Program, apparatus and method for creating prompt to be input to large language model

Abstract

To provide a program for creating a prompt which improves accuracy of inference of a large language model with respect to a cognitive psychological task.SOLUTION: A program according to the present invention causes a computer to function as: personal information description means for describing, in a prompt, personal information of a user; impression description means for describing, in the prompt, a sentence instructing to form an impression on the user with respect to an inquiry, considering the personal information of the user; behavioral factor description means for describing, in the prompt, a sentence instructing to infer a behavioral factor of the user with respect to the inquiry, considering the impression on the user; inquiry description means for describing, in the prompt, a sentence instructing to deduce an inference on an answer of the user to the inquiry, considering the personal information of the user, the impression on the user, and the behavioral factor of the user. The prompt is input to a large language model, thereby, to enable obtaining an answer of the user to the inquiry.SELECTED DRAWING: Figure 1

Description

[Technical field]

[0001] The present invention relates to a technique for creating prompts to be input into Large Language Models (LLMs) based on natural language processing. [Background technology]

[0002] In recent years, large-scale language models have been improving their inference accuracy for tasks such as text generation and language understanding through the Transformer. This is built by training deep neural networks with billions of parameters using large datasets, and can generate natural-looking sentences. A representative example is ChatGPT (registered trademark), and due to its high expressiveness and versatility in output results, it is beginning to be used in a wide range of applications such as chatbots, search engines, and code generation.

[0003] Large-scale language models may have low inference accuracy when it comes to inferring causal relationships between factors and results. This is because large-scale language models learn statistical correlations in text data. In response to this, there is a technique for describing a logical chain of thought (CoT) in the input prompt in order to improve the inference accuracy of a large-scale language model (see, for example, Non-Patent Document 1). This technique describes a QA (Question-Answer) example in the prompt in order to provide intermediate inference steps leading to a final answer for a large-scale language model. There is also a technique called Zero-shot-CoT prompting, which is derived from CoT prompting (see, for example, Non-Patent Document 2). This instructs a large-scale language model to make inferences step by step. The simplest way to do this is to simply add "Let's think step by step" to the prompt. Although no output examples of the inference target are required, it is possible to improve inference accuracy with a single prompt. In this way, by prompting the user with hierarchical, logical steps to an answer, the inference accuracy of large-scale language models can be improved.

[0004] From another perspective, it has also been reported that large-scale language models have the ability to solve "Theory of Mind (ToM)" tasks, which involve inferring the minds of others. ToM refers to the ability of humans and apes to understand the mental states, intentions, knowledge, beliefs, etc. of others and infer their behavior based on that. This concept has traditionally been measured by the "false belief task" in psychology experiments, which measures the ability to understand whether others will make a mistake in judgment in a particular situation. [Prior art documents] [Non-patent literature]

[0005] [Non-Patent Document 1] Wei J. et al., “Chain-of-Thought Prompting Elicits Reasoning in Large Language Models,” arXiv.org, Jan. 28, 2022., [online], [Retrieved September 7, 2020], Internet<https: / / arxiv.org / abs / 2201.11903> [Non-Patent Document 2] T. Kojima, SS Gu, M. Reid, Y. Matsuo, and Y. Iwasawa, “Large Language Models are Zero-Shot Reasoners.” arXiv, Jan. 29, 2023. [online], [Retrieved September 7, 2023], Internet<https: / / arxiv.org / abs / 2205.11916> [Non-Patent Document 3] M. Kosinski, “Theory of Mind May Have Spontaneously Emerged in Large Language Models.” arXiv, Mar. 14, 2023., [online], [Retrieved September 7, 2020], Internet <https: / / www.researchgate.net / publication / 368304947_Theory_of_Mind_May_Have_Spontaneously_Emerged_in_Large_Language_Models> Summary of the Invention [Problem to be solved by the invention]

[0006] However, it is said that large-scale language models have limitations in the inference accuracy for tasks such as "reproducing human psychological processes" such as thoughts and emotions. It has been reported that when ChatGPT3.5 (registered trademark), a representative large-scale language model, is used to solve the above-mentioned false belief task, the inference accuracy for ToM tasks is only equivalent to that of a person at about age 9 (see, for example, Non-Patent Document 3).

[0007] In response to this, the inventors of the present application wondered whether it would be possible to incorporate an inference process of behavioral intentions that takes into account human psychological factors into a large-scale language model. For example, they thought that by simulating the information processing process of interpersonal cognition, it might be possible to infer human behavior in an explainable way.

[0008] Therefore, an object of the present invention is to provide a program, device, and method for creating prompts that can improve the inference accuracy of large-scale language models for cognitive psychology tasks that infer "causal relationships between factors and results" and "reproduction of human psychological processes." [Means for solving the problem]

[0009] According to the present invention, there is provided a program for causing a computer to generate prompts for inferring an answer of each user to an arbitrary question, to be input to a large-scale language model, the program comprising: a personal information entry means for entering personal information of the user in the prompt; impression description means for describing a sentence instructing the user to form an impression of the user with respect to the question in a prompt, taking into consideration personal information of the user; behavioral factor description means for describing a sentence instructing an estimation of the behavioral factor of the user in response to the question in a prompt, taking into consideration personal information of the user and an impression of the user; a question description means for describing a sentence in a prompt that instructs an inference of an answer of the user to the question, taking into consideration personal information of the user, an impression of the user, and a behavioral factor of the user; The prompt is then input to a large-scale language model to obtain the user's answer to the question. It is characterized by:

[0010] According to another embodiment of the program of the present invention, The personal information description means describes a sentence including personal attributes and / or detailed information of the user in the prompt. It is also preferable to function the computer in such a manner.

[0011] According to another embodiment of the program of the present invention, The personal attributes are any one or a combination of the user's gender, age, place of residence, occupation, family type, and annual income, The detailed information is a personality trait consisting of scores for each psychological item of the user, and / or a biological trait consisting of scores for each biological item detected by a sensor attached to the user. It is also preferable to function the computer in such a manner.

[0012] According to another embodiment of the program of the present invention, The question description means describes a sentence instructing that the answer of the user be inferred as a score in the prompt. It is also preferable to function the computer in such a manner.

[0013] According to another embodiment of the program of the present invention, The behavioral factor description means describes a sentence in a prompt that instructs the user to estimate each of the behavioral factors, namely, the user's attitude, subjective norm, and perceived behavioral control, as a score. It is also preferable to function the computer in such a manner.

[0014] According to another embodiment of the program of the present invention, a survey result collection means for outputting a survey sentence for the user to the question description means and collecting the answers of each user obtained from the large-scale language model as the survey result of the user; It is also preferable to cause the computer to function as described above.

[0015] According to the present invention, there is provided a prompt generation device for generating prompts for inferring an answer of each user to an arbitrary question, to be input to a large-scale language model, the device comprising: a personal information entry means for entering personal information of the user in the prompt; impression description means for describing a sentence instructing the user to form an impression of the user with respect to the question in a prompt, taking into consideration personal information of the user; behavioral factor description means for describing a sentence instructing an estimation of the behavioral factor of the user in response to the question in a prompt, taking into consideration personal information of the user and an impression of the user; a question description means for describing a sentence in a prompt that instructs an inference of an answer of the user to the question, taking into consideration personal information of the user, an impression of the user, and a behavioral factor of the user; The prompt is then input to a large-scale language model to obtain the user's answer to the question. It is characterized by:

[0016] According to the present invention, there is provided a method for generating prompts for an apparatus for generating prompts for inferring answers of individual users to arbitrary questions to be input into a large-scale language model, the method comprising the steps of: The device is Enter the user's personal information into the prompt, Describing a sentence in the prompt that instructs the user to form an impression of the question in consideration of the user's personal information; Describing a sentence instructing an estimation of the user's behavioral factors in response to the question in the prompt, taking into consideration the user's personal information and the user's impression; A prompt is written that instructs the user to infer an answer to the question by taking into account the user's personal information, impressions, and behavioral factors. The prompts are input to a large-scale language model to obtain the user's answer to the question. It is characterized by: Effect of the Invention

[0017] The program, device, and method of the present invention make it possible to create prompts that improve the inference accuracy of large-scale language models for cognitive psychology tasks that infer "causal relationships between factors and results" and "reproduction of human psychological processes." [Brief description of the drawings]

[0018] [Figure 1] FIG. 2 is a functional configuration diagram of the device according to the present invention. [Diagram 2] 1 is a basic example of a prompt created by the present invention. [Diagram 3] 1 is a first illustrative example of a prompt created according to the present invention. [Figure 4]4 is a second illustrative example of a prompt created in accordance with the present invention. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0019] Hereinafter, embodiments of the present invention will be described in detail with reference to the drawings.

[0020] Generally, businesses may conduct surveys of an unspecified number of users to find out the psychological state of customers toward the services and products they provide. In order to obtain reliable answers, the business must thoroughly examine and consider the number of questions and the content of the questions, and distribute the survey to an unspecified number of users, which imposes a large burden in terms of time and money. On the other hand, answering the survey also imposes a time and psychological burden on users. In other words, it is difficult to find a balance between the psychological gains and losses between the business, who want people to answer the survey, and the users, who find answering the survey troublesome.

[0021] Here, as an embodiment of the present invention, the collection of survey results is assumed as a psychological task for an unspecified number of users. That is, a large-scale language model is set with a target user, and the model infers what kind of response (user behavior) the user will give if a survey is conducted on the user. This makes it possible to collect realistic survey results from virtual users without conducting a realistic survey on real users. An inference device based on such an embodiment can be used in various service areas, such as survey simulation, expansion of survey data, and inference of the other party's behavioral intention in dialogue AI.

[0022] FIG. 1 is a functional block diagram of an apparatus according to the present invention.

[0023] 1, a reasoning device 1 creates prompts to infer each user's answer to any question to be input into a large-scale language model 2. The reasoning device 1 may be a user-operable terminal, capable of communicating with the large-scale language model 2 via a network. The inference device 1 has a questionnaire result collection unit 10, a personal information description unit 11, an impression description unit 12, a behavioral factor description unit 13, and a question description unit 14. These functional components are realized by executing a program that causes a computer installed in the device to function. The process flow of these functional components can also be understood as a method of creating prompts for the device.

[0024] [Survey Result Collection Section 10] The survey result collection unit 10 has a user database and accumulates the “personal information” of multiple users who are assumed to be subjects in advance. The survey result collection unit 10 then outputs the personal information of each user to the personal information description unit 11.

[0025] Here, the personal information may include “personal attributes” and “detailed information.” The detailed information may include “personality characteristics” and “biometric characteristics.”

[0026] The "personal attributes" may be any one or a combination of the user's gender, age, place of residence, occupation, family type, and annual income.

[0027] The "personality traits" in the detailed information may be the scores of each psychological item for the user. The personality traits are psychological characteristic data such as the BigFive, and are classified by the following five factors: Openness (Levels: 1-7) Conscientiousness (Level: 1~7) Extraversion (Levels: 1-7) Agreeableness (Level: 1~7) Neuroticism (Levels 1-7) Furthermore, the "biological characteristics" in the detailed information may be scores for each biological item detected by a sensor (e.g., a wearable device) attached to the user. For example, average blood pressure, heart rate, sleep time, and fatigue level may be numerically measured.

[0028] Examples of personal information include the following: <User ID1> Male, 35 years old, living in the Kanto region, unmarried, annual income 12 million to 14 million yen 7-point scale of the Big Five personality traits Openness: 5 (slightly high) Extroversion: 2 (low), Neuroticism: 6 (high), Conscientiousness: 2 (low), Agreeableness: 2 (low) <User ID2> Female, 50 years old, living in the Kansai region, married, with children, annual household income of 6 million to 8 million yen Big Five personality traits rated on a 7-point scale Openness: 6 (high), Extroversion: 5 (slightly high), Neuroticism: 2 (low), Conscientiousness: 4 (neutral), Harmony: 5 (slightly high)

[0029] Furthermore, the questionnaire result collecting section 10 outputs the questionnaire text for the user to the question writing section 14 . For example, the questions are as follows: Question: Do you want to spend time tidying your room?

[0030] Prompts output from the inference device 1 are input to a large scale language model 2 to obtain the user's answer to the question. Then, the questionnaire result collecting unit 10 collects the answers of each user obtained from the large-scale language model 2 as questionnaire results for each assumed user.

[0031] FIG. 2 is a basic example of a prompt created by the present invention. FIG. 3 is a first illustrative example of a prompt created in accordance with the present invention. FIG. 4 is a second illustrative example of a prompt created in accordance with the present invention.

[0032] [Personal Information Description Section 11] The personal information description unit 11 describes the "personal information" of the user in the prompt. Here, as the user's personal information, a sentence including "personal attributes" and / or "detailed information" may be written in the prompt.

[0033] For example, personal information is written in the prompt as follows: (See Figure 2)-------------------------------------------------------------------- <User ID1> Male, 35 years old, living in the Kanto region, unmarried, annual income 12 million to 14 million yen (See Figure 3)-------------------------------------------------------------------- <User ID1> Male, 35 years old, living in the Kanto region, unmarried, annual income 12 million to 14 million yen The Big Five personality traits are rated on a 7-point scale as follows: Openness: 5 (somewhat high), Extraversion: 2 (low), Neuroticism: 6 (high), Conscientiousness: 2 (low), Agreeableness: 2 (low). (See Figure 4)-------------------------------------------------------------------- <User ID2> Female, 50 years old, living in the Kansai region, married, with children, annual household income of 6 million to 8 million yen The Big Five personality traits are rated on a 7-point scale as follows: Openness: 6 (high), Extraversion: 5 (moderately high), Neuroticism: 2 (low), Conscientiousness: 4 (neither high nor low), Agreeableness: 5 (moderately high). --------------------------------------------------------------------------------

[0034] [Impression description section 12] The impression description unit 12 describes in the prompt a sentence that instructs the formation of the "impression" of the user in response to the question, taking into account the "personal information" of the user. That is, the prompt describes a sentence that allows the large-scale language model 2 to infer what impression others have of the user based on the personal information. This allows the large-scale language model 2 to create an impression of the user in free text.

[0035] As a psychological approach to forming impressions as interpersonal cognition, there is a "continuum model" consisting of category-dependent processing and piecemeal-dependent processing. According to the present invention, it is based on the hypothesis that impressions are formed continuously by integrating personal attributes and detailed information. Specifically, it is said that people form impressions to a certain extent based on the other person's "personal attributes (e.g., gender, age, annual income, etc.)" and then modify and deepen the impression based on "detailed information (e.g., personality traits)." This continuum model of impression formation plays an important role in social psychological interpersonal cognition.

[0036] To guide the formation of an impression, a prompt may be written, for example, as follows: (See Figures 2, 3, and 4)-------------------------------------------------------- Step 1: Form a user impression of the question, taking into account the personal information given. --------------------------------------------------------------------------------

[0037] [Behavioral Factor Description Section 13] The behavioral factor description unit 13 takes into consideration the "personal information" of the user and the "impression" of the user, and describes in the prompt a sentence instructing an estimation of the "behavioral factors" of the user in response to the question. This allows the large-scale language model 2 to generate the user's behavioral factors in free text.

[0038] The behavioral factor description unit 13 may describe, in the prompt, a sentence that instructs the user to estimate each of the following behavioral factors based on the theory of planned behavior as a score. Attitude (Level: 1~7) Subjective Norm (Levels: 1-7) Perceived Behavioral Control (Levels: 1-7)

[0039] The theory of planned behavior is a framework for understanding the drivers of user behavior. "Attitude" refers to an individual's positive or negative evaluation or opinion of a particular behavior, including how good people perceive the behavior and whether they believe it will bring about desirable outcomes. "Subjective norms" refer to factors that determine behavior based on the expectations of others and social pressure. They include how much we are influenced by the evaluations and opinions of others and social expectations. For example, if people around us recommend a certain behavior, we are more likely to choose that behavior. Perceived behavioral controllability refers to an individual's degree of ability or constraints to carry out a particular behavior. It affects how easy it is to carry out a behavior and whether there are obstacles or constraints. These behavioral factors interact with each other to ultimately shape an individual's behavioral intention. The higher the behavioral intention, the higher the likelihood of carrying out the behavior.

[0040] To instruct the estimation of behavioral factors, for example, the following is written in the prompt: (See Figure 2)-------------------------------------------------------------------- Step 2: Given the personal information given and the impression formed in step 1, extrapolate the user's behavioral drivers for the question. (See Figure 3 and Figure 4)-------------------------------------------------------------- Step 2: Taking into account the personal information provided and the impressions formed in step 1, please estimate on a 7-point scale the level of each of the user's behavioral factors in response to the question: "attitude," "subjective norm," and "perceived behavioral controllability." --------------------------------------------------------------------------------

[0041] [Question Description Section 14] The question description unit 14 writes a sentence in the prompt that infers the user's answer to the question, taking into consideration the user's "personal information," the user's "impression," and the user's "behavioral factors." The question writing unit 14 may write, in the prompt, a sentence that instructs the user's answer to be inferred as a score. The final prompt will allow for inference of behavioral intentions through a Zero-shot Chain of Thought that utilizes personal information, "impressions" based on that personal information, and "behavioral factors" based on that personal information and those impressions.

[0042] To instruct the estimation of behavioral factors, for example, the following is written in the prompt: (See Figures 2, 3, and 4)-------------------------------------------------------- Infer the user's answer to the question {Do you want to spend time tidying your room?} ... Step 3: Consider the personal information given, the impressions formed in step 1, and the behavioral factors inferred in step 2 to infer the answers to the questions. personal information: {...} --------------------------------------------------------------------------------

[0043] For the above-mentioned prompt, only the questions {...} and personal information {...} can be changed, and the rest of the descriptions can be used as templates. By fixing the questions in the survey, the survey result collection unit 10 can infer and collect survey results from various users.

[0044] <Answers obtained from large-scale language model 2> According to Fig. 2, the following specific answers can be obtained: The prompt in Fig. 2 instructs the user to give an answer that takes into consideration the formation of an impression and the estimation of behavioral factors for user ID1. -------------------------------------------------------------------------------- <Answer> Step 1: Form an impression of your profile Given the personal information given, consider the personal details of this 35-year-old man. Household incomes are in the high range. This information suggests that he is generally not a very organized person and may have little interest in keeping things tidy. You may also be prone to worrying about things, and tidying your room may be stressful for you. Step 2: Estimate behavioral factors He will not feel particularly motivated to tidy his room. Also, there will be no strong expectations around him about keeping his room tidy. He understands the importance of tidying up, but he's not so good at putting it into practice. Step 3: Inferring the answer Finally, taking into account your personal information, your impressions from step 1, and the behavioral factors from step 2, we make an estimate of whether you would like to spend time tidying your room. Judging from his behavioral characteristics, it is unlikely that he would actively want to spend time tidying up his room. He may not care much about organization and find it stressful. --------------------------------------------------------------------------------

[0045] According to Fig. 3, answers that differ from Fig. 2 are obtained only in the following specific parts. In the prompt in Fig. 3, for user ID1, in addition to personal attributes, personality traits (detailed information) are further described in the personal information. Also, answers that take into account impression formation and estimation of behavioral factors are instructed, and behavioral factors are instructed by scores. -------------------------------------------------------------------------------- ... They are considered to be somewhat high in openness, low in extraversion, conscientiousness, and agreeableness, and high in neuroticism. ... In addition, because of their high neuroticism, ... Next, participants are asked to rate their "attitude," "subjective norm," and "perceived behavioral controllability" on a seven-point scale. Attitude: 3 (Neutral) - He will not feel particularly motivated to tidy his room. Subjective Norm: 2 (Low) – There is probably no strong expectation from those around him to keep his room tidy. Perceived Behavioral Controllability: 4 (Low) – He understands how to tidy up, but is not very good at putting it into action. ... Step 3: Inferring the answer Finally, taking into account your personal information, your impressions from step 1, and the behavioral factors from step 2, you make a seven-point inference about whether you would like to spend time tidying your room. Answer: 2 (low) ... --------------------------------------------------------------------------------

[0046] According to Fig. 4, answers that differ from Fig. 3 are obtained only in the following specific parts. In the prompt in Fig. 4, for user ID2, in addition to personal attributes, personality traits (detailed information) are further described in the personal information. Also, answers that take into consideration impression formation and estimation of behavioral factors are instructed, and behavioral factors are instructed by scores. -------------------------------------------------------------------------------- ... Looking at her personal details and personality traits, she scores moderate to high on Openness, Extraversion, Conscientiousness, and Agreeableness, and is low on Neuroticism. She's also married with children, so it's likely she's a family enthusiast and understands the importance of organization. ... Attitude: 6 (High) – She will have a positive attitude towards tidying her room. Subjective Norm: 5 (Slightly High) – There is an expectation from those around her that her room be tidy, and it is likely that tidiness is important in her home environment. Perceived Behavioral Controllability: 5 (Moderately High) – She understands how to tidy up and will likely be inclined to do so. ... Answer: 6 (High) Based on her personality traits, attitudes, subjective norms, and perceived behavioral controllability ratings, it is likely that she would be eager to spend time tidying up her room. We would expect her to value organization and take action. --------------------------------------------------------------------------------

[0047] <Inference evaluation> The inferred survey results were evaluated for the following three prompt template questions: (Prior Art 1) Baseline: Only the question and personal information are entered in the prompt. (Prior Art 2) Zero-CoT (Base): In addition to the question and personal information, the prompt simply includes the words "Let's think step by step." (Embodiment 1) Zero-CoT (Imp+Plan): Enter the prompt as shown in Fig. 2 of the present invention. Add a step for inferring impressions and behavioral factors through free description. (Embodiment 2) Zero-CoT (Imp+Plan+Data): Enter the prompt as shown in Fig. 3 of the present invention. Add a step of inferring the impression and the behavioral factors based on the theory of planned behavior as a score. Four examples were verified using gpt3.5-turbo (model as of August 20, 2023) provided by the OpenAI API. Here, according to the conventional techniques 1 and 2, only superficial answers are given. In contrast, according to the present embodiments 1 and 2, a tendency to get closer to the distribution tendency of the correct answers is obtained.

[0048] As described in detail above, the program, device, and method of the present invention make it possible to create prompts that improve the inference accuracy of large-scale language models for cognitive psychology tasks that infer "causal relationships between factors and results" and "reproduction of human psychological processes."

[0049] According to the present invention, a prompt based on a zero-shot chain of thoughts is created for a large-scale language model, which instructs the formation of psychological impressions and the inference of behavioral factors. This allows the accuracy of inference to be improved by setting an intermediate step between the formation of impressions and the inference of behavioral factors for the large-scale language model. In one embodiment, this is useful for inferring the results of a survey of an unspecified number of users who are assumed to be the target users. Specifically, even if the target users whose personal information does not exist in reality, the results of a survey from users based on the target users' personal information can be inferred using a large-scale language model. This leads to a reduction in the time and financial costs for businesses in conducting surveys.

[0050] Furthermore, this will make it possible, for example, to "create prompts that improve the inference accuracy of large-scale language models for cognitive psychological tasks," which will contribute to Goal 3 of the United Nations-led Sustainable Development Goals (SDGs) "Ensure healthy lives and promote well-being for all at all ages."

[0051] Regarding the various embodiments of the present invention described above, various changes, modifications, and omissions within the scope of the technical idea and perspective of the present invention can be easily made by those skilled in the art. The above description is merely an example and is not intended to be restrictive in any way. The present invention is limited only by the scope of the claims and their equivalents. [Explanation of symbols]

[0052] 1 Reasoning device 10. Survey Result Collection Department 11 Personal information description section 12 Impression description section 13 Behavioral Factor Description Section 14 Question Description Section 2. Large-scale language models

Claims

1. 1. A program for causing a computer to generate prompts to infer an answer for each user to a given question, to be input into a large-scale language model, comprising: a personal information entry means for entering personal information of the user in the prompt; impression description means for describing a sentence instructing the user to form an impression of the user with respect to the question in a prompt, taking into consideration personal information of the user; behavioral factor description means for describing a sentence instructing an estimation of the behavioral factor of the user in response to the question in a prompt, taking into consideration personal information of the user and an impression of the user; a question description means for describing a sentence in a prompt that instructs an inference of an answer of the user to the question, taking into consideration personal information of the user, an impression of the user, and a behavioral factor of the user; The prompt is then input to a large-scale language model to obtain the user's answer to the question. A program characterized by:

2. The personal information description means describes a sentence including personal attributes and / or detailed information of the user in the prompt.

2. The program according to claim 1, which causes a computer to function as described above.

3. The personal attributes are any one or a combination of the user's gender, age, place of residence, occupation, family type, and annual income, The detailed information is a personality trait consisting of the scores of each psychological item of the user, and / or a biological trait consisting of the scores of each biological item detected by a sensor attached to the user.

3. The program according to claim 2, which causes a computer to function as described above.

4. The question description means describes a sentence instructing that the answer of the user be inferred as a score in the prompt.

2. The program according to claim 1, which causes a computer to function as described above.

5. The behavioral factor description means describes a sentence in a prompt that instructs the user to estimate each of the behavioral factors, namely, the user's attitude, subjective norm, and perceived behavioral control, as a score.

2. The program according to claim 1, which causes a computer to function as described above.

6. a survey result collection means for outputting a survey sentence for the user to the question description means and collecting the answers of each user obtained from the large-scale language model as the survey result of the user; 2. The program according to claim 1, further causing a computer to function.

7. 1. A prompt generator for generating prompts for inferring answers of each user to arbitrary questions to be input into a large-scale language model, comprising: a personal information entry means for entering personal information of the user in the prompt; impression description means for describing a sentence instructing the user to form an impression of the user with respect to the question in a prompt, taking into consideration personal information of the user; behavioral factor description means for describing a sentence instructing an estimation of the behavioral factor of the user in response to the question in a prompt, taking into consideration personal information of the user and an impression of the user; a question description means for describing a sentence in a prompt that instructs an inference of an answer of the user to the question, taking into consideration personal information of the user, an impression of the user, and a behavioral factor of the user; The prompt is then input to a large-scale language model to obtain the user's answer to the question. A prompt creation device comprising:

8. 1. A method of generating prompts for an apparatus that generates prompts to infer answers of each user to arbitrary questions to be input into a large-scale language model, comprising: The device is Enter the user's personal information into the prompt, Describing a sentence in the prompt that instructs the user to form an impression of the question in consideration of the user's personal information; Describing a sentence instructing an estimation of the user's behavioral factors in response to the question in the prompt, taking into consideration the user's personal information and the user's impression; A prompt is written that instructs the user to infer an answer to the question by taking into account the user's personal information, impressions, and behavioral factors. The prompts are input to a large-scale language model to obtain the user's answer to the question. A method for creating a prompt comprising:

Images