Target respondent estimation system

Abstract

The objective is to provide a respondent response estimation system that allows the interviewer to conduct interviews in a natural manner, as if they were actually asking questions to the survey subjects. [Solution] The machine learning device 10 of the survey subject response estimation system 1 is configured to include the following functions: a learning model storage unit 11 that stores and holds a digital clone Ci of each survey subject Pi; a question input unit 13 that inputs question data indicating the content of the question given by the user Q to the required digital clone Cx (=C1, or C2, or ..., or CN) from among the digital clones C1 to CN; and a response output unit 14 that outputs the response obtained from the digital clone Cx according to the question data.

Description

Technical Field

[0001] The present invention relates to a survey respondent answer estimation system that estimates and outputs answers from survey respondents.

Background Art

[0002] In surveys in various fields such as marketing surveys and television rating surveys, as shown in Patent Document 1 below, it has been conventionally common to collect answers to questions from a plurality of pre-selected survey respondents (so-called panelists or panelers). The applicant of the present application and the inventor of the present invention have made a patented invention shown in Patent Document 2 below so as to ensure the same quality and reliability as when collecting answers from survey respondents while reducing the burden on survey respondents.

[0003] In the survey aggregation system of Patent Document 2 below, the answers of each of the plurality of survey respondents to the questions implemented for each of the plurality of survey respondents are compared with the answers generated by each of the plurality of learning models for the questions, and from the combination of each of the plurality of survey respondents and each of the plurality of learning models, a pair set is extracted, which is a set of survey respondents and learning models with a high degree of answer agreement for the questions of a predetermined threshold or more. It is configured so that, as a substitute for the answers of the survey respondents, the answers generated by each learning model of the pair set for the questions can be obtained.

Prior Art Documents

Patent Documents

[0004]

Patent Document 1

Patent Document 2

Summary of the Invention

Problems to be Solved by the Invention

[0005] Here, the inventors of the present invention have developed such a survey and data aggregation system and arrived at the realization that various survey data can be obtained in a natural manner, as if the questioner were exchanging questions with the actual survey subject.

[0006] In view of the above circumstances, the present invention aims to provide a subject response estimation system that enables the questioner to conduct the interview in a more natural manner, as if they were actually asking questions to the subject of the survey. [Means for solving the problem]

[0007] The first invention of the present invention is a survey subject response estimation system that estimates and outputs responses from survey subjects, The machine learning device comprises a machine learning device that constructs and stores learning models for multiple survey subjects using machine learning processing, The aforementioned machine learning device is A learning model storage unit stores multiple learning models, each constructed to be able to estimate the response given by each survey participant to a question by learning the responses of multiple survey participants, as digital clones of the multiple survey participants. A question input unit that obtains questions from multiple survey subjects and inputs question data representing the content of those questions into the multiple learning models, The question input unit receives question data and outputs an answer that the multiple learning models have estimated to be the answers of multiple survey subjects. It is equipped with, The response output unit is characterized by outputting multiple estimated responses from the multiple learning models to the same question data via digital clones that represent the personas of the multiple survey subjects, whose responses were used to construct the multiple learning models, using human images.

[0008] According to this invention, the answer output unit outputs multiple estimated answers from multiple learning models for the same question data at once, so that the questioner can naturally recreate a scenario where they are asking the same question to multiple survey subjects and receiving answers simultaneously (without selecting any special settings).

[0009] In such scenarios, multiple estimated answers from multiple learning models to the same question data are output via digital clones that represent the personas of multiple survey participants whose answers were used to train the multiple learning models, using human images.

[0010] In this way, since the answers are output from digital clones of human images that reflect the persona of the original research subject corresponding to the learning model, it is possible to achieve a more natural feel, as if the questioner were having a question-and-answer session with the actual research subject.

[0011] The survey participant response estimation system of the second invention is, in the first invention, The response output unit is characterized in that, when there is no question data, it outputs information about the survey subject, including a self-introduction associated with the persona.

[0012] According to this invention, while a conversation may not start if there is no question data from the questioner, if there is no question data, the digital clone outputs information about the research subject, including a self-introduction corresponding to the persona, allowing the questioner to understand the research subject (more precisely, the digital clone corresponding to the research subject) in a natural way, prompting them to ask questions and start a conversation.

[0013] The survey participant response estimation system of the third invention is, in the first invention, The response output unit is characterized by extracting characteristics of the question content from the question data and outputting a question corresponding to those characteristics of the question content.

[0014] According to this invention, by extracting the characteristics of the questions asked by the questioner and having a digital clone ask the questioner questions corresponding to the extracted characteristics, it is possible to achieve a natural conversation as if the questioner and the original subject of the investigation were asking each other questions.

[0015] The survey participant response estimation system of the fourth invention is, in the first invention, The aforementioned response output unit is characterized by extracting characteristics of the response content from the outputted predicted response and outputting a question corresponding to those characteristics.

[0016] According to this invention, by extracting the characteristics of the answer content from the outputted predicted answer and asking questions corresponding to the extracted characteristics from the digital clone to the questioner, it is possible to realize a natural conversation as if the questioner and the original subject of the survey were asking each other questions.

[0017] The survey subject response estimation system of the fifth invention is, in the third or fourth invention, The system is characterized by repeating a conversation by using the questioner's response to the question output from the response output unit as the question data.

[0018] According to this invention, by using the questioner's response to a question output from the response output unit as question data, the conversation can be repeatedly continued using the questioner's response as new question data, thereby achieving a more natural feel, as if the questioner were actually asking questions to the subject of the investigation.

[0019] The survey participant response estimation system of the sixth invention is, in the fifth invention, The machine learning device is characterized in that it stores a learning model, constructed to be able to estimate the answers that a questioner will give to a question by learning the answers from the questioner to the questions output by the answer output unit, as a digital clone of the questioner in the learning model storage unit.

[0020] According to such an invention, a natural state is realized as if the questioner is exchanging questions with the original respondent, so that, as various survey data, a learning model that has learned the questions and answers by the questioner can be constructed, and a digital clone of the questioner can be created.

[0021] In the respondent answer estimation system of the seventh invention, in the first invention, The answer output unit is characterized by including an attribute estimation unit that estimates the attributes of the questioner from the characteristics of the question content by extracting the characteristics of the question content from the question data.

[0022] According to such an invention, a natural state is realized as if the questioner is exchanging questions with the original respondent, so that, as various survey data, the attributes of the questioner (true attributes that the questioner himself / herself is not aware of) can be estimated from the characteristics of the question content of the questioner.

[0023] In the respondent answer estimation system of the eighth invention, in the third or fourth invention, It is characterized by including an attribute estimation unit that estimates the attributes of the questioner from the characteristics of the answer content of the answer from the questioner to the question output from the answer output unit.

[0024] According to such an invention, a natural state is realized as if the questioner is exchanging questions with the original respondent, so that, as various survey data, the attributes of the questioner (true attributes that the questioner himself / herself is not aware of) can be estimated from the characteristics of the answer content of the questioner.

[0025] In the respondent answer estimation system of the ninth invention, in the first invention, The learning model storage unit is characterized by changing the constructed plurality of learning models over time.

[0026] According to this invention, by changing multiple constructed learning models over time, the digital clone can be aligned with the passage of time, allowing the questioner to experience a more natural interaction, as if they were questioning the original subject, including the passage of time. [Brief explanation of the drawing]

[0027] [Figure 1] A block diagram showing the configuration of a survey participant response estimation system in an implementation form of the present invention. [Figure 2] Figure 1 is an explanatory diagram of the processing of the pair set extraction unit. [Figure 3] A block diagram showing the configuration of the machine learning system shown in Figure 1. [Figure 4] Figure 1 is an explanatory diagram showing the processing of the machine learning device. [Figure 5] Figure 1 is an explanatory diagram showing the processing of the machine learning device. [Modes for carrying out the invention]

[0028] One embodiment of the present invention will be described below with reference to Figures 1 to 5.

[0029] First, referring to Figure 1, the survey subject response estimation system 1 of this embodiment comprises a machine learning device 10 that constructs (generates) and stores multiple learning models LM1, LM2, ..., LMN capable of generating answers to questions using machine learning processing; an answer aggregation unit 20 that obtains and aggregates answers to questions from all or part of a pre-selected group of survey subjects P1, P2, ..., Pn; a question data generation unit 30 that generates question data indicating the content of questions for all or part of the survey subjects P1 to Pn; a survey subject database 40 (hereinafter simply referred to as database 40) that stores information about each of the survey subjects P1 to Pn; and a communication device 50 capable of communicating with an external network NW (wide area network) consisting of the Internet, a telephone communication network, etc. Hereafter, any one of the multiple learning models LM1 to LMN will be denoted as LMi, and any one of the multiple survey subjects P1 to Pn will be denoted as Pj.

[0030] In addition to being able to communicate with an external network NW, the communication device 50 can also communicate with each of the survey subjects P1 to Pn's communication terminals 60 via the external network NW.

[0031] Here, the communication terminal 60 used by each survey participant Pj consists of, for example, a smartphone, tablet, or personal computer, and includes a display unit 60a consisting of an LCD display or the like, and a sound output unit 60b consisting of a speaker or the like. A predetermined application for the survey (hereinafter referred to as the survey application) is pre-installed on this communication terminal 60.

[0032] Furthermore, the communication terminal 60 can communicate with the communication device 50 via an external network NW while the survey application is running. In this case, the communication terminal 60 has functions such as receiving question data transmitted from the survey subject response estimation system 1 via the communication device 50, informing the survey subject Pj of the content of the questions indicated by the question data, and receiving the survey subject Pj's answers to the questions and transmitting them to the communication device 50.

[0033] The machine learning device 10, response aggregation unit 20, question data generation unit 30, and database 40 of the survey participant response estimation system 1 are composed of one or more computers, including, for example, a processor such as a microcontroller, a storage device such as memory, and an interface circuit.

[0034] For example, the machine learning device 10, the response aggregation unit 20, the question data generation unit 30, and the database 40 may each be composed of separate computers capable of communicating with one another. In this case, the computer constituting the machine learning device 10 is configured to function as the machine learning device 10 through the hardware configuration and programs (software configuration) implemented therein. The same applies to the response aggregation unit 20, the question data generation unit 30, and the database 40.

[0035] However, the computer may be configured such that two or more of the machine learning device 10, the response aggregation unit 20, the question data generation unit 30, and the database 40 are included in, for example, one computer. Furthermore, the communication device 50 may be included in or attached to any of the computers comprising the machine learning device 10, the response aggregation unit 20, the question data generation unit 30, and the database 40. Alternatively, the communication device 50 may be included in or attached to each of the multiple computers comprising the machine learning device 10, the response aggregation unit 20, the question data generation unit 30, and the database 40.

[0036] Database 40 stores information about each of the pre-selected survey subjects P1 to Pn, including attribute information such as the place of residence, gender, age, and family structure of each survey subject Pj, as well as response history information showing the history of answers to questions given to each survey subject Pj. Database 40 can output arbitrary information (attribute information, response history information, etc.) about any survey subject Pj in response to requests from the machine learning device 10, the response aggregation unit 20, etc.

[0037] The question data generation unit 30 is configured to generate question data, which indicates the content of multiple questions for a survey, in response to instructions from an operator, when conducting a survey such as a marketing survey or a television viewership survey. This question data can be generated, for example, as text data or audio signal data. The question data generation unit 30 is also capable of storing a history of the question data generated in each survey that has been conducted.

[0038] The machine learning device 10 is configured to construct (generate) multiple learning models LM1 to LMN that can generate answers to questions by using machine learning processing that utilizes the life logs of multiple unspecified people (an unspecified number of individuals) as training data.

[0039] This life log contains a large amount of unique or distinctive information about an unspecified number of individuals. This life log may include information shared by individuals via social networking services (SNS), etc.

[0040] Such life logs can be input by the operator to the machine learning device 10 at any time via an appropriate input device. Alternatively, the machine learning device 10 can automatically collect them from an external network NW such as the internet.

[0041] The machine learning device 10 extracts life logs from a given number of life logs to be used as training data for each learning model LMi, and constructs each learning model LMi by performing machine learning processing using the extracted life logs as training data for each learning model LMi.

[0042] In this case, when extracting lifelogs for the training data of each learning model LMi, for example, the lifelogs for the training data of each learning model LMi can be extracted such that at least some of the lifelogs used as training data by each learning model LMi are different from each other in learning models LM1 to LMN. This allows each of learning models LM1 to LMN to be configured to have different characteristics in terms of generating answers to questions.

[0043] To elaborate, if life logs of subjects P1 to PN can be collected, then, for example, the life logs of any subject Pi may be used as training data for at least some of the learning models LM1 to LMN. In this case, if there are multiple learning models that use the life logs of subject Pi as training data, then the life logs of different subjects will be used as training data for each of those multiple learning models.

[0044] Furthermore, the input of new life logs into each LMi learning model and the subsequent machine learning processing will be carried out continuously. As a result, each LMi learning model will be updated to be able to generate answers that are consistent with the content of the data for various questions.

[0045] The machine learning device 10 is configured to construct multiple learning models LM1 to LMN through machine learning processing, as described above. More specifically, the technology proposed in, for example, Japanese Patent Publication No. 2018-190457, can be used to generate each learning model LMi.

[0046] The machine learning device 10 of this embodiment is further configured to include a function as a learning model storage unit 11 that stores the constructed learning models LM1 to LMN, and a function as a pair set extraction unit 12 that extracts pair sets of research subjects and learning models from combinations of each of the multiple research subjects P1 to Pn and each of the multiple learning models LM1 to LMN, in which the degree of agreement in the answers to the questions is higher than a predetermined threshold.

[0047] Here, we will explain the specific processing of the pair set extraction unit 12. The pair set extraction unit 12 performs the pair set extraction process, for example, at regular intervals, at intervals instructed by the operator, or at intervals after each survey has been conducted. In this process, the pair set extraction unit 12 obtains question data from the question data generation unit 30 from surveys conducted in the past (for example, the most recent survey or the most recent multiple surveys) for all or some of the survey subjects P1 to Pn, and inputs the multiple questions (text-format questions) indicated by that question data into the learning models LM1 to LMN.

[0048] Furthermore, the pair set extraction unit 12 retrieves the answers generated by each learning model LMi in response to each of the multiple questions, and also retrieves the answers obtained from all or part of the survey subjects P1 to PN for the multiple questions from the database 40, and compares the answers obtained from each of the learning models LM1 to LMN with the answers obtained from all or part of the survey subjects P1 to PN.

[0049] The pair set extraction unit 12 then extracts pairs of subjects and learning models from the combinations of each of the multiple subjects P1 to PN and each of the multiple learning models LM1 to LMN, in which the degree of agreement with the answers to the questions is higher than a predetermined threshold.

[0050] In this case, as an index value representing the degree of agreement between the responses of each learning model LMi and the responses of each survey subject Pj, for example, the ratio of the total number of responses that match between the learning model LMi and the survey subject Pj (the ratio to the total number of questions; hereafter referred to as the matching response ratio) can be used. The pair set extraction unit 12 then extracts a pair set (Pj, LMi) of a survey subject Pj from among survey subjects P1 to PN and a learning model LMi from among learning models LM1 to LMN if the matching response ratio for that pair is above a predetermined threshold (for example, if the matching response ratio is above a threshold close to 100%).

[0051] For example, referring to Figure 2, if the agreement rate for the pair of subject Pa and learning model LMa, the agreement rate for the pair of subject Pb and learning model LMb, and the agreement rate for the pair of subject Pc and learning model LMc are all above a predetermined threshold, then these three pairs (Pa,LMa), (Pb,LMb), and (Pc,LMc) are extracted as pair sets.

[0052] The learning model memory unit 11 then stores and retains the learning model LMi that forms a pair set with any of the research subject Pj from learning models LM1 to LMN, associating it with the research subject Pj that forms the pair set. For example, Figure 1 shows that learning model LM2 is stored in the learning model memory unit 11 as being associated with research subject P5, which forms a pair set with it.

[0053] To elaborate, the learning model LMi that forms a pair set with any of the research subject Pj exhibits a high degree of agreement with the research subject Pj in its answers to the questions, and can therefore be considered equivalent to artificial intelligence that simulates the thinking and behavioral patterns of the research subject Pj (in other words, a digital clone of the research subject Pj).

[0054] The response aggregation unit 20, as will be described in detail later, is configured to select multiple survey subjects from the survey subjects P1 to Pn registered in the database 40 as subjects from whom responses to questions should be obtained when conducting a required survey, and to obtain and aggregate responses to multiple questions from these multiple subjects.

[0055] In addition, if there are a certain number of respondents who make up a pair set among the selected respondents, the response aggregation unit 20 can also obtain and aggregate responses to multiple questions only from the learning models corresponding to the respondents who make up the pair set.

[0056] Next, we will describe the details and operation of the machine learning device 10.

[0057] As shown in Figure 3, the machine learning device 10 is configured to construct (generate) digital clones Ci as learning models LM1 to LMN, each constructed for each research subject P1 to Pn.

[0058] Furthermore, the machine learning device 10 is configured to include the following functions: a learning model storage unit 11 that stores and holds a digital clone Ci of each survey subject Pi; a question input unit 13 that inputs question data indicating the content of a question given by a user Q (corresponding to the questioner in this invention) via a communication terminal (not shown) to a required digital clone Cx (=C1, or C2, or ..., or CN) from among the digital clones C1 to CN; and an answer output unit 14 that outputs an answer obtained from the digital clone Cx according to the question data.

[0059] Next, we will specifically explain how the survey participant response estimation system 1 operates when conducting the necessary survey.

[0060] As shown in Figure 4, User Q first inputs the target group criteria. Specifically, User Q enters a group name, then selects gender, age, place of residence, and occupation, and finally enters the text "studying using a learning app" in the attribute details setting field. This allows the database 40 to select pairs of research participants and learning models that satisfy the target group criteria from the attribute information of research participants stored in the database. As a result, in this example, a group with attributes such as "student & learning app user" is set.

[0061] In contrast to conventional qualitative research, which can take several weeks just to recruit interviewees, the interviewee response estimation system 1 of the present invention allows for the selection of interviewees to be completed in just a few tens of seconds.

[0062] Here, the attribute details settings have no input restrictions and allow you to freely enter text, so you can describe "studying using a learning app" in a bulleted list, for example, "user of a learning app, enjoys studying, has been using the app for more than a year," or you can express it in a sentence, such as "Someone who uses a learning app on the train during their commute and finds studying with the app enjoyable."

[0063] Next, once user Q has finished selecting the survey subjects (gathered all the AI ​​consultants), they can use the chat function to ask questions. The questions are then input to the machine learning device 10 via the question input unit 13 (more precisely, input to the selected learning model), and the estimated answers of the survey subjects are output via the answer output unit 14 as the output result (the response of the selected learning model). In this way, the answers to the questions are output to user Q simultaneously and immediately.

[0064] To elaborate, the learning model corresponding to each respondent with a paired set is a learning model that can simulate the thinking and behavioral patterns of that respondent. Therefore, the answers that this learning model generates for each of the multiple questions are highly likely to match the answers of that respondent. Thus, the answers of each respondent with a paired set can be obtained virtually from the learning model corresponding to that respondent.

[0065] Figure 5 shows that all respondents answered the question, "How often do you read the newspaper?" immediately and simultaneously.

[0066] Also, if asked another question, "What are your goals for studying?", Ayaka Yamada (18-year-old female): "My goal is to get into university." Takashi Takada (19-year-old male): "I am studying psychology, which is my favorite subject, diligently with the goal of going to university." • Emi Toda (13-year-old female): "My goal is to improve my school grades and become more fluent in English." Ryoichi Nakamura (19-year-old male): "I'm aiming to take the university entrance exam, so I'm studying with the goal of acquiring the knowledge necessary to pass it." As you can see, each survey participant gives a different answer.

[0067] In this way, the response output unit 14 outputs multiple estimated responses from multiple learning models to the same question data at once, so that user Q can naturally experience a scenario where they are asking the same question to multiple survey subjects and receiving answers simultaneously (without selecting any special settings).

[0068] In this process, each digital clone corresponding to a research participant is a persona representing participant P1 to Pn, complete with their name and a photograph. Therefore, even though it's an AI, it creates an atmosphere that closely resembles a real conversation.

[0069] Thus, according to the survey subject response estimation system 1 of this embodiment, since the response is output from a digital clone of a person image that reflects the persona of the original survey subject corresponding to the learning model, user Q can achieve a more natural feel, as if they were exchanging questions with the original survey subject.

[0070] Furthermore, it is preferable that the learning models constructed for the survey subjects be modified over time to reflect the passage of time since their construction. In other words, by modifying multiple constructed learning models over time, the digital clones can be aligned with the passage of time, allowing the interviewer to experience a more natural interaction, as if they were asking questions to the original survey subjects, including the passage of time. For example, consider a scenario where one year has passed since the learning model was constructed. The subject of the learning model Li, Pi, is assumed to have aged +1 year. Here, we analyze and extract changes in attributes such as hobbies and preferences associated with aging from an overall analysis of subjects P1 to Pn. For example, we might find that subjects aged A+1 year tend to have a greater awareness of and interest in health compared to a given age A. Then, by incorporating (retraining) the changes in characteristics and trends of Pi's attributes, or similar attributes, at age +1 year into Pi's learning model Li as a one-year timescale, we can apply the effects of time progression.

[0071] Furthermore, if user Q wishes to have a more detailed conversation with a specific research participant, such as in a depth interview where a moderator and research participant discuss a specific topic in depth, they can use the "individual chat" function to ask questions only to that specific research participant.

[0072] For example, in the above example, the AI ​​consultant, Ms. Yamada, answered that she was "using a learning app to pass university entrance exams," Q. Which subjects do you use the app for studying? A. Mainly mathematics and English Q. Why math and English? A. Because I think it's important for entrance exams and future higher education. ...In this way, just like in a group interview with real people, it becomes possible to delve into Yamada's individual answers and explore the deeper aspects of his consciousness, such as "why he thinks that way."

[0073] Furthermore, if there is no question data, the response output unit 14 refers to the database 40 and outputs the respondent's information, including a self-introduction associated with the respondent's persona. Of course, if a self-introduction is requested as part of the question data, it will refer to the database 40 and output the same respondent information.

[0074] This means that while a conversation might not start if there is no question data from user Q, if there is no question data, the digital clone can output information about the research subject, including a self-introduction linked to the persona, allowing user Q to naturally understand the research subject (more precisely, the digital clone corresponding to the research subject), prompting questions and initiating a conversation.

[0075] In the case of no question data, it means that user Q has not entered, started, or completed a question within the given period.

[0076] Furthermore, the answer output unit 14 may output a question corresponding to the characteristics of the question content by extracting the characteristics of the question content from the question data of user Q's question.

[0077] Specifically, by employing various existing technologies related to language analysis and natural language processing, including text mining, the characteristics of the questions asked by user Q can be extracted. Then, questions corresponding to the extracted characteristics are asked back to user Q from a digital clone, thereby achieving a natural conversation that is as if user Q and the original research subject were asking each other questions.

[0078] In this case, the question genres may be selected from those of high interest according to the characteristics of the questions asked by user Q. Furthermore, if there are trends, characteristics, or biases in the characteristics of the questions asked by user Q (when information entropy is low), or if there are no trends, characteristics, or biases (when information entropy is high), the question genres may be selected from those of high interest according to the characteristics.

[0079] Furthermore, the response output unit 14 may output questions corresponding to the characteristics of the content by extracting the characteristics of the response content from the predicted response output.

[0080] Similarly, by employing various existing technologies related to language analysis and natural language processing, it is possible to extract features from the predicted output answers and then ask questions corresponding to those extracted features back to the user Q from the digital clone, thereby achieving a natural conversation as if the user Q and the original survey subject were asking each other questions.

[0081] In this way, when a question is output from the answer output unit 14, the conversation is repeated by using the answer from user Q to that question as new question data. This makes it possible to encourage user Q to ask questions.

[0082] In other words, by inputting the user Q's response to the question output from the response output unit 14 as question data via the question input unit 13, the conversation can be repeated and continued using the user Q's response as new question data, enabling a more natural exchange of questions between user Q and the actual survey subject.

[0083] Here, if the machine learning device 10 outputs a question using the answer output unit 14, it may build a learning model of user Q by learning from the user Q's answers. That is, a learning model built to be able to estimate the answers that user Q will give to a question may be stored in the learning model storage unit as a digital clone of user Q.

[0084] This allows user Q to interact with the actual survey subjects in a more natural way, as if they were asking questions directly. As a result, a learning model can be constructed that incorporates user Q's questions and answers into various survey data, enabling the creation of a digital clone of user Q.

[0085] In addition, the response output unit 14 may include an attribute estimation unit (not shown) that extracts characteristics of the question content from the user Q's question data and estimates the questioner's attributes from the characteristics of the question content.

[0086] As described above, by enabling a more natural interaction between user Q and the actual survey subjects, it becomes possible to estimate the interviewer's attributes (true attributes that even the interviewer themselves may not be aware of) from the characteristics of user Q's questions, as part of the various survey data.

[0087] The user attributes (personas) estimated here include not only demographic attributes such as gender and age group, but also psychological attributes such as interests.

[0088] For example, the attribute estimation unit consists of a machine learning model that has been trained using machine learning on feature data and ground truth data based on survey results such as questionnaires.

[0089] The above describes the details and operation of the survey subject response estimation system 1 and its machine learning device 10 according to this embodiment. This survey subject response estimation system 1 and its machine learning device 10 make it possible to achieve a more natural feel, as if the interviewer were having a question-and-answer session with the actual survey subject.

[0090] In this embodiment, the questioner Q is assumed to be a person, but the questioner Q may be a digital clone. In this case, since it is possible to create copies of the digital clone, for example, if the original digital clone was formed from a large-scale language model (learning model), and the copy destination is a slightly smaller (smaller number of parameters) large-scale language model, then by repeating the question conversation, a copy of the digital clone can be created with an even smaller language model. [Explanation of Symbols]

[0091] 1... Survey participant response estimation system, 10... Machine learning device, 11... Learning model storage unit, 12... Pair set extraction unit, 13... Question input unit, 14... Response output unit, 20... Response aggregation unit, 30... Question data generation unit, 40... Survey participant database, 50... Communication device, 60... Communication terminal, P... Survey participant, Q... User (questioner), LM... Learning model, C... Digital clone.

Claims

1. A survey participant response estimation system that estimates and outputs responses from survey participants, The machine learning device comprises a machine learning device that constructs and stores learning models for multiple survey subjects using machine learning processing, The aforementioned machine learning device is A learning model storage unit stores multiple learning models, each constructed to be able to estimate the response given by each survey participant to a question by learning the responses of multiple survey participants, as digital clones of the multiple survey participants. A question input unit that obtains questions from multiple survey subjects and inputs question data representing the content of those questions into the multiple learning models, The question input unit receives question data and outputs an answer that the multiple learning models have estimated to be the answers of multiple survey subjects. It is equipped with, The survey participant response estimation system is characterized in that the response output unit outputs multiple estimated responses from the multiple learning models for the same question data via digital clones that represent the personas of the multiple survey participants, whose responses were trained to construct the multiple learning models, using human images.

2. In the survey participant response estimation system described in claim 1, The respondent response estimation system is characterized in that, when the response output unit does not have the question data, it outputs information about the survey subject, including a self-introduction associated with the persona.

3. In the survey participant response estimation system described in claim 1, The aforementioned response output unit is characterized by extracting characteristics of the question content from the question data and outputting a question corresponding to those characteristics of the question content, thereby providing a target respondent estimation system.

4. In the survey participant response estimation system described in claim 1, The aforementioned response output unit is characterized by extracting characteristics of the response content from the outputted predicted response and outputting a question corresponding to those characteristics of the content.

5. In the survey participant response estimation system according to claim 3 or 4, A survey participant response estimation system characterized by repeating a conversation by using the responses from the questioner to the questions output by the response output unit as the question data.

6. In the survey participant response estimation system described in claim 5, The machine learning device is characterized in that it stores a learning model, constructed to be able to estimate the answers that an interviewer will give to a question by learning the answers from the interviewer to the questions output by the answer output unit, as a digital clone of the interviewer in the learning model storage unit.

7. In the survey participant response estimation system described in claim 1, The aforementioned response output unit is characterized by comprising an attribute estimation unit that estimates the attributes of the questioner from the characteristics of the question content by extracting characteristics of the question content from the question data.

8. In the survey participant response estimation system according to claim 3 or 4, A target respondent estimation system characterized by comprising an attribute estimation unit that estimates the attributes of the questioner based on the characteristics of the content of the answer from the questioner to the question output by the answer output unit.

9. In the survey participant response estimation system described in claim 1, The learning model memory unit is characterized by changing the constructed learning models over time, thereby providing a system for estimating responses from survey participants.

Images