Trained model generation method, trained model generation device, program, and recording medium

Abstract

Provided is a method for creating an LLM to which individuality is imparted.　A trained model generation method according to the present disclosure comprises an information acquisition step, an investigation step, an extraction step, and a retraining step. In the information acquisition step, user-provided information including a keyword selected by a user is acquired. In the investigation step, investigation based on the keyword is performed by using a large language model. In the extraction step, retraining information is extracted from the investigated result acquired by the investigation. In the retraining step, retraining the large language model is performed on the basis of the retraining information. Each step is performed by a computer.

Description

Trained Model Generation Method, Trained Model Generation Apparatus, Program, and Recording Medium

[0001] The present disclosure relates to a trained model generation method, a trained model generation apparatus, a program, and a recording medium.

[0002] Patent Document 1 discloses an information processing apparatus suitable for generating summary sentences using a large language model.

[0003] Japanese Unexamined Patent Publication No. 2023-073095

[0004] Generally, large language models (LLM) as in Patent Document 1 are specialized for functional responses and do not exhibit personality in interaction with users. If an LLM also has personality, the LLM can become an existence that users can feel affection for, allowing users to enjoy interaction with the LLM.

[0005] Accordingly, an object of the present disclosure is to provide a trained model generation method, a trained model generation apparatus, a program, and a recording medium for generating an LLM to which personality is imparted.

[0006] In order to achieve the above object, the trained model generation method of the present disclosure includes an information acquisition step, an investigation step, an extraction step, and a retraining step, wherein the information acquisition step acquires user-provided information including a keyword selected by a user, the investigation step performs an investigation based on the keyword using a large language model, the extraction step extracts information for retraining from an investigation result obtained by the investigation, and the retraining step performs retraining of the large language model based on the information for retraining, and each step is executed by a computer.

[0007] The trained model generation device of this disclosure includes an information acquisition unit, a research unit, an extraction unit, and a retraining unit, wherein the information acquisition unit acquires user-provided information including keywords selected by the user; the research unit conducts research based on the keywords using a large-scale language model; the extraction unit extracts retraining information from the research results obtained by the research; and the retraining unit retrains the large-scale language model based on the retraining information.

[0008] The pre-trained model generation program of this disclosure includes an information acquisition procedure, a survey procedure, an extraction procedure, and a retraining procedure, wherein the information acquisition procedure acquires user-provided information including keywords selected by the user; the survey procedure performs a survey based on the keywords using a large-scale language model; the extraction procedure extracts retraining information from the survey results obtained from the survey; and the retraining procedure retrains the large-scale language model based on the retraining information. The program is designed to cause a computer to execute each of these procedures.

[0009] The recording medium of this disclosure is a computer-readable recording medium on which the program of this disclosure is recorded.

[0010] According to this disclosure, it is possible to provide a method for generating trained models, a trained model generation apparatus, a program, and a recording medium for creating LLMs with individuality.

[0011] Figure 1 is a block diagram showing the configuration of an example of the pre-trained model generation device of the present disclosure. Figure 2 is a block diagram showing an example of the hardware configuration of the pre-trained model generation device of the present disclosure. Figure 3 is a flowchart showing an example of the process in the pre-trained model generation method of the present disclosure. Figure 4 is a block diagram showing the configuration of another example of the pre-trained model generation device of the present disclosure. Figure 5 is a flowchart showing another example of the process in the pre-trained model generation method of the present disclosure. Figure 6A is a schematic diagram showing an example of a method for retraining a large-scale language model using the pre-trained model generation method of the present disclosure. Figure 6B is a schematic diagram showing another example of a method for retraining a large-scale language model using the pre-trained model generation method of the present disclosure.

[0012] Embodiments of this disclosure will now be described. However, this disclosure is not limited to the embodiments described below. In the following figures, the same parts are denoted by the same reference numerals. Furthermore, unless otherwise specified, the descriptions of each embodiment can be used interchangeably. Furthermore, unless otherwise specified, the configurations of each embodiment can be combined. Also, in the programs of this disclosure described later, the term "procedure" can be read as "process," for example.

[0013] [Embodiment 1] The trained model generation apparatus, trained model generation method, and trained model generation program of the present disclosure will be described.

[0014] An example of the trained model generation device described herein will be explained with reference to Figures 1 and 2.

[0015] Figure 1 is a block diagram showing the configuration of an example of the trained model generation device 10 (the device 10) of the present disclosure. As shown in Figure 1, the device 10 includes an information acquisition unit 11, a research unit 12, an extraction unit 13, and a retraining unit 14.

[0016] The device 10 may be, for example, a single device including the aforementioned parts, or it may be a system device in which each of the aforementioned parts is a separate device that can be connected via a communication network. The device 10 can also be connected to an external device, which will be described later, via the communication network. The communication network is not particularly limited and can be a known network, for example, wired or wireless. Examples of the communication network include the Internet, WWW (World Wide Web), telephone lines, LAN (Local Area Network), SAN (Storage Area Network), DTN (Delay Tolerant Networking), LPWA (Low Power Wide Area), L5G (Local 5G), etc. Examples of wireless communication include Wi-Fi (registered trademark), Bluetooth (registered trademark), Local 5G, LPWA, etc. The wireless communication may be in the form of direct communication between devices (Ad Hoc communication), infrastructure communication, indirect communication via an access point, etc. The device 10 may be, for example, incorporated into a server as a system. Furthermore, the device 10 may be, for example, a personal computer (PC, e.g., desktop or notebook type), a smartphone, a tablet terminal, or digital signage on which the program disclosed herein is installed. The device 10 may also be in the form of cloud computing or edge computing, for example, in which at least one of the aforementioned parts is on a server and the other aforementioned parts are on a terminal.

[0017] Figure 2 illustrates a block diagram of the hardware configuration of the device 10. The device 10 includes, for example, a central processing unit (CPU, GPU, etc.) 101, memory 102, bus 103, storage device 104, input device 105, output device 106, communication device 107, etc. Each part of the device 10 is interconnected via the bus 103 through its respective interface (I / F).

[0018] The central processing unit 101 operates in coordination with other components via controllers (system controller, I / O controller, etc.) and is responsible for the overall control of the device 10. In the device 10, the central processing unit 101 executes, for example, the program disclosed herein and other programs, and also reads and writes various types of information. Specifically, for example, the central processing unit 101 functions as an information acquisition unit 11, an investigation unit 12, an extraction unit 13, and a relearning unit 14. The central processing unit 101 may be equipped with arithmetic units such as a CPU, GPU (Graphics Processing Unit), APU (Accelerated Processing Unit), NPU (Neural Processing Unit), or a combination thereof.

[0019] Bus 103 can also be connected to external devices, for example. Examples of such external devices include external storage devices such as external databases, printers, external input devices, external display devices, and external imaging devices. The device 10 can be connected to an external network (the aforementioned communication network) by a communication device 107 connected to bus 103, for example, and can also be connected to other devices via the external network.

[0020] Memory 102 may be, for example, main memory. When the central processing unit 101 performs processing, memory 102 reads various operational programs, such as the program of this disclosure, stored in the storage device 104 (described later), and the central processing unit 101 receives data from memory 102 and executes the program. The main memory may be, for example, RAM (random access memory). Alternatively, memory 102 may be, for example, ROM (read-only memory).

[0021] The storage device 104 is also called an auxiliary storage device, for example, in relation to the main memory (primary memory). As described above, the storage device 104 stores an operating program including the program of this disclosure. The storage device 104 may be, for example, a combination of a recording medium and a drive for reading and writing to the recording medium. The recording medium is not particularly limited and may be internal or external, for example, an HD (hard disk), CD-ROM, CD-R, CD-RW, MO, DVD, flash memory, memory card, etc. The storage device 104 may be, for example, a hard disk drive (HDD) in which the recording medium and the drive are integrated, or a solid state drive (SSD).

[0022] In this device 10, the memory 102 and storage device 104 can also store various types of information, such as log information, information obtained from an external database (not shown) or external devices, information generated by this device 10, and information used by this device 10 when executing processing. In this case, the memory 102 and storage device 104 may store, for example, user-provided information, survey results, and relearning information, which will be described later. At least some of the information may be stored on an external server other than the memory 102 and storage device 104, or it may be stored in a distributed manner across multiple terminals using blockchain technology or the like.

[0023] The device 10 further includes, for example, an input device 105 and an output device 106. The input device 105 may include, for example, a pointing device such as a touch panel, trackpad, or mouse; a keyboard; imaging means such as a camera or scanner; a card reader such as an IC card reader or magnetic card reader; an audio input means such as a microphone; and so on. The output device 106 may include, for example, a display device such as an LED display or liquid crystal display; an audio output device such as a speaker; a printer; and so on. In this disclosure, the input device 105 and the output device 106 are configured separately, but the input device 105 and the output device 106 may be configured as an integrated unit, such as a touch panel display.

[0024] Next, an example of the pre-trained model generation method of this disclosure will be explained based on the flowchart in Figure 3. The pre-trained model generation method of this disclosure is carried out as follows, for example, using the apparatus 10 in Figure 1 or Figure 2. However, the pre-trained model generation method of this disclosure is not limited to the use of the apparatus 10 in Figure 1 or Figure 2. In the pre-trained model generation method of this disclosure, for example, each of the steps described later may be performed by a computer. Also, the pre-trained model generation method of this disclosure can also be called, for example, the pre-trained model manufacturing method of this disclosure.

[0025] The information acquisition unit 11 acquires user-provided information, including keywords selected by the user (S11, information acquisition step). The keywords are not particularly limited, but may, for example, be keywords related to the user's preferences, hobbies, expertise, or interests, or requests for the large-scale language model, or they may be keywords related to whether or not the user is interested in the preferences, hobbies, or interests. The keywords may be one or more.

[0026] The user-provided information may further include, for example, an interest index indicating the degree of the user's interest in or concern with the keyword. The interest index is, for example, an index expressed as a numerical value. The numerical value is, for example, a score. The interest index may be determined, for example, by the user.

[0027] The user-provided information may further include, for example, personality information assigned to the large-scale language model. This personality information may be, for example, personality information classified by a five-factor model. Here, the five-factor model is a psychological framework for understanding and classifying human personality, which is classified into five indicators: openness, conscientiousness, extraversion, agreeableness, and neuroticism. The five-factor model can also be called the Big Five. This personality information may be, for example, numerical personality information, or it may be personality information expressed relatively, such as "high," "medium," or "low."

[0028] The investigation unit 12 conducts an investigation based on the keywords using a large-scale language model (S12, investigation process). The large-scale language model is, for example, a large-scale language model that will be retrained as described later. The investigation means, for example, inputting the keywords into the large-scale language model and obtaining output results (investigation results described later). The investigation may be performed only once or multiple times.

[0029] The research unit 12, for example, conducts a research based on the keywords using a large-scale language model, and further conducts an additional research using the large-scale language model based on the information obtained from the research results obtained in the first research. The additional research means, for example, inputting the research results into the large-scale language model and obtaining output results. The information obtained from the research results may be, for example, keywords obtained from the research results. The keywords obtained from the research results may be extracted from the research results by, for example, a known method.

[0030] The aforementioned additional investigation may, for example, involve investigating the remaining keywords by linking the investigation results related to any keyword selected from among the multiple keywords, if there are multiple keywords. For example, if there are keywords "A, B, C", the investigation result "a" obtained by investigating "A" would be linked to the remaining "B, C" (for example, "B of a", "C for a", etc.) and investigated.

[0031] The research unit 12 conducts the additional research on a scale determined, for example, based on the interest index. The scale may be calculated based on the interest index. Alternatively, the scale for the interest index may be defined in advance, and the additional research may be conducted based on the definition. Examples of the scale include frequency, number, volume, range, time, amount of information, hierarchy, and depth.

[0032] The research unit 12 may, for example, use a large-scale language model to conduct research based on the keywords at any frequency. The frequency is not particularly limited, but may be, for example, daily, weekly, monthly, seasonally, or yearly, or any frequency set by the user.

[0033] The extraction unit 13 extracts information for retraining from the results obtained from the investigation (S13, extraction step). The extraction unit 13 extracts information for retraining from the results obtained from, for example, the investigation and at least one of the additional investigation. The information for retraining may be, for example, keywords obtained from the investigation results. The keywords obtained from the investigation results may be extracted from the investigation results by, for example, a known method.

[0034] The extraction unit 13 may, for example, normalize the retraining information extracted from the survey results into a format usable for retraining. The extraction unit 13 may, for example, generate the retraining information as a machine-readable record based on the survey results (e.g., output of a large-scale language model, information obtained from an external device, log information, or database information).

[0035] The machine-readable record may include, for example, at least one of the following: input information, expected output information, evidence information (e.g., source or reference), confidence information, timestamp information, and target keyword information. The input information may be, for example, the keyword, or a sentence or question related to the keyword. The expected output information may be, for example, a response, summary, classification result, or generation template that the large-scale language model should output in response to the input information.

[0036] The extraction unit 13 may, for example, store the machine-readable records in a storage area such as memory 102 or storage device 104. Such storage may include, for example, accumulating multiple machine-readable records and managing them as a dataset.

[0037] The extraction unit 13 may, for example, perform verification processing on the machine-readable records. The extraction unit 13 may, for example, perform duplicate removal processing, inconsistency detection processing, formal consistency checking processing, or missing item completion processing. The inconsistency detection processing may, for example, include processing to determine whether there are multiple inconsistent expected output pieces of information for the same or similar input information.

[0038] The extraction unit 13 may, for example, perform processing to reduce confidential or personal information that may be included in the machine-readable record. The extraction unit 13 may, for example, perform processing to mask, anonymize, or delete descriptions that may constitute confidential or personal information based on a predetermined policy. Furthermore, the extraction unit 13 may, for example, perform filtering processing to detect prohibited words or inappropriate expressions, and processing to exclude all or part of the machine-readable record based on said filtering processing.

[0039] The extraction unit 13 may, for example, calculate an index representing quality (e.g., reliability, consistency, or source accuracy) for each machine-readable record. The extraction unit 13 may, for example, exclude machine-readable records whose index falls below a predetermined threshold from the candidates for retraining. This ensures, for example, the quality of the data used for retraining.

[0040] The relearning unit 14 retrains the large-scale language model based on the relearning information (S14, relearning step). The relearning unit 14 retrains the large-scale language model based, for example, on the relearning information and the personality information. The relearning may be performed, for example, by a known method.

[0041] The relearning unit 14 may, for example, generate a training dataset for relearning based on the machine-readable records stored in the memory area. The relearning unit 14 may, for example, generate pairs of inputs and teacher signals (expected output information) from the machine-readable records and configure these as a training dataset. The relearning unit 14 may, for example, use the training dataset in combination with the personality information.

[0042] The relearning performed by the relearning unit 14 is not particularly limited in terms of update targets. For example, the relearning unit 14 may update all or part of the parameters of the large language model, or may update additional parameters (e.g., adapters, additional layers, low-rank parameters, etc.). Furthermore, the relearning unit 14 may perform an update equivalent to relearning, for example, by updating a prompt template used for input to the large language model, an external search index referenced by the large language model, or a reference database.

[0043] The relearning unit 14 may, for example, evaluate the updated large language model after relearning. For example, the relearning unit 14 may perform automatic evaluation using a predetermined test question set, detection of output deviation (e.g., deviation from a predetermined policy or constraint), or monitoring of waiting time or memory usage in inference processing.

[0044] For example, when the result of the evaluation does not satisfy a predetermined criterion, the relearning unit 14 may suspend the update or revert to the large language model before the update. The relearning unit 14 may, for example, assign version information to the updated large language model, and store the version information in a storage area in association with the evaluation result. Thereby, for example, traceability of update history can be improved.

[0045] The relearning unit 14 may, for example, deploy the updated large language model to an inference environment. The inference environment may include, for example, a server, a cloud, or an edge terminal. The relearning unit 14 may, for example, perform processing such as model quantization, compression, or divided arrangement according to the deployment destination.

[0046] The program of the present disclosure is a program for causing a computer to execute each step of the present disclosure as procedures. Specifically, the trained model generation program of the present disclosure is a program for causing a computer to execute an information acquisition procedure, an investigation procedure, an extraction procedure, and a retraining procedure. The trained model generation program of the present disclosure can also be described as a program that causes a computer to function as the information acquisition procedure, the investigation procedure, the extraction procedure, and the retraining procedure. For the trained model generation program of the present disclosure, the descriptions in the trained model generation apparatus and the trained model generation method of the present disclosure can be incorporated by reference.

[0047] According to the present disclosure, in the information acquisition step, user-provided information including a keyword selected by a user is acquired; in the investigation step, an investigation based on the keyword is performed using a large language model; in the extraction step, retraining information is extracted from an investigation result obtained by the investigation; and in the retraining step, retraining of the large language model is performed based on the retraining information. This enables creation of an LLM to which personality is imparted.

[0048] [Embodiment 2] Next, the trained model generation apparatus, trained model generation method, and trained model generation program of the present disclosure will be further described.

[0049] Another example of the trained model generation apparatus of the present disclosure will be described with reference to FIG. 4. FIG. 4 is a block diagram showing a configuration of an example of a trained model generation apparatus 10A (present apparatus 10A) of the present disclosure. As shown in FIG. 4, in addition to the configuration of the present apparatus 10, the present apparatus 10A further includes a generation unit 15 and an output unit 16.

[0050] In the present apparatus 10A, a central processing unit 101 functions as the information acquisition unit 11, an investigation unit 12, an extraction unit 13, a retraining unit 14, the generation unit 15, and the output unit 16. Other hardware configurations are the same as those of the present apparatus 10, except that the central processing unit in FIG. 2 further includes the generation unit 15 and the output unit 16.

[0051] Next, an example of the pre-trained model generation method of this disclosure will be explained based on the flowchart in Figure 5. The pre-trained model generation method of this disclosure is carried out as follows, for example, using the apparatus 10A in Figure 4. However, the pre-trained model generation method of this disclosure is not limited to the use of the apparatus 10A in Figure 4. In the pre-trained model generation method of this disclosure, for example, each of the steps described later may be performed by a computer.

[0052] First, the information acquisition unit 11, the investigation unit 12, the extraction unit 13, and the retraining unit 14 perform the same processing as described in S11 to S14 above.

[0053] Next, the generation unit 15 generates profile information of the large-scale language model based on the retraining information (S15, generation step). The profile information includes, for example, the preferences of the large-scale language model and at least one piece of information relating to the generation characteristics of the information generated by the large-scale language model. The preferences are, for example, but are not limited to, preferences related to food, music, movies, fashion, travel, reading, sports, art, culture, lifestyle, and hobbies. The generation characteristics are, for example, but are not limited to, characteristics relating to sociability, friendliness, empathy, consideration, listening ability, politeness, humility, politeness of expression, assertiveness, and tolerance. The information may be, for example, numerical, or it may be text information relating to the preferences and generation characteristics.

[0054] Furthermore, if the user-provided information includes, for example, personality information assigned to the large-scale language model, the generation unit 15 generates profile information for the large-scale language model based on the retraining information and the personality information. The profile information includes, for example, at least one piece of information about preferences, generation characteristics, and the personality of the large-scale language model. Examples of such personality traits include, but are not limited to, introverted, extroverted, optimistic, pessimistic, impulsive, planned, self-centered, dedicated, empathetic, humble, and arrogant.

[0055] The output unit 16 outputs the profile information, for example (S16, output step). The profile information may be output in the form of at least one of the following: text, diagrams, audio, and video. The profile information may also be output as a story, for example. The output may be output by the output device 106, for example, or by another device. By outputting the profile information from the output unit 16, the user can, for example, understand the preferences of the generated large-scale language model. Furthermore, by understanding the preferences of the large-scale language model, for example, the user's interaction with the large-scale language model can become more familiar and enjoyable.

[0056] The program of this disclosure is a program that causes a computer to perform a generation procedure and an output procedure. The trained model generation program of this disclosure can also be described as a program that causes a computer to function as a generation procedure and an output procedure. The trained model generation program of this disclosure can be made by reference to the descriptions of the trained model generation apparatus and trained model generation method of this disclosure.

[0057] [Embodiment 3] An example and another example of a method for retraining a large-scale language model using the pre-trained model generation method of the present disclosure will be explained with specific examples shown in Figures 6A and 6B. In this disclosure, examples using the device 10 or 10A are shown, but the invention is not limited thereto.

[0058] As shown in Figure 6A, the device 10 acquires user-provided information. User-provided information includes keywords selected by the user. Next, the device 10 uses a large-scale language model to conduct research based on the keywords and extracts information for retraining from the research results. At this time, based on the information obtained from the research results, further research may be conducted, and information for retraining may be extracted from the research results obtained from the additional research. Subsequently, the device 10 retrains the large-scale language model based on this information for retraining.

[0059] As shown in Figure 6B, this device 10A can generate and output profile information for a large-scale language model based on retraining information, etc.

[0060] The following shows a more specific example of a large-scale language model retrained using the device 10 or 10A.

[0061] <Assigning Personality Combining "Hobbies and Expertise"> First, we will explain an example of how to assign personality combining "hobbies and expertise" by retraining a large-scale language model.

[0062] The device 10 acquires keywords selected by the user, namely "wine," "food," and "travel," and further acquires a numerical representation of the degree of interest in each keyword as an interest index (for example, wine: 7 points, food: 5 points, travel: 3 points). Regarding "wine," the device 10 uses a large-scale language model to investigate grape varieties, regions, and tasting methods. Next, it uses the large-scale language model to investigate related keywords associated with "food" and "travel," such as "regional cuisine," "restaurant exploration," and "overseas wineries." Subsequently, the device 10 extracts information for retraining from the investigation results and retrains the large-scale language model using this information and personality information (for example, information that enhances "extroversion (prefers interaction at overseas wineries)" and "openness (shows interest in new cooking methods)" in the five-factor model).

[0063] Furthermore, according to this device 10A, profile information such as "Has experience as a sommelier and diligently collects information on local ingredients and restaurants when traveling. Prefers free thinking and has a personality that always wants to try new recipes" can be generated from the retraining information and personality information. This device 10A can present the profile information to the user.

[0064] Based on the above, a large-scale linguistic model is completed, for example, for "a cheerful, charismatic food connoisseur character who enjoys wine, food, and travel."

[0065] <Assigning a personality with a growth element of "learning from small mistakes"> Next, we will explain an example of how to assign a personality with a growth element of "learning from small mistakes" to a large-scale language model through retraining.

[0066] The device 10 acquires keywords selected by the user, namely "funny failure stories," "fearful," and "still willing to try," and further acquires numerical values ​​representing the degree of interest in each keyword as an interest index (for example, funny failure stories: 6 points, fearful: 4 points, trying: 8 points). The device 10 investigates "funny failure stories" and obtains patterns of real-life episodes and stories. Next, it links the obtained research results with "fearful" and conducts additional research using a large-scale language model, focusing on, for example, "hesitation to fail" and "overcoming anxiety." Furthermore, it extracts psychological information that reinforces the motivation to "still want to try." Subsequently, the device 10 extracts information for relearning from the research results and retrains the large-scale language model using this relearning information and personality information (for example, information that enhances "slightly high neuroticism (easily anxious)" and "high conscientiousness (learns from failures and strives to improve)" in the five-factor model).

[0067] Furthermore, according to this device 10A, profile information such as "He shares past failures with others, turning them into humor, and despite being timid, he never gives up on challenges. He constantly ventures into new areas and tends to fail, but learns and grows each time" can be generated from the relearning information and personality information. This device 10A can then present the profile information to the user.

[0068] Based on the above, a large-scale language model is completed, for example, representing a character who is "vulnerable to failure but perseveres with effort." This allows users to converse with a large-scale language model that includes "human-like" dialogue, such as stories of failure.

[0069] <Assigning Personality Based on Regional and Cultural Background> Next, we will explain an example of assigning personality based on regional and cultural background by retraining a large-scale language model.

[0070] The device 10 acquires keywords selected by the user, such as "Japanese traditional culture," "Indian spirituality," and "music festival," and further acquires a numerical representation of the degree of interest for each keyword as an interest index. The device 10 investigates the cultural background and researches related knowledge such as tea ceremony and yoga. Next, from the research results obtained, it extracts keywords related to music festivals and local event information (such as the origins of traditional instruments and local festivals) using a large-scale language model and conducts additional research. Subsequently, the device 10 extracts information for retraining from the research results and retrains the large-scale language model using the retraining information and personality information. For example, the personality information could be set to a high level for "openness" in the five-factor model, resulting in a type that is curious about both tradition and new experiences, or set to a moderate level for "extroversion," resulting in a type that quietly speaks of deep knowledge.

[0071] Furthermore, according to this device 10A, profile information such as "A traveler who, from a young age, has been interested in combining foreign art with Japanese traditions, and has eclectically learned about the spirit of tea and spiritual thinking while traveling to various music festivals" can be generated from the relearning information and personality information. This device 10A can present the profile information to the user.

[0072] Based on the above, a large-scale language model is completed that, for example, describes a character as having a "global perspective" with a cultural background, yet possessing a gentle personality.

[0073] <Imparting a dual personality to balance "expert advisor" and "everyday friend"> Next, we will explain an example of how to impart a dual personality to balance "expert advisor" and "everyday friend" by retraining a large-scale language model.

[0074] The device 10 acquires keywords selected by the user, such as the following, and also acquires a numerical score for the degree of interest in each of them as an interest index: • Specialized field (e.g., management consulting) ... Interest level: 8 points • Everyday hobbies (e.g., gardening, crafts) ... Interest level: 4 points • Approachability ... Interest level: 6 points

[0075] First, the device 10 investigates the user's area of ​​expertise (management consulting) and acquires existing business methods and success stories. Next, it connects the acquired research results with the user's everyday hobbies, such as gardening and crafting, and conducts additional research from the perspectives of relaxation effects and "balancing business and hobbies." Subsequently, the device 10 extracts information for retraining from the additional research results and retrains a large-scale language model using this information and personality information. The personality information may, for example, be set to enhance "conscientiousness" and "openness" in the five-factor model, while simultaneously setting "cooperativeness" to a medium to high level. As a result, the retrained large-scale language model is adjusted to provide precise professional advice while maintaining a friendly tone in casual conversation.

[0076] Furthermore, according to this device 10A, it is possible to generate profile information such as, "A consultant who advises companies on weekdays but is engrossed in gardening and craft workshops on weekends. When you talk to him, he will use difficult business jargon, but once the consultation is over, he will talk to you in a friendly manner about gardening." The device 10A can then present this profile information to the user.

[0077] As a result, users can, for example, enjoy both "rigorous business questions" and "casual, hobby-based conversations" with a single large-scale language model.

[0078] <Dynamic Personality Evolution Triggered by "Continuous Seasonal Events and Occasions"> Next, we will explain an example of dynamic personality evolution triggered by "continuous seasonal events and occasions" through the retraining of a large-scale language model.

[0079] The device 10 acquires seasonal events such as "Spring Festival," "Christmas," and "Halloween" as keywords, and further acquires a moderate numerical value (for example, 5 to 6 points) representing the degree of interest in each keyword as an interest index. Next, in the initial stage, the device 10 investigates seasonal customs and origins using a large-scale language model. From the results obtained from the above investigation, it extracts information for retraining, and periodically retrains the model based on this information during the period of each event. The device 10 also continuously investigates recipes, costumes, and history related to the events, extracts information for retraining from these investigation results, and retrains the large-scale language model. Subsequently, the device 10 retrains the large-scale language model using the retraining information and personality information. For example, the personality information is set so that "extroversion" and "openness" in the five-factor model are higher than usual in conjunction with the timing of each event. As a result, the retrained large-scale language model exhibits behavior that actively interacts with people enjoying festivals in conjunction with the timing of each event. On the other hand, when each event is out of season (off-season), the topics output by the large-scale language model naturally shift, returning to a somewhat calmer, "normal" state.

[0080] Furthermore, this device 10A can generate profile information such as, "A character who enjoys cherry blossom viewing in spring and gets excited about costumes as Halloween approaches. A 'somewhat festive' character who feels excitement and nostalgia with each season." from the relearning information and personality information. This device 10A can then present the profile information to the user.

[0081] As a result, for example, users can always enjoy "interacting with a large-scale language model that is constantly changing," which helps prevent boredom during long-term use.

[0082] [Embodiment 4] The program of the present disclosure may be recorded on, for example, a computer-readable storage medium. The storage medium is, for example, a non-transitory computer-readable storage medium. The storage medium is not particularly limited and includes, for example, random access memory (RAM), read-only memory (ROM), hard disk (HD), flash memory (e.g., USB flash memory, SD / SDHC card, etc.), optical disc (e.g., CD-R / CD-RW, DVD-R / DVD-RW, BD-R / BD-RE, etc.), magneto-optical disk (MO), floppy disk (FD), etc. The program of the present disclosure (for example, also referred to as a programming product or program product) may be delivered, for example, from an external computer. The "delivery" may be, for example, delivered via a communication network or delivered via a wired device. The program of the present disclosure may be installed and executed on the delivered device, or it may be executed without being installed.

[0083] Although the present disclosure has been described above with reference to embodiments, the present disclosure is not limited to the embodiments described above. The configuration and conditions of the present disclosure can be modified in various ways that can be understood by those skilled in the art within the scope of the present disclosure.

[0084] <Note> Some or all of the above embodiments may be described as follows, but are not limited to the following. <Method for generating a trained model> (Note 1) A method for generating a trained model, comprising an information acquisition step, a research step, an extraction step, and a retraining step, wherein the information acquisition step acquires user-provided information including keywords selected by the user, the research step conducts a research based on the keywords using a large-scale language model, the extraction step extracts retraining information from the research results obtained by the research, and the retraining step retrains the large-scale language model based on the retraining information, each step being executed by a computer. (Note 2) The method for generating a trained model according to Note 1, wherein the research step conducts a research based on the keywords using a large-scale language model, and further conducts an additional research based on information obtained from the research results obtained by the research using the large-scale language model, and the extraction step extracts retraining information from the research results obtained from at least one of the research and the additional research. (Note 3) The method for generating a trained model according to Note 2, wherein the user-provided information further includes an interest index indicating the degree of the user's interest or concern with the keyword, and the research step conducts the additional research on a scale determined based on the interest index. (Note 4) The method for generating a trained model according to Note 3, wherein the scale is the number of times. (Note 5) The method for generating a trained model according to any one of Notes 1 to 4, wherein the user-provided information further includes personality information to be assigned to the large-scale language model, and the retraining step retrains the large-scale language model based on the retraining information and the personality information. (Note 6) The method for generating a trained model according to Note 5, wherein the personality information is personality information classified by a five-factor model. (Appendix 7) A method for generating a trained model according to any one of Appendix 1 to 6, further comprising a generation step and an output step, wherein the generation step generates profile information of the large-scale language model based on the retraining information, the profile information includes at least one piece of information relating to the preferences of the large-scale language model and the generation characteristics relating to the information generated by the large-scale language model, and the output step outputs the profile information.(Note 8) The method for generating a trained model according to Note 7, wherein the user-provided information further includes personality information assigned to the large-scale language model, the generation step generates profile information of the large-scale language model based on the retraining information and the personality information, and the profile information includes at least one piece of information of preferences, generation characteristics, and the personality of the large-scale language model. <Trained Model Generation Apparatus> (Note 9) A trained model generation apparatus comprising an information acquisition unit, a research unit, an extraction unit, and a retraining unit, wherein the information acquisition unit acquires user-provided information including keywords selected by the user, the research unit conducts a research based on the keywords using the large-scale language model, the extraction unit extracts retraining information from the research results obtained by the research, and the retraining unit retrains the large-scale language model based on the retraining information. (Note 10) The trained model generation device according to Note 9, wherein the research unit conducts a research based on the keywords using a large-scale language model, and further conducts an additional research based on the information obtained from the research results obtained from the research using the large-scale language model, and the extraction unit extracts information for retraining from the research results obtained from at least one of the research and the additional research. (Note 11) The trained model generation device according to Note 10, wherein the user-provided information further includes an interest index indicating the degree of the user's interest or concern with the keywords, and the research unit conducts the additional research on a scale determined based on the interest index. (Note 12) The trained model generation device according to Note 11, wherein the scale is the number of times. (Note 13) The trained model generation device according to any one of Notes 9 to 12, wherein the user-provided information further includes personality information assigned to the large-scale language model, and the retraining unit retrains the large-scale language model based on the information for retraining and the personality information. (Note 14) The aforementioned personality information is personality information classified by a five-factor model, as described in Note 13, and is a trained model generation device.(Note 15) The trained model generation device according to any one of Notes 9 to 14, further comprising a generation unit and an output unit, wherein the generation unit generates profile information of the large-scale language model based on the retraining information, the profile information includes at least one piece of information on the preferences of the large-scale language model and the generation characteristics relating to the information generated by the large-scale language model, and the output unit outputs the profile information. (Note 16) The trained model generation device according to Note 15, further comprising, as user-provided information, personality information assigned to the large-scale language model, wherein the generation unit generates profile information of the large-scale language model based on the retraining information and the personality information, the profile information includes at least one piece of information on the preferences, the generation characteristics, and the personality of the large-scale language model. <Trained Model Generation Program> (Note 17) A trained model generation program that causes a computer to execute each of the following steps: information acquisition procedure, investigation procedure, extraction procedure, and retraining procedure, wherein the information acquisition procedure acquires user-provided information including keywords selected by the user; the investigation procedure conducts an investigation based on the keywords using a large-scale language model; the extraction procedure extracts retraining information from the investigation results obtained from the investigation; and the retraining procedure retrains the large-scale language model based on the retraining information. (Note 18) A trained model generation program as described in Note 17, wherein the investigation procedure conducts an investigation based on the keywords using a large-scale language model, and further conducts an additional investigation based on information obtained from the investigation results obtained from the investigation using the large-scale language model; and the extraction procedure extracts retraining information from the investigation results obtained from at least one of the investigation and the additional investigation. (Note 19) The trained model generation program described in Note 18, wherein the user-provided information further includes an interest index indicating the degree of the user's interest or concern with the keyword, and the research procedure performs the additional research on a scale determined based on the interest index. (Note 20) The trained model generation program described in Note 19, wherein the scale is the number of times.(Note 21) The trained model generation program according to any one of Notes 17 to 20, wherein the user-provided information further includes personality information assigned to the large-scale language model, and the retraining procedure retrains the large-scale language model based on the retraining information and the personality information. (Note 22) The trained model generation program according to Note 21, wherein the personality information is personality information classified by a five-factor model. (Note 23) The trained model generation program according to any one of Notes 17 to 22, further including a generation procedure and an output procedure, wherein the generation procedure generates profile information of the large-scale language model based on the retraining information, the profile information includes at least one piece of information relating to the preferences of the large-scale language model and the generation characteristics relating to the information generated by the large-scale language model, and the output procedure outputs the profile information. (Note 24) The trained model generation program described in Note 23, wherein the user-provided information further includes personality information assigned to the large-scale language model, the generation procedure generates profile information of the large-scale language model based on the retraining information and the personality information, and the profile information includes at least one piece of information of preferences, generation characteristics, and the personality of the large-scale language model. <Recording medium for trained model generation program> (Note 25) A computer-readable recording medium recording a trained model generation program that causes a computer to execute each of the following procedures: information acquisition procedure, investigation procedure, extraction procedure, and retraining procedure, wherein the information acquisition procedure acquires user-provided information including keywords selected by the user, the investigation procedure performs an investigation based on the keywords using the large-scale language model, the extraction procedure extracts retraining information from the investigation results obtained by the investigation, and the retraining procedure retrains the large-scale language model based on the retraining information.(Note 26) The recording medium according to Note 25, wherein the investigation procedure involves conducting an investigation based on the keywords using a large-scale language model, and further conducting an additional investigation based on information obtained from the investigation results obtained using the large-scale language model, and the extraction procedure involves extracting information for retraining from the investigation results obtained from at least one of the investigation and the additional investigation. (Note 27) The recording medium according to Note 26, wherein the user-provided information further includes an interest index indicating the degree of the user's interest or concern with the keywords, and the investigation procedure involves conducting the additional investigation on a scale determined based on the interest index. (Note 28) The recording medium according to Note 27, wherein the scale is the number of times. (Note 29) The recording medium according to any one of Notes 25 to 28, wherein the user-provided information further includes personality information assigned to the large-scale language model, and the retraining procedure involves retraining the large-scale language model based on the retraining information and the personality information. (Note 30) The recording medium according to Note 29, wherein the personality information is personality information classified by a five-factor model. (Note 31) The recording medium according to any one of Notes 25 to 30, further comprising a generation procedure and an output procedure, wherein the generation procedure generates profile information of the large-scale language model based on the retraining information, the profile information includes at least one piece of information on the preferences of the large-scale language model and the generation characteristics relating to the information generated by the large-scale language model, and the output procedure outputs the profile information. (Note 32) The recording medium according to Note 31, wherein the user-provided information further comprises personality information assigned to the large-scale language model, the generation procedure generates profile information of the large-scale language model based on the retraining information and the personality information, and the profile information includes at least one piece of information on the preferences, the generation characteristics, and the personality of the large-scale language model.

[0085] This disclosure makes it possible to provide a method for generating trained models, a trained model generating apparatus, a program, and a recording medium for creating LLMs with individual characteristics. The fields to which this disclosure can be applied are not limited, and it is useful in various fields using the program of this disclosure.

[0086] This application claims priority based on Japanese Patent Application No. 2025-057045, filed on 28 March 2025, and incorporates all of its disclosures herein.

[0087] 10, 10A Trained Model Generation Device 11 Information Acquisition Unit 12 Investigation Unit 13 Extraction Unit 14 Retraining Unit 15 Generation Unit 16 Output Unit 101 Central Processing Unit 102 Memory 103 Bus 104 Storage Device 105 Input Device 106 Output Device 107 Communication Device

Claims

1. A method for generating a trained model, comprising an information acquisition step, a research step, an extraction step, and a retraining step, wherein the information acquisition step acquires user-provided information including keywords selected by the user; the research step conducts a research based on the keywords using a large-scale language model; the extraction step extracts retraining information from the research results obtained from the research; and the retraining step retrains the large-scale language model based on the retraining information, with each step being performed by a computer.

2. The method for generating a trained model according to claim 1, wherein the investigation step involves conducting an investigation based on the keywords using a large-scale language model, and further conducting an additional investigation based on the information obtained from the investigation results using the large-scale language model, and the extraction step involves extracting information for retraining from the investigation results obtained from at least one of the investigation and the additional investigation.

3. The method for generating a trained model according to claim 2, wherein the user-provided information further includes an interest index indicating the degree of the user's interest or concern with the keyword, and the research step conducts the additional research on a scale determined based on the interest index.

4. The method for generating a trained model according to claim 3, wherein the scale is the number of times.

5. The method for generating a trained model according to any one of claims 1 to 4, wherein the user-provided information further includes personality information assigned to the large-scale language model, and the retraining step retrains the large-scale language model based on the retraining information and the personality information.

6. The method for generating a trained model according to claim 5, wherein the personality information is personality information classified by a five-factor model.

7. A method for generating a trained model according to any one of claims 1 to 6, further comprising a generation step and an output step, wherein the generation step generates profile information of the large-scale language model based on the retraining information, the profile information includes at least one piece of information relating to the preferences of the large-scale language model and the generation characteristics relating to the information generated by the large-scale language model, and the output step outputs the profile information.

8. The method for generating a trained model according to claim 7, wherein the user-provided information further includes personality information assigned to the large-scale language model, the generation step generates profile information of the large-scale language model based on the retraining information and the personality information, and the profile information includes at least one piece of information of preferences, generation characteristics, and the personality of the large-scale language model.

9. A trained model generation device comprising an information acquisition unit, a research unit, an extraction unit, and a retraining unit, wherein the information acquisition unit acquires user-provided information including keywords selected by the user, the research unit conducts research based on the keywords using a large-scale language model, the extraction unit extracts retraining information from the research results obtained by the research, and the retraining unit retrains the large-scale language model based on the retraining information.

10. The trained model generation apparatus according to claim 9, wherein the research unit conducts a research based on the keywords using a large-scale language model, and further conducts an additional research based on the information obtained from the research results obtained from the research using the large-scale language model, and the extraction unit extracts information for retraining from the research results obtained from at least one of the research and the additional research.

11. The trained model generation apparatus according to claim 10, wherein the user-provided information further includes an interest index indicating the degree of the user's interest or concern with the keyword, and the research unit conducts the additional research on a scale determined based on the interest index.

12. The trained model generation apparatus according to claim 11, wherein the scale is the number of times.

13. The trained model generation device according to any one of claims 9 to 12, wherein the user-provided information further includes personality information assigned to the large-scale language model, and the retraining unit retrains the large-scale language model based on the retraining information and the personality information.

14. The trained model generation device according to claim 13, wherein the personality information is personality information classified by a five-factor model.

15. The trained model generation device according to any one of claims 9 to 14, further comprising a generation unit and an output unit, wherein the generation unit generates profile information of the large-scale language model based on the retraining information, the profile information includes at least one piece of information relating to the preferences of the large-scale language model and the generation characteristics relating to the information generated by the large-scale language model, and the output unit outputs the profile information.

16. The trained model generation device according to claim 15, wherein the user-provided information further includes personality information assigned to the large-scale language model, the generation unit generates profile information of the large-scale language model based on the retraining information and the personality information, and the profile information includes at least one piece of information of preferences, generation characteristics, and the personality of the large-scale language model.

17. A trained model generation program that causes a computer to execute each of the following steps: an information acquisition procedure, a research procedure, an extraction procedure, and a retraining procedure, wherein the information acquisition procedure acquires user-provided information including keywords selected by the user; the research procedure conducts a research based on the keywords using a large-scale language model; the extraction procedure extracts retraining information from the research results obtained from the research; and the retraining procedure retrains the large-scale language model based on the retraining information.

18. The trained model generation program according to claim 17, wherein the investigation procedure involves conducting an investigation based on the keywords using a large-scale language model, and further conducting an additional investigation based on the information obtained from the investigation results using the large-scale language model, and the extraction procedure involves extracting information for retraining from the investigation results obtained from at least one of the investigation and the additional investigation.

19. The trained model generation program according to claim 18, wherein the user-provided information further includes an interest index indicating the degree of the user's interest or concern with the keyword, and the research procedure performs the additional research on a scale determined based on the interest index.

20. A computer-readable recording medium that records a trained model generation program for causing a computer to execute each of the following procedures: an information acquisition procedure, a research procedure, an extraction procedure, and a retraining procedure, wherein the information acquisition procedure acquires user-provided information including keywords selected by the user; the research procedure conducts a research based on the keywords using a large-scale language model; the extraction procedure extracts retraining information from the research results obtained from the research; and the retraining procedure retrains the large-scale language model based on the retraining information.

Images