Agent system

Abstract

This system provides an agent that can provide an appropriate final answer tailored to the user's permissions when the user obtains an answer using an answer generation and output means capable of outputting answers according to the input task. [Solution] The agent system 1 includes storage units 16,6 that store information linked to user permissions, an authority information acquisition unit 17 that acquires information available under the user's permissions from the storage units 16,6, a task generation unit 15 that generates tasks corresponding to the user's input, and a task execution unit 18 that, if the user's request includes matters related to information stored in the storage units 16,6, uses only the information available under the user's permissions, including the information acquired by the authority information acquisition unit 17, to have the response generation output unit 4 execute the task, and presents the user with a final response based on the response from the response generation output unit 4.

Description

Technical Field

[0001] The present invention relates to an agent system that intervenes between a user and answer generation output means capable of generating and outputting an answer according to an input request, and assists in the use of the answer generation output means.

Background Art

[0002] In recent years, technologies related to AI (Artificial Intelligence) have been rapidly developing, and the use of AI is also being promoted in various organizations such as companies and public institutions. As a service using AI, for example, there is known a chat-type answer generation output means that generates and outputs an answer according to a task (such as a question or an instruction) input from a user (see, for example, Patent Document 1). <第0000010号>[ ]

Prior Art Documents

Patent Documents

[0003]

Patent Document 1

Summary of the Invention

Problems to be Solved by the Invention

[0004] By the way, in various organizations such as companies, it is necessary to strictly manage various types of information such as know-how, customer information, technical information, and personal information of employees. Here, in terms of the management of various types of information, not only is it necessary to prevent the leakage of information to the outside, but it is also important to appropriately manage the information within the organization, that is, to limit the accessible information according to the access rights of the users belonging to the organization.

[0005] The present invention has been made in view of the above circumstances, and its purpose is to provide an agent system that can present an appropriate final answer tailored to the user's authority when the user obtains an answer using an answer generation and output means capable of generating and outputting answers according to the input task. [Means for solving the problem]

[0006] Below, we will describe, in separate sections, each means suitable for achieving the above objectives. Furthermore, we will add notes on the effects and benefits specific to each means as needed.

[0007] Method 1. An agent system that presents a final answer to a user's request using an answer generation and output means capable of generating and outputting an answer according to the input task, A storage method that stores various information linked to user permissions, From the aforementioned storage means, an authorized information acquisition means is provided that can acquire information available under the authority of the user who made the request, A task generation means that generates a task corresponding to the request entered by the user, If the user's request includes matters relating to information stored in the storage means, the task execution means causes the response generation output means to execute a task generated by the task generation means using only the information available under the user's authority, including the information obtained from the storage means by the authority information acquisition means. A response receiving means that receives the response output from the response generation output means upon execution of a task by the response generation output means, An agent system characterized by comprising a final answer presentation means that presents a final answer to the user based on the answers received by the answer receiving means.

[0008] According to the above-described means 1, if a user's request includes matters relating to information stored in the storage means, the response generation output means executes a task corresponding to the user's input request using only information available under the user's authority, including information obtained from the storage means by the authority-based information acquisition means. The response generation output means then generates and outputs a response upon task execution, and a final response based on the output response is presented to the user. Therefore, it is possible to more reliably prevent the final response from including information that the user making the request does not have access to. This allows for the presentation of an appropriate final response tailored to the user's authority. Furthermore, it enables more thorough confidentiality management of information.

[0009] Means 2. The task generation means is configured to generate tasks such that the answers output from the answer generation output means include the basis used to generate those answers. The agent system according to means 1, characterized in that it includes a basis verification means that checks whether the basis contains information that cannot be used under the authority of the user who made the request, by comparing the basis contained in the answer output from the answer generation output means with the information stored in the storage means.

[0010] According to the above-described method 2, the evidence verification means confirms whether the evidence output from the response generation output means contains information that cannot be used under the authority of the user who made the request. Therefore, it is possible to more reliably prevent the final response from containing information that the user who made the request cannot access. Furthermore, the confidentiality management of information can be carried out more appropriately.

[0011] Means 3. The agent system according to Means 2, characterized in that the basis verification means is configured to verify whether or not the answer output from the answer generation output means is appropriate, using the basis output from the answer generation output means.

[0012] According to the above-described means 3, the evidence verification means can use the evidence output from the answer generation output means (for example, based on whether or not evidence is included) to verify whether or not the answer output from the answer generation output means is appropriate. Therefore, for example, it is possible to more accurately verify whether or not the answer generation output means is presenting a false answer without evidence. As a result, a more accurate final answer can be presented to the user.

[0013] Means 4. The task generation means is configured to generate tasks such that the answers output from the answer generation output means include the basis used to generate those answers. The agent system according to means 1, characterized in that the final answer presentation means is configured to present to the user, along with the final answer, the basis output from the answer generation output means.

[0014] According to the above-described method 4, the user is presented with the final answer along with the evidence output from the answer generation output means (i.e., the basis for the final answer). This allows the user to more easily grasp the basis for the final answer, and to more easily verify and examine the accuracy of the final answer based on the presented evidence. Therefore, convenience for the user can be enhanced.

[0015] Means 5. The agent system according to Means 1, characterized in that the task generation means is configured to generate tasks using the results of morphological analysis performed on the requests input by the user.

[0016] According to method 5 described above, morphological analysis is used as an example to more accurately extract the intent of the request entered by the user. By utilizing the results of morphological analysis, the task generation means can generate tasks that are easier for the answer generation output means to understand (i.e., tasks that make it easier for the answer generation output means to obtain an answer that aligns with the intent of the request). This makes it possible to more reliably present the user with a final answer that aligns with the intent of the request, thereby increasing convenience for the user.

[0017] Means 6. The agent system according to Means 1, characterized in that it includes an NG word verification means capable of performing a process to verify whether or not the response output from the response generation output means contains any inappropriate information, using a pre-configured list of NG words that are inappropriate to present to the user.

[0018] According to method 6 described above, it is possible to more reliably prevent inappropriate information related to the NG word list (such as personal information) from being included in the final answer. This makes it possible to present users with more appropriate final answers.

[0019] Means 7. The agent system according to Means 6, characterized in that the task generation means is configured to modify and regenerate a task corresponding to the request input by the user so that if the response output from the response generation output means contains the inappropriate information, the response from the response generation output means no longer contains the inappropriate information.

[0020] According to the above-described means 7, when the answer output from the answer generation output means contains inappropriate information, the task generation means modifies and regenerates the task so that the answer does not contain inappropriate information. Therefore, by causing the answer generation output means to execute the regenerated task, it is possible to more reliably obtain an answer that does not contain inappropriate information. As a result, the final answer presented to the user can be made more appropriate. In addition, in obtaining an appropriate final answer, the user does not need to perform any special operations, so the convenience for the user can be further improved.

[0021] In addition, the technical matters related to the above-described means may be appropriately combined. Therefore, for example, at least one of the technical matters related to the above-described means 4 to 7 may be combined with the technical matter related to the above-described means 2.

Brief Description of the Drawings

[0022] [Figure 1] It is a block diagram showing a schematic configuration of an agent system or the like. [Figure 2] It is a flowchart for explaining the process flow in the agent system. [Figure 3] It is an explanatory diagram for explaining information stored in association with the user's authority.

Embodiments for Carrying Out the Invention

[0023] Hereinafter, an embodiment will be described with reference to the drawings. As shown in FIG. 1, the agent system 1 is constructed on a server (cloud server) on the Internet 100 and has a role of assisting and managing the "communication" between the user terminal 3 and the answer generation output unit 4. In this embodiment, the answer generation output unit 4 corresponds to the "answer generation output means". In addition, the user terminal 3 is, for example, a smartphone or a personal computer that can perform information input / output and display, and is communicable with the agent system 1. Prior to the description of the agent system 1, first, the answer generation output unit 4 will be briefly described.

[0024] The response generation and output unit 4 is built on the Internet 100 and generates and outputs responses according to the input task. In this embodiment, the response generation and output unit 4 uses a large language model (LLM) called ChatGPT, developed by OpenAI (registered trademark). In this embodiment, ChatGPT is a pre-trained AI model, specifically a model called GPT-4, which can handle various natural language tasks. Depending on the input task, ChatGPT performs various processes such as text generation, code generation, question answering, summarization, translation, citation (evidence) indication, text classification, information retrieval (web search), data graphing, text analysis, creation of prompts to input to other AI models (such as image generation AI or AI that performs speech recognition / speech-to-text conversion), and image processing, and presents the user with a response according to the task. ChatGPT can also perform the above processes using data input (uploaded) along with the task. Note that the response generation and output unit 4 may be an AI other than ChatGPT (for example, an LLM other than ChatGPT).

[0025] Next, Agent System 1 will be explained. Agent System 1 is a ChatGPT API (Application Programming Interface) that provides a final response to user requests by generating appropriate tasks for the requests and requesting task execution from the response generation output unit 4.

[0026] Agent System 1 comprises an access management unit 11, a request receiving unit 12, a request analysis request unit 13, a request analysis result receiving unit 14, a task generation unit 15, an internal storage unit 16, an authority information acquisition unit 17, a task execution unit 18, a response receiving unit 19, a basis verification unit 20, a response analysis request unit 21, a response analysis result receiving unit 22, an NG word verification unit 23, an NG word verification result receiving unit 24, and a final response presentation unit 25. The various functions provided by these units are realized through the cooperation of the various hardware components of the server and the software pre-stored in the server.

[0027] In this embodiment, the task generation unit 15 corresponds to the "task generation means," similarly, the internal storage unit 16 and the external storage unit 6 (described later) correspond to the "storage means," the authorization information acquisition unit 17 corresponds to the "authorization information acquisition means," the task execution unit 18 corresponds to the "task execution means," the basis verification unit 20 corresponds to the "basis verification means," the NG word verification unit 23 corresponds to the "NG word verification means," and the final answer presentation unit 25 corresponds to the "final answer presentation means."

[0028] The access management unit 11 manages access from the user terminal 3 to the agent system 1. Based on the ID (e.g., employee number) and password entered from the user terminal 3, the access management unit 11 determines whether or not the user terminal 3 is allowed to log in to the agent system 1. If the user terminal 3 is allowed to log in to the agent system 1, the access management unit 11 obtains information regarding the user's privileges based on the ID entered during login. In this embodiment, the available information from the various types of information stored in the internal storage unit 16 and the external storage unit 6 (described later) (in this embodiment, various types of internal company information including product information, supplier information, and employee personal information) is restricted according to the user's privileges.

[0029] The request receiving unit 12 receives requests (i.e., requests from users, such as user questions or instructions) entered into the agent system 1 from user terminals 3 authorized to log in. Examples of user requests include, "Please tell me the current selling price of product model number xx-xxxx and product name yyyyy," or "A new customer has inquired about product (model number xx-xxxx and product name yyyyy), so I would like to create a quote request email to the supplier." In this embodiment, "product model number xx-xxxx and product name yyyyy" refers to a product of the organization (company) to which the user belongs, and "supplier" refers to the supplier of the product for that organization.

[0030] The request analysis request unit 13 requests the analysis unit 5 to analyze the request received by the request receiving unit 12. The analysis unit 5 is a Japanese morphological analysis engine (MeCab in this embodiment) built on the Internet 100, which uses a dictionary from a vast Japanese database to divide the input sentence into words in an appropriate form and outputs the result as an analysis result to the input source. For example, if the sentence "I would like to know the current selling price of product with model number xx-xxxx and product name yyyyy" is input to the analysis unit 5, the analysis result (generated natural language) "I would like to know the current selling price of product / product Furthermore, when the analysis unit 5 receives text such as, "A new customer has inquired about a product (model number xx-xxxx, product name yyyyy), so I would like to create a quote request email for the supplier," it outputs an analysis result such as, "A new / customer / inquired / about / product / ( / model number / xx / - / xxxx / , / product name / yyyyy / ), so I would like to create a quote request / email / for / the supplier."

[0031] The request analysis result receiving unit 14 receives the analysis results from the analysis unit 5 regarding the user's request.

[0032] The task generation unit 15 uses the analysis results output from the analysis unit 5 and received by the request analysis result receiving unit 14 (i.e., the results of morphological analysis performed on the request input by the user) to generate a task (prompt) corresponding to the request input by the user. In this embodiment, the task generation unit 15 functions as an AI prompt generator, and in order to obtain an appropriate answer that matches the user's request from the answer generation output unit 4, it considers prompt engineering and the recency of the information, and uses words included in the analysis results, various pre-set prompt templates and candidate sentences (sentences prepared for each anticipated user request that are completed by inserting words into the essential parts) to generate a task that conforms to the user's request.

[0033] Furthermore, the task generation unit 15 may have a function to generate a task by checking whether the user's request contains any ambiguities based on the analysis results (for example, checking whether the subject and object are appropriately stated in the user's request), and if there are ambiguities, by asking the user questions about the ambiguities in the request based on a prompt template, and by using the user's answers to those questions. In addition, when interacting with the user, an AI such as an LLM that has been asked to get the user to answer according to a prompt template may be used.

[0034] Furthermore, the task generation unit 15 generates tasks that correspond to the requests entered by the user, as well as tasks that include the basis used to generate the answer in the answer output from the answer generation output unit 4 (tasks related to clarifying the basis for the answer). In other words, the task generation unit 15 generates tasks that correspond to the requests, as well as tasks such as "clarify the supporting data."

[0035] Therefore, in response to a user request such as, "I would like to know the current selling price of product model number xx-xxxx and product name yyyyy," the task generation unit 15 uses the analysis results of the request to generate tasks such as, "- Search the web for the selling price of product model number xx-xxxx and product name yyyyy," "- Search the internal data for the selling price of product model number xx-xxxx and product name yyyyy," "- Exclude older sales dates," and "- Clearly state the supporting data." Similarly, in response to a user request such as, "A new customer has inquired about a product (model number xx-xxxx and product name yyyyy), so I would like to create a quote request email to our suppliers," the task generation unit 15 generates tasks such as, "- Pick out suppliers that handle product model number xx-xxxx and product name yyyyy from the internal data," "- Create a draft of the quote request email," and "- Clearly state the supporting data."

[0036] The internal memory unit 16 is an information storage device that stores information uploaded to the agent system 1 and makes this information readable within the agent system 1. The internal memory unit 16 stores various types of information such as txt, csv, word, excel, and pdf files (various files 1, 2, 3...n, n+1, n+2...n is a natural number greater than or equal to 4) and associates them with user permissions (see Figure 3). In Figure 3, "〇" means that the user has permission to use the information. Therefore, in Figure 3, for example, users A and C have permission to use file 1, and users A and E have permission to use file 3.

[0037] Furthermore, in this embodiment, in addition to the internal storage unit 16, an external storage unit 6 is provided outside the agent system 1. The external storage unit 6 is composed of a predetermined cloud storage or internal data storage, and, like the internal storage unit 16, stores various information linked to the user's permissions.

[0038] The authorization information acquisition unit 17 acquires information available under the authorization of the user who made the request from the internal storage unit 16 and the external storage unit 6. More specifically, based on the generated task and the user authorization information acquired by the access management unit 11, the authorization information acquisition unit 17 acquires from both storage units 16 and 6 information necessary for executing the task that is available under the authorization of the user. For example, if user A is logged into agent system 1, based on words etc. included in the task (e.g., model number xx-xxxx, product name yyyyy, sales price, etc.), information necessary for executing the generated task (e.g., an information file about the sales price of model number xx-xxxx and product name yyyyy) that is available under the authorization of user A (e.g., file n, etc.) is acquired from both storage units 16 and 6.

[0039] Furthermore, the information stored in memory units 16 and 6 is periodically updated, and an API is used, so that the authorization information acquisition unit 17 is always configured to acquire the latest information from memory units 16 and 6. As a result, the information source in the final answer is always kept up-to-date, and the output of a final answer based on outdated information is not ensured.

[0040] The task execution unit 18 has functions such as executing tasks generated by the task generation unit 15 at the response generation output unit 4, and checking whether there is any missing information in the task. The task execution unit 18 includes an execution request unit 18a and an information deficiency confirmation unit 18b. For the sake of explanation, the information deficiency confirmation unit 18b will be explained after the response receiving unit 19.

[0041] The execution request unit 18a requests the response generation output unit 4 to execute the task generated by the task generation unit 15, and causes the response generation output unit 4 to execute the task. Here, if the user's request includes matters related to information stored in the storage units 16,6 (for example, if the user's request includes matters related to internal company information), the execution request unit 18a causes the response generation output unit 4 to execute the task using only information available under the user's authority, including information acquired by the authority information acquisition unit 17. For example, when the response generation output unit 4 is to execute a task such as "Search for the sales price of model number xx-xxxx, product name yyyyy" in response to a request from user A, the execution request unit 18a inputs the task along with the information necessary to execute the task (for example, a file related to sales prices) that is available under user A's authority to the response generation output unit 4. In response to the task execution request, the response generation output unit 4 outputs a response to the agent system 1 using only information available under the user's authority.

[0042] The response receiving unit 19 receives the response output from the response generation output unit 4 in response to the task execution request from the execution request unit 18a. As described above, the tasks executed by the response generation output unit 4 include tasks related to clarifying the basis for the response, so the response from the response generation output unit 4 includes both a response to the user's request and a statement indicating the basis used to generate the response. In this embodiment, a list of information sources (e.g., file names and URLs) used to generate the response is output as the basis.

[0043] The information deficiency confirmation unit 18b checks whether there is any information deficiency in the content of the task generated by the task generation unit 15. More specifically, the information deficiency confirmation unit 18b checks whether there is any information deficiency in the task by having the task generated by the task generation unit 15 executed by the answer generation output unit 4 and checking whether there is any information deficiency in the answer output in relation to the task. For example, the information deficiency confirmation unit 18b compares the number of answers that should be output by the generated task with the number of answers that are actually output. If the numbers match, it determines that there is no information deficiency in the task; if the numbers do not match, it determines that there is an information deficiency in the task.

[0044] If there is insufficient information in a task, the execution request unit 18a will cause the response generation output unit 4 to execute the task again. In other words, even if there is insufficient information in a task, the agent system 1 automatically requests the task to be re-executed so that an appropriate final answer can be presented without requiring any action from the user. However, even if the same task is executed, the response generated by the response generation output unit 4 will not necessarily be the same. Therefore, as a result of this task re-execution, it may be determined that there is no insufficient information in the task.

[0045] On the other hand, if it is determined that there is insufficient information for the task even after re-executing the task, the response generation output unit 4 will present a response to the user (sent to the user terminal 3). Furthermore, information will be sent to the user terminal 3 requesting the user to make further requests (such as further questions) with more details. If the user makes further requests, the response generation output unit 4 will generate a task corresponding to the further requests and execute the said task.

[0046] The basis verification unit 20, when there is no lack of information for the task, compares the basis included in the answer output by the answer generation output unit 4 with the information stored in the storage units 16,6 to confirm whether the basis includes information that cannot be used under the authority of the user who made the request. For example, the basis verification unit 20 checks the user permissions that can access the information sources used to generate the answer from the basis (i.e., the list of information sources used to generate the answer), and confirms whether the information sources include information (files) that cannot be used under the authority of the user who made the request.

[0047] If the information sources include information sources that cannot be used under the user's authority, the task generation unit 15 regenerates the task so that such information is not used when the response generation output unit 4 executes the task. For example, the original task is modified and regenerated by adding instructions such as "Do not use this information" or explicit instructions on narrowing down the information sources used for response generation. Then, using the regenerated task, the various processes described above (i.e., the execution of the task by the response generation output unit 4, the confirmation of insufficient information in the task, the confirmation of the basis for the answer, etc.) are performed again.

[0048] Furthermore, the justification verification unit 20 uses the justification output from the answer generation output unit 4 to verify whether the answer output from the answer generation output unit 4 is appropriate. For example, the justification verification unit 20 verifies the appropriateness of the answer based on whether the justification for that answer is included in the answer output from the answer generation output unit 4. If the justification for the answer is not included and the answer cannot be said to be appropriate, the unit may regenerate the task so that the justification is appropriately included in the answer, or re-execute the various processes described above (such as the execution of the task in the answer generation output unit 4) using the regenerated task.

[0049] The response analysis request unit 21 requests the analysis unit 5 to analyze the response received from the response generation output unit 4. This analysis of the response can be considered preparation for determining whether or not the response contains inappropriate information (NG words described later). The analysis unit 5 divides the response into words in an appropriate format.

[0050] The response analysis result receiving unit 22 receives the analysis results for the responses received from the response generation output unit 4 from the analysis unit 5.

[0051] The NG word verification unit 23 checks against a predetermined NG word list (for example, stored in the internal memory unit 16) and performs processing to confirm whether the answer output from the answer generation output unit 4 contains any information that should not be presented to the user. The NG word list is data that lists multiple pieces of such inappropriate information (for example, personal information or information that identifies personal information such as email addresses). The NG word verification unit 23 uses the analysis results of the answer output from the answer generation output unit 4 and the NG word list to request the answer generation output unit 4 to confirm whether or not the answer contains any inappropriate information.

[0052] The NG word confirmation result receiving unit 24 receives confirmation results from the answer generation output unit 4 regarding whether or not the answer contains inappropriate information.

[0053] If it is confirmed that the answer contains inappropriate information, the task generation unit 15 modifies and regenerates the task corresponding to the user's request so that the answer from the answer generation output unit 4 does not contain the inappropriate information. For example, the task is modified and regenerated by adding instructions such as "Do not include personal names" to the original task. Then, using the regenerated task, the various processes described above (i.e., the execution of the task in the answer generation output unit 4, confirmation of insufficient information in the task, confirmation of the basis for the answer, etc.) are performed again.

[0054] The final answer presentation unit 25 presents the user with the final answer based on the answers received by the answer receiving unit 19. More specifically, the final answer presentation unit 25 presents to the user (displayed on the user terminal 3) the answers received by the answer receiving unit 19 that have been determined to be free of problems as a result of the above-mentioned checks by the basis verification unit 20 and the NG word verification unit 23. Therefore, the user will not be presented with any information that is not available to them under their authority, nor will any inappropriate information such as personal names be presented.

[0055] Furthermore, the final answer presentation unit 25 presents the user with the final answer along with the evidence output from the answer generation output unit 4 (i.e., a list of information sources used to generate the answer).

[0056] Next, the operation flow of the agent system 1 when a request is entered by a user will be explained with reference to the flowchart in Figure 2. It should be assumed here that the user is already logged into agent system 1 and that the access management unit 11 has obtained information regarding the user's permissions.

[0057] First, the processes in steps S1, S2, and S3 analyze the user's input request as preparation for task generation. Specifically, in step S1, the request receiving unit 12 receives the user's request; in step S2, the request analysis request unit 13 requests the analysis unit 5 to analyze the request; and in step S3, the request analysis result receiving unit 14 receives the analysis result of the request. This provides the result of morphological analysis of the request.

[0058] Then, in step S4, a task is generated by the task generation unit 15 using the analysis results obtained in step S3. Furthermore, in the following step S5, in order to obtain the information necessary for task execution, the authorization information acquisition unit 17 acquires information necessary for task execution from the storage units 16,6 that is available under the authorization of the logged-in user.

[0059] Next, in step S6, the execution request unit 18a causes the task to be executed by the response generation output unit 4 using only the information available under the user's authority, including the information obtained in step S5. Then, in step S7, the response receiving unit 19 receives the response from the response generation output unit 4. The response includes the basis for that response (a list of information sources).

[0060] Next, various checks are performed on the obtained response, namely, checks for insufficient information regarding the task, checks for consistency between the information included in the response and user permissions, and checks for prohibited words in the response.

[0061] Regarding the process of checking for insufficient information in a task, first, in step S8, the information deficiency confirmation unit 18b checks whether or not there is insufficient information in the task. If an information deficiency in a task is confirmed once (step S8: NO, step S9: NO), the execution request unit 18a causes the task to be executed again by the response generation output unit 4. On the other hand, if an information deficiency in a task is confirmed multiple times (step S8: NO, step S9: YES), the answer is presented to the user in step S10, and a request for further information (such as further questions) is made to the user in step S11.

[0062] If there is no lack of information regarding the task (Step S8: YES), in Step S12, a check is performed to verify the consistency between the information included in the answer and the user's permissions. In other words, it is checked whether the answer contains any information that cannot be used under the user's permissions. This check is performed by the basis verification unit 20, which compares the basis of the answer output by the answer generation output unit 4 with the information stored in the storage units 16,6.

[0063] If the answer contains information that cannot be used under the user's authority (i.e., the information in the answer is inconsistent with the user's authority) (Step S12: YES), the task generation unit 15 modifies and regenerates the task so that the information is not used (Step S13). Then, the processing from Step S6 onwards is performed again using the regenerated task.

[0064] On the other hand, if the answer does not contain information that cannot be used under the user's authority (step S12: NO), steps S14, S15, and S16 perform a check for the presence of NG words in the answer. Specifically, in step S14, the answer analysis request unit 21 requests the analysis unit 5 to analyze the answer, in step S15, the answer analysis result receiving unit 22 receives the analysis result of the answer, and in step S16, the NG word confirmation unit 23 checks whether or not the answer contains inappropriate information.

[0065] If the answer contains inappropriate information (NG words) (Step S16: YES), the task generation unit 15 modifies and regenerates the task so that the answer does not contain inappropriate information (Step S17). Then, the processing from Step S6 onwards is performed again using the regenerated task.

[0066] On the other hand, if the answer does not contain any inappropriate information (step S16: NO), in step S18, the final answer presentation unit 25 presents the user with the final answer and the basis for that final answer.

[0067] As detailed above, according to this embodiment, if a user's request includes information stored in the memory units 16,6, the response generation output unit 4 executes a task corresponding to the user's request using only the information available under the user's authority, including the information obtained from the memory units 16,6 by the authority information acquisition unit 17. Then, the user is presented with a final response based on the response output from the response generation output unit 4. Therefore, it is possible to more reliably prevent the final response from including information that the user who made the request does not have the authority to access. This makes it possible to present the user with an appropriate final response that matches the user's authority. Furthermore, it becomes possible to more thoroughly manage the confidentiality of information.

[0068] Furthermore, the evidence verification unit 20 verifies whether the evidence output from the response generation output unit 4 contains any information that cannot be used under the authority of the user who made the request. Therefore, it is possible to more reliably prevent the final response from containing information that the user who made the request cannot access. In addition, the confidentiality management of information can be carried out more appropriately.

[0069] In addition, the evidence verification unit 20 uses the evidence output from the answer generation output unit 4 to verify whether the answer output from the answer generation output unit 4 is appropriate. Therefore, it is possible to more accurately verify whether the answer generation output unit 4 has presented a false answer without evidence. As a result, a more accurate final answer can be presented to the user.

[0070] In addition, the user is presented with the basis for their final answer, along with the reasoning output from the answer generation output unit 4 (i.e., the basis for their final answer). This allows the user to more easily understand the basis for their final answer, and to more easily verify and examine the accuracy of their final answer based on the presented reasoning. Therefore, user convenience can be enhanced.

[0071] In addition, by utilizing the results of morphological analysis of the user's input requests, the task generation unit 15 generates tasks that are easier for the response generation output unit 4 to understand (i.e., tasks that make it easier for the response generation output unit 4 to obtain a response that aligns with the intent of the request). Therefore, the task generation unit 15 can generate appropriate tasks that align with the user's requests without the user having to learn how to use the system in advance or remember complex settings. This makes it possible to more reliably present the user with a final response that aligns with the intent of their request, thereby increasing user convenience.

[0072] In addition, the NG word verification unit 23 more reliably prevents inappropriate information (such as personal information) related to the NG word list from being included in the final answer. This allows for the presentation of a more appropriate final answer to the user.

[0073] In addition, if the response output from the response generation output unit 4 contains inappropriate information, the task generation unit 15 modifies and regenerates the task so that the inappropriate information is no longer included in the response. Therefore, it is possible to more reliably obtain a response that does not contain inappropriate information, and as a result, the final response presented to the user can be made more appropriate. Furthermore, since the user does not need to perform any special operations to obtain an appropriate final response, the convenience for the user is further enhanced.

[0074] Furthermore, the embodiment is not limited to the description above, and may be implemented as follows, for example. Of course, other applications and modifications not exemplified below are also possible.

[0075] (a) In the above embodiment, a response generation output unit 4 and an analysis unit 5 are provided separately from the agent system 1, but the agent system 1 may be configured to have the functions of the response generation output unit 4 and the analysis unit 5.

[0076] (b) In the above embodiment, if a problem is found in the check process regarding the consistency between the information included in the answer and the user's permissions, or in the check process regarding the presence or absence of prohibited words in the answer, the task is regenerated and the task is executed again by the answer generation output unit 4. Alternatively, if a problem is found once or multiple times in these check processes, the system may be configured to send information to the user indicating that it is not possible to provide an appropriate final answer. In this case, the system may also send the user information regarding the problematic points.

[0077] (c) In the above embodiment, the user is presented with the final answer and its rationale. However, the system may be configured to present the user with the final answer and the tasks used to obtain that final answer. [Explanation of Symbols]

[0078] 1... Agent system, 4... Answer generation output unit (answer generation output means), 6... External storage unit (storage means), 15... Task generation unit (task generation means), 16... Internal storage unit (storage means), 17... Authorized information acquisition unit (authorized information acquisition means), 19... Answer receiving unit (answer receiving means), 20... Basis verification unit (basis verification means), 23... NG word verification unit (NG word verification means), 25... Final answer presentation unit (final answer presentation means).

Claims

1. An agent system that presents a final answer to a user's request using an answer generation and output means capable of generating and outputting answers according to the input task, A storage method that stores various information linked to user permissions, From the aforementioned storage means, an authorized information acquisition means is provided that can acquire information available under the authority of the user who made the request, A task generation means that generates a task corresponding to the request entered by the user, If the user's request includes matters relating to information stored in the storage means, the task execution means causes the response generation output means to execute a task generated by the task generation means using only the information available under the user's authority, including the information obtained from the storage means by the authority information acquisition means. A response receiving means that receives the response output from the response generation output means upon execution of a task by the response generation output means, An agent system characterized by comprising a final answer presentation means that presents a final answer to the user based on the answers received by the answer receiving means.

2. The task generation means is configured to generate tasks such that the answers output from the answer generation output means include the basis used to generate those answers. The agent system according to claim 1, further comprising a basis verification means that checks whether the basis included in the response output from the response generation output means is not information that cannot be used under the authority of the user who made the request, by comparing the basis included in the response with the information stored in the storage means.

3. The agent system according to claim 2, characterized in that the basis verification means is configured to verify whether or not the answer output from the answer generation output means is appropriate, using the basis output from the answer generation output means.

4. The task generation means is configured to generate tasks such that the answers output from the answer generation output means include the basis used to generate those answers. The agent system according to claim 1, characterized in that the final answer presentation means is configured to present to the user, along with the final answer, the basis output from the answer generation output means.

5. The agent system according to claim 1, characterized in that the task generation means is configured to generate tasks using the results of morphological analysis performed on the requests input by the user.

6. The agent system according to claim 1, further comprising an NG word verification means capable of performing a process to verify whether or not the response output from the response generation output means contains inappropriate information, using a pre-configured list of NG words that are inappropriate to present to the user.

7. The agent system according to claim 6, characterized in that the task generation means is configured to modify and regenerate a task corresponding to the request input by the user so that if the response output from the response generation output means contains the inappropriate information, the response from the response generation output means no longer contains the inappropriate information.

Images