Nonsynchronous dialogue system and program

Abstract

To provide a nonsynchronous dialogue system and a program capable of appropriately answering with the details of dialogues with a plurality of users being taken into consideration.SOLUTION: A nonsynchronous dialogue system according to the present invention includes: first extracting means for referring to a second database that stores second question information and second answer information in association with each other, and for extracting, based on first question information, the second question information and the second answer information; second extracting means for referring to a first database that stores the first question information and first answer information in association with each other, and for extracting, based on the second question information, the first question information and the first answer information; first outputting means for inputting the first question information to a first base model, and for outputting, based on the second question information and the second answer information, the first answer information; and second outputting means for inputting the second question information to a second base model, and for outputting, based on the first question information and the first answer information, the second answer information.SELECTED DRAWING: Figure 3

Description

Technical Field

[0001] The present invention relates to an asynchronous dialogue system and a program.

Background Art

[0002] In recent years, for example, at construction sites such as civil engineering work sites, the introduction of ICT (Information and Communication Technology) technology has been demanded. In order to effectively introduce this ICT technology, it is necessary for ICT engineers to collect information on the construction site. However, in many cases, ICT engineers are office workers who rarely actually visit the construction site. For this reason, it has been difficult to have close communication with the workers at the construction site, and information collection has been difficult. For this reason, an interactive management system such as that shown in Patent Document 1, for example, has attracted attention.

[0003] In Patent Document 1, in an interactive management system using an AI robot that manages by having workers entering and leaving a construction site interact with the AI robot, the AI robot extracts a voice feature amount from the voice data of the worker, a conversation feature amount from the conversation data with the worker, a physical feature amount from the image data of the worker, or an action feature amount from the video data of the worker through dialogue with the worker, and includes an individual feature amount detection unit that stores the feature amount as individual feature amount data in a database. An interactive management system using an AI robot is disclosed.

Prior Art Documents

Patent Documents

[0004]

Patent Document 1

Summary of the Invention

Problems to be Solved by the Invention

[0005] On the one hand, the interactive management system disclosed in Patent Document 1 assumes interaction with workers at the construction site, but does not assume interaction with in-house workers such as ICT engineers who are not at the construction site. For this reason, the interactive management system disclosed in Patent Document 1 has a problem that it cannot appropriately answer based on the content of interactions with multiple users.

[0006] Therefore, the present invention has been devised in view of the above-described problems, and an object thereof is to provide an asynchronous dialogue system and a program that can appropriately answer based on the content of interactions with multiple users.

Means for Solving the Problems

[0007] The asynchronous dialogue system according to the first invention includes a first acquisition means for acquiring first question information indicating a question of a first user, a second acquisition means for acquiring second question information indicating a question of a second user, referring to a second database that stores the second question information and second answer information indicating an answer to the second user's question in association with each other, and based on the first question information acquired by the first acquisition means, a first extraction means for extracting the second question information and the second answer information, referring to a first database that stores the first question information and first answer information indicating an answer to the first user's question in association with each other, and based on the second question information acquired by the second acquisition means, a second extraction means for extracting the first question information and the first answer information, inputting the first question information acquired by the first acquisition means into a first base model, and based on the second question information and the second answer information extracted by the first extraction means, a first output means for outputting first answer information, inputting the second question information acquired by the second acquisition means into a second base model, and based on the first question information and the first answer information extracted by the second extraction means, a second output means for outputting second answer information.

[0008] The asynchronous dialogue system according to the second invention, in the first invention, further comprises: a first storage means for associating the first question information acquired by the first acquisition means with the first answer information output by the first output means and storing the same in the first database; and a second storage means for associating the second question information acquired by the second acquisition means with the second answer information output by the second output means and storing the same in the second database.

[0009] The asynchronous dialogue system according to the third invention, in the second invention, the first storage means associates first evaluation information indicating an evaluation of the first answer information associated with the first question information with the first question information and the first answer information and stores the same in the first database, the second storage means associates second evaluation information indicating an evaluation of the second answer information associated with the second question information with the second question information and the second answer information and stores the same in the second database, and further comprises: a first learning means for learning the first base model based on the first question information, the first answer information, and the first evaluation information stored in the first database; and a second learning means for learning the second base model based on the second question information, the second answer information, and the second evaluation information stored in the second database.

[0010] In the asynchronous dialogue system according to the fourth invention, in the second invention, the first storage means acquires first business information related to the business of the first user, associates the acquired first business information, the first question information, and the first answer information, and stores them in the first database. The second storage means acquires second business information related to the business of the second user, associates the acquired second business information, the second question information, and the second answer information, and stores them in the second database. The first output means inputs the first question information acquired by the first acquisition means into a first base model, and outputs the first answer information based on the second business information associated with the second question information and the second answer information extracted by the first extraction means. The second output means inputs the second question information acquired by the second acquisition means into a second base model, and outputs the second answer information based on the first business information associated with the first question information and the first answer information extracted by the second extraction means.

[0011] In the asynchronous dialogue system according to the fifth invention, in the first invention, the first acquisition means acquires the first question information indicating a question of a construction worker as the first user, and the second acquisition means acquires the second question information indicating a question of an office worker as the second user.

[0012] In the asynchronous dialogue system according to the sixth invention, in the first invention, the first output means inputs the first question information acquired by the first acquisition means into a first base model which is a large language model (Large Language Model), and outputs the first answer information based on the second question information and the second answer information extracted by the first extraction means. The second output means inputs the second question information acquired by the second acquisition means into a second base model which is a large language model, and outputs the second answer information based on the first question information and the first answer information extracted by the second extraction means.

[0013] The asynchronous dialogue program according to the seventh invention includes a first acquisition step of acquiring first question information indicating a question of a first user, a second acquisition step of acquiring second question information indicating a question of a second user, referring to a second database that stores the second question information and second answer information indicating an answer to the second user's question in association with each other, and based on the first question information acquired in the first acquisition step, a first extraction step of extracting the second question information and the second answer information, referring to a first database that stores the first question information and first answer information indicating an answer to the first user's question in association with each other, and based on the second question information acquired in the second acquisition step, a second extraction step of extracting the first question information and the first answer information, inputting the first question information acquired in the first acquisition step into a first base model, and based on the second question information and the second answer information extracted in the first extraction step, a first output step of outputting first answer information, inputting the second question information acquired in the second acquisition step into a second base model, and based on the first question information and the first answer information extracted in the second extraction step, a second output step of outputting second answer information, and causing a computer to execute the above steps.

Effect of the Invention

[0014] According to the first to seventh inventions, the asynchronous dialogue system and program of the present invention input first question information into a first base model, and based on second question information and second answer information, output first answer information, input second question information into a second base model, and based on first question information and first answer information, output second answer information. Thereby, the content of the dialogue of the first user is reflected in the answer to the second user, and the content of the dialogue of the second user is reflected in the answer to the first user. Therefore, it is possible to output an answer based on the content of the dialogue with a plurality of users.

[0015] In particular, according to the second invention, the asynchronous dialogue system of the present invention associates and stores the first question information and the first answer information, and associates and stores the second question information and the second answer information. As a result, it becomes possible to use the output answer information as further learning data. Therefore, it becomes possible to answer questions with higher accuracy.

[0016] In particular, according to the third invention, the asynchronous dialogue system of the present invention learns the first base model based on the first question information, the first answer information, and the first evaluation information, and learns the second base model based on the second question information, the second answer information, and the second evaluation information. As a result, for example, by using the second evaluation information input from the first user and the first evaluation information input from the second user, it becomes possible to reflect the evaluation of the dialogue from a plurality of users. Therefore, it becomes possible to answer questions with higher accuracy.

[0017] In particular, according to the fourth invention, the asynchronous dialogue system of the present invention inputs the first question information into the first base model, and outputs the first answer information based on the second business information associated with the second question information and the second answer information. The second question information is input into the second base model, and the second answer information is output based on the first business information associated with the first question information and the first answer information. As a result, it becomes possible to reflect information such as information on the construction site and the content of the work of the office workers. Therefore, it becomes possible to answer questions more flexibly.

[0018] In particular, according to the fifth invention, the asynchronous dialogue system of the present invention acquires, as the first user, the first question information indicating the questions of the construction workers, and acquires, as the second user, the second question information indicating the questions of the office workers. As a result, it becomes possible for the construction site workers and the office workers to have an asynchronous dialogue. Therefore, due to the difference in work patterns, even if the on-site workers and the office workers, who have difficulty gathering and having a dialogue at the same time, do not have a dialogue, they can still have a dialogue with each other and share information.

[0019] In particular, according to the sixth invention, the asynchronous dialogue system of the present invention outputs response information using a Large Language Model. This enables the model to be trained from a vast amount of data. Therefore, it becomes possible to respond to questions more flexibly.

Brief Description of the Drawings

[0020]

Figure 1

Figure 2

Figure 3

Figure 4

Figure 5

Figure 6

Figure 7

Modes for Carrying Out the Invention

[0021] 〈First Embodiment〉 Hereinafter, an example of an asynchronous dialogue system according to the first embodiment to which the present invention is applied will be described with reference to the drawings.

[0022] FIG. 1 is a schematic diagram showing an example of the configuration of an asynchronous dialogue system 100 according to the first embodiment. As shown in FIG. 1, for example, the asynchronous dialogue system 100 includes an asynchronous dialogue device 1, a server 3, and a plurality of user terminals 2 connected via a public communication network 4. Further, the asynchronous dialogue system 100 may include, for example, a user terminal 2a and a user terminal 2b connected via the public communication network 4.

[0023] The server 3 is a storage medium that stores various data such as question information and answer information transmitted from the asynchronous dialogue device 1 and the user terminal 2. Further, the server 3 transmits the stored various data to the asynchronous dialogue device 1 and the user terminal 2 as necessary. The server 3 may include, for example, at least some of the functions provided in the asynchronous dialogue device 1, and may perform at least some of the processes instead of the asynchronous dialogue device 1.

[0024] The public communication network 4 is, for example, the Internet network to which the asynchronous dialogue device 1 is connected via a communication circuit. The public communication network 4 may be configured by a so-called optical fiber communication network. Further, the public communication network 4 may be realized by a known communication technology such as a wireless communication network in addition to the wired communication network.

[0025] The user terminal 2 is, for example, owned by a user of a service using the asynchronous dialogue system 100 and is connected to the asynchronous dialogue device 1 via the public communication network 4. The user terminal 2 may represent, for example, an electronic device that generates a database. As the user terminal 2, for example, an electronic device such as a personal computer or a tablet terminal is used. The user terminal may include, for example, at least some of the functions provided in the asynchronous dialogue device 1. The user terminal 2 may include a display (not shown) or a speaker that can present information to the user. Further, the user terminals 2 are terminals owned by different users. Also, the user terminals 2 may be terminals provided at different locations. The user terminal 2a is, for example, owned by a first user such as a worker at a construction site, and the user terminal 2b is, for example, owned by a second user such as an in-office worker of an ICT engineer.

[0026] The asynchronous dialogue device 1 outputs first response information indicating an answer to the first user's question based on first question information indicating the first user's question, and outputs second response information indicating an answer to the second user's question based on second question information indicating the second user's question. The asynchronous dialogue device 1 is used as an electronic device such as a personal computer (PC), and may also be used as an electronic device such as a smartphone, a tablet terminal, a wearable terminal, an IoT (Internet of Things) device, or a single-board computer such as a Raspberry Pi (registered trademark).

[0027] Next, with reference to FIG. 2, an example of the asynchronous dialogue device 1 in the first embodiment will be described. FIG. 2 is a schematic diagram showing an example of the configuration of the asynchronous dialogue device 1 in the first embodiment, and FIG. 3 is a schematic diagram showing an example of the functions of the asynchronous dialogue device 1 in the first embodiment.

[0028] As shown in FIG. 2, for example, the asynchronous dialogue device 1 includes a housing 10, a CPU (Central Processing Unit) 101, a ROM (Read Only Memory) 102, a RAM (Random Access Memory) 103, a storage unit 104, and I / Fs 105 to 107. The CPU 101, the ROM 102, the RAM 103, the storage unit 104, and the I / Fs 105 to 107 are connected by an internal bus 110.

[0029] The CPU 101 controls the entire asynchronous dialogue device 1. The ROM 102 stores the operation code of the CPU 101. The RAM 103 is a work area used during the operation of the CPU 101. The storage unit 104 stores various information such as question information, answer information, and a base model. As the storage unit 104, for example, in addition to an HDD (Hard Disk Drive), a data storage device such as an SSD (Solid State Drive), an SD card, or a miniSD card is used. Note that, for example, the asynchronous dialogue device 1 may have a GPU (Graphics Processing Unit) not shown in the figure.

[0030] I / F105 is an interface for transmitting and receiving various types of information via the public communication network 4. I / F106 is an interface for transmitting and receiving information with the input unit 108. As the input unit 108, for example, a keyboard is used, and a user who uses the asynchronous dialogue device 1 inputs various types of information or control commands for the asynchronous dialogue device 1 via the input unit 108. I / F107 is an interface for transmitting and receiving various types of information with the display unit 109. The display unit 109 outputs various types of information such as the deterioration information stored in the storage unit 104 or the processing status of the asynchronous dialogue device 1. As the display unit 109, a display is used, and for example, it may be a touch panel type.

[0031] As shown in FIG. 3, the asynchronous dialogue device 1 includes a first processing unit 11, a second processing unit 12, a first storage unit 13, and a second storage unit 14 that are respectively connected. Note that the first processing unit 11, the second processing unit 12, the first storage unit 13, and the second storage unit 14 are realized, for example, when a CPU 101 executes a program stored in the storage unit 104 or the like using the RAM 103 as a work area, and may be controlled by, for example, artificial intelligence.

[0032] The first processing unit 11 outputs first response information based on the first question information. The first storage unit 13 stores a first database in which the first question information and the first response information are associated. The second processing unit 12 outputs second response information based on the second question information. The second storage unit 14 stores a second database in which the second question information and the second response information are associated.

[0033] In the first database, first business information related to the business of the first user may be stored in association with the first question information and the first response information. Also, in the first database, first evaluation information indicating the evaluation of the first response information associated with the first question information may be stored in association with the first question information and the first response information. Further, in the first database, external information such as information on past construction sites may be stored in association with the first question information and the first response information.

[0034] The second database may store the second business information related to the business of the second user in association with the second question information and the second answer information. Further, the second database may store the second evaluation information indicating the evaluation of the second answer information associated with the second question information in association with the second question information and the second answer information. Further, the second database may store external information such as information on past construction sites in association with the second question information and the second answer information.

[0035] Next, the operation of the asynchronous dialogue system 100 to which the first embodiment of the present invention is applied will be described. FIG. 4 is a flowchart showing the operation of the asynchronous dialogue system 100 to which the first embodiment is applied. FIG. 4(a) is a flowchart showing the operation of outputting the first answer information of the asynchronous dialogue system 100 to which the first embodiment is applied. FIG. 4(b) is a flowchart showing the operation of outputting the second answer information of the asynchronous dialogue system 100 to which the first embodiment is applied.

[0036] First, in step S11, the first processing unit 11 acquires the first question information. The first processing unit 11 acquires, for example, the first question information input from the first user via the user terminal 2a. Further, in step S11, the first processing unit 11 may further acquire various types of information such as the first business information or external information. The business information is information related to the user's business, and may be, for example, information on the construction site in charge of the user, information on the user's affiliation or attributes, etc. The first business information is business information related to the business of the first user.

[0037] Question information is information indicating a question or an instruction, and may be information in text or voice format. Further, the question information may be a prompt. Further, the question information may be information regarding a question about a construction site. Further, the question information may be information of an image or video input from a user. Further, the question information may be information obtained by morphological analysis of a question. The first question information is information indicating a question of a first user. Further, the question information may be information obtained by morphological analysis of a question. When the acquired question information is in voice or image format, the first processing unit 11 may convert the question information into text format using voice recognition or image recognition. When the acquired question information is in image format, the first processing unit 11 may calculate a feature amount from the image of the question information using R-CNN (Region Based Convolutional Neural Networks), YOLO (You Only Look Once), SSD (Single Shot MultiBox Detector), etc., and convert it into text format such as the name of an object, a place, etc. based on this feature amount. The first question information may be question information input by the first user. Further, the first question information may be question information input via the user terminal 2a held by the first user. The first user is, for example, a worker at an outdoor construction site, but is not limited thereto and may be any user.

[0038] Next, in step S12, the first processing unit 11 searches the second database based on the first question information obtained in step S11. The first processing unit 11 refers to the second database and extracts second question information and second answer information based on the first question information obtained in step S11. For example, the first processing unit 11 performs a search from the second database based on the first question information obtained in step S11, searches for second question information stored in the second database that is similar to the first question information obtained in step S11, and extracts the second answer information stored in association with the searched second question information. Further, for example, the first processing unit 11 performs a search from the second database based on the first question information obtained in step S11, searches for two or more pieces of second question information stored in the second database whose similarity to the first question information obtained in step S11 is higher than a reference value, and extracts two or more pieces of second answer information stored in association with the searched second question information. In such a case, the first processing unit 11 may also obtain similarity information indicating the similarity between the first question information obtained in step S11 and the second question information stored in the second database. Further, for example, the first processing unit 11 performs a search from the second database based on the first question information obtained in step S11, searches for the top K pieces of second question information having a high similarity to the first question information obtained in step S11, and extracts the K pieces of second answer information stored in association with the searched K pieces of second question information. The first processing unit 11 uses the second question information and the second answer information extracted by the search as search results for newly outputting the first answer information. Further, the first processing unit 11 extracts external information, second evaluation information, or second business information associated with the second question information stored in the second database, whose similarity is higher than the reference value, based on the first question information obtained in step S11, and may use the extracted various information as search results.

[0039] In step S12, the first processing unit 11 may perform a search based on the first question information acquired in step S11, using, for example, a maximum inner-product search (MIPS).

[0040] Furthermore, in step S12, the first processing unit 11 may perform a search by referring to a database stored in a server or the like that is communicable via a public communication network, not limited to the second database stored in the second storage unit 14. Furthermore, in step S12, if the first processing unit 11 has acquired external information or first business information in step S11, the first processing unit 11 may search for second question information and second answer information linked to the acquired information.

[0041] Next, in step S13, the first processing unit 11 outputs first answer information based on the first question information acquired in step S11. The answer information is information indicating an answer or response to a question included in the question information. The answer information may also include information indicating the basis of the answer. The information indicating the basis of the answer may be, for example, specifications such as the Highway Bridge Specifications and the Concrete Standard Specifications, guidelines for regular inspection of highway bridges, guidelines issued by the Ministry of Land, Road and Transport or the Road Bureau, bridge longevity repair plans of each prefecture, descriptions in papers, patent documents, etc. The answer information may also be information obtained by decomposing the answer into morphemes. The first answer information is an answer to a question included in the first question information.

[0042] In step S13, the first processing unit 11 inputs the first question information acquired in step S11 into the first base model, and outputs first answer information based on the second question information and second answer information extracted in step S12.

[0043] The base model is a trained model trained using, for example, question information and answer information as training data. The first base model is a base model for outputting first answer information.

[0044] As a method for generating a base model, for example, machine learning using a neural network as a model may be used to generate the base model. The base model may be a model using multi-modal. The base model is learned using, for example, machine learning with a neural network such as CNN (Convolution Neural Network) as a model, and any model may be used. Further, as a method for generating the base model, for example, Retrieval-Augmented Generation (RAG), seq2seq (Sequence To Sequence) linear discrimination, support vector machine, k-nearest neighbor method, random forest, deep learning, etc. may be used to generate the base model.

[0045] In such a case, the base model stores, for example, as shown in FIG. 5, the relevance having a degree of association between the question information as input data and the answer information as output data. Also, in such a case, the morphemes or words included in the question information may be used as input data, and the morphemes or words included in the answer information may be used as output data. The degree of association indicates the degree of connection between the input data and the output data. For example, it can be determined that the higher the degree of association, the stronger the connection between each data. The degree of association is indicated by, for example, three or more values or three or more levels such as a percentage, and may also be indicated by two values or two levels. Further, the question information and the answer information used for learning the degree of association are, for example, the question information and the answer information for use in the learning data acquired in advance, but are not limited thereto, and information acquired at any timing may be used.

[0046] For example, the relevance is constructed based on the degree of connection among a plurality of input data, pairs, and a plurality of output data. The relevance is appropriately updated during the machine learning process and represents, for example, a classifier using a function optimized based on the plurality of input data and the plurality of output data. Note that the relevance may have a plurality of relevance degrees indicating, for example, the degree of connection among each data. The relevance degree can correspond to a weight variable, for example, when the database is constructed by a neural network. The relevance may indicate, for example, the degree of connection between a plurality of input data and a plurality of output data as shown in FIG. 5. In this case, by using the relevance, the degree of relationship with a plurality of output data of "Answer Information A" to "Answer Information C" can be linked and stored for each input data of "Question Information A" to "Question Information C" in FIG. 5. For this reason, for example, a plurality of input data can be linked to one output data via the relevance. Thereby, it is possible to realize a multi-faceted selection of output data for the input data. Further, the input data and the output data are not limited to this, and any type of information may be further used. Further, the input data may be, for example, text information and / or image information included in the question information.

[0047] The relevance has a plurality of relevance degrees that respectively link each input data and each output data. The relevance degree is indicated by three or more levels such as a percentage, a 10-level scale, or a 5-level scale, and is indicated by, for example, a line feature (such as thickness). For example, "Question Information A" included in the input data indicates a relevance degree AA of "73%" with "Answer Information A" included in the output data and a relevance degree AB of "12%" with "Answer Information B" included in the output data. That is, the "relevance degree" indicates the degree of connection between each data. For example, the higher the relevance degree, the stronger the connection between each data.

[0048] Such a relevance degree of three or more levels shown in FIG. 5 is acquired in advance. That is, when actually discriminating the solution, accumulate the past data sets of which of the input data and the output data were adopted and evaluated, and create the relevance degree shown in FIG. 5 by analyzing and analyzing these.

[0049] For example, in the past, it is assumed that for input data called "question information B", "answer information B" is determined to be the most suitable and evaluated. By collecting and analyzing such data sets, the correlation between the input data and the output data becomes stronger.

[0050] This analysis and interpretation may be performed by artificial intelligence. In such a case, for example, when there are many cases where "answer information B" is estimated for input data called "question information B", the correlation connecting this "question information B" and "answer information B" is set higher.

[0051] Also, this correlation may be composed of nodes of a neural network in artificial intelligence. That is, the weight coefficient for the output of this neural network node will correspond to the above-described correlation. Also, not limited to neural networks, it may be composed of any decision-making factor constituting artificial intelligence.

[0052] Also, at least one or more hidden layers may be provided between the input data and the output data in the base model, and machine learning may be performed. The above-described correlation is set in either one or both of the input data or the hidden layer data, and this becomes the weighting of each data, and the output is selected based on this. And when this correlation exceeds a certain threshold value, the output may be selected.

[0053] Such a degree of association serves as learning data in the context of artificial intelligence. By pre-learning such learning data, in actual step S13, the first processing unit 11 refers to the first base model learned using the question information and the answer information as learning data, and based on the first question information obtained in step S11, outputs the first answer information. When outputting, for example, the degree of association shown in FIG. 5 obtained in advance is referred to. For example, when the newly obtained first question information is the same as or similar to "Question Information A", through the degree of association, the degree of association AA "73%" with "Answer Information A" and the degree of association AB "12%" with "Answer Information B" are associated. In this case, "Question Information A" with the highest degree of association is selected as the optimal solution. However, it is not essential to select the one with the highest degree of association as the optimal solution, and it is also possible to select "Answer Information B" whose degree of association is low but whose relevance itself is recognized as the optimal solution. Needless to say, it is also possible to select an output solution that is not connected by an arrow, and as long as it is based on the degree of association, it may be selected in any other priority order.

[0054] By referring to such a degree of association, in addition to the case where the question information is the same as or similar to the input data, even in the case where it is dissimilar, the output data suitable for the input data can be quantitatively selected.

[0055] Also, the base model may be a natural language model. The natural language model may be a dialogue-type, so-called chat-type or conversation-type model that alternately accepts the instruction sentences included in the question information and generates the response sentences included in the answer information. The natural language model may be a large language model (LLM) learned from a large amount of text data, or a model obtained by transfer learning of the large language model.

[0056] A large language model is a deep learning model that pre-learns from vast amounts of data a language model that models human spoken language by its probability of occurrence. That is, a large language model is a natural language processing model trained using a large amount of text data, which takes a text as input information for a question and outputs a text as answer information. When a large language model is applied to a question-and-answer system, when the question text is input into the large language model as question information, an answer text is output from the LLM as answer information.

[0057] When the first processing unit 11 receives text data (prompt) as question information, it statistically estimates the generation probability of the next word from the text included in the received prompt using a large language model, and outputs answer information based on the estimation result. As the large language model, for example, known technologies described on Internet sites such as "https: / / chatgpt-lab.com / n / n418d3aa56f0b" and "https: / / agirobots.com / chatgpt-mechanism-and-problem / " can be adopted. Also, as the large language model, for example, GPT-4 provided by OpenAI of the United States may be used.

[0058] In step S13, the first processing unit 11 may, for example, reset the relevance between the question information and the answer information of the first base model based on the search results extracted in step S12. In such a case, it may be set so that the relevance of the learning data similar to the second question information and the second answer information included in the search results becomes high. Also, when dealing with a large language model as the first base model, the first processing unit 11 may set the generation probability of generating the next word from the text included in the received prompt based on the search results. In such a case, the first processing unit 11 may be set so that the generation probability of the first base model of the words included in the second question information, the second answer information, and information such as external information included in the search results becomes high.

[0059] Also, in step S13, the first processing unit 11 may input the first question information acquired in step S11 into the first base model, and output first answer information based on the second business information associated with the second question information and the second answer information extracted in step S12. In such a case, the first processing unit 11 may be set such that, for example, the generation probability of the first base model for the words included in the second question information, the second answer information, and the second business information is increased.

[0060] Next, in step S14, the first storage unit 13 associates the first question information acquired in step S11 with the first answer information output in step S13 and stores them in the first database. Also, in step S14, the first storage unit 13 may associate and store the first business information or external information acquired in step S11 with the first question information acquired in step S11 and the first answer information output in step S13 in the first database. In such a case, the first storage unit 13 may assign a tag corresponding to the first business information or external information to the association of the first question information and the first answer information.

[0061] Thereby, the operation of outputting the first answer information of the asynchronous dialogue system 100 in the first embodiment ends. Next, the operation of outputting the second answer information of the asynchronous dialogue system 100 in the first embodiment will be described.

[0062] First, in step S21, the second processing unit 12 acquires second question information. The second processing unit 12 acquires, for example, second question information input from the second user via the user terminal 2b. Also, in step S21, the second processing unit 12 may further acquire various types of information such as second business information or external information. The second business information is business information related to the business of the second user.

[0063] The second question information is question information indicating the question of the second user. The second question information may be question information input by the second user. Further, the second question information may be question information input via the user terminal 2b held by the second user. The second user is, for example, an in-house ICT technician or the like, but is not limited thereto and may be any user.

[0064] Next, in step S22, the second processing unit 12 searches the first database based on the second question information acquired in step S21. Similar to step S12, the second processing unit 12 refers to the first database and extracts the first question information and the first answer information based on the second question information acquired in step S21.

[0065] Next, in step S23, the second processing unit 12 outputs second answer information based on the second question information acquired in step S21. The second answer information is answer information for the question included in the second question information.

[0066] In step S23, similar to step S13, the second processing unit 12 inputs the second question information acquired in step S21 into the second base model, and outputs second answer information based on the first question information and the first answer information extracted in step S22. The second base model is a base model for outputting the second answer information.

[0067] Next, in step S24, the second storage unit 14 associates the second question information acquired in step S21 with the second answer information output in step S23 and stores it in the second database. Further, in step S24, the second storage unit 14 may store the second business information or external information acquired in step S21 in the second database in association with the second question information acquired in step S21 and the second answer information output in step S23. In such a case, the second storage unit 14 may assign a tag corresponding to the second business information or external information to the one associating the second question information and the second answer information.

[0068] This concludes the operation of outputting the second response information of the asynchronous dialogue system 100 in the first embodiment. As a result, the content of the dialogue of the first user is reflected in the response to the second user, and the content of the dialogue of the second user is reflected in the response to the first user. Therefore, it becomes possible to output a response based on the content of the dialogue with a plurality of users. Also, by separating the dialogue of the first user and the dialogue of the second user into different databases, the accuracy of the search is improved in steps S12 and S22, so that it becomes possible to conduct a dialogue with higher accuracy.

[0069] <Second Embodiment> Hereinafter, an example of an asynchronous dialogue system in the second embodiment to which the present invention is applied will be described with reference to the drawings. The second embodiment is different from the first embodiment in that the base model is tuned using evaluation information indicating an evaluation from the user.

[0070] FIG. 6 is a schematic diagram showing an example of the functions of the asynchronous dialogue device 1 in the second embodiment. The first storage unit 13 stores first evaluation information indicating an evaluation from the second user in a first database in association with the first question information and the first response information. The second storage unit 14 stores second evaluation information indicating an evaluation from the first user in a second database in association with the second question information and the second response information.

[0071] Next, the operation of tuning the base model of the asynchronous dialogue system 100 to which the second embodiment of the present invention is applied will be described. FIG. 7 is a flowchart of the operation of tuning the base model of the asynchronous dialogue system 100 to which the second embodiment is applied. FIG. 7(a) is a flowchart of the operation of tuning the first base model of the asynchronous dialogue system 100 to which the second embodiment is applied. FIG. 7(b) is a flowchart of the operation of tuning the second base model of the asynchronous dialogue system 100 to which the second embodiment is applied.

[0072] First, the operation of tuning the first base model of the asynchronous dialogue system 100 to which the second embodiment is applied will be described. As shown in FIG. 7(a), in step S101, the first storage unit 13 transmits the first response information and the first question information associated with the first response information to the user terminal 2b held by the second user.

[0073] Next, in step S102, the first storage unit 13 acquires the first evaluation information. The evaluation information is information indicating the evaluation of the response information associated with the question information. The evaluation information may be, for example, information indicating the accuracy of the answer to the question included in the question information. Also, the evaluation information may be information indicating the corrected answer. The first evaluation information is evaluation information indicating the evaluation by the second user of the superiority or inferiority of the first response information associated with the first question information. The first evaluation information may be evaluation information acquired from the user terminal 2b held by the second user.

[0074] In step S102, the first storage unit 13 stores the acquired first evaluation information in association with the first question information and the first response information.

[0075] Next, in step S103, the first processing unit 11 performs re-learning of the first base model based on the first question information, the first response information, and the first evaluation information. In such a case, for example, the first processing unit 11 may re-set the relevance of the first base model based on, for example, the first evaluation information. In such a case, for example, when the first evaluation information is a higher evaluation than the reference value, the relevance between the question information and the response information of the first base model referred to when outputting the first response information based on the first question information may be set higher. Also, when the first base model is a large language model, the generation probability between words of the first question information, the first response information, and the first evaluation information associated with the first evaluation information may be changed according to the height of the evaluation of the first evaluation information.

[0076] Also, when the first evaluation information is information indicating the corrected answer, the question information and the response information indicating the corrected answer may be newly used as learning data for re-learning.

[0077] As a result, the operation of tuning the first base model of the asynchronous dialogue system 100 to which the second embodiment is applied is completed.

[0078] Next, an operation of tuning the second base model of the asynchronous dialogue system 100 to which the second embodiment is applied will be described. As shown in FIG. 7(b), in step S201, the second storage unit 14 transmits the second response information and the second question information associated with the second response information to the user terminal 2a possessed by the first user.

[0079] Next, in step S202, the second storage unit 14 acquires second evaluation information. The second evaluation information is evaluation information indicating the evaluation by the first user of the superiority or inferiority of the second response information associated with the second question information. The second evaluation information may be evaluation information acquired from the user terminal 2a possessed by the first user.

[0080] In step S202, the second storage unit 14 stores the acquired second evaluation information in association with the second question information and the second response information.

[0081] Next, in step S203, the second processing unit 12 performs relearning of the second base model based on the second question information, the second response information, and the second evaluation information. In such a case, for example, the second processing unit 12 may re-set the relevance of the second base model based on, for example, the second evaluation information. In such a case, for example, when the second evaluation information is a high evaluation compared to a reference value, the relevance between the question information and the response information of the second base model referred to when outputting the second response information may be set higher based on the second question information. Further, when the second base model is a large language model, the generation probability between words of the second question information, the second response information, and the second evaluation information associated with the second evaluation information may be changed according to the height of the evaluation of the second evaluation information.

[0082] Further, when the second evaluation information is information indicating a corrected response, relearning may be performed using the question information and the response information indicating the corrected response as learning data.

[0083] As a result, the operation of tuning the second base model of the asynchronous dialogue system 100 to which the second embodiment is applied ends. As a result, the evaluation of the first user's dialogue is reflected in the response to the second user, and the evaluation of the second user's dialogue is reflected in the response to the first user. For this reason, it becomes possible to answer questions with higher accuracy.

[0084] Although the embodiments of the present invention have been described, these embodiments are presented as examples and are not intended to limit the scope of the invention. These novel embodiments can be implemented in various other forms, and various omissions, replacements, and changes can be made without departing from the gist of the invention. These embodiments and their modifications are included in the scope and gist of the invention, and are included in the invention described in the claims and its equivalent scope.

Explanation of Reference Numerals

[0085] 1 Asynchronous dialogue device 2 User terminal 3 Server 4 Public communication network 11 First processing unit 12 Second processing unit 13 First storage unit 14 Second storage unit 100 Asynchronous dialogue system 101 CPU 102 ROM 103 RAM 104 Storage unit 105 I / F 106 I / F 107 I / F 108 Input unit 109 Display unit 110 Internal bus

Claims

1. a first acquisition means for acquiring first question information indicating a question of a first user; a second acquisition means for acquiring second question information indicating a question of a second user; referring to a second database that stores the second question information and second answer information associated with the answer to the second user's question, and based on the first question information acquired by the first acquisition means, a first extraction means for extracting the second question information and the second answer information; referring to a first database that stores the first question information and first answer information associated with the answer to the first user's question, and based on the second question information acquired by the second acquisition means, a second extraction means for extracting the first question information and the first answer information; inputting the first question information acquired by the first acquisition means into a first base model, and a first output means for outputting first answer information based on the second question information and the second answer information extracted by the first extraction means; inputting the second question information acquired by the second acquisition means into a second base model, and a second output means for outputting second answer information based on the first question information and the first answer information extracted by the second extraction means A non-synchronous dialogue system characterized by the above.

2. a first storage means for storing the first question information acquired by the first acquisition means and the first answer information output by the first output means in association with each other in the first database; A second storage means for storing the second question information acquired by the second acquisition means and the second answer information output by the second output means in association with each other in the second database is further provided The asynchronous dialogue system according to claim 1, characterized by the above.

3. The first storage means stores first evaluation information indicating an evaluation of the first answer information associated with the first question information in association with the first question information and the first answer information in the first database, The second storage means stores second evaluation information indicating an evaluation of the second answer information associated with the second question information in association with the second question information and the second answer information in the second database, a first learning means for learning the first base model based on the first question information, the first answer information, and the first evaluation information stored in the first database; A second learning means for learning the second base model based on the second question information, the second answer information, and the second evaluation information stored in the second database is further provided The asynchronous dialogue system according to claim 2, characterized by the above.

4. The first storage means acquires first business information related to the business of the first user, associates the acquired first business information, the first question information, and the first answer information, and stores them in the first database. The second storage means acquires second business information related to the business of the second user, associates the acquired second business information, the second question information, and the second answer information, and stores them in the second database. The first output means inputs the first question information acquired by the first acquisition means into a first base model, and outputs the first answer information based on the second business information associated with the second question information and the second answer information extracted by the first extraction means. The second output means inputs the second question information acquired by the second acquisition means into a second base model, and outputs the second answer information based on the first business information associated with the first question information and the first answer information extracted by the second extraction means. The asynchronous dialogue system according to claim 2, characterized in that.

5. The first acquisition means acquires, as the first user, the first question information indicating a question of a construction worker. The second acquisition means acquires, as the second user, the second question information indicating a question of an office worker. The asynchronous dialogue system according to claim 1, characterized in that.

6. The first output means inputs the first question information acquired by the first acquisition means into a first base model which is a large language model (Large Language Model), and outputs the first answer information based on the second question information and the second answer information extracted by the first extraction means. The second output means inputs the second question information acquired by the second acquisition means into a second base model which is a large language model, and outputs the second answer information based on the first question information and the first answer information extracted by the second extraction means. The asynchronous dialogue system according to claim 1, characterized in that.

7. A first acquisition step of acquiring first question information indicating a question of a first user; A second acquisition step of acquiring second question information indicating a question of a second user; Referring to a second database that stores the second question information and the second answer information indicating an answer to the second user's question in association with each other, and based on the first question information acquired in the first acquisition step, a first extraction step of extracting the second question information and the second answer information. Refer to a first database that stores the first question information and the first answer information indicating the answer to the question of the first user in association with each other, and based on the second question information acquired in the second acquisition step, perform a second extraction step of extracting the first question information and the first answer information. Input the first question information acquired in the first acquisition step into a first base model, and based on the second question information and the second answer information extracted in the first extraction step, perform a first output step of outputting the first answer information. Cause a computer to execute a second output step of inputting the second question information acquired in the second acquisition step into a second base model and outputting the second answer information based on the first question information and the first answer information extracted in the second extraction step. An asynchronous dialogue program characterized by the above.

Images