Data collecting system, and data collecting method

Abstract

To provide a data collecting system and a method which also collect response data of a user who has not spoken of his or her opinion on the Internet.SOLUTION: A fair data collecting system 1000 comprises: a user selection unit which selects a second user who corrects bias with respect to contribution, on the basis of an article and element information about a first user, and a database storing the second user having a predetermined element different from that of the first user, when receiving request which corrects the bias with respect to the contribution of the first user to the article of a predetermined site; a daily conversation generation unit which summarizes a content of the article, creates a question sentence using daily conversation with respect to the second user, about the summarized content, and acquires response data with respect to the created question; and a proxy contribution unit which contributes the response data on behalf of the second user, to a contribution destination of the first user.SELECTED DRAWING: Figure 3

Description

[Technical field]

[0001] The present invention relates to a data collection system and a data collection method. [Background technology]

[0002] As AI (Artificial Intelligence) is increasingly implemented in society, the level of demand for transparency and fairness in AI models is rising, which is starting to affect the trust in AI service providers. AI is trained using large amounts of data, so if there is bias in the training data, the AI ​​model will also be biased. Bias in an AI model can cause AI ethics issues. For example, an issue arose when an interview AI learned past hiring trends and produced unfair results depending on race and gender. This is an example of an issue involving AI ethics. In terms of how to provide fair data to AI, the direction of regulations is being considered, especially in Europe.

[0003] On the other hand, large-scale language models (LLMs) such as GPT (Generative Pre-trained Transformer) have steadily improved their performance, and are now able to handle a variety of tasks, including those performed by professionals. In 2022, ChatGPT was released, making it possible to use LLMs in a natural language dialogue format rather than an API (Application Programmable Interface), and it is being used by a wide range of users and is being used in a variety of applications. LLMs improve their performance by training huge models with huge amounts of data, and as society pays more attention to the issue, transparency in the operation of AI models and the collection of unbiased data will become issues that society as a whole must address.

[0004] Biased data and applications are mainly related to communication between people. For example, news may sometimes be mixed with an intention to steer public opinion, and on social media, people may express opinions that are likely to be popular with the masses out of a desire to have their own ideas recognized. When analyzing sentiment from data learned using such data, it is easy for bias to creep in, so care must be taken. One possible measure is to collect information from a variety of people and ensure fairness in the sense that "what everyone thinks is correct."

[0005] News portal sites provide information to help users interpret trends by learning users' sentiment data and displaying trends using an AI model that supports judgment. As an extension of this mechanism, Patent Literature 1 discloses a method for outputting sentiment data results using an interactive system even for information that is difficult to judge when entered by the user. [Prior art documents] [Patent documents]

[0006] [Patent Document 1] JP 2020-170239 A Summary of the Invention [Problem to be solved by the invention]

[0007] From the perspective of AI ethics, disclosure of what data was used and quality assurance against bias are important. In particular, while LLMs are highly powerful, the learning scale is huge, and at present, there are high hurdles in terms of re-learning and implementation. However, it is becoming clear that LLMs of a certain scale or larger can replace human work. Therefore, it is clear that companies will embark on this by limiting the model scale to the minimum necessary, subject to hardware constraints. In that case, from the perspective of AI ethics, it will be important to have a framework for transparency regarding the model and to be able to clearly declare the legitimacy of data acquisition.

[0008] Portal sites and SNS platforms that have collected information from users in the past have very strong assets in terms of the amount of data. On the other hand, the information published on the Internet is collected on the premise that users actively disseminate information. Therefore, the information found on portal sites and SNS differs in the attributes of the users gathered on each platform, and the collected data is in line with the preferences of a specific user group. This means that the tendency of the AI ​​model changes depending on which site the data obtained from is learned from. Therefore, there is a possibility of bias being mixed in. The tendency of users gathered on each platform is somewhat clear from questionnaire surveys, so it is a good idea to collect data that complements the questionnaire survey. In this case, there are two problems: the fact that the data available on each site is biased, and the fact that there are users who express opinions on the Internet and those who do not, both of which need to be addressed. In order to mitigate the tendency of this learning data to acquire site-specific tendencies, it is preferable to approach users who do not express opinions and actively collect sentiment analysis.

[0009] An object of the present invention is to provide a technique that makes it possible to collect reaction data from users who have not expressed their own opinions on the Internet. [Means for solving the problem]

[0010] The data collection system of the present invention is a data collection system implemented by a computer, and when requested to correct a bias in a post by a first user to an article on a specified site, the data collection system is characterized in having: a user selection unit that selects the second user for correcting the bias in the post based on specification information on the article and the first user and a database storing a second user having specifications different from that of a predetermined first user; a daily conversation generation unit that summarizes the content of the article, creates questions for the second user about the summarized content through daily conversation, and acquires reaction data to the created questions; and a proxy posting unit that posts the reaction data to the posting destination of the first user on behalf of the second user. Effect of the Invention

[0011] According to the present invention, it is possible to collect reaction data from users who have not expressed their own opinions on the Internet. [Brief description of the drawings]

[0012] [Figure 1] Overall Processing [Diagram 2] Diagram explaining the sequence [Diagram 3] Diagram explaining the functional blocks [Figure 4A] Diagram explaining database [Figure 4B] Diagram explaining database [Figure 4C] Diagram explaining database [Figure 5A] Diagram explaining the user interface [Figure 5B] Diagram explaining the user interface [Figure 5C] Diagram explaining the user interface [Figure 6A] A diagram explaining an example of a posting by a proxy posting bot [Figure 6B] A diagram explaining an example of a posting by a proxy posting bot [Figure 6C]A diagram explaining an example of a posting by a proxy posting bot [Figure 7] Hardware configuration diagram [Figure 8] A diagram explaining the flow of updating the way you talk DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0013] Hereinafter, an embodiment of the present invention will be described with reference to the drawings. The following description and drawings are examples for explaining the present invention, and appropriate omissions and simplifications have been made for clarity of explanation. The present invention can be implemented in various other forms. Unless otherwise specified, each component may be singular or plural.

[0014] In order to facilitate understanding of the invention, the position, size, shape, range, etc. of each component shown in the drawings may not represent the actual position, size, shape, range, etc. Therefore, the present invention is not necessarily limited to the position, size, shape, range, etc. disclosed in the drawings.

[0015] In the following explanation, various information may be described using expressions such as "database," "table," and "list," but the various information may be expressed in other data structures. To indicate independence of data structure, "XX table," "XX list," and the like may be referred to as "XX information." When describing identification information, when expressions such as "identification information," "identifier," "name," "ID," and "number" are used, these are interchangeable.

[0016] When there are multiple components having the same or similar functions, they may be described by using the same reference numerals with different subscripts, but when there is no need to distinguish between these multiple components, the subscripts may be omitted.

[0017] In addition, in the following description, processing performed by executing a program may be described, but the program is executed by a processor (e.g., a CPU, a Graphics Processing Unit (GPU)) to perform a defined process using storage resources (e.g., a memory) and / or an interface device (e.g., a communication port) as appropriate, so the subject of the processing may be the processor. Similarly, the subject of the processing performed by executing a program may be a controller, device, system, computer, or node having a processor. The subject of the processing performed by executing a program may be a calculation unit, and may include a dedicated circuit (e.g., a Field-Programmable Gate Array (FPGA) or an Application Specific Integrated Circuit (ASIC)) that performs a specific process.

[0018] A program may be installed in a device such as a computer from a program source. The program source may be, for example, a program distribution server or a computer-readable storage medium. When the program source is a program distribution server, the program distribution server may include a processor and a storage resource for storing a program to be distributed, and the processor of the program distribution server may distribute the program to be distributed to other computers. In the following description, two or more programs may be realized as one program, and one program may be realized as two or more programs.

[0019] In the following, an example is given in which the system is realized by a single computer; however, similar functions may be realized by distributing all or part of the functions of the system across one or more computers, such as a cloud, and communicating with each other via a network.

[0020] In the following embodiment, the steps to obtain highly fair data, as described in the problem solving method, will be described. In this embodiment, it is assumed that the data bias detection itself will be detected by a system other than this system. The fair data is data that is not biased and has a high fairness.

[0021] The overall flow of the process performed by the fair data collection system 1000 is shown in FIG. 1. The fair data collection system 1000 receives a bias correction request for data used in learning an AI model from an AI application that collects information from online media, such as a news analysis application, and is running (S1). When the fair data collection system 1000 receives the bias correction request, it selects an appropriate user group for bias correction (S2) and asks the user about the content of the data used in learning the AI ​​model in the form of everyday conversation (S3). Then, the fair data collection system 1000 reports the obtained results so that they can be posted on behalf of the user to online media or added to the AI ​​application database (S4). The specific process performed by the fair data collection system 1000 will be described later.

[0022] A method and system for realizing the above processing will be described using the sequence shown in Fig. 2 and the functional block diagram shown in Fig. 3. As shown in Fig. 3, the fair data collection system 1000 in this embodiment is configured to include a user selection unit 201, a daily conversation generation unit 202, dialogue interfaces 203 and 204, a reaction analysis and processing unit 205, a database (DB) 206, a proxy posting unit 207, and a consideration payment unit 208. Specific processing performed by each of these units will be described later.

[0023] When the system receives a request R1 for data bias correction from an AI app (application), it starts collecting information to correct the data bias. Information collection is performed interactively through a user device equipped with the system. Using the sequence in FIG. 2, an example will be described in which a request for data bias correction is received from an AI app that analyzes news. In this example, the AI ​​app collects data and requests the above correction, but a social networking site or portal site may directly request the above correction.

[0024] First, in step 200, online media distributes news to SNS or portal sites (SA). Here, we assume that the online media is news media that creates news articles. A first group of users browses the news (SB) and posts comments and sensitivities, such as sentiment information such as approval, disapproval, anger, sadness, and happiness, to the SNS or some portal sites that accept comments (SC).

[0025] The AI ​​app collects these sentiment data as reaction data (SD) and visualizes trend analysis and the fairness of the news (SE). In particular, news that attracts a large amount of attention, so-called buzz news, has a significant impact on people's thinking. This can be detected, for example, by the number of citations and comments on news sites and social media. SB, SC, and SD processing are performed by devices (e.g., smartphones) operated by the first group of users.

[0026] Buzz news attracts a lot of attention, so it is expected that there will be a high demand for accurate and fair information to ensure that public consciousness is not misled. In order to provide fair information, it is necessary to evaluate the accuracy and fairness of the news article itself, collect public opinion, reflect diverse opinions, and fill in any missing perspectives. News analysis using AI apps involves building AI models that judge or assist in these judgments.

[0027] At this time, if the AI ​​model used in the AI ​​app is trained using information obtained online and the source of the data is biased, the learned characteristics will be biased. To determine whether this bias is discriminatory, the accuracy and fairness of the news will be analyzed while taking AI ethics into consideration, and operation based on the results will be required (SE). In order to take AI ethics into consideration, ensuring the transparency of the data model is an important issue for AI apps, so it is expected that procedures will be taken to ensure transparency. Here, as one such procedure, a correction request to correct the bias in the data will be made to this system. In this way, on the assumption that the processing in step 200 has been performed, this system will perform the following processing.

[0028] The user selection unit 201 first receives a request to correct the bias of data from the AI ​​application and the specification information of the data collected by the application (S201, S202). The specification information may be, for example, information that indirectly conveys the characteristics of the site from which the data is obtained, such as the ID and URL of the site. It may also be information that directly conveys the characteristics of the site, such as the profile information of the accessing user, that is, information on the user's age group and gender. The specification information may also include the original article of the news or paper for which opinions are to be solicited, its summary, and information on trends in the world, and it is preferable that the user be able to specify how to correct the bias of the data and the balance. It is possible to specify conditions such as the number of cases being balanced by age group, wanting to collect more of a specific age group, or wanting to collect more of either men or women.

[0029] User selection unit 201 reads out database 401, which manages information on sites that are sources of data used for learning, as shown in Fig. 4A, and database 402, which manages a second user group, as shown in Fig. 4B. These databases are held in database 206. As shown in Fig. 4A, database 401 stores a site URL, a user profile, and a method for acquiring the user profile in association with each other.

[0030] The site URL is an example of the above-mentioned characteristic information, and is the URL of the site from which the application acquires the data collected. In FIG. 4A, "http: / / www.xxx.co.jp / yyy02.html" is stored. In addition, as a user profile, it is shown that the profile of the user who responded to the site of the URL was "19 years old or younger: {15% male, 20% female}, 20s: {20% male, 15% female}, ..., 70 years old or older: {5% male, 5% female}". Furthermore, it is shown that the user profile was obtained from a survey by "Xxx research organization". As described in FIG. 4A, the user profile refers to an approximate value of the ratio of site access by men and women of each age group. For a site for which the database already has user profile information, it is not necessary to include the user profile information in the characteristic information, and only the URL and ID of the site are notified. Conversely, if the user profile information does not exist in the database, the information is written to the database.

[0031] The information on the missing user group can be extracted from the information on how to balance the data bias included in the specification information and the user profile information of the site. Using the information on the missing user group as a key, a database 402 that manages a second user group is searched to select the second user group (S203).

[0032] As shown in Fig. 4B, the database 402 stores a user ID for identifying a second user in the system, an access route when the user profile of the user was acquired, the profile information of the user, and a user specific ID which is the identification information of the user according to the access route, in association with each other. The daily conversation generation unit 202 can acquire these pieces of information in the following manner.

[0033] For example, an application for using this system is installed via an application store installed on a smartphone or AI speaker. At this time, the smartphone or AI speaker basically accepts the user's operation and consents to the provision of information including the user profile. Alternatively, when a smartphone contract is concluded or a home appliance is purchased, an application for using this system may be pre-installed, and the smartphone or AI speaker may accept the user's operation and consent to the provision of information including the user profile, which is linked to real-offline actions. An example of an ID that identifies a user obtained at the time of installation is the smartphone account ID. The smartphone account ID often matches the email address linked to the OS (Operating System), but an email address may also be used. In addition, the installation ID or customer ID of the home appliance, the user ID of the provider of the AI ​​speaker, etc. can be considered as a user-specific ID. In either case, it is desirable that these user-specific IDs are linked to the actual data of the user at the time of the contract. Home appliances are particularly blessed with opportunities to ask whether real information can be provided in relation to product life guarantees.

[0034] 4B shows that, for example, a user identified by user ID "00000001" has profile information "40 years old, male" registered in access routes originating from three media sources: "smartphone," "home appliance," and "AI speaker." Additionally, it shows that "AccountID," "InstallID," and "UserID" are registered as user-specific IDs corresponding to the three access routes, respectively.

[0035] In addition to these pieces of information, the database 402 stores information on the response rate when an inquiry is made using each user specific ID. In this system, a user with a user specific ID with an extremely low response rate is determined to be highly likely to be causing trouble, and does not transmit information until requested. In FIG. 4B, for example, the response rates for the specific user IDs "AccountID", "InstallID", and "UserID" of a user identified by the user ID "00000001" are "25%, 25%, and 30%, respectively. Therefore, in this system, for the access route by the smartphone or home appliance corresponding to "AccountID" or "InstallID" that is below the predetermined threshold of 25%, the system determines that the access route will cause trouble to the user, and makes an inquiry using the AI ​​speaker corresponding to "UserID". The judgment based on the response rate will be described further below with reference to FIG. 8.

[0036] In addition to this information, the system may also store rewards given to users who respond to the system, in the hope of encouraging users to participate by using the game-like nature of collecting points. The rewards may simply be points, or tokens circulating within a blockchain network may be distributed. For example, FIG. 4B shows that a reward of "200 points" has been given to a user identified by the user ID "00000001."

[0037] In addition to this information, it is also a good idea to manage each user's preferred way of speaking.

[0038] For example, it is advisable to update the speaking style according to the flow shown in Fig. 8 (S207). Here, the speaking style is updated after S205 and S206 are performed, but it may be updated after S203 and S204 are performed. This allows the dialogue in S205 and S206 to be performed using the speaking style updated to the latest state.

[0039] In S207, the daily conversation generation unit 202 receives requests for each user, such as "please speak in a friendly manner," "please speak in a bossy manner," "please speak like a robot," "please speak in a dialect," etc., and devise ways to increase the response rate. To explain in detail below, the daily conversation generation unit 202 has daily conversations like S205 and S206 for a certain period of time (for example, one month) in a speaking style that has been requested in advance and stored in the database 402 together with the preferred speaking style, and tally up the response rate (S801).

[0040] The daily conversation generation unit 202 judges whether or not the response rate from the user in the requested speaking style satisfies a threshold value (S802), and if it is judged that the response rate satisfies the threshold value (S802; Yes), it ends the process without doing anything. On the other hand, if the daily conversation generation unit 202 judges that the response rate does not satisfy the threshold value (S802; No), it refers to the database 402 and identifies users from the second user group who have similar profile information. The judgment of whether or not the profile information is similar may be made by, for example, judging that the users are similar if their ages and genders match.

[0041] The daily conversation generation unit 202 selects a speech style with a response rate higher than a certain threshold from among the preferred speech styles corresponding to the identified user, and presents a dialogue scenario based on the selected speech style to the user (S803). One method of presentation is, for example, displaying the scenario on a screen of a terminal (e.g., a smartphone) operated by the second user group. The daily conversation generation unit 202 accepts a selection of a desired scenario from the presented scenarios from the user (S804), and updates the preferred speech style stored in the database 402 with the speech style of the accepted scenario (S805). By performing such processing, it is possible to increase the response rate while taking into account the user's request.

[0042] In this example, a preferred way of speaking is stored for each user (i.e., for each user ID), but a preferred way of speaking may also be stored for each user specific ID. This allows the way of speaking to be changed depending on the means used by the user to respond to inquiries. Users who often make inquiries and respond to those inquiries using a smartphone can respond more smoothly via a smartphone that they are accustomed to using, so even if you speak to them in a friendly manner, the user will respond in a good mood. On the other hand, with home appliances that you are not accustomed to using, you will not be able to respond smoothly, so it is preferable to speak to them politely. In this way, changing the way of speaking in consideration of the access route can contribute to improving the user's response rate.

[0043] Alternatively, as a modified example, when a human presence sensor provided in a medium such as an AI speaker or a home appliance such as a refrigerator that provides the access route shown in Fig. 4B detects the user, the daily conversation generation unit 202 may receive a signal indicating the detection from the home appliance and talk to the user in the preferred way of talking. In this case, the user can talk to the device at the timing when the user performs these operations, enabling an efficient conversation.

[0044] The daily conversation generation unit 202 creates a question for daily conversation using information stored in the database 402 as the preferred way of speaking for the second group of users (S204). For example, the daily conversation generation unit 202 summarizes the original article of the news or paper for which opinions are sought, its summary, and information on trends in the world, which are included in the specification information, and creates a question through daily conversation that is the preferred way of speaking for the summarized content. The question is preferably in a form in which information is collected little by little in casual conversation while speaking naturally, rather than in a concentrated answer format like a questionnaire. The daily conversation generation unit 202 prepares several scenarios of ways of speaking, and generates daily conversations with the user according to the prepared scenarios using the dialogue interfaces 203 and 204. The daily conversation generation unit 202 exchanges daily conversations after converting the generated daily conversation into the preferred way of speaking, and acquires reaction data to questions from the system (S205, S206). The daily conversation generation unit 202 may run by calling up a huge natural language understanding model such as LLM online when generating or converting daily conversations, or may run after securing an internal operating environment. The dialogue interfaces 203 and 204 have a function to convert voice to text and a function to convert text to voice. It is desirable for the voice to have a variety of voices, including the above-mentioned preferred speaking style and the type of voice that is preferred. There is a possibility that the quality of the voice can encourage users to use the system.

[0045] The dialogue interface 204 of the user-side device may be in the form of a screen 301 output by an application installed in a smartphone as shown in FIG. 5A, for example. If the system has a dialogue function, it does not matter whether the screen is output or not, but when the system operates without outputting the screen, the system does not actively talk to the user, but only starts talking when the user talks to it. The dialogue may be executed when the user wants to kill time, when another voice-input search is performed, etc. Whether the user wants to kill time or not can be determined by the system judging whether the user's utterance contains a word meaning "kill time".

[0046] 5A, a conversation is taking place on a screen 301 between a bot, which is the proxy posting unit 207 of the present system, and a user of the second user group about the topic "xxx." The daily conversation generation unit 202 refers to the database 402 and carries out a conversation in the preferred speaking style corresponding to the user.

[0047] When voice cannot be used, the only option is to display the information on a display, in which case dialogue may be achieved by a pop-up notification 302, as shown in FIG. 5B. In addition to smartphones, implementation in AI speakers and smart home appliances, as shown in FIG. 5C, is also possible. AI speakers are representative of interfaces operated by voice input, and because they are actually smartphones, it is easy to add applications and create user profiles. Therefore, they are considered the most natural implementation form for this system.

[0048] Other smart home appliances equipped with IT functions are available on the market, but home appliances that are accessed frequently, such as refrigerators, and that are in the user's vicinity for a certain period of time, such as while cooking, are well suited to installing the dialogue interface with the user that this system aims to provide. A feature of smart home appliances is that they use IT to perform predictive diagnosis and report problems in order to extend the product's lifespan, and they provide usage methods that allow the product to be used for a long time. In particular, when providing usage methods that allow the product to be used for a long time, it is natural to provide a speaker or a function to display tips. Therefore, these home appliances can be considered to function as a front end for functions that collect data while conducting everyday dialogue, such as this system.

[0049] When the reaction analysis and processing unit 205 receives reaction data from the second group of users in response to a question in the form of daily conversation in S206, it analyzes and processes the reaction data (S208). This process corresponds to anonymization processing for anonymizing profile information so that the identity of the user is not revealed when the system outputs proxy posting. If the information obtained from the user is directly transcribed and used, there is a possibility that the user may be identified from the information received from the user containing proper nouns or speaking habits. The method of converting only the speaking style without changing the meaning is a process that the generation AI excels at, and a conventionally known practical application may be used.

[0050] The proxy posting unit 207 posts the text output from the response analysis and processing unit 205 on behalf of the user to the data posting destination internet site or the data reception unit of the AI ​​application that made the request to correct the bias (S209). The data posting destination internet sites that the AI ​​application uses as information sources include SNS. By posting here, it is expected that information for indirectly eliminating the bias will flow to the data collection unit for the AI ​​model.

[0051] On the other hand, the proxy posting unit 207 can directly return the user's reaction data to the AI ​​application database by posting on behalf of the data receiving unit. The proxy posting unit 207 is not limited to one bot, and can have multiple instances. Instead of posting the reaction data of a specific user on behalf of a single bot, the proxy posting unit 207 can be configured with multiple bots in groups classified based on the user's profile information shown in the database 402, and can send the data while shuffling them.

[0052] For example, the proxy posting unit 207 automatically assigns a serial number 601 to an instance as shown in FIG. 6, and posts the data as a different user from the original user. In other words, one instance posts reaction data of multiple users. On a site on the Internet that is seen by many people, the shuffle function is effective as a means of preventing individuals from being identified. On the other hand, when returning an answer directly to the data receiving unit of the AI ​​application, shuffling is not necessary, as matching with a specific user is prioritized. As shown in FIG. 6A, the proxy posting unit 207 declares that it is a bot on the user's profile screen, but clearly indicates that it will post only articles created based on real answers 602. As shown in FIGS. 6B and 6C, the article posting screen clearly indicates an inquiry ID 603 for the reaction data output by this system. The inquiry ID 603 is automatically assigned by the proxy posting unit 207. The proxy posting unit 207 builds a system that can guarantee the authenticity of reaction data based on this number when it receives an inquiry about the post from a user or another person who has viewed the proxy post. For this purpose, the reaction analysis and processing unit 205 should associate the reaction data obtained from the user with the inquiry ID 603, and manage the converted history in a ledger in a traceable format. For example, it is possible to have it in a format that can be connected to a blockchain, and manage the ledger on the blockchain. Of course, a normal database format is also acceptable, but by implementing it with a blockchain, it becomes possible to trace the processing step by step.

[0053] The proxy posting unit 207 accumulates and holds in the database 206 a ledger database 403 capable of correlating and matching the proxy posted data with the original reaction data collected from the second user group as shown in Fig. 4C. In Fig. 4C, the ledger database 403 stores a collection ID for identifying the data collected in S202, the user ID of the second user selected in S203, the above-mentioned inquiry number, the reaction data acquired in S206, the results of the analysis in S208, the flavor and sentence model of the text to be posted by proxy, identification information (BOT instance) of the proxy posting unit 207, and the posting data actually posted by proxy in a corresponding manner.

[0054] In this way, when the inquiry ID 603 and BOT_instance are given, the proxy posting unit 207 is configured to be able to reverse-look up which user in the second user group returned what kind of reaction data. In other words, the reaction data acquired in S206 holds the original data of the reaction data (data before being processed in S208) and the intermediate processing result (result of analysis in S208) when the original data is deciphered by natural language processing as analysis data. Based on this intermediate processing result, the proxy posting unit 207 uses a predetermined flavor of the text or a text model to adjust the flavor of the text to be posted by proxy according to the result. The proxy posting unit 207 records the flavor to which the change was made at this time, or information on the text generation model. This makes it possible to ensure the transparency of the process by which the actual posted data was generated.

[0055] When the compensation payment unit 208 receives a request for cooperation compensation payment from the AI ​​application as a compensation with the bias of the data corrected (S210), the compensation payment unit 208 redistributes a part of the funds obtained according to the number of news views of SB to the second user side as cooperation compensation (S211). As shown in FIG. 4B, the database 206 manages which users of the second user group have been requested to provide information, and the compensation payment unit 208 pays the user compensation in the form of points or tokens. As described above, by making the second user find value in the game nature of collecting points or tokens and in the mechanism that allows the user to receive cooperation compensation just by talking to the bot as a way of killing time, an incentive can be given to the second user, and as a result, it is considered possible to increase the response rate.

[0056] Through the above process, we provide a mechanism for collecting fair data.

[0057] FIG. 7 shows the hardware configuration for realizing this system. The CPU 413 starts the program of this system and executes the processing of the functional blocks in FIG. 3. The data used for the calculation is temporarily stored in the memory 415. When high-speed calculation is required, the data may be transferred to the memory of the GPU 414 and the calculation support of the GPU may be used. In particular, in this system, high-load processing is performed for daily conversation, so the support by the GPU is important. The display information generated by this system is displayed on the display device 411, and the database 206 that can be previewed by the system operator may be stored on the storage device 412. However, the processing history of the user's reaction generated by the reaction analysis and processing unit 205 may be connected to an online environment through the communication interface 417 and managed in a blockchain ledger. Even when the cooperation fee to the user is paid with tokens, it is a necessary condition for implementation that the server of this system and the user device are connected to the same blockchain. The server side of this system may be provided as an on-premise server or may be implemented on a public cloud. When implementing on a public cloud, a configuration may be adopted in which the CPU 413, GPU 414, and memory 415 are dynamically secured and scaled according to the scale of calculation.

[0058] The implementation of an AI application that aims to collect fair data that interacts with this system is indicated by adding the symbol -A to the end. The news analysis process uses 411-A to 417-A to realize the function, but when receiving data directly from the proxy posting bot of the present invention, it is received through communication interface 417-A.

[0059] The server communicates with the user via the communication I / F 417 to transfer the news browsing website GUI. The user's terminal also has a similar hardware configuration. Any client terminal with a similar hardware configuration, such as a personal computer, smartphone, AI speaker, smart home appliance, or game console, that has a web browsing function may be used as a network participating terminal for this system. The user interface 416-U should preferably have a speaker and microphone function to enable voice input from the user.

[0060] With the above, we can provide a system that collects data from everyday conversations and allows users to receive compensation for their cooperation if they respond to the conversation.

[0061] As described above, as described with reference to FIG. 2 etc., the data collection system according to this embodiment includes a data collection system (fair data collection system 1000) executed by a computer, a user selection unit 201 (S201-S203) that, when receiving a request to correct bias in a post by a first user to an article on a predetermined site (SNS or portal site), selects the second user for which bias in the post is to be corrected based on specification information on the article and the first user and a database (e.g., database 402 in FIG. 4B) that stores a second user having specifications different from those of the first user, a daily conversation generation unit 202 (S204-S206) that summarizes the content of the article, creates a question sentence for the second user about the summarized content through daily conversation, and acquires reaction data to the created question, and a proxy posting unit 207 (S209) that posts the reaction data to the posting destination of the first user on behalf of the second user. This makes it possible to collect reaction data from users who have not expressed their own opinions on the Internet, thereby improving the fairness of sentiment data.

[0062] As explained with reference to Fig. 4B etc., the daily conversation generation unit 202 stores in the database the access route when the profile information of the second user was acquired, the profile information of the second user, a user specific ID which is the identification information of the second user according to the access route, and a response rate to the question based on the reaction data in association with each other. This makes it possible to grasp what kind of profile the second user has.

[0063] As described with reference to Fig. 4B and S205 in Fig. 2, the daily conversation generation unit 202 asks the second user the question in a preferred manner of speaking that is stored in advance in the database, thereby increasing the response rate of the reaction data.

[0064] As explained as a modified example of S205 in Fig. 4B and Fig. 2, when a sensor of a medium (e.g., an AI speaker, a home appliance such as a refrigerator) that is a source of profile information of the second user detects the second user, the daily conversation generation unit 202 acquires reaction data of the second user in the form of daily conversation. This allows the user to talk to the second user at the timing when the user operates these media, enabling efficient conversation.

[0065] As described with reference to Fig. 3 etc., the daily conversation generation unit 202 acquires the questions and the reaction data from the second user using a smartphone, an AI speaker, or a smart home appliance as an interface. This allows the user to converse with the second user through a device or home appliance that the user is familiar with and uses daily, allowing the user to post reaction data to the system without any burden.

[0066] 4B, S208 in FIG. 2, etc., the analysis and processing unit 205 conceals the profile information of the second user included in the reaction data acquired by the daily conversation generation unit 202, and the proxy posting unit 207 posts the concealed reaction data. This ensures the confidentiality of the second user.

[0067] 2, the proxy posting unit 207 is a bot having multiple instances, and posts the reaction data of multiple second users to one instance, thereby preventing individuals from being identified on a site on the Internet.

[0068] As described with reference to Fig. 4C etc., the proxy posting unit 207 associates the reaction data actually posted on behalf of the second user with the original reaction data collected from the second user and stores them in a ledger database (e.g., ledger database 403 in Fig. 4C). This makes it possible to verify which second user the reaction data was collected from.

[0069] 4C, S209 in FIG. 2, etc., the proxy posting unit 207 assigns an inquiry number to the reaction data, and stores the assigned inquiry number in the ledger database in association with the reaction data actually posted on behalf of the second user and the original reaction data collected from the second user. This makes it possible to ensure the authenticity of the reaction data even when an inquiry about the post is received from the user or another person who viewed the proxy post.

[0070] 2, the proxy posting unit 207 (1) receives the post from the SNS or portal site, which is the predetermined site, and posts the reaction data of the second user to the SNS or portal site, or (2) an application (AI application) that collects the posts from the SNS or portal site receives the post and posts the reaction data of the second user to the application. This allows the user to select (1) or (2) according to their environment and use the system.

[0071] 2, when a request for cooperation compensation payment is received from the application that collects the posts from the predetermined site as a compensation with the bias corrected, the compensation payment unit 208 redistributes a portion of the funds obtained according to the number of views of the article to the second user as cooperation compensation. This provides an incentive to the second user, making it possible to increase the response rate of reaction data.

[0072] In this way, this system makes it possible to collect data from a wide range of users, build AI models and operate services with a high level of fairness. Alternatively, by collecting extensive information on a targeted audience, it is possible to provide marketing information. If information is collected only from online media, there is a risk that opinions of "loud" users will be collected, which could lead to bias in the construction and operation of AI models and lead to significant liability in terms of AI ethics. However, this system makes it possible to avoid such situations. [Explanation of symbols]

[0073] 1000 Fair Data Collection System 201 User Selection Section 202 Daily Conversation Generation Unit 203 Dialogue Interface (Server Side) 204 Dialogue Interface (User Side) 205 Reaction Analysis and Processing Department 206 Database 207 Proxy Posting Bot 208 Payment Department 411 Display device 412 Storage device 413 CPU 414 GPU 415 Memory 416 User Interface 417 Communication Interface

Claims

1. 1. A computer implemented data collection system comprising: a user selection unit that, when receiving a request to correct a bias in a post made by a first user to an article on a predetermined site, selects the second user for which a bias in the post is to be corrected based on specification information on the article and the first user and a database that stores second users having specifications different from those of a predetermined first user; a daily conversation generation unit that summarizes the content of the article, creates a question sentence through daily conversation for the second user regarding the summarized content, and acquires reaction data to the created question; a proxy posting unit that posts the reaction data to a posting destination of the first user on behalf of the second user; A data collection system comprising:

2. 2. The data collection system of claim 1, the daily conversation generation unit stores in the database an access route when the profile information of the second user was acquired, the profile information of the second user, a user specific ID which is identification information of the second user according to the access route, and a response rate to the question based on the reaction data, in association with each other; A data collection system comprising:

3. 2. The data collection system of claim 1, the daily conversation generation unit asks the question to the second user in a preferred manner of speaking that is stored in advance in the database; A data collection system comprising:

4. 2. The data collection system of claim 1, the daily conversation generation unit, when a sensor of a medium that is a source of profile information of the second user detects the second user, acquires reaction data of the second user in the form of daily conversation; A data collection system comprising:

5. 2. The data collection system of claim 1, The daily conversation generation unit acquires the questions and the reaction data between the second user and the daily conversation generation unit through an interface of a smartphone, an AI speaker, or a smart home appliance. A data collection system comprising:

6. 2. The data collection system of claim 1, an analysis and processing unit that conceals profile information of the second user included in the reaction data acquired by the daily conversation generation unit; The proxy posting unit posts the anonymized reaction data. A data collection system comprising:

7. 2. The data collection system of claim 1, The proxy posting unit is a bot having a plurality of instances, and posts the reaction data of the plurality of second users to one instance. A data collection system comprising:

8. 2. The data collection system of claim 1, the proxy posting unit associates the reaction data actually posted on behalf of the second user with the original reaction data collected from the second user and accumulates the associated reaction data in a ledger database; A data collection system comprising:

9. 9. The data collection system of claim 8, the proxy posting unit assigns an inquiry number to the reaction data, and stores the assigned inquiry number in the ledger database in association with the reaction data actually posted on behalf of the second user and the original reaction data collected from the second user; A data collection system comprising:

10. 2. The data collection system of claim 1, The proxy posting unit (1) receives the post from the SNS or portal site, which is the predetermined site, and posts the reaction data of the second user to the SNS or portal site, or (2) an application that collects the posts from the SNS or portal site receives the post, and posts the reaction data of the second user to the application. A data collection system comprising:

11. 11. The data collection system of claim 10, a compensation payment unit that, when receiving a request for cooperation compensation payment as a compensation with the bias corrected from the application that collects the posts from the predetermined site, redistributes a portion of funds obtained according to the number of views of the article to the second user as cooperation compensation; A data collection system comprising:

12. 1. A computer implemented data collection method comprising: When receiving a request to correct a bias in a post of a first user to an article on a predetermined site, a second user for correcting a bias in the post is selected based on specification information on the article and the first user and a database storing a second user having specifications different from those of a predetermined first user; Summarizing the content of the article, creating a question sentence for the second user through everyday conversation regarding the summarized content, and acquiring reaction data to the created question; Posting the reaction data to a posting destination of the first user on behalf of the second user; A data collection method comprising:

Images