Information generation method, information generation device and information generation program

Abstract

To provide an information generation system and an information generation device that enable a user of any position to make statements (transmit information) with improved anonymity and safety.SOLUTION: In an information generation system 1, a server 20 includes a processing unit and a storage unit. The processing unit, in cooperation with the storage unit, receives opinion information of multiple anonymous users within a group, combines the opinion information of the multiple anonymous users, and generates representative opinion information of the group using the combined opinion information of the multiple anonymous users.SELECTED DRAWING: Figure 1

Description

[Technical Field]

[0001] The present disclosure relates to an information generation method, an information generation device, and an information generation program. [Background technology]

[0002] In recent years, it has been pointed out that in hierarchical organizations, such as between company management and field employees, members in a weaker position are reluctant to speak their true feelings because they are wary of those in a stronger position. Therefore, there is a demand for a communication tool platform that allows anyone to express their true opinions without having to consider the opinions of others, even if the weaker member is not known by the stronger member, or vice versa.

[0003] Patent Document 1 discloses a screening work document creation support device for financial institutions. This screening work document creation support device includes an item identification means for identifying specific input items from among the input items provided in the screening work document format, a question generation means for generating question information for a language model to respond to input information for the specific input items, a question transmission means for transmitting the question information to the language model, a response reception means for receiving response information corresponding to the question information from the language model, and a response input means for inputting input information based on the response information into the specific input items. This screening work document creation support device causes the language model to generate response information for the question information. Using such a language model, it is conceivable to create a platform that facilitates the exchange of honest opinions regardless of position. [Prior art documents] [Patent documents]

[0004] [Patent Document 1] Patent No. 7396582 Summary of the Invention [Problem to be solved by the invention]

[0005] However, even if the language model of Patent Document 1 is simply used, it is difficult to guarantee the anonymity and security of opinions (comments) from those who express them (for example, employees who are in a weak position to blow the whistle). For this reason, it is not necessarily possible to realize a platform where people can easily express their true opinions.

[0006] The present disclosure has been devised in consideration of the above-mentioned conventional circumstances, and aims to enable users of any position to make statements (disseminate information) with improved anonymity and security. [Means for solving the problem]

[0007] The present disclosure provides an information generation method comprising a processing unit and a memory unit, wherein the processing unit cooperates with the memory unit to receive opinion information of a plurality of anonymous users in a group, mix the opinion information of the plurality of anonymous users, and generate representative opinion information of the group using the mixed opinion information of the plurality of anonymous users. [Effects of the Invention]

[0008] According to the present disclosure, users of any position can make statements (transmit information) with improved anonymity and security. [Brief explanation of the drawings]

[0009] [Figure 1] 1 is a schematic diagram illustrating an information generating system according to an embodiment of the present invention; [Figure 2] 1 is a sequence diagram illustrating an information generation method according to an embodiment of the present invention in chronological order. [Figure 3] A contrasting diagram of the original post (text before tone adjustment) and the text after tone adjustment. [Figure 4] FIG. 1 is a flow diagram showing a method for generating anonymous group opinions according to an embodiment of the present invention. [Figure 5] A contrastive comparison of sentences before and after blending DETAILED DESCRIPTION OF THE INVENTION

[0010] Hereinafter, with reference to the drawings as appropriate, embodiments that specifically disclose the configuration and operation of an information generation method and an information generation device according to the present disclosure will be described in detail. However, more detailed description than necessary may be omitted. For example, detailed description of well-known matters or redundant description of substantially identical configurations may be omitted. This is to avoid unnecessary redundancy in the following description and to facilitate understanding by those skilled in the art. Note that the accompanying drawings and the following description are provided to enable those skilled in the art to fully understand the present disclosure and are not intended to limit the subject matter recited in the claims.

[0011] (Embodiment) 1. Information Generation System 1 First, an information generation system 1 according to the present embodiment will be described with reference to FIG. 1. FIG. 1 is a schematic diagram illustrating the information generation system 1 according to the present embodiment. The information generation system 1 includes one or more user terminals 10 and a server 20. Specifically, multiple user terminals 10 may be connected to the information generation system 1 so as to enable data communication (see FIG. 2). For example, assume that a chairperson (e.g., a boss or organizer) and multiple employees within an organization within a company each own a user terminal 10 with a chat function. The chairperson sets a talk topic or a questionnaire, and multiple employees post their opinions on the talk topic or questionnaire in the chat. In this case, the information generation system 1 provides anonymity and security for posted opinions so that it is not clear who posted which opinion, making it easier for even weaker employees to post their opinions freely.

[0012] The user terminal 10 and the server 20 that make up the information generation system 1 can transmit and receive information and data to and from each other via a wired cable or a communication network. Examples of wired cables include an HDMI (registered trademark) cable and a USB cable. Examples of communication networks include a wired LAN, a wireless LAN, LTE, 4G, 5G, the Internet, a Virtual Private Network (VPN), and the like.

[0013] The user terminal 10 has at least a processing unit 11 and a memory unit 12. The processing unit 11 includes a control unit 111 and a base model 112. The memory unit 12 stores a user input sentence 121 and an adjusted sentence 122. The user terminal 10 is, for example, an electronic device such as a personal computer, a server, a smartphone, or a tablet.

[0014] When the control unit 111 receives a user input, it adjusts the input sentence (i.e., the user's opinion) in cooperation with the base model 112 and transmits the adjusted sentence to the server 20. The control unit 111 is configured using at least one of, for example, a Central Processing Unit (CPU), a Micro Processing Unit (MPU), a controller, a Large Scale Integration (LSI), an Application Specific Integrated Circuit (ASIC), a Programmable Logic Device (), a Field-Programmable Gate Array (FPGA), and a Graphic Processing Unit (GPU).

[0015] The base model 112 is also called a foundation model, and is a machine learning model of Artificial Intelligence (AI) that has been trained (pre-learned) using a large amount of data, and is a large-scale deep learning neural network. In particular, in this embodiment, a language model (LM) or a large-scale language model (LLM) is included. The base model 112 adjusts the sentences input by the user in response to instructions from the control unit 111. The adjustment will be described later.

[0016] The memory unit 12 stores computer programs and data handled by the processing unit 11. The memory unit 12 may include, for example, a read only memory (ROM) and a random access memory (RAM). The user input sentence 121 stored in the memory unit 12 is a sentence input by a user via a user interface (UI) 13. The adjusted sentence 122 is a sentence input by a user after the sentence has been adjusted by the processing unit 11. The user input sentence 121 and the adjusted sentence 122 can each be considered opinion information that reflects the will of an anonymous user.

[0017] The UI 13 is a user interface and has the function of accepting various information or data input by the user and transferring it to the processing unit 11 .

[0018] The server 20 has at least a processing unit 21 and a storage unit 22. The processing unit 21 has a control unit 211 and a base model 212. The storage unit 22 stores a mixed sentence 221 and an output sentence 222. The server 20 may be configured as an on-premise server, a cloud server connected to a network, or any electronic device such as a personal computer. The server 20 may be, for example, a server shared by an organization such as within a company, or a cloud server located outside the organization.

[0019] When the control unit 211 receives adjusted sentences from the user terminal 10, it works with the base model 212 to mix the adjusted sentences and generate an output sentence. Like the control unit 111, the control unit 211 is configured using at least one of a Central Processing Unit (CPU), a Micro Processing Unit (MPU), a controller, a Large Scale Integration (LSI), an Application Specific Integrated Circuit (ASIC), a Programmable Logic Device (), a Field-Programmable Gate Array (FPGA), and a Graphic Processing Unit (GPU).

[0020] The base model 212 is a foundation model, similar to the base model 112. In response to an instruction from the control unit 211, the base model 212 mixes multiple sentences sent from the control unit 211 and outputs the generated output sentence.

[0021] The storage unit 22 stores computer programs and data handled by the processing unit 21. The storage unit 22 may include, for example, a read-only memory (ROM) and a random-access memory (RAM). The mixed sentence 221 stored in the storage unit 22 is a mixed sentence made by mixing a plurality of sentences, and can also be referred to as representative opinion candidate information here. The output sentence 222 is a sentence that is finally output, and can also be referred to as representative opinion information.

[0022] 2. Information generation method Next, the information generation method according to this embodiment will be described in detail with reference to Fig. 2 to Fig. 5. Fig. 2 is a sequence diagram chronologically explaining the information generation method according to this embodiment. Fig. 3 is a diagram contrastively comparing an original post, which is a sentence before tone adjustment, with a sentence after tone adjustment. Fig. 4 is a flow diagram showing a method for generating anonymous group opinions according to this embodiment. Fig. 5 is a diagram contrastively comparing a sentence before mixing with a sentence after mixing.

[0023] The information generation method performed by the server 20 receives opinion information of a plurality of anonymous users in a group, mixes the opinion information of the plurality of anonymous users, and generates representative opinion information of the group using the mixed opinion information of the plurality of anonymous users.

[0024] (1) How to represent input and output data when using language models (LM) and large-scale language models (LLM) First, the method of notating input and output data when using a language model (LM) or a large-scale language model (LLM) is explained below. The language model (LM) and the large-scale language model (LLM) are included in the base models 112 and 212. Here, an example of how to describe input data in the representation of Python program code when calling an Application Program Interface (API) from a Python (registered trademark) program is shown.

[0025] Examples of target sentences are as follows:

[0026] Examples of target sentences: My boss pointed out the safety of the button to delete all data. With the current design, there was a risk that if the button was pressed accidentally, all data would be deleted immediately. A design review found no problems, but... we decided to add a warning and confirmation process to prevent users from accidentally operating the button. Specifically, we changed the button design to a more visible color (red), and also used JavaScript to add a confirmation dialog asking, "Are you sure you want to delete the data?"

[0027] An example of a Python program for calling the API is shown below. The program function is call_llm_json, which internally calls the gpt-3.5-turbo API provided by OpenAI. Enter instructions for LLM in system_prompt. Enter the text you want LLM to process in user_input.

[0028] Example code for a program that calls LLM: system_prompt = """ Please extract the following elements from the notes below: before: Design content before change after: Design content after change Please output in JSON format. The JSON keys should be before and after #Memo """ user_input = """ My boss pointed out the safety of the button to delete all data. With the current design, there was a risk that if the button was pressed accidentally, all data would be deleted immediately. A design review found no problems, but... we decided to add a warning and confirmation process to prevent users from accidentally operating the button. Specifically, we changed the button design to a more visible color (red), and also used JavaScript to add a confirmation dialog asking, "Are you sure you want to delete the data?" """ def call_llm_json(input: str): from openai import OpenAI client = OpenAI() response = client.chat.completions.create( model="gpt-3.5-turbo-1106", messages=[ {"role": "system", "content": system_prompt}, {"role": "user", "content": input}, ], response_format={"type": "json_object"} ) return response.choices[0].message.content

[0029] By running this program, you can obtain the design contents before and after the change in JSON format, which is easy to process in subsequent programs. An example of an LLM response is shown below.

[0030] Example LLM response: { "before": "When the user presses the button, all data is deleted immediately", "after": "The button design will be changed to red and a confirmation dialog will be added using JavaScript asking, 'Are you sure you want to delete this?'" }

[0031] In addition to cloud-provided APIs, private LLMs installed on-premises can also be called. In addition to large-scale language models (LLMs), small-scale LLMs, or language models (LMs), can also be used. Alternatively, language models scaled down to a level that can be run on a local PC or smartphone can also be used. When using a language model executed on a local device, programming languages ​​used for mobile apps, such as Java (registered trademark), C++, Kotlin (registered trademark), Rust (registered trademark), Objective-C (registered trademark), and Swift (registered trademark), may be used instead of the Python code exemplified in this specification. However, the input / output information required when using an LM or LLM and how to use it are not significantly different from the Python code exemplified in this specification.

[0032] Since the above Python code example is long, in this embodiment, the following simplified notation is used:

[0033] Example of LLM instructions (Example of LLM call): Please extract the following elements from the notes below: before: Design content before change after: Design content after change Please output in JSON format. The JSON keys should be before and after #Memo My boss pointed out the safety of the button to delete all data. With the current design, there was a risk that if the button was pressed accidentally, all data would be deleted immediately. A design review found no problems, but... we decided to add a warning and confirmation process to prevent users from accidentally operating the button. Specifically, we changed the button design to a more visible color (red), and also used JavaScript to add a confirmation dialog asking, "Are you sure you want to delete the data?"

[0034] Please note that the above statements such as "Please output in JSON format", "Please use before and after as JSON keys", and "#Memo" may be omitted.

[0035] (2) Stylistic adjustments Next, we will explain stylistic adjustment to improve the anonymity of text. A writing style is a style that reflects the individual characteristics of the writer (opinion inputter). Stylistic adjustment may include, for example, tone adjustment, abstraction, synonym conversion, and reordering of text, and may also include other adjustments. In this embodiment, the processing unit 11 of the user terminal 10 adjusts the text input by the user. The processing unit 11 then transmits the adjusted text 122 to the server 20, so that the server 20 does not acquire or store the text that the user input to the user terminal 10 (i.e., the text in its state before adjustment by the processing unit 11). As a result, even if someone views the data on the server 20, they cannot see the text before adjustment, thereby ensuring the anonymity and security of the text input by the user. Therefore, while style adjustment may be performed by the server 20, it is preferable that it be performed by the user terminal 10.

[0036] Originally, the information generation system 1 includes multiple user terminals 10, each used by multiple users. If each user adjusts their writing style independently, there is a concern that differences will occur in the text after the style adjustment. If the text after the style adjustment is unique to each user, there is a concern that the user may be more likely to be identified and anonymity may be reduced. To avoid this, the organizer or administrator of the survey or chat system conducted by the information generation system 1 may send specific instructions for style adjustment to all user terminals 10. That is, the server 20 sends instructions to adjust the input information of the multiple anonymous users to the user terminals 10, which are the electronic devices of each of the multiple anonymous users, and receives the adjusted input information of the multiple anonymous users as opinion information.

[0037] For example, as shown in FIG. 2, first, the chairperson outputs a style adjustment instruction to the server 20 (step ST1), and the server 20 transmits the style adjustment instruction to each user terminal 10 (step ST2). The chairperson may use the user terminal 10 as a means for outputting the style adjustment instruction to the server 20. After that, when one of the users, person A, inputs a sentence to the user terminal 10 (step ST3), the processing unit 11 of the user terminal 10 adjusts the sentence based on the style adjustment instruction from the server 20 (step ST3). Sentence adjustment includes "tone adjustment," "abstraction," "synonym conversion," "reordering," and the like, which will be described in detail later. Then, as described above, the user terminal 10 transmits the sentence adjusted in step ST3 to the server 20 (step ST4). A series of processes similar to the series of processes from step ST2 to step ST4 is also executed for the sentence of another user (person B) (steps ST5 to ST7).

[0038] Before getting into the main topic, the chairperson may specify that a different tone be adopted than usual, such as using "nya" endings, speaking in an elementary school-aged girl's style, speaking in a gal-like style, or speaking in a passionate style, in order to break the ice. When there are multiple questions or themes in group training or questionnaires, adjusting the tone in this way can be effective as an ice-breaker or as practice to get familiar with the system.

[0039] Below is an example of instructions for the Foundation Model 112 (LLM). This is an example of ending the word with "nya." Example LLM instructions: Please change the following sentences to cat language (ending with "nya"). ###Input text Despite its low price, I think it's a very well-made product. The color is also luxurious. It comes with a face mask, but the only drawback is that it's a bit of a pain to swap out. I'd like to buy another one for my face. Example output from LLM: Despite its low price, I think it's a very well-made product. The color is also luxurious. It comes with a face mask, but the only downside is that it's a bit of a pain to swap out. I almost want to buy another one for my face.

[0040] In this way, the information generation system 1 executes a process for adjusting the writing style found in each user's unique writing input at each user terminal 10, thereby reducing the user's individual writing habits (word preferences and sentence structure preferences) at each user terminal 10, thereby increasing the difficulty of identifying the individual and enhancing anonymity. Furthermore, by facilitating the output of highly anonymous writing, users (participants) can confidently express their opinions in responses to questions set in a questionnaire or in conversations (threads) on an internal chat system. Here, "active opinions" includes objections to traditional organizational policies and comments about behavior that may be considered common sense within the organization but is considered unreasonable in society. It also includes comments about compliance violations, harassment violations, and other behavior that has been tolerated in small organizations, or behavior that was tolerated before changes in laws and regulations or in the past but is now considered problematic. Increasing the anonymity of users' comments in this way lowers the psychological barrier to speaking out, such as raising issues.

[0041] Next, the adjustment of the style of the text will be described in detail.

[0042] a. Tone adjustment

[0043] For style adjustment, the tone of the user's input sentences can be adjusted using LLM, as shown in Figure 3. The four sentences on the right or above the arrow (hereinafter referred to as the "right side of the arrow") in Figure 3 are tone-adjusted versions of the sentences of the four people on the left or below the arrow (hereinafter referred to as the "left side of the arrow").

[0044] Of the four examples of writing by the people to the left of the arrow, the first and second examples from the top use somewhat casual expressions and can be said to contain the user's writing habits. If a sentence containing such writing habits is posted to an internal survey or anonymous message board, there is a possibility that the user who wrote the sentence will be identified. In other words, it can be said that the anonymity is low. Therefore, as in the sentence to the right of the arrow, the processing unit 11 instructs the infrastructure model 112 (LLM) to correct the sentence to adjust the tone of the sentence. Note that, as a tone adjustment, it may be instructed to adjust the tone, style, tone, endings (desu, masu, dearu) etc. to be appropriate for the community where the sentence is to be posted or submitted, such as "Please adjust the expression to be appropriate for an internal chat" or "Please adjust the expression to be appropriate for a gathering of PTA members."

[0045] Below is an example of instructions for the Foundation Model 112 (LLM).

[0046] Example LLM instructions: Please revise each of the following sentences to make them more suitable for business communication. ###Input text <<User input text (multiple)>> ###Output format Output the split sentences in JSON format as an array.

[0047] When the above command is input, the modified sentences can be obtained in an array as shown to the right of the arrow in Figure 3. As shown in Figure 3, the processing unit 11 can also improve the efficiency of LLM processing by adjusting the tone of sentences from multiple users at once. Note that although an example of tone adjustment has been shown here, other style adjustments can also be performed in the same way.

[0048] b. Abstracting and superordinate concepts of proper nouns

[0049] The more detailed the information, the greater the possibility of identifying the user. In particular, the use of technical terms or proper nouns increases the risk of identifying the user, so it is a good idea to abstract or superordinate these terms as a stylistic adjustment. For example, it is a good idea to change specific model names or product names to category names (e.g., Lexus (registered trademark) → automobiles, e.g., certification registration test by XX organization → public certification test). It is also a good idea to change department names to larger organizations (e.g., XX Center, △△ Department, □□ Section → XX Center, △△ Department). Furthermore, if a personal name is included, it can be considered to delete it or replace it with a different name (e.g., Yamada-san → person in charge). The processing unit 11 can instruct the infrastructure model 112 (LLM) to perform these abstractions or superordinate concepts.

[0050] c. Synonym substitution

[0051] As a style adjustment, the processing unit 11 may instruct the platform model 112 (LLM) to perform a synonym replacement process. Because users have preferences and biases in the vocabulary they normally use, automatically converting words into synonyms can enhance anonymity. Note that methods other than LLM may also be used. Synonym replacement can be achieved by using a technology such as WORD2VEC, which represents text vectors, to replace each word in the target sentence with another word with a similar vector. Alternatively, a synonym table may be prepared. A single record in the synonym table contains multiple words with nearly the same meaning. If the target sentence contains a word included in the synonym table, it can be replaced with another word in that record. Alternatively, the processing unit 11 can instruct the platform model 112 (LLM) to create a sentence in which synonyms have been replaced.

[0052] Below is an example of instructions for the Foundation Model 112 (LLM).

[0053] Example LLM instructions: Please replace the following sentences with synonyms to generate another sentence with almost the same content. ###Input text This product is easy to operate with one hand, and the automatic operation means your arm won't get tired, making your daily massage extremely comfortable and providing a blissful experience. Example output from LLM: This product is easy to operate with one hand, and the automatic functioning of the device means no arm fatigue. Daily massage time is extremely comfortable and provides the ultimate relaxation.

[0054] In the above, the substitutions are "machine" to "device," "everyday" to "daily," and "bliss" to "exquisite."

[0055] d. Changing the order

[0056] As a style adjustment, the processing unit 11 may instruct the basic model 112 (LLM) to rearrange the order of the sentences as long as the meaning of the sentences remains unchanged. This can further enhance anonymity.

[0057] Below is an example of instructions for the Foundation Model 112 (LLM). Example LLM instructions: Perform a Random Swap on the following sentences to generate another sentence with roughly the same content. ###Input text This product is easy to operate with one hand, and the automatic operation means your arm won't get tired, making your daily massage extremely comfortable and providing a blissful experience. Example output from LLM: This product's automatic operation eliminates arm fatigue and allows for easy one-handed operation, providing a blissful experience and making daily massages extremely comfortable.

[0058] In the above example, the order of "one-handed operation" and "automatic machine operation" has been swapped, as has the order of "a moment of bliss" and "daily massage time."

[0059] The text adjustments described so far are based on the premise that they are performed in the base model 112 of the user terminal 10, but they may also be performed in the base model 212 of the server 20.

[0060] (3) Generating anonymous group opinions Next, the processing of the control unit 21 of the server 20 after style adjustment in the user terminal 10 will be described. The control unit 21 generates an anonymous group opinion using the adjusted sentences 122 received from the multiple user terminals 10. The anonymous group opinion is generated by mixing the multiple adjusted sentences 122 to generate multiple new sentences.

[0061] The generation of anonymous group opinions will be described in chronological order with reference to FIG. 4. First, the control unit 21 selects a plurality of users and creates an anonymous group of individuals (step ST11). The control unit 21 selects a plurality of users to create one group. In this group, it is preferable that the names of the users and the number of people in the group are unknown. For example, if there are 100 users, the control unit 21 creates 20 anonymous groups of 5 people each. Alternatively, the control unit 21 may create 10 anonymous groups of 10 people each.

[0062] Next, the control unit 21 mixes the sentences of multiple users in the group to generate multiple group opinions (step ST12). A group opinion is a candidate for a representative opinion. As shown in FIG. 5, a group opinion is a new sentence that combines the opinions of multiple people by dividing the opinions of each user into multiple sentences and mixing these sentences. Now, the "division" and "mixing" steps for creating a group opinion will be described in detail.

[0063] a. When dividing using the Basic Model 212 (LLM) Below is an example of instructions for the Foundation Model 212 (LLM).

[0064] Example LLM instructions: Divide the following sentences into information elements. For example, information elements include good points, bad points, problems, what the future should be, and solutions. There may be more than one of them. ###Input text The machine moves even when used with one hand, so you only need to support it and your arm doesn't get tired. Daily massage time is a very comfortable and blissful experience. ###Output format Output the split sentences in JSON format as an array. There is no need to output a viewpoint. Example output from LLM: [ "This product is easy to operate with one hand, and the machine moves automatically, so your arm doesn't get tired." "My daily massage is very comfortable and provides a blissful experience." ]

[0065] As described above, the control unit 21 can divide the opinion of one user into multiple opinions by instructing the basic model 212 (LLM).

[0066] b. When dividing by punctuation marks

[0067] Even if there is no instruction to the Foundation Model 212 (LLM), if the text is Japanese, you can separate it using punctuation marks (,.). If the text is English or other languages, you can separate it using periods.

[0068] c. Mixing with the Foundation Model 212 (LLM)

[0069] Below is an example of instructions to the basic model 212 (LLM). Here, an already segmented sentence is input.

[0070] Example LLM instructions: Please generate six sentences that mix the sentences from A to D below. Make sure your mixed sentences contain sentences of three or more people. ###Input text Person A: "When used for a long time, it can cause some irritation to the scalp." Person A: "Overall, the feeling of use is good." Person A: "If your hair is long, it may get a little tangled." Person A: "I am satisfied with my purchase." Person B: "Sometimes I wish I had a little more strength." Person B: "The current strength may be appropriate for people with long hair." C: "For the price, this product is very well made." C: "The color has a luxurious feel and I'm satisfied with the design." C: "It's a bit of a hassle to change the face attachments." C: "I'm also considering purchasing face products separately." D: "This product is easy to operate with one hand, the machine moves automatically, and my arm doesn't get tired." D: "My daily massage is very comfortable and provides a blissful experience." ###Output format Output in JSON format of the array. The elements of the array should include the sender (user) and mixed document (mixed). Example output from LLM: [ { "user": ["D", "C", "B"], "mixed": "My daily massage session is extremely comfortable and provides a moment of bliss. For the price, this product is very well made. The current strength may be appropriate for those with long hair." }, { "user": ["A", "C", "D"], "mixed": "After extended use, the scalp may feel a little irritated. For the price, this product is very well made. It's easy to operate with one hand, and the automatic mechanism keeps your arm from getting tired." }, { "user": ["B", "A", "C"], "mixed": "Sometimes I wish it had a bit more power. If you have long hair, it can get a little tangled. Changing the face attachment is a bit of a pain." }, { "user": ["D", "B", "C"], "mixed": "This product is easy to use with one hand, the mechanism is automatic, and it doesn't tire your arm. Sometimes I wish it had a bit more strength. For the price, this product is very well made." }, { "user": ["D", "A", "B"], "mixed": "This product is easy to use with one hand, and the automatic mechanism saves your arm. However, if you have long hair, it can get a little tangled. I sometimes wish it had a bit more power." }, { "user": ["B", "D", "C"], "mixed": "If you have long hair, the current strength may be appropriate. This product is easy to operate with one hand, and the mechanism works automatically, so your arm doesn't get tired. The color has a luxurious feel, and I'm pleased with the design." } ]

[0071] By issuing instructions to the basic model 212 (LLM) as described above, it is possible to generate a sentence that combines the divided opinions of multiple users.

[0072] d. Random mixing

[0073] The control unit 21 may randomly select and combine sentences from each user's divided sentences to create multiple mixed sentences 221. Since the foundation model 212 (LLM) may occasionally fail to follow instructions due to probabilistic reasons, processing the process programmatically can increase the reliability of the mixture. Since shuffling and combining programmatically can result in unnatural sentence connections, the control unit 21 may instruct the foundation model 212 (LLM) to proofread the combined document.

[0074] e. Mixing similar sentences

[0075] The control unit 21 may calculate the similarity of sentences using WORD2VEC and cosine similarity, and generate more similar sentences as the mixed sentence 221. If there is too much diversity, there is a concern that some sentences will not pass the representative opinion selection process described below, so mixing similar sentences in this way is effective.

[0076] f. When dividing and mixing at the same time using the Basic Model 212 (LLM)

[0077] Below is an example of instructions to the basic model 212 (LLM). Here, the sentence before division is input.

[0078] Example LLM instructions: Divide the following sentences from A to D into information elements, then create six mixed sentences. For example, information elements include good points, bad points, problems, what the future should be, and solutions. Mixed documents should always contain at least three authors. ###Input text Person A: "I'd like to share a few thoughts about my experience using the product. It can be a little irritating to the scalp after prolonged use, but overall it feels good to use. If you have long hair, please be aware that it may tangle a little. However, I'm very satisfied with my purchase." Person B: "Personally, I wish it was a little stronger. However, the current strength may be appropriate for people with long hair." C: "For the price, I think this product is very well made. The color is luxurious and I'm happy with the design. However, I think there's room for improvement in the fact that it's a bit of a hassle to change the face attachment. I'm considering purchasing a face attachment separately." D: "This product is easy to operate with one hand, and the automatic operation means my arm doesn't get tired. My daily massage is very comfortable and provides a blissful experience." ###Output format Output in JSON format of the array. The elements of the array should include the sender (user) and mixed document (mixed).

[0079] By inputting the above, it is possible to obtain the same results as the output obtained when mixing is performed using c. the base model 212 (LLM).

[0080] Although the above explanation shows an example of mixing sentences from four people, sentences from multiple people can be mixed. Also, when sentences from 100 people are given, it is possible to mix all 100 people, or to form small groups of 3 to 5 people and create a combined document within those groups. Furthermore, sentences written by the same user from different perspectives can be treated as opinions from different users and mixed together.

[0081] As a result of step ST12, the result shown on the right side of FIG. 5 is obtained.

[0082] (4) Determining the group's representative opinion (=anonymous group opinion)

[0083] Next, the control unit 21 determines the representative opinion of the group (=anonymous group opinion) (step ST13). In this step, the control unit 21 selects from a plurality of candidates of the mixed document by a method such as majority vote. In other words, the group representative opinion information may be generated by selecting one of the representative opinion candidate information of a plurality of groups generated from the opinion information of a plurality of mixed anonymous users. In addition, the control unit 21 may output at least one of the following a to e as a selection method for selecting the group representative opinion information from the representative opinion candidate information of a plurality of groups. Here, for example, the explanation will be given assuming that the sentences of four people, A to D, are mixed to create mixed sentences 221 M1 to M6.

[0084] a. Majority vote

[0085] A case where majority voting is used as a method for determining the representative opinion of the group (=anonymous group opinion) will be described. The control unit 21 has four people each select one of the mixed sentences 221 M1 to M6 that they think is best, and input this into their respective user terminals 10. The server 20 receives this input information and determines the mixed sentence 221 that was selected by the most people as the representative opinion of the group (=anonymous group opinion).

[0086] b. Marking method

[0087] A case where a scoring system is used as a method for determining the group's representative opinion (=anonymous group opinion) will be described. The control unit 21 rearranges the mixed sentences M1 to M6 of the mixed sentences 221 in the order that each of the four people consider most appropriate, and assigns scores to them. In this example, there are six sentences, so six points are added to the best sentence, and one point is added to the sentence with the lowest ranking. Each user inputs their score into their respective user terminal 10, and the server 20 selects the mixed sentence with the highest total score from the received input information as the group's representative opinion (=anonymous group opinion).

[0088] c.All consent method

[0089] A case where a unanimous agreement method is used to determine the group's representative opinion (=anonymous group opinion) will be described. Each user selects whether or not they agree (Agree, Disagree) for each of M1 to M6 of the mixed sentences 221 and inputs this into the user terminal 10. If there is one of the six sentences that all four people agree on, the control unit 21 selects it as the group's representative opinion. If all people agree on multiple mixed sentences 221, the control unit 21 may randomly select one of them, or may further perform the majority vote described above. If there is no candidate that all four people agree on, the anonymous group will not output a representative opinion.

[0090] d. A third party intervenes in the selection

[0091] We will now explain how to determine the group's representative opinion (=anonymous group opinion) by having a third party intervene in the selection. Aside from the opinions of A to D, person E (a third party) can also participate in the selection process. This is expected to reflect a greater variety of perspectives. However, if a specific person is involved in the selection of all mixed documents, that person's intentions will be strongly reflected. Therefore, it is best for the third party, person E, to be randomly selected from among people other than A to D. Person E inputs the selection results into user terminal 10, and the input information is sent from user terminal 10 to server 20, where it is received and processed.

[0092] You can also ask a second third party, a Foundation Model 212 (LLM), to participate in the selection process. Generally, LLMs are tuned to behave ethically and are therefore often suitable as a third party. You can also ask an LLM who has studied the statements and thoughts of a specific person, such as a founder or president, to participate in the selection process.

[0093] As described above, through the series of processes from step ST11 to step ST13, the server 20 divides and combines each user's opinion, enabling the generation of an anonymous group opinion. This enables democratic formation of a collective opinion in which the entire group's opinion is formed while maintaining the anonymity of each user's opinion. In this information generation system 1, multiple users are first selected to create a group in which the names and number of participants are unknown. Next, the opinions provided by each user are combined to create multiple candidates for the group's representative opinion. Finally, users select candidates that are closest to their own opinion by majority vote or sorting, and the server 20 aggregates these selections to determine the representative opinion for the entire group. This process encourages users to post honest opinions, suggestions, and information about harassment and compliance violations, which they would normally hesitate to post. Anonymity provides an environment in which individual users can freely express their opinions. As a result, there are more opportunities for diverse opinions and new proposals, which have traditionally been rarely expressed, to be exchanged within the organization. As a result, diversity of opinion within the organization is maintained, different perspectives are accepted, and the organization becomes more open.

[0094] If the agenda is set in advance, it is advisable to create a representative opinion candidate from style-adjusted and mixed sentences using only the information entered by each user before the group is formed. This reduces the psychological and time costs for users.

[0095] (5) User selection in groups Next, we will explain how to select users to be assigned to each group. Since the sentences of multiple users in each group are mixed together to form a representative opinion, the method of selecting users is also important. Below, we will explain the criteria that the server 20 uses to select users for each group.

[0096] a. Expertise and diversity

[0097] One of the goals of this embodiment is to improve innovation and problem-solving processes within an organization, particularly by placing organizational members with different areas of expertise and diverse backgrounds in the same group.

[0098] Traditional methods for internal organizational communication and idea discussion involve gathering opinions at the organizational level, so the opinions of members with similar fields of expertise are first collected. This increases specialization, but it has limitations in terms of creativity and diversity. On the other hand, there are methods that bring together members with diverse backgrounds, but differences in each other's backgrounds can lead to differences in word definitions and gaps in understanding of issues, and when people do not know each other, it is difficult to make in-depth comments until they have determined the appropriate level of distance. For these reasons, there is the problem that it takes too much time.

[0099] For example, imagine an organization of 100 people, with 10 members divided into two separate groups of five. These groups discuss a certain issue. The first group is made up of people from the same field or occupation, and they hold specialized discussions. The second group is made up of members with different attributes, and they aim to generate ideas by leveraging diversity. In the first group, advanced solutions and ideas are generated through the sharing of specialized knowledge and deep discussions. In the second group, creative and innovative ideas are generated through dialogue between members with different backgrounds. For example, relationships between different members could be those between sales and development, managers and employees, or veterans and junior employees. In this way, interaction between members with different backgrounds provides innovative perspectives that are often overlooked using traditional methods.

[0100] The second group tends to misinterpret words, but adjusting the tone of the text can help facilitate this. Also, each user within the group can be anonymous. This eliminates the need for hesitation, as they won't know who they're talking to. This means they can skip the step of gauging the distance between themselves and the other person and express their own opinions.

[0101] In this way, when dividing an organization into groups, it is not necessary to divide all groups using the same indicators.

[0102] b. Homogeneity and diversity

[0103] Here, two groups may be created: homogeneous groups and divergent groups, based solely on the user's adjusted sentences. For example, the user's sentences are converted into vector representations using the WORD2VEC or DOC2VEC methods. Next, users with similar vectors are grouped (merged). The similarity of the vectors can be quantified, for example, using cosine similarity. If the target number of people in a group is five, groups with more than five people will be excluded from merging. In this way, users who output similar sentences will belong to one of the groups with a target number of around five people.

[0104] Diverse groups can also be created in parallel within an organization. These can simply be groups of randomly selected users. Furthermore, the mutual similarity of sentences in randomly created groups can be calculated, and if multiple similar sentences exist, they can be swapped with a user from another group. For example, cosine similarity ranges from -1 to 1, with values ​​of 0.5 or greater considered similar. If the similarity between sentences of a given user in a given group is 0.5 or greater, an exchange attempt is made with a user from another group. This exchange swaps the sentences of the users belonging to the two groups. If the number of occurrences of similarities of 0.5 or greater between the two groups is lower than before the exchange, the exchange is accepted. Conversely, if the number of occurrences of 0.5 or greater remains the same or increases, the exchange has not decreased the similarity of the sentences, and the exchange is not carried out. This exchange attempt ends under one of the following three conditions: the first is when a predetermined number of attempts is reached; the second is when the exchange attempt is not accepted more than a certain number of times in a row. The third is that no similarity of 0.5 or more occurs between any of the user sentences in any of the groups.

[0105] Note that two groups (a) based on expertise and diversity and two groups (b) based on homogeneity and diversity may exist simultaneously. For example, suppose that A, B, C, and D are assigned to one group based on similar occupations (expertise). At the same time, A, M, O, and Y are assigned to one anonymous group based on the diversity of their writing. In this case, in step ST13 of FIG. 4, A is presented with multiple candidate sentences generated from users A, B, C, and D, as well as multiple candidate sentences generated from users A, M, O, and Y. The former and latter may be presented separately to A, or they may be presented mixed together to further enhance anonymity. This allows A to select from each group only once.

[0106] Also, if a person expresses an opinion that is unusual even within the organization, it is unlikely that that opinion will be selected as the representative opinion in step ST13 of Figure 4. However, there is a possibility that this will waste the time of other users making their selections. In this case, it may be possible to allow the person to request the re-creation of a group up to a maximum number of times set by the server 20, organizer, administrator, etc.

[0107] In this way, it is advisable to select a plurality of users with mutually different attributes as the plurality of anonymous users in the group.

[0108] (6) About the survey If the information generation system 1 is used, a highly anonymous survey can be conducted. When an organizer conducts an anonymous survey to elicit honest opinions, conventional methods do not allow the organizer to know who has responded. This has led to a problem in that the organizer is unable to ask questions of the commenter or hold a discussion about the comments written in the survey. If the information generation system 1 of this embodiment is used, the information generation system 1 can know the identity of the user who created the original text of the mixed text 221, even if the organizer does not know it. This allows communication between the organizer and the user (such as a reply from the organizer or a discussion).

[0109] a. If you want to communicate with the author of a mixed text

[0110] The creator (target user) of the mixed sentence 221 is, for example, Mr. A and Mr. D in the top sentence after mixing in FIG. 5. It is recommended that at least a button such as "Reply" be placed around the display of this mixed sentence 221 on the UI 13 of the organizer's user terminal 10. It is also recommended that when the organizer presses the button, the organizer can start a group chat with Mr. A and Mr. D. In this way, when the server 20 receives information via the user terminal 10 from the organizer (or other user) requesting communication with multiple anonymous users (here, Mr. A and Mr. D) involved in the representative opinion information, it transmits information related to communication to the multiple anonymous users involved in the representative opinion information. "Communication" refers to sending a questionnaire (question) that allows anonymous conversation and asking for responses, or opening a chat room, talk room, etc. "Transmitting information related to communication" includes inviting or starting a questionnaire, chat, talk, etc.

[0111] b. When you want to communicate to a group that has created a mixed text

[0112] Since it is not always possible to reliably identify the creator of a representative sentence included in the mixed sentence 221, it is also possible to start communication with all members of the group who created that representative sentence. In the case of Figure 5, when the organizer receives information that the topmost sentence after mixing has been selected, the server 20 opens a chat or the like with the organizer and all members A through D, and communication begins. Naturally, communication is anonymous.

[0113] c. Group communication

[0114] The details of group communication between the organizer and multiple target users as described above will now be explained. At this time, the organizer's and target users' text may be adjusted by the control units 111, 211, etc. The server 20 initiates a group chat involving the organizer, A, and D. This group chat is anonymous and the number of participants is unknown, meaning neither party knows who or how many other people are in the group. A and D are not obligated to reply to the organizer's additional text. In response to an instruction from the server 20, the UI 13 of the target user's user terminal 10 may display a warning message stating, "The organizer wishes to open a group chat. Even the organizer does not know who else is participating in this group chat or how many people are participating. By engaging in a conversation through this chat, there is a high risk that your identity may be identified. Chat responses with the organizer are not mandatory." Additionally, a message such as, "You can hide or leave this group chat. The organizer and other users will not know that you are hidden or have left," may be displayed to inform users that they can safely opt out of the group chat.

[0115] d. In the case of individual communications

[0116] While the above discussion focused on group communication between the organizer and all target users, individual communication between the organizer and each target user is also possible. In this case, the organizer's and target user's messages may be adjusted by the control units 111 and 211. The organizer's message is sent individually to A and D. Instead of a group chat, individual private chats are initiated between the organizer and A and between the organizer and D. The target user is not obligated to reply to the organizer's additional messages. In response to an instruction from the server 20, the organizer's message may be displayed on the UI 13 of the target user's user terminal 10, along with a warning message such as, "The organizer wishes to open a one-on-one private chat. One-on-one conversations carry a high risk of revealing your identity. Chat responses with the organizer are not mandatory." The target user can indicate whether or not they wish to accept the private chat using the OK / No button on the UI 13.

[0117] e. When communication is a closed question

[0118] To ensure the anonymity of the target user, there may be a mode in which communication from the organizer is limited to closed questions. In other words, this is an operating mode in which the target user's output in UI 13 is limited to Yes / No. In addition to Yes / No, options such as "Not sure" and "Neither" may be added to make it three or four choices. This eliminates the need for the user to write a sentence, reducing the risk of the user being identified from their writing habits.

[0119] f. When LLM generates answer candidates for the target user

[0120] To ensure the anonymity of the target user, the base model 112, 212 (LLM) may generate multiple candidate answer sentences for the target user, from which the target user may select. This mode may be used simultaneously with the closed question mode. That is, if the base model 112, 212 (LLM) generates, for example, three candidate answer sentences, the three answer sentences and two to four closed question responses are displayed on the UI 13 of the user terminal 10, and the user can simply select one of these five to seven to reply to the organizer. Since the user does not create the sentences themselves, the risk of the user being identified from the user's writing habits can be reduced. The candidate answer sentences for the target user may be created by at least one of the base models 112, 212.

[0121] The explanation so far has been based on the relationship between the organizer of a survey and the target users, but if we think of this in a company organization, it can also be applied to communication between superiors and subordinates. For example, the manager, who is the superior, is the organizer and can write any text in the chat, but the infrastructure model 112 (LLM) processes the text to prevent subordinates from knowing who is making the comments. Subordinates' comments and chats can also be written using the selective input method mentioned above, rather than by freely entering text. Subordinates can hide or leave the chat without the manager knowing.

[0122] In addition, an organization may set a theme and have multiple department heads each engage in anonymous chats with their department members. The number of comments made by subordinates is considered to be proportional to the trust between the department head and the department head. Therefore, the president and human resources department may use the number of comments made by subordinates as a metric representing the department head's trustworthiness. In this anonymous chat environment, where the department head does not exercise authority, subordinates feel less psychologically obligated to respond to questions from the department head, and therefore tend to respond with honest enthusiasm. If the number of comments and responses made by subordinates in this system is low, it may be possible to identify problems in the department, particularly those caused by the department head's personality. However, if the theme set for the entire company is specialized, there may be variations, such as some departments receiving high interest and others receiving low interest. In addition to the department head's personality, there are other factors that can affect the frequency of subordinates' comments.

[0123] Since the chat is anonymous, prompting those who have not responded should be carried out by the information generation system 1. The server 20 should transmit information prompting the user terminals 10 of the users who have not responded to respond on the day before or a few hours before the response deadline set by the organizer.

[0124] Furthermore, when conducting a survey using the information generation system 1, numerical information such as the number of subjects and the number of people who completed the survey can be displayed on each user's UI 13, and push notifications can be sent via chat or email periodically before the response deadline. In an organization with a strong sense of isolation, even if there is a chat room where anyone in the organization can speak, only a very small number of people actually speak there. In other words, the majority of people are uninterested and never look at the chat room or speak. On the other hand, in the case of a survey, the organizer may be able to increase the number of responses by calling for responses. Therefore, by sharing the survey response status with not only the organizer but also the users, users' motivation to fill out the survey can be increased.

[0125] Furthermore, to increase motivation to answer the questionnaire, rather than adjusting, dividing, or mixing each user's response into groups after the deadline for answering the questionnaire has passed, these processes may be performed in order starting with the users who have completed their responses. Each time a representative opinion of a group is determined, it may be displayed in real time in a designated chat thread. By displaying the opinion of the group of people who responded quickly, the remaining users can use it as a reference to respond more easily. Note that instructions on what to display on the UI 13 described above may be given by the processing unit 11 of the user terminal 10 or the processing unit 21 of the server 20.

[0126] To improve real-time performance, the organizer can set a theme and make it known to everyone that they are obligated to respond. Suppose that users A, B, C, D, E, F, G, H, I, and J post in order on a certain theme. Here, we create anonymous groups based on five users. The first users, A, B, C, D, and E, are placed in one anonymous group based on their posting order. Once a representative opinion for this anonymous group is determined, server 20 outputs the selected representative opinion to the chat thread set as the output destination for this theme. Since anonymity is the goal, the user whose opinion created or was selected as the representative opinion is not displayed. Next, another anonymous group is created with the next five users, F, G, H, I, and J, and a representative opinion is selected. Once the representative opinion is determined, it is displayed in the chat thread as well. In other words, creating anonymous groups in the order of posting improves real-time performance. In addition, groups can be created in parallel that bring together users with similar attributes or users whose written content is highly similar. Rather than waiting until everyone has finished posting before starting grouping, it is possible to create groups of highly similar users from a slightly larger number of people, for example, in units of 10 or 20 people, or several times that number if the basic number is 5 people.

[0127] Also, when chat postings are not active, you can use a sliding window. That is, create an anonymous group with A, B, C, D, and E. When postings in that chat are not active, create an anonymous group with B, C, D, E, and F. Then create an anonymous group with C, D, E, F, and G. In this way, when postings in the chat are not active, it is a good idea to create anonymous groups by shifting the members one by one in chronological order.

[0128] In this way, when multiple groups are created, the server 20 creates the groups using information regarding the order in which opinion information from multiple anonymous users is received, and outputs representative opinion information from the group for which representative opinion information was generated. The representative opinion information is included in the output text 222.

[0129] Furthermore, if real-timeness is enhanced, discussions can be held via chat. In other words, while a normal survey only allows one response per person, it is possible for one person to make multiple comments. However, if the same person posts repeatedly within a short period of time, it is desirable to take measures to prevent that person's comments from dominating the chat. For example, if multiple posts are made, the comments are linked together to form a single post. Furthermore, if multiple posts are made, only the most recent posting time is retained. Basically, an anonymous group is created from users who have posted the oldest, so if there are multiple posts, it is best to use the most recent one to reduce the exposure of people who post too much.

[0130] Additionally, various chat tools allow a title and summary to be written at the beginning of each thread. In the case of information generation system 1, including when using a survey, it is advisable to link the representative opinion of a determined anonymous group to these threads and automatically output it. At this time, the title and summary at the beginning of the thread may be automatically updated to the latest status using LLM. For example, the summary field may display a summary of the representative opinion. The summary field may also display responses and the number of users who have posted. The organizer can set whether or not to enable automatic updates.

[0131] <About the technology of the present disclosure> As described above, the present disclosure discloses the following technical ideas. (Item 1) The system includes a processing unit (processing unit 11) and a memory unit (memory unit 12), and the processing unit cooperates with the memory unit to: receiving opinion information from a plurality of anonymous users in the group; The opinion information of the plurality of anonymous users (user input sentences 121 or adjusted sentences 122) is mixed; generating representative opinion information of the group (output sentence 222) using the mixed opinion information of the plurality of anonymous users (mixed sentence 221); Information generation method. This allows users of any position to make statements (disseminate information) with improved anonymity and security.

[0132] (Item 2) The representative opinion information of the group (output sentence 222) is generated by selecting one of the representative opinion candidate information of the plurality of groups (mixed sentence 221) generated from the mixed opinion information of the plurality of anonymous users (user input sentence 121 or adjusted sentence 122). Item 1. The information generation method according to item 1. This allows users of any position to make statements (disseminate information) with improved anonymity and security, and also allows for the selection of representative opinions that are more widely accepted.

[0133] (Item 3) The processing unit Dividing the opinion information to obtain a plurality of divided opinion information; Mixing the plurality of divided opinion information pieces; generating representative opinion information of the group using the mixed opinion information of the plurality of anonymous users; Item 1. The information generation method according to item 1. This allows users of any position to make statements (disseminate information) with improved anonymity and security by dividing and then mixing opinion information.

[0134] (Item 4) The processing unit inputting the opinion information of the plurality of anonymous users and instruction information for generating opinion information in which the opinion information of the plurality of anonymous users is mixed into a base model which is a language model; Obtaining opinion information generated by the base model, which is a mixture of opinion information of the plurality of anonymous users; generating representative opinion information of the group using opinion information obtained by mixing opinion information of the plurality of anonymous users; Item 1. The information generation method according to item 1. This allows a user of any position to efficiently generate representative opinion information for the group to which the user belongs, with improved anonymity and security, by utilizing the basic model, which is a language model.

[0135] (Item 5) The processing unit outputs a selection method for selecting representative opinion information (output sentence 222) of the group from representative opinion candidate information (mixed sentence 221) of the plurality of groups. Item 2. The information generation method according to item 2. This allows users of any position to make statements (disseminate information) with improved anonymity and security, and also allows representative opinions to be selected in a more appropriate manner.

[0136] (Item 6) the processing unit selects a plurality of users having different attributes as the plurality of anonymous users in the group. Item 5: The information generating method according to any one of items 1 to 5. This allows users of any position to make statements (disseminate information) with improved anonymity and security, and also allows for the generation of more effective representative opinions.

[0137] (Item 7) The processing unit When other users receive information requesting communication with the plurality of anonymous users regarding the representative opinion information (output text 222), Sending information about the communication to the plurality of anonymous users related to the representative opinion information (output text 222); Item 6: The information generating method according to any one of items 1 to 6. This allows users of any position to make statements (disseminate information) with improved anonymity and security, enabling communication while maintaining that anonymity and security.

[0138] (Item 8) The processing unit When a plurality of groups are provided, A plurality of groups are formed using information regarding the order in which opinion information of a plurality of anonymous users is received, and the representative opinion information is output from the group for which the representative opinion information was generated among the plurality of groups. Item 5: The information generating method according to any one of items 1 to 5. This allows users of any position to make statements (disseminate information) with improved anonymity and security, thereby stimulating communication.

[0139] (Item 9) The processing unit Transmitting input information of the plurality of anonymous users and instruction information for adjusting the input information of the plurality of anonymous users (user input sentences 121) to a base model, which is a language model possessed by each electronic device (user terminal 10) of the plurality of anonymous users; receiving, from the foundation model, the adjusted input information of the plurality of anonymous users generated by the foundation model; Instruct the input information to be displayed as the opinion information (adjusted text 122); Item 6: The information generating method according to any one of items 1 to 6. This allows the system to display adjusted input information from multiple anonymous users, generated using a language model called the base model, enabling users of any position to make statements (disseminate information) with improved anonymity and security.

[0140] (Item 10) a processing unit and a storage unit, The processing unit cooperates with the storage unit to receiving opinion information from a plurality of anonymous users in the group; Mixing the opinion information of the plurality of anonymous users; generating representative opinion information of the group using the mixed opinion information of the plurality of anonymous users; Information generation device. This allows users of any position to make statements (disseminate information) with improved anonymity and security.

[0141] (Item 11) On the computer, receiving opinion information from a plurality of anonymous users in the group; Mixing the opinion information of the plurality of anonymous users; generating representative opinion information of the group using the mixed opinion information of the plurality of anonymous users; program. This allows users of any position to make statements (disseminate information) with improved anonymity and security.

[0142] Although various embodiments have been described above with reference to the accompanying drawings, the present disclosure is not limited to such examples. It is clear that those skilled in the art can conceive of various modifications, alterations, substitutions, additions, deletions, and equivalents within the scope of the claims, and it is understood that these also fall within the technical scope of the present disclosure. Furthermore, the components of the various embodiments described above may be combined in any manner without departing from the spirit of the invention. [Industrial Applicability]

[0143] The present disclosure is useful as a proposal for an information generation system and an information generation device that enable users of any position to make statements (transmit information) with improved anonymity and security. [Explanation of symbols]

[0144] 1 Information Generation System 10 User terminal 11 Processing section 111 Control Unit 112 Base Model 12 Storage section 121 User-input text 122 Adjusted text 20 servers 21 Processing section 211 Control Unit 212 Base Model 22 Memory section 221 Mixed sentences 222 Output text

Claims

1. a processing unit and a storage unit, The processing unit cooperates with the storage unit to receiving opinion information from a plurality of anonymous users in the group; Mixing the opinion information of the plurality of anonymous users; generating representative opinion information of the group using the mixed opinion information of the plurality of anonymous users; Information generation method.

2. the representative opinion information of the group is generated by selecting one of candidate representative opinion information of a plurality of groups generated from the mixed opinion information of the plurality of anonymous users; The information generating method according to claim 1 .

3. The processing unit Dividing the opinion information to obtain a plurality of divided opinion information; Mixing the plurality of divided opinion information pieces; generating representative opinion information of the group using the mixed opinion information of the plurality of anonymous users; The information generating method according to claim 1 .

4. The processing unit inputting the opinion information of the plurality of anonymous users and instruction information for generating opinion information in which the opinion information of the plurality of anonymous users is mixed into a base model which is a language model; Obtaining opinion information generated by the base model, which is a mixture of opinion information of the plurality of anonymous users; generating representative opinion information of the group using opinion information obtained by mixing opinion information of the plurality of anonymous users; The information generating method according to claim 1 .

5. The processing unit outputting a selection method for selecting representative opinion information of the group from the representative opinion candidate information of the plurality of groups; The information generating method according to claim 2 .

6. The processing unit Selecting a plurality of users having different attributes as the plurality of anonymous users in the group; 3. The information generating method according to claim 1 or 2.

7. The processing unit When another user receives information requesting communication with the plurality of anonymous users related to the representative opinion information, transmitting information about the communication to the plurality of anonymous users related to the representative opinion information; 3. The information generating method according to claim 1 or 2.

8. The processing unit When a plurality of groups are provided, A plurality of groups are formed using information regarding the order in which opinion information of a plurality of anonymous users is received, and the representative opinion information is output from the group in which the representative opinion information is generated among the plurality of groups.

3. The information generating method according to claim 1 or 2.

9. The processing unit Transmitting input information of the plurality of anonymous users and instruction information for adjusting the input information of the plurality of anonymous users to a base model, which is a language model possessed by each electronic device of the plurality of anonymous users; receiving, from the foundation model, the adjusted input information of the plurality of anonymous users generated by the foundation model; instructing to display the input information as the opinion information; 3. The information generating method according to claim 1 or 2.

10. a processing unit and a storage unit, The processing unit cooperates with the storage unit to receiving opinion information from a plurality of anonymous users in the group; Mixing the opinion information of the plurality of anonymous users; generating representative opinion information of the group using the mixed opinion information of the plurality of anonymous users; Information generation device.

11. On the computer, receiving opinion information from a plurality of anonymous users in the group; Mixing the opinion information of the plurality of anonymous users; generating representative opinion information of the group using the mixed opinion information of the plurality of anonymous users; Information generating program.

Images