WO2022034922A1

WO2022034922A1 - System, program and method for questionnaire surveys

Info

Publication number: WO2022034922A1
Application number: PCT/JP2021/029806
Authority: WO
Inventors: ホースーンカン
Original assignee: 株式会社オルツ; ホースーンカン
Priority date: 2020-08-14
Filing date: 2021-08-13
Publication date: 2022-02-17
Also published as: JP2022125096A

Abstract

The purpose of the present invention is to provide a system, etc., for questionnaire surveys which can execute digital-clone questionnaire surveys. This invention provides a system for questionnaire surveys, the system being provided with a receiving means which receives questions for the questionnaire survey, multiple personalized AIs which are each configured to output a response to a received question, and an aggregating means which aggregates output from each of the multiple personalized AIs.

Description

Systems, programs, and methods for surveys

The present invention relates to a system, a program, and a method for a questionnaire survey, and more specifically, to a system, a program, and a method for a questionnaire survey using a plurality of personalized AIs. The present invention also relates to a method for creating a personalized AI and the like. More specifically, it relates to a method of creating a personalized AI that can output an answer to a question.

Conventionally, for questionnaire surveys, it was necessary to have multiple people answer and aggregate the answers. For example, Patent Document 1 discloses a system in which members participating in a questionnaire are made to answer a questionnaire via the Internet.

Japanese Unexamined Patent Publication No. 2019-12511

An object of the present invention is to provide a system for a questionnaire survey, etc., which can realize a digital clone questionnaire survey. Furthermore, it is an object of the present invention to provide a method for creating a personalized AI that can be used in a system for questionnaire surveys and the like.

In one embodiment, the present invention provides, for example, the following items.
(Item 1)
It ’s a system for questionnaire surveys.
Receiving means to receive questions for questionnaire surveys,
A plurality of personalized AIs, each of which is configured to output an answer to the received question, and a plurality of personalized AIs.
A system including an aggregation means for aggregating the outputs from each of the plurality of personalized AIs.
(Item 2)
Each of the plurality of personalized AIs outputs a plurality of answers by repeating the output of the answer to the received question a plurality of times.
The system according to item 1, wherein the aggregation means aggregates the plurality of responses output from each of the plurality of personalized AIs.
(Item 3)
Each of the plurality of personalized AIs has learned the expressions of individuals different from each other, and each of the plurality of personalized AIs can generate sentences according to the expressions of the learned individuals.
Each of the plurality of personalized AIs outputs a sentence according to the expression of the learned individual as an answer to the question.
The system according to item 2.
(Item 4)
Further equipped with emotion recognition means that can recognize the emotions contained in the text,
The emotion recognition means is
Recognizing the multiple emotions contained in the plurality of answers output from each of the plurality of personalized AIs.
It is configured to generate and generate output based on the multiple perceived emotions described above.
The aggregation means aggregates the output based on the plurality of emotions.
The system described in item 3,
(Item 5)
The emotion recognition means is
When the first emotion is the most among the plurality of emotions, the first output is generated.
The second output is generated when the second emotion is the most among the plurality of emotions.
The system according to item 4.
(Item 6)
The system according to any one of items 3 to 5, wherein each of the plurality of personalized AIs is learning the expression of the individual without correcting the bias of the expression of the individual.
(Item 7)
The system according to item 6, wherein the expression of the individual is an expression by an existing individual.
(Item 8)
The system according to item 6, wherein the expression of the individual is an expression by a virtual individual.
(Item 9)
It ’s a method for questionnaire surveys.
Receiving questions for surveys and
Using multiple personalized AIs to output multiple answers to the question and
A method comprising aggregating the outputs from each of the plurality of personalized AIs.
(Item 10)
A program for a questionnaire survey, the program is executed in a computer system including a processor unit, and the program is
Receiving questions for surveys and
Using multiple personalized AIs to output multiple answers to the question and
A program that causes the processor unit to aggregate the outputs from each of the plurality of personalized AIs.
(Item 11)
How to create a personalized AI that learns personal expressions,
Preparing a model that can generate sentences,
A method comprising having the model learn the individual's expression without correcting the bias of the individual's expression.
(Item 12)
It is a method to create a personalized AI that can output the answer to the question.
To get the result of an individual answering a question,
Preparing multiple models that can generate sentences,
By training the plurality of models to learn the expressions of a plurality of individuals, a plurality of trained models are generated, and one of the plurality of models is one of the expressions of the plurality of individuals. By training, one trained model is generated, and the representations of the plurality of individuals include the representations of at least one virtual individual.
To output the answer to the question from each of the plurality of trained models,
A method comprising determining one trained model based on a plurality of answers output from the plurality of trained models and the results.
(Item 13)
The expressions of the plurality of individuals are
To get multiple life logs and
12. The method of item 12, wherein the plurality of lifelogs are prepared by extracting representations of a plurality of individuals from the plurality of lifelogs, the plurality of lifelogs comprising at least one virtual lifelog.
(Item 14)
13. The method of item 13, wherein the at least one virtual lifelog is generated from at least one personalized AI.
(Item 15)
The method according to any one of items 12 to 14, wherein determining one trained model comprises determining one trained model that outputs the answer closest to the result answered by the individual. ..
(Item 16)
Determining the one trained model mentioned above
(1) Outputting at least one virtual individual expression from at least one of the plurality of trained models.
(2) To generate at least one second trained model by having the plurality of models learn the expression of at least one virtual individual that is output, and among the plurality of models. By training one of the expressions of one virtual individual, one second trained model is generated.
(3) Outputting the answer to the question from the at least one second trained model, and
(4) One trained model is determined based on the plurality of answers output from the plurality of trained models, the answers output from the at least one second trained model, and the above results. The method according to any one of items 12 to 15, including the above.
(Item 17)
Determining the one trained model mentioned above
By repeating the above (1) to the above (3) until a predetermined condition is satisfied, the above (1) expresses at least one virtual individual from the at least one second trained model. Further including output, the above (4) is
The method according to item 16, wherein one trained model satisfying a predetermined condition is specified from the plurality of trained models and at least one second trained model.
(Item 18)
The predetermined condition is one of a plurality of answers output from the plurality of trained models and one of the answers output from the at least one second trained model, and the result of the individual's answer. 17. The method of item 17, wherein the similarity of is above a predetermined threshold.
(Item 19)
The system according to any one of items 1 to 8, wherein the plurality of personalized AIs include personalized AIs prepared according to the method according to items 11 to 17.
(Item 20)
A system for creating personalized AI that can output answers to questions.
How to get the result of an individual answering a question,
A means of preparing multiple models capable of generating sentences,
It is a means for generating a plurality of trained models by having the plurality of models learn the expressions of a plurality of individuals, and one of the plurality of models is one of the expressions of the plurality of individuals. By training, one trained model is generated, and the representation of the plurality of individuals includes means and means including the representation of at least one virtual individual.
A means for outputting an answer to the question from each of the plurality of trained models,
A system including a plurality of answers output from the plurality of trained models and a means for determining one trained model based on the results.
(Item 21)
A program for creating a personalized AI capable of outputting an answer to a question, the program is executed in a computer system including a processor unit, and the program is
To get the result of an individual answering a question,
Preparing multiple models that can generate sentences,
By training the plurality of models to learn the expressions of a plurality of individuals, a plurality of trained models are generated, and one of the plurality of models is one of the expressions of the plurality of individuals. By training, one trained model is generated, and the representations of the plurality of individuals include the representations of at least one virtual individual.
To output the answer to the question from each of the plurality of trained models,
A program for causing the processor unit to perform a process including determining one trained model based on a plurality of answers output from the plurality of trained models and the results.

According to the present invention, it is possible to provide a system for a questionnaire survey using a plurality of personalized AIs, and a digital clone questionnaire survey can be realized by these systems and the like. This makes it possible to conduct a questionnaire survey at a higher speed and at a lower price than a conventional questionnaire survey in which a plurality of people are asked to answer.

A diagram showing an example of the flow of the conventional questionnaire survey method by the system S for the conventional questionnaire survey that allows multiple people to answer the questionnaire. The figure which shows an example of the method of the digital clone questionnaire investigation by this invention. The figure which shows an example of the user interface for the digital clone questionnaire investigation by this invention. The figure which shows an example of the user interface for the digital clone questionnaire investigation by this invention. The figure which shows an example of the user interface for the digital clone questionnaire investigation by this invention. A diagram showing an example of the configuration of the system 100 for a questionnaire survey. The figure which shows an example of the structure of the processor part 120 The figure which shows an example of the processor part 120'which is an alternative embodiment of the processor part 120 A flowchart showing an example (process 500) of processing by the system 100 for a questionnaire survey. Flow chart showing the detailed flow of step S502 in process 500 A flowchart showing another example of processing by the system 100 for a questionnaire survey (processing 700). A flowchart showing another example of processing by the system 100 for a questionnaire survey (processing 800). Conceptual diagram of an example of Flow 900 for creating personalized AI using real personal information Conceptual diagram of an example of Flow 1000 for creating personalized AI using non-existent virtual personal information A flowchart showing another example of processing by the system 100 for a questionnaire survey (processing 1100). A flowchart showing an example of a specific step of step S1105 in the process 1100. A flowchart showing another specific example of the specific step of step S1105 in the process 1100.

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

1. 1. Digital Clone Questionnaire Survey Figure 1A shows an example of the flow of a conventional questionnaire survey method in which a plurality of people are asked to answer a questionnaire. In the example shown in FIG. 1A, it will be described that the questionnaire survey is performed using the conventional system S for the questionnaire survey.

In step S1, the questioner Q inputs a question to the system S. The questioner Q can input a question to the system S via the information terminal, for example, by inputting the question to his / her own information terminal. Upon receiving the question, the system S _sends the question to each of the plurality of respondents RM ₁ to RM in order to have each of the plurality of respondents RM ₁ to _RM answer the question. In the present specification, a single alphabetic character such as "M", "m", "N", "n" means an integer of 2 or more unless otherwise indicated.

In step S2, each of the plurality of respondents _R1 to _RM receives a question from the system S. Each of the plurality of respondents _R1 to _RM can receive a question via, for example, their own information terminal.

In step S3, each of the plurality of respondents R ₁ to _RM inputs the answer to the question into the system S. Each of the plurality of respondents R ₁ to _RM can input the answer to the system S via the information terminal, for example, by inputting the answer to the question into its own information terminal.

When the system S receives the answers from each of the plurality of respondents _R1 to _RM , the system S aggregates the answers.

In step S4, the system S provides the aggregated answer to the questioner Q. The questioner Q can receive, for example, the aggregated answers via his / her own information terminal. As a result, the questioner Q can obtain the result of the questionnaire survey for the question.

In this way, with the conventional questionnaire survey method, it is necessary for each respondent to answer the question, which is time consuming and costly. In particular, the greater the number of respondents, the greater the time and cost required to obtain results.

The inventor of the present invention has developed a new questionnaire survey method that replaces the conventional questionnaire survey. It is a digital clone survey that uses "digital clones" to answer questions on behalf of humans. A "digital clone" can be realized by a "personalized AI" that can perform an activity unique to an individual by learning personal information, for example, an individual's behavior or thought. Here, the personal information may be the information of a real person or the information of a virtual person who does not exist. For example, a "personalized AI" that has learned information obtained from a real person can perform activities that are more like that person. For example, even if the information obtained from a real person is not available, a personalized AI can be constructed by using the information of a virtual person.

FIG. 1B shows an example of the method of the digital clone questionnaire survey according to the present invention. In the example shown in FIG. 1B, the answers obtained by having a plurality of "personalized AIs" answer the questionnaire are aggregated, and the aggregated results are shown to the questioner as the result of the questionnaire survey.

In step S11, the questioner Q inputs a question to the system 100. The questioner Q can input a question to the system 100 via the information terminal, for example, by inputting the question to his / her own information terminal.

In step S12, a question is provided to each of the plurality of personalized AI PAI ₁ to PAI _M in the system 100. Multiple personalized AI PAI ₁ to PAI _M are AI (artificial intelligence) that have learned information of different individuals. When a question is asked to a plurality of personalized AI PAI ₁ to PAI _M , each of the plurality of personalized AI PAI ₁ to PAI _M can output an answer to the question.

In step S13, in the system 100, each of the plurality of personalized AI PAI ₁ to PAI _M outputs an answer to the question.

In step S14, the system 100 provides the aggregated answer to the questioner Q. The questioner Q can receive, for example, the aggregated answers via his / her own information terminal. As a result, the questioner Q can obtain the result of the questionnaire survey for the question.

Thus, in the digital clone questionnaire survey according to the present invention, a plurality of personalized AIs answer questions on behalf of a plurality of respondents, which can greatly reduce the time and cost required to obtain results. can. This makes it possible, for example, to lower the hurdles for questionnaire surveys and promote the use of questionnaire surveys in a wide variety of fields.

2A-2C show an example of a user interface for a digital clone questionnaire survey according to the present invention. The user interface shown in FIGS. 2A to 2C is displayed, for example, on the display unit of the information terminal of the questioner Q. The user interface shown in FIGS. 2A to 2C may be, for example, a screen displayed on the information terminal of the questioner Q by a Web application, or may be a screen displayed on the information terminal of the questioner Q by the application installed on the information terminal of the questioner Q. It may be a screen displayed on the information terminal of.

FIG. 2A shows an example of the question input screen 10 for the digital clone questionnaire survey.

The question input screen 10 includes a question input unit 11, a survey detail setting button 12, an estimated amount display unit 13, and an ASK button 14.

Questioner Q can enter a question in the question input unit 11. It doesn't matter how you enter the question. For example, the question may be input on a text basis via a keyboard or the like, or the question may be input on a voice basis via a microphone or the like.

The survey detail setting button 12 is a button selected when setting the details for the questionnaire survey, and when the survey detail setting button 12 is selected, the survey detail setting unit 15 shown in FIG. 1B is expanded. ..

The estimated amount for the questionnaire survey is displayed on the estimated amount display unit 13. Estimates can vary, for example, depending on the details for the survey.

The ASK button 14 is a button for starting the questionnaire survey. When the ASK button 14 is selected, the digital clone questionnaire survey is started, and after the digital clone questionnaire survey is completed, the result display screen 20 of FIG. 2C is displayed. Transition.

FIG. 2B shows an example of the question input screen 10 in which the survey detail setting unit 15 is expanded.

Questioner Q can enter the details of the survey in the survey details setting unit 15. It doesn't matter how you enter the question. For example, the survey details may be entered text-based via a keyboard or the like, or the survey details may be entered voice-based via a microphone or the like, or from a plurality of options. You may choose the details of the survey. In the example shown in FIG. 2B, as details for the survey, the number of samples of the subjects of the questionnaire survey, the accuracy of the questionnaire survey, the gender of the subjects of the questionnaire survey, and the age of the subjects of the questionnaire survey can be set. .. Any other matter can be set as details for the investigation.

FIG. 2C shows an example of the response display screen 20 by the digital clone questionnaire survey.

The result display screen 20 displays the result of the digital clone questionnaire survey. The result display screen 20 includes a condition display unit 21, a result display unit 22, a required time display unit 23, a graph display unit 24, and a Next Question button 25.

The condition display unit 21 displays the conditions under which the digital clone questionnaire survey was conducted. In the example shown in FIG. 2C, since the questionnaire survey was conducted on 10,000 women aged 20 to 25 years, the conditions are displayed on the condition display unit 21.

The result of the digital clone questionnaire survey is displayed in text on the answer display unit 22. In the example shown in Figure 2C, the result is displayed that 4000 people answered that they would support Mr. XX and 6000 people did not support the question "Do you support Mr. XX in the next presidential election?" Has been done.

The required time display unit 23 displays the time required for the digital clone questionnaire survey. In the example shown in FIG. 2C. It is displayed that the required time was 100 seconds.

The result of the digital clone questionnaire survey is displayed as a graph on the graph display unit 24. In the example shown in FIG. 2C, a pie chart shows the results of a digital clone questionnaire survey. The graph display unit 24 can display the results of the digital clone questionnaire survey in any other format of the graph.

The Next Question button 25 is a button for proceeding to the digital clone questionnaire survey in the next question. For example, when the Next Question button 25 is selected, the screen transitions to the question input screen 10.

In this way, the questioner Q can obtain the result of the questionnaire survey for the question only by inputting the question for the questionnaire survey on the question input screen 10.

The system 100 described above can be achieved by a system for a questionnaire survey described later. The system for questionnaire surveys utilizes multiple personalized AIs.

2. 2. Configuration of the System for Questionnaire Survey Figure 3 shows an example of the configuration of the system 100 for questionnaire survey.

The system 100 is connected to the database unit 200. Further, the system 100 is connected to at least one terminal device 300 via the network 400.

Although three terminal devices 300 are shown in FIG. 3, the number of terminal devices 300 is not limited to this. Any number of terminal devices 300 may be connected to the system 100 via the network 400.

The network 400 can be any kind of network. The network 400 may be, for example, the Internet or a LAN. The network 400 may be a wired network or a wireless network.

An example of the system 100 is a computer (for example, a server device) installed in a service provider that provides a service for a digital clone questionnaire survey, but the present invention is not limited to this. An example of the terminal device 300 is a computer (for example, a terminal device) used by the questioner, but the present invention is not limited to this. Here, the computer (server device or terminal device) can be any type of computer. For example, the terminal device can be any type of terminal device such as a smartphone, tablet, personal computer, smart glasses, smart watch or the like.

The system 100 includes an interface unit 110, a processor unit 120, and a memory unit 130.

The interface unit 110 exchanges information with the outside of the system 100. The processor unit 120 of the system 100 can receive information from the outside of the system 100 via the interface unit 110, and can transmit the information to the outside of the system 100. The interface unit 110 can exchange information in any format.

The interface unit 110 includes, for example, an input unit that enables information to be input to the system 100. It does not matter in what manner the input unit enables information to be input to the system 100. For example, when the input unit is a touch panel, the user may input information by touching the touch panel. Alternatively, when the input unit is a mouse, the user may input information by operating the mouse. Alternatively, when the input unit is a keyboard, the user may input information by pressing a key on the keyboard. Alternatively, when the input unit is a microphone, the user may input information by inputting voice into the microphone. Alternatively, when the input unit is a camera, the information captured by the camera may be input. Alternatively, when the input unit is a data reading device, the information may be input by reading the information from the storage medium connected to the system 100. Alternatively, when the input unit is a receiver, the receiver may input information by receiving information from the outside of the system 100 via the network. In this case, the type of network does not matter. For example, the receiver may receive the information via the Internet or may receive the information via the LAN.

For example, the input unit of the interface unit 110 may receive a question for a questionnaire survey from the terminal device. The question for the questionnaire survey may be, for example, a text-based question or a voice-based question.

For example, the input unit of the interface unit 110 may receive the results of a questionnaire survey conducted in the past. The results of the questionnaire survey conducted in the past may be received from, for example, the terminal device, the database unit 200, or another place. The questionnaire survey conducted in the past may be, for example, a multiple-choice questionnaire survey, or a descriptive or essay-type questionnaire survey.

The interface unit 110 includes, for example, an output unit that enables information to be output from the system 100. It does not matter in what mode the output unit enables the information to be output from the system 100. For example, when the output unit is a display screen, information may be output to the display screen. Alternatively, when the output unit is a speaker, the information may be output by the voice from the speaker. Alternatively, when the output unit is a data writing device, the information may be output by writing the information to the storage medium connected to the system 100. Alternatively, when the output unit is a transmitter, the transmitter may output information by transmitting information to the outside of the system 100 via a network. In this case, the type of network does not matter. For example, the transmitter may transmit information via the Internet or may transmit information via LAN.

For example, the output unit of the interface unit 110 may transmit the result of the questionnaire survey to the terminal device.

The processor unit 120 executes the processing of the system 100 and controls the operation of the entire system 100. The processor unit 120 reads the program stored in the memory unit 130 and executes the program. This makes it possible to make the system 100 function as a system that performs a desired step. The processor unit 120 may be implemented by a single processor or may be implemented by a plurality of processors.

The memory unit 130 stores a program required for executing the processing of the system 100, data required for executing the program, and the like. The memory unit 130 creates a program (for example, a program that realizes the processing shown in FIGS. 5 to 7 described later) and / or a personalized AI for causing the processor unit 120 to perform the processing for the questionnaire survey. A program for causing the processor unit 120 to perform the processing (for example, a program for realizing the processing shown in FIGS. 8, 11, 12A, and 12B described later) may be stored. Here, it does not matter how the program is stored in the memory unit 130. For example, the program may be pre-installed in the memory unit 130. Alternatively, the program may be installed in the memory unit 130 by being downloaded via the network. In this case, the type of network does not matter. The memory unit 130 may be implemented by any storage means.

The database unit 200 stores, for example, data for creating a personalized AI. Personalized AI is an AI that can perform personal activities by learning personal information. For example, a personalized AI can be a personalized sentence generation model capable of learning an individual's expression and generating a sentence according to the individual's expression. For example, in the database unit 200, when a prompt (writing a sentence) is input as data for creating a personalized AI, a sentence generation model capable of generating a subsequent sentence and an individual expression are learned. An algorithm for doing so and an individual expression used for learning can be stored. Here, the personal information may be the information of a real person or the information of a virtual person who does not exist. When personal information is information of a virtual person, personal information does not have to be human-understandable information. The information of the virtual person may be, for example, a list of characters or symbols having no meaning. The personal information used for learning can be, for example, information contained in a life log. The life log may be a life log of a real person or a virtual life log of a virtual person.

Alternatively, the database unit 200 stores, for example, a created personalized AI (for example, a trained model that has learned personal information).

In the example shown in FIG. 2, the database unit 200 is provided outside the system 100, but the present invention is not limited thereto. It is also possible to provide the database unit 200 inside the system 100. At this time, the database unit 200 may be implemented by the same storage means as the storage means on which the memory unit 130 is mounted, or may be implemented by a storage means different from the storage means on which the memory unit 130 is mounted. In any case, the database unit 200 is configured as a storage unit for the system 100. The configuration of the database unit 200 is not limited to a specific hardware configuration. For example, the database unit 200 may be composed of a single hardware component or may be composed of a plurality of hardware components. For example, the database unit 200 may be configured as an external hard disk device of the system 100, or may be configured as a storage on a cloud connected via a network.

FIG. 4A shows an example of the configuration of the processor unit 120.

The processor unit 120 includes a receiving means 121, an answer generating means 122, and an aggregation means 123.

The receiving means 121 is configured to receive a question for a questionnaire survey. The receiving means 121 receives, for example, a question for a questionnaire survey received from the outside of the system 100 via the interface unit 110 from the interface unit 110. The question for the questionnaire survey may be, for example, a text-based question or a voice-based question.

The received questionnaire survey questions may be converted into a processable format by subsequent means. For example, if the question for the questionnaire survey is a voice-based question, it may be converted into a text-based question.

The answer generation means 122 is configured to generate a plurality of answers to a question for a questionnaire survey by using a plurality of personalized AI PAI ₁ to PAI _M. Each of the plurality of personalized AI PAI ₁ to PAI _M is configured to output the answer to the question. Each of the plurality of answers generated by the answer generation means 122 may correspond to the answers output from each of the plurality of personalized AI PAI ₁ to PAI _M.

Each of the plurality of personalized AI PAI ₁ to PAI _M is learning information of different individuals. This allows each of the personalized AI PAI ₁ to PAI _M to carry out their own personal activities. For example, a personalized AI PAI _M can output a personalized answer to an input question. The answer may be, for example, a text-based answer or a voice-based answer.

For example, each of a plurality of personalized AI PAI ₁ to PAI _M can be a personalized sentence generation model capable of learning an individual's expression and generating a sentence according to the individual's expression. In this case, each of the plurality of personalized AI PAI ₁ to PAI _M can generate an answer to the input question with a sentence according to the expression of the learned individual. The inventor of the present invention considers that the answer generated by the personalized sentence generation model reflects the individual's thought or emotion because the expression of the individual reflects the individual's thought or emotion. I thought I could do it. By using the answers generated by the personalized sentence generation model for the questionnaire survey, it is possible to obtain highly accurate questionnaire survey results that reflect the thoughts or feelings of the individual.

For example, the personalized sentence generation model may be a model created from scratch by learning a large amount of personal information, or it may be created by fine-tuning a general-purpose sentence generation model with personal information. It may be a model. Preferably, the personalized sentence generation model can be a model created by fine-tuning a general-purpose sentence generation model with personal information. Creating a personalized sentence generation model by fine-tuning can significantly reduce the amount of training data required for fine-tuning compared to creating a model from scratch, and fine-tuning This is because the cost (time and memory) required for this can be much lower than when creating a model from scratch.

In one example, the personalized sentence generation model can be created by fine-tuning the general-purpose sentence generation model GPT-2model developed by OpenAI with personal information. GPT-2model is a model that can generate the following sentences when a prompt (writing of sentences) is given. Currently, GPT-2model is available in languages such as English, Chinese, Italian and Korean. When creating a model in a language other than the available languages, for example, Japanese, a sentence generation model that can generate sentences in that language by training GPT-2model with sentences in that language. And fine-tune the created sentence generation model with personal information, it is possible to create a personalized sentence generation model. Note that GPT-2model is an example, and the present invention is not limited to a personalized sentence generation model based on GPT-2model. The personalized sentence generation model can be created based on other models, for example, depending on the application or performance.

Various personal information can be used to create a personalized AI. Here, the personal information may be the information of a real person or the information of a virtual person who does not exist. When the personal information is the information of a real person, the personal information includes, but is not limited to, for example, sentences created by the real individual (eg, email, chat, posting to SNS). For example, personal information may include, for example, a video of an actual individual's behavior, a voice spoken by an actual individual, and the like. When the personal information is the information of a non-existent person, the personal information may include, for example, a virtually created sentence, a virtually created video, a virtually created audio, and the like. The personal information used to create a personalized AI preferably contains personal emotions. This allows the personalized AI to generate an answer that reflects an individual's emotions.

Normally, when learning learning materials, it is necessary to create learning materials without bias in order to prevent bias in machine learning. This is because if the learning material is biased, the trained model outputs a biased result. When learning personal information in order to create a personalized AI, unlike ordinary machine learning, the personal information is learned without correcting the bias of the personal information. The created personalized AI will output a biased result, and this "bias" can be regarded as "individuality". In this way, the personalized AI can provide an output that includes "individuality".

Each of the plurality of personalized AI PAI ₁ to PAI _M can generate one answer, for example, when one question is input. By repeating this a plurality of times, each of the plurality of personalized AI PAI ₁ to PAI _M can output a plurality of answers. The output of each personalized AI does not always answer the question accurately. By outputting a plurality of answers from each personalized AI, it is possible to generate statistically accurate answers from each personalized AI. The number of repetitions may be any number of times, for example, 5 times or more, 10 times or more, 50 times or more, 100 times or more, 1000 times or more, 10000 times or more. The number of iterations can be changed depending on the accuracy of the answer. In order to improve the accuracy, the number of repetitions can be increased (for example, 100 times or more).

The aggregation means 123 is configured to aggregate a plurality of answers. The aggregation means 123 can aggregate a plurality of answers by any method. The aggregation means 123 may, for example, calculate the number of the first answer (for example, the answer of Yes) and the number of the second answers (for example, the answer of No) among the plurality of answers. The ratio of the first answer (for example, the answer of Yes) and the ratio of the second answer (for example, the answer of No) among the plurality of answers may be calculated.

For example, the aggregation means 123 can aggregate the respective responses from each of the plurality of personalized AI PAI ₁ to PAI _M. For example, the aggregation means 123 regards one response from one personalized AI as a response from the personalized AI, and responds from each of M personalized AI PAI ₁ to PAI _M (that is, M). Answer) can be aggregated. For example, the aggregation means 123 regards N answers from one personalized AI as answers from the personalized AI, and answers from each of M personalized AI PAI ₁ to PAI _M (that is, N × M). Individual answers) can be aggregated.

For example, the aggregation means 123 can aggregate the response sentences from each of the plurality of personalized AI PAI ₁ to PAI _M by classifying them into one of the plurality of categories. The plurality of categories includes, but are not limited to, for example, positive answers, negative answers, and the like. The aggregation means 123 can classify the response sentences from each of the plurality of personalized AI PAI ₁ to PAI _M into one of a plurality of categories by using, for example, natural language processing.

The output by the aggregation means 123 is output to the outside of the system 100 via the interface unit 110.

FIG. 4B shows an example of the processor unit 120'which is an alternative embodiment of the processor unit 120. The processor unit 120'is different from the processor unit 120 in that the processor unit 120'provides the emotion recognition means 124. In FIG. 4B, the same reference numbers are assigned to the same components as those described with reference to FIG. 4A, and the description of these components will be omitted here.

The processor unit 120'includes a receiving means 121, an answer generating means 122, an emotion recognition means 124, and an aggregation means 123.

The emotion recognition means 124 is configured to be able to recognize the emotions contained in the text. The emotion recognition means 124 can be achieved by any known emotion analysis engine capable of recognizing the emotions contained in the text.

The emotion recognition means 124 may be, for example, an emotion classifier (Intent Classifier) that classifies the emotions contained in the input sentence into a plurality of classifications. Multiple classifications can be any number of classifications. For example, the plurality of classifications may be two classifications such as negative / positive or gentle / severe, three classifications such as joy / anger / sadness, and joy / anger / sadness. / There may be 4 classifications such as comfort, or more than 4 classifications, for example, 100 classifications.

The emotion classifier can classify the emotions contained in the text into multiple categories using machine learning. Emotion classifiers can be created using any machine learning technique. The machine learning method for creating an emotion classifier can be, for example, a decision tree, a support vector machine, a deep neural network, or the like. Emotion classifiers can be created, for example, by supervised learning.

As learning data for creating an emotion classifier, for example, a sentence expressing emotions can be used. Sentences expressing emotions should be labeled to indicate which classification of emotions they are expressing. For example, an emotion classifier can be created by learning a sentence expressing emotions as input teacher data and learning the label of the sentence as output teacher data. A sentence expressing emotions can be expressed as a vector and then used as input teacher data. For example, a method using a word bag model (bag-of-words model), a method using a semantic expression vector (contextualized word embedding) of a contextual word (for example, BERT, ElMo, etc.), MUSE (Multilingual Universal Sentence https) /arxiv.org/abs/1907.04307) allows sentences to be represented as a vector.

For example, if you want to create an emotion classifier that outputs whether the emotions expressed by a sentence are positive or negative, you will learn sentences that express negative emotions and sentences that express positive emotions. It will be. The pair of input teacher data and output teacher data when learning the first sentence expressing negative emotions, the second sentence expressing positive emotions, etc. is ( It can be a vector expression of the first sentence, negative), (vector expression of the second sentence, positive), .... When a vector expression of a certain sentence is input to the emotion classifier learned in this way, it becomes possible to output whether the emotion expressed by the sentence is positive or negative.

For example, when creating an emotion classifier that outputs whether the emotion expressed by a sentence is joy, anger, or sadness, the sentence expressing the emotion of joy, the sentence expressing the emotion of anger, and the sentence expressing the emotion of anger. You will learn sentences that express feelings of sadness. For input when learning the first sentence expressing the feelings of joy, the second sentence expressing the feelings of anger, the third sentence expressing the feelings of sadness, etc. The pair of teacher data and output teacher data is (vector expression of the first sentence, joy), (vector expression of the second sentence, anger), (vector expression of the third sentence, sadness), ...・ It can be etc. When a vector expression of a sentence is input to the emotion classifier learned in this way, it becomes possible to output whether the emotion expressed by the sentence is joy, anger, or sadness.

The emotion recognition means 124 can recognize the emotions contained in the respective answers from each of the plurality of personalized AI PAI ₁ to PAI _M. The emotion recognition means 124 recognizes one emotion as an emotion contained in the answer from the personalized AI, for example, by recognizing the emotion contained in one answer from one personalized AI. Can be done. Alternatively, the emotion recognition means 124, for example, recognizes the emotions contained in each of the plurality of responses from one personalized AI, so that the plurality of emotions are included in the responses from the personalized AI. Can be recognized as. Alternatively, the emotion recognition means 124 recognizes, for example, the emotion contained in each of the plurality of responses from one personalized AI, and some emotions among the plurality of emotions are converted into the responses from the personalized AI. It can be recognized as an emotion. Some emotions among the plurality of emotions may be, for example, some of the top emotions with a large number of recognitions among the plurality of emotions, emotions with a predetermined number or more recognized, and the like. Alternatively, the emotion recognition means 124 recognizes, for example, the emotion contained in each of the plurality of responses from one personalized AI, and incorporates the emotion of one of the plurality of emotions into the response from the personalized AI. It can be recognized as an emotion. One emotion among the plurality of emotions may be, for example, the emotion with the highest number of recognitions among the plurality of emotions.

The emotion recognition means 124 can generate an output based on the recognized emotion, for example. The output may be, for example, an output indicating what the perceived emotion is, or an output indicating what the answer from the personalized AI is. The output may be, for example, an output indicating the ratio of each emotion among the plurality of recognized emotions. For example, the emotion recognition means 124 generates a first output when the first emotion is the most among the recognized plurality of emotions, and when the second emotion is the most among the recognized plurality of emotions. A second output can be generated. In one example, the emotion recognition means 124 has a first output (for example, an output indicating that the recognized emotion is positive, or a response from a personalized AI) when the positive emotion is the most among the plurality of emotions. Is an output that indicates Yes, and a second output (eg, an output that indicates that the recognized emotion is negative, or personalized) when the negative emotions are the most among the multiple emotions. An output indicating that the answer from AI is No) can be generated. In one example, the emotion recognition means 124 produces an output indicating (positive 60%, negative 40%) when the positive emotion is 60% and the negative emotion is 40% among the plurality of emotions. Can be done.

In this way, the emotion recognition means 124 recognizes M emotions from each one answer (that is, a total of M answers) from each of the M personalized AIs, and outputs an output based on the recognized emotions. It may be generated, or M × N emotions are recognized from each of N answers (that is, M × N answers) from each of the M personalized AIs, and the recognized emotions are converted into the recognized emotions. It may generate an output based on, or it may recognize and recognize M × n emotions from each of N answers (ie, M × N answers) from each of the M personalized AIs. It may be possible to generate an output based on emotions (N> n).

The aggregation means 123 can aggregate the output of the emotion recognition means 124. For example, the aggregation means 123 may aggregate the emotions indicated by the output of the emotion recognition means 124, or may aggregate the emotions indicated by the output of the emotion recognition means 124 in association with the answer to the question. .. The aggregation means 123 can, for example, aggregate the responses from the personalized AI indicated by the output by the emotion recognition means 124.

For example, the aggregation means 123 may calculate the number of first emotions (eg, positive emotions) and the number of second emotions (eg, negative emotions) among the plurality of emotions. The proportion of the first emotion (eg, positive emotion) and the second emotion (eg, negative emotion) of the plurality of emotions may be calculated. For example, the aggregation means 123 associates a first emotion (for example, a positive emotion) among a plurality of emotions with a first answer (for example, Yes's answer) and a second emotion (for example, a negative emotion). ) May be associated with a second answer (eg, No answer) to calculate their number or proportion.

In the examples shown in FIGS. 4A and 4B described above, the components of the processor units 120 and 120'are provided in the same processor units 120 and 120', but the present invention is not limited thereto. A configuration in which each component of the processor units 120 and 120'is distributed to a plurality of processor units is also within the scope of the present invention. At this time, the plurality of processor units may be located in the same hardware component, or may be located in separate hardware components in the vicinity or remote.

Note that each component of the system 100 described above may be composed of a single hardware component or may be composed of a plurality of hardware components. When it is composed of a plurality of hardware parts, the mode in which each hardware part is connected does not matter. Each hardware component may be connected wirelessly or may be connected by wire. The system 100 of the present invention is not limited to a specific hardware configuration. It is also within the scope of the present invention that the processor units 120 and 120'are configured by an analog circuit instead of a digital circuit. The configuration of the system 100 of the present invention is not limited to the above-mentioned one as long as the function can be realized.

3. 3. Processing by the system for questionnaire survey FIG. 5 is a flowchart showing an example (processing 500) of processing by the system 100 for questionnaire survey. The process 500 may be performed in the processor unit 120 of the system 100.

In step S501, the receiving means 121 of the processor unit 120 receives the question for the questionnaire survey. The receiving means 121 receives, for example, a question for a questionnaire survey received from the outside of the system 100 via the interface unit 110 from the interface unit 110.

In step S502, the answer generation means 122 of the processor unit 120 outputs a plurality of answers to the question received in step S501 by using the plurality of personalized AIs. The answer generation means 122 may generate M answers using, for example, M personalized AIs, or may generate N answers from each of the M personalized AIs. M × N answers or M × n answers may be generated (N> n).

Each of the multiple personalized AIs is learning information about different individuals. As a result, each of the plurality of personalized AIs can output a personalized answer to the input question. Each of the plurality of personalized AIs can be, for example, a personalized sentence generation model capable of learning an individual's expression and generating a sentence according to the individual's expression. At this time, each of the plurality of personalized AI PAI ₁ to PAI _M can generate an answer to the input question with a sentence according to the expression of the learned individual.

For example, personalized AI, which is a personalized sentence generation model, can generate subsequent sentences according to the personal expression learned by giving a prompt (writing a sentence), so that the prompt related to the question is personalized. You can get the answer to your question by entering it in the AI. The prompt related to the question may be selected from preset templates according to the type of question, may be created based on the question, or may be set by the questioner. It may be set by the service provider of the questionnaire survey service. If the prompt related to the question is created based on the question, for example, the prompt may be created based on a preset template, or machine learning may be used to generate the prompt. You may.

Personalized AI can generate a continuation of these prompts according to the personal expression learned. The generated text will be the answer to the question.

In step S503, the aggregation means 123 of the processor unit 120 aggregates the outputs from each of the plurality of personalized AIs. For example, the aggregation means 123 regards one response from one personalized AI as a response from the personalized AI, and responds from each of M personalized AI PAI ₁ to PAI _M (that is, M). Answer) can be aggregated. For example, the aggregation means 123 regards N answers from one personalized AI as answers from the personalized AI, and answers from each of M personalized AI PAI ₁ to PAI _M (that is, N × M). Individual answers) can be aggregated. The aggregation means 123 may, for example, calculate the number of the first answer (for example, the answer of Yes) and the number of the second answers (for example, the answer of No) among the plurality of answers. The ratio of the first answer (for example, the answer of Yes) and the ratio of the second answer (for example, the answer of No) among the plurality of answers may be calculated.

Processing 500 allows the answer to the question for the questionnaire survey to be obtained from each of the plurality of personalized AIs. This makes it possible to conduct a questionnaire survey at a higher speed and at a lower price than a conventional questionnaire survey in which a plurality of people are asked to answer the questionnaire.

FIG. 6 shows a detailed flow of step S502 in the process 500. FIG. 6 illustrates the steps for obtaining N answers from each of the M personalized AIs. The subscript _M in each step indicates that the process uses the Mth personalized AI. Here, N is an arbitrary integer of 2 or more, and may be an arbitrary integer such as 5, 10, 50, 100, 1000, 10000, and the like. N can be changed according to the accuracy of the response of the questionnaire survey. When increasing the accuracy, N can be increased (for example, 100 or more).

In step S601 ₁ , the variable n = 1 is defined.

In step S602 ₁ , the question received in step S501 is input to the first personalized AI.

In step S603 ₁ , the nth answer is output from the first personalized AI. Here, the first answer is output. The first personalized AI is learning information about the first person, for example, and can behave like the first person. Therefore, the first personalized AI can output an answer as if the first person answered the input question. For example, the first personalized AI is learning the expression of the first person, and can generate and output a sentence according to the expression of the first person as an answer to the input question. The output answer is stored in, for example, a memory.

In step S604 ₁ , the variable n is incremented.

In step S605 ₁ , the variable n is compared with N. When the variable n is N or more, the process using the first personalized AI can be completed. If the variable n is smaller than N, the process returns to step S602 ₁ and steps S602 ₁ to S605 ₁ are repeated until the variable n becomes N or more. As a result, N answers can be obtained from the first personalized AI.

Before, after, or at the same time as the process using the first personalized AI (steps S601 ₁ to S605 ₁ ), the process using the second personalized AI (steps S601 ₂ to S605 ₂ ), the third Processing using the personalized AI (step S601 ₃ to step S605 ₃ ), ... Processing using the third personalized AI (step S601 _M to step S605 _M ) can be performed. As a result, N answers can be obtained from each of the M personalized AIs. That is, a total of M × N answers can be obtained.

If answers are obtained from all of the M personalized AIs, the process can proceed to step S503.

The output of each personalized AI does not always answer the question accurately. By outputting a plurality of answers from each personalized AI in step S502, a statistically accurate answer can be generated from each personalized AI. This improves the accuracy of the questionnaire survey.

FIG. 7 is a flowchart showing another example (process 700) of the process by the system 100 for the questionnaire survey. The process 700 may be performed in the processor unit 120'of the system 100.

In step S701, the receiving means 121 of the processor unit 120'receives a question for a questionnaire survey. Step S701 is the same as step S501 described above, and the description thereof will be omitted.

In step S702, the answer generation means 122 of the processor unit 120'outputs a plurality of answers to the question received in step S501 by using the plurality of personalized AIs. The answer generation means 122 can output an answer in the form of a sentence, for example. Step S702 is the same as step S502 described above, and the description thereof will be omitted.

In step S703, the emotion recognition means 124 of the processor unit 120'recognizes the emotions contained in the answers output from each of the plurality of personalized AIs. The emotion recognition means 124 can be, for example, an emotion classifier that classifies the emotions contained in the text into a plurality of classifications, and the emotions contained in the answers output from each of the plurality of personalized AIs are classified into the plurality of classifications. Emotions can be recognized by classifying them into one of them. Multiple classifications can be any number of classifications. For example, the plurality of classifications may be two classifications such as negative / positive or gentle / severe, three classifications such as joy / anger / sadness, and joy / anger / sadness. / There may be 4 classifications such as comfort, or more than 4 classifications, for example, 100 classifications.

The emotion classifier can classify emotions contained in a sentence into a plurality of classifications by using machine learning, for example. By inputting the sentences output from each of the plurality of personalized AIs into the emotion classifier, the emotions contained in the sentences output from each of the plurality of personalized AIs can be output. For example, when the text of the first answer output from the first personalized AI is input to the emotion classifier, the emotion contained in the text of the first answer is output and output from the first personalized AI. When the text of the second answer is input to the emotion classifier, the emotion contained in the text of the second answer can be output, and when the text of the Nth answer is input to the emotion classifier, the Nth N It is possible to output the emotions contained in the text of the answer. In this way, N emotions can be recognized from the output of the first personalized AI. For example, when the sentence output from the first personalized AI is input to the emotion classifier, the emotion contained in the sentence output from the first personalized AI is output and output from the second personalized AI. When the text is input to the emotion classifier, the emotion contained in the text output from the second personalized AI can be output, and the text output from the third personalized AI is classified into emotions. When input to the device, the emotions contained in the text output from the Mth personalized AI can be output. In this way, M emotions can be recognized from the output of M personalized AIs.

In step S704, the emotion recognition means 124 of the processor unit 120'generates an output based on the emotion recognized in step S703. The output may be, for example, an output indicating what the perceived emotion is, or an output indicating what the answer from the personalized AI is. The output may be, for example, an output indicating the ratio of each emotion among the plurality of recognized emotions.

For example, the emotion recognition means 124 may generate one output or N outputs when N emotions are recognized from one personalized AI. Then, n outputs may be generated (N> n). For example, the emotion recognition means 124 has a first output (for example, an output indicating that the recognized emotion is the first emotion, or an output indicating that the recognized emotion is the first emotion) when the first emotion is the most among the recognized N emotions. , Generates an output indicating that the answer from the personalized AI is the first answer) and produces a second output (eg, recognized) when the second emotion is the most recognized of the N emotions recognized. An output indicating that the emotion is the second emotion, or an output indicating that the answer from the personalized AI is the second answer) can be generated. For example, the emotion recognition means 124 has an output for the top n emotions out of the recognized N emotions (for example, an output indicating that the recognized emotions are the first to nth emotions, or personalization. An output indicating that the response from the AI is the first to nth answers) can be generated (N> n).

The emotion recognition means 124 can generate an output based on the recognized emotion for each of the M personalized AIs. That is, the emotion recognition means 124 may generate one output (total of M outputs) based on the emotion recognized from each of the M personalized AIs, or the emotion recognition means 124 may generate one output of the M personalized AIs. It may be possible to generate N outputs (total M × N outputs) based on the N emotions recognized from each, or the N emotions recognized from each of the M personalized AIs. It is also possible to generate n outputs (total M × n outputs) based on (N> n).

In step S705, the aggregation means 123 of the processor unit 120'aggregates the output based on the emotion generated in step S704. For example, the aggregation means 123 may aggregate the emotions indicated by the output of the emotion recognition means 124, or may aggregate the emotions indicated by the output of the emotion recognition means 124 in association with the answer to the question. .. The aggregation means 123 can, for example, aggregate the responses from the personalized AI indicated by the output by the emotion recognition means 124.

For example, the aggregation means 123 may calculate the number of first emotions (eg, positive emotions) and the number of second emotions (eg, negative emotions) among the plurality of emotions. The proportion of the first emotion (eg, positive emotion) and the second emotion (eg, negative emotion) of the plurality of emotions may be calculated. For example, the aggregation means 123 associates a first emotion (for example, a positive emotion) among a plurality of emotions with a first answer (for example, Yes's answer) and a second emotion (for example, a negative emotion). ) May be associated with a second answer (eg, No answer) to calculate their number or proportion. The counting means 123 can count emotions such as gentle / severe, joy / anger / sadness, and joy / anger / sadness / comfort in association with various answers to the questions.

The aggregated result is output to the outside of the system 100 via the interface unit 110. For example, the aggregated results are presented to the questioner.

An example of the processing of steps S701 to S705 will be described by taking as an example the case where the question "Do you support Mr. XX in the next presidential election?" Shown in FIG. 2A is input. For the sake of simplicity, an example of using three personalized AIs will be described.

In step S701, the receiving means 121 receives the question "Do you support Mr. XX in the next presidential election?"

In step S702, the answer generation means 122 outputs the answer to "Do you support Mr. XX in the next presidential election?" In the form of sentences using three personalized AIs. Each of the three personalized AIs can generate a continuation of the text upon entering a prompt. Here, in the case of the question "Do you support Mr. XX in the next presidential election?", The prompts related to the question are "Mr. XX in my opinion" and "I am Mr. XX." Whether you like it or hate it, "," I disagree with or agree with Mr. XX, "," Personally, Mr. XX, "etc., but are not limited to these.

For example, if you enter a question-related prompt (for example, "Mr. XX in my opinion") into a personalized AI that has learned the expression of the first individual, the prompt will continue according to the expression of the first individual. Can generate sentences (eg, "I think I'll win") and ask a personalized AI who has learned the expression of a second person (eg, "In my opinion, Mr. XX," If you enter "), you can generate a continuation of the prompt text (eg," works ^^ ") according to the expression of the second individual, and ask the personalized AI who learned the expression of the third individual. When you enter a prompt related to (for example, "Mr. XX in my opinion"), the following sentence of the prompt (for example, "I think it's not good") is given according to the expression of the third individual. Can be generated. For each personalized AI, you may enter multiple prompts to get multiple answers.

In step S703, the emotion recognition means 124 responds to the answers output from each of the three personalized AIs (for example, "I think I will win", "It works ^^", and "I think it is not good"). Recognize the emotions that have been put into it. The emotion recognition means 124 recognizes that the emotions contained in, for example, "I think I will win" and "it works ^^" are positive emotions, and is included in "I think it is not good". Emotions can be recognized as negative emotions.

In step S704, the emotion recognition means 124 generates an output based on the recognized emotion. In the emotion recognition means 124, for example, the emotion contained in the response from the first personalized AI is positive, the emotion contained in the response from the second personalized AI is positive, and the third personalized means. It is possible to generate an output indicating that the emotion contained in the response from the AI is negative.

In step S705, the aggregation means 123 aggregates the output. The aggregation means 123 can be aggregated as, for example, (positive: 66.7%, negative: 33.3%). Alternatively, it can be aggregated as (Yes: 66.7%, No: 33.3%) in association with the positive emotion = Yes answer and the negative emotion = No answer.

In this way, in response to the question "Do you support Mr. XX in the next presidential election?", A questionnaire survey by three personalized AIs was conducted (positive: 66.7%, negative: 33.3). %) Or (Yes: 66.7%, No: 33.3%).

FIG. 9 shows a conceptual diagram of an example of a flow 900 for creating a personalized AI by utilizing information of an existing individual. In the flow 900, a plurality of personalized AIs are created by using the information of a plurality of existing individuals. In the example shown in FIG. 9, the first personalized AI PAI ₁ is created from the information of the individual ₁ , and the second personalized AI PAI ₂ is created from the information of the individual ₂ . In this way, one personalized AI is created from the information of one individual.

First, multiple models used to create a personalized AI are prepared. The plurality of models can be, for example, a model capable of generating sentences. The sentence generation model can be, for example, a model capable of generating a subsequent sentence by inputting a prompt (writing a sentence).

Next, the information of each of multiple individuals is prepared. The information of each of multiple individuals is, for example, sentences created by each individual (for example, email, chat, posting to SNS), video of each individual's behavior, voice spoken by each individual, etc. including. Personal information is contained, for example, in a personal life log.

When the plurality of models and the information of the plurality of individuals are prepared, in step S91, each of the plurality of models learns the information of the individual. This creates a plurality of trained models that have learned personal information, that is, personalized AI. For example, the first model of the plurality of models learns the information of the individual ₁ , so that the first personalized AI PAI ₁ is created, and the second model of the plurality of models is the information of the individual ₂ . By learning, a second personalized AI PAI ₂ is created. The _first personalized AI PAI ₁ can perform activities that are unique to the individual 1, and can generate an answer as if the individual ₁ answered the questionnaire, for example. The _second personalized AI PAI ₁ can perform activities that are unique to the individual 2, for example, can generate an answer as if the individual ₂ answered the questionnaire.

The flow 900 can be carried out by, for example, the process 800 shown in FIG.

FIG. 8 is a flowchart showing another example (process 800) of the process by the system 100 for the questionnaire survey. The process 800 is a process for creating a personalized AI, and may be performed in the processor unit 120 or 120'of the system 100. In the following, it will be described as an example of what is done in the processor unit 120 of the system 100.

In step S801, the processor unit 120 prepares a model capable of generating sentences. For example, the processor unit 120 can acquire the sentence generation model stored in the database unit 200 from the database unit 200. The sentence generation model can be, for example, a model capable of generating a subsequent sentence by inputting a prompt (writing a sentence).

In step S802, the processor unit 120 makes a model capable of generating sentences learn an individual expression. This corresponds to step S91 in FIG. At this time, the individual expression is learned without correcting the bias of the individual expression. For example, a model capable of generating sentences can be fine-tuned with an individual expression. By fine-tuning an existing model, the amount of training data required can be reduced, and the time cost and memory cost can be reduced as compared with the case of creating a model from scratch.

Normally, when learning learning materials, it is necessary to create learning materials without bias in order to prevent bias in machine learning. On the other hand, in step S802, as described above, unlike ordinary machine learning, the individual's expression is learned without correcting the bias of the individual's expression. As a result, the created personalized AI will output a biased result. This "bias" can be regarded as "individuality", and personalized AI will be able to output sentences containing individuality.

FIG. 10 shows a conceptual diagram of an example of a flow 1000 for creating a personalized AI by utilizing information of a virtual individual that does not exist. In the flow 1000, a personalized AI of one individual is created by using the information of a plurality of non-existent virtual individuals. In the example shown in FIG. 10, the personalized AI of the individual ₁ is created based on the result of the individual ₁ answering the question and the information of the virtual individual.

First, in step S101, a question is presented to individual ₁ . The question can be any question.

Next, in step S102, the individual ₁ answers the question. Individual ₁ can answer the question in any manner. For example, the individual ₁ may answer the question via a paper questionnaire, may answer the question via an information terminal, or may answer the question verbally. good.

On the other hand, multiple models used to create personalized AI are prepared. The plurality of models can be, for example, a model capable of generating sentences. The sentence generation model can be, for example, a model capable of generating a subsequent sentence by inputting a prompt (writing a sentence).

Next, the information of each of the multiple virtual individuals is prepared. The information of each of the plurality of virtual individuals may include, for example, a virtually created sentence, a virtually created video, a virtually created audio, and the like. Text, video, audio, etc. can be randomly created, for example. Information about virtual individuals does not have to be human-understandable. The information of the virtual individual may be, for example, a list of characters or symbols having no meaning. For example, the information of the virtual individual may be the information output from the personalized AI that has already been created. The information of the virtual individual may be included in, for example, a virtual life log (fake log) created virtually. The fake log can be, for example, a life log output from a personalized AI that has already been created.

Virtual personal information can be created by any method. As described above, the information of the virtual individual does not have to be the information that can be understood by humans, and therefore, the information created by any method functions as the information of the virtual individual.

For example, virtual personal information can be created using a random function. In one example, virtual personal information is created by randomly selecting a letter or symbol from a table of letters (eg, 50-sound letters, alphabets, etc.) and / or symbols and enumerating the selected letters or symbols. Can be done.

For example, virtual personal information can be created from the life log of an existing individual. In one example, it can be created by shuffling the lifelogs of multiple real individuals. For example, a part of one of the lifelogs of a plurality of real individuals (for example, at least one character, at least one word, at least one phrase, at least one clause, at least one sentence, at least one paragraph, etc.) is the other one. Alternatively, virtual personal information can be created by replacing it with a part of multiple life logs. For example, virtual individual information can be created by randomly selecting characters, words, phrases, clauses, sentences, paragraphs, etc. from a pool that stores the life logs of a plurality of existing individuals. Not limited to the life logs of a plurality of real individuals, any other information (eg, text, video, audio, etc.) can be used to create information for a virtual individual.

For example, virtual personal information can be created using a personalized AI that has already been created. The text output when a prompt (writing text) is given to the already created personalized AI can be used as virtual personal information.

When the plurality of models and the information of the plurality of virtual individuals are prepared, in step S103, each of the plurality of models learns the information of the virtual individual. As a result, a plurality of trained models in which the information of the virtual individual is learned, that is, a personalized AI is created. For example, the first model among the plurality of models learns the information of the virtual individual ₁ , so that the first personalized AI PAI ₁ is created, and the second model among the plurality of models is the virtual individual ₂ . By learning the information of, a second personalized AI PAI ₂ is created. The first personalized AI PAI ₁ can perform activities like the virtual individual ₁ , and can generate an answer as if the virtual individual ₁ answered the questionnaire, for example. The second personalized AI PAI ₁ can perform activities like the virtual individual ₂ , and can generate an answer as if the virtual individual ₂ answered the questionnaire, for example.

When a plurality of personalized AIs are created, in step S104, each of the plurality of personalized AIs is presented with the same question as the question presented to individual ₁ in step S101.

In step S105, each of the plurality of personalized AIs outputs an answer to the question. Each of the plurality of personalized AIs outputs an answer as if the virtual individual answered the question.

Then, among the answers from each of the plurality of personalized AIs, the answer closest to the answer by the individual ₁ is specified, and the personalized AI that outputs the answer is determined to be the personalized AI of the individual ₁ . This is based on the speculation that a personalized AI that answers the same question in the same way would be able to generate an output based on the same idea as that of Individual ₁ . The more multiple or complex the questions asked to both the individual ₁ and the personalized AI, the more the personalized AI determined to be the individual ₁ 's personalized AI can be a more reflective of the individual ₁ 's ideas.

By creating a personalized AI using virtual personal information that does not exist, it is possible to create a personalized AI even when the information of a real individual is unavailable. For example, even a company that does not have an individual's life log can create a personalized AI if there is a result of a questionnaire conducted to an individual.

The personalized AI created in this way can be used, for example, for pseudo market research. For example, by having a plurality of personalized AIs answer the questionnaire without actually taking the questionnaire to the customer, it is possible to investigate the market trend easily and at low cost.

The flow 1000 can be implemented, for example, by the process 1100 shown in FIG.

FIG. 11 is a flowchart showing another example (process 1100) of the process by the system 100 for the questionnaire survey. Process 1100 is another process for creating a personalized AI, which may be performed in the processor unit 120 or 120'of the system 100. In the following, it will be described as an example of what is done in the processor unit 120 of the system 100.

In step S1101, the processor unit 120 acquires the result of the individual answering the question. This may correspond to step S102 in FIG. An individual is a real person. The question can be any question. The question may be, for example, a multiple-choice answer-type question, a descriptive-answer-type question, or an essay-answer-type question. The questions can be any number. The larger the number of questions, the better. This is because the larger the number of questions, the higher the accuracy of the personalized AI generated by the process 1100.

It does not matter where the processor unit 120 obtains the result of the individual answering the question. For example, when the result of an individual answering a question is stored in the database unit 200, the processor unit 120 can acquire the result of the individual answering the question from the database unit. Alternatively, for example, the processor unit 120 can acquire the result of the individual answering the question from the individual terminal device 300. Alternatively, for example, the processor unit 120 can obtain the result of an individual answering a question from another location connected to the network 400.

In step S1102, the processor unit 120 prepares a plurality of models capable of generating sentences. For example, the processor unit 120 can acquire the sentence generation model stored in the database unit 200 from the database unit 200. The sentence generation model can be, for example, a model capable of generating a subsequent sentence by inputting a prompt (writing a sentence).

In step S1103, the processor unit 120 generates a plurality of trained models by learning the expressions of a plurality of individuals from a plurality of models capable of generating sentences. This corresponds to step S103 in FIG. At this time, one model is made to learn the expression of one individual. That is, one trained model is generated by having one of the plurality of models learn one of the expressions of the plurality of individuals. For example, when n models are prepared in step S1102, the first trained model is generated by letting the first model learn the expression of the first individual, and the second model becomes the second model. A second trained model is generated by learning the expression of the individual, and ... By training the nth model of the expression of the nth individual, the nth trained model is generated. Will be done. As a result, n trained models are generated.

The expression of an individual may be an expression by a real person or an expression by a virtual person who does not exist. Preferably, the representation of an individual comprises the representation of at least one virtual individual. As a result, the created trained model becomes a model that can create sentences with virtual personal expressions. Here, the expression of a virtual individual does not necessarily have to be information that can be understood by humans, and the information of a virtual individual may be, for example, a list of meaningless characters or symbols.

The expressions of multiple individuals can be extracted from, for example, multiple life logs. That is, the expressions of a plurality of individuals can be prepared, for example, by acquiring a plurality of life logs and extracting the expressions of a plurality of individuals from the acquired plurality of life logs. At this time, preferably, the plurality of life logs may include at least one virtual life log. The virtual life log may be, for example, a record of information output from a created personalized AI.

When learning individual expressions, it is preferable to learn individual expressions without correcting the bias of individual expressions. Normally, when learning learning materials, it is necessary to create learning materials without bias in order to prevent bias in machine learning. On the other hand, in step S1003, as described above, unlike ordinary machine learning, the individual's expression is learned without correcting the bias of the individual's expression. As a result, the created trained model will output a biased result. This "bias" can be regarded as "individuality", and the trained model can output sentences including individuality. For example, when a trained model is given a prompt (sentence export), the trained model can generate subsequent sentences according to the expression of the trained individual.

When learning an individual's expression, for example, a model capable of generating a sentence may be fine-tuned with the individual's expression. By fine-tuning an existing model, the amount of training data required can be reduced, and the time cost and memory cost can be reduced as compared with the case of creating a model from scratch.

In step S1104, the processor unit 120 outputs an answer to the question from each of the plurality of trained models generated in step S1103. This corresponds to steps S104 and S105 in FIG. The question is the same question as the question obtained from the result answered by the individual in step S1101. For example, multiple trained models can generate subsequent sentences according to the expression of the individual who learned by giving a prompt (writing a sentence), so that the prompt related to the question can be generated by multiple trained models. By inputting to each, the answer to the question can be obtained from each of the plurality of trained models. The prompt related to the question may be selected from preset templates according to the type of question, may be created based on the question, or may be set for each question. It may be set by the creator of the personalized AI. If the prompt related to the question is created based on the question, for example, the prompt may be created based on a preset template, or machine learning may be used to generate the prompt. You may.

The trained model can generate a continuation of these prompts according to the personal expression of the trained individual. The generated text will be the answer to the question. Alternatively, the emotion recognition means 124 described above may recognize the emotion contained in the generated sentence, and the recognized emotion may be used as an answer to the question. Alternatively, the above-mentioned emotion recognition means 124 recognizes the emotions contained in the generated sentence, generates an output based on the recognized emotions, and the above-mentioned aggregation means 123 aggregates the generated outputs. It may be the answer to the question.

For example, when n trained models are generated, the first trained model outputs the first answer to the question, and the second trained model outputs the second answer to the question. ... Output the nth answer to the question from the nth trained model. As a result, n answers are generated.

In step S1105, the processor unit 120 determines one trained model based on the plurality of answers output from the plurality of trained models in step S1104 and the result obtained in step S1101. The processor unit 120 can determine, for example, one trained model that outputs the same answer as the result acquired in step S1101. Alternatively, the processor unit 120 determines one trained model based on the degree of similarity between the plurality of answers output from the plurality of trained models in step S1104 and the result obtained in step S1101, for example. Can be done. The processor unit 120 can determine, for example, one trained model that outputs the answer closest to the result acquired in step S1101. Here, any known index can be used as the degree of similarity.

For example, when the result obtained in step S1101 is an answer to a multiple-choice answer type question, a plurality of answers output from a plurality of trained models in step S1104 and a result obtained in step S1101. Can be quantitatively compared with. For example, when the result acquired in step S1101 is an answer to a question of a descriptive answer formula or an essay answer formula, a plurality of learnings are performed in step S1104 by converting the result acquired in step S1101 into quantitative data. It becomes possible to quantitatively compare the plurality of answers output from the completed model with the results obtained in step S1101. For example, the quantification of the answer to a descriptive answer type or essay answer type question is performed by analyzing the content of the answer. This may be done, for example, by humans or by AI. Alternatively, this may be done, for example, by analyzing the perception of the contained emotions by the emotion recognition means 124.

One determined trained model can be regarded as an individual's personalized AI because it can output the same or similar answer as the result of the individual's answer to the question. This is based on the speculation that a personalized AI that answers the same question in the same way as an individual would be able to generate an output based on the same idea as the individual's idea. When an answer to another question is output from one determined trained model, the answer is expected to be an answer as if it were answered by an individual.

According to the process 1100, a personalized AI can be created even if the expression of the individual used for learning includes the expression of at least one virtual individual. For example, a personalized AI can be created if an individual responds, even if he / she does not have enough personal information, or even if he / she does not have any personal information at all. It is. This opens up the possibility of creating personalized AI and, by extension, the possibility of digital clone surveys.

FIG. 12A shows an example of a specific step of step S1105 in the process 1100. In this example, at least one of the plurality of trained models created in the process of processing 1100 is made to output a virtual individual expression, and the trained model is used using the output virtual individual expression. Includes the process of creating.

For example, if the process shown in FIG. 12A does not have a trained model that can be considered as an individual personalized AI, for example, a plurality of answers output from the plurality of trained models in step S1104 and a step. It can be performed when none of the similarity with the result obtained in S1101 exceeds a predetermined threshold value.

In step S1201, the processor unit 120 outputs at least one virtual individual expression from at least one of the plurality of trained models created in step S1103. At least one of the plurality of trained models is determined, for example, based on the similarity between the plurality of answers output from the plurality of trained models in step S1104 and the result obtained in step S1101. be able to. For example, at least one of the plurality of trained models can be at least one trained model with the highest similarity. For example, at least one of the plurality of trained models may be at least one trained model having a similarity greater than a predetermined value. This is because by using a trained model that can output an answer that is closer to the answer of an existing individual, the output virtual expression of the individual can be close to that of an actual individual. This can lead to improved accuracy of the personalized AI of the actual individual that is finally created. Since at least one of the plurality of trained models created in step S1103 can output the following sentences when prompted, for example, a virtual personal expression can be used. Can be output.

In step S1202, at least one second trained model is generated by having a plurality of models capable of generating sentences learn at least one virtual individual expression output in step S1201. .. Step S1202 is the same process as step S1103. In step S1202, one model is made to learn the expression of one virtual individual. That is, one second trained model is generated by having one of the plurality of models learn at least one of the representations of a virtual individual. For example, when the expressions of n virtual individuals are output in step S1201, the first model is made to learn the expressions of the first virtual individual, so that the first second trained model is trained. Is generated, and by letting the second model learn the expression of the second virtual individual, a second second trained model is generated, ... The nth model is the nth virtual. By learning the expression of a typical individual, the nth second trained model is generated. As a result, n second trained models are generated.

When learning the expression of a virtual individual, it is preferable to learn the expression of the virtual individual without correcting the bias of the expression of the virtual individual. Normally, when learning learning materials, it is necessary to create learning materials without bias in order to prevent bias in machine learning. On the other hand, in step S1202, as described above, unlike ordinary machine learning, the virtual individual expression is learned without correcting the bias of the virtual individual expression. As a result, the created second trained model will output a biased result. This "bias" can be regarded as "individuality", and the second trained model can output a sentence containing a virtual individual individuality. For example, when a prompt (sentence writing) is given to the second trained model, the second trained model can generate a subsequent sentence according to the expression of the virtual individual who has learned.

When learning a virtual individual expression, for example, a model capable of generating a sentence may be fine-tuned with the individual expression. By fine-tuning an existing model, the amount of training data required can be reduced, and the time cost and memory cost can be reduced as compared with the case of creating a model from scratch.

In step S1203, the processor unit 120 outputs an answer to the question from at least one second trained model generated in step S1202. Step S1203 is the same process as step S1104. The question is the same question as the question obtained from the result answered by the individual in step S1101. For example, the second trained model can generate a continuation of a sentence according to the virtual personal expression learned by giving a prompt (writing a sentence), so that the prompt related to the question is given as a second. By inputting to the trained model of, the answer to the question can be obtained from the second trained model. The prompt related to the question may be selected from preset templates according to the type of question, may be created based on the question, or may be set for each question. It may be set by the creator of the personalized AI. If the prompt related to the question is created based on the question, for example, the prompt may be created based on a preset template, or machine learning may be used to generate the prompt. You may.

The second trained model can generate the continuation of these prompts according to the virtual personal representation learned. The generated text will be the answer to the question. Alternatively, the emotion recognition means 124 described above may recognize the emotion contained in the generated sentence, and the recognized emotion may be used as an answer to the question. Alternatively, the above-mentioned emotion recognition means 124 recognizes the emotions contained in the generated sentence, generates an output based on the recognized emotions, and the above-mentioned aggregation means 123 aggregates the generated outputs. It may be the answer to the question.

For example, when n second trained models are generated, the first answer to the question is output from the first second trained model, and the question is asked from the second second trained model. The second answer to the question is output, and the nth answer to the question is output from the nth second trained model. As a result, n answers are generated.

In step S1204, the processor unit 120 includes a plurality of answers output from the plurality of trained models in step S1104, at least one answer output from at least one second trained model in step S1203, and step S1101. Based on the results obtained in, one trained model is determined. The processor unit 120 may determine, for example, one trained model that outputs the same answer as the result obtained in step S1101 from the plurality of trained models and at least one second trained model. can. Alternatively, the processor unit 120 may, for example, have a plurality of answers output from the plurality of trained models in step S1104, at least one answer output from at least one second trained model in step S1203, and step S1101. One trained model can be determined from a plurality of trained models and at least one second trained model based on the degree of similarity to the obtained results. The processor unit 120 can determine, for example, one trained model that outputs the answer closest to the result acquired in step S1101. Here, any known index can be used as the degree of similarity.

For example, the processor unit 120 extracts at least two trained models from a plurality of trained models and at least one second trained model, and based on the extracted two trained models, the processor unit 120 extracts, for example. One trained model may be generated. For example, one trained model can be generated by averaging the two extracted trained models.

One determined trained model can be regarded as an individual's personalized AI because it can output the same or similar answer as the result of the individual's answer to the question.

FIG. 12B shows another example of a specific step of step S1105 in the process 1100. In this example, at least one of the plurality of trained models created in the process of processing 1100 is made to output a virtual individual expression, and the trained model is used using the output virtual individual expression. It is characterized by repeating the process until a trained model capable of outputting an answer similar to the result answered by an existing individual is generated.

Steps S1201 to S1203 are the same as the steps described above with reference to FIG. 12A, and description thereof will be omitted here.

In step S1205, the processor unit 120 determines whether or not a predetermined condition is satisfied.

The predetermined conditions are, for example, one of a plurality of answers output from the plurality of trained models in step S1104 and one of the answers output from at least one second trained model in step S1203, and step S1101. The similarity with the result obtained in can exceed a predetermined threshold. As mentioned above, any known index can be used as the degree of similarity. Further, an arbitrary threshold value can be set as a predetermined threshold value. The higher the predetermined threshold, the higher the accuracy of the personalized AI generated by process 1100. Alternatively, the predetermined condition may be, for example, that the number of times of repeating steps S1201 to S1203 exceeds a predetermined threshold value. By limiting the number of repetitions, it is possible to prevent the processing from diverging. Alternatively, the predetermined condition may be, for example, the elapsed time of the process exceeds a predetermined threshold. By limiting the processing time, it is possible to prevent the processing from diverging.

If it is determined in step S1205 that the predetermined condition is satisfied, the process proceeds to step S1204'. If it is determined in step S1205 that the predetermined condition is not satisfied, steps S1201 to S1203 are repeated.

When steps S1201 to S1203 are repeated, in step S1201, in addition to outputting the representation of at least one virtual individual from at least one of the plurality of learning models, at least one generated in step S1202. At least one virtual individual representation can be output from the two second trained models. As a result, in step S1202, in addition to the representation of the virtual individual output from at least one of the plurality of learning models, the representation of the virtual individual output from at least one second trained model. Will also be used for learning.

In step S1204', the processor unit 120 identifies a trained model that satisfies a predetermined condition from the plurality of trained models and at least one second trained model.

The predetermined condition may be, for example, that the similarity exceeds a predetermined threshold. Alternatively, the predetermined condition may be that the similarity is highest when the number of repetitions or the elapsed time of processing exceeds the threshold.

The identified trained model can be considered as an individual's personalized AI. It is expected that the personalized AI created in this way can generate an output based on an idea similar to an individual idea with higher accuracy.

In the above-mentioned example with reference to FIGS. 5 to 12, it has been described that the processes are performed in a specific order, but the order of each process is not limited to the described one, and any order that is logically possible is not limited. Can be done at.

In the above-described example with reference to FIGS. 5 to 8 and FIGS. 11 to 12B, the processing of each step shown in FIGS. 5 to 8 and 11 to 12B is performed by the processor unit 120 or 120'and the memory unit 130. Although it has been described that it is realized by a program stored in, the present invention is not limited to this. At least one of the processes of each step shown in FIGS. 5 to 8 and 11 to 12B may be realized by a hardware configuration such as a control circuit.

Using a specific example, explain how to create a personalized AI that can be regarded as an individual personalized AI. In this example, multiple personalized AIs are created using randomly created fake logs, and each of the created multiple personalized AIs is asked to answer the question, so that the same answer as the individual answer is output. Identify possible personalized AIs.

Create 100 personalized AIs using 100 sets of fake logs. Let one sentence generation model learn a set of fake logs. 100 sets of fake logs are created as random characters and symbols using a random function.

The questions given to individuals are questions that have a total of two questions that can be answered with two choices of ○ or ×. The answer will be one of four types: XX, XX, XX, and XX. Have an individual answer this question.

On the other hand, have each of the 100 personalized AIs answer the same question (a total of 2 questions to be answered with 2 choices of ○ or ×).

When the answer by an individual is XX, the personalized AI that answered XX out of 100 personalized AIs is identified as the personalized AI that can output the same answer as the individual's answer. ..

In this example, it is expected that a personalized AI that can output the same answer as an individual's answer will be identified with a one-quarter probability.

In this way, by creating a sufficiently large number of personalized AIs for the "number of cases" of the answer to the question, the personalized AI that can output the same answer as the individual answer may be identified. Is expected to be high enough. It may be preferable to limit the "number of cases" of the answers so that the number of personalized AIs to be created does not diverge or the amount of computation is not enormous.

In the above-mentioned example, the case where the system 100 is a computer (for example, a server device) installed in a service provider that provides a service for a digital clone questionnaire survey has been described as an example, but the present invention is limited thereto. Not done. The system 100 can be any information processing device including a processor unit. The system 100 may be, for example, a user's terminal device.

(Creating a personalized AI)
In this embodiment, a personalized AI capable of generating sentences according to individual expressions in Japanese was created based on an existing sentence generation model.

As an existing sentence generation model, the general-purpose sentence generation model GPT-2model developed by OpenAI was used. Since GPT-2model was not compatible with Japanese as of 2020, we first created a Japanese version of GPT-2model that can generate Japanese sentences.

In order to create training data for creating the Japanese version of GPT-2model, use the data of Japanese Wikipedia XML dump file (https://dumps.wikimedia.org/jamwiki/20200501) and use the corpus data. Generated. From the corpus data, training data including 90% of Wikipedia articles, validation data including 5% of Wikipedia articles, and test data including 5% of Wikipedia articles were generated.

GPT-2model was trained with training data and the trained model was validated and tested with validation and test data. Training was performed on a 1 GPU GeForce RTX ^TM 2080 using the PyTorch implementation available from the Huggingface's transformers library. The size of the model corresponds to GPT-2 small with N = 12 layers, except that we used 768 embedding size, position embedding size, and context size, which are smaller than 1024 of GPT-2 small. Batch size 2 of gradient accumulation step 16 corresponding to 32 batch sizes was used. The initial learning rate of 1e-4 was used. Trained the model with 5 epochs.

The trained GPT-2 model was evaluated with test data using the perplexity scale. The perplexity was 19.31.

Next, the trained Japanese version of GPT-2model was fine-tuned with the expression of an individual. First, we collected the text of the individual's expression. We collected text of the individual's representation from the individual's chat history and postings on social media. From the collected text, sentences that were too short, sentences that contained too few words, URLs, punctuation marks, etc. were removed. Table 1 shows the statistics of the collected texts.

The Japanese version of GPT-2model was fine-tuned using the collected texts. Fine tuning was performed at a learning rate of 2e-5 so as not to overfit the data used for fine tuning.

Table 2 shows sentences generated by the Japanese version GPT-2 model without fine tuning and sentences generated by the Japanese version GPT-2 model without fine tuning. The results of each of the three prompts "This report is", "This project is", and "Natural language processing is" are shown.

In Table 2, the Japanese version of GPT-2model without fine-tuning generates sentences with more rigid expressions, while the fine-tuned Japanese version of GPT-2model generates sentences according to individual expressions. You can see that it is doing. It should be noted that the fine-tuned Japanese version of GPT-2model even learns to use pictograms such as ^^ as personal expressions.

With the fine-tuned Japanese version of GPT-2model, it is possible to realize a personalized AI that can generate sentences with individual expressions.

The present invention is not limited to the above-described embodiment. It is understood that the invention should be construed only by the claims. It will be understood by those skilled in the art that from the description of the specific preferred embodiments of the present invention, the equivalent scope can be carried out based on the description of the present invention and common general technical knowledge.

The present invention is useful in that it can provide a system for questionnaire surveys and the like using a plurality of personalized AIs.

Q Questioner PAI ₁ ~ PAI _M Personalized AI
100 System 110 Interface part 120 Processor part 130 Memory part 200 Database part 300 Terminal equipment 400 Network

Claims

It ’s a system for questionnaire surveys.
Receiving means to receive questions for questionnaire surveys,
A plurality of personalized AIs, each of which is configured to output an answer to the received question, and a plurality of personalized AIs.
A system including an aggregation means for aggregating the outputs from each of the plurality of personalized AIs.
Each of the plurality of personalized AIs outputs a plurality of answers by repeating the output of the answer to the received question a plurality of times.
The system according to claim 1, wherein the aggregation means aggregates the plurality of responses output from each of the plurality of personalized AIs.
Each of the plurality of personalized AIs has learned the expressions of individuals different from each other, and each of the plurality of personalized AIs can generate sentences according to the expressions of the learned individuals.
Each of the plurality of personalized AIs outputs a sentence according to the expression of the learned individual as an answer to the question.
The system according to claim 2.
Further equipped with emotion recognition means that can recognize the emotions contained in the text,
The emotion recognition means is
Recognizing the multiple emotions contained in the plurality of answers output from each of the plurality of personalized AIs.
It is configured to generate and generate output based on the multiple perceived emotions described above.
The aggregation means aggregates the output based on the plurality of emotions.
The system according to claim 3,
The emotion recognition means is
When the first emotion is the most among the plurality of emotions, the first output is generated.
The second output is generated when the second emotion is the most among the plurality of emotions.
The system according to claim 4.
The system according to any one of claims 3 to 5, wherein each of the plurality of personalized AIs is learning the expression of the individual without correcting the bias of the expression of the individual.
The system according to claim 6, wherein the expression of the individual is an expression by an existing individual.
The system according to claim 6, wherein the expression of the individual is an expression by a virtual individual.
It ’s a method for questionnaire surveys.
Receiving questions for surveys and
Using multiple personalized AIs to output multiple answers to the question and
A method comprising aggregating the outputs from each of the plurality of personalized AIs.
A program for a questionnaire survey, the program is executed in a computer system including a processor unit, and the program is
Receiving questions for surveys and
Using multiple personalized AIs to output multiple answers to the question and
A program that causes the processor unit to aggregate the outputs from each of the plurality of personalized AIs.
How to create a personalized AI that learns personal expressions,
Preparing a model that can generate sentences,
A method comprising having the model learn the individual's expression without correcting the bias of the individual's expression.
It is a method to create a personalized AI that can output the answer to the question.
To get the result of an individual answering a question,
Preparing multiple models that can generate sentences,
By training the plurality of models to learn the expressions of a plurality of individuals, a plurality of trained models are generated, and one of the plurality of models is one of the expressions of the plurality of individuals. By training, one trained model is generated, and the representations of the plurality of individuals include the representations of at least one virtual individual.
To output the answer to the question from each of the plurality of trained models,
A method comprising determining one trained model based on a plurality of answers output from the plurality of trained models and the results.
The expressions of the plurality of individuals are
To get multiple life logs and
12. The method of claim 12, wherein the plurality of lifelogs are prepared by extracting representations of a plurality of individuals from the plurality of lifelogs, the plurality of lifelogs comprising at least one virtual lifelog.
The method of claim 13, wherein the at least one virtual lifelog is generated from at least one personalized AI.
13. Method.
Determining the one trained model mentioned above
(1) Outputting at least one virtual individual expression from at least one of the plurality of trained models.
(2) To generate at least one second trained model by having the plurality of models learn the expression of at least one virtual individual that is output, and among the plurality of models. By training one of the expressions of one virtual individual, one second trained model is generated.
(3) Outputting the answer to the question from the at least one second trained model, and
(4) One trained model is determined based on the plurality of answers output from the plurality of trained models, the answers output from the at least one second trained model, and the above results. The method according to any one of claims 12 to 15, including the above.
Determining the one trained model mentioned above
By repeating the above (1) to the above (3) until a predetermined condition is satisfied, the above (1) expresses at least one virtual individual from the at least one second trained model. Further including output, the above (4) is
16. The method of claim 16, comprising identifying one trained model that satisfies a predetermined condition from the plurality of trained models and at least one second trained model.
The predetermined condition is one of a plurality of answers output from the plurality of trained models and one of the answers output from the at least one second trained model, and the result of the individual's answer. 17. The method of claim 17, wherein the similarity of is above a predetermined threshold.
The system according to any one of claims 1 to 8, wherein the plurality of personalized AIs include personalized AIs created according to the method according to claims 11 to 17.
A system for creating personalized AI that can output answers to questions.
How to get the result of an individual answering a question,
A means of preparing multiple models capable of generating sentences,
It is a means for generating a plurality of trained models by having the plurality of models learn the expressions of a plurality of individuals, and one of the plurality of models is one of the expressions of the plurality of individuals. By training, one trained model is generated, and the representation of the plurality of individuals includes means and means including the representation of at least one virtual individual.
A means for outputting an answer to the question from each of the plurality of trained models,
A system including a plurality of answers output from the plurality of trained models and a means for determining one trained model based on the results.
A program for creating a personalized AI capable of outputting an answer to a question, the program is executed in a computer system including a processor unit, and the program is
To get the result of an individual answering a question,
Preparing multiple models that can generate sentences,
By training the plurality of models to learn the expressions of a plurality of individuals, a plurality of trained models are generated, and one of the plurality of models is one of the expressions of the plurality of individuals. By training, one trained model is generated, and the representations of the plurality of individuals include the representations of at least one virtual individual.
To output the answer to the question from each of the plurality of trained models,
A program for causing the processor unit to perform a process including determining one trained model based on a plurality of answers output from the plurality of trained models and the results.