Learning device and learning method

Abstract

A learning device (10) has a learning unit (15) whereby learning of LLM (Large Language Models) to be forgotten (11) is executed using tokens to be forgotten and explanation token sequences explaining the tokens to be forgotten so as to satisfy a first requirement that the LLM to be forgotten (11) do not output the tokens to be forgotten and a second requirement that the LLM (11) to be forgotten do not output token sequences that suitably explain the tokens to be forgotten.

Description

Learning device and learning method

[0001] The present invention relates to a learning device and a learning method.

[0002] Advances in computer technology and AI (Artificial Intelligence) research, along with the use of large-scale corpora, have significantly improved the performance of large language models (LLMs), making them applicable to a variety of tasks such as Q&A systems, translation, and analysis.

[0003] HENG XU, TIANQING ZHU, LEFENG ZHANG, WANLEI ZHOU, PHILIP S. YU, “Machine Unlearning:A Survey”, 2023, [Retrieved August 21, 2020], Internet <URL: https: / / arxiv.org / abs / 2306.03558> Peng Wang, Ningyu Zhang, Bozhong Tian, ​​Zekun Xi, Yunzhi Yao, Ziwen Xu, Mengru Wang, Shengyu Mao, Xiaohan Wang, Siyuan Cheng, Kangwei Liu, Yuansheng Ni, Guozhou Zheng, Huajun Chen, “EasyEdit:An Easy-to-use Knowledge Editing Framework for Large Language Models”, 2023, 2024, [Searched on August 21, 2024], Internet <URL: https: / / arxiv.org / abs / 2308.07269>

[0004] Data on the web contains information linked to individuals. If LLM (Live Language Management) were to collect such personally linked information, it could pose a privacy risk to the use of LLM.

[0005] As an approach to such problems, there is a research field called Machine Unlearning. Machine Unlearning is a field that aims to forget the data used in the learning of an AI model. If Machine Unlearning is realized, it is considered that it contributes to reducing the privacy risk by forgetting the data including information linked to an individual.

[0006] However, at present, the problem setting of Machine Unlearning for LLMs only focuses on realizing the forgetting of a certain data point and has not yet achieved the realization of forgetting information (concepts) linked to an individual.

[0007] The present invention has been made in view of the above, and an object thereof is to provide a learning device and a learning method that realize the forgetting of information linked to an individual for LLMs.

[0008] In order to solve the above-described problems and achieve the object, the learning device according to the present invention uses a token to be forgotten and a sequence of explanatory tokens that explain the token to be forgotten, and sets a first requirement that the language model to be forgotten does not output the token to be forgotten, and a second requirement that the language model to be forgotten does not output a sequence of tokens that well explain the token to be forgotten, and has a learning unit that executes the learning of the language model to be forgotten so as to satisfy these requirements.

[0009] The learning method according to the present invention is a learning method executed by a learning device. Using a token to be forgotten and a sequence of explanatory tokens that explain the token to be forgotten, it includes a step of executing the learning of the language model to be forgotten so as to satisfy a first requirement that the language model to be forgotten does not output the token to be forgotten and a second requirement that the language model to be forgotten does not output a sequence of tokens that well explain the token to be forgotten.

[0010] According to the present invention, the forgetting of information linked to an individual for LLMs is realized.

[0011] FIG. 1 is a diagram for explaining the explanatory nature of a token sequence. FIG. 2 is a diagram showing an example of the configuration of a learning device according to an embodiment. FIG. 3 is a diagram for explaining the processing of the first collection unit shown in FIG. 2. FIG. 4 is a diagram for explaining the processing of the first collection unit shown in FIG. 2. FIG. 5 is a diagram for explaining the processing of the second collection unit shown in FIG. 2. FIG. 6 is a diagram for explaining the processing of the second collection unit shown in FIG. 2. FIG. 7 is a diagram showing the processing procedure of a learning process according to an embodiment. FIG. 8 is a diagram showing an example of evaluation data. FIG. 9 is a diagram showing an example of the output of an LLM before forgetting. FIG. 10 is a diagram showing an example of the output of an LLM before forgetting. FIG. 11 is a diagram showing an example of the output of a forgetting target LLM after forgetting. FIG. 12 is a diagram showing an example of the output of a forgetting target LLM after forgetting. FIG. 13 is a diagram showing the number of outputs that do not meet the requirements for the first requirement and the second requirement. FIG. 14 is a diagram showing the number of outputs that do not meet the requirements for the first requirement and the second requirement. FIG. 15 is a diagram showing an output example of a forgetting target LLM. FIG. 16 is a diagram showing performance evaluation results for a general LLM. FIG. 17 is a diagram showing performance evaluation results for a general LLM. FIG. 18 is a diagram showing an example of a computer in which a learning device is realized when a program is executed.

[0012] Hereinafter, an embodiment of the present invention will be described in detail with reference to the drawings. Note that the present invention is not limited by this embodiment. Also, in the description of the drawings, the same parts are denoted by the same reference numerals.

[0013] [Embodiment] [Premise] Before explaining this embodiment, prerequisite knowledge will be explained.

[0014] [Machine Unlearning] Machine Unlearning is a research field aimed at forgetting information regarding desired training data from a learned AI model (machine learning model). The existing problem settings of Machine Unlearning for large language models (LLMs) (language models) (machine learning models) are shown below.

[0015] Define the training dataset as D, and the dataset to be forgotten as Dｆ Let ∈ D. Also, let A be the learning algorithm and U be the forgetting algorithm.

[0016] Machine Unlearning aims to satisfy equation (1).

[0017]

[0018] In other words, in Machine Unlearning, an AI model trained on training dataset D undergoes a forgetting process, and then the AI ​​model is subjected to the forgetting target dataset D. ｆ The result should be consistent with the AI ​​model trained on the excluding data points.

[0019] [Knowledge Editing] This is a field of study aimed at editing desired knowledge from previously learned LLM (Language Learning). This section explains the problem setting for Knowledge Editing.

[0020] In Knowledge Edit, the data to be edited is represented in the form of the knowledge graph-based equation (2).

[0021]

[0022] However, s is called the subject, o is called the object, and r is called the relation, and this data represents the relationship between the subject and the object.

[0023] Knowledge Edit aims to modify an object into another object by perturbing the parameters of the LLM (Language Learning Model).

[0024] While Knowledge Edit allows for modifying specific relationships, applying it to information forgetting in LLM would likely require multiple Knowledge Edit steps, resulting in a high cost for forgetting.

[0025] [Gradient Ascent] Therefore, in this embodiment, Gradient Ascent is used to achieve forgetting. When the model parameter is \(w\), the loss is \(L\), and the learning rate is \(\lambda\), the update of the model parameter by Gradient Ascent is expressed by Equation (3).

[0026]

[0027] Gradient Ascent is a process of varying the model parameter in the direction where the increase in loss is maximized, that is, the direction in which the loss deteriorates the most, and it is one of the main approaches in Machine Unlearning.

[0028] [Concept Unlearning] In this embodiment, forgetting of information (concepts) associated with an individual is proposed. First, as shown in Equation (4), it is carried out from the formulation of the LLM.

[0029]

[0030] [The First Requirement] In the embodiment, as requirements for concept forgetting, the first requirement and the second requirement are defined. The first requirement is that the LLM does not output the token \(y\) to be forgotten. The first requirement is expressed by Equation (5). ｔ

[0031]

[0032] [The Second Requirement] The second requirement is that the LLM does not output a token sequence that well explains the token \(y\) to be forgotten. The second requirement is expressed by Equation (6). \(U\) is a set of token sequences that well explain the token \(y\) to be forgotten. ｔ ｔ ｔ

[0033]

[0034] [Explanatoriness of Token Sequence] Figure 1 is a diagram for explaining the explanatoriness of a token sequence. A token sequence \(x \in V\) well explains the target token \(y\) to be forgotten means that for a certain LLM model \(\theta\) (oracle LLM), with the explanatory token sequence \(x\) Ｎ ｔ oracle (oracle LLM), with the explanatory token sequence \(x\)ｔ When you enter the following, forget the token y ｔ This means that the following output is obtained. In other words, equation (7) must be true.

[0035]

[0036] Let U be a set of x such as θoracle ｔ This is how it is defined. For example, as shown in Figure 1, when x is input to the reference LLM (Auxiliary Language Model) used as an aid to the forgetting process, the target forgotten token "P group" is output. Thus, x is a sequence of tokens that well explains the target forgotten token "P group". The reference LLM is a different LLM from the LLM to be forgotten.

[0037] [Loss Design] In this embodiment, we aim to realize Concept Unlearning by designing losses corresponding to the first and second requirements, and then performing Gradient Ascent on the designed losses. The LLM to be forgotten is called the LLMθ to be forgotten. obj Therefore, LLM is used as an aid in the forgetting process, and LLMθ is used as a reference. ref Let's assume that.

[0038] The first requirement is the LLMθ to be forgotten. obj However, the forgotten token y ｔ This means not outputting, resulting in equation (8).

[0039]

[0040] In this embodiment, the loss L shown in equation (9) is used to realize equation (8). １ We propose (the first loss function). Loss L １ This corresponds to the first requirement, and in response to the input of a sequence of descriptive tokens that describe the token to be forgotten, the forgotten LLMθ obj The value decreases as the number of forgotten tokens is reduced.

[0041]

[0042] The second requirement is the LLMθ to be forgotten. obj However, the forgotten token y ｔThe key is not to output the token sequence x that explains this well, resulting in equation (10).

[0043]

[0044] In this embodiment, the loss L shown in equation (11) is used to realize equation (10). ２ We propose a second loss function: Loss L ２ This corresponds to the second requirement, and the forgotten token y ｔ For the input, the LLMθ to be forgotten is obj However, the value becomes smaller as the number of explanatory tokens that explain the forgotten token is output.

[0045]

[0046] [Learning Device] Next, a learning device according to the embodiment will be described. Figure 2 is a diagram showing an example of the configuration of the learning device according to the embodiment.

[0047] The learning device 10 according to this embodiment is a generation device that is realized by loading a predetermined program into a computer or the like, which includes ROM (Read Only Memory), RAM (Random Access Memory), CPU (Central Processing Unit), etc., and having the CPU execute the predetermined program. The learning device 10 also has a communication interface for sending and receiving various information with other devices connected via a network or the like.

[0048] The learning device 10 forgets the token y ｔ and a sequence of explanatory tokens x that describes the tokens to be forgotten. ｔ Using this, the forgotten LLM11 is the forgotten token y ｔ The first requirement is that it will not be output, and the forgotten LLM11 is the forgotten token y ｔ The learning of LLM11 to be forgotten is performed in such a way that it satisfies the second requirement of not outputting a token sequence that explains well.

[0049] The learning device 10 includes a memory to be forgotten LLM 11, a reference LLM 12, a first collection unit 13, a second collection unit 14, and a learning unit 15.

[0050] LLM11, which is subject to forgetting, is an LLM that is subject to forgetting its concept.

[0051] Reference LLM12 is an LLM used as an aid in the forgetting process.

[0052] The first collection unit 13 collects data to be used for learning the first requirement. The first collection unit 13 also collects the forgotten token y in the reference LLM 12, which is used as an aid to the forgetting process. ｔ Of the first sequence of descriptive tokens (e.g., descriptive text) obtained by inputting, the forgotten token y is input to the forgotten LLM11. ｔ Collect the second sequence of explanatory tokens that will be output.

[0053] Figures 3 and 4 illustrate the processing of the first collection unit 13 shown in Figure 2. As shown in Figure 3, the first collection unit 13 forgets the token y in the reference LLM 12. ｔ The first collection unit 13 inputs the token sequence x (first descriptive token sequence) output by the reference LLM 12. The first collection unit 13 inputs the token sequence x (first token sequence) obtained from the reference LLM 12 into the forgetting target LLM 11. Then, the first collection unit 13 determines that the forgetting target LLM 11 is the forgetting target token y ｔ Obtain x (the second explanation token sequence) that outputs the following. The second explanation token sequence is Loss L １ This is the data used for training LLM11, the model to be forgotten.

[0054] Specifically, as shown in Figure 4, the first collection unit 13 inputs a descriptive statement (for example, "Who is the three-member techno-pop unit from Q?") to the LLM 11 to be forgotten, which will output the forgetting token "P group". When the first collection unit 13 outputs the forgetting token "P group" in response to this descriptive statement input, it acquires this descriptive statement as a second descriptive token sequence.

[0055] The second collection unit 14 collects data to be used for learning the second requirement. The forgotten token y is stored in the forgotten LLM 11. ｔ Of the third sequence of explanatory tokens (e.g., explanatory text) obtained by inputting, the token y to be forgotten is input into Reference LLM12. ｔCollect the fourth sequence of explanatory tokens that will be output.

[0056] Figures 5 and 6 illustrate the processing of the second collection unit 14 shown in Figure 2. As shown in Figure 5, the second collection unit 14 collects the forgotten token y from the forgotten LLM 11. ｔ The second collection unit 14 inputs the following and obtains x (the third token sequence) output by the LLM 11 to be forgotten. For example, as shown in Figure 6, the second collection unit 14 prompts the LLM 11 to be forgotten with the forgotten token "Tell me about group P", causing it to output a descriptive statement about the forgotten item.

[0057] As shown in Figure 5, the second collection unit 14 inputs x (the third token sequence) obtained from the LLM 11 to be forgotten into the reference LLM 12. Then, the second collection unit 14 determines that the reference LLM 12 contains the forgotten token y ｔ Obtain x (the fourth token sequence) that outputs the following. The fourth token sequence is the loss L ２ This is the data used for training LLM11, the model to be forgotten.

[0058] The learning unit 15 receives a second sequence of explanatory tokens (e.g., an explanatory text) as input and forgets the token y ｔ To prevent outputting the forgotten token y, the learning (Gradient Ascent) of the LLM11 to be forgotten is performed. The learning unit 15 checks that the LLM11 to be forgotten is the forgotten token y ｔ A sequence of explanatory tokens x that outputs the following: １ , x ｎ Using loss L １ The learning unit 15 calculates the loss L. １ The direction in which the rise is maximized, i.e., Loss L １ The model parameters of the LLM11 to be forgotten are updated in the direction that makes the situation worst. The learning unit 15 will update the LLM11 to be forgotten regardless of what text the reference LLM12 outputs, and will then update the forgotten token y ｔ The program is trained to avoid outputting that message.

[0059] Learning unit 15 is the forgotten token y ｔThe learning unit 15 performs training on the LLM 11 to be forgotten so that it does not output the fourth explanatory token sequence for the input y. ｔ A sequence of explanatory tokens x that outputs the following: １ , x ｎ Using loss L ２ The learning unit 15 calculates the loss L. ２ The direction in which the rise is maximized, i.e., Loss L ２ The model parameters of the LLM11 to be forgotten are updated in the direction that causes the worst. The learning unit 15 sets the LLM11 to be forgotten, the reference LLM12, and the forgotten token y ｔ The system is trained to generate sentences that do not produce output like that.

[0060] [Learning Process] Figure 7 is a diagram showing the processing procedure of the learning process according to the embodiment.

[0061] The learning device 10 has a first collection unit 13 that stores the forgotten token y in the reference LLM 12. ｔ The first collection unit 13 inputs the explanation token sequence (first token sequence) output by the reference LLM 12 and obtains it (step S11). The first collection unit 13 inputs the explanation token sequence (first token sequence) obtained from the reference LLM 12 into the forgettable LLM 11, and the forgettable LLM 11 is determined to be the forgettable token y ｔ A sequence of explanatory tokens (a second sequence of explanatory tokens) is collected that outputs the following (step S12).

[0062] The second collection unit 14 receives the forgotten token y from the forgotten LLM 11. ｔ The second collection unit 14 inputs the explanation token sequence (third token sequence) output by the forgettable LLM 11 and obtains it (step S13). The second collection unit 14 inputs the explanation token sequence (third token sequence) obtained from the forgettable LLM 11 into the reference LLM 12, and the reference LLM 12 receives the forgettable token y ｔ Collect a sequence of explanatory tokens (a fourth sequence of tokens) that outputs the following (step S14).

[0063] If, in steps S12 and S14, zero data points are collected (step S15: Yes), the learning device 10 terminates the learning process.

[0064] In steps S12 and S14, if the collected data is not zero (step S15: No), that is, if there is collected data, the learning unit 15 performs learning of the LLM 11 to be forgotten (step S16), and returns to step S11. In step S16, the learning unit 15 uses the second token sequence obtained in step S12 to perform loss L １ Calculate the loss L １ The model parameters of the LLM11 to be forgotten are updated in the direction in which the value of increases. In step S16, the learning unit 15 uses the fourth token sequence obtained in step S14 to calculate the loss L ２ Calculate the loss L ２ The model parameters of the LLM11 to be forgotten are updated in the direction of increasing value.

[0065] [Experiment] Experiment 1 was conducted to evaluate the forgetting performance of the LLM11 to be forgotten, and Experiment 2 was conducted to evaluate the impact on the general performance of the LLM11 to be forgotten.

[0066] First, I will explain the experiment to evaluate forgetting performance. In this experiment, we evaluated each of the first and second requirements mentioned above.

[0067] The LLM models used in this evaluation experiment will be the Mistral-7B-Instruction model and the Llama3-8B model, each using 20 data points for evaluation. The token to be forgotten will be "Google® (registered trademark)".

[0068] The first requirement, not outputting forgotten tokens, is explained below. In Experiment 1, articles related to forgotten tokens were collected from Wikipedia (registered trademark), divided into paragraphs, and each paragraph was used as evaluation data. Figure 8 shows an example of evaluation data. As shown in Figure 8, the evaluation data consists of subject y and its explanation x. ｉ It consists of the following. The number of data points in the dataset is 21918.

[0069] In evaluating the first requirement, the forgotten token y included in the evaluation data ｔ(For example, group P) is masked and given to the LLM11 to be forgotten. In the example in Figure 8, the subject part in the explanation is masked with [MASK]. Then, the task of having the LLM11 to be forgotten answer the masked part is assigned.

[0070] Next, we will explain the second requirement: not outputting a token sequence that clearly explains the token to be forgotten. We will use GPT-4o as a reference LLM12.

[0071] In Experiment 2, the LLM11 to be forgotten is given prompts, such as "Tell me about Google," "Do you know Google?", and "Tell me about Microsoft," instructing it to output a descriptive statement about the forgotten object. The forgotten object tokens included in the output obtained from the LLM11 are then masked and given to GPT-4o, which is tasked with answering the masked portion. In the actual Experiment 2, it is evaluated whether GPT-4o outputs the forgotten object tokens in response to the given prompts.

[0072] [General Performance Evaluation of LLM] We evaluated how the forgetting process according to the embodiment affects the general performance of the LLM 11 to be forgotten. In the general performance evaluation of LLM, we evaluate whether the LLM 11 to be forgotten outputs an explanatory text that adequately describes the token asked about in the prompt in response to a given prompt (for example, "Tell me about Microsoft (registered trademark)?").

[0073] In this evaluation, the performance of the LLM11 to be forgotten was assessed before and after forgetting using eight inference tasks that evaluate the performance of LLMs. The eight inference tasks were HellaSwag, Lambada, Winogrande, COPA, ARC-Easy, ARC-Challenge, MathQA, and PubmedQA.

[0074] [Experimental Results] Figures 9 and 10 show examples of LLM output before forgetting. Figure 9 is an example of output from LLM when "Tell me about Google." is entered into LLM. "Tell me about Google." is the prompt used during the actual forgetting process. Figure 10 is an example of output from LLM when "Do you know Google?" is entered into LLM.

[0075] As shown in Figures 9 and 10, the LLM before the data was forgotten output key descriptions about Google, including descriptions of Google's main services (Google Drive and Google Photos) and descriptions indicating that Google is a multinational corporation targeting the internet.

[0076] Figures 11 and 12 show examples of the output of the LLM 11 after forgetting. Figure 11 shows an example of the output from the LLM 11 after forgetting when "Tell me about Google." was entered into the LLM, which was used during the actual forgetting process. Figure 12 shows an example of the output from the LLM when "Do you know Google?" was entered into the LLM 11 after forgetting.

[0077] As shown in Figures 11 and 12, even for the prompt "Tell me about Google," which was actually used during the forgetting process, and other questions concerning the token "Google" to be forgotten, the forgotten LLM11 only outputs a niche explanation of Google's AMP (Accelerated Mobile Pages). Therefore, it can be considered that the forgotten LLM11 proceeds with forgetting for all questions after the forgetting process.

[0078] Next, we show the evaluation results of the forgetting performance of the LLM11 to be forgotten. Figures 13 and 14 show the number of outputs that do not meet the requirements for the first and second requirements. Figure 13 is for the Mistral-7B-Instruction model. Figure 14 is for the Llama3-8B model.

[0079] The results shown in Figures 13 and 14 indicate the number of outputs that do not satisfy the first and second requirements, respectively, before and after forgetting. In other words, in Figures 13 and 14, a smaller number of outputs indicates higher forgetting performance. As shown in Figures 13 and 14, forgetting performance improved for both the first and second requirements in both models. Therefore, it can be confirmed that forgetting performance is improved by the forgetting process (learning process) in the embodiment, regardless of whether Mistral-7B-Instruction or Llama3-8B is used as the LLM11 to be forgotten.

[0080] Next, the performance evaluation results for a typical LLM are shown below. Figure 15 shows an example of the output of LLM11, which is the model to be forgotten. Figures 16 and 17 show the performance evaluation results for a typical LLM. Figure 16 is for the Mistral-7B-Instruction model. Figure 17 is for the Llama3-8B model.

[0081] When asked about "Microsoft," as shown in Figure 15, the LLM11 targeted for forgetting output a sufficient explanatory text regarding "Microsoft." Furthermore, as shown in Figures 16 and 17, most indicators did not fluctuate by more than 1%, confirming that the forgetting process had little negative impact on the general performance of the LLM.

[0082] [Effects of the Embodiment] In this embodiment, the learning of the LLM 11 to be forgotten is performed in such a way that it satisfies the first requirement of not outputting the tokens to be forgotten, and the second requirement of not outputting a token sequence that clearly explains the tokens to be forgotten.

[0083] As a result, in this embodiment, as shown in the experimental results, the forgetting of the information (concept) associated with the individual in the LLM11 to be forgotten was achieved. Specifically, in this embodiment, the loss L corresponding to the first requirement and the second requirement, respectively, was achieved. １ , L ２ Designed and designed Loss L １ , L ２By performing Gradient Ascent on the LLM 11, concept forgetting of the LLM to be forgotten was achieved. Furthermore, according to this embodiment, the forgetting process does not affect the general performance of the LLM, and appropriate concept forgetting can be achieved.

[0084] [Regarding the System Configuration of the Embodiment] Each component of the learning device 10 is a functional concept and does not necessarily have to be physically configured as shown in the figure. In other words, the specific forms of distribution and integration of the functions of the learning device 10 are not limited to those shown in the figure, and all or part of it can be configured by functionally or physically distributing or integrating in any unit according to various loads and usage conditions.

[0085] Furthermore, each process performed in the learning device 10 may be implemented, in whole or in part, by a CPU, a GPU (Graphics Processing Unit), and a program that is analyzed and executed by the CPU and GPU. Alternatively, each process performed in the learning device 10 may be implemented as hardware using wired logic.

[0086] Furthermore, among the processes described in the embodiments, all or part of the processes described as being performed automatically can be performed manually. Alternatively, all or part of the processes described as being performed manually can be performed automatically by known methods. In addition, the processing procedures, control procedures, specific names, and information including various data and parameters described above and illustrated may be changed as appropriate unless otherwise specified.

[0087] [Program] Figure 18 shows an example of a computer in which the learning device 10 is realized when a program is executed. The computer 1000 has, for example, memory 1010 and CPU 1020. The computer 1000 also has a hard disk drive interface 1030, a disk drive interface 1040, a serial port interface 1050, a video adapter 1060, and a network interface 1070. These components are connected by a bus 1080.

[0088] Memory 1010 includes ROM 1011 and RAM 1012. ROM 1011 stores, for example, a boot program such as a BIOS (Basic Input Output System). The hard disk drive interface 1030 is connected to the hard disk drive 1090. The disk drive interface 1040 is connected to the disk drive 1100. For example, a removable storage medium such as a magnetic disk or optical disk is inserted into the disk drive 1100. The serial port interface 1050 is connected to, for example, a mouse 1110 and a keyboard 1120. The video adapter 1060 is connected to, for example, a display 1130.

[0089] The hard disk drive 1090 stores, for example, an OS (Operating System) 1091, an application program 1092, a program module 1093, and program data 1094. That is, the program that defines each process of the learning device 10 is implemented as a program module 1093 in which code executable by the computer 1000 is written. The program module 1093 is stored, for example, in the hard disk drive 1090. For example, a program module 1093 for performing the same processes as the functional configuration of the learning device 10 is stored in the hard disk drive 1090. Note that the hard disk drive 1090 may be replaced by an SSD (Solid State Drive).

[0090] Furthermore, the configuration data used in the processing of the above-described embodiment is stored as program data 1094 in, for example, memory 1010 or hard disk drive 1090. The CPU 1020 then reads the program module 1093 and program data 1094 stored in memory 1010 or hard disk drive 1090 into RAM 1012 as needed and executes them.

[0091] Furthermore, the program module 1093 and program data 1094 are not limited to being stored in the hard disk drive 1090; for example, they may be stored in a removable storage medium and read by the CPU 1020 via a disk drive 1100 or the like. Alternatively, the program module 1093 and program data 1094 may be stored in another computer connected via a network (LAN (Local Area Network), WAN (Wide Area Network), etc.). The program module 1093 and program data 1094 may then be read from the other computer by the CPU 1020 via a network interface 1070.

[0092] Although embodiments applying the invention made by the present inventors have been described above, the present invention is not limited by the descriptions and drawings that constitute part of the disclosure of the present invention in these embodiments. That is, all other embodiments, examples, and operational techniques made by those skilled in the art based on these embodiments are included in the scope of the present invention.

[0093] 10 Learning device 11 Forgetting target LLM 12 Reference LLM 13 First collection unit 14 Second collection unit 15 Learning unit

Claims

1. A learning device characterized by having a learning unit that performs learning of a language model to be forgotten, using a token to be forgotten and a sequence of explanatory tokens that explains the token to be forgotten, such that a first requirement is met: the language model to be forgotten does not output the token to be forgotten, and a second requirement is met: the language model to be forgotten does not output a sequence of tokens that well explains the token to be forgotten.

2. The learning device according to claim 1, further comprising a first collection unit that collects a second sequence of explanation tokens from a first sequence of explanation tokens obtained by inputting the forgetting target tokens into an auxiliary language model used as an aid to forgetting processing, the second sequence of explanation tokens being input into the forgetting target language model to output the forgetting target tokens, and the learning unit performing learning of the forgetting target language model so as not to output the forgetting target tokens in response to the input of the second sequence of explanation tokens.

3. The learning device according to claim 1, further comprising a second collection unit that collects a fourth sequence of explanation tokens from a third sequence of explanation tokens obtained by inputting a forgetting token into the language model to be forgotten, the fourth sequence of explanation tokens being input into an auxiliary language model used as an aid to the forgetting process, and the learning unit performing learning of the language model to be forgotten so as not to output the fourth sequence of explanation tokens in response to the input of a forgetting token.

4. A learning method performed by a learning device, comprising the step of performing learning of a language model to be forgotten, using a token to be forgotten and a sequence of explanatory tokens that explains the token to be forgotten, such that the language model to be forgotten does not output the token to be forgotten, and the language model to be forgotten does not output a sequence of tokens that explains the token to be forgotten well.

Images