Response generation apparatus, method, and program

Abstract

When the same question is entered into multiple conversational AIs to engage in multi-turn dialogues, the system maintains a natural conversation with each of the multiple conversational AIs. [Solution] The response generation device displays multiple dialogue screens side by side, each containing a question input by the user and an answer to that question, each from multiple types of response generation modules using different large-scale language models. When a question following the previous answer is input to the first dialogue screen, the previous answer displayed on the second dialogue screen is updated with the previous answer on the first dialogue screen. The response to the next question from the response generation module corresponding to the first dialogue screen is displayed on the first dialogue screen, and the response to the next question, reflecting the updated previous answer, from the response generation module corresponding to the second dialogue screen is displayed on the second dialogue screen.

Description

Technical Field

[0001] The present disclosure relates to a response generation device, a response generation method, and a response generation program.

Background Art

[0002] In recent years, high-performance large language models (LLMs) have been released one after another, and various conversational AIs (Artificial Intelligence) using LLMs have also been provided. In addition, services are provided that input the same text to multiple conversational AIs and output answers simultaneously.

[0003] As a technology related to multiple conversational AIs, for example, a task specialization type dialogue control unit that controls a task specialization type dialogue based on a user's input sentence and past dialogue history and outputs a task specialization type response, and a chit-chat type dialogue control unit that controls a chit-chat type dialogue based on the input sentence and dialogue history and outputs a chit-chat type response, and a response validity evaluation unit that obtains a response validity evaluation value for evaluating whether the chit-chat type response is a reasonable response from the dialogue history and the purpose of the task, and determines which of the chit-chat type response and the task specialization type response to adopt based on the response validity evaluation value have been proposed (see Patent Document 1).

[0004] Also, for example, a screen dialogue processing device that performs processing by interacting with a screen, and a screen dialogue processing device characterized by including a dialogue screen generation means for generating a plurality of dialogue screens based on one request has been proposed (see Patent Document 2).

[0005] Furthermore, a method has been proposed that includes, for example, the steps of: determining a natural language text string associated with a request based on a request received at an endpoint of a system of at least two or more communicably coupled computing devices; determining one or more guidance mechanisms associated with the request based at least on the natural language text string and a task associated with the endpoint; processing the natural language text string to generate a text result corresponding to the request using a language model and at least based on the one or more guidance mechanisms; and generating a response to the request using the endpoint, wherein the response includes at least the text result (see Patent Document 3). [Prior art documents] [Patent Documents]

[0006] [Patent Document 1] Japanese Patent Publication No. 2020-118842 [Patent Document 2] Japanese Patent Application Publication No. 11-3203 [Patent Document 3] Japanese Patent Publication No. 2024-43563 [Overview of the project] [Problems that the invention aims to solve]

[0007] Consider a multi-turn dialogue where the same question is input to multiple conversational AIs, and they simultaneously output answers. In this case, each conversational AI may output a different answer to the same question, so if the same question is input to each conversational AI in subsequent turns, the dialogue may break down. For example, if a question is asked in the next turn that is only included in the answer of one conversational AI, the answer from the other conversational AI may become inconsistent. Even if the other conversational AI manages to output an answer that allows the dialogue to barely be established in the next turn, the likelihood of the dialogue breaking down increases as the turns progress.

[0008] This disclosure is made in view of the above points, and aims to provide a response generation device, method, and program that can continue a natural conversation with each of the multiple conversational AIs when the same question is entered into multiple conversational AIs to engage in a multi-turn dialogue. [Means for solving the problem]

[0009] A response generation device according to a first aspect of this disclosure includes a processor which displays a plurality of dialogue screens in a row for each type of response generation module, each containing a question input by a user and an answer to the question, each answer provided by a plurality of types of response generation modules using different large-scale language models. When a question following a previous answer is input to the first dialogue screen, the previous answer displayed on the second dialogue screen is updated with the previous answer on the first dialogue screen, the answer to the next question provided by the response generation module corresponding to the first dialogue screen is displayed on the first dialogue screen, and the answer to the next question, reflecting the updated previous answer, provided by the response generation module corresponding to the second dialogue screen is displayed on the second dialogue screen.

[0010] A second aspect of the present disclosure relates to a response generation method in which a computer displays a plurality of dialogue screens in a row, each containing a question input by a user and an answer to the question, each containing an answer from a plurality of types of response generation modules using different large-scale language models. When a question following a previous answer is input to the first dialogue screen, the previous answer displayed on the second dialogue screen is updated with the previous answer on the first dialogue screen. The answer to the next question from the response generation module corresponding to the first dialogue screen is displayed on the first dialogue screen, and the answer to the next question, reflecting the updated previous answer, from the response generation module corresponding to the second dialogue screen is displayed on the second dialogue screen.

[0011] A response generation device according to a third aspect of this disclosure displays a plurality of dialogue screens in a row for each type of response generation module, each containing a question input by a user and an answer to the question, each answer provided by a plurality of types of response generation modules using different large-scale language models. When a question following a previous answer is input to the first dialogue screen, the computer updates the previous answer displayed on the second dialogue screen with the previous answer on the first dialogue screen, displays the answer to the next question provided by the response generation module corresponding to the first dialogue screen on the first dialogue screen, and displays the answer to the next question, reflecting the updated previous answer, provided by the response generation module corresponding to the second dialogue screen on the second dialogue screen. [Effects of the Invention]

[0012] According to the response generation device, method, and program described herein, when the same question is entered into multiple conversational AIs to conduct a multi-turn dialogue, it is possible to continue a natural dialogue with each of the multiple conversational AIs. [Brief explanation of the drawing]

[0013] [Figure 1]This is a block diagram showing the hardware configuration of the response generation device. [Figure 2] This is a block diagram showing an example of the functional configuration of a response generation device. [Figure 3] This figure shows an example of a dialogue screen. [Figure 4] This figure shows an example of a dialogue screen. [Figure 5] This figure shows an example of a dialogue screen. [Figure 6] This figure shows an example of a dialogue screen. [Figure 7] This diagram schematically illustrates one example of how to manage the history of a conversation using a tree structure. [Figure 8] This flowchart shows the flow of the response generation process. [Modes for carrying out the invention]

[0014] Hereinafter, an example of an embodiment of this disclosure will be described with reference to the drawings. In each drawing, identical or equivalent components and parts are given the same reference numerals. Also, the dimensions and proportions in the drawings are exaggerated for illustrative purposes and may differ from actual proportions.

[0015] Figure 1 is a block diagram showing the hardware configuration of the response generation device 10 according to this embodiment. As shown in Figure 1, the response generation device 10 includes a CPU (Central Processing Unit) 12, memory 14, storage device 16, input device 18, output device 20, storage medium reader 22, and communication I / F (Interface) 24. Each component is connected to the others via a bus 26 so as to be able to communicate with each other.

[0016] The memory device 16 stores a response generation program for executing the response generation process described later. The CPU 12 is a central processing unit that executes various programs and controls each component. That is, the CPU 12 reads a program from the memory device 16 and executes the program using the memory 14 as a work area. The CPU 12 controls each of the above components and performs various arithmetic processes according to the program stored in the memory device 16.

[0017] The memory 14 is composed of a RAM (Random Access Memory) and temporarily stores programs and data as a work area. The memory device 16 is composed of a ROM (Read Only Memory), HDD (Hard Disk Drive), SSD (Solid State Drive), etc., and stores various programs including an operating system and various data.

[0018] The input device 18 is a device for performing various inputs, such as a keyboard and a mouse. The output device 20 is a device for outputting various information, such as a display and a printer. By adopting a touch panel display as the output device 20, it may function as the input device 18.

[0019] The storage medium reader 22 reads data stored in various storage media such as a CD (Compact Disc)-ROM, DVD (Digital Versatile Disc)-ROM, Blu-ray disc, USB (Universal Serial Bus) memory, etc., and writes data to the storage medium. The communication I / F 24 is an interface for communicating with other devices, and for example, standards such as Ethernet (registered trademark), FDDI or Wi-Fi (registered trademark) are used.

[0020] Next, the functional configuration of the response generation device 10 according to this embodiment will be described.

[0021] Figure 2 is a block diagram showing an example of the functional configuration of the response generation device 10. As shown in Figure 2, the response generation device 10 includes, as a functional configuration, a reception unit 32, a display control unit 34, and a management unit 36. In addition, a response generation DB 40 and an interaction history DB 42 are stored in a predetermined storage area of ​​the response generation device 10. Each functional configuration is realized when the CPU 12 reads the response generation program stored in the storage device 16, expands it into memory 14, and executes it.

[0022] The response generation DB40 stores multiple types of response generation modules, each using a different LLM. The response generation module is the computational unit of an interactive AI that enables multi-turn dialogue by outputting an answer to an input question, and then outputting an answer to the next question based on past questions and answers. Examples of response generation modules include OpenAI's ChatGPT, Anthropic's Claude3, Meta's Llama3, Fugaku, and CyberAgent's OpenCalm.

[0023] The reception unit 32 receives questions from the user as text data. As will be described in detail later, in this embodiment, a question is entered for one of several conversational AIs, and this question is simultaneously reflected to the other conversational AIs. The reception unit 32 also receives information that identifies which conversational AI the entered question was entered for, for example, the identification information of the conversational AI. The question may also be entered by voice. In this case, the reception unit 32 converts the entered voice into text data using speech recognition.

[0024] Furthermore, the reception unit 32 receives various selection information input from the user. This selection information includes, for example, the type of conversational AI to use and the history of selections (details will be described later). The reception unit 32 then passes the received information to the display control unit 34.

[0025] The display control unit 34 displays multiple dialogue screens for each type of dialogue AI, including the question input by the user and the answer output by each dialogue AI to that question, on a display, which is an example of an output device 20.

[0026] Figure 3 shows an example of a dialogue screen 50. Note that Figure 3 is an example using two conversational AIs, but three or more conversational AIs may be used. As shown in Figure 3, the dialogue screen 50 includes a chat UI (User Interface) 52A of conversational AI of type A (hereinafter also referred to as "Chat A") and a chat UI 52B of conversational AI of type B (hereinafter also referred to as "Chat B"). In the following, when describing conversational AIs without distinguishing between types, the A and B at the end of the symbols will be omitted.

[0027] The chat UI 52 includes a selection component 54 for selecting the type of conversational AI. The selection component 54 may be, for example, a pull-down menu, and its options are identification information such as the name of the conversational AI whose response generation module is stored in the response generation DB 40. The type of conversational AI selected by the selection component 54 is received by the reception unit 32 as one of the selection pieces of information. The chat UI 52 also includes a dialogue display area 56 where the question entered by the user and the answer output by the conversational AI, i.e., the text data of the dialogue, are displayed. The chat UI 52 also includes a question input area 58 where the question entered by the user is entered. The question entered in the question input area 58 and the type of conversational AI selected by the selection component 54 are received by the reception unit 32.

[0028] The display control unit 34 also reflects the question entered in the first chat UI 52 to the second chat UI. Hereafter, the chat UI 52 in which the question was entered will be referred to as the "selected chat," and the other chat UI 52 will be referred to as the "non-selected chat." For example, suppose a question associated with the identification information of chat A is passed from the reception unit 32 to the display control unit 34. That is, as shown in Figure 4, a question is entered in the question input area 58A of chat UI 52A. In this case, as shown in S in Figure 4, the display control unit 34 copies the question entered in the question input area 58A to the question input area 58B of chat UI 52B.

[0029] The display control unit 34 uses the question entered in the question input area 58 as a prompt to input to the response generation module of the conversational AI selected by the selection component 54, and obtains the answer output from the response generation module. In other words, each conversational AI displaying the chat UI 52 on the dialogue screen 50 simultaneously generates an answer to the same question. The display control unit 34 displays the input question and the answers obtained from each of the response generation modules in the corresponding dialogue display area 56 of the chat UI 52.

[0030] In the example in Figure 4, we see example responses from Chat A and Chat B to the question, "Suggest some recommended tourist spots for the weekend." In this example, suppose that as the next question in the multi-turn, a question about "Hakone" (for example, "Hakone sounds nice. What should I go to?") that exists only in Chat A's response is entered into the question input area 58A of Chat UI 52A. In this case, since Chat B, which did not mention "Hakone" in its previous response, is also asked a similar question about "Hakone," Chat B's responses in subsequent turns may become inconsistent.

[0031] Therefore, in this embodiment, when a question following a previous answer is entered in the selected chat, the display control unit 34 updates the display in the dialogue display area 56 by overriding (overwriting) the previous answer displayed in the non-selected chat with the previous answer in the selected chat. Then, the display control unit 34 copies the question entered in the selected chat to the non-selected chat. As a result, when the response generation module outputs an answer to a newly entered question in each dialogue AI, reflecting the answers up to the previous time, the non-selected chat outputs an answer that reflects the updated answers up to the previous time.

[0032] For example, as shown in Figure 5, suppose the question, "Hakone sounds nice. What places should I go?" is entered into the question input area 58A of the chat UI 52A. In this case, the display control unit 34 overrides the previous answer portion in the dialogue display area 56A with the previous answer portion in the dialogue display area 56B, as shown in T in Figure 5. As a result, in chat B as well, the previous answer will include content related to "Hakone" (the dashed line portion in Figure 5). Therefore, as shown in S in Figure 5, even if the question entered in chat A is reflected in chat B, it is possible to naturally answer the question about "Hakone" in chat B as well.

[0033] As will be described in more detail later, in this embodiment, the history of questions and answers is managed in the dialogue history DB 42, and when an answer is updated, both the original answer and the updated answer are stored in the dialogue history DB 42. Therefore, as shown in Figure 6, the display control unit 34 displays a selection component 60 that allows the user to select between the original answer and the updated answer, corresponding to each answer displayed in the dialogue display area 56. In the example in Figure 6, the selection component 60 includes the notation "n / N" indicating that the nth answer out of N answers stored as answers to the same question is being displayed, and buttons to increase or decrease n. The information of the nth answer selected by the selection component 60 is received by the reception unit 32 as history selection information. The display control unit 34 deletes the questions and answers after the selected nth answer from the dialogue display area 56, and reads the nth answer from the dialogue history DB 42 to update the display in the dialogue display area 56.

[0034] The management unit 36 ​​manages the history of questions and answers in the dialogue history DB 42 for each type of response generation module, that is, for each type of conversational AI, as a tree structure history branched for each corresponding question, including the original answer output by the response generation module and the answer updated based on the selected chat.

[0035] Figure 7 schematically shows an example of a method for managing the history of a conversation using a tree structure. Qi is the question entered in the i-th turn, and Ai is the conversational AI's answer to Qi. When the answer is updated, the management unit 36 ​​assigns branch number 1 to the original answer, and each time the answer is updated, it increments the branch number and assigns it to the updated answer.

[0036] The example in Figure 7 is a tree structure showing the history of the conversation in Chat A in Figures 4 to 6. Q1 is "Suggest some recommended tourist spots for the weekend," and A1 is the answer to Q1, which includes content about "Hakone." In response to this answer, the next question Q2, "Hakone sounds nice. What should I go to?" is entered into Chat A, so Q2 is saved under A1. A2 is the answer to Q1, which does not include content about "Hakone Shrine," but Chat B's answer does include content about "Hakone Shrine." In response to this answer, the next question Q3, "Oh, Hakone Shrine, that's wonderful. What kind of power spot is it?" is entered into Chat B, in which case answer A2 is overwritten by Chat B's answer. Therefore, the management unit 36 ​​saves the original answer A2 as A2-1, and also saves the updated answer A2-2 under question Q2, branching off from A2-1. Then, the management unit 36 ​​saves question Q3 under answer A2-2, which is associated with it. In this case, a selection component 60 such as "← 2 / 2 →" is displayed corresponding to answer A2-2 in the interactive display area 56A. Note that "←" and "→" represent buttons in the example in Figure 6.

[0037] Next, the operation of the response generation device 10 according to this embodiment will be described.

[0038] Figure 8 is a flowchart showing the flow of the response generation process executed by the CPU 12 of the response generation device 10. The CPU 12 reads the response generation program from the storage device 16, loads it into memory 14, and executes it. As a result, the CPU 12 functions as one of the various functional components of the response generation device 10, and the response generation process shown in Figure 8 is executed.

[0039] First, in step S10, the display control unit 34 displays the dialogue screen 50 in its initial state, for example, as shown in Figure 3. At this time, the reception unit 32 receives the number of types of conversational AI to be used from the user, and the display control unit 34 displays the dialogue screen 50, which includes the number of chat UIs 52 received.

[0040] Next, in step S12, the reception unit 32 receives the type of interactive AI to be used from the selection component 54 of each chat UI 52 included in the dialogue screen 50. Next, in step S14, the reception unit 32 receives the input of the first question to the question input area 58 in any of the chat UI 52. Then, the reception unit 32 associates the question entered in the question input area 58 with the type of interactive AI selected in the selection component 54 included in the same chat UI as the question input area 58 in which the question was entered, and passes it to the display control unit 34.

[0041] Next, in step S16, the display control unit 34 copies the question entered in the question input area 58 of the selected chat to the question input area 58 of the unselected chat. Next, in step S18, the display control unit 34 uses the question entered in the question input area 58 as a prompt to input to the response generation module of the conversational AI selected by the selection component 54, and obtains the answer output from the response generation module. That is, each of the conversational AIs displaying the chat UI 52 on the dialogue screen 50 simultaneously generates an answer to the same question. Then, the display control unit 34 updates the dialogue screen 50 by displaying the input question and the answers obtained from each of the response generation modules in the corresponding dialogue display area 56 of the chat UI 52.

[0042] Next, in step S20, the management unit 36 ​​saves the question and answer history in a tree structure in the dialogue history DB 42 for each type of conversational AI. Then, in step S22, the reception unit 32 determines whether or not it has received a question for the next turn. If it has received a question for the next turn, it proceeds to step S24; otherwise, it proceeds to step S28.

[0043] In step S24, the display control unit 34 updates the display in the dialogue display area 56 by overriding the previous answer in the selected chat with the previous answer displayed in the non-selected chat. Next, in step S26, the management unit 36 ​​adds the overridden answer to the tree structure of the dialogue history, branching off from the original answer below the previous question, and returns to step S16.

[0044] Meanwhile, in step S28, the reception unit 32 determines whether or not it has received the selection information from the history based on the operation of any of the selection components 60 in the chat UI 52. If the selection information from the history has been received, the process proceeds to step S30; otherwise, it proceeds to step S32. In step S30, the display control unit 34 deletes the questions and answers from the dialogue display area 56 that are after the answer indicated by the selection information from the history. Then, the display control unit 34 updates the dialogue screen 50 by reading the selected answer from the tree structure managed in the dialogue history DB 42 and displaying it in the dialogue display area 56, and returns to step S22.

[0045] In step S32, the reception unit 32 determines whether to terminate the currently ongoing multi-turn dialogue. For example, it determines to terminate the dialogue if the termination of the dialogue screen 50 is instructed, or if a different theme from the currently ongoing dialogue is selected. If the dialogue is not terminated, the process returns to step S22; if the dialogue is terminated, the response generation process ends.

[0046] As described above, the response generation device according to this embodiment displays multiple chat UIs for each type of conversational AI, including a question input by the user and an answer to that question, each answer provided by multiple types of conversational AIs using different LLMs. Furthermore, when a question following a previous answer is input to the first chat UI (selection chat), the response generation device updates the previous answer displayed in the second chat UI (non-selection chat) with the previous answer in the selection chat. The response generation device then displays the answer to the next question from the conversational AI corresponding to the selection chat in the selection chat, and displays the answer to the next question from the conversational AI corresponding to the non-selection chat, reflecting the updated previous answer, in the non-selection chat. This allows for a natural conversation to be continued with each of the multiple conversational AIs when the same question is input to multiple conversational AIs and a multi-turn dialogue is performed.

[0047] In the above embodiment, we described the case where two types of conversational AI are used, but as mentioned above, three or more types of conversational AI may be used. In this case, the chat UI into which the question is entered is the selection chat, and the other two or more chat UIs are non-selection chats.

[0048] Furthermore, during a multi-turn conversation, the types of conversational AIs used may be added, removed, or changed. In this case, the newly added or changed conversational AI may be displayed in the chat UI 52 with the previous questions and conversations reflected, by referring to the conversation history.

[0049] Furthermore, the response generation process, which the CPU reads and executes in the above embodiment, may be executed by various processors other than the CPU. Examples of such processors include PLDs (Programmable Logic Devices) such as FPGAs (Field-Programmable Gate Arrays) whose circuit configuration can be changed after manufacturing, and dedicated electrical circuits that are processors with circuit configurations specifically designed to execute specific processes, such as ASICs (Application Specific Integrated Circuits). The response generation process may be executed by one of these various processors, or by a combination of two or more processors of the same or different types (for example, multiple FPGAs, and a combination of a CPU and an FPGA). A GPU (Graphics Processing Unit) may also be used as the processor. More specifically, the hardware structure of these various processors is an electrical circuit that combines circuit elements such as semiconductor elements.

[0050] Furthermore, although the above embodiment describes a configuration in which the response generation program is pre-stored (installed) in a storage device, the invention is not limited to this configuration. The program may be provided in a form recorded on a recording medium such as a CD-ROM, DVD-ROM (Digital Versatile Disc Read Only Memory), or USB (Universal Serial Bus) memory. Alternatively, the program may be provided in a form that can be downloaded from an external device via a network. [Explanation of symbols]

[0051] 10 Response generation device 12 CPU 14 memory 16 Storage device 18 Input device 20 Output device 22 Storage medium reader 24 Communication I / F 26 bus 32 Reception Department 34 Display Control Unit 36 Management Department 40 Response Generation DB 42 Dialogue History Database 50 Dialogue screen 52 Chat UI 54 Selectable parts 56 Dialogue display area 58 Question Input Area 60 Selectable Parts

Claims

1. Including the processor, The aforementioned processor, Multiple dialogue screens are displayed side-by-side, each containing a question entered by the user and an answer to that question, each from a different type of response generation module using a different large-scale language model, with each type of response generation module representing a different type of response generation module. When a follow-up question to the previous answer is entered in the first dialogue screen, the previous answer displayed in the second dialogue screen is updated with the previous answer in the first dialogue screen. The response generation module corresponding to the first dialogue screen displays the answer to the next question on the first dialogue screen, and the response generation module corresponding to the second dialogue screen displays the answer to the next question, reflecting the updated previous answer, on the second dialogue screen. Response generation device.

2. The response generation device according to claim 1, wherein for each type of response generation module, the history of questions and answers is managed as a tree structure history in which the original answer output by the response generation module and the answer updated based on the first dialogue screen are branched for the corresponding question.

3. For each answer displayed on the aforementioned dialogue screen, it is possible to select between the original answer and the updated answer. Based on the history of the aforementioned tree structure, the currently displayed answer is updated to the selected answer. The response generation device according to claim 2.

4. Computers Multiple dialogue screens are displayed side-by-side, each containing a question entered by the user and an answer to that question, each from a different type of response generation module using a different large-scale language model, with each type of response generation module representing a different type of response generation module. When a follow-up question to the previous answer is entered in the first dialogue screen, the previous answer displayed in the second dialogue screen is updated with the previous answer in the first dialogue screen. The response generation module corresponding to the first dialogue screen displays the answer to the next question on the first dialogue screen, and the response generation module corresponding to the second dialogue screen displays the answer to the next question, reflecting the updated previous answer, on the second dialogue screen. Response generation method.

5. On the computer, Multiple dialogue screens are displayed side-by-side, each containing a question entered by the user and an answer to that question, each from a different type of response generation module using a different large-scale language model, with each type of response generation module representing a different type of response generation module. When a follow-up question to the previous answer is entered in the first dialogue screen, the previous answer displayed in the second dialogue screen is updated with the previous answer in the first dialogue screen. The response generation module corresponding to the first dialogue screen displays the answer to the next question on the first dialogue screen, and the response generation module corresponding to the second dialogue screen displays the answer to the next question, reflecting the updated previous answer, on the second dialogue screen. A response generation program to perform a task.

Images