Knowledge extraction system and knowledge extraction method

Abstract

We provide technologies that can flexibly respond to diverse demands and changes in the environment and needs. [Solution] The AI ​​agent system 100 includes an episodic memory database 310 that stores the history of interactions with the user as episodic memory containing multiple messages, a semantic memory construction unit 510 that clusters the messages contained in the episodic memory into multiple clusters and extracts scenario branches based on the transitions between message clusters, and a semantic memory database 320 that stores the scenario branches as semantic memory.

Description

Technical Field

[0001] The present invention relates to a knowledge extraction system and a knowledge extraction method.

Background Art

[0002] Based on the decrease in the working population and labor shortages, task substitution by robots and virtual agents has been progressing. Such agents have been applied to relatively simple tasks such as transportation, cleaning, and security, but with the development of generative AI (LLM), their use has also spread to more complex and natural interaction tasks such as customer service guidance.

[0003] To make an AI agent behave appropriately, adjustment work such as prompt engineering is necessary. However, when responding to diverse needs such as customer service guidance in commercial facilities, it is difficult to prepare a prompt that realizes an appropriate response.

[0004] Patent Document 1 describes a technique for recording the process of successful tasks and calling them as specific examples when performing new tasks for reference. In this technique, high-level insights are extracted through natural language from past task experiences, and application and generalization to new tasks are realized without updating parameters.

Prior Art Documents

Non-Patent Documents

[0005]

Non-Patent Document 1

Summary of the Invention

Problems to be Solved by the Invention

[0006] However, Non-Patent Document 1 manages and searches experience and insights as flat string information, making it difficult to organize and interpret diverse requests from a large number of unspecified users and extract them as usable knowledge. In other words, Non-Patent Document 1 is intended for dealing with a single user or a specific task, and is not suitable for tasks that deal with a large number of unspecified users with diverse needs (tasks), such as in a commercial facility.

[0007] In environments such as commercial facilities, where it is necessary to respond to users with diverse demands, there is a need for AI agents that can continuously learn response methods through experience and flexibly respond to diverse demands and changes in the environment and needs.

[0008] Therefore, the present invention aims to provide a technology that can flexibly respond to diverse demands and changes in the environment and needs. [Means for solving the problem]

[0009] To solve the above problems, one representative knowledge extraction system of the present invention is a knowledge extraction system for accumulating experiential memories, comprising: an episodic memory unit that stores the history of dialogue with the user as an episodic memory containing multiple messages; a semantic memory construction unit that clusters the messages contained in the episodic memory into multiple clusters and extracts scenario branches based on the transitions between message clusters; and a semantic memory unit that stores the scenario branches as semantic memories. [Effects of the Invention]

[0010] According to the present invention, it is possible to flexibly respond to diverse demands and changes in the environment and needs.

[0011] Other issues, configurations, and effects not mentioned above will be clarified by the following description of the embodiments. [Brief explanation of the drawing]

[0012] [Figure 1] This block diagram shows an example of the configuration of an AI agent system to which the knowledge extraction system of this embodiment is applied. [Figure 2] This block diagram shows an example of the hardware configuration of an information processing device to which the knowledge extraction system of this embodiment is applied. [Figure 3] This flowchart shows an example of how the user interface handles customer interactions. [Figure 4] This flowchart shows an example of prompt generation processing by the memory retrieval unit. [Figure 5] This is a flowchart illustrating an example of the semantic memory construction process performed by the semantic memory construction unit. [Figure 6] This figure shows an example of the data structure of an episodic memory database. [Figure 7] This is a schematic diagram illustrating message clustering by the semantic memory construction unit. [Figure 8] This is a schematic diagram illustrating the extraction of scenario branching by the semantic memory construction unit. [Figure 9] This figure shows an example of a data structure for a semantic memory database. [Figure 10] This figure shows an example of a standard user interface screen displayed by the user interface unit. [Figure 11] This figure shows an example of a standard screen of the management interface displayed by the management interface unit. [Figure 12] This figure shows an example of a standard screen of the management interface displayed by the management interface unit. [Modes for carrying out the invention]

[0013] The following examples will be described with reference to the drawings. [Examples]

[0014] FIG. 1 is a block diagram showing an example of the configuration of an AI agent system to which the knowledge extraction system of this embodiment is applied. The AI agent system 100 is installed in devices such as, for example, computers, tablets, signage, robots, etc. In this embodiment, an AI agent system installed in a device installed at an airport or a commercial facility and providing guidance and product explanations through interaction with customers will be described as an example, but the present invention is not limited thereto.

[0015] The AI agent system 100 includes a user interface unit 110, a management interface unit 120, an AI agent unit 210, an episode memory database 310, a semantic memory database 320, a memory search unit 410, and a semantic memory construction unit 510.

[0016] The user interface unit 110 includes a GUI (graphical user interface) and provides functions for a user to input to and receive output from the AI agent system 100.

[0017] The management interface unit 120 provides functions for an administrator of the AI agent system 100 to check the experience memory accumulated in the system and perform management and adjustment as necessary. The experience memory includes episode memory and semantic memory.

[0018] The AI agent unit 210 utilizes AI technologies such as generative AI (large language model: LLM) to provide AI agent services to users.

[0019] The episodic memory database 310 (episodic memory unit) stores the history of user interactions as episodic memory. The user interaction history includes questions and requests from the user, intermediate outputs from the AI ​​agent unit 210, responses to the user, and the user's responses to those responses. In addition to linguistic information, the episodic memory also includes non-linguistic contextual information. Non-linguistic contextual information includes, for example, the situation in which the user is located (e.g., location, weather, season, time of day) and the user's attributes (age, gender, composition of users if there are multiple users).

[0020] Contextual information may be acquired by the device's cameras and sensors, or by an external system via network communication. Alternatively, the user may input contextual information via a GUI, or the user ID may be pre-associated with the contextual information. Based on the contextual information, the AI ​​agent system 100 can generate a response more appropriate to the user.

[0021] The semantic memory database 320 (semantic memory unit) stores semantic memories constructed based on episodic memories. Semantic memory is knowledge that is generally applicable to specific use cases or groups of tasks, or knowledge that is broadly applicable regardless of use cases or tasks. Furthermore, semantic memory includes information about the branching structure of conversation scenarios extracted from multiple episodes.

[0022] The memory retrieval unit 410 searches and extracts the necessary information from the episodic memory database 310 and the semantic memory database 320, and generates a prompt to send to the AI ​​agent unit 210 based on the extracted information.

[0023] The semantic memory construction unit 510 constructs semantic memories that represent more general knowledge based on episodic memories stored in the episodic memory database 310. The semantic memories include information for the AI ​​agent unit 210 to generate scenario branches of episodic memories and refined questions.

[0024] Scenario branching refers to information that indicates how the transitions between user and AI agent unit 210 messages contained in episodic memory branch. For example, let's explain using the episodic memory shown in Figure 10, which will be described later, in which the user asks "Are there any stand-up soba noodle shops?" and the AI ​​agent unit 210 responds "There are authentic soba noodle shops." If, in addition to this, there is an episodic memory in which the user says "I'm in a hurry" and another episodic memory in which the user says "Tell me about authentic soba noodle shops," the semantic memory construction unit 510 extracts this as a scenario branch.

[0025] A refined question is a question designed to narrow down the user's needs in order to provide a response that matches the user's intent. For example, in the above example, the semantic memory building unit 510 generates information for the AI ​​agent unit 210 to generate a refined question, "Confirm whether you are in a hurry or if you prefer soba noodles," based on the scenario branching. Based on this information, the AI ​​agent unit 210 responds to the user's message, "Are there any standing soba noodle shops?" with a refined question such as, "There are soba noodle shops, although they are not standing-only. Alternatively, if you are in a hurry, there are also cafes and food courts."

[0026] In this way, the AI ​​agent system 100 can estimate the user's potential intentions and suggest options by generating scenario branching and refined questions, thereby providing information that meets the user's needs more quickly.

[0027] The knowledge extraction system of this embodiment is implemented, for example, by an information processing device as shown in Figure 2. Figure 2 is a block diagram showing an example of the hardware configuration of an information processing device to which the knowledge extraction system of this embodiment is applied.

[0028] Figure 1 shows an example of the hardware configuration of an information processing device 1000 in which the user interface unit 110, management interface unit 120, AI agent unit 210, memory retrieval unit 410, and semantic memory construction unit 510 shown are in operation. The information processing device 1000 is a server or computer configured by interconnecting a CPU (Central Processing Unit) 1001, memory 1002, storage device 1003, communication unit 1004, input unit 1005, and output unit 1006 via an internal communication path 1007.

[0029] The CPU 1001 is the central processing unit, and it implements the necessary functions by executing programs stored in the memory 1002 (or storage device 1003). The programs include those that realize the user interface unit 110, the management interface unit 120, the AI ​​agent unit 210, the memory retrieval unit 410, and the semantic memory construction unit 510 shown in Figure 1.

[0030] Memory 1002 is the main memory used by the CPU 1001 when executing processing, and is composed of volatile memory elements such as RAM (Random Access Memory).

[0031] The storage device 1003 is an auxiliary storage device for storing input data provided to the CPU 1001 and output data output from the CPU 1001, and is composed of non-volatile memory elements such as an SSD (Solid State Drive). The storage device 1003 also stores data such as the episodic memory database 310 and the semantic memory database 320 shown in Figure 1.

[0032] The communication unit 1004 is an interface used by the information processing device 1000 to communicate with external devices, and consists of a network adapter, a communication module, etc. The communication unit 1004 is connected to a network (e.g., the Internet) and communicates with external devices via the network.

[0033] The input unit 1005 is an interface that receives input from the user (operator) and consists of a keyboard, touch panel, or voice input device (microphone), etc.

[0034] The output unit 1006 is an interface that outputs data to the operator and is composed of a display or an audio output device (speaker), etc.

[0035] The internal communication channel 1007 is a communication path for exchanging data between the various components of the information processing device 1000.

[0036] In this embodiment, the user interface unit 110, the management interface unit 120, the AI ​​agent unit 210, the memory retrieval unit 410, and the semantic memory construction unit 510 are executed on one or more information processing devices 1000 having the hardware configuration illustrated in Figure 2, thereby realizing the processes described later.

[0037] Next, the processes performed by the knowledge extraction system of this embodiment will be explained using Figures 3 to 5.

[0038] Figure 3 is a flowchart showing an example of response processing by the user interface unit 110.

[0039] The user interface unit 110 receives messages entered by the user in natural language (S1101). Messages may be entered by the user through operations on a standard user interface screen, such as the one shown in Figure 10 later, or by voice input from a microphone.

[0040] Subsequently, the received message is added to the episodic memory database 310 and updated (S1102).

[0041] Next, the user interface unit 110 determines whether the message from the user has a keyword indicating completion of the interaction (for example, "processing complete") or a data label indicating completion of the interaction (S1103). If, in S1103, the message from the user has a keyword indicating completion of the interaction or a data label indicating completion of the interaction, this process is completed.

[0042] On the other hand, if, in S1103, the user message does not have a keyword or data label indicating completion of the interaction, the user interface unit 110 accesses the memory retrieval unit 410 and obtains a prompt from the memory retrieval unit 410 to give to the AI ​​agent unit 210 (S1104). This prompt includes samples of past episodes similar to the user message and instructions for generating refined questions. The prompt generation process by the memory retrieval unit 410 will be described later in Figure 4.

[0043] Next, the user interface unit 110 accesses the AI ​​agent unit 210 using a prompt obtained from the memory retrieval unit 410 and obtains a response from the AI ​​agent unit 210 to the user's message (S1105).

[0044] Next, the user interface unit 110 presents the response obtained from the AI ​​agent unit 210 to the user (S1106), and returns to S1101. Here, the user interface unit 110 displays the response obtained from the AI ​​agent unit 210 on a standard user interface screen, such as the one shown in Figure 10, which will be described later.

[0045] Figure 4 is a flowchart showing an example of the prompt generation process by the memory retrieval unit 410.

[0046] The memory retrieval unit 410 receives a request from the user interface unit 110 to generate a prompt to be given to the AI ​​agent unit 210 (S4101). At this time, the memory retrieval unit 410 also receives a message from the user from the user interface unit 110 along with the prompt generation request.

[0047] Next, the memory retrieval unit 410 generates feature quantities based on the user's message using an embedded model, etc., and converts them into a format suitable for similarity search (S4102). The generated feature quantities are linked to the message and stored in the episodic memory database 310.

[0048] Next, the memory retrieval unit 410 uses the generated feature quantities to retrieve semantic memories similar to the user's message from the semantic memory database 320 (S4103). The memory retrieval unit 410 also retrieves episodic memories of past cases similar to the user's message from the episodic memory database 310 (S4104).

[0049] Furthermore, when the memory retrieval unit 410 searches for experiential memories similar to a message from the user, it does not search from the beginning of the conversation, but rather searches backward from the conclusion that satisfied the user. This narrows down the target of message comparison and reduces the computational load.

[0050] Additionally, messages can be weighted based on how frequently they are traversed during the search process. This allows for prioritizing the filtering of less reliable branches. The frequency of traversing a path is updated with each conversation. A feature to change the message weighting may also be included as needed.

[0051] Next, the memory retrieval unit 410 generates a prompt to be given to the AI ​​agent unit 210 using the acquired semantic memory and episodic memory (S4105). At this time, the acquired semantic memory is embedded in the prompt that generates the refinement question, and the acquired episodic memory is embedded in the prompt as a similar example.

[0052] Next, the memory retrieval unit 410 returns the generated prompt to the user interface unit 110 as a response (S4106), and terminates the process.

[0053] Figure 5 is a flowchart illustrating an example of the semantic memory construction process performed by the semantic memory construction unit 510. In this process, the semantic memory construction unit 510 constructs a semantic memory based on multiple episodic memories stored in the episodic memory database 310 and registers it in the semantic memory database 320.

[0054] The semantic memory construction unit 510 receives instructions for semantic memory construction from the management interface unit 120 (S5101). These instructions may be entered by the user through a standard screen of the management interface, such as the one shown in Figure 11, which will be described later, or they may be pre-configured to execute the semantic memory construction process periodically.

[0055] Next, the semantic memory construction unit 510 retrieves all episodic memories from the episodic memory database 310 (S5102).

[0056] Here, we will explain the episodic memory stored in the episodic memory database 310 using Figure 6.

[0057] Figure 6 shows an example of the data structure of the episodic memory database 310. The episodic memory database 310 stores natural language messages entered by the user from the user interface unit 110 and natural language messages obtained from the AI ​​agent unit 210. Episodes 600 are represented in a graph structure (chain structure) that links these messages together based on their temporal order. This structure allows for the storage of a vast amount of data on the context and conversation flow for each episode of an unspecified number of users.

[0058] Furthermore, episodic memories not used to generate prompts or responses may be unusable or irreproducible because they are individual. Therefore, it may be advisable to maintain the quality of such episodic memories by lowering their importance score.

[0059] Each message node 601 that makes up episode 600 stores one natural language message. By recording messages in detail at the node level, the entire dialogue can be saved. In addition, each episode 600 is assigned a unique episode ID 610, which makes it possible to uniquely identify the episode.

[0060] Each message node 601 has a property 620, which includes the index, type, message, and search features.

[0061] The index represents the order and position information of the nodes. The type indicates the nature of the message (e.g., user utterance, response from AI agent unit 210, etc.). The message is the natural language message itself, recording the specific content of the conversation.

[0062] Search features are used for similarity searches within the episodic memory database 310. These search features are numerical data representing the meaning and context of messages, enabling efficient calculation of similarity between messages.

[0063] Furthermore, although not shown in the diagram, by linking nonverbal contextual information to this episodic memory graph and applying convolutional features, it becomes possible to perform similar episode retrieval and response generation that also takes this contextual information into account.

[0064] Returning to Figure 5, at S5103, the semantic memory construction unit 510 clusters the messages that constitute multiple episodes included in the episodic memory using search features. In this case, the semantic memory construction unit 510 does not search for similar messages for each individual message and cluster them, but rather clusters several adjacent messages in a sliding window manner. Furthermore, when traversing adjacent messages, the semantic memory construction unit 510 performs clustering based on the movement trajectory and distance in the embedded space.

[0065] Furthermore, even if the progression of the conversation varies among users, it is believed that there is a correlation between the progression of the conversation and the distance between messages. Therefore, when measuring the similarity of messages, the semantic memory construction unit 510 reduces the computational load by narrowing the range of the similarity search using the distance between messages.

[0066] Furthermore, the semantic memory construction unit 510 performs processes to remove unnecessary parts of the utterances and to summarize the utterances as preprocessing for clustering. The semantic memory construction unit 510 also identifies proper nouns contained in the message, and uses a hierarchical structure of concepts to classify and generalize the proper nouns into abstract categories.

[0067] Here, we will explain message clustering using Figure 7.

[0068] Figure 7 is a schematic diagram showing message clustering by the semantic memory construction unit 510.

[0069] The semantic memory construction unit 510 groups message nodes 601 of multiple episodes 600 stored in the episodic memory database 310 into multiple clusters 701, with message nodes 601 of messages that are similar in content being grouped into the same cluster.

[0070] Phases 1, 3, and 5 are messages from the user, while phases 2, 4, and 6 are responses from the AI ​​agent unit 210. In Figure 7, the semantic memory construction unit 510 clusters the message nodes 601 of the user messages from phases 1, 3, and 5, and groups them into multiple clusters 701.

[0071] Furthermore, the semantic memory construction unit 510 compares the similarity between messages in corresponding phases of multiple episodes, as well as between messages in adjacent phases, and adopts the one with the highest similarity. This enables robust similarity search. Additionally, the similarity between the user's message and the AI ​​agent unit 210's message in the immediately preceding adjacent phase may also be calculated and the weighted average may be used as the message similarity. This allows for the calculation of similarity that takes into account the relationship between the user's response and the AI ​​agent unit 210's response.

[0072] Returning to Figure 5, in S5104, the semantic memory construction unit 510 extracts transitions between clusters as scenario branches based on the clustering results. Here, we will explain the extraction of scenario branches using Figure 8.

[0073] Figure 8 is a schematic diagram showing the extraction of scenario branching by the semantic memory construction unit 510. Figure 8 shows the feature space in which clusters exist in each of phases 1, 3, and 5. In Figure 8, clusters A to D each contain multiple messages that have been clustered.

[0074] In this feature space, the transitions between messages in clusters are observed. Specifically, the semantic memory construction unit 510 analyzes how conversations progress within or between episodes and what kinds of relationships arise between messages belonging to different clusters.

[0075] For example, messages clustered to cluster A in phase 1 transition to either cluster B or cluster C in phase 3. Similarly, messages clustered to cluster C in phase 3 transition to either cluster A or cluster B in phase 5. In this way, the semantic memory construction unit 510 extracts cases where messages belonging to the same cluster transition to different clusters as scenario branches and constructs semantic memory.

[0076] Furthermore, messages clustered to cluster B in phase 3 transition to cluster D in phase 5. Although not shown in the diagram, messages from different clusters may also transition to the same cluster. In such cases, no scenario branch is detected.

[0077] The semantic memory construction unit 510 stores the extracted scenario branches in the semantic memory database 320. This allows the memory retrieval unit 410 to search and extract past scenario branches similar to the user's message from the semantic memory database 320, anticipate the possible branches of the subsequent conversation, and generate prompts.

[0078] Returning to Figure 5, in S5105, the semantic memory construction unit 510 accesses the AI ​​agent unit 210 and obtains information to generate refined questions based on the scenario branching (S5105). For example, based on a scenario branch where the user's message "Are there any standing soba noodle shops?" transitions to the messages "I'm in a hurry" or "Tell me about a proper soba noodle restaurant," the unit generates information such as "Confirm whether they are in a hurry or if they prefer soba noodles."

[0079] Next, the semantic memory construction unit 510 registers the cluster information obtained in S5103, the scenario branching information extracted in S5104, and the information for generating the refined questions obtained in S5105 into the semantic memory database 320 (S5106). After that, it notifies the management interface unit 120 that the semantic memory update process is complete (S5107), and terminates the process. Here, the data structure in the semantic memory database 320 will be explained using Figure 9.

[0080] Figure 9 shows an example of the data structure of the semantic memory database 320. This semantic memory database 320 is used to manage information regarding the classification and transitions of episodic memory messages, and mainly holds information regarding the relationships between clusters and their transitions.

[0081] The semantic memory database 320 stores two main types of information. One is information about each cluster, 3201 and 3202, and the other is information about transitions between clusters, 3203 and 3204.

[0082] Information about each cluster, 3201 and 3202, includes the cluster ID and data on the centroid and variance of each cluster. The centroid and variance represent the distribution of message features within the cluster. This information is used to determine which cluster a new user message belongs to.

[0083] Information regarding transitions between clusters, 3203 and 3204, includes the cluster ID of the source cluster, information about the destination cluster, and information about the refinement question. The destination cluster information includes the cluster ID of the destination cluster and the probability of transitioning to each destination cluster. The refinement question information includes information for generating a refinement question to identify the destination cluster.

[0084] By having the AI ​​agent unit 210 generate these refined questions based on the information used to generate them, the destination of the message can be estimated, and a response that meets the user's needs can be quickly reached.

[0085] If there is a bias exceeding a threshold in the transition probability of the destination cluster, as in the information 3204 regarding transitions between clusters, the semantic memory construction unit 510 determines that a filtering query is unnecessary and does not retain information for the filtering query. In such cases, when generating a prompt, the memory retrieval unit 410 skips the branch of the information 3204 regarding transitions between clusters and identifies the next filtering query by referring to the branch information of cluster D, whose transition probability exceeds the threshold.

[0086] Figure 10 shows an example of a standard screen of the user interface displayed by the user interface unit 110.

[0087] The standard user interface screen 800 is displayed on the screens of devices such as computers, tablets, signage, and robots equipped with the AI ​​agent system 100.

[0088] When a user clicks the microphone icon 801 on the standard screen 800 and then speaks into the microphone provided by the device, the user interface unit 110 transcribes the spoken content and displays it on the standard screen 800 as spoken content 802.

[0089] Furthermore, the user interface unit 110 acquires responses to user speech from the AI ​​agent unit 210, displays them as responses 803 on the standard screen 800, and reads the acquired responses aloud from the device's speaker.

[0090] Figure 11 shows an example of a standard screen of the management interface displayed by the management interface unit 120.

[0091] The standard screen 900 of the management interface is displayed on the display of the device on which the AI ​​agent system 100 is installed, or on the display of an external device such as a computer or tablet that communicates with that device.

[0092] The management interface has two modes: episode list / semantic memory generation mode and episode playback mode. Switching modes on the standard screen 900 is done by the administrator clicking either episode list / semantic memory generation 901 or episode playback 902. Figure 11 shows an example where episode list / semantic memory generation mode is selected.

[0093] The episode list 903 on the left side of the standard screen 900 displays a list of episodes stored in the episode memory database 310. The episode list 903 includes a selection field 911 for the administrator to select an episode, an ID field 912 showing the ID of each episode, and a date and time field 913 showing the update date and time of each episode.

[0094] When an administrator clicks on the ID field 912 or the date and time field 913 of any episode, the management interface unit 120 displays the content of that episode in the preview 904 on the right side of the standard screen 900.

[0095] Furthermore, when the administrator selects one or more episodes in the selection field 911 and clicks the semantic memory generation 905, the semantic memory construction unit 510 starts the semantic memory construction process shown in Figure 5 and generates semantic memories from the selected episodes. In addition, the management interface unit 120 displays the progress of the semantic memory construction process in the status field 906 at the bottom of the standard screen 900. For example, when the semantic memory construction process is completed, the management interface unit 120 displays "Semantic memory generation complete" in the status field 906.

[0096] When the administrator clicks episode playback 902, the management interface unit 120 switches the standard screen 900 of the management interface to episode playback mode, as shown in Figure 12.

[0097] Figure 12 shows an example of a standard screen of the management interface displayed by the management interface unit 120. Figure 12 shows an example where the episode playback mode is selected.

[0098] In episode playback mode, it is possible to see how semantic memories stored in the semantic memory database 320 were used in episodes included in past episode memories stored in the episode memory database 310.

[0099] At the bottom of the standard screen 900, a seek bar 921 is displayed. The seek bar 921 shows the message transitions within a single episode. The administrator moves the cursor 924 by clicking back 922 or forward 923 to select the message for which they want to check the semantic memory usage.

[0100] The management interface unit 120 displays the chat status before and after the message selected by the seek bar 921 in the chat status 930 in the upper left of the standard screen 900.

[0101] Furthermore, the management interface unit 120 displays the semantic memory usage status 940 in the upper right corner of the standard screen 900, including the cluster 941 to which the user message immediately preceding the message selected by the seek bar 921 belongs in semantic memory, the cluster 942 which has been determined to potentially branch off from there, and the related filtering questions 943.

[0102] As described above, according to this embodiment, by extracting scenario branches based on a large amount of accumulated episodic memory and generating refined questions, it is possible to estimate the unspoken intentions not expressed in the user's statements and to provide responses that meet the user's true needs. Furthermore, by continuously learning appropriate ways to respond to diverse user requests through experience, it is possible to flexibly respond to diverse requests and changes in the environment and needs.

[0103] It should be noted that the present invention is not limited to the embodiments described above, and various modifications are included. For example, the embodiments described above are described in detail to make the present invention easier to understand, and are not necessarily limited to those having all the configurations described. Furthermore, it is possible to replace parts of the configuration of one embodiment with the configuration of another embodiment, and it is also possible to add configurations from other embodiments to the configuration of one embodiment. In addition, it is possible to add, delete, or replace parts of the configuration of each embodiment with other configurations. [Explanation of Symbols]

[0104] 100: AI Agent System 110: User Interface Department 120: Management Interface Section 210: AI Agent Department 310: Episodic Memory Database 320: Semantic Memory Database 410: Memory Retrieval Department 510: Semantic Memory Construction Department

Claims

1. In a knowledge extraction system that accumulates experiential memories, An episodic memory unit that stores the history of interactions with the user as an episodic memory containing multiple messages, A semantic memory construction unit clusters the messages contained in the episodic memory into multiple clusters and extracts scenario branches based on the transitions between the clusters of messages, A knowledge extraction system comprising a semantic memory unit that stores the aforementioned scenario branches as semantic memories.

2. In the knowledge extraction system described in claim 1, A user interface unit that receives natural language messages from the aforementioned user, A memory retrieval unit searches for and retrieves episodic memories and semantic memories that are similar to the message from the user, and generates a prompt based on at least one of the retrieved episodic memories and semantic memories. The AI ​​agent department is further equipped with, The user interface unit is Using the prompt, the AI ​​agent unit obtains a natural language message in response to the message from the user. A knowledge extraction system that presents the message from the AI ​​agent unit to the user.

3. In the knowledge extraction system described in claim 2, The aforementioned semantic memory construction unit, A knowledge extraction system that generates information for the AI ​​agent unit to generate refined questions to present to the user based on the aforementioned scenario branching.

4. In the knowledge extraction system described in claim 2, The episodic memory stored in the episodic memory unit is a knowledge extraction system in which messages from the user and messages from the AI ​​agent unit are linked together in a graph structure based on their chronological order.

5. In the knowledge extraction system described in claim 4, The memory retrieval unit is, The meaning and context of the aforementioned message are quantified to generate feature quantities, A knowledge extraction system that stores the aforementioned feature quantities in the episode memory unit, linked to the aforementioned message.

6. In the knowledge extraction system described in claim 4, The aforementioned semantic memory construction unit, A knowledge extraction system that calculates the similarity between messages contained in the episodic memory based on the aforementioned features, and clusters the messages based on the similarity.

7. In the knowledge extraction system described in claim 2, The aforementioned episodic memory is a knowledge extraction system that includes nonverbal contextual information.

8. In a knowledge extraction method for a knowledge extraction system that accumulates experiential memories, The episodic memory unit stores the user's interaction history as an episodic memory containing multiple messages. The semantic memory construction unit clusters the messages contained in the episodic memory into multiple clusters, extracts scenario branches based on the transitions between the clusters of messages, A knowledge extraction method in which a semantic memory unit stores the aforementioned scenario branching as semantic memory.

Images