Information processing method, program and information processing device

Abstract

To provide an information processing method, etc., lightening the burden on a medical counselor.SOLUTION: An information processing method comprises: acquiring a question related to medical counseling; specifying a classification of the acquired question; determining a prompt corresponding to the specified classification; and inputting the prompt including the question to a language model to output an answer. Preferably, the information processing method includes determining whether an improper term is included in the output answer. Preferably, the information processing method includes specifying the classification by inputting the acquired question to the learning model which specifies the classification that the question belongs to among a plurality of classifications on inputting the question.SELECTED DRAWING: Figure 1

Description

[Technical Field]

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

[0002] Patent Document 1 discloses a genetic counseling support system that supports the routine work of genetic counselors. [Prior art documents] [Patent documents]

[0003] [Patent Document 1] Patent Application No. 2021-64706 Summary of the Invention [Problem to be solved by the invention]

[0004] However, the invention of Patent Document 1 has a problem in that it cannot output answers to questions related to medical counseling.

[0005] In one aspect, an object is to provide an information processing method etc. that reduces the burden on medical counselors. [Means for solving the problem]

[0006] An information processing method according to one aspect includes acquiring a question related to medical counseling, identifying a classification for the acquired question, determining a prompt according to the identified classification, and inputting the prompt including the question into a language model to output an answer. [Effects of the Invention]

[0007] On the one hand, it will be possible to reduce the burden on medical counselors. [Brief explanation of the drawings]

[0008] [Figure 1]FIG. 1 is an explanatory diagram showing an overview of a medical counseling system. [Figure 2] FIG. 2 is a block diagram illustrating an example of the configuration of a server. [Figure 3] FIG. 10 is an explanatory diagram illustrating an example of a common format file. [Figure 4] FIG. 10 is an explanatory diagram showing an example of a record layout of a training DB. [Figure 5] FIG. 1 is an explanatory diagram of a learning model. [Figure 6] FIG. 10 is an explanatory diagram showing an example of a record layout of a question and answer DB. [Figure 7] FIG. 1 is a block diagram illustrating an example of the configuration of a computer. [Figure 8] FIG. 10 is an explanatory diagram showing a first example of a prompt. [Figure 9] FIG. 10 is an explanatory diagram showing a second example of a prompt. [Figure 10] FIG. 10 is an explanatory diagram showing an example of displaying answers. [Figure 11] 10 is a flowchart showing the procedure for constructing a learning model. [Figure 12] 10 is a flowchart illustrating an example of a processing procedure of the medical counseling system. [Figure 13] FIG. 10 is a block diagram showing an example of the configuration of a server according to the second embodiment. [Figure 14] FIG. 10 is an explanatory diagram showing an example of an inappropriate term file. [Figure 15] FIG. 10 is an explanatory diagram of a second learning model. [Figure 16] FIG. 10 is an explanatory diagram showing an example of an alert display. [Figure 17] FIG. 10 is an explanatory diagram of a second learning model that outputs an inappropriateness score. [Figure 18] 10 is a flowchart showing an example of a processing procedure of a medical counseling system according to a second embodiment. DETAILED DESCRIPTION OF THE INVENTION

[0009] (Embodiment 1) FIG. 1 is an explanatory diagram showing an overview of a medical counseling system. The medical counseling system identifies a category of a question related to medical counseling, and outputs an answer by inputting a prompt corresponding to the identified category into a language model. The medical counseling system includes an information processing device 1 and an information processing device 2. The medical counseling system transmits and receives information via a network N. Medical counseling includes genetic counseling, mental counseling, and psychological counseling. In this embodiment, medical counseling will be described as genetic counseling.

[0010] The information processing device 1 is an information processing device that processes, stores, and transmits / receives various types of information. The information processing device 1 is, for example, a server device, a personal computer, or a general-purpose tablet PC (personal computer). The information processing device 1 may also be a cloud server device that provides functions included in the information processing device 1 as cloud services. In this embodiment, the information processing device 1 will be described as a server 1.

[0011] The information processing device 2 is, for example, a server device, a smartphone, a tablet, a personal computer (hereinafter referred to as a computer), a general-purpose tablet PC (personal computer), etc. In this embodiment, the information processing device 2 will be described as a computer 2.

[0012] 2 is a block diagram showing an example of the configuration of a server. The server 1 includes a control unit 11, a storage unit 12, a communication unit 13, a mass storage unit 14, and a reading unit 15. The above-mentioned units are connected to each other via a bus. The control unit 11 is configured using one or more processors such as a CPU (Central Processing Unit), an MPU (Micro-Processing Unit), a GPU (Graphics Processing Unit), a GPGPU (General-purpose computing on graphics processing units), a TPU (Tensor Processing Unit), or an AI chip (semiconductor for AI).

[0013] The storage unit 12 includes a RAM (Random Access Memory) or a ROM (Read Only Memory), etc. The storage unit 12 stores various data and the like necessary for a control program 12P (program product) executed by the control unit 11. The storage unit 12 also temporarily stores data and the like generated when the control program 12P is executed. The control unit 11 executes the control program 12P stored in the storage unit 12 as appropriate, thereby performing various information processing and control processing related to the server 1.

[0014] Furthermore, the storage unit 12 stores a language model 12M. The language model 12M is, for example, a general-purpose large-scale language model (LLM) constructed by performing unsupervised pre-learning using a large set of sentences or the like. The language model 12M is, for example, CALMS-7B-Chat, Med-PalM, GPT (Generative Pre-trained Transformer)-3, GPT-3.5, GPT-4, RWKV (Receptance Weighted Key Value), PaLM2, or LLaMa (Large Language Model Meta AI). When a prompt according to a classification described below is input, the language model 12M performs a calculation to generate an answer according to the prompt and outputs the generated answer.

[0015] The language model 12M may be configured by combining a plurality of algorithms. Instead of storing the language model 12M in the storage unit 12, the control unit 11 may access a language processing server that stores the language model 12M and input a prompt to the language model 12M.

[0016] The communication unit 13 is a communication module that transmits and receives information to and from the computer 2 via the network N. The mass storage unit 14 includes a RAM (Random Access Memory) or a ROM (Read Only Memory), etc. The mass storage unit 14 stores a common format file 141, a training DB 142, a learning model 143, a question and answer DB 144, and a reference information file 145. The common format file 141, the training DB 142, the learning model 143, and the question and answer DB 144 will be described later. The reference information file 145 stores reference information data. The reference information data includes, for example, guidelines for genetic testing, an ophthalmological glossary, and the Journal of the Japanese Ophthalmological Society. The reference information file 145 stores, for example, original reference information data and vectorized reference information data.

[0017] A method for storing reference information data in the reference information file 145 will be described. The control unit 11 reads the reference information data via the communication unit 13. The reference information data includes data published on the Internet, data converted from specialized books into text files, independently created data, and in-hospital examination data. The control unit 11 divides the read reference information data into chunks. A chunk is a component of the reference information data, such as a document, a paragraph, a sentence, a clause, and a word. The control unit 11 vectorizes the divided reference information data. The control unit 11 stores the original reference information data and the vectorized reference information data in the reference information file 145.

[0018] The control unit 11 may convert the data format of the read reference information data. In this case, the control unit 11 vectorizes the converted reference information data and stores the vectorized reference information data in the reference information file 145. The reference information file 145 may be stored in an external storage device (not shown).

[0019] In this embodiment, the storage unit 12 and the mass storage unit 14 may be configured as an integrated storage device. The mass storage unit 14 may be configured by a plurality of storage devices. The mass storage unit 14 may be an external storage device connected to the server 1.

[0020] The reading unit 15 reads information stored in the portable storage medium 1a. The portable storage medium 1a is, for example, a CD (Compact Disc), a DVD (Digital Versatile Disc), a USB (Universal Serial Bus) memory, or an SD (Secure Digital). The reading unit 15 reads the control program 12P from the portable storage medium 1a. The control unit 11 stores the read control program 12P in the storage unit 12. The control unit 11 may download the control program 12P from another computer via the network N. In this case, the control unit 11 stores the downloaded control program 12P in the storage unit 12. The control unit 11 may store the read control program 12P in the mass storage unit 14.

[0021] In this embodiment, the server 1 may be configured with multiple servers. The server 1 may be a virtual machine virtually constructed by software within a single device. The server 1 may be a local server installed within the facility where the server 1 is located. The server 1 may be a cloud server connected for communication via a network N. Furthermore, the control program 12P may be executed on a single server, or may be distributed and executed on multiple servers interconnected via the network N.

[0022] FIG. 3 is an explanatory diagram showing an example of a common format. The common format file 141 stores a common format 3 for creating prompts to be input into the language model 12M. The common format 3 includes a common role information field 3a, a guideline field 3b, and a reference information field 3c. The common role information field 3a contains common role information including instructions for the language model 12M. The common role information field 3a in FIG. 3 contains the message, "You are a certified genetic counselor. Please answer the questions according to the following guidelines." The guideline field 3b contains multiple guidelines for preventing the language model 12M from outputting inappropriate answers.

[0023] The guideline column 3b in Figure 3 lists the following guidelines, along with specific examples: 1. Do not recommend prenatal testing, 2. Do not provide numerical values ​​such as the probability of inheritance, 3. Do not sympathize with anxious questions, 4. Do not use expressions that incite anxiety, 5. Do not make statements that criticize the questioner's relatives, 6. Do not make statements that take excessive responsibility, and 7. Do not share incorrect information. The reference information column 3c lists information that assists the language model 12M in outputting answers. The administrator of the medical counseling system can edit the contents of the guideline column 3b through the input unit 24.

[0024] The reference information field 3c in FIG. 3 states, "Information related to the patient's question will be searched for and entered from an external database." When the control unit 11 identifies a portion of the reference information data related to the patient's question from the mass storage unit 14, it replaces the above-mentioned text in the reference information field 3c with the portion of the reference information data. The content of the common format 3 is not limited to FIG. 3. The content of the common format 3 can be changed as appropriate to suit the embodiment.

[0025] FIG. 4 is an explanatory diagram showing an example of a record layout of a training DB. The training DB 142 stores training data used to construct the learning model 143. The training DB 142 includes a question column and a classification column. The question column stores questions to be input to the learning model 143. The classification column stores classifications for the questions. In the example of FIG. 4, the input data, the question "Is there anything I should be careful about in terms of lifestyle habits when it comes to retinitis pigmentosa?" is output in correspondence with the output data, the classification "lifestyle."

[0026] FIG. 5 is an explanatory diagram of a learning model. The learning model 143 is a trained model that has been trained to output information indicating a classification (hereinafter referred to as classification information) to which a question related to medical counseling (hereinafter referred to as a question) belongs from among multiple classifications when the question is input. For example, BERT (Bidirectional Encoder Representations from Transformers) is used for the learning model 143. The learning model 143 includes an input layer that receives a question, an intermediate layer that performs calculations, and an output layer that outputs classification information. The classification includes, for example, lifestyle, genetic testing preference, research, treatment, prognosis, and genetics. The classification is not limited to the above example. The classification information is, for example, a probability value that estimates the classification of the question or a classification to which the question belongs. In this embodiment, the classification information is described as a probability value that estimates the classification of the question. The learning model 143 divides the question into tokens (e.g., words) generated by morphological analysis or the like. The learning model 143 performs calculations on the divided tokens to output a probability value that estimates the classification of the question. The probability values ​​output from the output layer are, for example, values ​​between 0 and 1. The sum of the probability values ​​output from the output layer is 1.0 (100%).

[0027] In FIG. 5, when the control unit 11 inputs the question "Does retinitis pigmentosa occur more frequently in men?" into the learning model 143, the control unit 11 outputs probability values ​​"lifestyle: 0.1, desire for genetic testing: 0.01, research: 0.04, treatment: 0.03, prognosis: 0.02, and genetics: 0.8" that estimate the classification of the question from the learning model 143. The control unit 11 identifies the classification with the highest probability value from among the multiple probability values ​​output from the output layer as the classification for the question. In FIG. 5, the control unit 11 identifies "genetics" as the classification for the question.

[0028] The control unit 11 may use a predetermined threshold to identify the category to which the question belongs. In the example of Fig. 5, the control unit 11 identifies "genetics" as the category to which the question belongs when the probability value of "genetics" (0.8) is equal to or greater than a predetermined threshold (e.g., 0.75). The predetermined threshold can be changed as appropriate depending on the embodiment.

[0029] The procedure for constructing the learning model 143 will be described. The control unit 11 reads out training data in which questions are associated with classifications for the questions from the training DB 142. The control unit 11 uses the above-mentioned training data to train the learning model 143. Specifically, the control unit 11 optimizes the parameters of the learning model 143 so that there is no difference between the classification output from the learning model 143 when a question is input and the classification for the question in the training data. The control unit 11 optimizes the parameters using, for example, backpropagation or steepest descent. The parameters are weights between neurons in the learning model 143. The control unit 21 constructs the learning model 143 by repeating the above process using the training data. The control unit 21 stores the constructed learning model 143 in the mass storage unit 14.

[0030] The construction process of the learning model 143 may be performed by an external computer (not shown), etc. In that case, the constructed learning model 143 is deployed to the server 1 from the external computer, etc. via the network N.

[0031] The learning model 143 may use a classification algorithm such as LSTM (Long Short Term Memory), RNN (Recurrent Neural Network), GPT (Generative Pre-trained Transformer), SVM (Support Vector Machine), CNN (Convolution Neural Network), Seq2Seq (Sequence to Sequence), random forest, or decision tree.

[0032] FIG. 6 is an explanatory diagram showing an example of a record layout of a question and answer database. The question and answer database 144 includes a category column and a question and answer column. The category column stores a category for each question. The question and answer column stores instructions for each category for the language model 12M and questions and answers corresponding to the category. In the example of FIG. 6, the category "daily life," the instruction for each category "The question is about daily life. Please refer to the following example questions and answers about daily life and provide an answer that will resolve the question's question about daily life," and the question and answer "Question: For retinitis pigmentosa, are there any lifestyle habits that should be taken into consideration? Answer: Most cases of retinitis pigmentosa are..." are associated and stored. In the example of FIG. 6, one question and answer is stored for each category, but this is not limiting. Multiple questions and answers may be stored for each category.

[0033] 7 is a block diagram showing an example configuration of a computer. The computer 2 includes a control unit 21, a storage unit 22, a communication unit 23, an input unit 24, and a display unit 25. The above-mentioned units are connected to each other via a bus. The control unit 21 is configured using one or more processors such as a CPU, an MPU, or a GPU. The storage unit 22 includes a RAM or a ROM. The storage unit 22 stores various data and the like required for a control program 22P (program product) executed by the control unit 21. The storage unit 22 temporarily stores data and the like generated when executing the control program 22P. The control unit 21 executes the control program 22P stored in the storage unit 22 as needed to perform information processing and control processing related to the computer 2.

[0034] The communication unit 23 is a communication module that transmits and receives information to and from the server 1 via the network N. The input unit 24 accepts operational inputs from a genetic counseling user. The input unit 24 includes, for example, a keyboard, a mouse, and a microphone for voice input. The display unit 25 is a liquid crystal display or an organic EL display. The display unit 25 outputs various information in accordance with instructions from the control unit 21. The input unit 24 and the display unit 25 may be a touch panel configured as an integrated unit.

[0035] The processing of embodiment 1 will be described. A patient using the medical counseling system inputs a question via the input unit 24. An input question is, for example, "Does retinitis pigmentosa occur more frequently in men?" The control unit 21 acquires the question input by the user. The control unit 21 transmits the acquired question to the server 1.

[0036] The control unit 11 receives a question sent from the computer 2. The control unit 11 inputs the question into the learning model 143, thereby obtaining a probability value that estimates the classification of the question. The control unit 11 identifies the classification of the question based on the obtained probability value. For example, when the question "Does retinitis pigmentosa occur more frequently in men?" is input into the learning model 143, the control unit 11 identifies the classification of the question as "genetic." The control unit 11 reads out a question and answer corresponding to the identified classification "genetic" from the question and answer DB 144. For example, when the identified classification is "genetic," the control unit 11 reads out instructions for each classification, such as "The question is about genetics. Please refer to the following example questions and answers about genetics and provide an answer that will resolve your concerns in daily life," and a question and answer, such as "Question: If the genetic test is negative, then...? Answer: No, the genetic test shows that it is the cause..." from the question and answer DB 144.

[0037] The control unit 11 may identify the category of a question using keywords included in the question instead of the learning model 143. In this case, the control unit 11 stores in advance in the mass storage unit 14 categories such as lifestyle, genetic testing preference, research, treatment, prognosis, and genetics, and keywords corresponding to each category. For example, "diet" is stored in the mass storage unit 14 as a keyword corresponding to "lifestyle." The control unit 11 identifies multiple keywords included in the question, for example, by dividing the question into words. The control unit 11 refers to the mass storage unit 14 and counts to which categories the identified multiple keywords belong. The control unit 11 identifies the category with the most occurrences among the counted keyword categories as the category of the question.

[0038] The control unit 11 vectorizes the question. The control unit 11 refers to the reference information file 145 and calculates the similarity between the vectorized question and each portion of the vectorized reference information data. When multiple pieces of reference information data are stored in the reference information file 145, the control unit 11 performs the same process on the multiple pieces of reference information data. For example, cosine similarity or k-nearest neighbor algorithm is used to calculate the similarity between vectors. The control unit 11 identifies a portion of the vectorized reference information data (a portion of the reference information data related to the content of the question) whose calculated similarity is equal to or greater than a predetermined threshold value among the multiple pieces of vectorized reference information data. The predetermined threshold value can be changed as appropriate depending on the embodiment. The control unit 11 reads a portion of the original reference information data (hereinafter referred to as reference information) corresponding to the identified portion of the reference information data from the reference information file 145. The reference information is a portion of the original reference information data related to the content of the question and assists the language model 12M in outputting an answer. The control unit 11 reads the common format 3 from the common format file 141. The control unit 11 determines a prompt to be input to the language model 12M based on the question, the question and answer, the read reference information, and the common format 3.

[0039] FIG. 8 is an explanatory diagram showing a first example of a prompt. The prompt 30 in FIG. 8 is generated when the classification is "genetic." The prompt 30 includes a common role information field 3a, a guideline field 3b, a reference information field 3c, a question and answer field 3d, and a question field 3e. The content of the common role information field 3a and the guideline field 3b in FIG. 8 is the same as that in FIG. 3. The reference information field 3c contains reference information read from the reference information file 145. The reference information field 3c in FIG. 8 contains the following: "Retinal degenerative diseases are the term for diseases that cause degeneration of photoreceptors. All diseases except age-related macular degeneration, which occurs in elderly people, are hereditary. Inherited retinal degenerations (IRDs) can be caused by genetic mutations in photoreceptors or RPE. ..." The question and answer field 3d contains questions and answers according to the classification. The question and answer section 3d in Figure 8 states, "The question is about genetics. Please answer. Question: If the genetic test is negative, then... Answer: No, the genetic test shows that it is the cause..." The question section 3e states the question entered by the user. The question section 3e in Figure 8 states, "Does retinitis pigmentosa occur more frequently in men?"

[0040] Next, a prompt 30 for a question classified as "lifestyle" will be described. FIG. 9 is an explanatory diagram showing a second example of a prompt. The prompt 30 in FIG. 9 is generated when the control unit 11 receives the question, "Please tell me how men can prevent retinitis pigmentosa." The prompt 30 includes a common role information field 3a, a guideline field 3b, a reference information field 3c', a question and answer field 3d', and a question field 3e'. The contents of the common role information field 3a and the guideline field 3b in FIG. 9 are the same as those in FIG. 8. The question field 3e' states, "Please tell me how men can prevent retinitis pigmentosa." The question and answer field 3d' in FIG. 9 contains instructions for each category: "The question is about lifestyle. Please refer to the example questions and answers below about lifestyle and provide an answer that resolves the question." and a question and answer: "Question: Are there any lifestyle habits I should be careful about when it comes to retinitis pigmentosa? Answer: In most cases of retinitis pigmentosa, the disease progresses in a ~ manner." The reference information field 3c' contains reference information related to the question in the question field 3e'.

[0041] 8 to the language model 12M. The control unit 11 acquires the answer output by the language model 12M. The control unit 11 transmits the acquired answer to the computer 2.

[0042] The control unit 21 receives the answer sent from the server 1. The control unit 21 displays a screen including the received answer on the display unit 25. FIG. 10 is an explanatory diagram showing an example of an answer display. The answer shown in FIG. 10 is an answer to the prompt in FIG. 8. The screen d04 in FIG. 10 includes an answer field d41. The answer output by the language model 12M is displayed in the answer field d41. The answer displayed in the answer field d41 in FIG. 10 is "There are some genetic types of retinitis pigmentosa that are more likely to cause symptoms in men, but this is only a small proportion of the total. For other types, there is no difference in the probability of inheritance between men and women. If you would like to know which genetic type applies to you or your family, please use genetic counseling."

[0043] In this embodiment, the control unit 11 generates a prompt based on the classification with the highest probability value among the multiple probability values ​​output from the output layer, but this is not limited to this. The control unit 11 may also generate a prompt based on the classification of the second or subsequent probability values. In this case, the control unit 11 can obtain multiple answers by inputting multiple prompts generated from one question into the learning model 143.

[0044] 11 is a flowchart showing the procedure for constructing a learning model. The control unit 11 reads out training data in which questions are associated with classifications of the questions from the training DB 142 (step S101). The control unit 11 uses the read out training data to construct a learning model 143 that takes a question as input and outputs a probability value that estimates the classification of the question (step S102). The control unit 11 stores the constructed learning model 143 in the mass storage unit 14 (step S103).

[0045] 12 is a flowchart showing an example of a processing procedure of the medical counseling system. The control unit 21 acquires a question input by a patient (step S201). The control unit 21 transmits the question to the server 1 (step S202).

[0046] The control unit 11 receives a question sent from the computer 2 (step S301). The control unit 11 inputs the question to the learning model 143 (step S302). The control unit 11 acquires a probability value that estimates the classification of the question output by the learning model 143 (step S303). The control unit 11 identifies the classification of the question based on the acquired probability value (step S304). The control unit 11 reads out a question and answer corresponding to the identified classification from the question and answer DB 144 (step S305). The control unit 11 vectorizes the question (step S306). The control unit 11 refers to the reference information file 145 and calculates the similarity between the vectorized question and each part of the vectorized reference information data (step S307). The control unit 11 identifies a part of the reference information data, from the vectorized multiple reference information data, whose calculated similarity is equal to or greater than a predetermined threshold (step S308). The control unit 11 reads out reference information corresponding to a portion of the identified reference information data from the reference information file 145 (step S309). The control unit 11 reads out the common format 3 from the common format file 141 (step S310). The control unit 11 determines a prompt to be input to the language model 12M based on the question, the question and answer, the read out reference information, and the common format 3 (step S311). The control unit 11 inputs the determined prompt into the language model 12M (step S312). The control unit 11 acquires the answer output by the language model 12M (step S313). The control unit 11 transmits the acquired answer to the computer 2 (step S314).

[0047] The control unit 21 receives the response transmitted from the server 1 (step S203). The control unit 21 displays a screen including the received response on the display unit 25 (step S204).

[0048] According to the first embodiment, the medical counseling system can identify the category of a question related to medical counseling and output an answer based on the identified category.

[0049] According to the first embodiment, the medical counseling system can reduce the burden on medical counselors by outputting answers to questions.

[0050] (Embodiment 2) The second embodiment describes a method in which the medical counseling system determines whether an output answer contains an inappropriate term.

[0051] 13 is a block diagram showing an example of the configuration of a server according to embodiment 2. The mass storage unit 14 includes a second learning model 146 and an inappropriate term file 147. The second learning model 146 is an estimator (output device) that outputs information about inappropriate terms in an answer when an answer is input. The inappropriate term file 147 stores a list of inappropriate terms and expressions that have been registered in advance.

[0052] FIG. 14 is an explanatory diagram showing an example of an inappropriate term file. The inappropriate term file 147 includes an inappropriate term field 147a and an inappropriate expression field 147b. The inappropriate term field 147a stores inappropriate terms that cannot be included in the answers of the language model 12M. The inappropriate term field 147a in FIG. 14 stores phrases such as "~%," "discrimination," "crime," "don't worry," and "it's okay." The inappropriate expression field 147b in FIG. 14 stores phrases or sentence-wide expressions such as "I understand very well," "By undergoing prenatal testing, you can check the risk of ...," and "It is your own responsibility if you have done something wrong."

[0053] The processing of the second embodiment will be described. Methods for determining whether an answer contains an inappropriate term include, for example, a second learning model or a rule base. First, a method using a learning model will be described. FIG. 15 is an explanatory diagram of the second learning model. The second learning model 146 is a trained model that has been trained to output information about inappropriate terms in an answer when an answer is input. The information about inappropriate terms in an answer is, for example, inappropriate terms or inappropriate expressions in the answer, a score indicating whether an inappropriate term or inappropriate expression is included, and a probability value indicating whether an inappropriate term or inappropriate expression is included and a probability value indicating whether an inappropriate term or inappropriate expression is not included. In this embodiment, the information about inappropriate terms in an answer will be described as inappropriate terms or inappropriate expressions in the answer.

[0054] For example, Bidirectional Encoder Representations from Transformers (BERT) is used for the second learning model 146. The second learning model 146 may use Long Short Term Memory (LSTM), Recurrent Neural Network (RNN), Generative Pre-trained Transformer (GPT), or the like.

[0055] For example, the inappropriate term file 147 or a Japanese harmful expression dataset is used as training data for the second learning model 146. The second learning model 146 includes an input layer that accepts input of an answer, an intermediate layer that performs calculations, and an output layer that outputs inappropriate terms in the answer. The second learning model 146 divides the answer into tokens (e.g., words) generated by morphological analysis or the like. The second learning model 146 performs calculations on the divided tokens to output inappropriate terms or inappropriate expressions in the answer.

[0056] In FIG. 15, the control unit 11 determines a prompt 30 including the question, "Are you worried that your genetic information may fall into the hands of a malicious person and be misused?" using the same procedure as in the first embodiment. The control unit 11 inputs the determined prompt 30 into the language model 12M. The control unit 11 acquires the answer, "...If your genetic information is passed on to a third party, there is a possibility that the third party may unfairly discriminate against you or use it for criminal purposes..." output from the language model 12M. The control unit 11 inputs the acquired answer into the second learning model 146. The control unit 11 extracts the inappropriate terms "discrimination" and "crime" from the answer output from the second learning model 146. The control unit 11 generates an alert. The alert may be, for example, a statement or the like that warns the patient who input the question. Alternatively, the alert may be in the form of highlighting (e.g., bold or hatching) or underlining inappropriate terms or expressions in the answer instead of a statement or the like. The control unit 11 transmits the acquired alert and question to the computer 2.

[0057] The control unit 21 receives the alert and question sent from the server 1. The control unit 21 displays a screen including the received alert and question on the display unit 25. FIG. 16 is an explanatory diagram showing an example of an alert display. The screen d05 in FIG. 16 includes a question field d51, an alert field d52, and a re-acquire button d53. The question field d51 displays a question entered by the patient. The question field d51 in FIG. 16 displays, "I have a chronic condition called epilepsy. The cause is unknown, and my symptoms are currently stable. I would like to become pregnant. Will this be inherited by my child?" The alert field d52 displays an alert. The alert field d52 in FIG. 16 displays, "For the above question, please directly consult a genetic counselor." The re-acquire button d53 is a button that instructs the patient to re-acquire the answer.

[0058] The contents of the alert field d52 are not limited to the example in Fig. 16. For example, the control unit 21 may display, in the alert field d52, a message or the like that alerts the patient who input the question and an answer in which inappropriate terms are highlighted.

[0059] After displaying the screen d05 of FIG. 16, the control unit 21 determines whether a predetermined time (for example, 60 seconds) has elapsed. If the control unit 21 determines that the predetermined time has elapsed, it terminates the processing. If the control unit 21 determines that the predetermined time has not elapsed, it determines whether the reacquire button d53 has been selected. If the control unit 21 determines that the reacquire button d53 has not been selected, it displays the screen d05 of FIG. 16 as is. If the control unit 21 determines that the reacquire button d53 has been selected, it sends a notification to the server 1 that it will acquire the answer again (hereinafter referred to as a reacquire notification).

[0060] The control unit 11 receives the reacquisition notification sent by the computer 2. The control unit 11 re-determines the prompt using the same procedure as in embodiment 1. The re-determined prompt contains, for example, a statement to prevent the inappropriate term or expression contained in the first answer from being included in the next answer. The subsequent processing is the same as in embodiment 1, and therefore a description thereof will be omitted.

[0061] Next, a method using a rule base will be described. The control unit 11 inputs a prompt similar to that in the second learning model 146 to the language model 12M. The control unit 11 acquires the answer in FIG. 15 output by the language model 12M, "If this information is passed on to a third party, the third party may unfairly discriminate against me or use it for criminal purposes..." The control unit 11 refers to the inappropriate term file 147 and determines whether the acquired answer contains an inappropriate term or an inappropriate expression. The control unit 11 generates an alert when it determines that the acquired answer contains an inappropriate term or an inappropriate expression. Specifically, the control unit 11 generates an alert by determining that the answer contains the inappropriate terms "discrimination" and "crime." The control unit 11 transmits the generated alert and question to the computer 2. Furthermore, when it determines that the acquired answer does not contain an inappropriate term or an inappropriate expression, the control unit 11 transmits the acquired answer to the computer 2. The subsequent processing is similar to that of the second learning model 146, and therefore description thereof will be omitted.

[0062] (Variation) In the second embodiment, an example in which the second learning model 146 outputs inappropriate terms or expressions in an answer has been described, but the present invention is not limited to this. The second learning model 146 may output a score indicating whether or not an inappropriate term is included in an answer (hereinafter referred to as an inappropriateness score) as information regarding the inappropriate term in an answer. FIG. 17 is an explanatory diagram of a second learning model that outputs an inappropriateness score. The inappropriateness score output from the second learning model 146 is, for example, a value between 0 and 1. The inappropriateness score is output as a numerical value close to "1" when an inappropriate term is included in an answer, and close to "0" when an inappropriate term is not included. The second learning model 146 divides the answer into tokens (e.g., words) generated by morphological analysis or the like. The second learning model 146 performs calculations on the divided tokens to output the inappropriateness score. The second learning model 146 is constructed, for example, from training data in which the answers output by the language model 12M are associated with an inappropriateness score of "1 (inappropriate terms included)" or "0 (inappropriate terms not included)" for the answers.

[0063] In FIG. 17 , the control unit 11 determines a prompt 30 including the question, "I have epilepsy. The cause is unknown, and my symptoms are currently stable. I would like to become pregnant. Will this be passed on to my child?" The control unit 11 inputs the determined prompt 30 into the language model 12M. The control unit 11 acquires the answer output by the language model 12M, "If you have any chronic illnesses or medical history, not limited to epilepsy, ... Also, prenatal testing can help determine the risk of your child having a congenital disease." The control unit 11 inputs the acquired answer into the second learning model 146. The control unit 11 acquires the inappropriateness score (0.8) output by the second learning model 146. The control unit 11 determines whether the inappropriateness score (0.8) is equal to or greater than a predetermined threshold (e.g., 0.75). The predetermined threshold can be changed to a value between 0 and 1 depending on the embodiment. If the inappropriateness score exceeds the predetermined threshold, the control unit 11 determines that the answer contains an inappropriate term or expression. The control unit 11 generates an alert when it determines that an answer contains an inappropriate term or an inappropriate expression. Furthermore, the control unit 11 determines that an answer does not contain an inappropriate term or an inappropriate expression when the inappropriateness score does not exceed a predetermined threshold. The subsequent processing is the same as in the second embodiment, and therefore a description thereof will be omitted.

[0064] In the second embodiment, the control unit 11 determines whether an answer contains an inappropriate term or an inappropriate expression using either the second learning model or the rule base, but this is not limited to this. The control unit 11 may also determine whether an answer contains an inappropriate term or an inappropriate expression using both the second learning model 146 and the rule base. In this case, the control unit 11 can more accurately determine whether an answer contains an inappropriate term or an inappropriate expression by applying both the second learning model 146 and the rule base.

[0065] In the second embodiment, the control unit 11 inputs a prompt to one type of language model 12M to obtain an answer, but this is not limiting. The control unit 11 may input a prompt to multiple types of language models 12M to obtain multiple answers. In this case, the control unit 11 inputs the obtained multiple answers to the second learning model 146 of FIG. 17. The control unit 11 obtains multiple inappropriateness scores output by the second learning model 146. The control unit 11 identifies, from the obtained multiple inappropriateness scores, an answer that is below a predetermined threshold and has the smallest inappropriateness score. The control unit 11 transmits the identified answer to the computer 2. The control unit 21 receives the answer transmitted from the server 1. The control unit 21 displays the received answer on the display unit 25. The medical counseling system can select the best answer for the patient from multiple answers.

[0066] FIG. 18 is a flowchart showing an example of a processing procedure of the medical counseling system according to the second embodiment. In the flowchart shown in FIG. 18, step S313 and subsequent steps in the processing shown in FIG. 12 are replaced with steps S601 to S605 and steps S501 to S505. Descriptions of steps similar to those in FIG. 12 will be omitted. The control unit 11 determines whether the acquired answer contains an inappropriate term or expression using the second learning model 146 or the rule base (step S601). If the control unit 11 determines that the acquired answer does not contain an inappropriate term or expression (step S601: NO), the control unit 11 proceeds to step S314 in FIG. 12. If the control unit 11 determines that the acquired answer contains an inappropriate term or expression (step S601: YES), the control unit 11 generates an alert (step S602). The control unit 11 transmits the acquired alert and question to the computer 2 (step S603).

[0067] The control unit 21 receives the alert and question sent from the server 1 (step S501). The control unit 21 displays a screen including the received alert and question on the display unit 25 (step S502). The control unit 21 determines whether a predetermined time (e.g., 60 seconds) has elapsed (step S503). If the control unit 21 determines that the predetermined time has elapsed (step S503: YES), it ends the process. If the control unit 21 determines that the predetermined time has not elapsed (step S503: NO), it determines whether the reacquisition button d53 has been selected (step S504). If the control unit 21 determines that the reacquisition button d53 has not been selected (step S504: NO), it returns the process to step S502. If the control unit 21 determines that the reacquisition button d53 has been selected (step S504: YES), it transmits a reacquisition notification to the server 1 (step S505).

[0068] The control unit 11 receives the reacquisition notification sent by the computer 2 (step S604). The control unit 11 determines the prompt again (step S605) in the same procedure as in the first embodiment. The control unit 11 proceeds to step S312 in FIG.

[0069] According to the second embodiment, the medical counseling system can determine whether or not an inappropriate term is included in the answer output by the language model 12M.

[0070] According to the second embodiment, the medical counseling system can prevent the display of inappropriate answers.

[0071] According to the second embodiment, the patient can ask the medical counselor for a correct judgment without viewing inappropriate answers.

[0072] The features described in each of the above embodiments can be combined with each other. Furthermore, the independent claims and dependent claims described in the claims can be combined with each other in any and all combinations, regardless of the reference format. Furthermore, the claims use a format in which a claim references two or more other claims (multiple claim format), but this is not limited to this. Multiple claims (multi-multi claims) that reference at least one other multiple claim may also be used.

[0073] The embodiments disclosed herein are illustrative in all respects and should not be considered limiting. The scope of the present invention is defined by the claims, not by the above meaning, and is intended to include all modifications within the meaning and scope of the claims. [Explanation of symbols]

[0074] 1. Information processing device (server) 11 Control section 12 Storage section 12P control program 12M language models 13 Communications Department 14 Mass storage 141 Common Format File 142 Training DB 143 Learning Model 144 Q&A DB 145 Reference Information File 146 Second Learning Model 147 Inappropriate Term File 15 Reading unit 2. Information processing equipment (computers) 21 Control section 22 Memory section 22P control program 23 Communications Department 24 Input section 25 Display section 1a Portable storage media D Reference information 3 Common format 30 prompts

Claims

1. Get medical counseling questions, Identifying a classification for the acquired question; determining a prompt according to the identified classification; Inputting the prompt, including the question, into a language model to output an answer. Information processing methods.

2. Determine whether the outputted answer contains inappropriate words The information processing method according to claim 1 .

3. When the question is input, the acquired question is input to a learning model that identifies a category to which the question belongs from among a plurality of categories, thereby identifying the category.

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

4. The categories include lifestyle, genetic testing preference, research, treatment, prognosis, and genetics.

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

5. The prompt includes a number of guidelines to prevent the output of an inappropriate answer.

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

6. When an answer is input, the answer output from the language model is input to a second learning model that outputs information about inappropriate terms in the answer, thereby outputting information about inappropriate terms in the answer. The information processing method according to claim 2 .

7. If the answer output from the language model contains an inappropriate term that has been registered in advance, an alert is output. The information processing method according to claim 2 .

8. The prompt includes common role information, the guidelines, reference information to assist in outputting the answer, questions and answers according to the classification, and the question. The information processing method according to claim 5 .

9. Get medical counseling questions, Identifying a classification for the acquired question; determining a prompt according to the identified classification; Inputting the prompt, including the question, into a language model to output an answer. A program that causes a computer to perform a process.

10. An information processing device having a control unit, The control unit Get medical counseling questions, Identifying a classification for the acquired question; determining a prompt according to the identified classification; Inputting the prompt, including the question, into a language model to output an answer. Information processing device.

Images