows such troubles to be input by the users themselves, and refers to academic literature using search words obtained by actively dividing these troubles.

[0026] Therefore, the present invention can provide users with divination texts that are persuasive, reflect appropriate knowledge corresponding to the troubles, have a sense of mystery, while encouraging the input of user attributes related to the troubles. As a result, the present invention can further encourage the input of additional user attributes.

[0027] Thus, the present invention can provide technical means for giving users divination results and the like that reflect a large amount of up-to-date knowledge and encouraging the input of additional user attributes.

[0028] The invention relating to the fourth feature of the present invention relates to one or more features selected from the first, second, or third features, wherein the fortune-telling text generation unit causes the large-scale language model to generate fortune-telling text that further includes wording prompting the user to input additional user attributes, the attribute acquisition unit is configured to further acquire additional user attributes that have been added based on the wording, the search word generation unit is configured to further generate additional search words based on the additional user attributes, and the fortune-telling text generation unit inputs a prompt further including the additional search words to the large-scale language model, thereby providing a generation device.

[0029] The invention generates fortune-telling text that includes wording prompting the user to input additional user attributes, using a fortune-telling text generation unit. The invention then generates further fortune-telling text based on the user attributes that the user has additionally input after seeing this wording.

[0030] Therefore, this invention can provide a technical means to encourage users to input further user attributes by providing them with results such as fortune-telling that reflect a large amount of up-to-date knowledge.

[0031] The invention relating to the fifth feature of the present invention relates to one or more features selected from the first, second, third, or fourth features, and provides a generation device in which the large-scale language model in the fortune-telling text generation unit refers to the academic literature using Retrieval-Augmented Generation (RAG) that has been pre-trained based on the academic literature.

[0032] As mentioned above, large-scale language models can experience hallucinations. Hallucinations can be broadly categorized into intrinsic hallucinations, which output facts different from those in the training data, and extrinsic hallucinations, which output facts not present in the training data. The search term generation unit contributes to suppressing extrinsic hallucinations by generating appropriate search terms, thereby encouraging the referencing of academic literature present in the training data. However, referencing academic literature using large-scale language models presents the challenge of intrinsic hallucinations.

[0033] This invention utilizes a RAG that has been pre-trained based on academic literature, allowing the large-scale language model to obtain the descriptions related to academic literature from outside the large-scale language model, instead of generating the descriptions related to academic literature itself.

[0034] By acquiring the descriptions themselves from academic literature, it is prevented that facts different from the training data (content of academic literature) will be output. Therefore, this invention suppresses not only extrinsic hallucinations but also intrinsic hallucinations, and can appropriately reflect a large amount of up-to-date knowledge in the results of fortune-telling.

[0035] Therefore, this invention can provide a technical means to encourage users to input further user attributes by providing them with results such as fortune-telling that reflect a large amount of up-to-date knowledge.

[0036] The invention relating to the sixth feature of the present invention relates to one or more features selected from the first, second, third, fourth, or fifth features, and further comprises an attribute management unit that manages the user attributes in association with their scope of publication, thereby providing a generation device.

[0037] By acquiring a large number of user attributes, it is possible to appropriately narrow down the user profile. However, acquiring a large number of user attributes can raise concerns regarding the protection of personal information. Such concerns may cause users to hesitate in entering those attributes.

[0038] This invention appropriately manages personal information by having an attribute management unit manage user attributes in association with their disclosure scope. This eliminates user concerns regarding personal information. Furthermore, according to this invention, users can control the disclosure scope themselves, for example, disclosing only a pseudonym and a part of their profile to acquaintances only online, and disclosing their date of birth or address to close friends, thus managing personal information according to the depth of their relationship. Such personal information management encourages users to input more user attributes.

[0039] Therefore, this invention can provide a technical means to encourage users to input further user attributes by providing them with results such as fortune-telling that reflect a large amount of up-to-date knowledge. [Effects of the Invention]

[0040] Based on the above, the present invention provides a technical means that can prompt users to input further user attributes by providing them with a large amount of results from fortune-telling and other sources that reflect the latest knowledge. [Brief explanation of the drawing]

[0041] [Figure 1] Figure 1 is a block diagram showing an example of the hardware and software configuration of system S in this embodiment. [Figure 2] Figure 2 shows an example of attribute database 131. [Figure 3] Figure 3 is a main flowchart showing an example of a preferred flow of the generation process in this embodiment. [Figure 4] Figure 4 is a continuation of the previous figure. [Figure 5] Figure 5 is a continuation of the previous figure. [Figure 6]Figure 6 is a continuation of the previous figure. [Modes for carrying out the invention]

[0042] Firstly, although the following disclosures, figures, and / or claims are described either individually or in combination with one or more other aspects, the subject matter of the immediate disclosure is not intended to be limited in that way. That is, the immediate disclosures, figures, and claims are intended to encompass the various aspects described herein, either individually or in one or more combinations with each other. For example, even if the immediate disclosure describes and illustrates the first, second, and third embodiments in such a way that the first embodiment is described and illustrated particularly in relation to the second embodiment, or the second embodiment is described and illustrated only in relation to the third embodiment, the immediate disclosures and illustrations are not limited in that way and may include only the first embodiment, only the second embodiment, only the third embodiment, or one or more combinations of the first, second, and / or third embodiments, such as the first and second embodiments, the first and third embodiments, the second and third embodiments, or the first, second, and third embodiments.

[0043] In this text, the phrase "or" is used to mean a "non-exclusive" arrangement unless explicitly specified otherwise. For example, when we say "item x is A or B," it means either (1) item x is either A or B, or (2) item x is both A and B. In other words, the word "or" is not used to define an "exclusive" arrangement.

[0044] Furthermore, when the phrases "contain at least one" or "contain at least one of the following" are used in the text, they mean that the system or element contains one or more of the elements listed after the phrase. For example, if there are three types of elements, from element 1 to element 3, the phrases "contain at least one" or "contain at least one of the following" are interpreted as any of the following structural arrangements: a device containing element 1, a device containing element 2, a device containing element 3, a device containing element 1 and element 2, a device containing element 1 and element 3, a device containing element 2 and element 3, or a device containing element 1, element 2, and element 3.

[0045] The same interpretation is intended when the phrase "used in at least one of the following" is used in the text. Furthermore, "and / or" as used in the text is used as a linguistic conjunction to indicate that one or more of the listed elements or conditions are included or occur. For example, a device containing the first element, the second element, and / or the third element is interpreted as any of the following structural arrangements: a device containing the first element, a device containing the second element, a device containing the third element, a device containing the first and second elements, a device containing the first and third elements, a device containing the second and third elements, or a device containing the first, second, and third elements.

[0046] Furthermore, the use of the phrase "and / or" in this text signifies a "non-exclusive" arrangement, as stipulated in the Japanese Industrial Standard (JIS) "Format and Preparation Method of Standards Documents JIS Z 8301".

[0047] The following describes in detail an example of an embodiment of the present invention with reference to the drawings.

[0048] <System S> Figure 1 is a block diagram showing an example of the hardware and software configuration of System S according to this embodiment. The fortune-telling text generation system (System S) according to this embodiment is configured to include a fortune-telling text generation device 1. In the following, the fortune-telling text generation device 1 will also be simply referred to as "generation device 1".

[0049] In this embodiment, System S, in the generation device 1, estimates unentered attribute values ​​from entered attribute values, and searches the large amount of up-to-date knowledge stored in the academic literature database D using search words generated based on these attribute values. Then, in the generation device 1, System S generates fortune-telling text based on the search results using a large-scale language model and provides it to Terminal T.

[0050] This allows System S to encourage the input of additional user attributes through compelling fortune-telling texts that reflect a large amount of up-to-date knowledge, and contributes to the realization of an application that more appropriately refines the user's profile based on the additional attribute values ​​and estimated unentered attribute values.

[0051] [Fortune-telling and fortune-telling texts] In this embodiment, "fortune-telling, etc." (or simply "fortune-telling") is not limited to anything titled "fortune-telling," but refers to all applications that refer to the user's personality, as exemplified by fortune-telling, personality tests, psychological tests, and personality assessments. Similarly, in this embodiment, "fortune-telling text" is not limited to text titled "fortune-telling," but refers to all types of text that include descriptions of the user's personality, as exemplified by fortune-telling results, personality test results, psychological test results, and personality assessment results.

[0052] [Generation device 1] The fortune-telling text generation device 1 comprises a control unit 11, a storage unit 13, and a communication unit 14, etc. The type of generation device 1 is not particularly limited and may be, for example, a server device, a cloud server, etc. The generation device 1 is configured to communicate with the academic literature database D and terminal T via the network N.

[0053] [Control Unit 11] The control unit 11 includes a Central Processing Unit (CPU), Random Access Memory (RAM), and Read Only Memory (ROM), among other things.

[0054] The control unit 11 cooperates with at least one of the storage unit 13 and the communication unit 14 as needed. The control unit 11 then implements the software components of the program of this embodiment executed by the generation device 1, such as the attribute management unit 111, attribute acquisition unit 112, input reflection unit 113, machine learning unit 114, attribute estimation unit 115, estimation reflection unit 116, search word generation unit 117, fortune-telling text generation unit 118, display command unit 119, etc.

[0055] [Storage section 13] The storage unit 13 is a device on which data and / or files are stored, and has a storage unit that stores data non-temporarily using a hard disk, semiconductor memory, recording medium, and memory card, etc. The storage unit 13 stores programs executed by a microcomputer, an attribute database 131, data for actively partitioning user attributes, data for estimating unentered attribute values, data for generating fortune-telling text etc. using a large-scale language model, a bibliographic database, etc.

[0056] (Attribute Database 131) The attribute database 131 stores attribute datasets related to the user's profile. Each attribute dataset contains attribute values ​​belonging to a predetermined set of user attributes. The attribute datasets and attribute values ​​stored in the attribute database 131 are managed by the attribute management unit 111. Each attribute value belonging to the attribute dataset may be an attribute value entered by the user, or it may be an estimated result of an unentered attribute value that has not been entered by the user.

[0057] Each attribute value belonging to the attribute dataset is stored in a format that distinguishes between unentered and entered values. Unentered attribute values ​​are further stored in a format that distinguishes between unestimated attribute values ​​and estimated attribute values. In other words, each attribute value belonging to the attribute dataset is managed by the attribute management unit 111 in a format that distinguishes between entered attribute values ​​and estimated attribute values. As a result, the attribute estimation unit 115 of this embodiment can generate estimation results for unentered attribute values ​​based only on entered attribute values. That is, this configuration suppresses the inclusion of errors in the estimation results due to the estimation results being fed back into the estimation results.

[0058] Furthermore, this configuration leaves room for the fortune-telling text generation unit 118 to generate fortune-telling text using only the knowledge retrieved by the input attribute values ​​or search words based on the input attribute values, and without using information related to the estimation results. As a result, the generation device 1 can provide fortune-telling text that does not utilize such estimation results when the reliability of the estimation results is low. In this case, it is preferable that the estimation results are stored in association with the accuracy of the estimation so that processing can be performed according to the reliability.

[0059] Incidentally, as mentioned above, the attribute management unit 111 manages attribute datasets that include estimated results for unentered attribute values ​​and entered attribute values. Therefore, users may have a stronger concern about such attribute datasets, which provide a more detailed profile of a person, being leaked as personal information than they would with attribute datasets that do not include estimated results.

[0060] Therefore, it is preferable for the attribute management unit 111, which manages the attribute database 131, to manage user attributes in association with their scope of public access. In order to individually control the scope of public access for user attributes, it is preferable for the attribute management unit 111 to manage some or all of each user attribute belonging to the attribute dataset in association with a scope of public access that is independent of the scope of public access for other user attributes.

[0061] The attribute management unit 111 manages personal information appropriately by associating user attributes with their disclosure scope. This eliminates user concerns regarding personal information. Furthermore, this configuration allows users to control the disclosure scope themselves, for example, disclosing only a pseudonym and a portion of their profile to acquaintances only online, and disclosing their date of birth or address to close friends, thus managing personal information according to the depth of their relationship. This type of personal information management encourages users to input more user attributes.

[0062] Furthermore, with this kind of personal information management in place, terminal T can execute a program that, in contacts via long-distance communication such as internet communication, only discloses certain user attributes or the person profile related to the fortune-telling text generation unit 118 that are to be shared with such acquaintances, and in face-to-face contacts, discloses other user attributes that are to be shared only with acquaintances that have actually met. This provides users with an incentive to input user attributes that they do not want to share in contacts via long-distance communication but do want to share in face-to-face contacts.

[0063] Furthermore, regarding the above-mentioned concerns, it is preferable that the attribute dataset be managed by the attribute management unit 111 in the form of an anonymous dataset that does not contain personal information that can identify users.

[0064] As described above, the attribute management unit 111 manages the attribute dataset in an anonymous dataset format that does not contain personal information that can identify users, thereby preventing the attribute dataset from being leaked as personal information. Therefore, this configuration manages attribute datasets that allow for a more detailed identification of the user's profile, while eliminating concerns about the leakage of personal information and further meeting the standards required by laws and regulations regarding personal information. As a result, this configuration contributes to encouraging the input of further user attributes.

[0065] Furthermore, this configuration prevents users from hesitating to disclose the fortune-telling text generated by the fortune-telling text generation unit 118 to the public due to the inclusion of personal information. When disclosing fortune-telling text containing personal information, users may hesitate to do so due to concerns about the leakage of such personal information. Disclosure to the public includes, for example, disclosure via social media. Therefore, this configuration contributes to the widespread adoption of the generation device 1 by encouraging the disclosure of fortune-telling text via social media, etc.

[0066] The "predefined set of attributes" relating to attribute values ​​belonging to the attribute dataset are not particularly limited. For example, the attribute dataset includes attribute values ​​belonging to 10 predetermined attributes, such as gender, height, blood type, age, personality, date of birth, occupation, dominant hand, area of ​​origin, and area of ​​workplace. To appropriately narrow down the user profile, it is preferable that the attribute dataset includes multiple attributes related to psychological diagnoses, such as those exemplified by the results of a questionnaire related to the Myers-Briggs Type Indicator (MBTI). To appropriately narrow down the user profile from a commercial use perspective, it is preferable that the attribute dataset includes multiple attributes related to targeted advertising.

[0067] To achieve the management of the anonymous dataset described above, it is preferable that the multiple attributes do not include any attributes that constitute personal information that can identify a user on their own. Attributes that constitute personal information include, for example, name, account name, membership number, email address, telephone number, and exact address.

[0068] Attribute datasets should preferably satisfy k-anonymity. This ensures that even if an attribute dataset is leaked externally, there will always be at least k users with similar attribute datasets, thus protecting the privacy of each user.

[0069] Figure 2 shows an example of attribute database 131. This example stores attribute datasets containing attribute values ​​belonging to 10 predetermined attributes. The 10 predetermined attributes are gender, height, blood type, age, personality, date of birth, occupation field, dominant hand, area of ​​origin (prefecture), and area of ​​work (prefecture). In this example, for explanatory purposes, each attribute dataset is assigned an attribute ID to identify that attribute dataset, but please note that in attribute database 131, information to identify an attribute dataset is not necessarily attached to the attribute dataset.

[0070] The attribute dataset for attribute ID "A0001" in this example includes the following attribute values: Gender "Male", Height "Not entered, estimated 165-169cm", Blood type "Type O", Age "27 years old", Personality "Not entered, estimated introverted", Date of birth "October 30th", Occupation "IT-related", Dominant hand "Not entered, estimated right-handed", Place of origin "Not entered, estimated born in Tokyo", and Workplace "Works in Tokyo".

[0071] This attribute dataset does not contain personal information such as names or email addresses. Therefore, the personal information and privacy of users corresponding to this attribute dataset are protected. For the unentered attribute "height," this attribute dataset stores an estimated result of "estimated 165-169cm" based on the entered attribute values ​​"male," "blood type O," "27 years old," "born October 30th," "IT-related," and "works in Tokyo." Similarly, for the unentered attributes "personality," "dominant hand," and "place of origin," this attribute dataset stores estimated results of "estimated introverted," "estimated right-handed," and "estimated born in Tokyo," respectively, based on the entered attribute values.

[0072] By storing the estimation results, the attribute dataset now presents a more accurately defined profile of the user than if the estimation results were not stored. As a result, various applications, such as the generation of fortune-telling texts and targeted advertising on social media, will run more effectively than when the estimation results are not stored.

[0073] The attribute dataset for attribute ID "A0002" in this example includes the following attribute values: gender "male", height "not entered", blood type "O", age "27", personality "not entered", date of birth "October 30", occupation "IT-related", dominant hand "not entered", place of origin "not entered", and workplace "works in Tokyo".

[0074] This attribute dataset is similar to the attribute dataset with attribute ID "A0001," but it does not store the estimation results. For example, in the generation process described later, such an attribute dataset may be created after the input attributes acquired by the attribute acquisition unit 112 have been reflected in the attribute database 131 by the input reflection unit 113, but before the estimation by the attribute estimation unit 115 has been performed, or before the estimation has been performed but the estimation results have been reflected in the attribute database 131 by the estimation reflection unit 116.

[0075] The attribute dataset for attribute ID "A0003" in this example includes the following attribute values: gender "male", height "168cm", blood type "O", age "27", personality "introverted", date of birth "October 30", occupation "IT-related", dominant hand "left-handed", place of origin "born in Hyogo Prefecture", and workplace "works in Tokyo".

[0076] This attribute dataset is similar to the attribute dataset with attribute ID "A0001," but all attribute values ​​are entered. For example, if a user who has viewed the fortune-telling text generated by the fortune-telling text generation unit 118 enters further user attributes, such an attribute dataset may be created.

[0077] By storing these attribute datasets in the example, the generator 1 can, for example, provide a user associated with gender ID "A0001" with fortune-telling text based on findings obtained from academic literature searched using search terms such as "male in his 20s" and "desk work" based on the entered attribute values, and search terms such as "introverted" and "born in the city" based on the estimation results.

[0078] (Data for actively segmenting user attributes) The memory unit 13 stores data for actively segmenting user attributes. User attributes may include the estimated results of entered attribute values ​​and unentered attribute values ​​mentioned above, as well as various data entered by the user being analyzed, such as data indicating the user's concerns.

[0079] Data for actively segmenting user attributes includes, for example, data from a conversion table that associates attribute values ​​to be converted with converted search terms, and data from a machine learning model that outputs converted search terms when an attribute value to be converted is input. Machine learning model data is, for example, data from a neural network.

[0080] In this case, in order to improve the conversion accuracy, it is preferable that the machine learning model is pre-trained using training data that includes the attribute values ​​to be converted as explanatory variables and the converted search words as the target variable. The pre-training procedure is not particularly limited and may include, for example, deep learning. In order to further improve the conversion accuracy through supervised learning, it is preferable that the above-mentioned training data includes the attribute values ​​to be converted as explanatory variables and the search words labeled to indicate that the conversion is appropriate or inappropriate as the target variable.

[0081] (Data for estimating unentered attribute values) The memory unit 13 stores data for estimating unentered attribute values. This data is not particularly limited as long as it shows the correlation between multiple attribute values ​​belonging to the attribute dataset. Examples of data for estimating unentered attribute values ​​include data from a conversion table that associates a vector having one or more entered attribute values ​​with the estimated results of unentered attribute values, and a machine learning model configured to output the estimated results of unentered attribute values ​​when entered attribute values ​​are input.

[0082] While not mandatory, it is preferable that the data for estimating unentered attribute values ​​includes a machine learning model configured to output estimated results for unentered attribute values ​​when entered attribute values ​​are provided. This machine learning model is pre-trained using training data that includes entered attribute values ​​but does not include estimated results for unentered attribute values. The data related to the machine learning model is, for example, neural network data.

[0083] As a result, the attribute estimation unit 115 can utilize such a pre-trained machine learning model in the attribute estimation step described later. Therefore, the attribute estimation unit 115 can generate estimation results that reflect the complex correlations between many attribute values. Furthermore, in order to improve estimation accuracy through supervised learning, it is preferable that the above-mentioned training data includes the input attribute values ​​as explanatory variables and, as the objective variable, estimation results labeled to indicate appropriate estimation or estimation results labeled to indicate inappropriate estimation.

[0084] (Large-scale language models) The large-scale language model of this embodiment is not particularly limited. Preferably, the large-scale language model is configured to generate text that reflects the knowledge related to the academic literature referenced by the search words or search phrases included in the prompt. The large-scale language model may refer to the academic literature using the search words or search phrases included in the prompt as they are, or it may refer to the academic literature after performing appropriate processing, such as machine translating it into a language corresponding to the academic literature database D.

[0085] Examples of large-scale language models used in this embodiment include GPT-4, GPT-4o, and Perplexity®, which are used in ChatGPT®. To provide more persuasive fortune-telling texts, it is preferable that the large-scale language model of this embodiment undergoes fine-tuning or pre-training related to the generation of fortune-telling texts. Such pre-training is achieved, for example, by training data that includes fortune-telling texts. To prevent the generation of expressions that are inappropriate as fortune-telling, it is preferable that the training data is supervised learning data that includes fortune-telling texts labeled as desirable and fortune-telling texts labeled as undesirable.

[0086] (Literature Database) The literature database stores scholarly literature published in the external scholarly literature database D. The literature database is configured to use Retrieval-Augmented Generation (RAG) to allow scholarly literature or its contents corresponding to search terms to be retrieved from a large-scale language model.

[0087] By storing such a literature database in the memory unit 13, the large-scale language model can use RAG to retrieve the descriptions of academic literature from the literature database, which is an external resource to the large-scale language model, instead of generating the descriptions of academic literature itself. This prevents the output of incorrect information that differs from the training data (content of academic literature) from the large-scale language model.

[0088] The bibliographic database may be stored in the storage unit 13 of the generation device 1, or it may be stored in a server configured separately from the generation device 1. The separate configuration contributes to efficient retrieval of academic literature by a server with hardware configuration specialized for database processing. The separate server is configured to communicate with the generation device 1.

[0089] To ensure efficient referencing, RAG prefers to utilize an embedding model. An embedding model transforms high-dimensional text data into a more manageable, lower-dimensional format by converting words or phrases into fixed-length numerical vectors. This allows RAG to reference scholarly literature more effectively and quickly than by using high-dimensional information directly as queries.

[0090] The embedding model is preferably one that reflects the semantic relationships between words or phrases. Such an embedding model allows RAG to perform more accurate references that reflect semantic relationships. For example, such an embedding model would transform words with semantic relationships, such as "young people," "children," and "minors," into similar numerical vectors.

[0091] To ensure efficient referencing, it is preferable that the RAG undergoes pre-training based on academic literature. Such pre-training can be achieved, for example, by creating an embedding model based on several initially stored academic literature articles and updating the embedding model based on one or more additional academic literature articles.

[0092] In this process, RAG, for example, converts search terms handed over from a large-scale language model into indexes in a literature database using an embedding model, retrieves academic literature or descriptions corresponding to the search terms from the literature database using these indexes, converts them into human-readable language, and hands them over to the large-scale language model, thereby enabling the large-scale language model to access the academic literature or descriptions.

[0093] [Communications Section 14] The communication unit 14 is not particularly limited as long as it is used to connect the generating device 1 to the network N and perform communication. Examples of the communication unit 14 include a network card compatible with the Ethernet standard and a communication device compatible with wireless LAN.

[0094] [Academic Literature Database D] The academic literature database D is not particularly limited as long as it is publicly available outside of the generation device 1. The method of publication is not particularly limited as long as it is accessible from the generation device 1, and examples include free publication and member-only publication. The number of academic literature databases according to this embodiment is not particularly limited. In order to generate fortune-telling texts that reflect a larger amount of knowledge, it is preferable that the number of academic literature databases D be two or more.

[0095] The academic literature database D is accessed from the generating device 1 via an API or the like, and searches for recorded academic literature based on a given search request, and transmits the search results to the generating device 1. The search request has a search phrase that includes at least one search word. The search words for the academic literature database D may be keywords placed at the beginning of the academic literature, or free words related to the abstract or main text of the academic literature. The search request may also include one or more parameters such as the publication date of the academic literature to be searched, the journal or international conference to be searched, and the academic field to be searched.

[0096] [Network N] The type of network N is not particularly limited, as long as it is used for communication by the generating device 1, etc. Examples of network N include the internet, mobile phone networks, wireless LANs, etc.

[0097] [Terminal T] Terminal T is used by users who utilize the fortune-telling service. Terminal T performs processes such as transmitting user information to the generation device 1 based on user input, and displaying fortune-telling text transmitted from the generation device 1. The types of terminal T include, for example, mobile terminals such as smartphones, tablet terminals, and laptop computers, and stationary terminals such as desktop computers.

[0098] [Main flowchart of the generation process] Figure 3 is a main flowchart showing an example of a preferred flow of the generation process in this embodiment. Figure 4 is a continuation from the previous figure. Figure 5 is a continuation from the previous figure. Figure 6 is a continuation from the previous figure. The following is a description of an example of a preferred flow in the generation process of this embodiment, using Figures 3 to 6.

[0099] The generation process includes a series of operations to retrieve attribute values ​​that belong to the attribute dataset and have been entered by the user, and to reflect these entered attribute values ​​in the attribute dataset. Steps S1 to S3 are an example of this process. This allows the attribute management unit 111 to manage the attribute values ​​entered by the user.

[0100] [Step S1: Determine whether to retrieve user attributes] The control unit 11 works in cooperation with the storage unit 13 and the communication unit 14 to execute the attribute acquisition unit 112. The control unit 11 then performs a process to determine whether to acquire user attributes using the attribute acquisition unit 112 (attribute acquisition determination step). If it determines that to acquire the attributes, the control unit 11 moves the process to step S2; otherwise, the control unit 11 moves the process to step S9.

[0101] In the attribute acquisition determination step, the control unit 11 determines, for example, that user attributes should be acquired when attribute values ​​related to user attributes entered at terminal T are received by the generation device 1.

[0102] [Step S2: Get entered attribute values] The control unit 11 executes a process to acquire attribute values ​​that belong to the attribute dataset related to the user's profile and that have been entered by the user, using the attribute acquisition unit 112 (attribute acquisition step). That is, the attribute acquisition unit 112 acquires one or more user attributes related to the user. The control unit 11 then moves the process to step S3.

[0103] [Step S3: Reflect the entered attribute values ​​in the attribute dataset] The control unit 11 works in cooperation with the storage unit 13 and the communication unit 14 to execute the input reflection unit 113. Then, the control unit 11 performs a process to reflect the acquired input attribute values ​​in the attribute dataset managed by the attribute management unit 111 using the input reflection unit 113 (input reflection step). The control unit 11 then moves the process to step S4.

[0104] In the input reflection step, the input reflection unit 113 reflects the acquired input attribute values ​​in the attribute dataset by procedures such as storing the input attribute values ​​in the attribute database 131. At this time, input attribute values ​​entered by the same user in a series of sessions are linked to each other by being stored in the attribute database 131 as belonging to a single attribute dataset related to that user. The input reflection unit 113 reflects the above-mentioned input attribute values ​​in the attribute dataset in a format that allows for the distinction between the input attribute values ​​and the estimated attribute values.

[0105] [Anonymization Step] To protect user privacy and make it difficult to identify personal information, the generation process preferably includes an anonymization step to ensure that the attribute dataset in the attribute database 131 meets the k-anonymity requirement if it does not. The anonymization step is implemented, for example, by a process that includes a procedure to merge any entered attribute value with a similar value. The value of k is specified by the administrator of the generation device 1, etc., depending on the sensitivity of the data or the privacy requirements of the system.

[0106] The following is an example of the procedure described above, where the attribute datasets for users 1 through 4 have genders and heights of "male, 169cm," "male, 163cm," "male, 173cm," and "male, 175cm," respectively, and the attribute datasets satisfy the 2-anonymity requirement for gender and height. In this case, a third party who knows that the subject is "male, 163cm" can uniquely identify the subject as user 2 based on the combination of gender and height.

[0107] Therefore, the generator 1 performs a procedure to merge the height attribute value among the input attribute values ​​with similar values ​​to create "160cm range" and "170cm range". This procedure creates attribute datasets for users 1 to 4 as "male, height 160cm range", "male, height 160cm range", "male, height 170cm range", and "male, height 170cm range", respectively. As a result, a third party who knows that the subject is "male, height 163cm" will not be able to determine whether the subject is user 1 or user 2. Thus, this procedure ensures that the attribute dataset satisfies the requirement of 2-anonymity.

[0108] Although not a mandatory configuration, the generation process preferably includes a series of steps to pre-train the machine learning model used by the attribute estimation unit 115 (described later) using training data that includes input attribute values ​​but does not include the estimation results for uninput attribute values. Steps S4 to S5 are an example of such a process. This makes it possible to improve the estimation accuracy of the machine learning model during the operation of the generation device 1.

[0109] [Step S4: Determine whether to perform machine learning] The control unit 11 works in cooperation with the storage unit 13 and the communication unit 14 to execute the machine learning unit 114. The control unit 11 then performs a process to determine whether to perform machine learning on a machine learning model that outputs estimation results for uninputted attribute values ​​using the machine learning unit 114 (machine learning determination step). If it determines that to perform machine learning, the control unit 11 moves the process to step S5; otherwise, the control unit 11 moves the process to step S6.

[0110] In the machine learning discrimination step, the machine learning unit 114 determines whether to perform machine learning, for example, when the amount of input attribute values ​​reflected in the attribute dataset after the previous machine learning session exceeds a given threshold.

[0111] [Step S5: Perform machine learning] The control unit 11 executes a machine learning process related to a machine learning model that outputs estimated results for uninputted attribute values, using training data that includes input attribute values ​​but does not include estimated results for uninputted attribute values ​​(machine learning step). The control unit 11 then moves the process to step S6.

[0112] In the machine learning step, the machine learning unit 114 uses, for example, data obtained by removing the estimated results of unentered attribute values ​​from an attribute dataset that reflects the input attribute values ​​after the previous machine learning step, as training data.

[0113] After the entered attribute values ​​are reflected in the attribute dataset, the generation device 1 generates estimation results based on the entered attribute values ​​for unentered attribute values ​​that belong to the attribute dataset and for which the user has not entered an attribute value, and performs a series of processes to reflect the estimation results in the attribute dataset. Steps S6 to S8 are an example of this process. As a result, the attribute management unit 111 can manage not only the entered attribute values ​​entered by the user, but also the estimation results for unentered attribute values.

[0114] [Step S6: Determine whether to estimate the unentered attribute value] The control unit 11 works in cooperation with the storage unit 13 and the communication unit 14 to execute the attribute estimation unit 115. The control unit 11 then performs a process to determine whether to estimate or determine an uninputted attribute value using the attribute estimation unit 115 (attribute estimation determination step). If it determines that estimation is necessary, the control unit 11 moves the process to step S7; otherwise, the control unit 11 moves the process to step S9.

[0115] In the attribute estimation and determination step, the attribute estimation unit 115 determines, for example, that it should estimate unentered attribute values ​​if there are unentered attribute values ​​in the attribute dataset to which entered attribute values ​​have been reflected since the previous estimation.

[0116] (Regarding autocomplete of attribute values) In order to assist the user in inputting attribute values, in the attribute estimation and determination step, it is preferable that the attribute estimation unit 115 determines whether to estimate uninputted attribute values ​​if the input attribute values ​​on terminal T satisfy the given conditions related to autocomplete.

[0117] The "given conditions" referred to here are conditions related to the timing at which the user does not feel distrust regarding personal information management and feels convenience to the user. These conditions are exemplified by conditions such as the input of a given number of pre-filled attribute values ​​required for autocomplete, or the input of a given combination of pre-filled attribute values ​​required for autocomplete. The given number is not particularly limited and may be, for example, 5 or 10. The given combination is not particularly limited and may be, for example, a combination consisting of all attribute values ​​on the same page of the attribute value input user interface.

[0118] At this time, the attribute estimation unit 115 transmits the unentered attribute values ​​estimated in the attribute estimation step described later to the terminal T. This allows the terminal T to use the estimated unentered attribute values ​​as the initial values ​​for the input fields, thereby assisting the user in entering attribute values.

[0119] If the fortune-telling text generated in the fortune-telling text generation step described later is text related to fortune-telling or the like (for example, MBTI diagnosis) that requires a large number of attribute values, the process of the user entering these attribute values ​​becomes cumbersome. Therefore, the above-mentioned support is particularly effective when such fortune-telling text is generated.

[0120] [Step S7: Generate estimated results for unentered attribute values] The control unit 11 uses the attribute estimation unit 115 to perform a process to estimate uninputted attribute values ​​based on the input attribute values ​​(attribute estimation step). The control unit 11 then moves the process to step S8. At this time, the attribute estimation unit 115 generates estimation results based on the correlation between multiple attribute values ​​belonging to the attribute dataset.

[0121] To perform estimation based on input attribute values, it is preferable that the attribute estimation unit 115 performs estimation based on the correlation between input attribute values ​​and the input attribute values. To perform estimation based on externally collected data, it is preferable that the attribute estimation unit 115 performs estimation based on the correlation between attribute values ​​derived from said data and the input attribute values. Externally collected data is, for example, statistical data collected by various organizations such as government agencies. As a result, the generation device 1 can estimate uninputted attribute values ​​based on input attribute values, based on the correlation between user attributes derived from the statistical data.

[0122] In order to generate estimation results that reflect the complex correlations between many attribute values, it is preferable that the attribute estimation unit 115 performs the above estimation by a process that involves inputting the input attribute values ​​into a machine learning model that has been pre-trained using training data that includes the input attribute values ​​but does not include the estimation results for the uninput attribute values, and then outputting the estimation results for the uninput attribute values.

[0123] [Step S8: Reflect the estimation results in the attribute dataset] The control unit 11 works in cooperation with the storage unit 13 and the communication unit 14 to execute the estimation reflection unit 116. Then, the control unit 11 performs a process to reflect the above estimation results in the attribute dataset using the estimation reflection unit 116 (estimation reflection step). The control unit 11 then moves the process to step S9.

[0124] In the estimation and reflection step, the estimation and reflection unit 116 reflects the estimation results by storing the estimation results of unentered attribute values ​​in the attribute database 131, etc. The estimation and reflection unit 116 reflects the estimation results in the attribute dataset in a format that allows for the distinction between the entered attribute values ​​and the estimated attribute values.

[0125] After the attribute values ​​entered are reflected in the attribute dataset, the generation device 1 generates fortune-telling text based on the search results from the academic literature database D and performs a series of processes to display it on a terminal T or the like. Steps S9 to S15 are an example of this process. As a result, the generation device 1 can provide the user with fortune-telling results that reflect a large amount of up-to-date knowledge and encourage them to input further user attributes.

[0126] [Step S9: Determine whether to generate fortune-telling text] The control unit 11, in cooperation with the memory unit 13 and the communication unit 14, performs a process to determine whether to generate a fortune-telling text (fortune-telling text generation determination step). If it determines that a text should be generated, the control unit 11 moves the process to step S10; otherwise, the control unit 11 returns the process to step S1 and repeats the process from step S1 to step S18.

[0127] In the attribute acquisition determination step, the control unit 11 determines to generate fortune-telling text, for example, when the generation device 1 receives an instruction to generate fortune-telling text input at terminal T.

[0128] In order to generate fortune-telling text that answers the user's concerns, the generation process preferably includes a series of processes that generate search terms from the concerns entered by the user. Steps S10 to S11 are an example of such processes.

[0129] [Step S10: Determine if a problem has been entered] The control unit 11 works in cooperation with the storage unit 13 and the communication unit 14 to execute the attribute acquisition unit 112. The control unit 11 then uses the attribute acquisition unit 112 to determine whether a problem has been entered by the user (problem input determination step). If it is determined that a problem has been entered, the control unit 11 moves the process to step S11; otherwise, the control unit 11 moves the process to step S12.

[0130] In the problem input determination step, the attribute acquisition unit 112 determines that a problem has been entered by a user, for example, when data related to a problem entered at terminal T is received by the generation device 1, and acquires user attributes indicating the user's problem.

[0131] [Step S11: Generate search terms from your problem] The control unit 11 works in cooperation with the storage unit 13 and the communication unit 14 to execute the search word generation unit 117. Then, the control unit 11 uses the search word generation unit 117 to perform the process of generating problem search words from the problem (problem search word generation step). The control unit 11 then moves the process to step S12.

[0132] In the problem search term generation step, it is preferable that the search term generation unit 117 generates search terms by a process that includes a procedure for generating problem search terms for searching the academic literature database D based on the problem.

[0133] In order to generate search terms that take user attributes into account, the procedure for generating search terms for problems by the search term generation unit 117 preferably generates search terms for problems based on a combination of one or more user attributes and problems. As an example of such a conversion, for example, if the problem is "love (marriage)" and the user attribute "female" is reflected in the attribute dataset as an input attribute value, a conversion can be made to generate search terms for problems that correspond to women's worries about love or marriage, such as fear of career changes, dealing with spouse's relatives, and childbirth.

[0134] [Step S12: Generate search words from attribute values] The control unit 11 works in cooperation with the storage unit 13 and the communication unit 14 to execute the search word generation unit 117. The control unit 11 then uses the search word generation unit 117 to generate search words from attribute values ​​(search word generation step). In other words, in the search word generation step, the search word generation unit 117 generates search words used to refer to academic literature published in the external academic literature database D, based on user attributes. The control unit 11 then moves the process to step S13.

[0135] In the search term generation step, it is preferable that the search term generation unit 117 generates search terms by a process that includes a procedure for performing active segmentation to convert user attributes into search terms for referencing academic literature. The user attributes referred to here may be either the input attribute values ​​or the estimated results of the uninputted attribute values. Active segmentation referred to here means segmenting user attributes using criteria suitable for referencing academic literature, instead of segmenting user attributes using well-known criteria related to fortune-telling.

[0136] The following is an example of active segmentation by the search word generation unit 117. This example does not limit this embodiment to segments that are exemplified. Furthermore, the active segmentation by the search word generation unit 117 is not limited to those exemplified.

[0137] Dividing age not by the Chinese zodiac sign such as "Year of the Dragon," but by categories such as "born in the 1970s," "born in the 1980s," "born in the 1990s," "born in the 2000s," or by categories such as "baby boomers," "bubble generation," "job-hunting ice age generation," "yutori generation," "Generation Z," etc. Dividing birth dates not by ruling zodiac signs such as "Aries," but by categories such as "born in spring," "born in summer," etc. Dividing birth dates by categories such as "born early," "born late," etc., or by categories such as "born in April," "born in May," etc. Dividing place of origin by categories such as "born by the sea," "born inland," etc., or by categories such as "born in the countryside," "born in the city," etc.

[0138] Preferably, the search term generation unit 117 can generate search terms based on combinations of multiple user attributes. As an example of such generation, for instance, the search terms "migrant worker" and "immigrant worker" can be generated from the combination of user attributes "born in Hyogo Prefecture" and "works in Tokyo." As a result, the generation device 1 can provide a more persuasive fortune-telling text than a fortune-telling text based solely on the insights gained from a search term based on a single user attribute, by adding insights gained from search terms based on combinations of multiple user attributes.

[0139] The generation process preferably includes a step (S13) of passively segmenting user attributes based on criteria related to fortune-telling. This allows the fortune-telling text generation unit 118 to generate fortune-telling text that includes descriptions that correlate the results of the passive segmentation based on the criteria related to fortune-telling with findings obtained from academic literature.

[0140] [Step S13: Passively split the entered attribute values] The control unit 11 works in cooperation with the storage unit 13 and the communication unit 14 to execute the fortune-telling text generation unit 118. The control unit 11 then performs a process of passively dividing the input attribute values ​​using the fortune-telling text generation unit 118 (passive division step). That is, in the passive division step, the fortune-telling text generation unit 118 passively divides the user attributes according to the criteria related to fortune-telling. The control unit 11 then moves the process to step S14.

[0141] Examples of passive segmentation of user attributes based on divination criteria include dividing age into the twelve zodiac signs or the ten heavenly stems and twelve earthly branches, or dividing the date of birth into the ruling constellation of the twelve zodiac signs. This allows the generator 1 to provide users with compelling divination texts that reflect a large amount of up-to-date knowledge without compromising the mystique of divination exemplified by zodiac divination, Four Pillars of Destiny, and ruling constellation divination.

[0142] [Step S14: Generate fortune-telling text] The control unit 11 works in cooperation with the storage unit 13 and the communication unit 14 to execute the fortune-telling text generation unit 118. The control unit 11 then uses the fortune-telling text generation unit 118 to generate fortune-telling text related to the user based on the attribute dataset described above (fortune-telling text generation step). The control unit 11 then moves the process to step S15.

[0143] In order to generate fortune-telling text that reflects the estimation results of a wide variety of input and uninputted attribute values ​​included in the attribute dataset, it is preferable that the fortune-telling text generation unit 118 generates the above-mentioned fortune-telling text by a process that includes inputting a prompt to a large-scale language model that includes an instruction to generate fortune-telling text for a user whose profile is represented by these attribute values, based on the estimation results of input and uninputted attribute values ​​related to the attribute dataset.

[0144] If the search word generation step is being executed, in the fortune-telling text generation step, it is preferable that the fortune-telling text generation unit 118 inputs instructions to generate fortune-telling text and prompts including the generated search words into a large-scale language model that references academic literature, thereby generating fortune-telling text relevant to the user. This allows the generation device 1 to generate compelling fortune-telling text that reflects a large amount of up-to-date knowledge while suppressing hallucination through search words suitable for referencing academic literature.

[0145] In order to reference academic literature in the generation of fortune-telling texts, it is preferable that the large-scale language model involved in the search word generation step is a model that has been pre-trained using academic literature published in an external academic literature database D as training data. Pre-training may be performed using various methods to enable the large-scale language model to refer to the descriptions in the literature, such as self-supervised learning or transfer learning.

[0146] The large-scale language model involved in the search word generation step is preferably configured to refer to academic literature using a RAG that has been pre-trained on academic literature. In this case, the RAG retrieves data from the literature database based on the search words, converts this data into human-readable language, and passes it to the large-scale language model. The large-scale language model then generates fortune-telling text related to the user based on this academic literature.

[0147] The procedure for retrieving academic literature from a literature database using RAG suppresses hallucination related to large-scale language models. As a result, the generator 1 can provide users with results such as fortune-telling that more accurately reflect a large amount of up-to-date knowledge.

[0148] In a configuration that uses a pre-trained RAG to reference academic literature, the literature database referenced by the RAG is preferably implemented in the generation device 1 to avoid a complex hardware configuration, but it may also be implemented on a server or the like, separate from the generation device 1.

[0149] If the search word generation step is being executed, the fortune-telling text generation unit 118 may generate fortune-telling text relating to the user by inputting a prompt into the large-scale language model that includes an instruction to search the academic literature database D using a search phrase containing the above-mentioned search words and generate fortune-telling text.

[0150] When the problem input discrimination step and the problem search word generation step are being executed, it is preferable that the fortune-telling text generation unit 118 inputs a prompt further containing the problem search word into the large-scale language model during the fortune-telling text generation step. This allows the generation device 1 to provide the user with a persuasive fortune-telling text that reflects appropriate insights corresponding to the problem and possesses the mystical quality of fortune-telling, while prompting the input of user attributes related to the problem.

[0151] The fortune-telling text generation unit 118 can, for example, generate fortune-telling text that presents insights obtained through searches related to search terms for the user's concerns, such as age at first marriage and marriage rate, as if they were fortune-telling results related to love.

[0152] If the passive segmentation step is performed, it is preferable that the fortune-telling text generation unit 118 generates fortune-telling text in the fortune-telling text generation step that corresponds to the results of passively segmenting user attributes according to the criteria related to fortune-telling. This allows the generation device 1 to provide fortune-telling text that presents insights obtained from academic literature as if they were predictions obtained through the mystical means of fortune-telling. Therefore, the generation device 1 can provide the user with compelling fortune-telling text that reflects a large amount of up-to-date knowledge without compromising the mystical nature of fortune-telling, and can further encourage the input of additional user attributes.

[0153] The fortune-telling text generation unit 118 generates fortune-telling text that, for example, for a user born on April 1, 2000, uses insights obtained from searches using search terms such as Gen Z and early birth generated by active segmentation to present the results as if they were fortune-telling results for a person whose zodiac sign is the Dragon and whose ruling constellation is Aries.

[0154] To improve search accuracy through search-specific processing, the fortune-telling text generation unit 118 may, in the fortune-telling text generation step, perform a search of the academic literature database D using the search phrase, obtain the search results, and then generate the fortune-telling text by inputting a prompt containing the search results into a large-scale language model. Such a pre-generation search may be a search using a regular API, or a search using a language model for searching such as Perplexity®.

[0155] To encourage the input of further attribute values, it is preferable that the fortune-telling text generation unit 118 generates fortune-telling text that includes a statement requesting further input of attribute values. The generation of such fortune-telling text can be achieved, for example, by having the fortune-telling text generation unit 118 have the large-scale language model generate fortune-telling text that further includes wording prompting the user to input additional user attributes, when the fortune-telling text generation unit 118 has the large-scale language model generate fortune-telling text.

[0156] At this time, it is preferable that the generation device 1 executes the additional attribute acquisition step described later, the attribute acquisition unit 112 is configured to further generate additional search words based on the additional user attributes, and the fortune-telling text generation unit 118 inputs a prompt containing the additional search words to the large-scale language model. This allows the generation device 1 to generate further fortune-telling text based on user attributes additionally entered by the user who sees this text.

[0157] In other words, the fortune-telling text generation unit 118 can generate fortune-telling text such as, "If you were born in the Year of the Dragon and are a sign of Aries, then... If you would like a more detailed reading, please also tell us about the area where you work," prompting the user to input the user attribute "workplace." The generation device 1 can then provide the user with even more persuasive fortune-telling text based on insights further refined by actively segmenting the additionally entered workplace "Saitama City, Saitama Prefecture" using search terms such as "inland area" and "metropolitan area." This allows the generation device 1 to further encourage the input of even more user attributes.

[0158] The generation process preferably includes a process to display the generated fortune-telling text on terminal T or the like (step S15).

[0159] [Step S15: Command to display the fortune-telling text] The control unit 11 works in cooperation with the storage unit 13 and the communication unit 14 to execute the display command unit 119. The control unit 11 then executes a process to instruct the terminal T, etc., to display the fortune-telling text generated in the fortune-telling text generation step (display command step). The control unit 11 then moves the process to step S16.

[0160] If the generated fortune-telling text includes further wording prompting the user to input additional user attributes, the generation process preferably includes a series of processes to acquire the additionally input user attributes (additional user attributes) and reflect them in the attribute dataset. Steps S16 to S18 are an example of such processes. As a result, the generation device 1 can reflect the user attributes that the user has additionally input after seeing the wording in the attribute dataset, and provide a more accurate fortune-telling text based on the attribute dataset after the reflection.

[0161] [Step S16: Determine if additional input has been made] The control unit 11 works in cooperation with the storage unit 13 and the communication unit 14 to execute the attribute acquisition unit 112. The control unit 11 then performs a process to determine whether additional input has been made based on the wording prompting for additional input as described above (additional input determination step). If it is determined that additional input has been made, the control unit 11 moves the process to step S17; otherwise, the control unit 11 returns the process to step S1 and repeats the process from step S1 to step S18.

[0162] In the attribute acquisition determination step, the control unit 11 determines that additional input has been made, for example, when an attribute value relating to a user attribute that has been additionally input at terminal T is received by the generation device 1.

[0163] [Step S17: Get additional user attributes] The control unit 11 executes a process to acquire additional user attributes that have been added based on the wording prompting for additional input as described above, using the attribute acquisition unit 112 (additional attribute acquisition step). The control unit 11 then moves the process to step S18.

[0164] [Step S18: Reflect additional user attributes in the attribute dataset] The control unit 11 works in cooperation with the storage unit 13 and the communication unit 14 to execute the input reflection unit 113. Then, the control unit 11 performs a process to reflect the additional user attributes acquired by the input reflection unit 113 into the attribute dataset managed by the attribute management unit 111 (additional attribute reflection step). The control unit 11 then moves the process to step S6.

[0165] In the additional attribute reflection step, the input reflection unit 113 reflects the additional user attributes in the attribute dataset by procedures such as storing the entered attribute values, which are additional user attributes, in the attribute database 131. At this time, the entered attribute values ​​and additional user attributes entered by the same user in a series of sessions are linked to each other by being stored in the attribute database 131 as belonging to a single attribute dataset related to that user.

[0166] [Effects of the generation process] [Challenge 1: Providing users with better results in fortune-telling, etc.] The generation process of this embodiment can take various forms to provide the user with a large amount of results from fortune-telling or the like that reflect the latest knowledge, and to prompt the user to input further user attributes.

[0167] (Background to Issue 1) Based on user attributes entered by the user, the user's profile is narrowed down and used for personality tests, fortune-telling, and other applications. In such cases, if there are few user attributes entered, the user's profile may not be sufficiently narrowed down, and the application may not function properly.

[0168] However, attempting to implement a complete system by requiring users to input many user attributes from the start raises concerns that users may abandon using the application. This is because users may hesitate to input user attributes due to concerns such as the complexity of the process or the potential leakage of personal information.

[0169] Therefore, there is a need for technological means to further attract user interest and encourage them to willingly input more user attributes.

[0170] Regarding technologies that further attract user interest and encourage users to voluntarily input more user attributes in conventional services and technologies, Japanese Patent Publication No. 2007-236755 discloses a fortune-telling device characterized by comprising: means for displaying a screen prompting a fortune-telling requester to input personal information; means for displaying a screen prompting the fortune-telling requester to select the content of the fortune-telling requester from two or more pre-prepared standard phrases; means for receiving input of the personal information and input for selecting either of the standard phrases, and reading one or more question sentences corresponding to the selected standard phrase and two or more answer choices corresponding to the question sentences from a pre-prepared table; means for displaying the read question sentences and answer choices on a screen and prompting input for selecting an answer choice; and means for receiving input for selecting an answer choice, and outputting a fortune-telling result based on the received personal information input, the standard phrase selection input and the answer choice input. Furthermore, the above-mentioned document shows a method of searching a psychology database that provides different answers based on known psychology, and outputting a mixture of search results from the fortune-telling database and the psychology database.

[0171] The techniques described in the aforementioned literature can enable users of fortune-telling to feel satisfied that they have received direct answers to the specific problems they are facing, which is what they were looking for in the fortune-telling.

[0172] (Task 1) Incidentally, a vast amount of knowledge related to psychology has been discovered by diligent psychologists, and moreover, such knowledge is constantly being updated by psychologists. Reflecting this vast amount of knowledge in the psychology database of Patent Document 1 would require considerable effort. Furthermore, even if a psychology database reflecting this vast amount of knowledge were to be built once, adding questions and answer choices based on newly discovered knowledge to such a database and continuously reflecting the latest findings would place a heavy burden on the administrator of the psychology database.

[0173] Therefore, the technology described in the aforementioned literature can only provide users with answers that reflect psychological insights to the extent permitted by the economic circumstances of the application. Thus, the technology described in Patent Document 1 has room for further improvement in that it provides users with fortune-telling results that reflect a large amount of up-to-date knowledge and encourages the input of further user attributes.

[0174] The object of the present invention is to provide a technical means that prompts users to input further user attributes by providing them with results of fortune-telling, etc., that reflect a large amount of up-to-date knowledge. In this invention, "fortune-telling, etc." is not limited to those titled "fortune-telling," but refers to all kinds of analyses related to the user's character, exemplified by fortune-telling, personality tests, psychological tests, and personality assessments.

[0175] (Means to solve Problem 1) In response to the aforementioned challenges, a portion of academic literature databases containing knowledge from various academic fields has been made public in order to make researchers' findings known to the world. To utilize this knowledge, a large-scale language model has been realized that has the ability to search academic literature databases and generate answers based on the search results. Therefore, at first glance, it seems that by simply instructing such a large-scale language model to search academic literature databases using the parameters used for fortune-telling and to generate fortune-telling text, it would be possible to provide users with fortune-telling results that reflect a large amount of up-to-date knowledge.

[0176] In other words, for example, for a user born on April 1, 2000, it seems that the above problem can be solved simply by implementing a system that passively partitions the birth year into the Chinese zodiac to classify 2000 as the "Year of the Dragon," and passively partitions the birth date into the ruling constellation to classify April 1, 2000 as "Aries," generating a prompt that reads, "For subjects who are both in the Year of the Dragon and Aries, please search the academic literature database for relevant information and generate astrology text," and then inputting this prompt into a large-scale language model.

[0177] However, it is known that large-scale language models may sometimes respond as if non-existent search results exist when they cannot obtain search results that directly correspond to a given question. In the research, development, and application of large-scale language models, this phenomenon is called hallucination. Hallucination is thought to be caused by various factors, exemplified by the properties of the training data, training process, or architecture.

[0178] The inventors noticed that when they passively segmented user attributes according to criteria related to fortune-telling and then prompted a large-scale language model to generate fortune-telling texts based on these results, a phenomenon resembling hallucination occurred frequently. The inventors then performed the same processing on multiple types of large-scale language models and, despite referencing the same academic literature database, the generated fortune-telling texts differed in meaning, leading them to believe that hallucination was indeed occurring.

[0179] The frequent occurrence of hallucinations in this process is thought to stem from the fact that, given the general perception that fortune-telling is unscientific, knowledge directly corresponding to the results of passive division regarding fortune-telling is either not stored in academic literature databases, or if it is, only a small amount. Frequent hallucinations mean that fortune-telling texts cannot be said to reflect a large amount of up-to-date knowledge.

[0180] Therefore, in order to solve the above problems, the inventors diligently investigated and explored means of solving the hallucination problem. As a result, they found that the above objectives can be achieved by generating search words for an academic literature database from user attributes, inputting instructions to generate fortune-telling text and prompts containing the search words into a large-scale language model that references academic literature. Thus, the inventors completed the present invention.

[0181] (Aspect relating to the first feature corresponding to Problem 1) The generation device 1 of this embodiment, which performs the generation process described above, may take the following configuration: the user attributes acquired by the attribute acquisition unit 112 (steps S1 to S3) are converted by the search word generation unit 117 into search words used to reference academic literature published in an external academic literature database D (step S12); and the fortune-telling text generation unit 118 inputs a prompt including this search word and an instruction to generate fortune-telling text into a large-scale language model to generate fortune-telling text related to the user (step S14). In other words, in this configuration, the generation device 1 may take the configuration in which the search word generation unit 117 generates fortune-telling text based on references to academic literature using search words obtained by actively segmenting the user attributes.

[0182] In the fortune-telling text generation unit 118, the large-scale language model references academic literature using search terms used for searching academic literature, rather than relying on the results of passive segmentation related to fortune-telling that is prone to hallucination (step S14). The fortune-telling text generation unit 118 then generates fortune-telling texts that reflect a large amount of up-to-date knowledge from various academic fields, such as psychology or sociology, using the large-scale language model (step S14).

[0183] As a result, the generation device 1 in this embodiment can provide the user with compelling fortune-telling texts that reflect a large amount of up-to-date knowledge, while suppressing hallucination, and can encourage the input of further user attributes.

[0184] This embodiment may be configured to generate fortune-telling text related to the user by referencing academic literature using search words generated based on the estimated results of unentered attribute values, which are estimated based on the entered attribute values.

[0185] When there are many conditions required for the literature, the likelihood of finding corresponding literature decreases. Therefore, in such cases, the risk of hallucination is considered to increase. In the above configuration, the generation device 1 can suppress hallucination and provide the user with a large amount of compelling fortune-telling text that reflects the latest knowledge, even in applications with an even higher risk of hallucination, such as generating fortune-telling text based on a user profile further narrowed down by estimation.

[0186] (Aspects relating to the second feature corresponding to Issue 1) In addition to the above-described embodiment corresponding to Problem 1, the generation process of this embodiment may also take the form of a process in which the search word generation unit 117 generates search words by performing an active segmentation that converts user attributes into search words for referencing academic literature (step S12), and the fortune-telling text generation unit 118 generates fortune-telling text that includes a description that associates the results of passive segmentation of user attributes according to criteria related to fortune-telling (step S13) with knowledge obtained from academic literature (step S14).

[0187] In this embodiment, the fortune-telling text generation unit 118 generates fortune-telling text that includes a description that associates the results of passive division based on criteria related to fortune-telling, such as passive division that converts the year of birth into the Chinese zodiac, and passive division that converts the date of birth into a guardian constellation, with knowledge obtained from academic literature.

[0188] Therefore, this embodiment can provide users with compelling fortune-telling texts that reflect a large amount of up-to-date knowledge without compromising the mystique of fortune-telling. As a result, this embodiment can further encourage the input of additional user attributes.

[0189] This embodiment may be configured to generate fortune-telling text by referencing academic literature using search words generated by active segmentation, even for the estimated results of unentered attribute values ​​that have been estimated based on entered attribute values.

[0190] When there are many conditions required for the literature, the likelihood of finding corresponding literature decreases. Therefore, in such cases, the risk of hallucination is considered to increase. In the above configuration, generation device 1 can suppress hallucination even in applications with an even higher risk of hallucination, such as generating fortune-telling text based on a user profile further narrowed down by estimation. Furthermore, generation device 1 can provide users with persuasive fortune-telling text that reflects a large amount of up-to-date knowledge, and in which the mystique of fortune-telling is enhanced by associating the results of passive division related to fortune-telling with said knowledge.

[0191] (An aspect relating to the third feature corresponding to Issue 1) In addition to any of the above-described embodiments corresponding to Problem 1, the generation process of the present invention may also take the form of: an attribute acquisition unit 112 acquiring user attributes indicating the user's concerns (step S10); a search word generation unit 117 further generating concern search words based on these concerns (step S11); and a fortune-telling text generation unit 118 inputting a prompt further containing the concern search words into a large-scale language model (step S14).

[0192] Those who seek fortune-telling often look to it for clues to solving their problems. This particular method involves having the user input their problems themselves, and then using actively segmented search terms based on these problems to refer to academic literature.

[0193] Therefore, this embodiment can provide users with persuasive fortune-telling text that reflects appropriate insights tailored to their concerns, while also possessing the mystical quality of fortune-telling, and while encouraging them to input user attributes related to their concerns. This embodiment provides a technical means to encourage users to input further user attributes by providing them with fortune-telling results that reflect a large amount of up-to-date knowledge.

[0194] This embodiment can be configured to generate fortune-telling text related to the user by inputting a prompt into a large-scale language model that further includes the estimated results of uninputted attribute values, which have been estimated based on the inputted attribute values, as well as search words generated by actively splitting the estimated results, search words related to the inputted attribute values, and problem search words.

[0195] When there are many conditions required for the literature, the likelihood of finding corresponding literature decreases. Therefore, in such cases, the risk of hallucination is considered to increase. In the above configuration, the generation device 1 can suppress hallucination and provide the user with a large amount of compelling fortune-telling text that reflects the latest knowledge, even in applications where the risk of hallucination is even higher, such as generating fortune-telling text based on a user's profile and concerns that have been further narrowed down by estimation.

[0196] (Aspects relating to the fourth feature corresponding to Issue 1) In addition to any of the above-described embodiments corresponding to Problem 1, the generation process of the present invention may also take the following configuration: the fortune-telling text generation unit 118 generates fortune-telling text in a large-scale language model that further includes phrases prompting the user to input additional user attributes (step S14); the attribute acquisition unit 112 is configured to further acquire additional user attributes that have been added based on these phrases (steps S16 to S18); the search word generation unit 117 is configured to further generate additional search words based on the additional user attributes (step S12); and the fortune-telling text generation unit 118 inputs a prompt further including the additional search words into the large-scale language model (step S14).

[0197] In this embodiment, the fortune-telling text generation unit 118 generates fortune-telling text that further includes wording prompting the user to input additional user attributes (step S14). Then, this embodiment generates further fortune-telling text based on the user attributes that the user has additionally input after seeing this wording (steps S16 to S18) (steps S6 to S15). Thus, this embodiment can provide a technical means to prompt the user to input further user attributes by giving them fortune-telling results that reflect a large amount of up-to-date knowledge.

[0198] This embodiment can be configured to generate fortune-telling text for a user by inputting a prompt into a large-scale language model that includes instructions to search the academic literature database D using a search phrase that includes search words generated by actively splitting the estimated results of unentered attribute values ​​based on entered attribute values ​​and additional user attributes, search words related to entered attribute values, and search words related to additional user attributes, thereby generating fortune-telling text for the user.

[0199] When there are many conditions required for the literature, the likelihood of finding corresponding literature decreases. Therefore, in such cases, the risk of hallucination is considered to increase. In the above configuration, even in applications where the risk of hallucination is even higher, such as generating fortune-telling texts based on a user profile further refined by additional user attributes and estimation results, the generator 1 can suppress hallucination and provide users with compelling fortune-telling texts that reflect a large amount of up-to-date knowledge.

[0200] (Aspects relating to the fifth feature corresponding to Issue 1) In addition to any of the above embodiments corresponding to Problem 1, the generation process of the present invention may also take the form in which the fortune-telling text generation unit 118 uses a large-scale language model that has been pre-trained on academic literature to refer to academic literature (step S14).

[0201] As mentioned above, large-scale language models can experience hallucinations. Hallucinations can be broadly categorized into intrinsic hallucinations, which output facts different from those in the training data, and extrinsic hallucinations, which output facts not present in the training data. The search term generation unit contributes to suppressing extrinsic hallucinations by generating appropriate search terms, thereby encouraging the referencing of academic literature present in the training data. However, referencing academic literature by large-scale language models presents challenges related to intrinsic hallucinations.

[0202] This embodiment utilizes a RAG that has been pre-trained based on academic literature, allowing the large-scale language model to obtain the descriptions related to academic literature from external resources of the large-scale language model, instead of generating the descriptions related to academic literature itself.

[0203] By acquiring the descriptions themselves from academic literature, it is possible to prevent the output of incorrect information that differs from the training data (content of academic literature). Therefore, this invention suppresses not only extrinsic hallucinations but also intrinsic hallucinations, and can appropriately reflect a large amount of up-to-date knowledge in the results of fortune-telling.

[0204] Therefore, this embodiment can provide a technical means to encourage users to input further user attributes by providing them with results such as fortune-telling that reflect a large amount of up-to-date knowledge.

[0205] (Aspects relating to the sixth feature corresponding to Issue 1) In addition to any of the above embodiments corresponding to Problem 1, the generation process of the present invention may further include an attribute management unit 111 that manages user attributes in association with their publication scope.

[0206] By acquiring a large number of user attributes, it is possible to appropriately narrow down the user profile. However, acquiring a large number of user attributes can raise concerns regarding the protection of personal information. Such concerns may cause users to hesitate in entering those attributes.

[0207] In this configuration, the attribute management unit 111 manages user attributes in association with their disclosure scope, thereby appropriately managing personal information. This eliminates user concerns regarding personal information. Furthermore, according to this configuration, users can control the disclosure scope themselves, for example, disclosing only a pseudonym and a portion of their profile to acquaintances only online, and disclosing their date of birth or address to close friends, thus managing personal information according to the depth of their relationship. Such personal information management encourages users to input more user attributes.

[0208] Therefore, this embodiment can provide a technical means to encourage users to input further user attributes by providing them with results such as fortune-telling that reflect a large amount of up-to-date knowledge.

[0209] (Summary of the embodiments corresponding to Problem 1) Based on the above, each embodiment corresponding to Problem 1 can provide a technical means to give users fortune-telling results that reflect a large amount of up-to-date knowledge and encourage them to input further user attributes.

[0210] [Challenge 2: Realizing services based on a more comprehensive personal profile] The generation process of this embodiment contributes to the realization of a service based on a person profile that contains more information than the person profile explicitly stated by the entered user attributes themselves, and can take various forms to encourage the input of further user attributes.

[0211] (Background to Issue 2) Based on user attributes entered by the user, the user's profile is narrowed down and used for personality tests, fortune-telling, and other applications. In such cases, if there are few user attributes entered, the user's profile may not be sufficiently narrowed down, and the application may not be able to operate properly.

[0212] However, if users are asked to enter a large amount of user attributes from the start, it is likely that they will abandon using the application. This is because users may hesitate to enter their user attributes due to reasons such as the complexity of the process or concerns about the leakage of personal information.

[0213] Therefore, there is a need for technical means to allow the application to run as much as possible, even when only a small amount of user attributes have been entered, thereby increasing the likelihood of further user attribute input.

[0214] Regarding technologies that increase the likelihood of user attribute input in conventional services and technologies, Japanese Patent Publication No. 2013-171477 discloses that various bots responsible for Four Pillars of Destiny, animal fortune-telling, blood type fortune-telling, astrology, and compatibility fortune-telling ask questions to SNS users who have requested to become friends, according to a prepared scenario. Patent Document 1 describes how, regarding the collection of personal information by the bot, it also takes into account personal information already stored in the personal information database and is configured to prevent the duplicate collection of personal information that has already been collected from the same user.

[0215] The technology described in the aforementioned literature can enable the collection of personal information from social media users that has not yet been collected from users, through fortune-telling bots, which each bot should collect.

[0216] (Task 2) By the way, in applications that can provide higher-quality services tailored to specific target individuals by narrowing down the target person profile based on methods such as fortune-telling and personality assessments, the more user attributes that can be utilized, the higher the quality of the service that can be provided.

[0217] However, the technology described in the aforementioned literature is limited to providing users with various fortune-telling results based on collected personal information. Therefore, the technology described in the aforementioned literature has room for further improvement in order to realize a service based on a profile that contains more information than the profile explicitly stated by the entered user attributes themselves.

[0218] The objective of this invention is to contribute to the realization of a service based on a person profile that contains more information than the person profile explicitly stated by the entered user attributes themselves, and to provide a technical means that encourages the input of further user attributes.

[0219] (Means to solve problem 2) As a result of diligent research to solve the above problems, the inventors have found that the above objectives can be achieved by generating estimation results for unentered attribute values ​​from entered attribute values ​​based on the correlation between multiple attribute values ​​belonging to an attribute dataset, managing the attribute dataset in a format that allows the estimated attribute values ​​and the entered attribute values ​​to be distinguished from each other, and generating fortune-telling text based on an attribute dataset that may contain such estimation results. Thus, the inventors have completed the present invention.

[0220] (An embodiment relating to the first feature corresponding to Issue 2) The generation process of the present invention may take the following forms: an attribute management unit 111 manages an attribute dataset related to the user's profile; an attribute acquisition unit 112 acquires attribute values ​​that belong to this attribute dataset and have been entered by the user (steps S1 to S2); an input reflection unit 113 reflects the entered attribute values ​​in the attribute dataset (step S3); an attribute estimation unit 115 generates estimation results based on the entered attribute values ​​for unentered attribute values ​​that belong to this attribute dataset and have not been entered by the user (steps S6 to S7); an estimation reflection unit 116 reflects these estimation results in the attribute dataset (step S8); and a fortune-telling text generation unit 118 generates fortune-telling text based on the attribute dataset (step S14). In this embodiment, the attribute management unit 111 manages an attribute dataset in a format that includes each attribute value belonging to a predetermined set of attributes and that allows for the distinction between input attribute values ​​and estimated attribute values ​​(see the section on attribute database 131), the attribute estimation unit 115 generates estimation results based on the correlation between the multiple attribute values ​​belonging to the attribute dataset (step S7), and the fortune-telling text generation unit 118 generates fortune-telling text that includes a description requesting further input of attribute values ​​(step S14).

[0221] In this configuration, the attribute management unit 111 manages attribute datasets related to the user's profile. This attribute management unit 111 manages attribute datasets that contain attribute values ​​belonging to a predetermined set of attributes. As a result, various applications that utilize this attribute dataset can use the predetermined attribute dataset as is to narrow down the profile necessary for the implementation of their application. In other words, this configuration enables various applications to provide services in a simple procedure without requiring the complex processing steps of converting unstructured attribute datasets to suit their applications.

[0222] In this embodiment, the attribute management unit 111 manages the attribute dataset in a format that allows it to distinguish between attribute values ​​reflected in the attribute dataset by the input reflection unit 113 and the estimated results of unentered attribute values ​​reflected in the attribute dataset by the attribute estimation unit 115. This enables the attribute estimation unit 115 to generate estimated results of unentered attribute values ​​based only on entered attribute values ​​(step S7). In other words, this embodiment suppresses the inclusion of errors in the estimation results due to the estimation results being fed back into the estimation results.

[0223] In this embodiment, the attribute estimation unit 115 generates estimation results for uninputted attribute values ​​based on the correlation between multiple attribute values ​​belonging to the attribute dataset (step S7). This enables the embodiment to generate reasonable estimation results based on correlation. Then, the fortune-telling text generation unit 118 generates fortune-telling text based on this attribute dataset (step S14).

[0224] In addition, the fortune-telling text generation unit 118 in this embodiment generates fortune-telling text that includes a description requesting further input of attribute values ​​(step S14). By generating fortune-telling text based on the attribute dataset including the estimation results described above, and by generating fortune-telling text that includes the description described above, this embodiment realizes fortune-telling text based on a more appropriately narrowed profile of the user than when estimation is not performed, and increases the likelihood that further attribute values ​​will be input.

[0225] Therefore, this embodiment contributes to the realization of a service based on a person profile that contains more information than the person profile explicitly stated by the entered user attributes themselves, and can provide a technical means to encourage the input of further user attributes.

[0226] In this configuration, the search word generation unit 117 generates search words used to reference publicly available academic literature based on user attributes (step S12), and the fortune-telling text generation unit 118 inputs an instruction to generate fortune-telling text and a prompt containing these search words into a large-scale language model to generate fortune-telling text related to the user (step S14).

[0227] By using a large-scale language model configured to search the academic literature database D, it is possible to generate fortune-telling texts that reflect a large amount of up-to-date knowledge from various academic fields. However, as mentioned above, in fortune-telling text generation using a large-scale language model, the risk of hallucination increases as the person's profile is narrowed down.

[0228] This configuration, by generating search terms used for searching in the academic literature database D, suppresses the occurrence of hallucination, even in situations where hallucination is likely to occur, such as generating fortune-telling text about a person's profile that contains more information than the person's profile explicitly stated by the entered user attributes themselves. This contributes to providing fortune-telling text that appropriately reflects knowledge about the person's profile.

[0229] (An embodiment relating to the second feature corresponding to Issue 2) In addition to any of the above-described embodiments corresponding to Problem 2, the generation process of this embodiment may also take the form of generating estimation results by a process (steps S4 to S7) in which the attribute estimation unit 115 generates estimation results by inputting the input attribute values ​​into a machine learning model that has been pre-trained using training data that includes input attribute values ​​and does not include estimation results for uninput attribute values, and then outputs estimation results for uninput attribute values.

[0230] This embodiment utilizes a machine learning model that has been pre-trained using training data that includes input attribute values ​​but does not include estimated results for uninputted attribute values, thereby enabling the generation of estimation results that reflect the complex correlations between many attribute values.

[0231] As a result, this embodiment can output estimation results with higher accuracy than generating estimation results based on correlations between a limited number of attribute values, especially when many attribute values ​​are input through additional input. Therefore, this embodiment contributes to the realization of services based on a person profile that contains even more information than the person profile explicitly stated by the input user attributes themselves.

[0232] Therefore, this embodiment contributes to the realization of a service based on a person profile that contains more information than the person profile explicitly stated by the entered user attributes themselves, and can provide a technical means to encourage the input of further user attributes.

[0233] In this configuration, the search word generation unit 117 generates search words used to reference publicly available academic literature based on user attributes (step S12), and the fortune-telling text generation unit 118 inputs an instruction to generate fortune-telling text and a prompt containing these search words into a large-scale language model to generate fortune-telling text related to the user (step S14).

[0234] By using a large-scale language model configured to search the academic literature database D, it is possible to generate fortune-telling texts that reflect a large amount of up-to-date knowledge from various academic fields. However, as mentioned above, in fortune-telling text generation using a large-scale language model, the risk of hallucination increases as the person's profile is narrowed down. Such hallucination cannot be prevented by improving the accuracy of estimation using machine learning models.

[0235] This configuration, by generating search terms used for searching in the academic literature database D, suppresses the occurrence of hallucination, even in situations where hallucination is likely to occur, such as generating fortune-telling text about a person's profile that contains more information than the person's profile explicitly stated by the entered user attributes themselves. This contributes to providing fortune-telling text that appropriately reflects knowledge about the person's profile.

[0236] (An embodiment relating to the third feature corresponding to Issue 2) In addition to any of the above-described embodiments corresponding to Problem 2, the generation process of this embodiment may also take the form of generating fortune-telling text for a user by having a process in which the fortune-telling text generation unit 118 inputs a prompt to a large-scale language model that includes an instruction to generate fortune-telling text for a user of a person profile indicated by these attribute values, based on the estimation results of inputted attribute values ​​and uninputted attribute values ​​related to the attribute dataset (step S14).

[0237] In simple bot-based fortune-telling, a given data table is referenced based on a limited number of user attributes such as date of birth and blood type to provide fortune-telling text. However, there is room for further improvement in providing fortune-telling text that reflects the wide variety of attribute values ​​contained in the attribute dataset.

[0238] This embodiment allows for the generation of fortune-telling text that reflects the estimation results of a wide variety of input and uninputted attribute values ​​included in the attribute dataset, using a large-scale language model. As described above, the estimation results of uninputted attribute values ​​are reasonable based on correlations. Therefore, despite reflecting the estimation results of uninputted attribute values ​​in addition to input attribute values ​​in the fortune-telling text, this invention can provide users with persuasive fortune-telling text.

[0239] In this way, by having a large-scale language model trained on a large amount of text data generate fortune-telling text that takes into account not only the estimated results of entered attribute values ​​but also reasonable estimated results of unentered attribute values, it can be expected that users will trust the fortune-telling results shown in the text and enter further user attributes.

[0240] Therefore, this embodiment contributes to the realization of a service based on a person profile that contains more information than the person profile explicitly stated by the entered user attributes themselves, and can provide a technical means to encourage the input of further user attributes.

[0241] In this configuration, the search word generation unit 117 generates search words used to reference publicly available academic literature based on user attributes (step S12), and the fortune-telling text generation unit 118 inputs an instruction to generate fortune-telling text and a prompt containing these search words into a large-scale language model to generate fortune-telling text related to the user (step S14).

[0242] By using a large-scale language model configured to search the academic literature database D, it is possible to generate fortune-telling texts that reflect a large amount of up-to-date knowledge from various academic fields. However, as mentioned above, in fortune-telling text generation using a large-scale language model, the risk of hallucination increases as the person's profile is narrowed down. It is difficult to suppress such hallucination by training a large-scale language model.

[0243] This configuration, by generating search terms used for searching in the academic literature database D, suppresses the occurrence of hallucination, even in situations where hallucination is likely to occur, such as generating fortune-telling text about a person's profile that contains more information than the person's profile explicitly stated by the entered user attributes themselves. This contributes to providing fortune-telling text that appropriately reflects knowledge about the person's profile.

[0244] (Aspects relating to the fourth feature corresponding to Issue 2) In addition to any of the above-described embodiments corresponding to Problem 2, the generation process of this embodiment may also take the form in which the attribute management unit 111 manages the attribute dataset in an anonymous dataset form that includes estimated results of unentered attribute values ​​and entered attribute values, and does not contain personal information that can identify a user (see the section on attribute database 131).

[0245] As described above, the generation device 1 of this embodiment manages attribute datasets, including estimated results for unentered attribute values, using the attribute management unit 111. Therefore, users may have a stronger concern about such attribute datasets, which provide a more detailed profile of a person, being leaked as personal information than they would with attribute datasets that do not include estimated results.

[0246] In this configuration, the attribute management unit 111 manages the attribute dataset in an anonymous dataset format that does not contain personal information that can identify the user, thereby preventing the attribute dataset from being leaked as personal information. Therefore, this configuration eliminates concerns about the leakage of personal information, even though it manages an attribute dataset that allows for a more detailed identification of the user's profile, and furthermore, it can meet the standards required by laws and regulations regarding personal information. As a result, this configuration encourages the input of further user attributes.

[0247] Therefore, this embodiment contributes to the realization of a service based on a person profile that contains more information than the person profile explicitly stated by the entered user attributes themselves, and can provide a technical means to encourage the input of further user attributes.

[0248] (Summary of the embodiments corresponding to Problem 2) Based on the above, each embodiment corresponding to Problem 2 contributes to the realization of a service based on a person profile that contains more information than the person profile explicitly stated by the entered user attributes themselves, and can provide a technical means to encourage the input of further user attributes.

[0249] <Usage example> The following is an example of how to use the generation device 1 of this embodiment.

[0250] [Fortune-telling, etc. designated] The user, via terminal T, selects the type of fortune-telling or divination from multiple options related to the content of the divination, such as MBTI diagnosis, Enneagram, fortune-telling: love (life), and fortune-telling: love (marriage).

[0251] [When specifying a fortune teller] The user selects "Romance (Marriage)" from the above options via terminal T. The generation device 1 displays text on terminal T requesting input of user attributes related to the selected option. The generation device 1 also receives the specified content to be predicted as the user's concern and generates search terms corresponding to that concern, such as "How you met your partner" and "Methods used to find a marriage partner."

[0252] (Entering user attributes) The user accesses the generator 1 via terminal T and inputs user attributes, specifically gender ("male"), age ("27 years old"), blood type ("Type O"), and date of birth ("October 30th"). The generator 1 receives the input attribute values ​​(gender, age, blood type, and date of birth) from terminal T and reflects them in the attribute dataset of attribute database 131.

[0253] (Estimation of unentered attribute values) Based on the input attribute values ​​of gender "male", age "27 years old", blood type "Type O", and date of birth "October 30th", the generator 1 estimates the uninputted attribute values ​​such as height "170-174cm", personality "normal", occupation "part-time worker", and dominant hand "right-handed". At this time, the generator 1 estimates the uninputted attribute values ​​from the input attribute values ​​using the correlation between the input attribute values. Then, the generator 1 reflects the estimation results in the attribute dataset of attribute database 131.

[0254] (Generating search terms) The generation device 1 generates search words for referencing academic literature using data for actively segmenting user attributes, based on both the input attribute values ​​and the estimated results of uninputted attribute values.

[0255] (Generating fortune-telling text (love fortune-telling)) The generator 1 inputs instructions to generate fortune-telling text and prompts including the aforementioned search words into a large-scale language model that references academic literature, thereby generating fortune-telling text relevant to the user. The generator 1 also associates the results of passively dividing the age "27 years old" into the twelve zodiac signs ("Year of the Ox"), the birth date "October 30th" into the ruling zodiac sign ("Scorpio"), and the blood type "Type O" with the content of the fortune-telling text. The generator 1 then displays the generated fortune-telling text, "You, born in the Year of the Ox and with Scorpio as your ruling zodiac sign, and blood type O, your love fortune this month is... If you would like a more detailed reading, please also tell us where you were born and where you currently live," on terminal T.

[0256] (Enter additional user attributes) Based on a well-understood profile of the user, the user is drawn to a fortune-telling result written in a mystical style beginning with "You, born in the Year of the Ox, with Scorpio as your ruling sign and blood type O, your love fortune this month is..." and incorporates a large amount of up-to-date knowledge obtained from searching the academic literature database D. Users who are intrigued by this result then add their place of origin "Hyogo Prefecture" and workplace "Tokyo". The generator 1 receives the additional user attributes of place of origin and workplace from terminal T and reflects them in the attribute dataset of attribute database 131. The generator 1 also estimates any unentered attribute values ​​and reflects the estimation results in the attribute dataset of attribute database 131. Subsequently, the generator 1 generates search words based on these user attributes, generates fortune-telling text that reflects the search results of academic literature database D related to these search words, and displays it on terminal T. Users who are further intrigued by the more accurate fortune-telling text then input even more user attributes.

[0257] [When specifying the MBTI test] The user selects "MBTI Diagnosis" from the above options via terminal T. The generating device 1 displays text on terminal T requesting input of user attributes related to the selected option.

[0258] (Entering user attributes) The user accesses the generator 1 via terminal T and inputs user attributes related to the questions on the first page of the MBTI diagnostic test. The reason for requiring input across multiple pages is that the MBTI diagnostic test has a large number of questions, 93 in total. The generator 1 receives the input attribute values ​​related to the questions on the first page from terminal T and reflects them in the attribute dataset of the attribute database 131.

[0259] (Estimation of unentered attribute values) The generator 1 estimates the unentered attribute values ​​for the questions on page 2 based on the entered attribute values ​​for the questions on page 1. Then, the generator 1 provides the user with the estimated results as the initial values ​​for the fields where attribute values ​​are entered. In other words, the generator 1 performs autocomplete based on the entered attribute values. The user modifies the initial values ​​as needed.

[0260] The generation device 1 repeats this procedure to obtain attribute values ​​entered by the user or estimated results that the user did not modify for all question items on every page.

[0261] (Generating search terms) The generation device 1 generates search words for referencing academic literature using data for actively segmenting user attributes, based on both the input attribute values ​​and the estimated results of uninputted attribute values.

[0262] (Generating fortune-telling text (MBTI diagnostic results)) The generator 1 inputs instructions to generate fortune-telling text and prompts including the aforementioned search words into a large-scale language model that references academic literature, thereby generating fortune-telling text related to the user, i.e., the diagnostic results related to the MBTI diagnosis. The generator 1 then displays the generated fortune-telling text, "Your basic type is INFP (Mediator). INFP types are idealistic, highly empathetic, and true to their own values. They are introverted, highly creative, sensitive, and empathetic towards the feelings of others..." on terminal T.

[0263] 〔Use of Attribute Dataset〕 The attribute dataset collected in this way is stored in the attribute database 131 in the form of an anonymous dataset that does not contain personal information. The administrator of the generation device 1 can use the attribute dataset, which is an anonymous dataset, not only for generating fortune-telling texts but also for SNS advertisements of the attribute information input type, etc. In the SNS advertisement of the attribute information input type, the content of the advertisement displayed is optimized based on the user attributes input in this way.

[0264] In the scope of the idea of the present invention, those skilled in the art can conceive of various modification examples and correction examples. Therefore, those modification examples and correction examples are understood to belong to the scope of the present invention. For example, for the above-described embodiments, those obtained by appropriately adding, deleting, or changing the design of components by those skilled in the art, or those obtained by adding, omitting, or changing the conditions of the process, are also included in the scope of the present invention as long as they have the gist of the present invention.

Explanation of Signs

[0265] S System 1 Generation Device 11 Control Unit 111 Attribute Management Unit 112 Attribute Acquisition Unit 113 Input Reflection Unit 114 Machine Learning Unit 115 Attribute Estimation Unit 116 Estimation Reflection Unit 117 Search Word Generation Unit 118 Fortune-Telling Text Generation Unit 119 Display Instruction Unit 13 Storage Unit 131 Attribute Database 14 Communication Unit D Academic Literature Database N Network T Terminal

Claims

1. An attribute acquisition unit that acquires one or more user attributes related to a user, A search term generation unit generates search terms used to reference academic literature published in external academic literature databases based on the user attributes, A fortune-telling text generation unit inputs instructions to generate fortune-telling text and prompts including the search words into a large-scale language model that references academic literature, and generates fortune-telling text relating to the user. A device for generating fortune-telling texts, equipped with the necessary features.

2. The search term generation unit generates the search term by a process that includes a procedure for performing active segmentation to convert the user attributes into search terms for referring to academic literature, The fortune-telling text generation unit generates fortune-telling text associated with the results of passively dividing the user attributes according to the criteria related to fortune-telling. The generating apparatus according to claim 1.

3. The attribute acquisition unit is configured to acquire user attributes that indicate the user's concerns. The aforementioned search term generation unit is configured to further generate search terms for problems based on the aforementioned problems, The fortune-telling text generation unit is configured to input prompts further containing the aforementioned problem search words to the large-scale language model. The generating apparatus according to claim 1 or claim 2.

4. The fortune-telling text generation unit causes the large-scale language model to generate fortune-telling text that further includes wording prompting the user to input additional user attributes. The attribute acquisition unit is configured to acquire additional user attributes that have been added based on the aforementioned text. The search term generation unit is configured to further generate additional search terms based on the additional user attributes, The fortune-telling text generation unit is configured to input a prompt further containing the additional search words to the large-scale language model. The generating apparatus according to claim 1 or claim 2.

5. The generation apparatus according to claim 1 or 2, wherein the fortune-telling text generation unit uses a Retrievable-Augmented Generation (RAG) that has been pre-trained based on the academic literature to refer to the academic literature in the large-scale language model.

6. The generation apparatus according to claim 1 or 2, further comprising an attribute management unit for managing the user attributes in association with their publication scope.

Images