JP2023114460A - Expert answer prediction system - Google Patents

Abstract

To perform the pseudo-output of a level requested by a user from an expert to the user who wants to obtain the answer to a question from the expert in a specific field.SOLUTION: A machine learning device 10 of an expert answer prediction system 1 includes: a learning model storage part 11 which stores, as a digital clone Ci of each individual expert Ei, a learning model constructed so as to be able to predict the answer given by each individual expert Ei to the question by learning the life log of each of individual experts E1 to EN; a question input part 12 for inputting the question data to any personal expert Ex into a digital clone Cx corresponding to the personal expert Ex; an answer output part 13 which outputs the answer predicted by the digital clone Cx as the answer of the individual expert Ex according to the inputted question data; and a learning model evaluation part 15 which evaluates the digital clones C1 to CN as a plurality of learning models constructed by the machine learning device 10.SELECTED DRAWING: Figure 1

Description

The present invention relates to an expert answer prediction system that predicts answers from experts who have expertise in a specific field.

2. Description of the Related Art Conventionally, as a system for outputting answers to questions given by users, for example, a system seen in Patent Literature 1 is known. In this system, answers to various questions are prepared and stored in advance, and answers matching the content of the questions given by the user are extracted and output.

Japanese Patent No. 6770283

By the way, in many fields, users with questions about a field often want answers (including advice, recommendations, etc.) from a reputable individual expert in that field.

However, it is generally difficult for a general user to contact a prominent individual expert and obtain an answer from the individual expert. In addition, in the system as seen in Patent Document 1, in many cases, only general-purpose answers can be obtained, and the usefulness and quality of the answers tend to be poor for the user.

Furthermore, even if a virtual expert simulating an expert were constructed using artificial intelligence, it would be difficult to verify the certainty of the expert, and as a result, the usefulness of the level of answers required by the user would be difficult. There is a risk that the product will be of poor quality.

In view of the above circumstances, the present invention provides a specialized technology capable of pseudo-outputting an answer of a level desired by the user from an expert for a user who wishes to obtain an answer to a question from an expert in a specific field. The purpose is to provide a home answering system.

The expert answer prediction system of the first invention is an expert answer prediction system that predicts answers from experts who have expertise in a specific field,
A machine learning device that builds and stores learning models of a plurality of experts by machine learning processing,
The machine learning device
By learning the lifelogs of a plurality of said experts, a plurality of learning models constructed so as to be able to predict the answers given by each expert to questions are stored as digital clones of said experts. a learning model storage unit for
a question input unit that acquires questions for the plurality of experts and inputs question data indicating the content of the questions to the plurality of learning models;
an answer output unit that outputs answers predicted by the plurality of learning models as answers of the plurality of experts in response to the question data input from the question input unit;
a learning model evaluation unit that evaluates the plurality of learning models,
The learning model evaluation unit
An expert side answer group, which is a set of answers of each of the plurality of experts to questions asked by each of the plurality of experts by changing the group of the plurality of experts, and the question configured to evaluate the degree of approximation with a model-side answer group, which is a set of answers generated by each of the plurality of learning models, identifying a plurality of groups of the experts with different degrees of approximation, It is characterized in that it is configured so that the answers to the questions to be asked in the training can be obtained from the plurality of learning models corresponding to the plurality of identified groups of experts.

According to the expert answer prediction system of the first invention, the digital clone learns the transmissions of individual experts in a specific field (remarks issued by experts, SNS, blogs, publications, books, papers, etc.) Since it is a learning model constructed by Therefore, when the question data is input to the learning model, the learning model can predict a certain highly reliable answer as the answer given by the individual expert. can be output from the question output unit.

Here, for the multiple individual experts constructed by the learning model, the learning model evaluation unit changes the group of multiple experts (creates a group of experts with different sizes), By evaluating the degree of approximation between the actual answer group of the group (set) and the answer group of the learning model, it is possible to verify the certainty of the expert's answer as the group (set).

Then, a plurality of groups of experts with different degrees of approximation (degrees of approximation) are specified, and answers to newly-implemented questions can be obtained from a plurality of learning models corresponding to the specified groups of experts. , it is possible to obtain answers on the level of usefulness and quality desired by users from a group of experts whose certainty has been verified.

As a result, for example, if you create a group of gastroenterologists for specialists such as multiple internists, the degree of approximation will be high, but if you create a group of infectious disease internal medicine, the degree of approximation will be low. Answers can be obtained from a trained model of a group of gastroenterology specialists with a high degree of variability, while a trained model of an infectious disease specialist group with a small number of specialists and a high degree of variability and low approximation (with variability) ), it is possible to receive an answer.

As described above, according to the expert answer prediction system of the first invention, for a user who wishes to obtain an answer to a question from an expert in a specific field, a pseudo answer from the expert of the level desired by the user is provided. can be output.

The expert answer prediction system of the second invention is, in the first invention,
The learning model evaluation unit identifies a plurality of small groups included in the plurality of groups of experts, re-evaluates the degree of approximation for the identified small groups, and determines the degree of approximation equal to or greater than a predetermined threshold. A group of a plurality of experts is reconstructed as a large group by combining the small groups having

According to the expert answer prediction system of the second invention, in particular, by reconstructing a group of a plurality of experts as a large group by combining small groups having the degree of approximation equal to or higher than a predetermined threshold, the likelihood can be calculated. It is possible to create a group of experts as a large group of collateral.

For example, in the case of reconstructing a population with a higher degree of approximation, in the previous example, the degree of approximation is evaluated with a small group subdivided in the department of gastroenterology, and only the small group with a high degree of approximation is evaluated. By reconstructing , it is possible to build a learning model of a group of specialists in gastroenterology with higher certainty.

As described above, according to the expert answer prediction system of the second invention, for a user who wishes to obtain an answer to a question from an expert in a specific field, a pseudo high-level answer desired by the expert is provided. can be output to

The expert answer prediction system of the third invention is, in the first invention,
The learning model evaluation unit identifies a plurality of small groups included in the plurality of groups of experts, re-evaluates the degree of approximation for the identified small groups, and determines the degree of approximation that is less than a predetermined threshold. A group of a plurality of experts is reconstructed as a large group by combining the small groups having

According to the expert answer prediction system of the third invention, in particular, by reconstructing a group of a plurality of experts as a large group by combining small groups having the degree of approximation less than a predetermined threshold, You can create a group of experts as the missing mass.

That is, by intentionally creating a learning model of an expert with a low degree of approximation, it is possible to identify fields/regions in which there is no certainty of answers (fields/regions where expertise is not established). Furthermore, it is possible to dare to answer the learning model of a group of experts in such fields and areas.

For example, it is possible to evaluate the degree of approximation for each small group subdivided from the department of infectious disease, identify the group of infectious disease department with extremely low degree of similarity as an unexplored area, and examine the answers. can.

As described above, according to the expert answer prediction system of the third invention, for a user who wishes to obtain an answer to a question from an expert in a specific field, the expert can provide an unknown multidimensional answer desired by the user. Can be output in a pseudo manner.

The expert answer prediction system of the fourth invention is, in any one of the first to third inventions,
In the machine learning device, the question input unit selects one of the plurality of learning models according to the content of the acquired question item and the degree of approximation requested by the questioner regarding the question item. The learning model is selected and the question data is input to the selected learning model.

According to the expert answer prediction system of the fourth invention, in the present invention, the machine learning device stores a plurality of learning models respectively corresponding to a plurality of experts in the learning model storage unit, and The question input unit selects one learning model from among the plurality of learning models according to the content of the obtained question item and the degree of approximation requested by the questioner regarding the question item, and selects the selected learning model. may be configured to input said question data to a.

According to this, it is possible to cause the question output unit to output the reliability answer desired by the questioner as the answer from the expert in each field to the question in various fields.

Thus, according to the expert answer prediction system of the fourth invention, a user who wishes to obtain an answer to a question from an expert in a specific field inputs question data in response to a request given by the questioner. It is possible to select a learning model to be used and output a pseudo answer from an expert at the level desired by the user.

The block diagram which shows the structure of the expert answer prediction system of embodiment of this invention.

One embodiment of the invention is described below with reference to FIG. The expert answer prediction system 1 of this embodiment includes a machine learning device 10 that configures digital clones C1, C2, . . . , CN of a plurality of experts E1, E2, . Hereinafter, any one of the experts E1, E2, .

Each expert Ei is a real individual expert with expertise in a particular field. The individual expert can be, for example, a widely known (prominent) expert in a particular field or in the general public.

A specific field in which each expert Ei has expertise (hereinafter sometimes referred to as a specialized field) may be any field regardless of its type or breadth. The specialized fields include, for example, academic fields such as physics and linguistics, fields related to specific organisms such as dogs, cats, fish, and insects, fields related to daily life such as cleaning, cooking, and laundry, and each expert It may be a field that Ei claims to be his/her own field of expertise or field of expertise, or a field that is generally recognized as a field of expertise or specialty of each expert Ei.

The specialist fields of the experts E1 to EN may be different from each other, but the specialist fields of the experts E1 to EN may overlap in whole or in part. Further, the specialized field of each expert Ei may span multiple fields.

The machine learning device 10 is composed of, for example, one or more computers including a processor such as a microcomputer, a memory, an interface circuit, a communication device, etc., and an external network NW (wide area network) composed of the Internet, a telephone communication network, etc. It is configured to be able to communicate with the communication terminal 20 used by the user P via the external network NW.

Then, the machine learning device 10 can build (generate) a digital clone Ci as a learning model for each expert Ei by the functions implemented by the implemented hardware configuration and program (software configuration). and functions as a learning model storage unit 11 for storing and holding the digital clones Ci of each expert Ei, and question data indicating the content of questions given by the user P via the communication terminal 20 are stored digitally. A function as a question input unit 12 for inputting to a required digital clone Cx (=C1, or C2, or, . . . , or CN) among clones C1 to CN, and a learning model evaluation unit 15 that evaluates digital clones C1 to CN as a plurality of learning models constructed by the machine learning device 10. there is

In this case, the machine learning device 10 learns, for each expert Ei, a lifelog including at least transmissions of each expert Ei (remarks made by experts, SNS, blogs, publications, books, papers, etc.). Using it as data, a machine learning process constructs a digital clone Ci of each expert Ei. The lifelog of each expert Ei is information that includes a lot of information unique to each expert Ei or characteristic information. The life log includes, for example, attribute information of each expert Ei (age, gender, location, specialty field, hobby, etc.), SNS (Social Networking Service) information sent by each expert Ei, and meetings of each expert Ei. It is possible to use recording data or video recording data etc. at meetings such as.

The worker can input such a life log to the machine learning device 10 through an appropriate input device at any time. Alternatively, the machine learning device 10 can automatically collect information from the external network NW such as the Internet for each expert Ei.

The machine learning device 10 constructs a learning model corresponding to each expert Ei by machine learning processing using life logs given to each expert Ei as learning data, and applies the learning model to each expert Ei is stored in the learning model storage unit 11 as a digital clone Ci of . In this case, the learning model as a digital clone Ci of each expert Ei functions as artificial intelligence capable of simulating the thinking and behavior of the expert Ei. is given, an answer (response information) predicted as the answer of the expert Ei himself can be generated.

The input of new lifelogs of each expert Ei to the machine learning device 10 and the machine learning process for each expert Ei are continuously performed. Along with this, the digital clone Ci of each expert Ei is updated so as to increase its reliability as an artificial intelligence that simulates the thoughts and behaviors of each expert Ei.

As described above, the machine learning device 10 constructs a plurality of digital clones C1 to CNi as learning models for the respective experts E1 to EN, and stores and holds them in the learning model storage unit 11. FIG. As the technology for generating the digital clones C1 to CN, for example, the technology proposed in Japanese Unexamined Patent Application Publication No. 2018-190457 can be adopted.

The learning model evaluation unit 15 creates a group (set) by combining the constructed digital clones C1 to CN, and evaluates the degree of approximation between the actual answer group of each group (set) of experts and the answer group of the learning model. By evaluating high and low, it is possible to verify the certainty of the expert's answer as a group (group).

In other words, an expert side response group that is a set of answers obtained from all or part of experts E1 to En to questions in a survey conducted in the past on all or part of experts E1 to En and the degree of approximation to the model-side answer group, which is a set of answers generated by each of the digital clones C1 to CN, which are learning models, for the question.

The learning model evaluation unit 15 evaluates the degree of approximation between the acquired expert response group and the model response group as follows. In the following description, any one question out of a plurality of questions may be referred to as question Qx.

That is, for each of a plurality of questions, the learning model evaluation unit 15 uses the expert side answer group corresponding to each question Qx and the model side answer group to determine the feature amount of the tendency of the answer to the question Qx by the expert. It is specified for each of the side response group and the model side response group. As the characteristic amount of the tendency of answers to each question Qx (hereinafter referred to as the characteristic amount of answer tendency), for example, the content of the answer with the highest frequency (the number of answers) A set of answer contents for each of a plurality of ranking frequencies (for example, first and second ranking frequencies) including answers with the frequency of .

Then, for each question Qx, the learning model evaluation unit 15, when the answer tendency feature amount in the expert side answer group and the answer tendency feature amount in the model side answer group match, It is determined that the degree of approximation between the group and the model side answer group is high, and if their answer tendency feature values do not match, it is determined that the degree of approximation between the expert side answer group and the model side answer group for question Qx is low. .

For example, it is assumed that there are four types of answers a, b, c, and d to a certain question Qx, and the answer content of the first order frequency (maximum frequency) is adopted as the answer tendency feature amount for the question Qx. do. In this case, when the answer content of the first order frequency in the expert side answer group for the question Qx and the answer content of the first order frequency in the model side answer group are mutually identical (for example, both If the answer content is a), it is determined that the degree of approximation between the expert side answer group and the model side answer group for the question Qx is high.

On the other hand, when the answer content of the first-order frequency in the expert side answer group and the answer content of the first-order frequency in the model side answer group for the question Qx are different from each other (for example, one side answer If the content is a and the answer content of the other side is c), it is determined that the degree of approximation between the expert side answer group and the model side answer group for the question Qx is low.

Further, for example, there are four types of answer contents a, b, c, and d to a certain question Qx, and as the answer tendency feature quantity for the question Qx, the answer contents of the first order frequency (maximum frequency) and the second Assume a case where a combination with the answer contents of the frequency of ranking (the frequency next to the maximum frequency) is adopted. In this case, a set of the first-order frequency response content and the second-order frequency response content in the expert side answer group for the question Qx, and the first-rank frequency response content and the second-order frequency in the model side answer group (For example, if the set of the answer content of the first order and the set of the answer content of the second order are in both the expert side answer group and the model side answer group In both cases, it is determined that the degree of approximation between the expert side answer group and the model side answer group for the question Qx is high.

On the other hand, a set of the response content of the first order frequency and the response content of the second order frequency in the expert side answer group for the question Qx, and the answer content of the first order frequency and the second order frequency in the model side answer group If the set of answer contents of the frequency is different from each other (for example, the set of answer contents on one side is a set of a (: first order) and b (: second order), and the answer on the other side If the set of contents is a (: first order) and c (: second order)), it is determined that the degree of approximation between the expert side answer group and the model side answer group for the question Qx is low.
Supplementally, the answer tendency feature amount does not need to be the same for all of the plurality of questions, and may be different according to the type and content of the question.

In addition, the answer tendency feature amount for each question Qx is not limited to the content of the answer with the frequency of the first rank or the set of the content of the answers with the frequency of a plurality of ranks. For example, the answer tendency feature amount for question Qx is obtained in addition to the first-order frequency answer content for question Qx, or in addition to a set of answer content for each of a plurality of rank frequencies including the first-order frequency. , the proportion of each of the one or more rank frequencies, such as the proportion of first rank frequencies. Here, the ratio of the frequency of an arbitrary rank, such as the ratio of the frequency of the first rank, is the frequency of the rank with respect to the total number of answers to the question Qx (the total number of each of the expert side answer group and the model side answer group) means the ratio (ratio) of

In this way, when the ratio of frequency of one or more ranks for question Qx is included in the answer tendency feature amount, it is determined whether or not the answer tendency feature amount for question Qx matches between the target side answer group and the model side answer group. In order to do so, in addition to the judgment result of whether or not the response content (or the set of response content) matches, the frequency ratio of the one or more rankings matches between the target side response group and the model side response group (Including the case where they almost match) is required.

In this case, whether or not the frequency ratio of any one rank matches between the target response group and the model response group is determined by, for example, the ratio of the frequency in the target response group and the model response group Depending on whether the difference between the two ratios and the ratio of the frequency in is within a predetermined range near zero, or whether the ratio of the two ratios is within a predetermined range near "1" It can be determined by whether or not

As described above, the learning model evaluation unit 15 evaluates (determines) the degree of approximation between the expert-side answer group and the model-side answer group for each of a plurality of questions. Furthermore, the learning model evaluation unit 15 selects a group of answers from the expert side and a If the ratio of the total number of questions determined to be highly similar to the answer group (ratio to the total number of questions) is greater than or equal to a predetermined threshold, the entire expert side answer group and the model side answer group It is determined that the degree of approximation with the whole is high. In addition, if the ratio of the total number of questions judged to have a high degree of similarity between the expert side response group and the model side response group is lower than a predetermined threshold, the total of the expert side response group and the model side response group It is determined that the degree of approximation to the whole is low.

In addition, for example, a weight value (1.0, 1.5, 0.8, etc.) is set for each of multiple questions according to the type of question, etc., and the degree of approximation between the expert side answer group and the model side answer group is high. Whether or not the value obtained by summing the weight values of each question determined as is above a predetermined threshold is used to determine whether or not the degree of approximation between the entire expert response group and the entire model response group is high. It is also possible to evaluate In this case, for example, it is possible to set a high weight value for a question that is likely to have characteristics as a tendency of answers from experts.

Then, the learning model evaluation unit 15 performs the above learning model evaluation with the actual answer group of each expert group (set) by changing the group (set) combining the constructed digital clones C1 to CN. Evaluate the degree of approximation with the answer group of the learning model.

In particular, the learning model evaluation unit 15 reconstructs a group of a plurality of experts (a large group that guarantees certainty) as a large group that combines small groups having the degree of approximation equal to or greater than a predetermined threshold, while A group of the plurality of experts is also reconstructed as a large group (a large group lacking certainty) by combining the small groups having the degree of approximation of less than the threshold of .

When combining small groups and large groups, the hierarchical relationship of specialist types (for example, internal medicine⊃gastrointestinal medicine⊃gastrointestinal medicine) is appropriately referred to.

Although an example of the evaluation method is shown in the present embodiment, the evaluation method is not limited to this, and for example, a known statistical test method such as the t-test may be used.

Next, the communication terminal 20 used by the user P is configured by, for example, a smartphone, tablet terminal, personal computer, etc., and includes a display section 20a configured by a liquid crystal display or the like and a sound generating section 20b configured by a speaker or the like. A predetermined service application (hereinafter simply referred to as a service application) for using the machine learning device 10 is pre-installed in the communication terminal 20 .

Then, the communication terminal 20 can communicate with the machine learning device 10 via the external network NW while the service application is activated. In this case, the communication terminal 20 has a function of receiving a question from the user P and transmitting it to the machine learning device 10, or receiving an answer corresponding to the question from the machine learning device 10 and notifying the user P It has functions.

Next, the operation of the expert answer prediction system 1 of this embodiment will be described. When a user P wants to obtain an answer to a required question from one of a plurality of experts E1 to EN, an expert Ex, the user P who asks the question connects the Launch an installed service app. Then, the user P inputs a question to the communication terminal 20 by operating the operation unit (touch panel, operation switch, keyboard, etc.) of the communication terminal 20 or by voice input operation. When the user P wants to designate a desired expert as the question target expert Ex, the user P inputs the identification information (name, etc.) of the desired expert into the communication terminal 20 together with the questions. Further, an expert group can be designated in place of the desired expert.

First, the case of asking a question to a specific expert will be explained.

For example, when user P wants to ask a specific expert who is good at cleaning and tidying up how to tidy up his room, he can say, "Mr. I think so, but what should I do?” is input into the communication terminal 20.

Also, for example, when user P wants to ask a prominent expert in drama critics, etc., about the evaluation of a particular drama, he can say, "Mr. XX (expert's name), How is the content?” is input into the communication terminal 20.
If the expert to be questioned is not specified, the input of identification information such as the name of the expert is omitted.

The communication terminal 20 to which the questions are input as described above transmits question transmission data indicating the input questions to the machine learning device 10 via the external network NW. The question transmission data is transmitted to the machine learning device 10 as text data or voice signal data, for example.

In the machine learning device 10 that has received the question transmission data, the question input unit 12 generates text-format question data indicating the content of the question from the received question transmission data. When the received question transmission data is text data, the question transmission data may be used as it is as the question data.

Then, if a desired expert is specified in the received question transmission data, the question input unit 12 selects the specified digital clones C1 to CN stored and held in the learning model storage unit 11. Enter the question data into the digital clone Cx of the expert Ex. As a result, the question data is input to the digital clone Cx determined according to the request of the user P who is the questioner.

In addition, if the desired expert is not specified in the received question transmission data, the question input unit 12, from the question transmission data or the generated question data, based on the keyword etc. included in the question content, Identify the field of question content. Then, the question input unit 12 selects a digital clone of one expert corresponding to the specified field from among the digital clones C1 to CN stored and held in the learning model storage unit 11. Clone Cx is selected, and question data is input to the selected digital clone Cx. As a result, the question data is input to the digital clone Cx determined according to the content of the question.

The digital clone Cx to which the question data is input generates, in text data format, an answer that is expected to be answered by the expert Cx to the question content indicated by the question data, and the answer output unit 13 receives the generated answer. hand over. The reply output unit 13 transmits reply transmission data indicating the contents of the reply to the communication terminal 20 of the user P via the external network NW. The reply transmission data is transmitted to the communication terminal 20 as text data or voice signal data, for example.

Then, the communication terminal 20 that has received the reply transmission data notifies the user P of the content of the reply indicated by the reply transmission data. In this case, the content of the answer is displayed as character information on the display unit 20a of the communication terminal 20, or is output from the pronunciation unit 20b of the communication terminal 20 as voice information. When notifying the user P of the content of the answer, the image of the questioned expert Ex may be displayed on the display unit 20a of the communication terminal 20. FIG. Further, when the content of the answer is output from the communication terminal 20 as voice information, an image that looks as if the expert Ex is actually speaking may be displayed on the display unit 20a.

Next, we describe the case of asking a question to a digital clone of a particular group of experts.

For example, when asking a question about a group of specialists consisting of multiple physicians, for the specialist group of physicians, the evaluation result "the degree of approximation between the entire expert-side response group and the model-side response group as a whole is determined to be high." In addition, answers to questions can be obtained from some or all of the specialist group of physicians. In addition, when obtaining answers from all, the majority value of the answers may be shown in several patterns from the largest one.

In addition, if there is a group of gastroenterology and neurology in internal medicine as a group that is judged to have a particularly high degree of similarity between the entire response group on the expert side and the entire response group on the model side, which is the evaluation result In addition to obtaining answers from a group of experts in gastroenterology and neurology, which are highly similar groups, if a group of infectious disease internal medicine is created, if the degree of similarity is low (lack of certainty, Answers can also be obtained from a panel of infectious disease specialists (knowingly at times).

In addition, in addition to obtaining answers from a large group that combines groups with a high degree of similarity (small groups), it is also possible to obtain answers from a large group that combines groups with a low degree of similarity (small groups). For example, an answer can be obtained from a group of highly similar specialists combining gastroenterology and neurology, which are groups with a high degree of similarity.

As described above, the user P can obtain the answer to the question for the question target expert Ex from the digital clone Cx corresponding to the expert Ex and the digital clones of the group of experts, which is a group of these, based on the likelihood that the user wants. It can be obtained in a form that guarantees the level. In this case, since the digital clone Cx is constructed by machine learning processing so as to be able to simulate the thinking and behavior of the expert Ex, the user P can give an answer equivalent to the answer obtained from the expert Ex himself. can be obtained artificially. In particular, when the user P designates an expert Ex to be questioned, the user P obtains an answer from the digital clone Cx as if an answer was actually obtained from the desired expert Ex. can.

In the above-described embodiment, the communication terminal 20 used by the user P is used to input questions and notify the content of answers. 10 may be attached.

Further, in the above-described embodiment, even if the user P does not specify the desired expert, for example, the user P can specify (request) attributes such as the gender and location of the expert to be questioned. Question data may be input to digital clones corresponding to experts and expert groups selected according to specified attributes from among experts corresponding to the field of question content.

1... expert answer prediction system, 10... machine learning device, 11... learning model storage unit, 12... question input unit, 13... answer output unit, 15... learning model evaluation unit.

Claims (4)

An expert answer prediction system that predicts answers from experts who have expertise in a specific field,
A machine learning device that builds and stores learning models of a plurality of experts by machine learning processing,
The machine learning device
A plurality of learning models constructed so as to be able to predict the answers given by each expert to a question by learning the transmissions of the plurality of experts are stored as digital clones of the experts. a learning model storage unit for
a question input unit that acquires questions for the plurality of experts and inputs question data indicating the content of the questions to the plurality of learning models;
an answer output unit that outputs answers predicted by the plurality of learning models as answers of the plurality of experts in response to the question data input from the question input unit;
a learning model evaluation unit that evaluates the plurality of learning models,
The learning model evaluation unit
An expert side answer group, which is a set of answers of each of the plurality of experts to questions asked by each of the plurality of experts by changing the group of the plurality of experts, and the question configured to evaluate the degree of approximation with a model-side answer group, which is a set of answers generated by each of the plurality of learning models, identifying a plurality of groups of the experts with different degrees of approximation, 1. An expert answer prediction system, wherein answers to questions to be asked to an expert can be acquired from the plurality of learning models corresponding to the plurality of identified groups of experts. In the expert answer prediction system according to claim 1,
The learning model evaluation unit identifies a plurality of small groups included in the plurality of groups of experts, re-evaluates the degree of approximation for the identified small groups, and determines the degree of approximation equal to or greater than a predetermined threshold. An expert answer prediction system characterized by reconstructing a group of a plurality of experts as a large group combining the small groups having In the expert answer prediction system according to claim 1,
The learning model evaluation unit identifies a plurality of small groups included in the plurality of groups of experts, re-evaluates the degree of approximation for the identified small groups, and determines the degree of approximation that is less than a predetermined threshold. An expert answer prediction system characterized by reconstructing a group of a plurality of experts as a large group combining the small groups having In the expert answer prediction system according to any one of claims 1 to 3,
In the machine learning device, the question input unit selects one of the plurality of learning models according to the content of the acquired question item and the degree of approximation requested by the questioner regarding the question item. An expert answer prediction system configured to select and input the question data into the selected learning model.

Images