Natural language generation device, natural language generation method, and program

Abstract

To provide a technology that enables more efficient correction of a dialogue system that generates answers based on information provided by a user. [Solution] An information processing device that updates constraint information that affects answers generated in a dialogue with a user by a dialogue device that provides a dialogue interface that generates answers based on information provided by the user, and is equipped with an extraction unit that extracts areas requiring improvement based on a record of the dialogue, a generation unit that accepts input of knowledge information useful for improving the areas requiring improvement and generates update instruction information for updating the constraint information based on the knowledge information, and an update unit that updates the constraint information based on the update instruction information.

Description

[Technical field]

[0001] The present invention relates to a natural language generation device, a natural language generation method, and a program technology. [Background technology]

[0002] Conventionally, dialogue systems using large language models (LLMs) have been studied. For example, Patent Literature 1 describes a system that generates a dialogue model by performing machine learning using chat dialogue information as training data, and answers questions from users using the dialogue model. [Prior art documents] [Patent documents]

[0003] [Patent Document 1] Patent No. 7542904 Summary of the Invention [Problem to be solved by the invention]

[0004] However, in the conventional technology, when a dialogue system cannot give an appropriate response to a specific pattern, a lot of effort is spent on correcting the dialogue system. For example, it takes time to collect failure cases for analysis. In addition, for example, when correcting a dialogue model, it takes time because expertise is required to collect learning data and re-learn the model. In addition, expertise is also required to correct an external data source used by the dialogue system, and it is also time-consuming to formulate a policy on how to make the correction. As such, in the conventional technology, there are cases where it is not possible to efficiently correct a dialogue system using natural language.

[0005] In view of the above circumstances, an object of the present invention is to provide a technique that enables more efficient correction of a dialogue system that generates answers based on information provided by a user. [Means for solving the problem]

[0006] One aspect of the present invention is an information processing device that updates constraint information that affects answers generated in a dialogue with a user by a dialogue device that provides a dialogue interface that generates answers based on information provided by the user, and the information processing device includes an extraction unit that extracts areas requiring improvement based on a record of the dialogue, a generation unit that accepts input of knowledge information useful for improving the areas requiring improvement and generates update instruction information for updating the constraint information based on the knowledge information, and an update unit that updates the constraint information based on the update instruction information.

[0007] One aspect of the present invention is the above-mentioned information processing device, wherein the extraction unit presents the extracted areas needing improvement to an administrator of the dialogue device, and the generation unit accepts input of the knowledge information by the administrator.

[0008] One aspect of the present invention is the above-mentioned information processing device, further comprising an input unit for inputting the knowledge information to the generation unit, wherein the extraction unit supplies the areas needing improvement to the input unit, and the input unit extracts knowledge information useful for improvement for the areas needing improvement supplied by the extraction unit from an external data source and inputs the knowledge information to the generation unit.

[0009] One aspect of the present invention is the above information processing device, wherein the input unit is configured using a dialogue model that is trained to respond to an input from the user.

[0010] One aspect of the present invention is the above-mentioned information processing device, further comprising a dialogue unit that provides an interface for dialogue with an administrator of the dialogue device, wherein the dialogue unit identifies information required as the knowledge information based on the areas needing improvement and presents it to the administrator, accepts input of the knowledge information from the administrator, and inputs the knowledge information input by the administrator to the generation unit.

[0011] One aspect of the present invention is the above information processing device, wherein the dialogue unit is configured using a dialogue model that is trained to provide a reply to an input from the user.

[0012] One aspect of the present invention is the above-mentioned information processing device, wherein when feedback information regarding modifications to the generated update instruction information is input from the administrator, the generation unit modifies the update instruction information based on the feedback information.

[0013] One aspect of the present invention is the above-mentioned information processing device, wherein the generation unit stores both first constraint information, which is the constraint information before the update, and second constraint information, which is the constraint information after it has been updated by update instruction information that has been modified based on the feedback information, and the dialogue unit presents to the administrator outputs obtained when the same dialogue is executed using each of the first constraint information and the second constraint information in a comparable manner.

[0014] One aspect of the present invention is the above-mentioned information processing device, wherein the extraction unit and the generation unit are configured using a dialogue model trained to provide a response to an input from the user.

[0015] One aspect of the present invention is the above-mentioned information processing device, wherein the generation unit performs preprocessing on the input knowledge information to shape the format or extract key information for updating the constraint information, and inputs the preprocessed knowledge information to the dialogue model.

[0016] One aspect of the present invention is the above-mentioned information processing device, wherein the extraction unit classifies the causes of the dialogue extracted as the area requiring improvement, into categories corresponding to the modification policy of the constraint information, and presents the categories to the administrator.

[0017] One aspect of the present invention is the above-mentioned information processing device, wherein the dialogue device dialogues with the user using language-based dialogue information exchanged between the user and the dialogue device and visual information that supplements the dialogue information, and the record of the dialogue includes historical information relating to the dialogue information and the visual information.

[0018] One aspect of the present invention is the above-mentioned information processing device, wherein the dialogue interface is provided as a robot capable of physical movements, and the historical information regarding the visual information includes historical information regarding the physical movements performed by the robot.

[0019] One aspect of the present invention is the above-mentioned information processing device, wherein the dialogue device is equipped with a camera for reading the user's facial expressions, and the record of the dialogue includes the dialogue information and an image captured by the camera or historical information regarding the user's facial expressions recognized from the image.

[0020] One aspect of the present invention is the above-mentioned information processing device, wherein the dialogue device is equipped with a microphone for sensing the state of the sound environment in the space in which the dialogue device interacts with the user, and the record of the dialogue includes the dialogue information and audio captured by the microphone or historical information regarding the state of the sound environment recognized from the audio.

[0021] One aspect of the present invention is the above-mentioned information processing device, wherein the dialogue record includes the dialogue information and historical information regarding the user's voice or the user's evaluation of the dialogue device recognized from the voice.

[0022] One aspect of the present invention is the information processing device described above, wherein the record of the dialogue includes history information regarding a communication error of the dialogue device.

[0023] One aspect of the present invention is an information processing method in which a computer updates constraint information that influences answers generated in a dialogue with a user by a dialogue device that provides a dialogue interface in which answers are generated based on information provided by the user, in which the computer extracts areas requiring improvement based on a record of the dialogue, accepts input of knowledge information that is useful for improving the areas requiring improvement, generates update instruction information for updating the constraint information based on the knowledge information, and updates the constraint information based on the update instruction information.

[0024] One aspect of the present invention is a program for causing a computer to execute an update process for updating constraint information that influences answers generated in a dialogue with a user by a dialogue device that provides a dialogue interface for generating answers based on information provided by the user, the update process including a process for extracting areas requiring improvement based on a record of the dialogue, a process for accepting input of knowledge information useful for improving the areas requiring improvement, and generating update instruction information for updating the constraint information based on the knowledge information, and a process for updating the constraint information based on the update instruction information. Effect of the Invention

[0025] The present invention allows for more efficient modification of dialogue systems that generate responses based on information provided by users. [Brief description of the drawings]

[0026] [Figure 1] 1 is a diagram illustrating an example of a system configuration of a dialogue system 1 according to an embodiment. [Diagram 2] 1 is a diagram illustrating an example of a configuration of a dialogue device 100 according to an embodiment. [Diagram 3] FIG. 11 is a diagram (part 1) showing an example of a process flow in which the dialogue device 100 according to the embodiment updates constraint information of a dialogue model through a dialogue with a user. [Figure 4]FIG. 2 is a diagram (part 2) illustrating an example of a process flow in which the dialogue device 100 according to the embodiment updates constraint information of a dialogue model through a dialogue with a user. [Diagram 5] FIG. 11 is a diagram (part 3) illustrating an example of a processing flow in which the dialogue device 100 according to the embodiment updates constraint information of a dialogue model through a dialogue with a user. [Figure 6] FIG. 4 is a diagram (part 4) illustrating an example of a process flow in which the dialogue device 100 according to the embodiment updates constraint information of a dialogue model through a dialogue with a user. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0027] Hereinafter, an embodiment of the present invention will be described in detail with reference to the drawings.

[0028] FIG. 1 is a diagram showing an example of a system configuration of a dialogue system 1 according to an embodiment. The dialogue system 1 is a system that accepts input of natural language from a user and outputs a response to the natural language. The dialogue system 1 includes, for example, a dialogue device 100, a user terminal device 200, and an administrator terminal device 300. The dialogue device 100, the user terminal device 200, and the administrator terminal device 300 can communicate with each other via a network NW. The network NW may be a network using wireless communication or a network using wired communication. The network NW may be configured using, for example, the Internet or a local area network (LAN). The network NW may be configured by combining a plurality of networks.

[0029] The user terminal device 200 is a terminal device used by a user U of the dialogue system 1. For example, the user terminal device 200 may be a terminal device such as a smartphone, a tablet, or a personal computer. A user interface operates on the user terminal device 200. By operating the user interface, the user U can input any natural sentence to the dialogue device 100 and obtain a response to the natural sentence from the dialogue device 100. The user interface may be a dedicated application program or a web application provided via a web browser.

[0030] The administrator terminal device 300 is a terminal device used by an administrator M of the dialogue system 1. For example, the administrator terminal device 300 may be a terminal device such as a smartphone, a tablet, or a personal computer. An administrator interface runs on the user terminal device 300. The administrator M can transmit and receive information to and from the dialogue device 100 by operating the administrator interface. The administrator interface may be a dedicated application program or a web application provided via a web browser.

[0031] FIG. 2 is a diagram showing an example of the configuration of the dialogue device 100 of the embodiment. The dialogue device 100 includes, for example, a communication unit 110, a storage unit 120, and a control unit 130. The control unit 130 is configured using, for example, a processor such as a CPU (Central Processing Unit) and a memory. The control unit 130 functions as a dialogue unit 131, an improvement-required part extraction unit 132, an update instruction generation unit 133, and an update execution unit 134 by the processor executing a program. All or part of the functions of the control unit 130 may be realized using hardware such as an ASIC (Application Specific Integrated Circuit), a PLD (Programmable Logic Device), or an FPGA (Field Programmable Gate Array). The above program may be recorded on a computer-readable recording medium. The computer-readable recording medium is, for example, a portable medium such as a flexible disk, a magneto-optical disk, a ROM, a CD-ROM, or a semiconductor storage device (for example, an SSD: Solid State Drive), or a storage device such as a hard disk or a semiconductor storage device built into a computer system. The above program may be transmitted via an electric communication line.

[0032] The communication unit 110 is a communication device that connects the interactive device 100 to the network NW. The communication unit 110 is, for example, a network interface. The network interface may be a device that performs wireless communication or a device that performs wired communication.

[0033] The storage unit 120 is configured using a storage device such as a magnetic hard disk device or a semiconductor storage device. The storage unit 120 stores, for example, a dialogue log 121, an update patch 122, and an alternative file 123. Details of the dialogue log 121, the update patch 122, and the alternative file 123 will be described later.

[0034] The control unit 130 acquires information on natural language input by a user to the user terminal device 200 from the user terminal device 200, generates a response to the natural language, and returns the response to the user terminal device 200. More specifically, the control unit 130 includes a dialogue unit 131, an improvement-requiring part extraction unit 132, an update instruction generation unit 133, and an update execution unit 134.

[0035] The dialogue unit 131 executes dialogue processing to generate a response to a natural sentence supplied from the user terminal device 200 and transmit the response to the user terminal device 200. For example, the dialogue unit 131 may be configured with a large language model (LLM) trained to respond to an input natural sentence. For example, GhatGPT (registered trademark), Claude (registered trademark), Gemini (registered trademark), or the like may be used as the LLM. In this case, the dialogue unit 131 inputs the supplied natural sentence into the LLM, and can obtain a response to the natural sentence as the output.

[0036] More specifically, prompt engineering (for example, a technique called "Policy Prompt") is applied to the LLM of the dialogue unit 131, which controls the output of the LLM using constraint information based on specific policies, rules, knowledge, etc. In the present embodiment, it is assumed that the constraint information is stored in an external data source. Note that a technique that enables the LLM to refer to an external data source is, for example, a technique called "RAG (Retrieval-Augmented Generation)." In this way, the dialogue device 100 of the present embodiment stores the constraint information in the external data source and allows the LLM to refer to it. As a result, the dialogue device 100 of the present embodiment can tune the LLM without relearning it by changing the constraint information stored in the external data source.

[0037] In addition to the above, the dialogue unit 131 stores information related to the dialogue processing as a dialogue log 121 in the storage unit 120 so that it can be referenced later. For example, the dialogue log 121 includes natural sentences input from the user terminal device 200 and the contents of the responses to the natural sentences by the dialogue unit 131. Furthermore, the dialogue log 121 may also include a history of data processing performed in the process of generating responses to the input natural sentences, a reference history of the external data source 40, and the like.

[0038] The improvement-requiring part extraction unit 132 extracts parts that should be improved (improvement-requiring parts) from the contents of the dialogue recorded in the dialogue log 121. The improvement-requiring parts can be referred to as cases in which the dialogue has failed (failure cases) among a group of dialogue cases conducted with the user. For example, the improvement-requiring part extraction unit 132 can be configured with an LLM. In this case, for example, the improvement-requiring part extraction unit 132 can input the dialogue log 121 and a natural sentence that instructs to extract improvement-requiring parts from the contents of the dialogue log 121 to the LLM, and obtain the improvement-requiring parts of the dialogue as the output.

[0039] For example, the improvement-required part extraction unit 132 inputs the dialogue log 121 to the LLM, and inputs an instruction to the LLM such as "Extract cases where location guidance failed from the input dialogue log." For example, if an exchange in which the question "Where is the restroom?" is answered with "It's at the south end of the second floor of the main building," is recorded in the dialogue log, the LLM is expected to determine that an appropriate dialogue has taken place since a location has been answered to the question "Where?". On the other hand, if an exchange in which the same question is answered with "Look around, you'll see it," is recorded in the dialogue log, the LLM may determine that an appropriate dialogue has not taken place since a location has not been answered to the question "Where?". In this way, it is expected that the LLM may recognize that a context-dependent dialogue needs improvement even though it is an appropriate dialogue.

[0040] In anticipation of such a case, the improvement-requiring part extraction unit 132 may be configured to determine the improvement-requiring parts using the following rule-based method. (1) Cases in which the dialogue time is too long (greater than a threshold time) are extracted as failure cases. (2) Cases in which user U repeatedly asks the same question are extracted as failure cases. (3) Cases in which the dialogue was terminated midway (for example, cases in which there was no phrase suggesting the end of the dialogue, such as "thank you" or "see you later") are extracted as failure cases. The improvement-needing part extraction unit 132 may extract the improvement-needing parts by instructing the LLM to extract failure cases based on the above-mentioned rules.

[0041] Furthermore, the improvement-required part extraction unit 132 may be configured to extract, when a failure of internal processing related to a dialogue is detected, the dialogue taking place at that time as a failure case. For example, when the dialogue system 1 has a function that enables the LLM to call and execute an external function, such as the "Function Calling" function of ChatGPT (registered trademark), the dialogue unit 131 is configured to save a communication log with the external function in addition to the dialogue log 121. In this case, the improvement-required part extraction unit 132 may be configured to determine whether the call to the external function was successful based on the communication log, and to extract, as a failure case, a dialogue when the call to the external function failed (a request error occurred).

[0042] In addition to these methods, the improvement-requiring area extraction unit 132 may be configured to determine whether or not a dialogue is a failure case by inputting the dialogue log of the dialogue to be judged into a learned model that has been trained by machine learning to input the dialogue log of each dialogue and output the success or failure of the dialogue.

[0043] The update instruction generating unit 133 generates an instruction (update instruction) for updating the constraint information stored in the external data source. More specifically, the update instruction generating unit 133 receives input of knowledge information (knowledge information) that serves as material for determining how to improve the LLM for the improvement-requiring parts of the dialogue extracted by the improvement-requiring part extracting unit 132, and generates an update instruction based on the knowledge information.

[0044] For example, the update instruction generating unit 133 generates, as an update instruction, an alternative file 123 describing the updated constraint information and an update patch 122, which is a program for updating the constraint information stored in the external data source to the content described in the alternative file. For example, when a dialogue scene in which the location of the toilet in the premises cannot be properly guided is extracted as an area requiring improvement, text information such as "The toilet is located at the south end of the second floor of the main building. Please generate an update patch and an alternative file for adding this knowledge to the external data source" is input to the update instruction generating unit 133. In this case, for example, the update instruction generating unit 133 recognizes the content of the knowledge to be added based on the text information, and generates an alternative file 123 containing the knowledge and an update patch 122 for adding the knowledge of the alternative file 123 to the external data source. The update instruction generating unit 133 stores the alternative file 123 and the update patch 122 thus generated in the storage unit 120.

[0045] The update execution unit 134 updates the constraint information stored in the external data source based on the update instruction generated by the update instruction generation unit 133. For example, the update execution unit 134 executes the update patch 122 to update the constraint information to the content described in the alternative file 123.

[0046] [First Example] Fig. 3 is a diagram (part 1) showing an example of a process flow in which the dialogue device 100 of the embodiment updates constraint information of a dialogue model (LLM) through dialogue with a user. In Fig. 3, S101 to S105 represent phases in which the user U and the dialogue device 100 dialogue with each other (hereinafter referred to as "dialogue phases"), and S201 to S206 represent phases in which the dialogue function is improved by modifying constraint information stored in the external data source 400 based on the content of the dialogue performed in the dialogue phase (hereinafter referred to as "improvement phases").

[0047] [Dialogue Phase] In the dialogue phase, first, the user U inputs a natural sentence to the dialogue device 100 via the user terminal device 200 (S101). In the dialogue device 100, the dialogue unit 131 generates an answer to the natural sentence input in S101 based on the constraint information (policy 41 and knowledge 42) stored in the external data source 400 (S102). The dialogue unit 131 presents the answer generated in S102 to the user U (S104). Meanwhile, the dialogue unit 131 records information related to the dialogue processing of S102 in the dialogue log 121 (S105). S101 to S105 may be executed repeatedly.

[0048] [Improvement Phase] Next, in the improvement phase, first, the improvement-needing part extraction unit 132 refers to the dialogue log 121 and extracts parts (improvement-needing parts) that are thought to require improvement from the dialogues performed in the dialogue phase (S201). The improvement-needing part extraction unit 132 presents a list of the extracted improvement-needing parts (dialogue list) to the manager M (S202). For example, the improvement-needing part extraction unit 132 transmits the dialogue list to the manager terminal device 300 to display it. The manager M checks the contents of the dialogue list presented in S202, and inputs knowledge information required to improve the listed dialogues to the dialogue device 100 (S203). Note that the improvement-needing part extraction unit 132 may determine knowledge information (insufficient knowledge information) that is lacking for providing an appropriate response to the dialogue of the improvement-needing part, and include the insufficient knowledge information in the dialogue list.

[0049] Next, in the interactive device 100, the update instruction generating unit 133 generates update instruction information INS (including the update patch 122 and the alternative file 123) for reflecting the knowledge information input in S203 in the constraint information (S204A). The update instruction information INS may reflect the knowledge information input by the administrator M in the policy 41, may reflect the knowledge information in the knowledge 42, or may reflect the knowledge information in both the policy 41 and the knowledge 42. The type of information to which the constraint information is updated is appropriately determined by the judgment of the update instruction generating unit 133 according to the knowledge information input by the administrator M.

[0050] The update instruction generating unit 133 may be configured to refer to the current constraint information as necessary when generating the update instruction information INS (S204B). For example, the update instruction generating unit 133 may generate information to be added to the current constraint information as the alternative file 123, taking into consideration the current constraint information and new knowledge information input by the administrator M, or may generate the current constraint information plus the new knowledge information as the alternative file 123.

[0051] Next, the update execution unit 134 updates the constraint information stored in the external data source 40 based on the update instruction information INS generated in S204A. For example, in this embodiment, the update execution unit 134 executes the update patch 122 (S205) to cause the content of the alternative file 123 to be reflected in the constraint information (S206).

[0052] Once the constraint information has been updated, the dialogue unit 131 becomes able to dialogue with the user U based on the updated constraint information. In the dialogue after the constraint information has been updated, the knowledge 42 and the policy 41 for improving the areas requiring improvement identified in the immediately preceding improvement phase are accumulated as constraint information, so that in the current dialogue phase, the dialogue unit 131 is expected to conduct a more favorable dialogue on the topic of the areas requiring improvement than in the dialogue phase before improvement. By repeating such dialogue phases and improvement phases, the dialogue system 1 is updated so that it can have a smoother conversation with the user U.

[0053] [Second Example] FIG. 4 is a diagram (part 2) showing an example of a process flow in which the dialogue device 100 of the embodiment updates constraint information of a dialogue model (LLM) through dialogue with a user. The flow in FIG. 4 differs from that in FIG. 3 in that, instead of the configuration in which the manager M inputs knowledge information on the parts requiring improvement to the update instruction generating unit 133, the dialogue device 100 inputs the knowledge information. That is, in this case, the control unit 130 of the dialogue device 100 includes a knowledge information input unit 135 in addition to the configuration in FIG. 2. The knowledge information input unit 135 may be configured by an LLM. In FIG. 4, the same process flow as in FIG. 3 is assigned the same reference numerals as in FIG. 3, and thus the description of the same process flow as in FIG. 3 will be omitted.

[0054] The knowledge information input unit 135 inputs a list of improvement-requiring parts (dialogue list) extracted by the improvement-requiring part extraction unit 132 in S202, and acquires knowledge information useful for improving the improvement-requiring parts from the second external data source 50 based on the dialogue list (S203A). The knowledge information input unit 135 inputs the knowledge information acquired in S203A to the update instruction generation unit 133 (S203B). Here, the second external data source 50 may be any data source that is accessible from the knowledge information input unit 135 and has information useful for improving the improvement-requiring parts. For example, the second external data source 50 is the Internet. In this case, the knowledge information input unit 135 may acquire knowledge information by searching for information useful for improving the improvement-requiring parts, for example, on a search site on the Internet.

[0055] For example, the knowledge information input unit 135 can recognize knowledge (or rules) necessary for improving the dialogue of the failure case by inputting the dialogue log 121 of the failure case to the LLM and inputting an instruction such as "Please extract the lack of knowledge (or rules) that caused the dialogue to fail" to the LLM. Also, for example, the knowledge information input unit 135 may be configured to extract frequently occurring words from the dialogue log 121 of the failure case and recognize the extracted frequently occurring words as knowledge (or rules) necessary for improvement. In this case, the extraction of the frequently occurring words may be realized by rule-based processing or by instructing the LLM to extract the frequently occurring words.

[0056] In this way, in the example of FIG. 4, since the administrator does not need to input knowledge information, the costs associated with system operation can be reduced.

[0057] [Third Example] Fig. 5 is a diagram (part 3) showing an example of a process flow in which the dialogue device 100 of the embodiment updates constraint information of a dialogue model (LLM) through dialogue with a user. The flow in Fig. 5 differs from that in Fig. 3 in that, instead of the configuration in which the manager M directly inputs knowledge information to the update instruction generation unit 221, the manager M inputs knowledge information to the update instruction generation unit 133 via the second dialogue unit 136. That is, in this case, the control unit 130 of the dialogue device 100 includes the second dialogue unit 136 in addition to the configuration in Fig. 2. For example, the second dialogue unit 136 can be configured by an LLM.

[0058] In this case, for example, the second dialogue unit 136 inputs a dialogue list and acquires knowledge information useful for improving the improvement-requiring parts based on this dialogue list through dialogue with the manager M. For example, the second dialogue unit 136 presents the improvement-requiring parts listed in the dialogue list to the manager M (S203C) and requests the manager M to input knowledge information useful for improving the presented improvement-requiring parts. In response to this request, the manager M responds (inputs) knowledge information useful for improvement of the presented improvement-requiring parts to the second dialogue unit 136. Note that the knowledge information input by the manager M to the second dialogue unit 136 may be knowledge that the manager M himself originally knew, or may be knowledge that he has learned as a result of conducting an investigation in response to the request of S203C.

[0059] Also, for example, the second dialogue unit 136 may be configured to determine knowledge information (insufficient knowledge information) that is lacking for an appropriate response to a dialogue on an improvement-requiring part, and to request the manager M to input the insufficient knowledge information. At this time, the second dialogue unit 136 may be configured to determine the insufficient knowledge information based on the constraint information and the contents of the dialogue list stored in the external data source 40. Also, the interface between the second dialogue unit 136 and the manager M may be in the form of a text-based chat, or may be configured as a voice-based interface using a microphone, intercom, or the like.

[0060] According to the second dialogue unit 136 configured in this manner, for example, the manager M can input knowledge information useful for improving areas requiring improvement to the second dialogue unit 136 by carrying out the following exchanges (dialogues) such as those shown in (1) to (3). (1) Dialogue 2 136: “I don’t know if there is a diaper changing station in the bathroom on the second floor.” (2) Manager M: "There aren't any (diaper changing tables) in the (restrooms) on the second floor. But there are (diaper changing tables) in the restrooms on the first floor." (3) Dialogue 2 136: “Thank you. I’ll update my knowledge with that information.”

[0061] [Fourth Example] Fig. 6 is a diagram (part 4) showing an example of a process flow in which the interactive device 100 of the embodiment updates constraint information of the interactive model (LLM) through an interaction with a user. In Fig. 3, when the update instruction generating unit 133 generates update instruction information INS, the update executing unit 134 updates the constraint information based on the generated update instruction information INS, whereas the flow in Fig. 6 differs from that in Fig. 3 in that the constraint information is updated when the administrator M checks the generated update instruction information INS and then decides to execute the update.

[0062] In this case, even if the update instruction information INS is generated in S204A, the update execution unit 134 does not immediately update the constraint information, but waits for an update execution instruction from the update instruction generation unit 133. Meanwhile, the administrator M acquires the update instruction information INS generated in S204A from the interactive device 100, and checks the validity of its contents (S204C). Here, if the administrator M determines that there is something to be corrected in the update instruction information INS, the administrator M inputs the correction instruction to the update instruction generation unit 133 (S204D). For example, the correction instruction includes feedback information for allowing the update instruction generation unit 133 to recognize the points to be corrected and the viewpoint from which the correction should be made. When the correction instruction is input in S204D, the update instruction generation unit 133 recognizes the contents to be corrected based on the feedback information, and reflects them in the update instruction information INS (S204E). S204C to S204E are repeated until the administrator M can check the validity of the update instruction information INS. In this way, by repeatedly inputting feedback information and correcting the update instruction information INS, it is possible to increase the possibility of obtaining a desired output.

[0063] The update instruction information INS may be modified by editing the update instruction information INS that has already been generated. For example, the update instruction generation unit 133 may modify the update instruction information INS by changing a parameter of a process for outputting the update instruction information INS based on feedback information and re-executing output of the update instruction information INS. The update instruction generation unit 133 may determine whether or not to re-output the update instruction information INS based on the update instruction information INS that has been the subject of validity confirmation and the content of the feedback information. The update instruction generation unit 133 may also perform additional learning based on such an output (update instruction information INS) and the feedback information. This can improve the accuracy of the inference of the update instruction generation unit 133.

[0064] In general, the method of additional learning of a trained model through human feedback is called Reinforcement Learning from Human Feedback (RLHF). In RLHF, humans give OK (positive feedback) or NG (negative feedback) to the output of a trained model, and the trained model is updated based on that feedback information. This makes it possible to adjust the output of the trained model so that it is more in line with human intentions and expectations.

[0065] In contrast to this, the fourth embodiment accumulates feedback information as the same session information (e.g., prompt history, session memory, metadata, etc.), and sequentially optimizes the output based on the accumulated feedback information. That is, in the fourth embodiment, it is not necessary to re-learn the LLM based on the feedback information (e.g., update the weight parameters). Therefore, according to the fourth embodiment, it is possible to efficiently generate update instruction information INS with higher accuracy.

[0066] On the other hand, if the administrator M confirms the validity of the update instruction information INS in S204C, the administrator M instructs the update instruction generation unit 133 to update the constraint information (S204F). Upon receiving the update execution instruction of S204F, the update instruction generation unit 133 instructs the update execution unit 134 to update the constraint information (S204G). Then, upon receiving the update execution instruction of S204G, the update execution unit 134 updates the constraint information stored in the external data source 40 based on the generated update instruction information INS (S205).

[0067] If there is nothing to be corrected in the update instruction information INS generated in S204A, the manager M may input the update execution instruction in S204F without inputting the correction instruction in S204D. In this case, S204D and S204E are omitted, and the update execution unit 134 updates the constraint information based on the update instruction information INS generated in S204A.

[0068] According to the dialogue system 1 of the embodiment described above, it becomes possible to more efficiently correct the dialogue system using natural language.

[0069] <Modification> In the above embodiment, the dialogue unit 131, the improvement-required part extraction unit 132, the update instruction generation unit 133, the update execution unit 134, the knowledge information input unit 135, and the second dialogue unit 136 are provided in the dialogue device 100. However, among these functional units, some or all of the functional units other than the dialogue unit 131 may be configured separately from the dialogue device 100. For example, the dialogue device 100 may include the dialogue unit 131, and be communicatively connected to a separate information processing device including the improvement-required part extraction unit 132, the update instruction generation unit 133, the update execution unit 134, the knowledge information input unit 135, and the second dialogue unit 136.

[0070] In the above embodiment, the dialogue unit 131 of the dialogue device 100 mainly provides a dialogue interface in natural language, but the dialogue unit 131 may be configured to provide a so-called multimodal dialogue interface capable of recognizing information other than natural language. For example, the dialogue unit 131 may be configured to input text, images, voice, etc. directly or via a file, and output voice, images, etc. indicating a response to the input.

[0071] In the above embodiment, the case where the output of the dialogue unit 131 is controlled by the constraint information has been described, but such control of output by the constraint information may be applied to functional units other than the dialogue unit 131. For example, the output of the improvement-required part extraction unit 132, the update instruction generation unit 133, the knowledge information input unit 135, or the second dialogue unit 136, which may be configured using an LLM like the dialogue unit 131, may be controlled by applying dedicated constraint information to each of them. In this way, by applying dedicated constraint information to control the output (for example, by specifying an output format, NG rules, file extensions, etc.), it is possible to suppress the input and output of unnecessary information and to make it easier to obtain a desired output, thereby improving the performance of the improvement phase.

[0072] [Update instruction generation unit] (1) Preprocessing of knowledge information The update instruction generating unit 133 may be configured to instruct the LLM to execute preprocessing in order to improve the accuracy of the update instruction information INS. For example, the update instruction generating unit 133 may instruct the LLM to format the data input as knowledge information. Also, for example, when knowledge information in text format is input, the update instruction generating unit 133 may instruct the LLM to extract key information from the text data. For example, the update instruction generating unit 133 may input an instruction to the LLM, "Please organize and extract the key and value for the knowledge to be added" for the input text data. As a result, when knowledge information, for example, "The toilet is at the south end of the second floor of the main building," is input, the knowledge information can be converted into a form such as "Info_Type=Location, Name="Toilet", value="South end of the second floor of the main building". By such preprocessing, the knowledge information is converted in advance into a form that is easy for the LLM to handle, thereby improving the accuracy of inference by the LLM.

[0073] (2) Feedback information input support function In the fourth embodiment, a configuration has been described in which the administrator M inputs feedback information for the update instruction information INS generated by the update instruction generating unit 133, but it is also assumed that the administrator M may need to have a certain level of expertise to determine the validity of the update instruction information INS. In order to enable the administrator M to more easily determine the validity of the update instruction information INS, the dialogue device 100 may be configured to present to the administrator M an example of how the content of the dialogue changes as a result of feedback information input by the administrator M.

[0074] In this case, for example, the update instruction generating unit 133 may be configured to store both the constraint information before the update and the constraint information after the update by the update instruction information INS corrected based on the input feedback information, and present to the administrator M the outputs when the same dialogue is executed with both of them in a comparable manner. For example, the same dialogue may be a failure pattern extracted from the dialogue log 121. According to the update instruction generating unit 133 configured in this manner, the contents of the dialogue based on each of the constraint information before and after the update are presented to the administrator M, so that the administrator M can check the validity of the feedback information from the change in the contents of the dialogue before and after the update even if he does not have expertise in the update instruction information INS.

[0075] [Improvement required area extraction section] (1) Classification of reasons for dialogue failure The improvement-required part extraction unit 132 may be configured to have a function of classifying the cause of the failure case in addition to the above functions. For example, the improvement-required part extraction unit 132 may be configured to input the dialogue log 121 of the failure case to the LLM and instruct the LLM to classify the cause of the failure of the dialogue of the failure case. Also, for example, the improvement-required part extraction unit 132 may be configured to obtain a classification result (cause type) of the cause of the failure of the dialogue of the failure case by inputting the dialogue log 121 of the failure case to a learned model trained by machine learning to input the dialogue log 121 of the failure case and output a classification result of the cause of the failure of the dialogue of the failure case.

[0076] According to the improvement-required part extraction unit 132 configured in this manner, the type of constraint information to be corrected when improving the dialogue of the failure case can be recognized, so that the constraint information can be corrected more accurately, and the improvement accuracy of the improvement-required part can be improved. For example, when the cause type is an error in the answering policy, the improvement accuracy can be improved by notifying the update instruction generation unit 133 that the correction target is the policy 41. Also, for example, when the cause type is a lack of knowledge or an error in inference, the improvement accuracy can be improved by notifying the update instruction generation unit 133 that the correction target is the knowledge 42. Also, for example, when the cause type is a communication error or an error in voice recognition, the update instruction generation unit 133 can be notified that the correction of the constraint information is unnecessary, so that the accuracy of the dialogue of the dialogue unit 131 can be prevented from being reduced by unnecessary correction.

[0077] (2) Output of evaluation report In addition to the above functions, the improvement-requiring part extraction unit 132 may be configured to have a function of generating an evaluation report on the performance of the dialogue unit 131 based on information on the extracted failure cases. For example, the evaluation report may display a list of the number of extracted failure cases, the ratio of the extracted failure cases to the total dialogue cases, specific examples of the failure cases, cause types, etc. The improvement-requiring part extraction unit 132 may be configured to format the evaluation report using a rule base or LLM, or may be configured to summarize information to be included in the evaluation report using LLM. The improvement-requiring part extraction unit 132 configured in this way makes it possible to measure the performance of the dialogue system 1 and the dialogue model (LLM) and evaluate the validity of customer satisfaction based on the measurement results.

[0078] [Dialogue section] The dialogue unit 131 may be configured to output visual information that supplements the dialogue information, in addition to inputting and outputting language-based dialogue information (such as voice and chat). In this case, the dialogue device 100 may be provided with a display device for outputting the visual information. For example, the visual information may be displayed in the form of a map. For example, when the user U inputs a question such as "Where is the toilet?" to the dialogue unit 131, the dialogue unit 131 may display on the display unit a map showing a route from the current location to the nearest toilet, together with a language-based answer such as "Look here." In this way, the dialogue unit 131 provides visual information together with the dialogue information, so that the dialogue device 100 can improve the accuracy of information transmission to the user U.

[0079] In the above embodiment, the user U inputs voice into the user terminal device 200, and the dialogue device 100 acquires the voice data from the user terminal device 200. However, the dialogue unit 131 may have a microphone for voice input, and in that case, the user terminal device 200 may be omitted. For example, the dialogue unit 131 may be configured as a robot having a physicality such as a human type or an animal type. In this case, the robot may have an input / output device such as a microphone, a speaker, and a touch panel in addition to a communication interface for communicating with the main body side of the dialogue device 100. In addition to such input / output devices, the robot may be configured to transmit information by bodily movements. For example, when the robot is asked about a target location in a dialogue with the user U, the robot may be configured to move its arms to point in the direction of the target location. Such movements can also be considered a type of visual information in the sense that they are recognized by the user U's vision.

[0080] In addition, it is assumed that the aspect of the dialogue information changes due to the supplementation of visual information, and it becomes difficult to judge whether the dialogue has been successful or not only based on the dialogue information log. For example, in the above example, when the answer "Please look here" is given, it is difficult to judge whether the dialogue has been successful or not based on text, since it is not possible to know what the visual information at that time was. Therefore, when supplementing the dialogue information with visual information, the dialogue unit 131 may be configured to record in the dialogue log 121 the contents of the dialogue that was carried out and the visual information displayed when the dialogue was carried out in association with each other. In this case, the dialogue unit 131 may record the image of the visual information itself in the dialogue log 121, may record text information summarizing the contents of the visual information in the dialogue log 121, or may convert the visual information into a feature and record it in the dialogue log 121. This enables the improvement-required part extraction unit 132 in the subsequent stage to extract improvement-required parts based on both the dialogue information and the visual information.

[0081] [Extended conversation logs] (1) Expansion using image data Even if it is confirmed from the contents of the dialogue log 121 that the user U has said "Thank you" to the dialogue device 100 on a text basis, it is assumed that the user U is not necessarily satisfied with the dialogue with the dialogue device 100. In such a case, the dialogue system 1 may be configured to include a camera for capturing an image of the user U's facial expression during the dialogue, and to input image data captured by the camera to the improvement-required part extraction unit 132 together with the dialogue log 121. In this case, the improvement-required part extraction unit 132 may be configured to read the facial expression of the user U from the input image, and to extract a dialogue failure case based on the read facial expression and the contents of the dialogue log 121. The method of reading a person's facial expression from an image may be realized using any existing technology. In this case, the improvement-required part extraction unit 132 may be configured to execute a process of detecting the facial expression of the user U from image data, and to input the detection result to the LLM instead of the image data itself. In addition, if the LLM supports multimodality, the improvement-required part extraction unit 132 may input the image data itself to the LLM.

[0082] (2) Expansion with voice data The dialogue system 1 may also be configured to include a microphone for inputting a voice uttered by the user U, and to input data of the voice recorded by the microphone to the improvement-required part extraction unit 132 together with the dialogue log 121. In this case, the improvement-required part extraction unit 132 may be configured to read the satisfaction level of the user U with the dialogue device 100 from the input voice, and to extract a dialogue failure case based on the read satisfaction level and the content of the dialogue log 121. For example, the improvement-required part extraction unit 132 may extract, from the voice data, feature amounts related to features that may be influenced by the thoughts and emotions of the user U, such as tone of voice, speaking style, intonation, and speed, and compare or classify the extracted feature amounts or index values ​​acquired based on them with thresholds, thereby estimating the user U's evaluation (satisfaction, dissatisfaction, etc.) of the dialogue with the dialogue device 100. In this case, the improvement-required part extraction unit 132 may be configured to input the estimated evaluation result to the LLM instead of the voice data itself. In addition, if the LLM supports multimodality, the improvement-required part extraction unit 132 may input the voice data itself to the LLM.

[0083] In addition, when the user terminal device 200 is owned by each individual user, the above-mentioned camera and microphone may be mounted on the user terminal device 200. In addition, when the user terminal device 200 is shared by an unspecified number of users (for example, a guide terminal), a camera that captures images of the surroundings of the user terminal device 200 and a microphone that inputs voices around the user terminal device 200 may be separately provided. With this configuration, the text-based dialogue log 121 can be expanded by non-verbal information such as the voice and images of the user U during the dialogue, so that the improvement-requiring part extraction unit 132 can more appropriately determine whether or not the extraction of improvement-requiring parts is a failure case.

[0084] (3) Augmentation using sound environment information In a situation where the sound environment of the dialogue space is poor, such as when there is a lot of noise, the user has difficulty hearing the voice, and the dialogue system 1 also has a low accuracy in recognizing the user's voice, which may cause the dialogue to fail. Therefore, the dialogue device 100 may be configured to input time-series information (sound environment information) indicating the state of the sound environment at the time of the dialogue in the dialogue space, together with the dialogue log 121, to the improvement-required part extraction unit 132. In this case, the dialogue system 1 may include a microphone for sensing the sound environment of the dialogue space (hereinafter referred to as an "environmental microphone") in addition to the microphone used for dialogue with the user. In this case, the dialogue system 1 generates and stores sound environment information based on the voice data input from the environmental microphone. In addition, in this case, the improvement-required part extraction unit 132 may be configured to recognize the state of the sound environment at the time of the dialogue based on the sound environment information, and extract a dialogue failure case based on the recognition result and the contents of the dialogue log 121. According to such a configuration, the text-based dialogue log 121 can be expanded by the sound environment information of the dialogue space, so that the improvement-needing part extraction unit 132 can more appropriately determine whether or not a case is a failure when extracting improvement-needing parts.

[0085] (4) Expansion by operation control information As described above, the dialogue unit 131 may be configured in the form of a robot capable of performing physical actions. In this case, visual information presented by the robot through physical actions can be used as information for expanding the dialogue log 121, together with visual information that can be presented by the above-mentioned image display. For example, when the robot performs an action of pointing to a place asked by the user U by the direction of its arm, the dialogue device 100 stores control information of such an action together with the dialogue log 121. In this case, the improvement-required part extraction unit 132 may be configured to recognize the action of the robot during the dialogue based on the control information, and to extract a dialogue failure case based on the recognition result and the contents of the dialogue log 121. According to such a configuration, the text-based dialogue log 121 can be expanded by information on the action (output) performed by the dialogue unit 131, so that the improvement-required part extraction unit 132 can more appropriately determine whether or not the improvement-required part is a failure case when extracting the improvement-required part.

[0086] Although an embodiment of the present invention has been described in detail above with reference to the drawings, the specific configuration is not limited to this embodiment, and designs that do not deviate from the gist of the present invention are also included. [Industrial Applicability]

[0087] The present invention is applicable to a system that provides a dialogue interface with a person based on constraint information. [Explanation of symbols]

[0088] 1 Dialogue System 40 External Data Sources 41 Policy 42 Knowledge 50 Secondary external data source 100 Interactive device 110 Communications Department 120 Storage section 121 Dialogue Log 122 Update Patch 123 Alternate File 130 Control section 131 Dialogue Section 132 Extraction of areas requiring improvement 133 Update instruction generation unit 134 Update Execution Department 135 Knowledge Information Input Unit 136 Dialogue Part 2 200 User terminal device 221 Update instruction generation unit 300 Administrator terminal device 300 User terminal device 400 External Data Source

Claims

1. An information processing device for updating constraint information that affects an answer generated in a dialogue between a dialogue device providing a dialogue interface that generates an answer based on information provided by a user and the user, the information processing device comprising: an extraction unit that extracts improvement-required points based on the dialogue record; a generating unit that receives input of knowledge information useful for improving the improvement-requiring portion, and generates update instruction information for updating the constraint information based on the knowledge information; an update unit that updates the constraint information based on the update instruction information; An information processing device comprising:

2. The extraction unit presents the extracted improvement-required points to an administrator of the dialogue device, The generation unit receives an input of the knowledge information from the administrator. The information processing device according to claim 1 .

3. An input unit for inputting the knowledge information to the generation unit, The extraction unit supplies the improvement-required portion to the input unit, The input unit extracts knowledge information useful for improvement from an external data source for the improvement-requiring portion supplied from the extraction unit, and inputs the knowledge information to the generation unit. The information processing device according to claim 1 .

4. The input unit is configured using a dialogue model trained to respond to the input of the user. The information processing device according to claim 3 .

5. A dialogue unit that provides an interface for dialogue with an administrator of the dialogue device, The dialogue unit includes: Identifying information necessary as the knowledge information based on the improvement-required points and presenting it to the manager; accepting an input of the knowledge information from the administrator, and inputting the knowledge information input by the administrator to the generation unit; The information processing device according to claim 1 .

6. The dialogue unit is configured using a dialogue model trained to respond to input from the user. The information processing device according to claim 5 .

7. When feedback information regarding a correction to the generated update instruction information is input from the administrator, the generation unit corrects the update instruction information based on the feedback information. The information processing device according to claim 5 .

8. the generation unit stores both first constraint information, which is constraint information before the update, and second constraint information, which is constraint information after the update by update instruction information corrected based on the feedback information, and the dialogue unit presents to the administrator outputs obtained when the same dialogue is executed using each of the first constraint information and the second constraint information in a comparable manner. The information processing device according to claim 7.

9. The extraction unit and the generation unit are configured using a dialogue model trained to respond to the user's input. The information processing device according to claim 1 .

10. the generation unit performs preprocessing on the input knowledge information to shape a format or to extract information that is key for updating the constraint information, and inputs the preprocessed knowledge information to the dialogue model. The information processing device according to claim 9.

11. the extraction unit classifies, for the dialogue extracted as the improvement-requiring area, causes of the dialogue extracted as the improvement-requiring area into categories corresponding to a correction policy of the constraint information, and presents the categories to the manager. The information processing device according to claim 2 .

12. The dialogue device communicates with the user by language-based dialogue information exchanged between the user and the dialogue device and visual information supplementing the dialogue information; The record of the interaction includes historical information regarding the interaction information and the visual information. The information processing device according to claim 1 .

13. the dialogue interface is provided as a robot capable of physical actions; The history information regarding the visual information includes history information regarding the physical motion performed by the robot. The information processing device according to claim 12.

14. the dialogue device includes a camera for reading the facial expression of the user; The dialogue record includes language-based dialogue information exchanged between the user and the user, and historical information regarding images captured by the camera or facial expressions of the user recognized from the images. The information processing device according to claim 1 .

15. the dialogue device includes a microphone for sensing a state of a sound environment in a space in which the dialogue device and the user are to dialogue; The dialogue record includes language-based dialogue information exchanged between the user and the user, and history information regarding the voice captured by the microphone or the state of the sound environment recognized from the voice. The information processing device according to claim 1 .

16. The dialogue record includes language-based dialogue information exchanged between the user and the dialogue device, and history information regarding the user's voice or the user's evaluation of the dialogue device recognized from the voice. The information processing device according to claim 1 .

17. The dialogue record includes historical information regarding communication errors of the dialogue device. The information processing device according to claim 1 .

18. The computer 1. An information processing method for updating constraint information that affects an answer generated in a dialogue between a dialogue device providing a dialogue interface that generates an answer based on information provided by a user and the user, comprising: The computer, Based on the records of the dialogue, areas that need improvement are extracted, receiving input of knowledge information useful for improving the improvement-requiring portion, and generating update instruction information for updating the constraint information based on the knowledge information; updating the constraint information based on the update instruction information; Information processing methods.

19. On the computer, A program for causing a dialogue device providing a dialogue interface for generating a response based on information provided by a user to execute an update process for updating constraint information that affects a response generated in a dialogue with the user, the program comprising: The update process includes: A process of extracting points requiring improvement based on the records of the dialogue; A process of receiving an input of knowledge information useful for improving the improvement-requiring portion, and generating update instruction information for updating the constraint information based on the knowledge information; updating the constraint information based on the update instruction information; Including, program.

Images