Information processing device and information processing method

Abstract

A processor of this information processing device: acquires first answer data generated by a first trained model and second answer data generated by a second trained model with a learning level different from the first trained model, on the basis of input prompt data; performs an authenticity analysis of the first answer data using the second answer data as a correct answer; and, as the authenticity analysis result, outputs correction information for false information included in the first answer data, and calculates and outputs a score based on the number of pieces of the false information.

Description

Information processing device and information processing method

[0001] The present invention relates to an information processing device and an information processing method.

[0002] In recent years, interactive generative AI using large-scale language models (LLMs) has been developed, and search services using ChatGPT and the like are being provided. These large-scale language models are artificial intelligence models generated through large-scale pre-training of natural language contained in numerous documents and texts existing in the human world. Some artificial intelligence models have more than a billion parameters. In addition to this, there are also models that have undergone reinforcement learning based on feedback from humans. An example of a base model is a model called a Transformer.

[0003] These large-scale language models are capable of natural language translation, natural language text proofreading, natural language text summarization, and the like. Advanced models among them are capable of natural language question answering (also known as dialogue or conversation), natural language suggestion generation, and programming code generation. Because the number of parameters in these artificial intelligence models is very large, training requires vast amounts of data and computational resources. Therefore, training this level of artificial intelligence for a specific application is extremely resource-inefficient. Therefore, models are generated through large-scale pre-training as foundation models that can be applied to various applications. For example, some large-scale language models are provided on a large-scale language model server and configured to be accessible on various terminals via an API (Application Programming Interface).

[0004] Known examples of such large-scale language models include benchmark-based methods such as MMLU (Massive Multitask Language Understanding, see Non-Patent Document 1) and GLUE (General Language Understanding Evaluation, see Non-Patent Document 2).

[0005] Dan Hendrycks, "Measuring Massive Multitask Language Understanding," [online], [Retrieved June 9, 2024], Internet<URL:https: / / arxiv.org / abs / 2009.03300>

[0006] Alex Wang, "GLUE: A Multi-Task Benchmark and Analysis Platform for Natural Language Understanding," [online], [Retrieved June 9, 2024], Internet<URL:https: / / arxiv.org / abs / 1804.07461>

[0007] One of the challenges in evaluating the performance of large-scale language models (LLMs) is the consistency and fairness of evaluation criteria. To compare different models fairly, the evaluation criteria must be consistent. Currently, however, while several evaluation criteria have been proposed, no unified criteria have been established. As various services using LLMs become more widespread, an increasing number of users are expected to rely on LLMs' answers. In this case, there is a concern that users may believe answers generated by LLMs that contain errors without realizing the errors.

[0008] Therefore, the present invention has been made in consideration of the above-mentioned situation, and aims to provide a technology that increases the accuracy of answers output from an interactive generation AI and updates the further machine learning of an interactive generation AI that outputs an incorrect answer.

[0009] In order to solve the above problems, the present invention has the configurations described in the claims. As an example, the present invention is an information processing device including a processor, which acquires first answer data generated by a first trained model and second answer data generated by a second trained model having a different learning level from the first trained model based on input prompt data, performs an authenticity analysis of the first answer data using the second answer data as a correct answer, and outputs correction information for false information included in the first answer data as a result of the authenticity analysis, and calculates and outputs a score based on the number of pieces of false information.

[0010] According to the present invention, it is possible to provide a technology that increases the accuracy of answers output from an interactive generation AI and updates the machine learning of an interactive generation AI that has output an incorrect answer. Note that the objects, configurations, and effects of the present invention other than those described above will be clarified in the following embodiments.

[0011] FIG. 1 is an explanatory diagram of a use case of an information processing device according to the present embodiment. FIG. 2 is a hardware configuration diagram of an information processing device (smartphone). FIG. 3 is a block diagram showing the functional configuration of the information processing device according to the first embodiment. FIG. 4 is a flowchart showing the processing flow of the information processing device according to the first embodiment. FIG. 5 is an example of a prompt input screen. FIG. 6 is a partial enlarged view of the information processing device after prompt input. FIG. 7 is a partial enlarged view of the information processing device 10 displaying first answer data from the generation AI. FIG. 8 is a diagram showing display example 1 of an authenticity analysis result screen. FIG. 9 is a diagram showing display example 2 of an authenticity analysis result screen. FIG. 10 is a block diagram showing the functional configuration of an information processing device according to a second embodiment. FIG. 11 is a flowchart showing the processing flow of the information processing device according to the second embodiment.

[0012] The information processing device according to the present invention performs an authenticity analysis of the answer output by the generation AI, corrects any false information, and provides the user with a more accurate answer, and is also capable of relearning the generation AI that output the false information. Therefore, the present invention can increase the commercial value of the information processing device to which the present invention is applied, and is expected to contribute to Goal 8.2 of the Sustainable Development Goals (SDGs) advocated by the United Nations (increasing economic productivity through diversification, technological improvement, and innovation, particularly in industries that increase the value of goods and services and labor-intensive industries).

[0013] Hereinafter, an embodiment of the present invention will be described with reference to the drawings. The same components are designated by the same reference numerals throughout the drawings, and duplicated explanations will be omitted.

[0014] First Embodiment FIG. 1 is an explanatory diagram of a use case of an information processing device according to this embodiment.

[0015] The information processing device 10 shown in FIG. 1 is configured to be communicatively connected to a server 1 via a network 2. The server 1 includes a generative AI 3 including a first trained model. The information processing device 10 includes a local AI 311 including a second trained model. The first trained model and the second trained model may use, for example, large language models (LLMs) to realize functions as an interactive generative AI. Alternatively, the first trained model and the second trained model may be configured using a combination of natural language processing (NLP; Neuro Linguistic Programming) and machine learning. These NLP processes can process, understand, and generate responses in a natural way through machine learning processes that continuously improve AI algorithms. The Generative AI3 and Local AI311 are configured to use natural language understanding (NLU) to decipher the meaning and understand the intent of information input either as voice or text, and then use natural language generation (NLG) and NLP components to generate an answer.

[0016] The algorithms of the first trained model and the second trained model may be different, but the learning level of the second trained model is more advanced than that of the first trained model, and the accuracy of the answer sentence to the prompt is assumed to be higher. When evaluating the learning levels of the first trained model and the second trained model, if the first trained model and the second trained model are LLMs, publicly known LLM performance evaluation criteria may be used. The LLM performance evaluation criteria mainly focus on learning ability and prediction ability. In other words, the evaluation criteria are "how much data the LLM can accurately predict" and "how few errors it makes," and the evaluation targets are "how much data it can learn and how accurately it can make predictions based on the learning results."

[0017] Benchmark tests and competitive analysis play an important role in evaluating the performance of LLMs. Benchmark tests are a method of measuring performance based on established standards. This allows for a comparison of the performance of models under the same conditions, allowing for a clear comparison of their merits.

[0018] In addition, competitive analysis allows comparing one model with other models to identify limitations and areas for improvement for each model, and appropriate use of these evaluation methods can lead to improved model performance.

[0019] With reference to these performance evaluation indicators, a second trained model having a more advanced learning level than the first trained model is provided in the local AI 130. Here, a "trained model having an advanced learning level" is, for example, a trained model trained using a relatively large amount of teacher data, a trained model with a relatively large number of learning trials, or a trained model trained using a teacher model with relatively good accuracy due to relatively little noise or relatively little missing data, which are likely to be trained models having an advanced learning level.

[0020] The information processing device 10 transmits the prompt (corresponding to prompt data) input by the user to the server 1. The server 1 inputs the received prompt data to the generation AI 3 and transmits the answer data output by the generation AI 3 to the information processing device 10. The local AI 311 performs an authenticity analysis process on the received answer data and converts the results of the authenticity analysis into a score. If false information is further included, the answer data is corrected and processed, and the corrected answer data and score are provided to the user 5.

[0021] Furthermore, the information processing device 10 transmits the score and the content of the false information (collectively referred to as "genuine analysis result data") to the server 1. As a result, the server 1 can advance machine learning of the generation AI 3 based on the genuine analysis result data and improve the learning level of the first trained model.

[0022] FIG. 2 is a diagram showing the hardware configuration of the information processing device 10 (smartphone).

[0023] The smartphone as the information processing device 10 includes an outer camera 111, an inner camera 112, a ranging sensor 113, an RTC (real-time clock) 114, an acceleration sensor 115, a gyro sensor 116, a geomagnetic sensor 117, a GPS receiver 118, a display 119, a touch panel 120, a microphone 121, a speaker 122, an antenna 123, a communication interface (I / F) 124, a processor 125, a memory 128, and a telephone network communication device 129, and these components are connected to each other via a bus 140 that connects the components.

[0024] The display 119 is an example of an output device that can display and output authenticity analysis result data as a video. A touch panel 120 is layered on the display 119. The touch panel 120 functions as an input device that accepts operations from the user.

[0025] The communication interface (I / F) 124 is a communication interface that performs wireless communication between at least the information processing device 10 and the server 1 by short-range wireless communication, wireless LAN, or base station communication, includes a communication processing circuit corresponding to various predetermined communication interfaces, and is connected to the antenna 123. The short-range wireless communication is performed using a wireless LAN such as Bluetooth (registered trademark), IrDA (Infrared Data Association, registered trademark), Zigbee (registered trademark), HomeRF (Home Radio Frequency, registered trademark), or Wi-Fi (registered trademark). Furthermore, the base station communication may be performed using long-range wireless communication such as 4G, 5G, or LTE.

[0026] The processor 125 is configured by, for example, a CPU (Central Processing Unit). Alternatively, the processor 125 may be equipped with a GPU (Graphics Processing Unit) suitable for AI processing.

[0027] The memory 128 is configured by a flash memory and a non-volatile memory, and stores various programs 126 including an OS (Operating System), an interactive generation AI application 200 (see FIG. 3 ), and a fact-checking application 210 (see FIG. 3 ), as well as data 127 used by the processor 125.

[0028] Although a smartphone is exemplified as the information processing device 10, the type of device is not limited, and may be a tablet terminal, a personal computer, or the like.

[0029] FIG. 3 is a block diagram showing the functional configuration of the information processing apparatus 10 according to the first embodiment.

[0030] An interactive generation AI application 200 for using the generation AI 3 of the server 1 is installed in the information processing device 10. When the generation AI 3 of the server 1 is used on a web service, a web browser may be used instead of the interactive generation AI application 200.

[0031] Furthermore, a fact-checking application 210 according to this embodiment is installed in the information processing device 10. The fact-checking application 210 has a function of acquiring answers generated by the generation AI 3 from the interactive generation AI application 200 or a web browser and performing an authenticity analysis of the answers. Furthermore, the results of the authenticity analysis may be fed back to the generation AI 3 to advance machine learning of the generation AI 3.

[0032] The components that realize the functions provided by the interactive generation AI application 200 include an input detection unit 301, a prompt transmission control unit 302, and a first response data receiving unit 303.

[0033] The information processing device 10 also includes a display control unit 305 that displays display data on the display 119 .

[0034] The fact check app 210 also includes a local AI control unit 310, a local AI 311, an authenticity analysis unit 314, an answer correction unit 315, and a feedback processing unit 316. Components of the local AI 311 include a second answer data generation unit 312 and a second trained model storage unit 313. The functions of each unit will be described later with reference to a flowchart.

[0035] Fig. 4 is a flowchart showing the processing flow of the information processing device of the first embodiment. Fig. 5 is an example of a prompt input screen. Fig. 6 is a partial enlarged view of the information processing device 10 after prompt input. Fig. 7 is a partial enlarged view of the information processing device 10 displaying first response data from the generation AI. Fig. 8 is a diagram showing a display example 1 of an authenticity analysis result screen. Fig. 9 is a diagram showing a display example 2 of an authenticity analysis result screen.

[0036] First, the user opens the prompt input screen 400 of the generated AI 3 shown in FIG. 5 . The prompt input screen 400 includes a text box 401, a "Save" button 402, a "Verify Authenticity" button 403, a "Send" button 404, and a keyboard area 405. When the user 5 operates the keyboard area 405 to input a prompt to query the generated AI 3, the input detection unit 301 generates prompt data. Then, when the user 5 taps the "Send" button 404, the prompt transmission control unit 302 transmits the prompt data to the server 1 (S101). After transmission, the prompt data is temporarily stored in the memory 128 of the information processing device 10.

[0037] In the example of FIG. 6, "Please tell me the nearest station to Building 8 of W University (W Campus)." is entered as the prompt.

[0038] The prompt may be input as a voice input from the microphone 121. In this case, the input detection unit 301 performs a voice recognition process, converts the voice picked up by the microphone into text data, and generates prompt data.

[0039] The server 1 receives the prompt data and inputs it into the first trained model of the generation AI 3. The generation AI 3 outputs response data to the prompt (referred to as first response data). The server 1 transmits the first response data to the information processing device 10, which receives it and displays it on its display (S102).

[0040] 7 corresponds to an example of the screen display displayed in this step. In the example of FIG. 7, the first response data display screen 406 displays the response from generation AI3 as "'W Station A2 Exit' on the Toei Subway Oedo Line or 'W Station 1b Exit' on the Tokyo Metro Tozai Line."

[0041] In this state, when the user 5 taps the "authenticity confirmation button" 403, the fact-checking process by the information processing device 10 starts (S103).

[0042] The second response data generation unit 312 of the local AI control unit 310 reads the second trained model data from the second trained model storage unit 313, inputs prompt data having the same content as the prompt data sent to the generation AI 3 into the second trained model data, and generates second response data (S103). Then, the second response data is passed to the local AI control unit 310.

[0043] The authenticity analysis unit 314 compares the first response data with the second response data sentence by sentence, and performs authenticity analysis processing on the first response data, assuming the second response data to be the correct answer (S104). As an example of an authenticity analysis result, it is assumed that the authenticity of the following four items is determined. <Example of authenticity analysis result> Toei Subway Oedo Line "W Station": false "Exit A2": false Tokyo Metro Tozai Line "W Station": true "Exit 1b": false

[0044] The result of the authenticity determination is not displayed on the display 119 of the information processing device 10, but is temporarily stored as data 127 in the memory 128 within the information processing device 10.

[0045] The authenticity analysis unit 314 converts the result of the authenticity analysis process into a score (S105). In the above <Example of Authenticity Analysis Result>, of the four items "Toei Subway Oedo Line 'W Station'," "Exit A2," "Tokyo Metro Tozai Line 'W Station'," and "Exit 1b," three are determined to be "false" and one is determined to be "true," so the authenticity score is calculated as one out of four is "true," that is, an authenticity score of 25%.

[0046] If the answer correction unit 315 checks the authenticity analysis result and determines that the first answer data contains false information (S106: Yes), it corrects the content of the first answer data to match the content of the second answer data (S107). If there are no corrections, the display control unit 305 displays information based on the first answer data, or if there are corrections, the corrected answer data, together with the authenticity score (S108).

[0047] For example, based on the above <Example of authenticity analysis result>, the answer correction unit 315 generates corrected data indicating the three points determined to be false and authenticity analysis result data including an authenticity score, and displays these on the authenticity analysis result screen 408 of Figure 8.

[0048] Alternatively, the answer correction unit 315 may generate a corrected answer sentence "Toden Arakawa Line "W Station Exit" or Tokyo Metro Tozai Line "W Station Exit 3a," by correcting false information from the first answer data "It is "W Station A2 Exit" on the Toei Subway Oedo Line or "W Station 1b Exit" on the Tokyo Metro Tozai Line," based on the above <Example of Authenticity Analysis Result>, and display this together with the authenticity score on the authenticity analysis result screen 408 of FIG. 9 . At this time, the corrected parts may be highlighted, underlined, or the like to make clear the changes from the first answer data display screen 406.

[0049] If the score and answer are corrected, the feedback processing unit 316 transmits the corrections to the server 1 for feedback to the generation AI 3 (S109).

[0050] The administrator of server 1 or generation AI 3 performs machine learning of the first trained model using the fed-back information (S110).

[0051] According to this embodiment, the following effects can be expected. First, one of the challenges in evaluating the performance of LLMs is the consistency and fairness of evaluation criteria. To make fair comparisons between different models, the evaluation criteria must be consistent. However, evaluation criteria are not currently standardized, and a solution to this problem is being sought: standardization of standards across the industry. In response to this current situation, according to this embodiment, the same prompt is input to multiple trained models that can be ranked in terms of superiority through a performance evaluation, and first and second response data are obtained. The response data output from the trained model evaluated as superior is considered to be the correct answer, and the number of false information contained in the response data output from the trained model evaluated as inferior is converted into a score.

[0052] The error rate output from the LLM is an important indicator for evaluating the performance of the LLM, and the score based on the number of false pieces of information described above allows performance evaluation based on the error rate, which is expected to have the effect of enabling a fact-based evaluation of the LLM's performance based on the fact that an error occurred.

[0053] In addition, output data from a trained model that is evaluated as inferior can be corrected using output data from a trained model that is evaluated as superior and presented to the user, making it easier to prevent problems such as users being misled by false information.

[0054] Second Embodiment The second embodiment is an embodiment in which a self-learning function is added to the local AI. Fig. 10 is a block diagram showing the functional configuration of an information processing device 10 of the second embodiment. Fig. 11 is a flowchart showing the processing flow of the information processing device of the second embodiment.

[0055] As shown in FIG. 10 , the information processing device 10 according to the second embodiment further includes a self-learning unit 320 in the local AI 311 .

[0056] Then, as shown in FIG. 11 , the authenticity analysis process is performed in step S104, similar to the first embodiment. If the authenticity analysis process results in the authenticity analysis unit 314 finding no errors in the first response data (S106: No), the authenticity analysis unit 314 outputs a dataset in which the prompt data is input data and the first response data is output data to the self-learning unit 320. The self-learning unit 320 performs machine learning of the second trained model using this dataset (S120). Then, the process proceeds to step S108, where the same process as the first embodiment is performed.

[0057] According to this embodiment, if there is absolutely no false information in the first response data, it is determined that the LLM performance of the first trained model is equal to or greater than its own (second trained model) performance, and the first response data received from the generation AI 3 can be learned as its own knowledge. This allows the LLM performance of the second trained model of the local AI 130 to maintain a level equal to or greater than the LLM performance of the first trained model.

[0058] The present invention is not limited to the above-described embodiments, and it is possible to replace part of the configuration of one embodiment with another embodiment. It is also possible to add the configuration of another embodiment to the configuration of one embodiment. These all fall within the scope of the present invention, and the numerical values, messages, etc. appearing in the text and figures are merely examples, and the use of different ones does not impair the effects of the present invention.

[0059] Furthermore, some or all of the functions of the invention may be implemented in hardware, for example, by designing an integrated circuit, a general-purpose processor, or an application-specific processor. A processor includes transistors and other circuits and is considered to be circuitry or processing circuitry. Alternatively, the invention may be implemented in software by a microprocessor unit, processor, etc., interpreting and executing an operating program. Furthermore, the scope of software implementation is not limited, and hardware and software may be used together.

[0060] For example, as a fact determination process performed in the authenticity analysis process in step S104, the local AI 331 may individually confirm the contents of the first response data. Furthermore, machine learning may be performed on the second trained model using the usage history, search history, and multimodal input data of user 5, to generate second response data that is in line with the preferences and tendencies of user 5 and use this as the correct answer. This allows the criteria for authenticity analysis to be changed depending on user 5.

[0061] In the above, instead of the local AI, another server may be connected to the network 2, a second trained model may be provided on the other server, and genuine analysis processing may be performed by the information processing device 10.

[0062] Furthermore, in the above embodiment, a first trained model and a second trained model are described, but this is not intended to limit the use of two trained models; the present invention also includes a configuration in which three or more trained models are combined to execute the information processing device according to this embodiment and the information processing method described in this embodiment.

[0063] The embodiments include the following inventions: (Supplementary Note 1) An information processing device comprising: a processor, which acquires, based on input prompt data, first answer data generated by a first trained model and second answer data generated by a second trained model having a different learning level from the first trained model, performs an authenticity analysis of the first answer data using the second answer data as a correct answer, and outputs correction information for false information included in the first answer data as a result of the authenticity analysis, and calculates and outputs a score based on the number of pieces of false information. (Supplementary Note 2) An information processing method, wherein a processor executes the steps of: acquiring first answer data generated by a first trained model and second answer data generated by a second trained model having a different learning level from the first trained model based on input prompt data; performing an authenticity analysis of the first answer data using the second answer data as a correct answer; and outputting correction information for false information contained in the first answer data as a result of the authenticity analysis, and calculating and outputting a score based on the number of pieces of false information.

[0064] 1: Server 5: User 10: Information processing device 111: Outer camera 112: Inner camera 113: Distance measurement sensor 115: Acceleration sensor 116: Gyro sensor 117: Geomagnetic sensor 118: GPS receiver 119: Display 120: Touch panel 121: Microphone 122: Speaker 123: Antenna 125: Processor 126: Program 127: Data 128: Memory 129: Telephone network communication device 140: Bus 200: Interactive generation AI application 210: Fact check application 301: Input detection unit 302: Prompt transmission control unit 303: First response data receiving unit 305: Display control unit 310: Local AI control unit 312: Second response data generation unit 313 : Second trained model storage unit 314 : Authenticity analysis unit 315 : Answer correction unit 316 : Feedback processing unit 320 : Self-learning unit 400 : Prompt input screen 401 : Text box 402 : Save button 403 : Authenticity confirmation button 404 : Send button 405 : Keyboard area 406 : First answer data display screen 408 : Authenticity analysis result screen

Claims

1. An information processing device comprising a processor, which acquires, based on input prompt data, first answer data generated by a first trained model and second answer data generated by a second trained model having a different learning level from the first trained model, performs an authenticity analysis of the first answer data using the second answer data as a correct answer, and, as a result of the authenticity analysis, outputs correction information for false information contained in the first answer data, and calculates and outputs a score based on the number of pieces of false information.

2. An information processing device according to claim 1, wherein the first trained model is a large-scale language model, and the second trained model comprises a large-scale language model having a different learning level from the large-scale language model provided in the first trained model.

3. An information processing device according to claim 1, wherein, if the result of the authenticity analysis indicates that the first response data contains absolutely no false information, the processor performs machine learning of the second trained model using a dataset in which the prompt data is input data and the first response data containing absolutely no false information is output data.

4. An information processing device according to claim 1, wherein the processor outputs corrected response data obtained by correcting the first response data based on the result of the authenticity analysis as the correction information.

5. An information processing device according to claim 1, wherein the processor outputs the correction information describing corrections to the first response data based on the result of the authenticity analysis.

6. An information processing device according to claim 1, wherein the learning level of the second trained model is more advanced than that of the first trained model.

7. An information processing device according to claim 2, wherein the learning level of the large-scale language model provided in the second trained model is more advanced than that of the large-scale language model provided in the first trained model.

8. An information processing method, wherein a processor executes the following steps: acquiring first answer data generated by a first trained model and second answer data generated by a second trained model having a different learning level from the first trained model based on input prompt data; performing an authenticity analysis of the first answer data using the second answer data as the correct answer; and outputting correction information for false information contained in the first answer data as a result of the authenticity analysis, and calculating and outputting a score based on the number of false information pieces.

9. An information processing method according to claim 8, wherein the learning level of the second trained model is more advanced than that of the first trained model.

10. An information processing method according to claim 9, wherein the learning level of the large-scale language model provided in the second trained model is more advanced than that of the large-scale language model provided in the first trained model.

Images