Information processing method, information processing program, and information processing device
Patent Information
- Application Number
- PCT/JP2025/044505
- Authority / Receiving Office
- WO · WO
- Patent Type
- Applications
- Current Assignee / Owner
- Priority Date
- 2025-02-20
- Filing Date
- 2025-12-19
- Publication Date
- 2026-08-27
Smart Images
Figure JP2025044505_27082026_PF_FP_ABST
Abstract
Description
Information Processing Method, Information Processing Program, and Information Processing Apparatus ,
[0008] ,
[0007] , ,
[0006] , ,
[0005] ,
[0001] The present invention relates to an information processing method, an information processing program, and an information processing apparatus, and can be applied to, for example, a dialogue system using a graph document.
[0002] In recent years, with the development of large language models (LLMs), the performance of dialogue systems has also improved, and various technologies have been developed.
[0003] For example, Patent Document 1 discloses an apparatus that uses a language model to search for a document to be searched by dialogue. Also, Patent Document 2 discloses an apparatus that gives an instruction to edit a document by dialogue and corrects the document.
[0004] By the way, Non-Patent Document 1 states that there are various advantages in using a graph document instead of a text document as public information of the Industrial Japanese Language Research Society. A graph document represents the document structure as a graph, represents a sentence of about one sentence granularity as one node, connects the relationship between sentences with an edge, and attaches a label representing the meaning of the relationship to the edge, thereby expressing the structure and meaning of the document. A graph document can be said to be compressed information of a text document (including dialogue history).
[0005] Japanese Unexamined Patent Application Publication No. 2023-076413, Japanese Unexamined Patent Application Publication No. 2023-017938
[0006] "Industrial Japanese Language Research Society, Document Creation Support Subcommittee Public Information" [Searched on February 7, 2025], [Online], INTERNET, <URL:https: / / tech-jpn.jp / wp-content / uploads / symposium / 13 / 13-2-2.pdf>
[0007] By using a language model as in the above Patent Document 1, it is possible to create a response sentence that is somewhat appropriate from the dialogue system for the utterance from the user.
[0008] However, because there is a length limit to the prompts used to create response sentences to send to the language model, if the dialogue history to be sent becomes long, it is necessary to delete older dialogue history before sending. In that case, the content of the deleted dialogue history is not passed to the language model, resulting in a problem where the dialogue does not take into account the information contained in the deleted dialogue history.
[0009] Therefore, there is a need for an information processing method, information processing program, and information processing device that can conduct dialogues using graph documents, which are compressed pieces of dialogue history.
[0010] The first information processing method of the present invention is an information processing method that outputs information for displaying a dialogue history, which is a history of the content of the dialogue with a large-scale language model, and a graph document, which is a graph of the document structure of the dialogue content, on a display unit, characterized in that when an operation on the graph document is received, information for changing the display mode of the dialogue history or information for changing the display mode of the graph document is output.
[0011] The second information processing program of the present invention is characterized by causing a computer to execute the information processing method of the first present invention.
[0012] The third information processing device of the present invention is characterized by executing the information processing program of the second present invention.
[0013] According to the present invention, it is possible to have a conversation using a graph document.
[0014] This is a block diagram showing the configuration of the graph document dialogue system according to the first embodiment. This is an explanatory diagram showing an example of the screen configuration of the graph document dialogue input / output unit according to the first embodiment. This is an explanatory diagram showing an example of the configuration of a prompt template (for dialogue) according to the first embodiment. This is an explanatory diagram showing an example of the dialogue history according to the first embodiment. This is an explanatory diagram showing an example of a graph document according to the first embodiment. This is an explanatory diagram showing an example of the configuration of a prompt template (for searching) according to the first embodiment. This is an explanatory diagram (part 1) showing an example of the search results retrieved by the graph document storage and search unit according to the first embodiment. This is an explanatory diagram (part 2) showing an example of the search results retrieved by the graph document storage and search unit according to the first embodiment. This is a flowchart showing the characteristic operation of the graph document dialogue device (dialogue control unit) according to the first embodiment. This is a flowchart showing the characteristic operation of the graph document dialogue input / output unit according to the first embodiment. This is a block diagram showing the configuration of the graph document dialogue system according to the second embodiment. This is a flowchart showing the characteristic operation of the graph document dialogue device (dialogue control unit) according to the second embodiment. This is a flowchart showing the characteristic operation of the graph document dialogue input / output unit according to the second embodiment. This is an explanatory diagram (part 1) showing an example of a graph document according to the second embodiment. This is an explanatory diagram (part 2) showing an example of a graph document according to the second embodiment. This is a block diagram showing the configuration of the graph document dialogue system according to the third embodiment. This is a flowchart showing the characteristic operation of the graph document dialogue device according to the third embodiment. This is an explanatory diagram showing an example of the configuration of a prompt template (for existing document conversion) according to the third embodiment. This is a block diagram showing the configuration of the graph document dialogue system according to the fourth embodiment. This is an explanatory diagram showing an example of the configuration of a prompt template (for this template) according to the fourth embodiment. This is a flowchart showing the characteristic operation (cooperation with external systems) of the graph document dialogue device according to the fourth embodiment. This is a block diagram showing the configuration of the graph document dialogue system according to the fifth embodiment. This is a flowchart showing the characteristic operation of the graph document dialogue device according to the fifth embodiment. This is an explanatory diagram showing an example of the screen configuration of the aggregation and analysis result input / output unit 25 according to the fifth embodiment.
[0015] (A) The first embodiment and subsequent embodiments of the information processing method, information processing program, and information processing apparatus according to the present invention will be described in detail with reference to the drawings.
[0016] In the first embodiment, the objective is to improve the quality of dialogue by clearly indicating the correspondence between the dialogue content and the graph document when engaging in dialogue using a graph document and a language model.
[0017] (A-1) Diagram 1 of the configuration of the first embodiment is a block diagram showing the configuration of the graph document dialogue system according to the first embodiment.
[0018] In Figure 1, the graph document dialogue system 1 includes a graph document dialogue device 10 as an information processing device, a graph document dialogue input / output unit 20, and a large-scale language model 30.
[0019] Furthermore, in the graph document dialogue system 1, the graph document dialogue device 10, the graph document dialogue input / output unit 20, and the large-scale language model 30 can be connected via a network using various communication methods, whether wired or wireless. As a modification, the graph document dialogue device 10 may internally store the large-scale language model 30.
[0020] The graph document dialogue input / output unit 20 is a functional unit that can input user utterances, display user utterances and system utterances, and display and edit graph documents. The graph document dialogue input / output unit 20 can be implemented on, for example, a PC, tablet, or smartphone.
[0021] Figure 2 is an explanatory diagram showing an example of the screen configuration of the graph document dialogue input / output unit according to the first embodiment.
[0022] In Figure 2, the graph document interactive input / output screen 200 has a graph document display / editing field 201 and an interactive history display / input field 202.
[0023] The graph document display / editing section 201 allows for the display and editing of graph documents. While editing is intended to allow the addition of nodes and edges on the screen, display-only functionality is also acceptable. The edited graph document can be sent to the graph document dialogue device 10, for example, by pressing the send button 205, or it can be sent to the graph document dialogue device 10 at any time after editing is complete; however, there are no particular limitations on the trigger for sending the graph document.
[0024] The dialogue display area 203 of the dialogue history display / input area 202 displays the dialogue between the user and the system. The method of displaying the dialogue is not particularly limited, but for example, in Figure 2, the speech bubble from the right shows the user's utterance, and the speech bubble from the left shows the system's utterance. It is assumed that the user can input the dialogue by entering arbitrary text in the text input area 204 at the bottom and pressing the send button 205.
[0025] The graph document interaction device 10 comprises an interaction control unit 11, a storage unit 12, a graph document storage and retrieval unit 41, and an interaction history and graph document interaction management unit 50.
[0026] The graph document dialogue device according to this embodiment may be configured as hardware such as a dedicated IC chip equipped with each component shown in Figure 1, or it may be configured as software centered on a CPU and a program executed by the CPU, but functionally it can be represented as shown in Figure 1.
[0027] The dialogue control unit 11 is a functional unit that, in response to utterances from the user input via the graph document dialogue input / output unit 20, creates response sentences (system utterances) in cooperation with the large-scale language model 30 (described later) and controls the dialogue with the user. As a display control unit, the dialogue control unit 11 transmits the created system utterances and graph documents to the graph document dialogue input / output unit 20 and controls it to display their contents.
[0028] Furthermore, the dialogue control unit 11 uses the large-scale language model 30 to create search query statements for searching for relevant graph documents from the group of graph documents 42 managed by the graph document storage and search unit 41 based on the dialogue content.
[0029] The memory unit 12 is a functional unit that holds data used by the dialogue control unit 11, and includes prompt templates 13 (13A, 13B), dialogue history 14, the current graph document 15, and a group of graph documents 42.
[0030] The prompt templates 13 (13A, 13B) are template data for the large-scale language model 30 to construct command sentences (prompts) in order to generate system utterances. In this embodiment, as will be described later, a prompt template 13A for dialogue and a prompt template 13B for searching that is referenced during dialogue are used.
[0031] Figure 3 is an explanatory diagram showing an example configuration of a prompt template (for dialogue) according to the first embodiment.
[0032] In Figure 3, the prompt template 13A includes an overall prompt 13A-1, a variable portion for each domain 13A-2, a detailed prompt 13A-3, the current graph document 13A-4, the dialogue history 13A-5, and the search results 13A-6.
[0033] The overall prompt 13A-1 and the detailed prompt 13A-3 are fixed settings that are independent of the domain.
[0034] The overall prompt 13A-1 should contain instructions to interact with the user while creating a graph document. For example, the instructions might be, "Please interact with the user while summarizing the content of the conversation in a graph document."
[0035] Detailed prompt 13A-3 contains specific graph document formatting and notes. For example, it might say, "A graph document is a collection of nodes and labeled directed edges. Edge labels should be selected from 'Result', 'Cause', 'Background', 'Example', 'Conclusion', 'Constraint', 'Purpose', and 'Contrast'. Node content should be a single, relatively long sentence."
[0036] The domain-specific variable portion 13A-2 is a set of instructions that vary depending on the domain and must be finalized before execution. For example, the instructions might be: "You (large-scale language model 30) are a paper writer. Based on the information the user provides about the paper they want to write, extract the necessary information from the user. Provide appropriate information as needed."
[0037] The current graph document 13A-4 and dialogue history 13A-5 will be populated by the dialogue control unit 11 at runtime with the values of the current graph document 15 and dialogue history 14, respectively.
[0038] The search result 13A-6 contains the search results obtained by the graph document storage and search unit 41, which will be described later.
[0039] Figure 4 is an explanatory diagram showing an example of a dialogue history according to the first embodiment. The dialogue history 14 shown in Figure 4 is a representation of the content displayed in the dialogue display area 203 of Figure 2 in a predetermined data format. In Figure 4, the dialogue history 14 is shown in JSON format, but the data format used is not particularly limited. As the dialogue between the user and the system progresses, the dialogue history 14 becomes longer as the dialogue content (user utterances and system utterances) is embedded. That is, the dialogue history 14 held in the storage unit 12 is updated as needed by instructions from the dialogue control unit 11.
[0040] Figure 5 is an explanatory diagram showing an example of a current graph document according to the first embodiment. The current graph document 15 shown in Figure 5 shows the contents displayed in the graph document display / editing section 201 of Figure 2 in a predetermined data format (JSON format, as described above).
[0041] As shown in Figure 5, list the node contents and edge contents in "nodes" and "edges" respectively. The format does not have to be exactly as in this example; it is sufficient if similar content is included.
[0042] The large-scale language model 30 assumes a language model that is generally distributed or available via an API (Application Programming Interface), and is, for example, a generative AI such as ChatGPT. When a prompt is input, the large-scale language model 30 generates and outputs text data (system utterance sentences for user utterances, graph documents in which the dialogue history is structured) corresponding to the prompt. In the first embodiment, the large-scale language model 30 shows an example of handling only text data, but as a modification, not only text but also information such as images and sounds may be handled.
[0043] FIG. 6 is an explanatory diagram showing a configuration example of a prompt template for search according to the first embodiment.
[0044] In FIG. 6, the prompt template 13B for search has an overall prompt 13B-1, a variable part 13B-2 for each domain, a current graph document 13B-3, and a dialogue history 13B-4.
[0045] The overall prompt 13B-1 describes an instruction to create a search query sentence necessary for searching for a graph document for reference from the dialogue content. The variable part 13B-2 for each domain has instruction contents that change depending on the domain, and describes instructions such as the field and purpose of the dialogue. The current graph document 13B-3 and the dialogue history 13B-4 are such that the dialogue control unit 11 embeds the values of the current graph document 15 and the dialogue history 14 respectively at the time of execution.
[0046] The graph document group 42 is a plurality of graph documents managed (held) by the graph document storage / search unit 41.
[0047] The graph document storage / search unit 41 vectorizes the content of each fragment of the graph (one node, one node and one edge connected to that node, or two connected nodes and one edge, etc.) of each graph document in the held graph document group 42 so that it can be searched.
[0048] The graph document storage and search unit 41 performs the search in two stages. First, it creates a list of graph documents or fragments of graph documents that have content similar to the search query sentence. Specifically, the graph document storage and search unit 41 performs vector search using a method such as approximate nearest neighbor search using the vector representation of the character string. Not only the text representation of an entire graph document is converted into a vector representation, but vectors of fragments of the graph document (only the content of one node, only the label of the edge connecting to the node content, only the content of two connected nodes and the label of that edge, etc.) are also stored so that they can be searched using fragments of the graph document. Thereby, the vector search result shown in FIG. 7 is obtained.
[0049] Furthermore, based on the vector search result, the graph document storage and search unit 41 includes in the final search result the group of nodes connected by a specific label starting from the nodes included therein. For example, a final search result as shown in FIG. 8 is created including the group of nodes connected by labels such as "reference", "summary", "function", etc. By making this the final search result, not only the vector search result but also information on the group of nodes such as the reference destination can be included in the prompt, and the result returned by the large language model 30 can be made better. Generally, when the information necessary for answering is not included in the prompt, the large language model 30 tends to generate incorrect content. Therefore, by sufficiently providing the necessary information in the prompt in this way, the result returned by the large language model 30 can be made more accurate.
[0050] The dialogue history-graph document dialogue management unit 50 manages the correspondence between the dialogue history 14 and the current graph document 13A-4.
[0051] First, let me explain why the dialogue history / graph document dialogue management unit 50 manages the correspondence between the dialogue history 14 and the current graph document 13A-4. Systems using large-scale language models (LLMs) (such as chatbots) may randomly embed plausible-sounding lies (hallucinations) in the generated content. If a user finds a suspicious part of the generated content that contains hallucinations, it is crucial to access the information on which the generated content was based and verify its contents.
[0052] Therefore, the dialogue history / graph document dialogue management unit 50 creates a correspondence between each node of the generated content, the graph document (current graph document 15), and the corresponding location in the dialogue history 14 that served as the basis for generating each node, when the graph document is created. For example, by adding an instruction such as "...please show the correspondence between each node of the graph document and the dialogue history" to the detailed prompt 13A-3 of the prompt template 13A, information that explicitly shows the correspondence between each node of the graph document and the dialogue history is added in addition to the basic information of the graph document shown in Figure 5. Here, explicit means, for example, showing the node ID that identifies each node and the corresponding location in the dialogue history (creating corresponding information).
[0053] Furthermore, the data format and content used to store information indicating the correspondence between each node in the graph document and the dialogue history are not particularly limited and may be managed in a format other than that shown in Figure 5.
[0054] (A-2) Operation of the First Embodiment Next, the operation of the graph document dialogue system 1 according to the first embodiment having the above configuration will be described.
[0055] (A-2-1) The processing diagram 9 of the graph document dialogue device 10 is a flowchart showing the characteristic operation of the graph document dialogue device (dialogue control unit) according to the first embodiment.
[0056] <S101> The dialogue control unit 11 acquires user utterances and graph documents from the graph document dialogue input / output unit 20. Initially, the user utterance and graph document may be empty. If the user utterance is empty, the system utterance will be the starting point.
[0057] <S102, S103> The dialogue control unit 11 adds the acquired user utterance to the end of the dialogue history 14. Similarly, the dialogue control unit 11 stores the acquired graph document in the current graph document 15.
[0058] <S104> After step S103 described above, the dialogue control unit 11 creates a search prompt by embedding the current graph document 15 in the location of the current graph document 13B-3 in the search prompt template 13B shown in Figure 6, and embedding the dialogue history 14 in the location of the dialogue history 13B-4.
[0059] <S105> The dialogue control unit 11 sends the search prompt created in step S104 above to the large-scale language model 30. The large-scale language model 30 creates a search query statement based on the search prompt.
[0060] <S106> The dialogue control unit 11 obtains a search query statement from the large-scale language model 30.
[0061] <S107> The dialogue control unit 11 sends the acquired search query statement to the graph document storage and search unit 41, causing it to search the graph document group 42.
[0062] <S108> The dialogue control unit 11 obtains the search results shown in Figure 8 from the graph document storage and search unit 41.
[0063] <S109> The dialogue control unit 11 creates a dialogue prompt by embedding the current graph document 15 in the current graph document 13A-4 location of the dialogue prompt template 13A, embedding the dialogue history 14 in the dialogue history 13A-5 location, and embedding the search results obtained in step S108 above in the search results 13A-6 location.
[0064] In the first embodiment, the dialogue history / graph document dialogue management unit 50 instructs the dialogue control unit 11 to create a dialogue prompt that includes information indicating the correspondence between the graph document (each node) and the dialogue history.
[0065] <S110> The dialogue control unit 11 sends the prompt created in step S104 above to the large-scale language model 30. The large-scale language model 30 generates a graph document and a system utterance based on the prompt.
[0066] <S111> The dialogue control unit 11 obtains graph documents and system utterances from the large-scale language model 30.
[0067] <S112, S113> The dialogue control unit 11 stores the graph document acquired in step S111 in the current graph document 15. Similarly, the dialogue control unit 11 adds the acquired system utterance to the end of the dialogue history 14.
[0068] <S114> The dialogue control unit 11 transmits the current graph document 15 to the graph document storage and retrieval unit 41, which stores it in the graph document group 42 of the storage unit 12.
[0069] <S115> The dialogue control unit 11 then transmits the current graph document 15 and the dialogue history 14 to the graph document dialogue input / output unit 20.
[0070] After the processing in step S115, the process returns to step S101 and the series of processes is repeated. The process may then be terminated at any point, such as due to system shutdown.
[0071] (A-2-2) The processing diagram 10 of the graph document dialogue input / output unit 20 is a flowchart showing the characteristic operation of the graph document dialogue input / output unit according to the first embodiment.
[0072] <S201> The graph document interactive input / output unit 20 creates a graph document display / editing field 201 on the left side of the screen and an interactive history display / input field 202 on the right side, as shown in the example screen in Figure 2 above, and displays information. Initially, both the graph document and the interactive history start from an empty state.
[0073] <S202> When the user presses the send button 205, the graph document dialogue input / output unit 20 sends the user utterance (text entered in the text input field 204) and the graph document to the dialogue control unit 11. At the same time as sending, the graph document dialogue input / output unit 20 clears the text input field 204.
[0074] <S203> Subsequently, the graph document dialogue input / output unit 20 obtains the dialogue history 14 and the current graph document 15 from the dialogue control unit 11.
[0075] <S204> The graph document dialogue input / output unit 20 reflects the acquired dialogue history 14 and the current graph document 15 in the graph document display / editing field 201 and the dialogue history display / input field 202, making it possible to input user utterances.
[0076] After the processing in step S204, the process returns to step S202 and the series of processes is repeated. The process may then be terminated at any point, such as due to system shutdown.
[0077] (A-2-3) The processing diagram 11 of the graph document storage and search unit 41 is a flowchart showing the characteristic operation of the graph document storage and search unit according to the first embodiment.
[0078] <S301> The graph document storage and search unit 41 obtains the search query statement generated by the large-scale language model 30 in step S106 via the dialogue control unit 11.
[0079] <S302> The graph document storage and search unit 41 converts the search query statement into a vector, searches for a graph fragment with a vector similar to the vector of the search query statement from the graph document group 42 in the storage unit 12, and creates a list of vector search results as shown in Figure 6.
[0080] <S303> The graph document storage and search unit 41 adds graph fragments connected by specific labels such as "reference," "summary," and "function" to the list of nodes included in the list created in step S103 above, creating a final list of search results as shown in Figure 7, and transmits this list to the dialogue control unit 11.
[0081] (A-3) Effects of the First Embodiment According to the first embodiment, the following effects are achieved.
[0082] The graph document dialogue system 1 allows for dialogue using the large-scale language model 30 while displaying and storing the dialogue content in the form of a graph document. In other words, at any given time, the content of the dialogue up to that point is displayed and stored in the form of a graph document, allowing the user to confirm what the dialogue was about. Furthermore, by specifying any node displayed in the graph document and engaging in dialogue with that node, it is also possible to engage in dialogue with that node. In this case, when sending user utterances to the large-scale language model 30, the ID(s) of the specified nodes are also included, informing the large-scale language model 30 of the target node for dialogue and explicitly indicating that the dialogue is directed at that node, thereby enabling the acquisition of appropriate system utterances and graph documents. In addition, by informing the large-scale language model 30 not only of the target node ID but also of the dialogue history corresponding to that node, even more appropriate system utterances and graph documents can be acquired.
[0083] Furthermore, in the first embodiment, since the graph document is created with each node of the graph document and the underlying dialogue history of each node already associated, when a user refers to a node in the displayed graph document, it is possible to display the corresponding dialogue history on the screen with predetermined processing (e.g., focusing the corresponding history, highlighting the text). This predetermined processing can be achieved, for example, by obtaining the node ID when a node is referred to and applying existing special processing (focusing the corresponding history, highlighting the text) to the dialogue history associated with that node ID. Alternatively, when a user refers to a displayed dialogue history, the node of the corresponding graph document may be displayed on the screen with predetermined processing.
[0084] For example, even if a large-scale language model 30 causes hallucination at a node, the user can specify the suspicious node and immediately check the original text (corresponding dialogue history) to determine whether or not it is hallucination (i.e., improve dialogue quality).
[0085] Furthermore, since prompts sent to the large-scale language model 30 generally have length limitations, if the dialogue history to be sent becomes long, it is necessary to delete older dialogue history before sending. In this case, the content of the deleted dialogue history is not passed to the large-scale language model 30, resulting in a dialogue that does not take into account the information in the deleted dialogue history. In this embodiment, since the content of the dialogue is compactly represented in a graph document, the dialogue takes into account the information in the deleted dialogue history, and a more accurate system utterance can be obtained.
[0086] In this case, as described in Non-Patent Document 1, graph documents can represent the content of a dialogue in a structured and well-organized manner. Therefore, the large-scale language model 30 refers not only to the dialogue history but also to graph documents in which the dialogue content is appropriately organized, and performs system utterances that are in line with the graph documents. This results in a more focused and accurate dialogue that is closer to the actual dialogue content.
[0087] (B) Second Embodiments and subsequent embodiments of the information processing method, information processing program, and information processing apparatus according to the present invention will be described in detail with reference to the drawings.
[0088] In the second embodiment, the objective is to also utilize the summarized content of the graph document as a means of improving the dialogue quality.
[0089] (B-1) Diagram 12 of the configuration of the second embodiment is a block diagram showing the configuration of the graph document dialogue system according to the second embodiment.
[0090] In Figure 12, the graph document dialogue system 1A comprises a graph document dialogue device 10A, the graph document dialogue input / output unit 20 described above, and the large-scale language model 30 described above. Below, the graph document dialogue device 10A will be described, focusing on the differences from the first embodiment.
[0091] In addition to the above-mentioned configuration of the dialogue control unit 11, storage unit 12, and graph document storage / retrieval unit 41, the graph document summary unit 60 is added to the graph document dialogue device 10A.
[0092] The graph document summarization unit 60 has the function of summarizing the graph document created as the user continues to interact with the system. Naturally, as the interaction between the user and the system lengthens, the graph document becomes larger. In other words, as the relationships between nodes become more complex, it becomes difficult to grasp the content of the graph document at a glance.
[0093] As a countermeasure to the above, the graph document summarization unit 60 of the second embodiment transmits the graph document created by the large-scale language model 30 to the graph document dialogue input / output unit 20, and then performs a process to generate a graph document that summarizes the created graph document using the large-scale language model 30. The summarized graph document is transmitted to the graph document dialogue input / output unit 20 and can be referenced by the user as needed. Further details will be described in the operation section.
[0094] (B-2) Operation of the second embodiment Next, the operation of the graph document dialogue system 1A according to the second embodiment having the above configuration will be described.
[0095] Figure 13 is a flowchart showing the characteristic operation of the graph document dialogue device (dialogue control unit) according to the second embodiment.
[0096] (B-2-1) Processing of the graph document dialogue device 10A <S401> The dialogue control unit 11 executes the processing of steps S101 to S115 described above to generate a graph document and system utterances, and transmits the generated graph document and dialogue history to the graph document dialogue input / output unit 20.
[0097] <S402> After executing step S401 described above, the graph document summarization unit 60 generates a graph document summarizing the current graph document 15 using the large-scale language model 30. The method for summarizing the graph document (the content instructed in the prompt template 13A) is not particularly limited. For example, the number of nodes in the graph document may be limited (80% reduction from the number of nodes in the current graph document 15), or the graph document may be summarized taking into account the importance and centrality of sentences.
[0098] If necessary, a loop can be used to create a further summarized graph document (second stage) from the first summarized graph document (first stage). In this case, the prompt template may contain the summarized graph document (first stage) rather than the original graph document. This reduces the amount of information in the prompt template without omitting important details.
[0099] (B-2-2) The processing diagram 14 of the graph document dialogue input / output unit 20 is a flowchart showing the characteristic operation of the graph document dialogue input / output unit according to the second embodiment.
[0100] <S501> The graph document dialogue input / output unit 20 determines whether or not it has received the summarized graph document transmitted from the dialogue control unit 11 in step S402 described above. If the graph document dialogue input / output unit 20 has received the summarized graph document, it executes the next step S502; otherwise, it proceeds to step S503.
[0101] <S502> The graph document interaction input / output unit 20 holds the previously received unsummarized graph document together with the summarized graph document newly received in step S401 described above.
[0102] <S503> The graph document interaction input / output unit 20 determines whether it has received an operation from the user requesting to enlarge (or reduce) the graph document (for example, an instruction to enlarge or reduce using the mouse wheel and a specific key on the keyboard). If the graph document interaction input / output unit 20 has received a predetermined operation from the user requesting to enlarge (or reduce) the graph document, it executes the next step S504. On the other hand, if it has not received a predetermined operation, it returns to step S501.
[0103] <S504> When the graph document interactive input / output unit 20 receives a request to enlarge the graph document, it displays a graph document that has been reduced by one level from the current graph document in the graph document display / editing field 201. On the other hand, when the graph document interactive input / output unit 20 receives a request to reduce the graph document, it displays a graph document that has not been reduced by one level from the current graph document in the graph document display / editing field 201.
[0104] For example, if the graph document 100 shown in Figure 15 is currently displayed in the graph document display / editing area 201, and the user requests to enlarge the graph document, the graph document 00A shown in Figure 16, which is a summary of the graph document in Figure 15, will be displayed. If the graph document 100A shown in Figure 16 is currently displayed in the graph document display / editing area 201, and the user requests to reduce the graph document, the graph document 100 shown in Figure 15 will be displayed.
[0105] Although Figures 15 and 16 only show a one-stage summary, summarization may be performed in multiple stages. Furthermore, in the second embodiment, the graph document is displayed first, and then the summarized graph document is created in the background. However, as a variation, the summarized graph document may be created at the time a request for zooming in on the graph document (a request to roughly check the graph document) is received from the user. In this case, if the user has also provided a range to be summarized, the large-scale language model 30 may be instructed to summarize the graph document within that specified range.
[0106] The method of displaying the summarized nodes (nodes Y1 to Y3 in Figure 16) is not particularly limited. For example, as shown in Figure 16, the size (font size) of the summarized nodes Y1 to Y3 may be increased, or the font color, formatting, etc., may be changed. Furthermore, the summarized nodes Y1 to Y3 may be displayed overlaid on the original nodes N7 to N9, N10 to N12, and N17 to N19, respectively. In any case, the processing and representation methods for the summarized nodes are not particularly limited.
[0107] After the processing in step S504, the process returns to step S502 and the series of processes is repeated. The process may then be terminated at any point, such as due to system shutdown.
[0108] (B-3) Effects of the Second Embodiment According to the second embodiment, in addition to the effects of the first embodiment, the content of the graph document can be displayed in a summarized form according to the user's request, allowing the user to review the content at a glance. Furthermore, since the summaries of the graph document can be created and displayed in multiple stages, the most appropriate summary can be provided according to the user's request (contributing to user convenience). In addition, since the graph document included in the prompt sent to the large-scale language model can be summarized as appropriate, the amount of information in the prompt can be reduced without omitting important content.
[0109] (C) Third Embodiment and subsequent embodiments of the information processing method, information processing program, and information processing apparatus according to the present invention will be described in detail with reference to the drawings.
[0110] In the third embodiment, the objective is to use metadata such as the layout of an existing document when importing it as a graph document, as a means of improving dialogue quality.
[0111] (C-1) Diagram 17 of the configuration of the third embodiment is a block diagram showing the configuration of the graph document dialogue system according to the third embodiment.
[0112] In Figure 17, the graph document dialogue system 1B includes the graph document dialogue device 10B, the graph document dialogue input / output unit 20 described above, and the large-scale language model 30 described above. Below, the graph document dialogue device 10A will be described, focusing on the differences from the first embodiment.
[0113] In addition to the above-mentioned configuration of the dialogue control unit 11, storage unit 12, and graph document storage / retrieval unit 41, the graph document conversion unit 70 is added to the graph document dialogue device 10B.
[0114] The graph document conversion unit 70, upon importing an existing document T, works in conjunction with the graph document storage and search unit 41 to convert it into a graph document and store it in the graph document group 42. At this time, when the graph document conversion unit 70 imports the existing document T as a graph document, it also imports metadata of the existing document T (for example, layout (hierarchical information indicating which chapter, section, location, page, etc., each node of the graph document was located in), keywords, author, creation / modification date, etc.). Further details will be described in the operation section.
[0115] (C-2) Operation of the Third Embodiment Next, the operation of the graph document dialogue system 1 according to the third embodiment having the above configuration will be described.
[0116] Figure 18 is a flowchart showing the characteristic operation of the graph document dialogue device according to the third embodiment.
[0117] <S601> The graph document conversion unit 70 acquires an existing document T via an external storage medium or a communication network.
[0118] <S602> The dialogue control unit 11 creates a prompt for converting an existing document by, for example, embedding the existing document T in the existing document 13C-4 section of the prompt template 13C (Figure 19) for converting an existing document to a graph document.
[0119] <S603> The dialogue control unit 11 sends the prompt created in step S602 above to the large-scale language model 30. The large-scale language model 30 generates a graph document based on the prompt.
[0120] <S604> The dialogue control unit 11 acquires a graph document converted from the large-scale language model 30.
[0121] <S605, S606> The dialogue control unit 11 then transmits the converted graph document to the graph document storage and retrieval unit 41, which stores it in the graph document group 42 of the storage unit 12. When storing the converted graph document in the format of the graph document group 42 (vector conversion, etc.), metadata of the existing document T is also stored.
[0122] (C-3) Effects of the Third Embodiment According to the third embodiment, in addition to the effects described in the first embodiment, the following effects are achieved.
[0123] For example, when interacting with the system, the user can refer to (search) a group of graph documents 42 that hold metadata for existing documents T before engaging with the system. That is, the system can present the user with the basis for its utterances (which page, chapter, or section of the existing document they were found on). At that time, by simultaneously displaying the corresponding section of the existing document T based on that metadata, the user can make decisions based on the correct information in the existing document, even if the output of the large-scale language model includes hallucination.
[0124] (D) Fourth Embodiment and subsequent embodiments of the information processing method, information processing program, and information processing apparatus according to the present invention will be described in detail with reference to the drawings.
[0125] In the fourth embodiment, the objective is to improve the quality of the dialogue by collaborating with an external system (for example, by using information from the external system in a graph document).
[0126] (D-1) Diagram 20 of the configuration of the fourth embodiment is a block diagram showing the configuration of the graph document dialogue system according to the fourth embodiment.
[0127] In Figure 20, the graph document dialogue system 1C comprises a graph document dialogue device 10C, the graph document dialogue input / output unit 20 described above, the large-scale language model 30 described above, and an external system 85. Below, the graph document dialogue device 10C will be described, focusing on the differences from the first embodiment.
[0128] In addition to the above-mentioned configuration of the dialogue control unit 11, storage unit 12, and graph document storage / retrieval unit 41, the graph document dialogue device 10C also includes an external system linkage unit 80.
[0129] The external system linkage unit 80 accesses the external system 85 using information (generated by the large-scale language model 30) provided by the dialogue control unit 11 for accessing the external system 85.
[0130] The external system 85 is not particularly limited, but for example, it could be a patent information platform that allows searching of patent documents. In the fourth embodiment, the explanation will be based on the premise that the external system 85 is a site that handles intellectual property rights such as patents, utility models, designs, and trademarks.
[0131] Furthermore, in the fourth embodiment, a prompt template 13D, such as the one shown in Figure 21, is used to cause the large-scale language model 30 to generate information for coordinating with the external system 85.
[0132] In Figure 21, the prompt template 13D includes, in addition to the configuration shown in Figure 3, usage data 13D-1, the execution conditions for this prompt 13D-2, the next prompt template ID list 13D-3, and the prompt type 13D-4.
[0133] The data used, 13D-1, is used when data obtained from an external system, 85, etc. For example, patent search result data (CSV tabular data, etc.) is used.
[0134] The prompt execution conditions 13D-2 describe the execution conditions for each prompt (prompt template ID). The execution conditions include, for example, the execution of a prompt template for creating a patent search query, and the case where this prompt has not yet been executed. Here, "this prompt" refers to a prompt that is executed when the stage of the dialogue meets the conditions (execution conditions) (determined in step S702 described later). The prompt template ID is information that identifies each prompt template 13D. For example, since the functions used (such as creating a patent search query) differ depending on the stage of the dialogue, multiple prompt templates 13D with different prompt execution conditions are required. The prompt template ID is used to identify these multiple prompt templates 13D.
[0135] The next prompt template ID list 13-3 for this prompt shows information about the prompt template to be executed next. For example, in the case of a prompt template for creating a patent search query, it is the "prompt template ID for reading patent search results and creating a summary of each line".
[0136] Prompt type 13D-4 is used to specify the type of prompt that behaves differently from normal dialogue. For example, it can be used to specify a type that applies a prompt to each row of the data being used and stores the results in a table.
[0137] Further details will be provided in the section on operation.
[0138] (D-2) Operation of the fourth embodiment Next, the operation of the graph document dialogue system 1C according to the fourth embodiment having the above configuration will be described.
[0139] Figure 22 is a flowchart showing the characteristic operation (cooperation with external systems) of the graph document dialogue device according to the fourth embodiment.
[0140] (D-2-1) Processing of the graph document dialogue device 10C <S701> The dialogue control unit 11 executes the processing of steps S101 to S115 described above.
[0141] <S702> While the interaction between the user and the system (step S701) continues, the interaction control unit 11 determines the current stage of the interaction. Specifically, the interaction control unit 11 refers to the prompt execution condition 13-2 of the prompt template 13D and determines whether the stage of the interaction conforms to the prompt execution condition 13-2.
[0142] If the conditions are not met, the system will perform the dialogue and predetermined operations described in step S701 above (such as indicating the desire to use an external system from the graph document dialogue input / output unit 20, and specifying the content and conditions for using the external system), and the dialogue will continue until the conditions are met. Below, the explanation will continue assuming that the conditions for executing this prompt have been met.
[0143] <S703> The dialogue control unit 11 embeds information into each item of the prompt template 13D (Figure 21) related to the prompt ID determined by the determination in step S702 above, as necessary.
[0144] <S704> The dialogue control unit 11 sends the prompt created in step S703 described above to the large-scale language model 30. Based on the prompt, the large-scale language model 30 creates information (cooperation information) for cooperating with the external system 85.
[0145] <S705> The external system linkage unit 80 acquires information for linking with the external system 85 via the dialogue control unit 11.
[0146] <S706> The external system linkage unit 80 exchanges information with the external system 85 using information that links with the external system 85 as needed (such as patent search query statements).
[0147] <S707> The dialogue control unit 11 holds the information (patent search results) obtained from the external system 85 in step S706 described above. The held data will be used to generate a graph document, for example, by becoming information to be used in the next step.
[0148] After the processing in step S707, the process returns to step S701 and the series of processes is repeated. That is, a stage determination is made according to the conditions described in the next stage (the next prompt template ID list 13-3 of this prompt), and the same processing is performed. After that, the process may be terminated at any point, such as when the system shuts down.
[0149] (D-4) Effects of the Fourth Embodiment According to the fourth embodiment, in addition to the effects described in the first embodiment, high-quality dialogue can be performed by coordinating with external systems during interaction with the system. Furthermore, complex functions (searching for patent documents, etc., and arbitrary processing of search results) can be realized by switching the prompt template used for each stage. In this case, even if the actual search results are enormous in volume, the search results can be appropriately divided by adding conditions to the search and repeatedly sending them to the large-scale language model using the same prompt template, thereby enabling the same processing to be performed on all search results without any problems. Once the repeated processing is complete, the system proceeds to the next stage based on the stage determination, thereby enabling the realization of complex functions.
[0150] (E) Fifth Embodiment and subsequent embodiments of the information processing method, information processing program, and information processing apparatus according to the present invention will be described in detail with reference to the drawings.
[0151] In the fifth embodiment, the objective is to perform statistical processing as a means of improving the quality of the dialogue.
[0152] (E-1) Diagram 23 of the configuration of the fifth embodiment is a block diagram showing the configuration of the graph document dialogue system according to the fifth embodiment.
[0153] In Figure 23, the graph document dialogue system 1D includes the graph document dialogue device 10D, the graph document dialogue input / output unit 20 described above, the aggregation and analysis result input / output unit 25, the large-scale language model 30 described above, and the external system 85 described above. Below, the graph document dialogue device 10D will be described, focusing on the differences between the first and fourth embodiments.
[0154] In addition to the above-mentioned dialogue control unit 11, storage unit 12, graph document storage / retrieval unit 41, and external system linkage unit 80, the graph document dialogue device 10D also includes a dialogue status aggregation and analysis unit 90.
[0155] The dialogue status aggregation and analysis unit 90 is a functional unit that performs aggregation and analysis processing (statistical processing) of dialogue status using information from all users of the graph document dialogue device 10D (all users' dialogue history 14 and the current graph document 15).
[0156] The aggregation and analysis result input / output unit 25 inputs user utterances, displays user utterances and system utterances, and outputs the results of statistical processing performed by the dialogue situation aggregation and analysis unit 90.
[0157] (E-2) Operation of the Fifth Embodiment Next, the operation of the graph document dialogue system 1D according to the fifth embodiment having the above configuration will be described.
[0158] Figure 24 is a flowchart showing the characteristic operation of the graph document dialogue device according to the fifth embodiment.
[0159] (E-2-1) Processing of the graph document dialogue device 10D <S801> The dialogue control unit 11 executes the processing of steps S101 to S115 described above.
[0160] <S802> While the interaction between the user and the system continues, the dialogue control unit 11 and the dialogue status aggregation and analysis unit 90 receive a request for dialogue status analysis from the user via the aggregation and analysis result input / output unit 25 (for example, a verbal instruction for aggregation and analysis, or the pressing of a dedicated button to request aggregation and analysis).
[0161] <S803> The dialogue control unit 11 embeds information as necessary into each item of the prompt template 13C related to the prompt ID for the purpose of aggregation and analysis. In this case, the dialogue history 14 and the current graph document 15 use information held by all users (or information within the narrowed range if specific conditions are applied to narrow down the users, etc.).
[0162] <S804> The dialogue control unit 11 sends the prompt created in step S803 described above to the large-scale language model 30. The large-scale language model 30 creates aggregate analysis results based on the prompt.
[0163] <S805> The dialogue status aggregation and analysis unit 90 acquires the aggregation and analysis results via the dialogue control unit 11.
[0164] <S806> The dialogue status aggregation and analysis unit 90 transmits the aggregation and analysis results to the dialogue status aggregation and analysis unit 90.
[0165] (E-2-2) Processing of the Aggregation Analysis Result Input / Output Unit 25: When the Aggregation Analysis Result Input / Output Unit 25 receives the interactive aggregation analysis result transmitted in step S806 described above, it outputs it to the screen.
[0166] Figure 25 is an explanatory diagram showing an example of the screen configuration of the aggregation and analysis result input / output unit according to the fifth embodiment.
[0167] In Figure 25, the aggregation and analysis result input / output screen 300 has a dialogue history display / input field 301 and an aggregation and analysis result output field 302.
[0168] The dialogue history display / input field 301 is the same as the dialogue history display / input field 202 in Figure 2 above, so no explanation is provided.
[0169] The aggregated analysis results output field 302 is a field for outputting results based on aggregated analysis requests from users. In the example in Figure 25, user requests are aggregated by month based on the dialogue history and graph document information of all users. The analysis results are displayed in tabular format, and the requests are displayed as pie charts, but the method of representation is not particularly limited. Furthermore, statistical processing may also be performed on information obtained from the external system 85 shown in the fourth embodiment. For example, aggregated analysis may be performed on patent search results, and the aggregated analysis results may be displayed in the aggregated analysis results output field 302.
[0170] (E-3) Effects of the Fifth Embodiment According to the fifth embodiment, in addition to the effects described in the first and fourth embodiments, providing users with statistical results makes it possible to conduct high-quality conversations. In the first and fourth embodiments, conversations could be conducted on a per-user basis, but by using statistical results, it is possible to conduct conversations based on the conversation results of all users (or a narrowed range of users), so it is possible to conduct advanced conversations such as understanding the overall trends of conversations, what kind of conversations took place, and asking questions about the content. This makes it possible to consider the management and strategy of the entire conversation system.
[0171] (F) Other Embodiments The present invention is not limited to the first to fifth embodiments described above, and modified embodiments such as those exemplified below can also be cited.
[0172] (F-1) In the first to fifth embodiments described above, the graph document storage and search unit 41 was used to search the graph document group 42 for graph documents that could be used as reference during the interaction, and the results were used for the interaction. However, as a modification, the configuration of the graph document storage and search unit 41 and the graph document group 42, and the corresponding processing may be omitted.
[0173] (F-2) In the third embodiment described above, an example was shown in which the existing document T is held as a graph document group 42 independently of interaction with the system. However, interaction with the system may also be performed while specifying the existing document T via the graph document interaction input / output screen 200. That is, after converting the existing document T and holding it as a graph document group 42, the interaction is performed using the graph document information of the converted existing document T.
[0174] (F-3) In the fifth embodiment described above, a configuration specific to the fourth embodiment (external system linkage unit 80 and external system 85) is shown, but these configurations and functions may be omitted.
[0175] 1 (1A-1D)...Graph document dialogue system, 10 (10A-10D)...Graph document dialogue device, 11...Dialogue control unit, 12...Storage unit, 13...Prompt template, 14...Dialogue history, 15...Graph document, 20...Graph document dialogue input / output unit, 25...Aggregation and analysis result input / output unit, 30...Large-scale language model, 41...Search unit, 42...Graph document group, 50...Graph document dialogue management unit, 60...Graph document summarization unit, 70...Graph document conversion unit, 80...External system linkage unit, 85...External system, 90...Dialogue status aggregation and analysis unit, 200...Graph document dialogue input / output screen, 201...Editing field, 202...Input field, 203...Dialogue display field, 204...Text input field, 205...Send button, 300...Aggregation and analysis result input / output screen, 301...Input field, 302...Aggregation and analysis result output field, T...Existing document.
Claims
1. An information processing method that outputs information for displaying a dialogue history, which is a history of the content of the dialogue with a large-scale language model, and a graph document, which is a graph of the document structure of the dialogue content, on a display unit, characterized in that when an operation on the graph document is received, information for changing the display mode of the dialogue history or information for changing the display mode of the graph document is output.
2. The information processing method according to claim 1, characterized in that, upon receiving a first operation on the graph document, it outputs information for indicating the corresponding location of the dialogue history to the node of the graph document.
3. The information processing method according to claim 2, characterized in that the first operation is an operation that references the node.
4. The information processing method according to claim 1, characterized in that, upon receiving a second operation on the graph document, it outputs information for displaying the graph document with the reduced summarization level on the display unit.
5. The information processing method according to claim 4, characterized in that the second operation is an operation to request a change in the degree of summarization of a graph document.
6. The information processing method according to claim 4, characterized in that, upon receiving a first operation on the graph document whose summarization level has been changed, it outputs information to indicate the corresponding location in the dialogue history for the node of the graph document whose summarization level has been changed.
7. The information processing method according to claim 1, characterized in that an existing document is converted into a graph document containing metadata of the existing document using the large-scale language model.
8. The information processing method according to claim 1, characterized in that it uses the large-scale language model to create collaboration information for interacting with an external system, and uses the collaboration information to access the external system and obtain information from the external system.
9. The information processing method according to claim 1, characterized in that it performs aggregate analysis processing on the graph documents and dialogue history of all users using the large-scale language model, creates the results of the aggregate analysis processing, and outputs information for displaying the results of the aggregate analysis processing on the display unit.
10. An information processing program characterized by causing a computer to execute the information processing method described in any one of claims 1 to 9.
11. An information processing apparatus characterized by executing the information processing program described in claim 10.