Information processing methods, information processing systems, information processing devices, and programs

Abstract

This invention provides an information processing method, information processing system, information processing device, and program that enable the stable generation of responses from a large-scale language model by determining the information of the memory region to be referenced by the large-scale language model based on priority control rules, in cases where inconsistencies may occur in information between multiple memory regions. [Solution] An information processing method is provided for an information processing system that controls a large-scale language model having multiple memory areas, comprising: a contradiction detection step that detects whether or not there is a contradiction in the information stored in each of the multiple memory areas; a reference target determination step that, if the contradiction detection step detects a contradiction, determines a reference memory area from the multiple memory areas based on a priority control rule; and an interaction control step that controls the reference memory area determined by the reference target determination step as a reference target in the large-scale language model.

Description

Technical Field

[0001] The present invention relates to an information processing method, an information processing system, an information processing apparatus, and a program using a large language model.

Background Art

[0002] In recent years, technologies related to document generation, dialogue processing, and inference support using large language models (LLMs) have been developed. For example, Patent Document 1 discloses an information processing apparatus that generates a response sentence for input text using a large language model. Also, in a system using a large language model, a retrieval augmented generation technology (RAG) that generates a response by referring to an external database or a storage area within the system has become widespread. As a reference destination for such a retrieval augmented generation technology, there may be a configuration in which a temporary storage area that is valid only during a dialogue session and a long-term storage area that is held across multiple sessions coexist.

Prior Art Documents

Patent Documents

[0003]

Patent Document 1

Summary of the Invention

Problems to be Solved by the Invention

[0004] However, information obtained from different storage areas may be content-wise contradictory or may not be logically consistent. In conventional retrieval augmented generation technology, when such a contradiction occurs, the determination of which information should be prioritized has been left to probabilistic calculations or black-box algorithms inside the large language model.

[0005] Therefore, conventional technologies had the following problems. Firstly, the priority given to conflicts between memory regions was unclear, leading to inconsistencies in the responses of large-scale language models and a lack of reproducibility. For example, unstable behavior was observed where the same question sometimes prioritized long-term memory, and at other times prioritized the previous conversation input. Secondly, there was a lack of means for users and system administrators to explicitly specify priorities. For example, in certain tasks such as legal and medical work, there is a demand to always prioritize the latest regulations, but currently, one is forced to rely on unstable methods such as prompt engineering. Thirdly, when priorities cannot be determined, the large-scale language model may select uncertain information through probabilistic calculations, increasing the risk of hallucination.

[0006] The present invention was made to solve the above-mentioned problems, and aims to provide an information processing method, information processing system, information processing device, and program that enable the stable generation of responses from a large-scale language model by determining the information of the memory area to be referenced by the large-scale language model based on priority control rules when inconsistencies may occur in the information between multiple memory areas. [Means for solving the problem]

[0007] In order to achieve the above objective, An information processing method performed in an information processing system that controls a large-scale language model having multiple memory areas, A contradiction detection step for detecting whether or not there is a contradiction in the information stored in each of the aforementioned multiple memory areas, If the inconsistency detection step detects the inconsistency, a reference destination determination step is performed to determine a reference storage area from the plurality of storage areas based on a priority control rule. The process involves an interaction control step that controls the reference memory area determined in the reference destination determination step as a reference target in the large-scale language model. To provide an information processing method. [Effects of the Invention]

[0008] According to the information processing method of the present invention, when inconsistencies may occur in information between multiple memory areas, the information of the memory area to be referenced by the large-scale language model can be determined based on priority control rules. This avoids the selection of reference targets by probability calculations within the conventional large-scale language model, and since the reference targets are selected based on priority control rules, the response of the large-scale language model can be stabilized.

[0009] Other issues, configurations, and effects will be clarified in the embodiments for carrying out the invention described later. [Brief explanation of the drawing]

[0010] [Figure 1] This is a schematic diagram showing an example of information processing system 1. [Figure 2] This is a block diagram showing an example of the LLM server 20 in the first embodiment. [Figure 3] This is a hardware configuration diagram showing an example of the Computer 900. [Figure 4] This is a flowchart showing the processing flow of the information processing method in the first embodiment. [Figure 5] This is a conceptual diagram of a control system that determines the reference storage area based on priority control rules. [Figure 6] This is a block diagram showing an example of the LLM server 20 in the second embodiment. [Figure 7] This is a flowchart showing the processing flow of the information processing method in the second embodiment. [Modes for carrying out the invention]

[0011] Before providing a detailed description of the present invention, a general explanation of the common behavior of learning models, as well as the basic concepts, configuration, and operation related to their use, will be given with reference to the drawings. This will deepen the understanding of the specific embodiments for carrying out the present invention, which will be described later. In the following, the scope necessary for explaining how to achieve the objectives of the present invention will be schematically shown, and the scope necessary for explaining the relevant parts of the present invention will be mainly explained, with any parts that are omitted from explanation being based on prior art.

[0012] (Overview of Information Processing System 1) Figure 1 is a schematic diagram showing an example of information processing system 1. Information processing system 1 functions as a system that performs information processing such as document generation, dialogue processing, and reasoning support using a large-scale language model 40.

[0013] The information processing system 1, as shown in Figure 1, comprises a user terminal 10, an LLM server 20, external storage 30, and a large-scale language model 40. Each of the devices 10 to 40 is, for example, a general-purpose or dedicated computer (see Figure 3 below) and is connected to a wired or wireless network N, enabling the mutual transmission and reception of various types of data. The number of each device 10 to 40 and the connection configuration of the network N are not limited to the example in Figure 1 and may be changed as appropriate.

[0014] The user terminal 10 is an information processing device in which the user uses a Large Language Model (LLM) 40 to input various instructions and perform interactive operations. The user terminal 10 has the function of sending prompts to the LLM server 20, receiving responses generated by the LLM server 20, and outputting information such as display and audio playback. It also has the function of sending predetermined information to the LLM server 20 based on the user's operations.

[0015] The user terminal 10 may be an information processing terminal in any form, such as a desktop or notebook personal computer, a smartphone, a tablet terminal, a wearable terminal (such as smart glasses), or a VR / AR device, etc.

[0016] The LLM server 20 is an information processing device that provides an environment for the large language model 40 to execute inference processing based on the prompt sent from the user terminal 10. The LLM server 20 may execute the large language model 40 within its own device, or may function by calling an external model providing API. Also, the LLM server 20 manages the storage area referred to by the large language model 40 and plays a role of providing it to the large language model 40 during inference. The LLM server 20 includes a plurality of storage areas for appropriately managing external work products. The specific configuration of the plurality of storage areas will be described later with reference to FIG. 2.

[0017] The external storage 30 is an external storage device physically or logically separated from the large language model 40, and for example, it may be implemented as a cloud storage, a storage area of another server device, a network-connected storage (NAS), a storage area provided in a local device (e.g., encrypted storage, etc.), or a dedicated data vault area.

[0018] Note that "external" in the present disclosure does not necessarily mean physically separate enclosures. The external storage 30 may be a storage area (e.g., a separate partition or a separate volume, etc.) that is logically separated while being within the same enclosure as the LLM server 20.

[0019] The large-scale language model 40 is a pre-trained model (e.g., a neural network model) that has learned from a vast amount of text data, and functions as the execution unit for natural language processing in the information processing system 1. The large-scale language model 40 includes, for example, a transformer-type generative model (GPT® model, LLAMA model, etc., or an alternative architecture) with billions to hundreds of billions of parameters, and performs advanced language processing such as text generation, dialogue response, summarization, translation, and information extraction.

[0020] The prompt sent from the user terminal 10 is formatted as an inference request in the LLM server 20 and then input to the large-scale language model 40. The large-scale language model 40 generates a response to the input based on the information in the memory area determined based on the priority control rules described later. This ensures a consistent interaction experience with the user. The large-scale language model 40 may consist of a single model, or it may be configured as multiple models (e.g., an inference model and a summarization model) working together.

[0021] In this specification, a session refers to a logical unit of communication and interaction established between the user terminal 10 and the large-scale language model 40. Specifically, one session is defined as the period from the instruction to launch the application, log in, or start an interaction, to log out, terminate the application, or perform an explicit reset operation (clear the context).

[0022] Network N is a communication network for data communication between devices. Network N consists of the Internet, LAN, WAN, mobile communication network (5G, 6G, etc.), or a combination thereof, and can be wired or wireless. In cases where the LLM server 20 and the large-scale language model 40 are implemented within the same device, some communication may be replaced by internal bus connections or inter-memory transfers within the device.

[0023] When inconsistencies may arise in the information between multiple memory areas, the information processing system 1 can determine the information of the memory area to be referenced by the large-scale language model 40 based on priority control rules. This avoids the large-scale language model 40 selecting reference targets based on probability calculations, and instead selects reference targets based on priority control rules, thereby stabilizing the response of the large-scale language model 40. The configuration and processing details of the LLM server 20 of the information processing system 1 are described below.

[0024] <First Embodiment> (LLM Server 20 Configuration) Figure 2 is a block diagram showing an example of an LLM server 20. The LLM server 20 includes a control unit 21 for performing inference processing using a large-scale language model 40 and various other processing functions, and a storage unit 22 for holding various information. Note that the functional blocks shown in Figure 2 represent a logical configuration and may be implemented on a single piece of hardware or distributed across multiple pieces of hardware.

[0025] The control unit 21 includes a contradiction detection unit 211, a reference destination determination unit 212, a dialogue control unit 213, and an alert generation unit 214. The control unit 21 executes a series of control processes in the LLM server 20.

[0026] The inconsistency detection unit 211 compares information recorded in multiple memory areas, as described later, and detects whether or not there are inconsistencies in their contents. Specifically, the inconsistency detection unit 211 may use the internal inference function, semantic understanding function, or consistency judgment function inherent in the large-scale language model 40, or it may use an external algorithm independent of the large-scale language model 40 or a lighter identification model.

[0027] In this invention, "contradiction" is defined as a state of inconsistency or lack of coherence. For example, a contradiction is determined to exist not only when the information is simply different (e.g., two different facts, A and B), but also when it includes information that cannot be simultaneously true (e.g., "A is true" and "A is not true").

[0028] When a contradiction is detected by the contradiction detection unit 211, the reference destination determination unit 212 determines, from among multiple memory areas, which the large-scale language model 40 should preferentially refer to when generating a response, based on a priority control rule that is maintained in at least one of the first memory area and the second memory area, as described later.

[0029] Furthermore, the reference determination unit 212 has the function of updating or disabling the currently applied priority control rules for the current session only, based on the user's interactive input during the ongoing session. "Updating" the priority control rules refers to the process of temporarily rewriting the priority, for example, when the user instructs through interactive input, "For this session, set the newly uploaded priority control rules." "Disabling" the priority control rules refers to the process of removing all currently applied priority control rules, for example, when the user instructs through interactive input, "For this session, disable the priority control rules." This allows the user to perform exceptional verification based on a specific hypothesis for the ongoing session only, without modifying the existing priority control rules. When the session ends, the existing priority control rules will be applied.

[0030] The dialogue control unit 213 is a functional unit that sets the reference memory area determined by the aforementioned reference destination determination unit 212 as the reference target of the large-scale language model 40, and performs response generation based on the information in the reference memory area. Specifically, the dialogue control unit 213 combines the text information contained in the reference memory area with the user's input prompt and causes the large-scale language model 40 to perform inference. Furthermore, if the contradiction detection unit 211 does not detect a contradiction, the dialogue control unit 213 causes the large-scale language model 40 to perform normal response generation.

[0031] Furthermore, the basic implementation of the technology (RAG) in which the large-scale language model 40 in the dialogue control unit 213 retrieves relevant information from memory and uses it as the basis for response generation is known to those skilled in the art, and therefore a detailed explanation is omitted in this specification.

[0032] The alert generation unit 214 generates alert information that includes wording indicating a contradiction between information in multiple memory areas when a contradiction is detected by the contradiction detection unit 211. For example, the alert generation unit 214 generates alert information such as, "*A contradiction exists between multiple memory information." The alert information may also include a statement indicating that one of the pieces of information was adopted based on a priority control rule, for example, "*Because there is a contradiction between the company regulations and the user's dialogue input, the response prioritized the user's dialogue input according to the priority control rule."

[0033] The memory unit 22 is configured to have multiple memory areas for holding various context information that the LLM server 20 references when performing inference processing. In this embodiment, the memory unit 22 is logically or physically divided into a long-term memory 221 that holds information across multiple sessions and a short-term memory 222 that is immediately used in ongoing interactions.

[0034] The long-term memory 221 is an example of the first memory area in the claims, and is a memory area that permanently or long-term retains information across multiple dialogue sessions. The information in the long-term memory 221 is not erased even when a session ends and can be referenced in the next session. For example, by storing the information "My favorite character is A" as a user setting in the long-term memory 221, the information "My favorite character is A" can be retained even when the session switches.

[0035] Short-term memory 222 is an example of the second memory area in the claims, which is generated or referenced for each session and holds information that should be processed immediately within the session. The information in short-term memory 222 is discarded when the session ends, or is automatically removed from reference in the next session. For example, by storing the information "My favorite character is B" in short-term memory 222 during a session, the information "My favorite character is B" can be held only within that session and is erased when the session ends.

[0036] The short-term memory 222 may include a session-dedicated area 2221 and a user input-dedicated area 2222, and may store each piece of information separately. In this case, the session-dedicated area 2221 is an example of the second storage area in the claims, and the user input-dedicated area 2222 is an example of the third storage area in the claims.

[0037] The session-dedicated area 2221 is a memory area that holds information other than the information held in the user input-dedicated area 2222, which will be described later. For example, it holds external artifacts in a predetermined format or information read from the external storage 30. The basic configuration is the same as that of the short-term memory 222 described above, so the details are omitted.

[0038] The user input area 2222 is a memory area that holds the conversation history and the most recent prompt information entered by the user in real time during an ongoing session. The information in the user input area 2222 is discarded at the end of the session, or is automatically removed from reference in the next session, similar to the short-term memory 222 described above. For example, if a user interactively inputs "My favorite character is C" during an ongoing session, this information is stored in the user input area 2222. This allows the system to retain the most recent information, "My favorite character is C," only within that session, and it is deleted when the session ends.

[0039] The short-term memory 222 may hold information that needs to be processed immediately within the session, such as the user's most recent utterance, recent dialogue history, the purpose of the current dialogue, or target data read from external storage 30; auxiliary information carried over from past sessions; and data that remains without being erased at the end of the previous session (so-called cache).

[0040] Here, the first to third memory areas each function as independent sources of information, and their contents may contradict each other. For example, a situation may arise where the long-term memory 221 is set to "My favorite character is A," while the session-specific area 2221 is written to "My favorite character is B." In addition, a situation may arise where the user claims "My favorite character is C" in the user interaction input. The present invention assumes such contradictions between memory areas with different logical hierarchies and determines which one to reference in the large-scale language model 40 using priority control rules described later.

[0041] (Priority control rules) In this embodiment, priority control rules refer to declarative guidance information that specifies which memory area's information should be referenced by the large-scale language model 40 when a contradiction in the information between multiple memory areas is detected. These priority control rules are composed of information that describes priority relationships, such as "prioritize B over A" or "prioritize A above all else," and are composed of, for example, conditional branching, natural language, or structured data, similar to conventional programs. The following four types of priority control rules are given as examples. These can be applied individually or in combination.

[0042] (1) The rule that prioritizes the most recent information presented within the session: This is the default rule that is applied when no explicit priority is specified. Typically, this corresponds to the general contextual judgment criteria inherent in large language models.

[0043] (2) Rules that prioritize user dialogue input: These rules are dynamically specified by the user's dialogue input during the session. For example, the user's dialogue input, "My favorite character is C," stored in the user input-only area 2222, is set as the highest priority rule.

[0044] (3) Rule prioritizing information in the first memory area: This rule is applied permanently to a specific user account. For example, a user setting instruction stored in long-term memory 221, such as "My favorite character is A," is set as the rule with the highest priority. In this case, for example, by defining in the company's compliance policy that "In all cases, information in internal regulations (long-term memory 221) shall take precedence over external input information (session-only area 2221)," it is possible to prevent hallucination or unintended guidance by external input information while complying with the internal regulations.

[0045] (4) Rules that prioritize information in the second memory area: These are rules that are contained within the documents or external data themselves that are loaded during the session. For example, the instruction resulting from loading the document "My favorite character is B" stored in the session-specific area 2221 is considered the rule that takes precedence. In this case, for example, by defining "The definition in the contract (session-specific area 2221) takes precedence over the definition in general laws and regulations," it can function as a local rule that is valid only in that contract.

[0046] The four types described in (1) to (4) above are merely preferred examples of priority control rules in this embodiment, and the priority control rules of the present invention are not limited to these. For example, rules based on user attributes and authority levels, rules based on the confidence score of information, rules based on the specificity and amount of description of information may be adopted. Any rule based on any conditions or parameters that allows the reference storage area to be determined from multiple storage areas is included in the priority control rules of the present invention.

[0047] Furthermore, the priority control rules in this embodiment are stored in at least one of the first memory area (e.g., long-term memory 221) and the second memory area (e.g., the session-only area 2221 of the short-term memory 222). This makes it possible to flexibly change the priority of references in the large-scale language model 40 simply by updating the text or configuration file that defines the priority control rules in the memory area, without modifying the program.

[0048] As described above, with the configuration of the LLM server 20, if inconsistencies may arise in the information between multiple memory areas, the information of the memory area to be referenced by the large-scale language model 40 can be determined based on priority control rules. This avoids the large-scale language model 40 selecting reference targets based on probability calculations, and instead selects reference targets based on priority control rules, thereby stabilizing the response of the large-scale language model 40.

[0049] According to the alert generation unit 214, if a contradiction is detected by the contradiction detection unit 211, alert information can be generated and presented to the user. This makes it clear that the response generated by the large-scale language model 40 is not the only correct answer, but rather the result selected based on priority control rules. This reduces the business risks associated with blindly accepting the response of the large-scale language model 40 and improves reliability in collaborative work with humans.

[0050] (Hardware configuration of each device) Figure 3 is a hardware configuration diagram showing an example of a computer 900 as an information processing device. The user terminal 10, LLM server 20, external storage 30, and large-scale language model 40 in the information processing system 1 are configured using general-purpose or dedicated computers 900.

[0051] As shown in Figure 3, the computer 900 comprises, as its main components, a bus 910, a processor 912, memory 914, an input device 916, an output device 917, a display device 918, a storage device 920, a communication interface unit 922, an external device interface unit 924, an I / O device interface unit 926, and a media input / output unit 928. Note that the above components may be omitted as appropriate depending on the intended use of the computer 900.

[0052] The processor 912 consists of one or more arithmetic processing units (CPU (Central Processing Unit), MPU (Micro-processing unit), DSP (digital signal processor), GPU (Graphics Processing Unit), etc.) and operates as a control unit that oversees the entire computer 900. The memory 914 stores various data and programs 930 and consists of volatile memory (DRAM, SRAM, etc.) that functions as main memory, and non-volatile memory (ROM), flash memory, etc.

[0053] The input device 916 consists of, for example, a keyboard, mouse, numeric keypad, electronic pen, microphone, etc., and functions as an input unit. The output device 917 consists of, for example, a sound (voice) output device, a vibration device, etc., and functions as an output unit. The display device 918 consists of, for example, a liquid crystal display, organic EL display, electronic paper, projector, etc., and functions as an output unit. The input device 916 and the display device 918 may be configured as an integrated unit, such as a touch panel display. The storage device 920 consists of, for example, an HDD, SSD, etc., and functions as a storage unit. The storage device 920 stores various data necessary for the execution of the operating system and program 930.

[0054] The communication I / F unit 922 is connected to a network 940 such as the Internet or an intranet by wire or wireless, and functions as a communication unit that sends and receives data with other computers according to a predetermined communication standard. The external device I / F unit 924 is connected to external devices 950 such as cameras, printers, scanners, and reader / writers by wire or wireless, and functions as a communication unit that sends and receives data with external devices 950 according to a predetermined communication standard. The I / O device I / F unit 926 is connected to I / O devices 960 such as various sensors and actuators, and functions as a communication unit that sends and receives various signals and data with the I / O devices 960, such as detection signals from sensors and control signals to actuators. The media input / output unit 928 is composed of a drive device such as a DVD drive or CD drive, and reads and writes data to media (non-temporary storage medium) 970 such as DVDs and CDs.

[0055] In the computer 900 having the above configuration, the processor 912 calls and executes the program 930 stored in the storage device 920 in the memory 914, and controls various parts of the computer 900 via the bus 910. The program 930 may also be stored in the memory 914 instead of the storage device 920. The program 930 may be recorded on the media 970 in an installable or executable file format and provided to the computer 900 via the media input / output unit 928. The program 930 may also be provided to the computer 900 by downloading it via the network 940 through the communication interface unit 922. Furthermore, the computer 900 may implement the various functions realized by the processor 912 executing the program 930 using hardware such as an FPGA or ASIC.

[0056] Computer 900 is an electronic device of any form, consisting of, for example, a stationary computer or a portable computer. Computer 900 may be a client computer, a server computer, a cloud computer, or an embedded computer such as a control panel or controller (including microcontrollers, programmable logic controllers, and sequencers).

[0057] (Processing flow of Information Processing System 1) Figure 4 is a flowchart showing the processing flow of the information processing method in the first embodiment. As an example, the contradiction detection step (S101), reference destination determination step (S103), dialogue control step (S104), and alert generation step (S105) are executed. Note that the execution order of these steps is not necessarily limited to this order. The information processing system 1 may execute them in parallel or in an event-driven manner.

[0058] First, in the inconsistency detection step (S101), the inconsistency detection unit 211 acquires information related to response generation from multiple memory areas and determines whether or not there is an inconsistency between their contents. If an inconsistency is detected (S101: Yes), the process proceeds to step S102; if no inconsistency is detected (S101: No), the process proceeds to step S106.

[0059] Next, in step S102, the reference determination unit 212 checks whether a priority control rule to be applied to the first or second storage area is stored, and if a priority control rule is stored, it reads out the priority control rule.

[0060] Next, in the reference location determination step (S103), the reference location determination unit 212 determines, based on the read priority control rule, the reference memory area that should be preferentially referenced as the basis for the response from among multiple conflicting memory areas. For example, if the priority control rule "prioritize user input" is applied, the information in the user input-only area 2222 is determined as the reference memory area, rather than the information in the long-term memory 221. This avoids the selection of the reference target by probability calculation of the large-scale language model 40, and the reference target is identified based on the priority control rule.

[0061] Next, in the dialogue control step (S104), the dialogue control unit 213 inputs the information of the reference memory area determined in the reference destination determination step (S103) into the large-scale language model 40 as a valid context and generates a response based on that information. At this time, information from memory areas that are not given priority is either excluded from the context or explicitly weighted down as information that should be ignored.

[0062] Next, in the alert generation step (S105), the alert generation unit 214 generates alert information that includes at least wording indicating that there is a discrepancy in the information between multiple storage areas.

[0063] The response generated by the dialogue control process (S104) and the alert information generated by the alert generation process (S105) are output to the user terminal 10. The information may be output together, or one may be output first and the other later.

[0064] On the other hand, if the inconsistency detection step (S101) does not detect an inconsistency, in step S106, the dialogue control unit 213 performs normal reference control, that is, inputs the retrieved information as context into the large-scale language model 40 and performs general response generation processing.

[0065] As described above, in the case where inconsistencies may occur in the information between multiple memory areas, the information processing system 1 of this embodiment can determine the information of the memory area to be referenced by the large-scale language model 40 based on priority control rules. This avoids the selection of reference targets by probability calculations within the conventional large-scale language model, and since the reference targets are selected based on priority control rules, the response of the large-scale language model 40 can be stabilized.

[0066] Furthermore, the information processing system 1 of this embodiment can generate alert information and present it to the user when a contradiction is detected in the alert generation process. This makes it clear that the response generated by the large-scale language model 40 is not the only correct answer, but rather the result selected based on priority control rules. This reduces the business risks associated with blindly accepting the response of the large-scale language model 40 and improves reliability in collaborative work with humans.

[0067] Furthermore, the priority control rules in this embodiment have the function of updating or disabling the currently applied priority control rules for the session only, based on interactive input from the user during the session. This allows the user to perform exceptional verification based on a specific hypothesis for the ongoing session only, without modifying the existing priority control rules.

[0068] Figure 5 is a conceptual diagram of the control system that determines the reference storage area based on priority control rules. A specific example of the control system based on priority control rules will be explained with reference to Figure 5.

[0069] The inconsistency detection unit 211 determines whether or not there is an inconsistency in the information held between the first memory area and the second memory area. In Figure 5, an inconsistency exists in the information held between the first memory area and the second memory area.

[0070] The reference destination determination unit 212 refers to the priority control rule. In Figure 5, the priority control rule includes a condition stating that "information from the first memory area is given priority." Therefore, the reference destination determination unit 212 determines the first memory area as the reference memory area.

[0071] The dialogue control unit 213 generates a response by referring to the information stored in the first memory area determined by the reference destination determination unit 212.

[0072] In the example shown in Figure 5, the control between two memory areas, a first memory area and a second memory area, was described. Specifically, this could be a combination of the first memory area being a long-term memory 221 and the second memory area being a short-term memory 222, but it is not limited to this, and control between any two memory areas is also possible.

[0073] Alternatively, the control may be between three memory areas: a first memory area, a second memory area, and a third memory area. Specifically, this could be a combination where the first memory area is a long-term memory 221, the second memory area is a session-dedicated area 2221 of the short-term memory 222, and the third memory area is a user input-dedicated area 2222 of the short-term memory 222, but it is not limited to this, and control may be between any three memory areas.

[0074] Furthermore, control may be extended to four or more memory areas. In this case, in addition to the memory areas mentioned above, a fourth memory area independent of the other memory areas may be treated and controlled, for example, a memory area that stores the AI's past statements during a session, or new memory areas with different time axes and uses that may be added in the future (for example, medium-term memory that is only valid for a specific project period of several days to several weeks, emotional memory linked to the user's emotional state, or search results dynamically obtained via an external API).

[0075] <Second Embodiment> Next, a second embodiment of the present invention will be described. In the first embodiment, a system was described in which a priority control rule is applied when a contradiction is detected to generate a response. In contrast, the second embodiment differs in that, when a priority control rule is not set or when the rule is disabled by user instruction, it forcibly prohibits the generation of a response by selecting a reference target through probability calculation of the large-scale language model 40 and outputs an error.

[0076] (LLM Server 20 Configuration) Figure 6 is a block diagram showing an example of the LLM server 20 in the second embodiment. In addition to the configuration of the first embodiment, it is further configured to include a prohibition unit 215 and an error generation unit 216.

[0077] First, the reference determination unit 212, similar to the first embodiment, has the function of updating or disabling the currently applied priority control rule for that session only, based on interactive input from the user of the ongoing session.

[0078] The prohibition unit 215 prohibits the large-scale language model 40 from selecting multiple conflicting memory regions as reference targets when a contradiction is detected by the contradiction detection unit 211 and the priority control rule is not set or is disabled by the reference target determination unit 212. This prevents the large-scale language model 40 from exhibiting black-box behavior, such as "choosing the seemingly more likely option without any basis for judgment."

[0079] The error generation unit 216 generates error information that includes at least wording indicating a contradiction in the information between multiple memory areas when the prohibition process by the prohibition unit 215 is executed. For example, the error generation unit 216 generates error information such as, "Cannot respond due to a contradiction. Please specify a priority." This error information is not merely a system error, but clearly indicates a logical deadlock in the large-scale language model 40. Therefore, the user can confirm that the large-scale language model 40 is not selecting reference targets based on probability calculations, and can confidently give the next instruction.

[0080] As described above, with the configuration of the LLM server 20, if inconsistencies may occur in the information between multiple storage areas, and if a priority control rule is not set or the priority control rule is disabled, the large-scale language model 40 can be prevented from selecting multiple inconsistent storage areas and generating a response. This ensures a fail-safe function that guarantees safe behavior by deliberately not responding when no effective priority control rule exists.

[0081] Figure 7 is a flowchart showing the processing flow of the information processing method in the second embodiment. Since steps S201, S203-S205, and S208 in Figure 7 are the same as steps S101, S103-S105, and S106 in Figure 4 described above, a detailed explanation will be omitted, and the explanation will focus on the processing specific to this embodiment (steps S202, S206, and S207). As an example, the inconsistency detection step (S201), reference destination determination step (S203), prohibition step (S206), and error generation step (S207) are executed. Note that the execution order of these steps is not necessarily limited to this order. The information processing system 1 may execute these in parallel or in an event-driven manner.

[0082] In the inconsistency detection step (S101), if the inconsistency detection unit 211 detects an inconsistency (S201: Yes), in step S202, the reference destination determination unit 212 determines whether or not a valid priority control rule exists.

[0083] In step S202, the determination result is "Yes" if a priority control rule is set and no user has issued a disabling instruction. In this case, the process proceeds to step S203 and proceeds as described in the first embodiment.

[0084] On the other hand, in step S202, the judgment result is "No (no valid rules)" if no priority control rules are set, or if priority control rules are set but have been disabled by user interaction input.

[0085] Next, in the prohibition step (S206), if no priority control rule is set, or if a priority control rule is set but has been disabled by user interaction input (S202: No), the prohibition unit 215 prohibits the large-scale language model 40 from selecting one of the conflicting memory areas. Alternatively, the prohibition unit 215 may perform control only if a priority control rule is set but has been disabled by user interaction input.

[0086] Next, in the error generation step (S207), the error generation unit 216 generates error information that includes at least wording indicating a contradiction in the information between multiple memory areas. The error information generated in the error generation step (S207) is output to the user terminal 10.

[0087] As described above, in this embodiment, the information processing system 1 can prevent the large-scale language model 40 from selecting multiple conflicting memory areas and generating a response if a conflict may occur in the information between multiple memory areas, and if a priority control rule is not set or the priority control rule is disabled. This enables a fail-safe function that ensures safe behavior by deliberately not responding when no effective priority control rule exists.

[0088] <Other Embodiments> This disclosure is not limited to the embodiments described above, and can be implemented with various modifications without departing from the spirit of this disclosure. All such modifications are included in the technical concept of this disclosure.

[0089] The large-scale language model 40 does not necessarily need to be configured as a device independent of the LLM server 20. For example, the large-scale language model 40 may be integrated into the LLM server 20, and inference processing and context management processing may be executed together on the same computing resources. Conversely, the large-scale language model 40 may be placed on an external cloud inference platform or distributed learning platform, and the LLM server 20 may only perform context information management and prompt formatting processing. Furthermore, an on-device AI configuration may be implemented in which a lightweight language model (SLM) is installed inside the user terminal 10 (smartphone, PC, edge device, etc.), and inference processing and context management are completed in the local environment, or a hybrid configuration may be implemented in which only sensitive processing is performed locally.

[0090] Although embodiments of the present invention have been described above, the information processing system according to the present invention is not limited to being provided as a chat application (BtoC service) with a screen that end users can directly operate. For example, the LLM server 20 shown in Figure 1 may be configured as a backend infrastructure system equipped with memory governance functions (inconsistency detection, priority control, fail-safe, etc.), and may be provided to external client systems (third-party application servers, etc.) via an interface such as a web API (Application Programming Interface) as a cloud service (SaaS or PaaS, etc.).

[0091] In this API provisioning model, the external client system stores context information corresponding to the aforementioned first, second, and third storage areas within the parameters or payload (e.g., data in JSON format) of the API request, and attaches the priority control rules to be applied as metadata, before sending it to the LLM server 20. The control unit 21 of the LLM server 20 (inconsistency detection unit 211, reference destination determination unit 212, prohibition unit 215, etc.) parses the data structure of the received API request and performs inconsistency detection and priority control processing between each storage area. Then, it returns the generated response text, accompanying alert information, or error information status as an API response to the client system.

[0092] Thus, even in a loosely coupled system architecture where multiple memory areas and priority control rules of the present invention are dynamically injected from an external system via a network as API requests, and the control results are returned as API responses, the technical scope of the present invention is still included.

[0093] Furthermore, the information processing system 1 may be applied not only to a single agent (user vs. LLM) but also to a multi-agent system in which multiple autonomous AI agents cooperate to perform tasks.

[0094] In the embodiments described above, "multiple memory areas" refers to multiple logically distinct information sources that can be referenced as context when a large-scale language model generates a response. These do not necessarily need to be physically separated into different storage media (such as HDDs and SSDs); they only need to be logically separated by data retention period, access rights, or the source of the information.

[0095] The embodiments described above illustrate an example of detecting inconsistencies using the inference capabilities of the large-scale language model itself, but the present invention is not limited thereto. The inconsistency detection unit 211 may be implemented using an external algorithm independent of the LLM or a lighter discrimination model.

[0096] For example, the contradiction detection unit 211 may use vector search technology (Embeddings) to calculate the semantic vectors of sentences contained in multiple memory areas, and determine a "contradiction" if their cosine similarity is above a predetermined threshold (i.e., the topics are the same) and the affirmative and negative attribute values ​​are different. This makes it possible to detect conflicts between memories at high speed while suppressing the token consumption of the LLM.

[0097] Furthermore, the contradiction detection unit 211 may be configured as a rule-based program based on predetermined keyword matching or symbolic logic. For example, it may be configured to mechanically set a contradiction flag when a statement containing the negative word "is not A" exists in a different memory area in response to the assertion "is A".

[0098] Thus, the specific means of inconsistency detection can be either probabilistic determination by a neural network or deterministic determination by a program; any known method can be employed as long as it has the function of identifying inconsistencies in content among multiple sources.

[0099] Furthermore, the present invention can be implemented not only as the information processing system 1 described above, but also as a program for a computer to execute it, or as a computer-readable recording medium that stores the program. That is, each function of the control unit 21 shown in Figure 2 or Figure 6 may be implemented by a processor such as a CPU reading and executing a software program stored in memory.

[0100] While several embodiments of this disclosure have been illustrated above, these embodiments are presented as examples only and are not intended to limit the scope of the invention. These novel embodiments can be implemented in various other forms, and various omissions, substitutions, and modifications can be made without departing from the spirit of the invention. These embodiments and their variations are included in the scope and spirit of the invention, as well as in the claims and their equivalents. Furthermore, the embodiments described above can be implemented in combination with each other. [Explanation of Symbols]

[0101] 1... Information processing system, 10... User terminal, 20... LLM server, 21...Control unit, 211...Inconsistency detection unit, 212...Reference destination determination unit, 213...Dialogue control unit, 214… Alert generation unit, 215… Prohibition unit, 216… Error generation unit, 22...Memory unit, 221...Long-term memory, 222...Short-term memory, 2221...Session-only area, 2222...User input-only area, 30…External storage, 40…Large-scale language models

Claims

1. An information processing method performed in an information processing system that controls a large-scale language model having multiple memory areas, A contradiction detection step that detects whether or not there is a contradiction in the information stored in each of the aforementioned multiple memory areas, If the inconsistency detection step detects the inconsistency, a reference destination determination step is performed to determine a reference storage area from the plurality of storage areas based on a priority control rule. The process involves an interaction control step that controls the reference memory area determined in the reference destination determination step as a reference target in the large-scale language model. Information processing methods.

2. The plurality of storage areas comprises at least a first storage area and a second storage area that is logically or physically separated from the first storage area. The first memory area is, It is a long-term memory area that holds information across multiple sessions. The second memory area described above is It is a short-term memory area that holds information only within the ongoing session. The aforementioned session, The aforementioned large-scale language model and the user are a series of interaction units, The information processing method according to claim 1.

3. The plurality of memory areas further include a third memory area that holds information for user interaction input. The information processing method according to claim 2.

4. The aforementioned priority control rule is: A rule that prioritizes the latest information presented within the session, a rule that prioritizes the information from the user's interactive input, a rule that prioritizes the information from the first memory area, or a rule that prioritizes the information from the second memory area. The information processing method according to claim 2 or claim 3.

5. The aforementioned information processing method further, If the inconsistency detection step detects the inconsistency, an alert generation step is performed to generate alert information that includes wording indicating that there is an inconsistency in the information between the multiple storage areas. The information processing method according to claim 1.

6. The aforementioned priority control rule is: This is information describing the priority relationships between the aforementioned multiple memory areas. The information processing method according to claim 1.

7. The aforementioned priority control rule is: The data is stored in at least one of the first and second storage areas. The information processing method according to claim 2 or claim 3.

8. The aforementioned reference determination step further includes: Based on user interaction input during the aforementioned session, the priority control rules may be disabled or updated for that session only. The information processing method according to claim 2 or claim 3.

9. The aforementioned information processing method further, If the inconsistency detection step detects the inconsistency, and the priority control rule is not set or the priority control rule is disabled, a prohibition step is made to prohibit selecting the inconsistent storage area as the reference target. The process involves generating an error that indicates that control over the inconsistent memory area is impossible, and performing an error generation step. The information processing method according to claim 1.

10. The error information includes wording indicating that there is a discrepancy in the information between the multiple storage areas, The information processing method according to claim 9.

11. The aforementioned reference determination step is, The user's interactive input during a session allows for the disabling or updating of the priority control rules, but only for that session. The aforementioned prohibited step is, This applies when the priority control rule is disabled by the user's interactive input. The information processing method according to claim 9 or claim 10.

12. An information processing system for controlling a large-scale language model having multiple memory areas, A contradiction detection unit that detects whether or not there is a contradiction in the contents stored in each of the plurality of storage areas, When the inconsistency detection unit detects the inconsistency, the reference destination determination unit determines a reference storage area from the plurality of storage areas based on the priority control rule, The system includes an interaction control unit that controls the reference storage area determined by the reference destination determination unit as a reference target in the large-scale language model, Information processing system.

13. An information processing device included in an information processing system for controlling a large-scale language model having multiple memory areas, A contradiction detection unit that detects whether or not there is a contradiction in the contents stored in each of the plurality of storage areas, When the inconsistency detection unit detects the inconsistency, the reference destination determination unit determines a reference storage area from the plurality of storage areas based on the priority control rule, The system includes an interaction control unit that controls the reference storage area determined by the reference destination determination unit as a reference target in the large-scale language model, Information processing device.

14. A program for causing an information processing device, which comprises a processor and multiple memory areas, to execute a process for controlling a large-scale language model, The aforementioned processor, The steps include detecting whether or not there is a contradiction in the contents stored in each of the aforementioned multiple memory areas, If the aforementioned inconsistency is detected, the step of determining a reference storage area from the plurality of storage areas based on the priority control rule, The steps of controlling the determined reference memory area as a reference object in the large-scale language model are performed. program.

Images