Assist Server

Abstract

To provide a server intermediation system, an intermediation server, and an AI server assist system that can reduce the labor required for inputting data into AI and process items that have not yet been learned. [Solution] The assist server 2 is equipped with a trained model that is trained to input data to be processed received from an external server connected via a communication line, process the data to be processed, and output processed data to be sent to the external server, and a control unit that, when the trained model outputs a response indicating that it has not yet learned about the input data to be processed, sends keywords related to matters that the trained model has not yet learned to a search server 5 having a search engine to perform a search, accesses the URL of an external site 6 listed in the search results output by the search, acquires the content of the external site 6, trains the acquired content into the trained model, and after learning, re-inputs the data to be processed into the trained model for processing, acquires the processed data that has been processed again by the trained model, and sends it to the external server.

Description

[Technical field]

[0001] The present invention relates to an assist server. [Background technology]

[0002] Recently, the technology known as AI (Artificial Intelligence) has been developing rapidly. Furthermore, AI, specifically server systems equipped with trained models and communication functions (hereinafter, these server systems are referred to as "AI servers"), have been made public and developers and others have provided an environment for free use.

[0003] One type of AI that has recently attracted attention is so-called chat-type AI, in which natural language data is input and a trained model outputs a dialogue in natural language that is valid for the input data (e.g., Non-Patent Document 1).

[0004] This chat-type AI learns from conversation data from various chat services available on the Internet and other platforms, and users can obtain the desired output through conversation with the chat-type AI.

[0005] Therefore, in order to obtain the desired output from a chat-type AI, in addition to the underlying circumstances, situation, or conditions, it is necessary to input information requesting the type of output desired.

[0006] For example, chat AI can be used to acquire an MBA (Master of Business Administration) If a student wants to obtain sample questions and model answers for the final exam in marketing theory at the graduate school where he or she is studying to obtain a bachelor's degree, the student must interactively input into his or her computer or other terminal the following background information: that he or she is aiming to obtain an MBA, that he or she is taking a marketing theory course at graduate school, the scope of the final exam, etc., along with a request to be "taught" the sample questions and model answers for the final exam.

[0007] However, this input requires a great deal of effort. Even if you take the trouble to input the dialogue, the chat AI cannot output the desired information about the items that the trained model of the chat AI has not yet learned, and the current situation is that it simply outputs that the information has not yet been learned.

[0008] This type of input labor and the need to deal with processing unlearned items arise in AI other than chat-type AI as well.

[0009] Therefore, there is a demand for a server intermediation system, intermediation server, assist system, and assist server that can reduce the labor required to input data into AI and that can process items that have not yet been learned. [Prior art documents] [Patent documents]

[0010] [Non-Patent Document 1] ChatGPT published at https: / / openai.com / Summary of the Invention [Problem to be solved by the invention]

[0011] The problem that this invention aims to solve is to provide a server intermediation system, an intermediation server, and an AI server assist system that can reduce the amount of work required to input data into AI and that can process items that have not yet been learned.

[0012] The contents of the above "Background Art" and "Problems to be Solved by the Invention" indicate the opportunity (trigger) that led to the invention, and do not limit the technical scope of the invention, nor do they permit a limited interpretation of the technical scope of the invention (see Heisei 17 (Gyo-Ke) No. 10042 and the Japan Patent Office Examination Guidelines as of the filing date, Part II, Chapter 2, Section 2, 3.2.1). [Means for solving the problem]

[0013] The present invention relates to a trained model that is trained to input processing target data received from an external server connected via a communication line, process the processing target data, and output processed data to be transmitted to the external server; a system that extracts keywords from the processing target data, transmits the extracted keywords to a search server having a search engine to perform a search, accesses a URL of an external site listed in a search result output by the search, acquires content of the external site, trains the acquired content into the trained model, and, after training, converts the processing target data into the trained model Invest in Input and process the trained model But and a control unit that acquires the processed data and transmits the processed data to the external server. Effect of the Invention

[0014] According to the present invention, it is possible to provide a server intermediation system, an intermediation server, a learned model assist system, and an assist server that can reduce the amount of work required to input data to be fed into AI and process items that have not yet been learned. [Brief description of the drawings]

[0015] [Figure 1] FIG. 2 is a diagram illustrating the configuration of an intermediation server and an assist server according to the first embodiment. [Diagram 2] FIG. 4 is a diagram illustrating an example of a data configuration of a user database. [Diagram 3] FIG. 2 is a diagram showing a data structure of a question database. [Figure 4] FIG. 13 is a diagram showing the data structure of an additional information database. [Diagram 5] FIG. 2 is a diagram showing a data structure of a format database. [Figure 6] FIG. 4 is a diagram showing a data structure of an account database. [Figure 7] FIG. 4 is a diagram showing the data structure of a session record database. [Figure 8] FIG. 13 is a diagram showing the configuration of information stored as an AI group. [Figure 9] 3 is a flowchart showing the operation of the server intermediation system, the intermediation server, and the AI ​​server assist system of the first embodiment. [Figure 10] FIG. 13 is a diagram showing an example of a reply / request input screen displayed on a user terminal. [Figure 11] FIG. 13 is a diagram showing the configuration of a mediation server realizing the server mediation system of the second embodiment, and a user terminal and an assist server that communicate with the mediation server. [Figure 12] FIG. 4 is a diagram showing a data structure of a site database. [Figure 13] 13 is a flowchart showing the operation of the server intermediation system, the intermediation server, and the AI ​​server assist system of the second embodiment. [Figure 14] FIG. 13 is a diagram showing configurations of an intermediation server, a user terminal, an assist server, and a portfolio server in the third embodiment. [Figure 15] FIG. 2 is a diagram illustrating an example of a data configuration of a portfolio database. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0016] The server intermediation system, the intermediation server, and the AI ​​server assist system according to one embodiment of the present invention will be described in detail below.

[0017] (First embodiment) In the first embodiment, the intermediary server 1 generates pre-transmission data by adding additional data necessary to obtain the desired output from the trained model held by the assist server 2 to the initial data received from the user terminal 3, further prepares data to be processed from the pre-transmission data in accordance with an input format specific to the trained model, instructs the assist server 2 to input the data to be processed into the trained model and transmits it, receives processed data from the assist server 2 that the trained model has processed and output from the trained model, prepares output data to be output to the user terminal 3 based on the output format specific to the processed data output by the trained model, and transmits the output data to the user terminal.

[0018] The intermediary server 1 mediates data exchange between the user terminal 3 and the assist server 2, and the assist server 2 assists the operation of the trained model.

[0019] FIG. 1 is a diagram showing the configuration of a mediation server 1 which realizes the server mediation system of this embodiment, and a user terminal 3 and an assist server 2 which communicate with the mediation server 1. As shown in FIG.

[0020] As shown in FIG. 1, the intermediary server 1 includes an intermediary server control unit 11 , an intermediary server storage unit 12 , and an intermediary server input / output unit and communication unit 13 .

[0021] The intermediate server control unit 11 includes a processor such as a CPU (Central Processing Unit). The intermediation server control unit 11 sequentially reads out and executes the software stored in the intermediation server storage unit 12.

[0022] The relay server storage unit 12 includes a memory, a solid state drive (SSD), a hard disk drive (HDD), The device is equipped with a storage device such as a USB 3.0 storage device (USB 3.0 storage device).

[0023] The intermediation server storage unit 12 stores the software executed by the intermediation server control unit 11, as well as a user DB 121 (hereinafter, database is referred to as DB), an additional information DB 122, and a format DB 123.

[0024] The user DB 121 stores user information, which is information on the attributes of the user, for each user, and round information, which indicates a history of a series of information exchanges with the assist server 2.

[0025] The additional information DB122 stores additional information such as prerequisites necessary to obtain the desired output from the trained model possessed by the assist server 2, and request content indicating what type of output is desired, which are added to the initial data, which is the data received by the intermediary server control unit 11 from the user terminal 3, as well as additional example sentences, which are example sentences set for each additional information.

[0026] The format DB123 stores, for each assist server 2, an input format, which is the format required for data to be input to a learned model specific to the learned model stored in the assist server 2, and an output format specific to the processed data output by the learned model.

[0027] The intermediation server input / output unit and communication unit 13 includes input / output devices such as a keyboard, mouse, microphone, and display, and a communication device that communicates with external devices via a network.

[0028] The assist server 2 includes an assist server control unit 21 , an assist server storage unit 22 , and an assist server input / output unit and communication unit 23 .

[0029] The assist server control unit 21 includes a calculation device such as a CPU, etc. The assist server control unit 21 sequentially reads out and executes the software stored in the assist server storage unit 22.

[0030] The assistant server storage unit 22 includes a storage device such as a memory, an SSD, or an HDD.

[0031] The assist server storage unit 22 stores therein the software executed by the assist server control unit 21, an account DB 221, a session record DB 222, and an AI group 223.

[0032] The account DB 221 stores an account, which is an identifier uniquely assigned to each user of the assist server 2, and various information related to each account.

[0033] The session record DB 222 stores session information indicating a history of a series of information exchanges with the proxy server 1, and chat information such as the contents of chats exchanged within each session.

[0034] The AI ​​group 223 stores information about each trained model and the trained model for each trained model. The AI ​​group 223 can also include a plurality of trained models and an AI operation unit that is software that determines how to select and use the plurality of trained models and sequentially combines and operates the selected trained models.

[0035] The assist server input / output unit and communication unit 23 includes input / output devices such as a keyboard, a mouse, a microphone, and a display, and a communication device that communicates with external devices via a network.

[0036] The user terminal 3 and the intermediation server 1, and the intermediation server 1 and the assist server 2 are connected via a public communication network 4 such as the Internet, and communicate with each other.

[0037] Fig. 2 is a diagram showing an example of the data configuration of the user DB 121. As shown in Fig. 2, the user DB 121 stores a user ID, which is an identifier uniquely assigned to each user, user information, which is information related to the user such as name and contact information, a round ID, which is an identifier uniquely assigned to each round, round information, such as the access date and time, a question ID identifying a question posed to the user, the content of the user's answer to the question, an assist server ID identifying the assist server 2 in which the round occurred, and a session ID identifying the session in which the round occurred.

[0038] Here, we will explain rounds and sessions. A session refers to a series of information exchanges from when a user starts exchanging information with the assist server 2 via the intermediary server 1 using the user terminal 3 to when the user ends the series of information exchanges with the assist server 2. Usually, one session occurs regarding one topic, but the topic may change during the session.

[0039] A round refers to one information exchange between the user terminal 3 and the assist server 2 via the intermediary server 1 that occurs during one session. One round begins with a flow of information in the sending direction sent from the user terminal 3 to the assist server 2 via the intermediary server 1, and ends with a flow of information in the receiving direction received from the assist server 2 via the intermediary server 1. Thus, one session may include multiple rounds.

[0040] 3 is a diagram showing the data structure of the question DB. The intermediation server storage unit 12 of the intermediation server 1 is an arbitrary database that may or may not store the question DB. Therefore, it is not shown in FIG.

[0041] The question DB stores a question ID, which is an identifier uniquely assigned to each question; question attribute information, which is information about the attributes of questions such as non-cognitive skill measurement type questions that ask the user about their words and actions when assuming a situation that would normally not occur to the user, and response ability measurement type questions that ask the user about their response when an unexpected situation is set for the user; and the question text.

[0042] Fig. 4 is a diagram showing the data configuration of the additional information DB 122. As shown in Fig. 4, the additional information DB 122 stores an additional example sentence ID, which is an identifier uniquely assigned to each example sentence that the intermediation server 1 adds to the initial data, additional information, and an additional example sentence.

[0043] Fig. 5 is a diagram showing a data configuration of the format DB 123. As shown in Fig. 5, the format DB 123 stores an assist server ID, which is an identifier assigned to each assist server 2, an input format, and an output format.

[0044] Fig. 6 is a diagram showing the data configuration of the account DB 221. As shown in Fig. 6, the account DB 221 stores an account, which is an identifier uniquely assigned to each user of the assist server 2, identification information for identifying the user, and account attribute information indicating attributes of the account, such as educational use by a school or the like, or commercial use by a company or the like.

[0045] Fig. 7 is a diagram showing a data configuration of the session record DB 222. As shown in Fig. 7, the session record DB 222 stores a session ID which is an identifier uniquely assigned to each session, an account which performed the session, a session start date and time, a session end date and time, a chat ID which is an identifier uniquely assigned to each chat (comment) which occurred in the session, the chat content of each chat, and a read flag which is turned ON when the chat has been read by the other party (meaning that the chat has been displayed on the user terminal 3 and that the chat has been input to the trained model on the assist server 2).

[0046] FIG. 8 is a diagram showing the configuration of information stored as the AI ​​group 223. As shown in FIG. 8, the AI ​​group 223 stores a trained model ID, which is an identifier uniquely assigned to each trained model as a collection of indexed data such as a database, input format information regarding the input format such as the data type (e.g., plain text, XML, binary data, etc.) and data length that the data input to the trained model must have, output format information regarding the output format such as the data type and data length of the data output by the trained model, trained model information indicating the attributes of the trained model such as image judgment, character judgment, natural language analysis and interpretation, translation, and dialogue, a specific functional description of the trained model, and the main body of the trained model. The AI ​​group 223 may also include an AI operation unit.

[0047] 9 is a flowchart showing the operation of the server intermediation system, intermediation server 1, and AI server assist system of this embodiment. The server intermediation system is realized by the cooperation of hardware and software equipped in the intermediation server 1. The AI ​​server assist system is realized by the cooperation of hardware and software equipped in the assist server 2, and cooperation with the server intermediation system enables advanced use of the trained model equipped in the assist server 2.

[0048] As shown in FIG. 9, in step 901, the control unit of the user terminal 3 transmits an instruction to the intermediation server 1 requesting a question after exchanging various information to start a session with the intermediation server 1.

[0049] In step 902, the intermediary server control unit 11 transmits a question read out from the question DB at random or by selecting from questions that have not yet been administered to the user.

[0050] In step 903, the control unit of the user terminal 3 transmits the answer input by the user and, if necessary, a request to the intermediation server 1.

[0051] Here, steps 901 to 903 can be replaced with an operation in which the intermediary server 1 simply receives data input from the user terminal 3 to start a dialogue, rather than requesting and sending a question.

[0052] In step 904, the mediation server control unit 11 selects an assist server 2 having a trained model to be used from one or more assist servers 2 based on the answer (or data for starting a dialogue; hereinafter, these are referred to as initial data) received from the user terminal 3 and the request optionally received. This selection may be made by the mediation server control unit 11 searching a database that stores information about the assist server 2 that the mediation server 1 stores separately in the mediation server storage unit 12.

[0053] In step 905, the intermediary server control unit 11 searches the user DB based on the user ID received from the user terminal 3, and reads out the assist server ID, session ID, and round information if they are stored. The intermediary server control unit 11 identifies which assist server 2, which session, and which round the user wishes to continue using the assist server ID, session ID, and round information. In this case, the intermediary server control unit 11 selects the identified assist server 2 instead of the assist server 2 selected in step 904.

[0054] In step 906, the intermediation server control unit 11 adds the read session ID to the initial data.

[0055] In step 907, the intermediary server control unit 11 generates pre-conversion data by adding additional data such as a request message to the initial data. This request message may be a message created by the user and sent from the user terminal 3, or may be an additional example message retrieved by the intermediary server control unit 11 from the additional information DB 122 based on additional information selected by the user via the user terminal 3.

[0056] Specific examples are given below. (See Figure 10.) (1) Questions posed to users Q1. Katsu Kaishu is on the Kanrin Maru, sailing to America. You are a passenger on the ship. What would you suggest to Katsu Kaishu that he should do when he gets to America? (2) User answers (initial data) A1. I suggest to Katsu Kaishu that he invite the President of the United States to dinner. Even though the President likes hamburgers, he should have dinner. This shows that they are on equal footing. (3) Request selected by the user (Request 3) I would like to be evaluated from an objective perspective In this case, the intermediation server control unit 11 generates the following pre-conversion data based on the above (1) to (3). (Data before conversion: The underlined part is additional data added by the intermediation server control unit 11 to the initial data.) I was asked the following questions: Q1. Katsu Kaishu is on the Kanrin Maru, sailing to America. You are a passenger on the ship. What would you suggest to Katsu Kaishu that he should do when he gets to America? To this question, I replied as follows: A1. I would suggest to Katsu Kaishu that he invite the President of the United States to dinner. Even though the President likes hamburgers, they should have dinner. To show that they are on equal footing. Please evaluate my answer to Q1, A1, from an objective point of view.

[0057] In step 908, the intermediary server control unit 11 searches the format DB 123 based on the assist server ID of the assist server 2 selected in steps 904 and 905, and reads out the input format. The intermediary server control unit 11 converts the pre-conversion data based on the read input format, and generates data to be processed.

[0058] In step 909 , the intermediate server control unit 11 transmits the processing target data to the selected assist server 2 .

[0059] In step 910, the assistant server control unit 21 inputs the received processing target data into the learned model, and causes the processing target data to be processed by the learned model.

[0060] In step 911, the assist server control unit 21 transmits to the intermediation server 1 the processed data that the learned model has output after processing the target data.

[0061] In step 912, the relay server control unit 11 searches the format DB 123 based on the assist server ID, reads out the output format, converts the processed data based on the read out output format, and generates response data to be displayed on the user terminal 3.

[0062] In step 913 , the intermediation server control unit 11 transmits the response data to the user terminal 3 .

[0063] In step 914 , the control unit of the user terminal 3 causes the display of the user terminal 3 to display the received response data.

[0064] Of the above operations, steps 901 to 908 are referred to as a round start operation, and steps 912 to 914 are referred to as a round end operation.

[0065] Fig. 10 is a diagram showing an example of the answer / request input screen 100 displayed on the user terminal 3. When transmitting a question in step 902 of Fig. 9, the intermediation server control unit 11 generates data for the answer / request input screen 100, transmits it to the user terminal 3, and causes it to be displayed.

[0066] As shown in FIG. 10, the answer / request input screen 100 comprises a question display field 1001 that displays the question, an answer input field 1002 where the user inputs the answer, a request selection field 1003 which is a pull-down menu for selecting a request, a request details field 1004 that is displayed by clicking on the request selection field 1003, and a send button 1005 that instructs the input / selected information to be sent to the intermediary server 1.

[0067] As described above, in the server intermediation system, intermediation server, and AI server assist system of this embodiment, the intermediation server 1 has an additional information DB 122 that stores additional information, which is data to be added to initial data transmitted from the user terminal 3 by a user, in order to obtain a desired output from a trained model held by the assist server 2, and a format DB 123 that stores, for each assist server 2, an input format, which is a format required for data to be input to the trained model specific to the trained model, and an output format specific to processed data output by the trained model, and reads out additional information selected from the additional information DB 122 and adds the additional information to the initial data as additional data. an intermediary server control unit 11 that adds the pre-conversion data to the input format to generate pre-conversion data, reads an input format from the format DB 123 based on the selected assist server 2, converts the pre-conversion data based on the input format to generate data to be processed, transmits the data to be processed to the assist server 2 and inputs it to the learned model for processing, receives processed data from the assist server 2 that is output by the learned model after processing the data to be processed, reads an output format from the format DB 123 based on the assist server 2, converts the processed data based on the output format to generate response data to be displayed on the user terminal 3, and transmits the response data to the user terminal 3.

[0068] Therefore, according to this embodiment, it is possible to provide a server intermediation system, an intermediation server, and an AI server assist system that can reduce the amount of work required to input data to be input into the AI.

[0069] Second embodiment In the second embodiment, the configuration and operation of the intermediation server 1 are similar to the configuration and operation of the intermediation server 1 in the first embodiment.

[0070] In contrast to this, the assist server 2 of the second embodiment further stores a site DB 224 in the assist server storage unit 22.

[0071] In the second embodiment, when the assist server control unit 21 outputs a response indicating that the learned model has not yet learned the input data to be processed, it reads the site URL (Uniform Resource Locator) of the search server 5 from the site DB 224, sends keywords related to the matter that has not been learned to the read site URL to perform a search, accesses the URL of the external site 6 listed in the search results output by the search to obtain the content of that site, learns the obtained content into the learned model, and after learning, re-inputs the data to be processed and processes it, thereby obtaining the processed data. This is in terms of the configuration and operation of the assist server 2 in the first embodiment.

[0072] FIG. 11 is a diagram showing the configuration of the intermediation server 1, the user terminal 3 that communicates with the intermediation server 1, and the assist server 2, which realize the server intermediation system of this embodiment.

[0073] The assist server 2 of this embodiment further stores a site DB 224 in the assist server storage unit 22. The assist server 2 is further connected to a search server 5 and an external site 6 via a public communication network 4 such as the Internet, and communicates with each other.

[0074] Here, the search server 5 has a search engine that, when a keyword is input, outputs a list of external sites 6 related to the keyword and the URL. Also, when the URL is accessed, the external site 6 transmits the content to the access source.

[0075] Fig. 12 is a diagram showing the data configuration of the site DB 224. As shown in Fig. 12, the site DB 224 stores a site ID, which is an identifier uniquely assigned to a site, site attribute information such as a specialty field in the case where a search engine has a unique field, and a site URL indicating the URL of the site.

[0076] 13 is a flowchart showing the operation of the server intermediation system, intermediation server 1, and AI server assist system of this embodiment. The server intermediation system is realized by the cooperation of hardware and software equipped in the intermediation server 1. The AI ​​server assist system is realized by the cooperation of hardware and software equipped in the assist server 2, and by cooperation with the server intermediation system, further advanced use of the trained model equipped in the assist server 2 is possible.

[0077] Specifically, in the prior art, trained models had difficulty responding to matters that had not yet been learned, but according to this embodiment, the assist server 2 uses a search server 5 having a search engine to find the URL of an external site 6 that describes an item that has not yet been learned by the trained model, accesses the external site 6 to obtain content related to that item, and has the trained model learn the obtained content, thereby assisting in further learning of the trained model and achieving more advanced use of the trained model.

[0078] In this embodiment, it is possible to use the content of all sites around the world that can be accessed via a public communication network 4 such as the Internet as the background for learning the trained model, making it possible to further improve the performance of the trained model. It is also expected to eliminate the time-consuming and laborious tasks of conventionally entering keywords and checking search results.

[0079] As shown in FIG. 13, in step 1301, the control unit of the user terminal 3 and the proxy server control unit 11 execute a round starting operation.

[0080] In step 1302 , the intermediate server control unit 11 transmits the processing target data to the assist server 2 .

[0081] In step 1303, the assist server control unit 21 inputs the processing target data received from the intermediation server 1 into the learned model.

[0082] In step 1304, the assist server control unit 21 acquires the processed data that the learned model has output after processing the target data.

[0083] In step 1305, the assist server control unit 21 judges whether the processed data includes content indicating that the data has not been learned. If the assist server control unit 21 judges that the processed data does not include content indicating that the data has not been learned (N in step 1305), the process proceeds to step 1313, and if the assist server control unit 21 judges that the processed data includes content indicating that the data has not been learned (Y in step 1305), the process proceeds to step 1306.

[0084] Here, an example of a method for the assistant server 2 to determine whether or not the processed data includes a content indicating that the data has not been learned (hereinafter, referred to as an "unlearned response") will be described.

[0085] Example 1 of judgment method (when the trained model is a generative AI that generates text) In the case where the generation AI that generates text using the assist server control unit 21 as a trained model is equipped with a trained model, the text data of the following content is added in advance to the additional example sentences stored in the additional information DB 122.

[0086] [Example of additional text data] If there is any information contained in my question or necessary information to answer my question that you have not yet studied, please respond as follows and clearly indicate the information that you have not studied. Also, after [Contents of unstudied information (JSON)], please list the information that you have not yet studied in bullet points in JSON format. "[Unlearned information exists] There is some information in your question, or in the information needed to answer your question, that I have not yet studied. The information I have not studied is as follows: [Contents of unlearned information] (Write the numbering here.) (Write the content of your study information here.) [Contents of unlearned information (JSON)] (Write the contents of your learning information here in JSON format.)

[0087] If a request sentence such as the above is added to the additional example sentence in advance, the assist server 2 can easily extract (1) an "unlearned response," which is a response indicating that the trained model held by the assist server 2 has not been trained, and (2) the content of the items that have not been trained, from the processed data.

[0088] Example 2 of judgment method (when the trained model is not a generative AI that generates text) In this case, the model is trained as follows. (1) The system is trained in advance to output the probability of correct answers for the processed data itself. In the conventional generation of trained models, for example, when determining whether input image data is an image of a "cat," the probability of correct answer is output, and a model with a high probability of correct answer is adopted as the trained model. Using such a known method, the model is first made to output the probability of correct answer. (2) The system is trained in advance to explain each noun contained in the data to be processed and output the probability of correct answer. A single "noun" is used as the explanatory variable, and an explanation for that "noun" and the probability of that explanation being correct are used as the objective variables to train the model using the training data.

[0089] The trained model is further trained as described above, or a model trained separately as described above is made to process the processed data of the trained model as input data, and the output data includes the probability of correct answer for the processed data itself and the probability of correct answer for the explanation of the noun included in the data to be processed.

[0090] Therefore, when the probability of correct answer of the processed data itself is equal to or lower than a first threshold, the assist server 2 (1) can extract a response indicating that the trained model possessed by the assist server 2 has not been trained, and further, when the probability of correct answer of the explanation of the noun contained in the data to be processed is equal to or lower than a second threshold, (2) can easily extract the content of the untrained item from the processed data.

[0091] Example 3 of the judgment method (Another example of a case where the trained model is not a generative AI that generates text.) In the field of AI, it is common to extract words from sentences and assign them meaning.

[0092] For example, in JP 2022-47665 A (hereinafter referred to as Reference 1), as shown in Fig. 8, the control unit 101 tokenizes the input text using a tokenizer (converts it into a string of words in the smallest unit) (paragraph

[0081] ), vectorizes it, then converts it into a tensor (paragraphs

[0082] to

[0084] ), and extracts features. The control unit 101 then standardizes and scales the features, and outputs the obtained data as training data.

[0093] First, the model is trained using the above training data. Next, the model is additionally trained using training data in which tens to hundreds of words not included in the above training data are prepared, and these words are used as explanatory variables, and a word indicating "unlearned" is used as a target variable (for example, "This word has not been learned yet," or simply "unlearned")

[0094] Such techniques are common technical knowledge commonly used in the field of AI (see, for example, Patent Document 1 (JP 2018-169795) described in Reference 1), and a person skilled in the art can easily construct such a mechanism.

[0095] The trained model obtained after this two-stage training includes "unlearned responses" in the processed data. Therefore, the assist server 2 can easily extract the contents of the unlearned items from the processed data by comparing the specific text of the "unlearned responses" stored in advance in the assist server storage unit 22 with the trained model and the processed data.

[0096] In step 1306, the assist server control unit 21 extracts unlearned items, particularly nouns, from the processed data as keywords, and retrieves the URL of the search server 5 having a search engine that specializes in the field of the keyword, or a search engine that can search for general items if no such search engine is found, from the site DB 224. The assist server control unit 21 then accesses the retrieved search server 5 and transmits / inputs the keyword to the search engine as search information.

[0097] In step 1307, the search server 5 transmits to the assist server control unit 21 the search results including the URLs of the external sites 6 where information related to the keyword is described.

[0098] In step 1308 , the assist server control unit 21 reads out the URL of the external site 6 from the search result, and accesses the external site 6 .

[0099] In step 1309 , the control unit of the external site 6 transmits the content to the assist server 2 .

[0100] In step 1310, the assistant server control unit 21 generates training data with the data to be processed as an explanatory variable and the content as a target variable, and causes the trained model to learn the generated training data.

[0101] In step 1311, the assistant server control unit 21 re-enters the processing target data into the trained model.

[0102] In step 1312, the assist server control unit 21 reacquires the processed data that the learned model has output after processing the target data, and transmits it to the intermediation server 1.

[0103] In step 1313, the control unit of the user terminal 3 and the mediation server control unit 11 execute a round ending operation.

[0104] As described above, in the server intermediation system, intermediation server, and AI server assist system of this embodiment, the assist server 2 further stores a site DB 224 in the assist server memory unit 22, and when the assist server control unit 21 outputs a response indicating that the learned model has not yet learned the input data to be processed, it reads out the site URL of the search server 5 from the site DB 224, sends keywords related to matters that have not been learned to the read site URL to perform a search, accesses the URL of the external site 6 listed in the search results output by the search to obtain the content of the external site 6, trains the obtained content into the learned model, and after learning, re-inputs the data to be processed and processes it, thereby obtaining the processed data.

[0105] Therefore, the assist server 2 can use a search server 5 having a search engine to find the URL of an external site 6 that describes an item that has not yet been learned in the learned model, access the external site 6 to obtain content related to that item, have the learned model learn the obtained content, and after learning, re-input the data to be processed and process it, thereby obtaining the processed data.

[0106] (Third embodiment) In the third embodiment, the configuration and operation of the assist server 2 are similar to those of the assist server 2 in the second embodiment.

[0107] In contrast, the intermediary server 1 in the third embodiment is connected to the portfolio server 8 via the public communication line network 4, and the intermediary server control unit 11 acquires from the portfolio server 8 a portfolio that the user has previously stored in the portfolio DB 821 of the portfolio server 8, and when generating pre-conversion data, adds the acquired portfolio together with additional data to the initial data, which differs from the operation of the intermediary server 1 in the first and second embodiments.

[0108] Below, first the configuration of the portfolio server 8 will be described, and then the operation of the intermediation server control unit 11 to acquire a portfolio from the portfolio DB 821 of the portfolio server 8 will be described.

[0109] FIG. 14 is a diagram showing the configuration of the mediation server 1 which realizes the server mediation system of this embodiment, and the user terminal 3, the assist server 2, and the portfolio server 8 which communicate with the mediation server 1.

[0110] The configurations of the intermediation server 1 and the assist server 2 have already been explained, so here, the configuration of the portfolio server 8 will be explained.

[0111] As shown in FIG. 14, the portfolio server 8 includes a portfolio server control unit 81 , a portfolio server storage unit 82 , and a portfolio server input / output unit and communication unit 83 .

[0112] The portfolio server control unit 81 includes an arithmetic unit such as a CPU, etc. The portfolio server control unit 81 sequentially reads out and executes the software stored in the portfolio server storage unit 82.

[0113] The portfolio server storage unit 82 includes a storage device such as a memory, an SSD, or an HDD.

[0114] The portfolio server storage unit 82 stores the software executed by the portfolio server control unit 81 as well as a portfolio DB 821 .

[0115] The portfolio server input / output unit and communication unit 83 includes input / output devices such as a keyboard, a mouse, a microphone, and a display, and a communication device that communicates with external devices via a network.

[0116] Fig. 15 is a diagram showing an example of the data configuration of the portfolio DB 821. As shown in Fig. 15, the portfolio DB 821 stores, for each user ID, a portfolio ID that is an identifier uniquely assigned to the portfolio, and portfolio data that is the main body of the portfolio.

[0117] Next, the operation of the mediation server control unit 11 to acquire a portfolio from the portfolio DB 821 of the portfolio server 8 will be described, focusing only on the parts that differ from the operation of the mediation server control unit 11 in the first embodiment.

[0118] The operation of step 906 in the first embodiment is changed as follows. "In step 906, the mediation server control unit 11 transmits an instruction to the portfolio server 8 to read out the portfolio data from the portfolio DB 821 based on the user ID received from the user terminal 3 and transmit the read portfolio data to the mediation server 1. The mediation server control unit 11 adds the session ID and the portfolio data received from the portfolio server 8 to the initial data."

[0119] As described above, in the server intermediation system, intermediation server, and AI server assist system of this embodiment, the intermediation server 1 is connected to the portfolio server 8 via the public communication line network 4, and the intermediation server control unit 11 acquires from the portfolio server 8 the portfolio that the user has previously stored in the portfolio DB 821 of the portfolio server 8, and adds the acquired portfolio together with the additional data to the initial data when generating pre-conversion data.

[0120] Now, I will explain about portfolios. In recent years, in entrance examinations such as university entrance examinations, or in personnel evaluations or personnel selection at companies, attempts have been made to include not only academic tests but also portfolios that compile information on activities such as participation in volunteer activities, awards received in sports, and overseas travel experiences, in addition to various activities at school during student days, in order to include the portfolio in the evaluation.

[0121] A portfolio can be created using spreadsheet or word processing software, but recently, the applicant and several other companies have been offering cloud-based portfolio creation systems that not only reduce labor costs but also enable users to inspire each other through contact with other users.

[0122] This portfolio then reflects the user's values, interests, and even a part of their personality.

[0123] Therefore, according to this embodiment, the intermediary server 1 adds the user's portfolio data to the initial data to generate pre-conversion data, converts the format of this pre-conversion data to generate data to be processed, and inputs this data to be processed into the trained model, thereby having the effect of outputting processed data that is more suited to the user's personality than if the portfolio data is not added. [Explanation of symbols]

[0124] 1. Intermediary Server 2 Assist Server 3. User terminal 4 Public Telecommunications Network 5 Search Server 6. External Sites 8 Portfolio Server 11. Intermediation server control section 12. Intermediary server storage unit 13 Intermediary server input / output section / communication section 21 Assist server control unit 22 Assist server memory unit 23 Assist server input / output section / communication section 81 Portfolio Server Control Unit 82 Portfolio Server Storage Unit 83 Portfolio Server Input / Output Section / Communication Section 100 Answer / Request Input Screen 223 AI group 1001 Question display field 1002 Answer input field 1003 Request Selection Column 1004 Request details 1005 Send button

Claims

1. A trained model that is trained to input processing target data received from an external server connected via a communication line, process the processing target data, and output processed data to be transmitted to the external server; a control unit that extracts keywords from the data to be processed, transmits the extracted keywords to a search server having a search engine to perform a search, accesses a URL of an external site listed in a search result output by the search to acquire content of the external site, trains the acquired content into the trained model, inputs the data to be processed into the trained model after training and processes it, acquires the processed data processed by the trained model, and transmits the processed data to the external server; An assist server comprising:

2. A plurality of the trained models are provided, The control unit is The assist server according to claim 1, further comprising an AI operation unit that determines how to select and use the plurality of trained models and sequentially combines and operates the selected trained models.

3. The trained model is The assist server according to claim 1 or claim 2, wherein the trained model is trained to input the data to be processed, process the data to be processed, and output the processed data such that a dialogue can be established for the data to be processed.

4. The external server An additional information database that stores additional information that is data to be added to initial data transmitted by a user from a user terminal in order to obtain a desired output from the trained model possessed by the assist server; A format database that stores, for each assistant server, an input format that is a format required for data to be input to the trained model that is specific to the trained model, and an output format that is specific to the processed data output by the trained model; an intermediary server control unit that reads the selected additional information from the additional information database, generates pre-conversion data by adding the additional information to the initial data as additional data, reads the input format from the format database based on the selected assist server, converts the pre-conversion data based on the input format to generate the data to be processed, transmits the data to be processed to the assist server and inputs it to the trained model for processing, receives from the assist server the processed data that the trained model has processed and output from the trained model, reads the output format from the format database based on the assist server, converts the processed data based on the output format to generate response data to be displayed on the user terminal, and transmits the response data to the user terminal; The assist server according to claim 3, which is an intermediary server comprising:

Images