Prompt creation system

Abstract

This system provides a prompt creation system capable of generating prompts of appropriate size. [Solution] When the server device 3 of the prompt creation system 1 receives a question entered by the user device 2, the category learning model 4 is used to estimate the category to be assigned to the question, and the interactive learning model 5 is used to create a prompt (category prompt) that generates an answer to the question by referring to the provided information corresponding to the estimated category. In addition, keywords are extracted from the question, and the database unit 11 is searched using the keywords, and the search results are obtained. Then, by linking the category prompt and the search results, a prompt (linked category prompt) that generates an answer to the question by referring to the provided information and the search results is created.

Description

Technical Field

[0001] The present invention relates to a prompt creation system for creating a prompt for causing an interactive learning model to generate an answer to a question.

Background Art

[0002] In recent years, interactive learning models have come to be used in various operations. For example, a VPA (Virtual Personal Assistant) that accepts inquiries by voice, chat, etc. via a dedicated device or application and immediately provides an appropriate answer or response by utilizing AI is also becoming widespread (see Patent Document 1).

Prior Art Documents

Patent Documents

[0003] <a000016>

Patent Document 1

Summary of the Invention

Problems to be Solved by the Invention

[0004] In order to obtain an appropriate answer from AI, it is necessary to input an appropriate prompt to the AI. Certainly, if all the information (a large amount of information) required to generate an appropriate answer is included in the prompt, an appropriate answer can be obtained from the AI.

[0005] However, if the size (information amount) of the prompt is too large, it may take time for the AI to process and the response may be delayed. Also, a large amount of AI resources will be consumed, and there is a risk of reduced efficiency. On the other hand, if the size (information amount) of the prompt is too small, there is a risk that an appropriate answer cannot be obtained from the AI. Therefore, the development of a system that can create a prompt of an appropriate size (information amount) has been desired.

[0006] This invention has been made in view of the above-mentioned problems, and aims to provide a prompt creation system capable of creating prompts of appropriate size. [Means for solving the problem]

[0007] The prompt generation system of the present invention is a prompt generation system that generates prompts for an interactive learning model to generate answers to questions when questions are input to a contact center from a user device, the prompt generation system includes: a question receiving unit that receives questions input by the user device; a category estimation unit that uses a category learning model that has learned the relationship between questions to the contact center and the categories to be assigned to those questions, to take the received question as input and estimate and output the categories to be assigned to the received question; a provision information registration unit in which provision information provided by a service provider regarding the questions to the contact center is pre-registered for each predetermined category; and the prompt generation system generates prompts for an interactive learning model by referring to the provision information registered in the provision information registration unit based on the estimated category. The learning model includes a category prompt creation unit that creates a category prompt for generating an answer to the received question by referring to the provided information corresponding to the estimated category; a keyword extraction unit that extracts keywords contained in the received question; a database unit in which detailed information related to the question to the contact center is registered and linked to predetermined keywords; a search processing unit that searches the database unit using the extracted keywords to obtain detailed information corresponding to the keywords as search results; and a linked category prompt creation unit that links the category prompt and the search results to create a linked category prompt for generating an answer to the received question by referring to the provided information and the search results in the interactive learning model.

[0008] In this configuration, when a question is received from a user device, a category learning model is used to estimate the category to which the question (the question received from the user device) should be assigned. Then, an interactive learning model is used to create a prompt (category prompt) that generates an answer to the question by referring to the provided information (pre-registered provided information) corresponding to the estimated category. Furthermore, keywords are extracted from the question (the received question), and the database is searched using these keywords to obtain search results. Finally, by concatenating the category prompt and the search results, a prompt (concatenated category prompt) of an appropriate size is created that generates an answer to the question by referring to the provided information and the search results. In this way, a prompt of an appropriate size can be created.

[0009] Furthermore, in the prompt creation system of the present invention, the questions sent to the contact center include questions about tourist destinations, and the database unit comprises a tourist destination database in which detailed information of the tourist destination and identification information of the tourist destination are registered and linked to each other, and a vector database in which vector information of the name of the tourist destination and identification information of the tourist destination are registered and linked to each other, and the search processing unit generates vectorized keyword information by vectorizing the extracted keywords, and uses the generated vectorized keyword information to perform a search in the vector database to obtain identification information of tourist destinations linked to vector information similar to the vectorized keyword information, and uses the obtained identification information of tourist destinations to perform a search in the tourist destination database to obtain detailed information of tourist destinations linked to the identification number as the search result.

[0010] In this configuration, when a search is performed in the database section, first, a keyword (e.g., "Tokyo Tower") is vectorized, and vectorized keyword information (e.g., "Vector numbers representing Tokyo Tower: [0.021, -0.033, 0.087, ..., 0.072]") is generated. Next, a search is performed within the vector database using this vectorized keyword information, and tourist destination identification information (e.g., "Tokyo Tower identification number: T0001") linked to vector information similar to the vectorized keyword information is obtained. Then, a search is performed within the tourist destination database using this identification information (tourist destination identification information), and detailed information about the tourist destination linked to the identification information (e.g., "Tokyo Tower is...") is obtained as a search result. In this way, by using a vector database, variations in keywords (e.g., "Tokyo Tower" and "Tokyo Tower") can be absorbed, and appropriate search results (e.g., detailed information about "Tokyo Tower") can be obtained as search results.

[0011] Furthermore, in the prompt creation system of the present invention, if the search processing unit is unable to obtain the search results, the linked category prompt creation unit may create a linked category prompt for the interactive learning model to generate an answer to the received question by referring only to the category prompt.

[0012] In this configuration, if search results cannot be obtained, a concatenated category prompt is created using only the category prompt without using the search results. In this way, a prompt (concatenated category prompt) that generates an answer to a question by referring only to the provided information is created with an appropriate size.

[0013] Furthermore, in the prompt creation system of the present invention, if the category prompt creation unit does not have any information to provide that corresponds to the estimated category registered in the information provision registration unit, the category prompt creation unit may create a category prompt to cause the interactive learning model to generate an answer indicating that it does not have any information to provide in response to the received question.

[0014] In this configuration, if the information provided for the estimated category is not pre-registered, a category prompt is created to generate an answer indicating that there is no information to provide for the question. By concatenating this category prompt with the search results, a concatenated category prompt of appropriate size is created to generate an answer indicating that there is no information to provide for the question.

[0015] The present invention is a method performed by a prompt generation system that generates prompts for an interactive learning model to generate answers to questions when questions are input to a contact center from a user device, wherein the prompt generation system comprises: a provision information registration unit in which provision information provided by a service provider regarding questions to the contact center is pre-registered for predetermined categories, and a database unit in which detailed information related to questions to the contact center is registered in association with predetermined keywords, wherein the method comprises the steps of receiving a question input by the user device, and using a category learning model that has learned the relationship between questions to the contact center and categories to be assigned to those questions, the received question is used as input to generate prompts for a category to be assigned to the received question The method includes the steps of: estimating and outputting a category; creating a category prompt for the interactive learning model to generate an answer to the received question by referring to the provided information registered in the provided information registration unit based on the estimated category; extracting keywords contained in the received question from the question; obtaining detailed information corresponding to the keywords as search results by searching the database unit using the extracted keywords; and creating a linked category prompt for the interactive learning model to generate an answer to the received question by referring to the provided information and the search results.

[0016] This method, as with the system described above, also involves receiving a question entered by the user device. A category learning model is used to estimate the category to be assigned to that question (the received question), and an interactive learning model is used to create a prompt (category prompt) that generates an answer to the question by referencing the provided information (pre-registered provided information) corresponding to the estimated category. Keywords are also extracted from the question (the received question), and the database is searched using these keywords to obtain search results. Finally, by concatenating the category prompt with the search results, a prompt (concatenated category prompt) of appropriate size is created that references the provided information and the search results to generate an answer to the question. In this way, a prompt of appropriate size can be created.

[0017] The present invention is a program that runs on a prompt creation system that creates a prompt for an interactive learning model to generate an answer to a question when a question is input to a contact center from a user device, wherein the prompt creation system comprises: a provision information registration unit in which provision information provided by a service provider regarding the question to the contact center is pre-registered for predetermined categories; and a database unit in which detailed information related to the question to the contact center is registered in association with predetermined keywords, wherein the program instructs the prompt creation system to process a question input by the user device and to use a category learning model that has learned the relationship between the question to the contact center and the category to be assigned to that question, to use the received question as input, The system performs the following processes: estimating and outputting a category to be assigned to a received question; creating a category prompt for the interactive learning model to generate an answer to the received question by referring to the provided information registered in the provided information registration unit based on the estimated category; extracting keywords contained in the received question from the question; obtaining detailed information corresponding to the keywords as search results by searching the database unit using the extracted keywords; and creating a linked category prompt for the interactive learning model to generate an answer to the received question by referring to the provided information and the search results.

[0018] Similar to the system described above, this program, upon receiving a question entered by the user device, uses a category learning model to estimate the category to be assigned to that question (the received question). Using an interactive learning model, it creates a prompt (category prompt) to generate an answer to the question by referring to the provided information (pre-registered provided information) corresponding to the estimated category. Keywords are also extracted from the question (the received question), and the database is searched using these keywords to obtain search results. By concatenating the category prompt with the search results, a prompt (concatenated category prompt) of appropriate size is created to generate an answer to the question by referring to the provided information and the search results. In this way, a prompt of appropriate size can be created. [Effects of the Invention]

[0019] According to the present invention, it is possible to create prompts of an appropriate size. [Brief explanation of the drawing]

[0020] [Figure 1] This is a block diagram showing the configuration of a prompt creation system in an embodiment of the present invention. [Figure 2] This is a sequence diagram illustrating the operation of the prompt generation system in an embodiment of the present invention. [Modes for carrying out the invention]

[0021] The prompt creation system according to an embodiment of the present invention will be described below with reference to the drawings. In the present embodiment, the case of a prompt creation system used in a system for assisting a contact center or the like will be exemplified. The prompt creation system according to the present embodiment has a function of creating a prompt of an appropriate size. This function is realized by hardware of the prompt creation system (for example, a processor of a server device) cooperating with a program stored in a storage area of the prompt creation system (for example, a memory of the server device).

[0022] The configuration of the prompt creation system according to an embodiment of the present invention will be described with reference to the drawings. FIG. 1 is a block diagram showing the configuration of the prompt creation system according to the present embodiment. As shown in FIG. 1, the prompt creation system 1 includes a server device 3 connected via a network N1 to a user device 2, and the server device 3 is connected via a network N2 to a category learning model 4 and a dialogue learning model 5. This prompt creation system 1 has a function of creating a prompt for causing the dialogue learning model 5 to generate an answer to a question when the question to the contact center is input from the user device 2.

[0023] The user device 2 is, for example, a smartphone or a personal computer, and the server device 3 is, for example, a cloud server. The category learning model 4 is a learning model that has learned the relationship between questions to the contact center and categories to be assigned to those questions. Note that the networks N1 and N2 may be different networks or the same network.

[0024] Referring to Figure 1, the configuration of the server device 3 will be explained in detail. As shown in Figure 1, the server device 3 comprises a question receiving unit 6, a category estimation unit 7, a provided information registration unit 8, and a category prompt creation unit 9. The question receiving unit 6 has the function of receiving questions entered by the user device 2. The category estimation unit 7 uses a category learning model 4 to take the received question as input and has the function of estimating and outputting the category to be assigned to the received question.

[0025] For example, when the question receiving unit 6 receives the question "Can you recommend a good place to see cherry blossoms?", the category assignment unit uses the category learning model 4 to estimate and output the category "tourist destination". Also, when the question receiving unit 6 receives the question "What are the opening hours of Tokyo Tower?", the category assignment unit uses the category learning model 4 to estimate and output the category "tourist destination".

[0026] The Information Registration Section 8 contains pre-registered information provided by service providers regarding inquiries to the contact center, categorized accordingly. For example, under the category of "Tourist Destinations," the Information Registration Section 8 has pre-registered information such as "The following are famous tourist destinations in Tokyo... The following are seasonal recommendations..."

[0027] The category prompt creation unit 9 has the function of creating a prompt (category prompt) for the interactive learning model 5 (for example, ChatGPT®) to generate an answer to a received question by referring to the provided information registered in the provided information registration unit 8 based on the estimated category.

[0028] For example, if the category "tourist destination" is inferred, a category prompt is created to generate an answer to the received question by referring to the provided information corresponding to that category, such as "The following are famous tourist destinations in Tokyo... The following are seasonal recommendations..."

[0029] Furthermore, if the provided information corresponding to the estimated category is not pre-registered in the provided information registration unit 8, the category prompt creation unit 9 creates a category prompt to cause the interactive learning model 5 to generate an answer indicating that it does not have the information to provide for the received question.

[0030] Furthermore, the server device 3 includes a keyword extraction unit 10, a database unit 11, a search processing unit 12, and a linked category prompt creation unit 13. The keyword extraction unit 10 has the function of extracting keywords contained in a question from a received question. For example, the keyword extraction unit 10 extracts the keyword "cherry blossom viewing" from the question "Tell me some good places to see cherry blossoms." Also, the keyword extraction unit 10 extracts the keyword "Tokyo Tower" from the question "What are the opening hours of Tokyo Tower?".

[0031] The database unit 11 stores detailed information related to inquiries to the contact center (for example, inquiries about tourist destinations), linked to predetermined keywords. In this embodiment, the database unit 11 includes a vector database 110 and a tourist destination database 111. The vector database 110 stores vector information representing the names of tourist destinations (for example, a vector number representing "Tokyo Tower": [0.031,-0.024,0.102,···,0.087]) and identification information for those tourist destinations (for example, "Identification number for Tokyo Tower: T0001"), linked to each other. The tourist destination database 111 stores detailed information about tourist destinations and identification information for those tourist destinations, linked to each other.

[0032] The search processing unit 12 has a function (RAG search function) that retrieves detailed information corresponding to the keywords as search results by searching the database unit 11 using the extracted keywords. In this embodiment, the search processing unit 12 generates vectorized keyword information (for example, "Vector numbers representing Tokyo Tower: [0.021, -0.033, 0.087, ..., 0.072]") by vectorizing the extracted keywords (for example, "Tokyo Tower"), and retrieves tourist destination identification information (for example, "Identification number for Tokyo Tower: T0001") associated with vector information similar to the vectorized keyword information by performing a search in the vector database 110 using the generated vectorized keyword information, and retrieves detailed information about the tourist destination associated with the identification number (for example, "Tokyo Tower is...") as search results by performing a search in the tourist destination database 111 using the retrieved tourist destination identification information.

[0033] The linked category prompt creation unit 13 has the function of creating a prompt (linked category prompt) that allows the interactive learning model 5 to generate an answer to a received question by referring to the provided information and search results, by linking category prompts with search results. In this embodiment, if the search processing unit 12 is unable to obtain search results, the linked category prompt creation unit 13 creates a linked category prompt that allows the interactive learning model 5 to generate an answer to a received question by referring only to category prompts.

[0034] Furthermore, the server device 3 includes a dialogue processing unit 14 that uses the created linked category prompts to cause the interactive learning model 5 to generate answers to questions by referring to the provided information and search results. The dialogue processing unit 14 inputs the created linked category prompts to the interactive learning model 5 along with the questions entered by the user device 2, thereby causing the interactive learning model 5 to generate answers to questions by referring to the provided information and search results. The dialogue processing unit 14 also has a function to send the answers to questions generated by the interactive learning model 5 to the user device 2 (the user device 2 that entered the question).

[0035] The operation of the prompt generation system 1, configured as described above, will be explained with reference to the sequence diagram in Figure 2.

[0036] When the prompt creation system 1 of this embodiment is used, first, a question (for example, a question about a tourist destination) is input from the user device 2 to the contact center (S1). The question input by the user device 2 is transmitted to the server device 3 via the network N1 (S2).

[0037] When server device 3 receives a question entered by user device 2 (S3), it uses category learning model 4 to estimate the category to be assigned to the question (S4), and uses interactive learning model 5 to create a prompt (category prompt) to generate an answer to the question by referring to the provided information corresponding to the estimated category (S5).

[0038] Furthermore, keywords are extracted from the question (S6), and the database unit 11 is searched using the keywords to obtain search results (S7). Then, by linking the category prompt and the search results, a prompt (linked category prompt) is created to generate an answer to the question by referring to the provided information and the search results (S8).

[0039] The server device 3 then inputs the created linked category prompt, along with the question entered by the user device 2, into the interactive learning model 5 (S9), causing the interactive learning model 5 to generate an answer to the question by referring to the provided information and search results (S10). When the server device 3 receives the answer to the question generated by the interactive learning model 5 from the interactive learning model (S11), it sends the answer to the user device 2 (the user device 2 that entered the question) (S12). The user device 2 then displays the answer received from the server device 3 (S13).

[0040] According to the prompt creation system 1 of this embodiment, when a question is received by the user device 2, the category learning model 4 is used to estimate the category to be assigned to the question (the question received from the user device 2), and the interactive learning model 5 is used to create a prompt (category prompt) for generating an answer to the question by referring to the provided information (pre-registered provided information) corresponding to the estimated category. In addition, keywords are extracted from the question (the received question), and the database unit 11 is searched using the keywords to obtain the search results. Then, by linking the category prompt and the search results, a prompt (linked category prompt) for generating an answer to the question by referring to the provided information and the search results is created in an appropriate size. In this way, a prompt of an appropriate size can be created.

[0041] Furthermore, in this embodiment, when a search is performed in the database unit 11, first, a keyword (for example, "Tokyo Tower") is vectorized, and vectorized keyword information (for example, "Vector numbers representing Tokyo Tower: [0.021, -0.033, 0.087, ..., 0.072]") is generated. Next, a search is performed in the vector database 110 using this vectorized keyword information, and tourist destination identification information (for example, "Identification number for Tokyo Tower: T0001") linked to vector information similar to the vectorized keyword information is obtained. Then, a search is performed in the tourist destination database 111 using this identification information (tourist destination identification information), and detailed information about the tourist destination linked to the identification information (for example, "Tokyo Tower is...") is obtained as a search result. In this way, by using the vector database 110, it is possible to absorb keyword variations (for example, "Tokyo Tower" and "Tokyo Tower") and obtain appropriate search results (for example, detailed information about "Tokyo Tower") as search results.

[0042] Furthermore, in this embodiment, if search results cannot be obtained, a concatenated category prompt is created using only the category prompt without using the search results. In this way, a prompt (concatenated category prompt) for generating an answer to a question by referring only to the provided information is created with an appropriate size.

[0043] Furthermore, in this embodiment, if the information to be provided corresponding to the estimated category is not pre-registered, a category prompt is created to generate an answer indicating that there is no information to provide for the question. By linking this category prompt with the search results, a linked category prompt of an appropriate size is created to generate an answer indicating that there is no information to provide for the question.

[0044] Although embodiments of the present invention have been described above by example, the scope of the present invention is not limited to these, and modifications and alterations can be made within the scope described in the claims depending on the purpose.

[0045] For example, the above description described an example where the question to the contact center was about a tourist destination, but the scope of the present invention is not limited to this. The same can be done even if the question to the contact center is about a variety of things other than tourist destinations. [Industrial applicability]

[0046] As described above, the prompt creation system according to the present invention has the effect of being able to create prompts of an appropriate size and is useful as a system to support contact centers, etc. [Explanation of Symbols]

[0047] 1. Prompt creation system 2. User devices 3 Server equipment 4 Category Learning Models 5. Interactive Learning Models 6. Question Reception Section 7. Category Estimation Section 8. Information Registration Department 9. Category Prompt Creation Section 10 Keyword Extraction Section 11. Database Section 110 Vector Databases 111 Tourist Destination Database 12 Search Processing Unit 13. Linked Category Prompt Creation Section 14 Interaction Processing Unit N1, N2 Network

Claims

1. A prompt generation system that, when a question is input to a contact center from a user device, generates prompts for an interactive learning model to generate an answer to that question, The prompt generation system is, A question receiving unit that receives questions entered by the user device, A category estimation unit that uses a category learning model that has learned the relationship between questions sent to the contact center and the categories to be assigned to those questions, takes the received question as input, estimates the category to be assigned to the received question, and outputs it as an output. The information provided by the service provider regarding the questions sent to the aforementioned contact center is registered in the information registration unit, which is pre-registered for each predetermined category. A category prompt creation unit creates a category prompt for the interactive learning model to generate an answer to the received question by referring to the provided information registered in the provided information registration unit based on the estimated category, From the received question, a keyword extraction unit extracts keywords contained in the question, A database section in which detailed information related to the questions sent to the aforementioned contact center is registered and linked to predetermined keywords, A search processing unit that searches the database using the extracted keywords and obtains detailed information corresponding to the keywords as search results, A linked category prompt creation unit creates a linked category prompt that causes the interactive learning model to generate an answer to the received question by referring to the provided information and the search results, by linking the category prompt and the search results. A prompt creation system equipped with the following features.

2. The questions sent to the aforementioned contact center included questions about tourist destinations. The aforementioned database unit is A tourist destination database in which detailed information about the aforementioned tourist destination and identification information about the said tourist destination are linked to each other and registered, A vector database in which vector information representing the names of the aforementioned tourist destinations and identification information for those tourist destinations are registered and linked to each other, Equipped with, The aforementioned search processing unit, By vectorizing the extracted keywords, vectorized keyword information is generated. By performing a search within the vector database using the generated vectorized keyword information, identification information of tourist destinations linked to vector information similar to the vectorized keyword information is obtained. The prompt creation system according to claim 1, wherein a search is performed within the tourist destination database using the acquired tourist destination identification information, and detailed information of the tourist destination linked to the identification information is obtained as the search result.

3. The linked category prompt creation unit, The prompt creation system according to claim 1, wherein if the search processing unit fails to obtain the search results, it creates a linked category prompt for the interactive learning model to generate an answer to the received question by referring only to the category prompt.

4. The category prompt creation unit described above is: The prompt creation system according to claim 1, wherein if no information corresponding to the estimated category is pre-registered in the information provision registration unit, a category prompt is created to cause the interactive learning model to generate an answer indicating that it does not have information to provide in response to the received question.

5. A method is used in a prompt generation system to create prompts for an interactive learning model to generate answers to questions entered into a contact center from a user device. The prompt generation system is, The information provided by the service provider regarding the questions sent to the aforementioned contact center is registered in the information registration unit, which is pre-registered for each predetermined category. A database section in which detailed information related to the questions sent to the aforementioned contact center is registered and linked to predetermined keywords, Equipped with, The aforementioned method, The steps include receiving a question entered by the user device, The step of using a category learning model that has learned the relationship between questions sent to the contact center and the categories to be assigned to those questions, to take the received question as input and estimate and output the category to be assigned to the received question. The steps include: creating a category prompt for the interactive learning model to generate an answer to the received question by referring to the provided information registered in the provided information registration unit based on the estimated category; From the received question, the steps include: extracting keywords contained in the question; The steps include: searching the database section using the extracted keywords to obtain detailed information corresponding to the keywords as search results; The steps include creating a linked category prompt by linking the category prompt and the search results, thereby causing the interactive learning model to generate an answer to the received question by referring to the provided information and the search results, Methods that include...

6. A program executed in a prompt generation system that, when a question is input to a contact center from a user device, creates a prompt to cause an interactive learning model to generate an answer to that question. The prompt generation system is, The information provided by the service provider regarding the questions sent to the aforementioned contact center is registered in the information registration unit, which is pre-registered for each predetermined category. A database section in which detailed information related to the questions sent to the aforementioned contact center is registered and linked to predetermined keywords, Equipped with, The program in the prompt generation system The process of receiving a question entered by the user device, The process involves using a category learning model that has been trained on the relationship between questions sent to the contact center and the categories to be assigned to those questions, to take the received questions as input, estimate the categories to be assigned to those questions, and output them as output. Based on the estimated category, the interactive learning model performs a process to create a category prompt for generating an answer to the received question by referring to the provided information registered in the provided information registration unit, and by referring to the provided information corresponding to the estimated category. The process involves extracting keywords from the received question, The process involves searching the database section using the extracted keywords to obtain detailed information corresponding to the keywords as search results. By linking the category prompt and the search results, the interactive learning model is given the process of creating a linked category prompt that generates an answer to the received question by referring to the provided information and the search results, A program that executes something.

Images