Proposal System

Abstract

The goal is to obtain a proposal system that generates appropriate and beneficial suggestions while suppressing hallucination and making it difficult to suppress the occurrence of serendipity. [Solution] The trend extraction unit 23 uses LLM4a to generate input information related to the proposal based on the trend information. The input information selection unit 24 identifies the input information selected by the user from the input information generated by the trend extraction unit 23. The input information expansion unit 25 acquires unique data corresponding to the proposal for the input information selected by the user. The proposal generation unit 26 uses LLM4a to generate a proposal corresponding to the proposal request, the input information selected by the user, and the acquired unique data.

Description

Technical Field

[0001] The present invention relates to a proposal system.

Background Art

[0002] When creating a document that proposes products or the like to a customer, an information processing device creates a document considering the purpose of the proposing side (see, for example, Patent Document 1).

[0003] On the other hand, in recent years, large language models (LLMs) such as OpenAI's GPT have been put into practical use, and users can input a question (prompt) such that a desired result can be obtained as an answer into such an LLM and obtain an answer to the question from the LLM.

Prior Art Documents

Patent Documents

[0004]

Patent Document 1

Summary of the Invention

Problems to be Solved by the Invention

[0005] Although it is conceivable to generate appropriate proposal content (such as a document) for a request (proposal purpose) from a user using an LLM, the LLM has a problem of hallucination in which inappropriate answers (such as answers different from facts) are generated, and there is a possibility that an appropriate proposal may not be made to the user.

[0006] Furthermore, a technique called RAG (Retrieval-Augmented Generation) is sometimes introduced to suppress hallucination. In RAG, relevant information for a request is extracted from proprietary data not used in the LLM's training, and this relevant information is added to the request to obtain a response from the LLM. As a result, the response obtained from the LLM is limited by the relevant information, which may suppress the occurrence of serendipity.

[0007] This invention has been made in view of the above problems, and aims to provide a proposal system that generates appropriate and beneficial proposals while suppressing hallucination and making it difficult to suppress the occurrence of serendipity. [Means for solving the problem]

[0008] The proposal system according to the present invention is a proposal system that provides proposals in response to a desired proposal request, and comprises: a trend extraction unit that generates input information related to the proposal based on trend information using a large-scale language model; an input information selection unit that identifies input information selected by the user from the input information generated by the trend extraction unit; an input information expansion unit that acquires unique data corresponding to the proposal for the input information selected by the user; and a proposal generation unit that generates a proposal corresponding to the proposal request, the input information selected by the user, and the acquired unique data using a large-scale language model. [Effects of the Invention]

[0009] According to the present invention, a proposal system is obtained that generates appropriate and beneficial proposals while suppressing hallucination and making it difficult to suppress the occurrence of serendipity.

[0010] The above or other objects, features, and advantages of the present invention will become even more apparent from the following detailed description in conjunction with the accompanying drawings. [Brief explanation of the drawing]

[0011] [Figure 1]Figure 1 is a block diagram showing the configuration of the proposed system according to an embodiment of the present invention. [Figure 2] Figure 2 is a diagram illustrating the data flow of the proposed system shown in Figure 1. [Figure 3] Figure 3 shows an example of input information and target information obtained by the proposed system according to Embodiment 2. [Figure 4] Figure 4 shows an example of input information and target information obtained by the proposed system according to Embodiment 3. [Figure 5] Figure 5 shows an example of input information and target information obtained by the proposed system according to Embodiment 4. [Figure 6] Figure 6 shows other examples of input information and target information obtained by the proposed system according to Embodiment 4. [Figure 7] Figure 7 shows an example of input information and target information obtained by the proposed system according to Embodiment 5. [Modes for carrying out the invention]

[0012] Embodiments of the present invention will be described below with reference to the figures.

[0013] Embodiment 1.

[0014] Figure 1 is a block diagram showing the configuration of the proposed system according to an embodiment of the present invention. Figure 2 is a diagram illustrating the data flow of the proposed system shown in Figure 1. The proposed system 1 shown in Figure 1 is a system that provides proposals in response to proposal requests desired by the user. Specifically, the proposed system 1 is a system that provides answers (i.e., proposals) in response to prompts for proposal requests from the user using a large-scale language model (LLM) 4a, and comprises a computing device 11 as a computer, a communication device 12, and a storage device 13. In this case, the proposed system 1 is implemented on a single computer device, but it may instead be distributed and implemented on multiple computer devices.

[0015] The communication device 12 is a device (such as a network interface) capable of data communication with other devices (here, user terminal devices 3, servers 4, etc.) via a computer network 2 such as the Internet or an intranet. The user terminal device 3 is a network - communicable device such as a personal computer or a smartphone operated by the user. The server 4 incorporates an LLM 4a such as GPT. When receiving a prompt, it generates an answer corresponding to the prompt with the LLM 4a and transmits it as a response to the prompt. Here, although the LLM 4a is used via the computer network 2, the proposed system 1 may also incorporate the LLM 4a.

[0016] The storage device 13 is a non - volatile storage device such as a hard disk drive or a flash memory, and stores programs, data, etc. In the storage device 13, a trend database 13a, a knowledge database 13b, etc. are stored.

[0017] Specifically, the trend information included in the trend database 13a is stored in the storage device 13, and a database server for registering, editing, deleting, searching, etc. the trend information is executed by the arithmetic processing unit 11.

[0018] Similarly, the unique data included in the knowledge database 13b is stored in the storage device 13, and a database server for registering, editing, deleting, searching, etc. the unique data is executed by the arithmetic processing unit 11.

[0019] In this embodiment, the trend database 13a and the knowledge database 13b are implemented within the proposed system 1, but they may also be implemented as an external database server connected to the computer network 2.

[0020] Here, the arithmetic processing unit 11 executes the program in the storage device 13, It operates as a user interface processing unit 21, a prompt processing unit 22, a trend extraction unit 23, an input information selection unit 24, an input information expansion unit 25, and a proposal generation unit 26.

[0021] The user interface processing unit 21 communicates with the user terminal device 3 at the front end, uses the input device and display device of the user terminal device 3 as a user interface, obtains user operation information (information input by the user) from the input device of the user terminal device 3, and displays various information on the display device of the user terminal device 3.

[0022] In addition, an input device and a display device may be provided in the proposal system 1, and the user interface processing unit 21 may similarly obtain user operation information and display various information using the input device and the display device.

[0023] The prompt processing unit 22 generates a prompt as needed according to the proposal request and user attributes, and the user interface processing unit 21 displays the prompt to the user and obtains initial input information (here, the initial keyword list) based on the user operation. The proposal request and user attributes may be preset or input by the user. When the initial input information is not used, the prompt processing unit 22 is not particularly necessary. This prompt may be generated by the LLM4a from the proposal request and user attributes.

[0024] For example, the initial input information is one or more keywords.

[0025] The trend extraction unit 23 uses the LLM4a to generate input information related to the proposal (information input to the LLM4a when outputting the proposal) based on the trend information.

[0026] For example, the input information is one or more keywords, the trend information includes at least one of television broadcast program information, internet news, and social networking service posts, and the trend extraction unit 23 uses LLM4a to extract keywords related to the current trend from the trend information.

[0027] For example, program information for television broadcasts uses information provided as television metadata. Television metadata is data that transcribes the content of broadcast programs and commercials into text, and includes company names, product names, service names, and spot names (place names / facility names) mentioned in the programs and commercials.

[0028] For example, if there is no initial input information, the trend extraction unit 23 inputs a prompt (an inquiry to obtain input information), a suggestion request, and trend information into LLM4a, and causes LLM4a to output as an answer the products, services, companies, spots, etc. related to the suggestion request that are predicted to become trends (i.e., keywords as input information).

[0029] Furthermore, for example, if initial input information (initial keyword list) is available, the trend extraction unit 23 inputs a prompt (a query to obtain input information), a suggestion request, the initial input information, and trend information into LLM4a, and has LLM4a select the items from the initial input information that are predicted to become trends (i.e., keywords as input information) and output them as an answer. Note that here, LLM4a is made to select keywords from the initial input information, but it may also be made to extract keywords other than those in the initial input information.

[0030] The input information selection unit 24 identifies the input information selected by the user from the input information generated by the trend extraction unit 23.

[0031] For example, the input information selection unit 24 uses the user interface processing unit 21 to display a keyword list as input information generated by the trend extraction unit 23 to the user, detects a user operation in which the user selects a keyword desired by the user from that keyword list, and identifies the keyword selected by the user (i.e., the selected input information) based on that user operation.

[0032] The input information extension unit 25 acquires proprietary data corresponding to the suggestions for the input information selected by the user. This proprietary data may include, for example, data not publicly available on the internet (such as company-specific data), or data entered by the user. It may be specified by the user or pre-configured. The proprietary data may also be specific data publicly available on the internet.

[0033] The proposal generation unit 26 uses LLM4a to generate proposals corresponding to the proposal request, user-selected input information, and acquired proprietary data.

[0034] In other words, since RAG generates suggestions based on proprietary data using LLM4a, it becomes easier to generate appropriate suggestions. On the other hand, the trend extraction unit 23 extracts keywords (input information) using LLM4a without using the aforementioned proprietary data, so keywords are extracted without suppressing serendipity.

[0035] Next, the operation of the proposed system according to Embodiment 1 will be described.

[0036] The trend extraction unit 23 uses LLM4a to generate input information (in this case, a keyword list) related to the proposal based on the trend information. For example, the trend information is extracted from the trend database 13a based on the proposal request and initial input information. In other words, data containing or related to keywords included in the proposal request and initial input information is extracted from the trend database 13a as trend information.

[0037] Furthermore, the trend extraction unit 23 inputs predetermined prompts, suggestion requests, initial input information, trend information, etc., into the LLM4a, causes the LLM4a to output keywords that are predicted to become trends, and acquires such keywords.

[0038] Next, the input information selection unit 24 identifies the input information selected by the user from the input information generated by the trend extraction unit 23.

[0039] Furthermore, the input information extension unit 25 acquires proprietary data corresponding to the proposal for the input information selected by the user (i.e., the input information identified by the input information selection unit 24).

[0040] The proposal generation unit 26 then inputs the proposal request, user-selected input information, and acquired proprietary data into the LLM4a, causes the LLM4a to output objective information (i.e., a response to the proposal request), generates a proposal (such as text) containing that objective information, and presents it to the user.

[0041] As described above, according to Embodiment 1, the trend extraction unit 23 generates input information related to the proposal based on the trend information using LLM4a. The input information selection unit 24 identifies the input information selected by the user from the input information generated by the trend extraction unit 23. The input information expansion unit 25 acquires unique data corresponding to the proposal for the input information selected by the user. The proposal generation unit 26 generates a proposal corresponding to the proposal request, the input information selected by the user, and the acquired unique data using LLM4a.

[0042] This suppresses hallucination while making it difficult to suppress serendipity, thereby generating appropriate and beneficial suggestions. In other words, since the input information generated without suppressing serendipity is manually selected by the user and input into LLM4a to generate suggestions, appropriate and beneficial suggestions are obtained.

[0043] Embodiment 2.

[0044] The proposed system according to Embodiment 2 is a system that proposes a list of potential clients for a specific industry. Here, as an example, a proposal for a list of potential clients for a frozen warehouse (frozen warehouse business) is required. Therefore, the proprietary data mentioned above includes a list of companies that are potential clients. In the case of the frozen warehouse business, the proprietary data used is a list of companies that manufacture or distribute goods that require freezing.

[0045] Figure 3 shows an example of input information and target information obtained by the proposed system according to Embodiment 2.

[0046] In Embodiment 2, the prompt processing unit 22 generates a prompt (input information request) as shown in Figure 3, for example, according to the suggestion request and user attributes, and the user interface processing unit 21 displays the prompt to the user. When the user enters a keyword list or news list, the prompt processing unit 22 and the user interface processing unit 21 acquire the initial input information.

[0047] The trend extraction unit 23 then uses TV metadata, Google News, and Yahoo Trend Search as trend information to obtain trend information (i.e., data content in which the keyword appears) for each keyword in the initial input information (initial keyword list), inputs this trend information and keyword list (along with a prompt to provide the keyword list described later) into LLM4a, and obtains a list of keywords from LLM4a that are predicted to become trends from the initial keyword list. For example, in Figure 3, three keywords from the initial keyword list, "frozen vegetables," "frozen meal kits," and "frozen sweets," are obtained from LLM4a.

[0048] The input information selection unit 24 displays the list of keywords predicted to become trends, obtained in this manner, to the user and identifies the keyword selected by the user from that list. For example, in Figure 3, two keywords, "frozen vegetables" and "frozen sweets," have been selected by the user from the list of keywords predicted to become trends by LLM4a.

[0049] The input information extension unit 25 then acquires customer lists (company lists) and external company data as proprietary data from the knowledge database 13b and an external server. The proposal generation unit 26 inputs the two keywords and their proprietary data (along with prompts to provide target information) into the LLM 4a, and outputs a list of companies to be targeted for sales and their information (information related to those two keywords) as target information to the LLM 4a, for example, as shown in Figure 3. Once the proposal generation unit 26 acquires this target information, it displays a proposal containing that target information to the user.

[0050] Furthermore, the other configurations and operations of the proposed system according to Embodiment 2 are the same as those of Embodiment 1, so their description will be omitted.

[0051] Embodiment 3.

[0052] The proposed system according to Embodiment 3 is a system that proposes a list of investment targets in a specific industry. Here, as an example, a proposal for a list of investment targets in the food industry is required. Therefore, the proprietary data mentioned above includes a list of companies in that specific industry, and company information for each company in the company list (IR (Investor Relations) information, stock price data, financial results data, economic indicator data, lists of products, services, and businesses, etc.).

[0053] Figure 4 shows an example of input information and target information obtained by the proposed system according to Embodiment 3.

[0054] In Embodiment 3, the prompt processing unit 22 generates a prompt (input information request) as shown in Figure 4, for example, according to the suggestion request and user attributes, and the user interface processing unit 21 displays the prompt to the user. When the user enters a keyword list or news list, the prompt processing unit 22 and the user interface processing unit 21 acquire the initial input information.

[0055] The trend extraction unit 23 then uses TV metadata, Google News, and Yahoo Trend Search as trend information to obtain trend information (i.e., data content in which the keyword appears) for each keyword in the initial input information (initial keyword list), inputs this trend information and keyword list (along with a prompt to answer with the keyword list described later) into LLM4a, and obtains a list of keywords from LLM4a that are predicted to become trends from the initial keyword list. For example, in Figure 4, three keywords from the initial keyword list, "sustainability," "mergers and acquisitions," and "organic food," are obtained from LLM4a.

[0056] The input information selection unit 24 displays to the user a list of keywords predicted to become trends, and identifies the keywords selected by the user from that list. For example, in Figure 4, three keywords, "sustainability," "mergers and acquisitions," and "organic food," are selected by the user from the list of keywords predicted to become trends by LLM4a.

[0057] The input information extension unit 25 then acquires proprietary data, including a list of companies in the food industry and company information for each company in the list, from the knowledge database 13b or an external server. The proposal generation unit 26 inputs the three keywords and their proprietary data (along with prompts to provide the target information) into the LLM 4a, causing it to output a list of companies to be invested in, along with trend information and reasons for investment decisions for those companies, as target information, as shown in Figure 4, for example, from the LLM 4a. Once the proposal generation unit 26 acquires this target information, it displays a proposal containing that target information to the user.

[0058] The other configurations and operations of the proposed system according to Embodiment 3 are the same as those of Embodiment 1, so their description will be omitted.

[0059] Embodiment 4.

[0060] The proposed system according to Embodiment 4 is a system that proposes content for a tourism website introducing tourist attractions. Here, as an example, a proposal for an introductory text for Glover Garden is requested. In this case, the proprietary data mentioned above includes basic information about Glover Garden. For example, if there is a webpage for Glover Garden on an existing website, the content already published on that webpage will be included in the proprietary data. In that case, for example, content to be added to the existing website will be proposed.

[0061] Figure 5 shows an example of input information and target information obtained by the proposed system according to Embodiment 4. Figure 6 shows another example of input information and target information obtained by the proposed system according to Embodiment 4.

[0062] In Embodiment 4, without using initial input information, the trend extraction unit 23 uses television metadata, SNS (Social Networking Service) posts published on the internet, etc., as trend information to obtain trend information about the suggestion request (Glover Garden) (i.e., data content in which the keyword (Glover Garden) included in the suggestion request appears), inputs that trend information into LLM4a (along with a prompt to answer with the keyword list described later), and obtains a list of keywords that are predicted to become trends based on that trend information from LLM4a. For example, in Figure 5, keywords such as "retro photos," "European nobility," and "medieval costumes" are obtained from LLM4a.

[0063] For trend information, you may use data already registered in the trend database 13a, or you may use data publicly available on the internet.

[0064] The input information selection unit 24 displays to the user a list of keywords that are predicted to become trends, and identifies the keyword selected by the user from that list. For example, in Figure 5, two keywords, "medieval clothing" and "Masonic gates," have been selected by the user from the list of keywords predicted to become trends by LLM4a.

[0065] The input information extension unit 25 then obtains basic information about Glover Garden (such as existing web pages) as proprietary data from the knowledge database 13b or an external server. The suggestion generation unit 26 inputs the two keywords mentioned above and their proprietary data (along with prompts to provide the target information) into the LLM4a, and outputs an introductory text as the target information to the LLM4a, for example, as shown in Figure 5. Once the suggestion generation unit 26 obtains the target information, it displays a suggestion containing that information to the user.

[0066] Furthermore, if the user selects two keywords, "Ryoma Sakamoto" and "exotic atmosphere," from the list of keywords predicted to trend by LLM4a, different target information will be obtained compared to the case in Figure 5, as shown in Figure 6, for example. In this way, content is generated with input information that reflects the user's judgment as an expert, thus not hindering serendipity and reducing the likelihood of hallucination.

[0067] Furthermore, the other configurations and operations of the proposed system according to Embodiment 4 are the same as those of Embodiment 1, so their description will be omitted.

[0068] Embodiment 5.

[0069] The proposed system according to Embodiment 5 is a system that proposes travel plans for a specific tourist destination. Here, as an example, a proposal for a travel plan for Nagasaki is requested. In this case, the proprietary data mentioned above includes tourist settings specified by the user.

[0070] Figure 7 shows an example of input information and target information obtained by the proposed system according to Embodiment 5.

[0071] In Embodiment 5, the prompt processing unit 22 generates a prompt (input information request) as shown in Figure 7, for example, and the user interface processing unit 21 displays the prompt to the user. When the user inputs a keyword list, the prompt processing unit 22 retrieves the initial input information from the user interface processing unit 21. For example, a list of keywords related to a specific tourist destination (in this case, Nagasaki) is displayed in the input information request, and the user selects keywords of interest from that list. This results in a keyword list desired by the user (in Figure 7, "exotic atmosphere," "gourmet," and "anime pilgrimage site").

[0072] The trend extraction unit 23 then uses television metadata, SNS (social networking service) posts, etc., as trend information to obtain trend information related to the keyword list, inputs this trend information into LLM4a (along with a prompt to answer with the keyword list described later), and obtains a list of keywords that are predicted to become trends from LLM4a based on that trend information. For example, in Figure 7, three keywords are obtained from LLM4a: "Movie 'Kimi no Iro'", "Lemon Steak", and "Nagasaki Confucius Temple / Chinese Historical Museum".

[0073] The input information selection unit 24 displays to the user a list of keywords that are predicted to become trends, and identifies the keyword selected by the user from that list. For example, in Figure 7, two keywords, "The movie 'Kimi no Iro'" and "Lemon Steak," have been selected by the user from the list of keywords predicted to become trends by LLM4a.

[0074] Furthermore, the input information extension unit 25 acquires the sightseeing settings specified by the user as proprietary data. For example, the sightseeing settings include setting items such as the starting location (within or outside Nagasaki), the number of sightseeing days (half-day, one day, one night and two days, etc.), the number of spots (few, normal, many, etc.), and the amount of travel (little, round trip, etc.). The input information extension unit 25 acquires the sightseeing settings by displaying the options for each setting item in the user interface processing unit 21 and identifying the option selected by the user for each setting item.

[0075] The suggestion generation unit 26 then inputs the two keywords and their unique data (along with prompts to provide target information) into the LLM4a, causing it to output a sightseeing plan as target information, for example, as shown in Figure 7. Once the suggestion generation unit 26 obtains this target information, it displays a suggestion containing that information to the user.

[0076] The other configurations and operations of the proposed system according to Embodiment 5 are the same as those of Embodiment 1, so their description will be omitted.

[0077] Furthermore, various changes and modifications to the embodiments described above will be obvious to those skilled in the art. Such changes and modifications may be made without deviating from the spirit and scope of the subject matter and without diminishing the intended advantages. In other words, such changes and modifications are intended to be included in the claims. [Industrial applicability]

[0078] The present invention is applicable, for example, to a proposal system that generates and provides proposals using LLM. [Explanation of Symbols]

[0079] 1. Proposed System 4a Large-scale language models 23 Trend Extraction Section 24 Input Information Selection Section 25 Input Information Expansion Unit 26 Proposal generation section

Claims

1. In a proposal system that provides proposals in response to a desired request for proposals, A trend extraction unit that generates input information related to the proposal based on trend information using a large-scale language model, An input information selection unit identifies the input information selected by the user from the input information generated by the trend extraction unit, An input information extension unit that acquires proprietary data corresponding to the proposal with respect to the input information selected by the user, A proposal generation unit that uses a large-scale language model to generate proposals corresponding to the proposal request, the input information selected by the user, and the acquired proprietary data, A proposed system characterized by comprising the following features.

2. The aforementioned input information consists of one or more keywords. The aforementioned trend information includes at least one of the following: television broadcast program information, internet news, and social networking service posts. The trend extraction unit extracts keywords related to the current trend from the trend information using a large-scale language model. The proposed system according to claim 1, characterized by the following:

3. The proposed system according to claim 1, characterized in that the trend extraction unit extracts keywords related to the current trend from the trend information using the large-scale language model without using the proprietary data.

4. The aforementioned proposal is a list of potential clients in a specific industry. The aforementioned proprietary data includes a list of companies. The proposed system according to claim 1, characterized by the following:

5. The aforementioned proposal is a list of investment targets in specific industries, The aforementioned proprietary data includes a list of companies and company information about the companies in the company list. The proposed system according to claim 1, characterized by the following:

6. The aforementioned proposal is content for a tourism website that introduces tourist attractions. The aforementioned proprietary data includes basic information about the aforementioned tourist attractions. The proposed system according to claim 1, characterized by the following:

7. The above proposal is a tourism plan, The aforementioned proprietary data includes the tourism conditions entered by the user. The proposed system according to claim 1, characterized by the following:

8. The aforementioned proposal is content that should be added to the existing website. The aforementioned proprietary data includes content already published on the aforementioned existing website. The proposed system according to claim 1 or claim 6, characterized by the above.

Images