Information processing device, information processing method, and information processing program

Abstract

This information processing device (100) constructs a de-biased artificial intelligence model. A training unit (111) trains an artificial intelligence model (51) by using bias data (52) that includes a bias. A bias vector calculation unit (112) uses a bias model (53) obtained through training to calculate a bias vector (54) that is a model weight representing the bias. A de-biased model calculation unit (120) corrects the artificial intelligence model (51) by using the bias vector (54), and calculates a de-biased model (55) in which the artificial intelligence model (51) is de-biased.

Description

Information Processing Apparatus, Information Processing Method, and Information Processing Program

[0001] The present disclosure relates to an information processing apparatus, an information processing method, and an information processing program.

[0002] In recent years, artificial intelligence models such as language models have been constructed. Specific examples include artificial intelligence models such as GPT-3.5 or Llama-2. These artificial intelligence models may be referred to as generative AI. GPT is an abbreviation for Generative Pretrained Transformer. Llama is an abbreviation for Large Language Model Meta AI. AI is an abbreviation for Artificial Intelligence.

[0003] Artificial intelligence models may often generate text containing biases such as racial discrimination or gender discrimination. For practical applications of artificial intelligence models, it is necessary to generate training data such as text with the bias removed. Such biases are regarded as problems not only in dialogue tasks but also in all generative tasks such as translation tasks.

[0004] Patent Document 1 discloses a method of identifying a group of vectors and generating a debiased artificial intelligence model based on bias classification criteria and a bias definition vector.

[0005] Japanese Patent Application Laid-Open No. 2020-170504

[0006] Debiasing methods include techniques for determining whether data generated by an artificial intelligence model contains bias. Other methods include techniques for cleansing training data used to pre-train an AI model to remove biased data. However, all of these methods require debiased data. Patent document 1 discloses a method for creating debiased vectors by subtracting word embedding vectors. This method also requires debiased data and data that serves as a bias classification criterion. Debiased training data, i.e., debiased data, is often created manually by first preparing biased data and then constructing debiased data based on that. Therefore, a challenge with debiased data is the high cost of data preparation.

[0007] This disclosure aims to construct a debiased artificial intelligence model without using debiased training data.

[0008] The information processing device relating to this disclosure includes a bias vector acquisition unit that learns an artificial intelligence model using bias data including bias, and uses the bias model obtained through learning to acquire bias vectors which are model weights representing the bias, and a debias model calculation unit that corrects the artificial intelligence model using the bias vectors and calculates a debias model in which the artificial intelligence model has been debiased.

[0009] The information processing device described herein can construct a debiased artificial intelligence model without using debiased training data.

[0010] A diagram showing an example configuration of the information processing device according to Embodiment 1. A flowchart showing an example operation of the information processing device according to Embodiment 1. A schematic diagram showing an example operation of the information processing device according to Embodiment 1. A schematic diagram showing examples of each step in the operation example of the information processing device according to Embodiment 1. A diagram showing an example configuration of the information processing device according to Modification 1 of Embodiment 1. A flowchart showing an example operation of the information processing device according to Modification 1 of Embodiment 1. A diagram showing an example of obtaining the utilization ratio according to Modification 1 of Embodiment 1. A diagram showing an example configuration of the information processing device according to Modification 4 of Embodiment 1.

[0011] The following description of this embodiment will be illustrated with reference to the figures. In each figure, identical or corresponding parts are denoted by the same reference numerals. In the description of the embodiment, the explanation of identical or corresponding parts will be omitted or simplified as appropriate. The arrows in the figures mainly indicate the flow of data or processing.

[0012] Embodiment 1. ***Description of Configuration*** Figure 1 shows an example configuration of the information processing device 100 according to this embodiment. The information processing device 100 is a computer. The information processing device 100 includes a processor 910, as well as other hardware such as a memory 921, an auxiliary storage device 922, an input interface 930, an output interface 940, and a communication device 950. The processor 910 is connected to the other hardware via signal lines and controls this other hardware.

[0013] The information processing device 100 includes, as functional elements, a bias vector acquisition unit 110, a debias model calculation unit 120, and a storage unit 150. The bias vector acquisition unit 110 includes a learning unit 111 and a bias vector calculation unit 112. The storage unit 150 stores an artificial intelligence model 51, bias data 52, a bias model 53, a bias vector 54, and a debias model 55.

[0014] The functions of the bias vector acquisition unit 110 and the debias model calculation unit 120 are implemented by software. The storage unit 150 is provided in the memory 921. The storage unit 150 may also be provided in the auxiliary storage device 922, or it may be distributed between the memory 921 and the auxiliary storage device 922.

[0015] The processor 910 is a device that executes an information processing program. The information processing program is a program that realizes the functions of the bias vector acquisition unit 110 and the debias model calculation unit 120. The processor 910 is an IC that performs arithmetic processing. Specific examples of the processor 910 are a CPU, DSP, and GPU. IC is an abbreviation for Integrated Circuit. CPU is an abbreviation for Central Processing Unit. DSP is an abbreviation for Digital Signal Processor. GPU is an abbreviation for Graphics Processing Unit.

[0016] Memory 921 is a storage device that temporarily stores data. Specific examples of memory 921 include SRAM or DRAM. SRAM is an abbreviation for Static Random Access Memory. DRAM is an abbreviation for Dynamic Random Access Memory. Auxiliary storage device 922 is a storage device that stores data. Specific examples of auxiliary storage device 922 include HDD. Alternatively, auxiliary storage device 922 may be a portable storage medium such as an SD® memory card, CF, NAND flash, flexible disk, optical disk, compact disk, Blu-ray® disc, or DVD. HDD is an abbreviation for Hard Disk Drive. SD® is an abbreviation for Secure Digital. CF is an abbreviation for CompactFlash (registered trademark). DVD is an abbreviation for Digital Versatile Disk.

[0017] The input interface 930 is a port to which input devices such as a mouse, keyboard, or touch panel are connected. Specifically, the input interface 930 is a USB terminal. The input interface 930 may also be a port connected to a LAN. USB is an abbreviation for Universal Serial Bus. LAN is an abbreviation for Local Area Network.

[0018] The output interface 940 is a port to which the cable of an output device, such as a display, is connected. Specifically, the output interface 940 is a USB terminal or an HDMI® terminal. Specifically, the display is an LCD. The output interface 940 is also called the display interface. HDMI® is an abbreviation for High Definition Multimedia Interface. LCD is an abbreviation for Liquid Crystal Display.

[0019] The communication device 950 has a receiver and a transmitter. The communication device 950 is connected to a communication network such as a LAN, the Internet, a telephone line, or Wi-Fi (registered trademark). Specifically, the communication device 950 is a communication chip or NIC. NIC is an abbreviation for Network Interface Card.

[0020] The information processing program is executed in the information processing device 100. The information processing program is loaded into the processor 910 and executed by the processor 910. The memory 921 stores not only the information processing program but also the OS. OS is an abbreviation for Operating System. The processor 910 executes the information processing program while executing the OS. The information processing program and the OS may also be stored in the auxiliary storage device 922. The information processing program and the OS stored in the auxiliary storage device 922 are loaded into the memory 921 and executed by the processor 910. Note that part or all of the information processing program may be incorporated into the OS.

[0021] The information processing device 100 may include multiple processors that replace the processor 910. These multiple processors share the task of executing the information processing program. Each processor is a device that executes the information processing program, just like the processor 910.

[0022] Data, information, signal values, and variable values ​​used, processed, or output by the information processing program are stored in memory 921, auxiliary storage device 922, or registers or cache memory within the processor 910.

[0023] The "parts" in the bias vector acquisition unit 110 and the debias model calculation unit 120 may be read as "circuit," "process," "procedure," "process," or "circuitry." The information processing program causes the computer to execute the bias vector acquisition process and the debias model calculation process. The bias vector acquisition process includes a learning process and a bias vector calculation process. The "process" in the bias vector acquisition process and the debias model calculation process may be read as "program," "program product," "computer-readable storage medium storing the program," or "computer-readable recording medium recording the program." Furthermore, the information processing method is a method performed by the information processing device 100 executing the information processing program. The information processing program may be provided stored on a computer-readable recording medium. The information processing program may also be provided as a program product.

[0024] ***Explanation of Operation*** Next, the operation of the information processing device 100 according to this embodiment will be explained. The operation procedure of the information processing device 100 corresponds to the information processing method. Furthermore, the program that realizes the information processing, which is the operation of the information processing device 100, corresponds to the information processing program.

[0025] Figure 2 is a flowchart showing an example of the operation of the information processing device 100 according to this embodiment.

[0026] <Bias vector acquisition process: Steps S101 to S102> The bias vector acquisition unit 110 learns the artificial intelligence model 51 using bias data 52 which includes bias. The bias vector acquisition unit 110 uses the bias model 53 obtained through learning to acquire bias vectors 54, which are model weights representing bias. Specifically, it is as follows.

[0027] <<Learning Process: Step S101>> In step S101, the learning unit 111 learns the artificial intelligence model 51 using bias data 52. By learning the artificial intelligence model 51 using bias data 52, the learning unit 111 obtains a biased model 53 that includes bias.

[0028] Here, the artificial intelligence model 51 is the original model that is debiased by the information processing device 100 according to this embodiment. The artificial intelligence model 51 as the original model is, for example, a language model. Examples of artificial intelligence model 51 for text include language models such as BERT or GPT. BERT is an abbreviation for Bidirectional Encoder Representations from Transformers. The bias data 52 is data that contains bias. In this embodiment, it is sufficient to prepare only bias data 52 that contains bias as training data.

[0029] Figure 3 is a schematic diagram showing an example of the operation of the information processing device 100 according to this embodiment. Figure 4 is a schematic diagram showing an example of each step in the operation example of the information processing device 100 according to this embodiment.

[0030] As shown in Figures 3 and 4, in step S101, the learning unit 111 trains the artificial intelligence model 51 using biased data 52. For example, the learning unit 111 learns a method of predicting the next word in the sentences of the biased data 52. Alternatively, the learning unit 111 learns a method of predicting words that are randomly masked within the sentences of the biased data 52. By training the artificial intelligence model 51 using only the biased data 52 in this way, a biased model 53 that contains a lot of bias is obtained.

[0031] <<Bias vector calculation process: Step S102>> In step S102, the bias vector calculation unit 112 calculates the bias vector 54 by subtracting the weights of the artificial intelligence model 51 from the weights of the bias model 53 obtained by training the artificial intelligence model 51 with the bias data 52.

[0032] In Figure 4, the weights of the artificial intelligence model 51 are denoted as θbase, the weights of the bias model 53 as θbias, and the bias vector as τ. Here, the model weights are the parameters of the model. For example, in a neural network, the model weights are information associated with the synapses between neurons.

[0033] The bias vector calculation unit 112 calculates the bias vector 54 by subtracting the weights of the artificial intelligence model 51 from the weights of the bias model 53. That is, the following equation 1 is obtained: τ = θbias - θbase (Equation 1) The bias vector τ is the difference between the weights of the bias model 53 and the weights of the artificial intelligence model 51.

[0034] <Debiased Model Calculation Process: Step S103> In step S103, the debiased model calculation unit 120 corrects the artificial intelligence model 51 using the bias vector 54 and calculates a debiased model 55 in which the artificial intelligence model 51 has been debiased. Specifically, the debiased model calculation unit 120 calculates the weights of the debiased model 55 by subtracting the bias vector 54 from the weights of the artificial intelligence model 51. The debiased model calculation unit 120 constructs the debiased model 55 based on the weights of the debiased model 55.

[0035] In Figure 4, the weight of the debias model 55 is further defined as θdebias. The debias model calculation unit 120 subtracts the bias vector 54 from the weight of the artificial intelligence model 51 to calculate the weight of the debias model 55. That is, the following equation 2 is obtained: θdebias = θbase - τ (Equation 2) Here, the bias vector 54 is simply subtracted from the weight of the artificial intelligence model 51. However, a configuration that allows adjustment of the utilization ratio of the bias vector 54 may be used. A method for adjusting the utilization ratio of the bias vector 54 will be described later.

[0036] In step S104, the debiased model calculation unit 120 stores the debiased model 55, which has been calculated based on the weights of the debiased model 55, in the storage unit 150.

[0037] ***Other Configurations*** <Modification 1> In the embodiment described above, when subtracting the bias vector 54 from the weights of the artificial intelligence model 51, the bias vector 54 is subtracted as is. In Modification 1 of this embodiment, the debiased model calculation unit 120 obtains the utilization ratio of the bias vector 54. Then, a method is described in which the debiased model calculation unit 120 corrects the artificial intelligence model 51 using the bias vector obtained by multiplying the bias vector 54 by the utilization ratio.

[0038] Figure 5 shows an example of the configuration of an information processing device 100 according to modification 1 of this embodiment. In the information processing device 100 of Figure 5, in addition to the configuration of the information processing device 100 of Figure 1, the utilization ratio 56 is stored in the storage unit 150.

[0039] Figure 6 is a flowchart showing an example of the operation of the information processing device 100 according to Modification 1 of this embodiment. In Figure 6, the process in step S103 in Figure 3 is changed to steps S103a and S103b. The processes in steps S101, S102, and S104 are the same as in Figure 3.

[0040] In step S103a, the debias model calculation unit 120 obtains the utilization ratio 56. Specifically, it is as follows:

[0041] Figure 7 shows an example of obtaining the utilization ratio 56 according to Modification 1 of this embodiment. For example, as shown in Figure 7, the utilization ratio 56 is input by the user via the display and stored in the storage unit 150. Alternatively, the user may store the utilization ratio 56 in the storage unit 150 in advance.

[0042] In step S103b, the debiased model calculation unit 120 corrects the artificial intelligence model 51 using the bias vector obtained by multiplying the bias vector 54 by the utilization ratio 56.

[0043] In FIG. 4, the utilization ratio is further denoted as α. The bias model calculation unit 120 subtracts a bias vector obtained by multiplying the utilization ratio 56 from the weights of the artificial intelligence model 51. That is, the following Equation 3 is obtained. θdebias = θbase - α * τ (Equation 3) If α is increased, the artificial intelligence model 51 is more biased. On the other hand, when α approaches 0, θdebias becomes a weight close to the weight θbase of the original artificial intelligence model 51.

[0044] As described above, in the first modification example of the present embodiment, the user can adjust the utilization ratio to be subtracted through the display. For example, if the utilization ratio to be subtracted is set to 0, the original model is not biased. On the other hand, if the utilization ratio is set to 1, the bias vector is directly subtracted from the original model.

[0045] <Modification Example 2> In the above-described present embodiment, the bias model calculation unit 120 calculates the weights of the bias model 55 by subtracting the bias vector 54 from the weights of the artificial intelligence model 51. In the second modification example of the present embodiment, the bias model calculation unit 120 may calculate the weights of the bias model 55 by adding the bias vector 54 to the weights of the artificial intelligence model 51. Alternatively, the bias model calculation unit 120 may calculate the weights of the bias model 55 by multiplying or dividing the weights of the artificial intelligence model 51 and the bias vector 54. Any correction method may be used as long as the original artificial intelligence model 51 can be corrected using the bias vector 54 to generate the bias model 55.

[0046] Here, an example of calculating the weights of the bias model 55 by adding or subtracting the bias vector 54 to / from the weights of the artificial intelligence model 51 will be described. For example, there are multiple types of discrimination biases, such as racial discrimination, gender discrimination, or ethnic discrimination, and there is not always only one type of bias. These biases are related to each other. For example, there may be biases related to both gender and occupation. Here, consider the case of subtracting the gender vector and the occupation vector from the base model. In this case, usually, the following Equation 4 is considered. θgender represents the gender vector, and θprofession represents the occupation vector. θdebias = θbase - θgender - θprofession (Equation 4)

[0047] However, as presented earlier, gender and profession are related to each other. Therefore, θgen-prof, which is the weight of the bias related to these two, will be subtracted twice by the process of Equation 4. The detailed process is as follows in Equation 5. θdebias = θbase - (θgender + θgen-prof) - (θprofession + θgen-prof) (Equation 5) Here, θgen-prof is, similar to the bias vector, a vector constructed using, for example, both gender and occupation bias data.

[0048] To handle the process of subtracting twice as described above, addition and subtraction operations are performed as in the following Equation 6. θdebias = θbase - θgender - θprofession + θgen-prof (Equation 6)

[0049] <Modification Example 3>In the above-described embodiment, a language model was used as the artificial intelligence model 51 for explanation, but this embodiment can also be applied to other artificial intelligence models. For example, the artificial intelligence model 51 may be an image model.

[0050] <Modification 4> In this embodiment, the functions of the bias vector acquisition unit 110 and the debias model calculation unit 120 are implemented in software. As a modification, the functions of the bias vector acquisition unit 110 and the debias model calculation unit 120 may be implemented in hardware. Specifically, the information processing device 100 includes an electronic circuit 909 instead of a processor 910.

[0051] Figure 8 shows an example configuration of the information processing device 100 according to modification 4 of this embodiment. The electronic circuit 909 is a dedicated electronic circuit that realizes the functions of the bias vector acquisition unit 110 and the debias model calculation unit 120. Specifically, the electronic circuit 909 is a single circuit, a composite circuit, a programmed processor, a parallel programmed processor, a logic IC, a GA, an ASIC, or an FPGA. GA is an abbreviation for Gate Array. ASIC is an abbreviation for Application Specific Integrated Circuit. FPGA is an abbreviation for Field-Programmable Gate Array.

[0052] The functions of the bias vector acquisition unit 110 and the debias model calculation unit 120 may be implemented in a single electronic circuit, or they may be implemented in a distributed manner across multiple electronic circuits.

[0053] As another variation, some functions of the bias vector acquisition unit 110 and the debias model calculation unit 120 may be implemented by electronic circuits, while the remaining functions are implemented by software. Alternatively, some or all functions of the bias vector acquisition unit 110 and the debias model calculation unit 120 may be implemented by firmware.

[0054] The processor and the electronic circuit are also called processing circuits. In other words, the functions of the bias vector acquisition unit 110 and the debias model calculation unit 120 are realized by the processing circuits.

[0055] ***Explanation of the Effects of This Embodiment*** As described above, in the information processing device according to this embodiment, the learning unit further trains the original model, which is a pre-trained artificial intelligence model stored in the memory unit, based on bias data. The bias vector calculation unit calculates the bias vector by subtracting the weights of the original model from the weights of the bias model obtained through the additional training. The debiased model calculation unit obtains a debiased model by subtracting the bias vector from the weights of the original model. Here, the utilization ratio used for the subtraction can be defined before the subtraction of the bias vector, and the utilization ratio is stored in the memory unit.

[0056] As described above, according to the information processing device of this embodiment, by preparing only bias data that includes bias, a debiased model, which is an artificial intelligence model that does not contain bias, can be constructed without preparing debiased data. Therefore, the information processing device of this embodiment has the effect of eliminating the need to prepare debiased data, which is training data that does not contain bias and is therefore costly. Furthermore, the information processing device of this embodiment allows setting the utilization ratio of bias vectors used to correct the artificial intelligence model. Therefore, the amount of debias correction in the artificial intelligence model can be adjusted according to the information processing device of this embodiment.

[0057] In Embodiment 1 described above, each part of the information processing device was described as an independent functional block. However, the configuration of the information processing device does not have to be as described in the above-described embodiment. The functional blocks of the information processing device can be configured in any way as long as they can realize the functions described in the above-described embodiment. Furthermore, the information processing device does not have to be a single device, but a system composed of multiple devices. Also, multiple parts of Embodiment 1 may be combined and implemented. Alternatively, only one part of Embodiment 1 may be implemented. In addition, Embodiment 1 may be combined and implemented in any way, either as a whole or in part. That is to say, in Embodiment 1, it is possible to freely combine each embodiment, modify any component of each embodiment, or omit any component in each embodiment.

[0058] The embodiments described above are essentially preferred examples and are not intended to limit the scope of the Disclosure, the scope of the Applications of the Disclosure, or the scope of Uses of the Disclosure. The embodiments described above can be modified in various ways as needed. For example, the procedures described using flowcharts or sequence diagrams may be modified as appropriate.

[0059] The various aspects of this disclosure are summarized below as an appendix.

[0060] (Note 1) An information processing device comprising: a bias vector acquisition unit that learns an artificial intelligence model using bias data including bias, and uses the bias model obtained through learning to acquire a bias vector which is a model weight representing the bias; and a debiased model calculation unit that corrects the artificial intelligence model using the bias vector and calculates a debiased model in which the artificial intelligence model has been debiased. (Note 2) The information processing device according to Note 1, wherein the bias vector acquisition unit comprises: a learning unit that learns the artificial intelligence model using the bias data; and a bias vector calculation unit that subtracts the weights of the artificial intelligence model from the weights of the bias model obtained by learning the artificial intelligence model using the bias data and calculates the bias vector. (Note 3) The information processing device according to Note 1 or Note 2, wherein the debiased model calculation unit acquires the utilization ratio of the bias vector and corrects the artificial intelligence model using the bias vector obtained by multiplying the bias vector by the utilization ratio. (Note 4) The information processing device according to any one of Notes 1 to 3, wherein the debiased model calculation unit calculates the weights of the debiased model by subtracting the bias vector from the weights of the artificial intelligence model. (Note 5) The debiased model calculation unit is an information processing device according to any one of Notes 1 to 3, wherein the debiased model calculation unit calculates the weights of the debiased model by adding the bias vector to the weights of the artificial intelligence model. (Note 6) The artificial intelligence model is a language model, as described in any one of Notes 1 to 5. (Note 7) An information processing method in which a computer learns an artificial intelligence model using biased data that includes bias, uses the biased model obtained through learning to obtain a bias vector which is a model weight representing the bias, corrects the artificial intelligence model using the bias vector, and calculates a debiased model in which the artificial intelligence model has been debiased.(Note 8) An information processing program that causes a computer to perform the following: a bias vector acquisition process that learns an artificial intelligence model using biased data that includes bias, and uses the biased model obtained through learning to obtain bias vectors which are model weights representing the bias; and a debiased model calculation process that corrects the artificial intelligence model using the bias vectors and calculates a debiased model in which the artificial intelligence model has been debiased.

[0061] 51 Artificial intelligence model, 52 Biased data, 53 Biased model, 54 Biased vector, 55 Debiased model, 56 Utilization ratio, 100 Information processing device, 110 Biased vector acquisition unit, 111 Learning unit, 112 Biased vector calculation unit, 120 Debiased model calculation unit, 150 Storage unit, 909 Electronic circuit, 910 Processor, 921 Memory, 922 Auxiliary storage device, 930 Input interface, 940 Output interface, 950 Communication device.

Claims

1. An information processing device comprising: a bias vector acquisition unit that learns an artificial intelligence model using biased data that includes bias, and uses the biased model obtained through learning to acquire bias vectors which are model weights representing the bias; and a debiased model calculation unit that corrects the artificial intelligence model using the bias vectors and calculates a debiased model in which the artificial intelligence model has been debiased.

2. The information processing apparatus according to claim 1, wherein the bias vector acquisition unit comprises a learning unit that learns the artificial intelligence model using the bias data, and a bias vector calculation unit that calculates the bias vector by subtracting the weights of the artificial intelligence model from the weights of the bias model obtained by learning the artificial intelligence model using the bias data.

3. The information processing apparatus according to claim 1 or 2, wherein the debiased model calculation unit obtains the utilization ratio of the bias vector and corrects the artificial intelligence model using the bias vector obtained by multiplying the bias vector by the utilization ratio.

4. The information processing apparatus according to any one of claims 1 to 3, wherein the debias model calculation unit calculates the weights of the debias model by subtracting the bias vector from the weights of the artificial intelligence model.

5. The information processing apparatus according to any one of claims 1 to 3, wherein the debias model calculation unit calculates the weights of the debias model by adding the bias vector to the weights of the artificial intelligence model.

6. The information processing device according to any one of claims 1 to 5, wherein the artificial intelligence model is a language model.

7. An information processing method in which a computer learns an artificial intelligence model using biased data that includes bias, obtains a bias vector which is a model weight representing the bias using the bias model obtained through learning, corrects the artificial intelligence model using the bias vector, and calculates a debiased model in which the artificial intelligence model has been debiased.

8. An information processing program that causes a computer to perform the following: a bias vector acquisition process, which involves training an artificial intelligence model using biased data that includes bias, and using the biased model obtained through training to acquire bias vectors, which are model weights representing the bias; and a debiased model calculation process, which involves correcting the artificial intelligence model using the bias vectors and calculating a debiased model in which the artificial intelligence model has been debiased.

Images