Artificial intelligence system and program for artificial intelligence system

Abstract

To provide an artificial intelligence system that avoids previously learned artificial intelligence means from remembering character information having a word attribute capable of identifying an individual person.SOLUTION: An artificial intelligence system 100 includes a user terminal 110, a first server 120 and a second server 130. When data DT including a character having a word attribute of a preset masking is input in a browser 111 of the user terminal and is transmitted to the first server, a bot 121 converts the character having the word attribute of the masking into another character having the same word attribute and transmits the converted data to the second server, and the second sever uses previously learned artificial intelligence means 131 to execute text data generation of a content based on the converted data to transmit the generated text data to the first server, and the bot executes a change of returning another character part converted upon transmission to the second server to the character before conversion to transmit changed generation text data GT1 to the user terminal.SELECTED DRAWING: Figure 1

Description

[Technical field]

[0001] The present invention relates to an artificial intelligence system comprising a user terminal, a first server having a bot, and a second server having a pre-trained artificial intelligence means, in which the user terminal displays or outputs audio based on the first server, and a program for the system. [Background technology]

[0002] 2. Description of the Related Art Conventionally, there is known an information processing device that functions as a summary generating device having a summary generating unit that accesses a large-scale language model (for example, see Patent Document 1). [Prior art documents] [Patent documents]

[0003] [Patent Document 1] Patent Publication No. 2023-73095 Summary of the Invention [Problem to be solved by the invention]

[0004] However, the above-mentioned conventional information processing devices were configured to simply send the data to be summarized to GPT-3, which is an example of a large-scale language model. Therefore, if the data to be summarized contained character information with identifiable word attributes, there was a risk that GPT-3 would memorize the character information with word attributes that could identify an individual, resulting in the leakage of personally identifiable information to other devices via GPT-3.

[0005] Therefore, the present invention solves the problems of the prior art as described above, that is, the object of the present invention is to provide an artificial intelligence system and a program thereof that can prevent a pre-trained artificial intelligence means, which is a large-scale language model, from memorizing character information of word attributes that can identify an individual. [Means for solving the problem]

[0006] The invention according to claim 1 is an artificial intelligence system comprising a user terminal, a first server having a bot, and a second server having a pre-trained artificial intelligence means, in which the user terminal displays or outputs audio based on the first server, the user terminal communicates with the second server via the bot of the first server, and on an administration screen of an administrator terminal capable of communicating with the first server, masking of word attributes contained in data sent from the user terminal to the bot of the first server can be set freely, and when data including characters of a word attribute of a masking preset in the browser of the user terminal is input and sent to the first server, the bot of the first server that receives the data determines that the received data includes characters of a word attribute of the masking. The above-mentioned problem is solved by the configuration in which, in the received data, characters with a masking word attribute are converted into other arbitrary characters with the same word attribute, and the converted data is then transmitted to a second server, the second server uses the pre-trained artificial intelligence means of the second server to generate text data based on the converted data, and transmits the generated text data obtained by this generation to the first server, the bot of the first server changes the other arbitrary characters that were converted when transmitting the received generated text data to the second server back to the characters before conversion, and transmits the changed generated text data to a user terminal, and the user terminal displays the received changed generated text data in a browser or outputs it as audio.

[0007] The invention of claim 2 further solves the above-mentioned problem by, in addition to the configuration of the artificial intelligence system described in claim 1, configuring the system so that prohibited words can be freely set on the management screen of an administrator terminal capable of communicating with the first server, and when data containing pre-set prohibited words is entered into the browser of the user terminal and sent to the first server, the bot of the first server determines that the received data contains prohibited words, does not send the received data to the second server, and sends information to the user terminal that the data contains unavailable words.

[0008] The invention of claim 3 further solves the above-mentioned problem by being configured in such a way that, in addition to the configuration of the artificial intelligence system described in claim 2, the exclusion of inappropriate remarks can be freely set on the management screen of an administrator terminal capable of communicating with the first server, and when data including a preset inappropriate remark is input in the browser of the user terminal and sent to the first server, the bot of the first server determines that the received data includes an inappropriate remark, does not send the received data to the second server, and sends information to the user terminal indicating that the inappropriate remark is included, and on the other hand, the second server uses the pre-trained artificial intelligence means of the second server to generate text data based on the instruction command information included in the received data, and sends the generated text data obtained by execution to the first server, and the bot of the first server determines whether the received generated text data includes an inappropriate remark, and if it determines that the received generated text data includes an inappropriate remark, it does not send the received generated text data to the user terminal, and instructs the pre-trained artificial intelligence means of the second server to generate text data based on the instruction command information again without using the inappropriate remark.

[0009] The invention of claim 4 further solves the above-mentioned problem by being configured in addition to the configuration of the artificial intelligence system described in any one of claims 1 to 3, such that the user terminal is free to specify a main topic for the bot of the first server, and when instruction command information as data is input by the user in the browser while specifying the main topic and sent to the first server, the bot of the first server adds main topic information to the received instruction command information and sends it to the second server, and the second server generates text data based on the instruction command information and the main topic information using the pre-trained artificial intelligence means of the second server, and sends the generated text data obtained by the execution to the first server.

[0010] The invention of claim 5 further solves the above-mentioned problem by, in addition to the configuration of the artificial intelligence system described in claim 4, being configured so that on the management screen of an administrator terminal capable of communicating with the first server, users who can use the bot of the first server in the browser of a user terminal can be freely set based on user identification information.

[0011] The invention according to claim 6 is a program for an artificial intelligence system in which a user terminal displays or outputs sound based on a first server, comprising: a masking setting step in which the first server sets masking when there is a masking setting for a word attribute included in data from the user terminal to a bot of the first server on an administration screen of an administrator terminal capable of communicating with the first server; a masking target determination step in which, when data is input to a browser of the user terminal and transmitted to the first server, the bot of the first server receives the data and determines whether the received data contains a character with the word attribute of the masking; a masking conversion execution step in which, when the bot of the first server determines that the received data contains a character with the word attribute of the masking, the character with the word attribute of the masking is converted into another arbitrary character with the same word attribute in the received data; and The above-mentioned problem is solved by having a masking data transmitting step in which a bot of the server transmits the converted data to a second server, a data generating step in which the second server uses a pre-trained artificial intelligence means of the second server to generate text data based on the converted data, a generated data transmitting step in which the second server transmits the generated text data obtained by the execution to the first server, an inverse conversion step in which the bot of the first server changes a part of the received generated text data, which was converted when transmitting it to the second server, back to the character before conversion, an inverse conversion data transmitting step in which the bot of the first server transmits the changed generated text data to a user terminal, and a display / audio output step in which the user terminal displays the received changed generated text data in a browser or outputs it as audio. Effect of the Invention

[0012] The artificial intelligence system of the present invention comprises a user terminal, a first server having a bot, and a second server having a pre-trained artificial intelligence means, so that a user using the user terminal can not only interact with the pre-trained artificial intelligence means via the bot, but also achieve the following unique effects.

[0013] According to the artificial intelligence system of the invention of claim 1, for example, by setting in advance masking of word attributes that can identify individuals, such as name, address, telephone number, and email address, for data from a user terminal to a bot on the first server, if the data entered by the user contains characters with these word attributes, the bot on the first server converts the characters with these word attributes into any other characters with the same word attributes and then transmits the converted data to the second server, and text data generation is performed based on the converted data, thereby preventing the chat GPT, which is an example of a pre-trained artificial intelligence means of the second server, from memorizing information about characters with word attributes that can identify individuals. In other words, since the content sent to Chat GPT, which is an example of the second server's pre-trained artificial intelligence means, is converted into characters that cannot identify individuals, it is possible to avoid the leakage of personally identifiable information to other terminals via Chat GPT, which is an example of the second server's pre-trained artificial intelligence means. Furthermore, when the bot of the first server receives generated text data from the second server, the part of the received generated text data that was converted into another arbitrary character when it was sent to the second server is modified to return it to the character before conversion, and the modified generated text data is sent to the user terminal, so that modified generated text data that basically follows the intention of the instruction command information, which is the prompt included in the data entered by the user, can be obtained. In other words, the characters converted when the bot on the first server sends to the second server are converted back to the original characters when the bot on the first server receives from the second server, thereby preventing unwanted information leakage to the second server or to other terminals via the second server, and ensuring consistency between the instructions and commands from the user terminal and the responses to the user terminal.

[0014] According to the artificial intelligence system of the invention of claim 2, in addition to the effects of the invention of claim 1, for example, by presetting words such as "confidential," "personal information," and "My Number" as prohibited words, if the data entered by the user contains prohibited words, the bot of the first server detects the prohibited words and stops communication with Chat GPT, an example of the pre-trained artificial intelligence means of the second server, and information indicating that the data contains unusable words is sent to the user terminal, thereby prohibiting undesirable use by the user. In other words, since the transmission of information to chat GPT, which is an example of the second server's pre-trained artificial intelligence means, is restricted, it is possible to avoid information leakage to other terminals via chat GPT, which is an example of the second server's pre-trained artificial intelligence means.

[0015] According to the artificial intelligence system of the invention of claim 3, in addition to the effect achieved by the invention of claim 2, if an inappropriate remark is included in an exchange between a user terminal and a pre-trained artificial intelligence means, the bot of the first server detects the inappropriate remark and blocks it, and notifies the sender that the remark contains an inappropriate remark, thereby making it possible to eliminate the exchange containing the inappropriate remark. In other words, if a user uses an inappropriate remark, the content of that remark is not sent to Chat GPT, which is an example of a pre-trained artificial intelligence means of the second server, thereby avoiding the risk of Chat GPT, an example of a pre-trained artificial intelligence means, learning the inappropriate remark. Furthermore, if the user uses an inappropriate remark, the bot in the first server returns a reply to the user to that effect, allowing the user to recognize that he or she has used an inappropriate remark. In addition, if inappropriate remarks are included in the generated text data generated by Chat GPT, an example of a pre-trained artificial intelligence means, the content of the data will not be transmitted to the user's terminal, thereby avoiding the risk that a user will use the generated text data containing inappropriate remarks generated by Chat GPT, an example of a pre-trained artificial intelligence means, as is. Furthermore, if the generated text data contains inappropriate remarks, the bot on the first server sends an instruction command to Chat GPT, an example of a pre-trained artificial intelligence means, to generate text data again based on the instruction command information without using the inappropriate remarks, so that generated text data that does not contain inappropriate remarks can be obtained and displayed on the user terminal.

[0016] According to the artificial intelligence system of the invention of claim 4, in addition to the effects achieved by the invention of any one of claims 1 to 3, when a main topic is specified in the browser, main topic information is always included in the information from the bot to the pre-trained artificial intelligence means in the interaction between the user and the pre-trained artificial intelligence means, and the generated text data generated by the pre-trained artificial intelligence means has content that is in line with the main topic, so that the content of consultations, etc., can be consistent (the content does not vary) in the interaction between the user and the pre-trained artificial intelligence means. In other words, even if a user asks a question multiple times, the answers provided will be in line with the main topic, preventing the user from gradually deviating from the main topic as the user continues to ask questions.

[0017] According to the artificial intelligence system of the invention of claim 5, in addition to the effects achieved by the invention of claim 4, users are identified based on user identification information, and users who can interact with Chat GPT, which is an example of a pre-trained artificial intelligence means of the second server, are specified and restricted via a bot on the first server, so that administrators can easily set usage restrictions and permissions for Chat GPT for each employee member.

[0018] According to the program of the artificial intelligence system of the invention of claim 6, similar to the effect achieved by the invention of claim 1, for example, by setting in advance masking of word attributes that can identify individuals, such as name, address, telephone number, and email address, for data from a user terminal to a bot on the first server, if characters of those word attributes are included in the data entered by the user, the bot on the first server converts those characters of those word attributes into other arbitrary characters of the same word attribute and then transmits the converted data to the second server, and text data generation is performed based on the converted data, thereby preventing the chat GPT, which is an example of a pre-trained artificial intelligence means of the second server, from memorizing information about characters of word attributes that can identify individuals. In other words, since the content sent to Chat GPT, which is an example of the second server's pre-trained artificial intelligence means, is converted into characters that cannot identify individuals, it is possible to avoid the leakage of personally identifiable information to other terminals via Chat GPT, which is an example of the second server's pre-trained artificial intelligence means. Furthermore, when the bot of the first server receives generated text data from the second server, the part of the received generated text data that was converted into another arbitrary character when it was sent to the second server is modified to return it to the character before conversion, and the modified generated text data is sent to the user terminal, so that modified generated text data that basically follows the intention of the instruction command information, which is the prompt included in the data entered by the user, can be obtained. In other words, the characters converted when the bot on the first server sends to the second server are converted back to the original characters when the bot on the first server receives from the second server, thereby preventing unwanted information leakage to the second server or to other terminals via the second server, and ensuring consistency between the instructions and commands from the user terminal and the responses to the user terminal. [Brief description of the drawings]

[0019] [Figure 1] FIG. 1 is a diagram showing the concept of an artificial intelligence system according to a first embodiment of the present invention. [Diagram 2] FIG. 2 is a chart showing an example of the operation of the artificial intelligence system according to the first embodiment of the present invention. [Diagram 3] 1A and 1B are diagrams showing an example of masking settings on the management screen of the artificial intelligence system according to the first embodiment of the present invention, and a diagram showing what happens when characters to be masked are entered on the artificial intelligence service screen displayed on the browser of a user terminal. [Figure 4] FIG. 2 is a diagram for explaining an example of the masking operation of the artificial intelligence system according to the first embodiment of the present invention. [Diagram 5] (A) and (B) are figures showing the state when voice input is selected on the artificial intelligence service screen displayed on the browser of a user terminal of the artificial intelligence system which is the first embodiment of the present invention, and figures showing the state when voice is input on the artificial intelligence service screen displayed on the browser of the user terminal. [Figure 6] (A) and (B) are figures showing an example of setting prohibited words on the management screen of the artificial intelligence system which is the first embodiment of the present invention, and a figure showing what happens when a prohibited word is entered on the artificial intelligence service screen displayed on the browser of a user terminal. [Figure 7] 1A and 1B are diagrams showing an example of settings for eliminating inappropriate comments on the administration screen of the artificial intelligence system according to the first embodiment of the present invention, and a diagram showing an example of interaction on the artificial intelligence service screen displayed on the browser of a user terminal. [Figure 8](A) and (B) are explanatory diagrams of the operation of the artificial intelligence system according to the first embodiment of the present invention when an inappropriate remark is included in a user input, and an explanatory diagram of the operation when an inappropriate remark is included in text data generated by Chat GPT, an example of a pre-trained artificial intelligence means. [Figure 9] FIG. 2 shows a state in which a main topic is specified on an artificial intelligence service screen displayed on a browser of a user terminal of the artificial intelligence system according to the first embodiment of the present invention. [Figure 10] FIG. 2 is a diagram for explaining the operation of the main topic designation function of the artificial intelligence system according to the first embodiment of the present invention. [Figure 11] FIG. 2 is a diagram showing an example of settings for users who can use a bot on the management screen of the artificial intelligence system according to the first embodiment of the present invention. [Figure 12] FIG. 2 is a diagram showing a log displayed on the management screen of the artificial intelligence system according to the first embodiment of the present invention. [Figure 13] 13(A) and (B) are diagrams showing an example of input for a sending message window, which is a message sending and receiving window displayed by a browser on a user terminal of an artificial intelligence system according to a second embodiment of the present invention, and a diagram showing an example of a display for a receiving message window, which is a message sending and receiving window. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0020] The artificial intelligence system of the present invention comprises a user terminal, a first server having a bot, and a second server having pre-trained artificial intelligence means, the user terminal communicates with the second server via the bot of the first server, and on an administration screen of an administrator terminal capable of communicating with the first server, masking of word attributes included in data sent from the user terminal to the bot of the first server can be freely set, and when data including characters of a masking word attribute preset in the browser of the user terminal is input and sent to the first server, the bot of the first server that receives the data determines that the received data includes characters of the masking word attribute, converts the characters of the masking word attribute in the received data to other arbitrary characters of the same word attribute, and then sends the converted data to the second server. the second server uses the pre-trained artificial intelligence means of the second server to generate text data based on the converted data, and transmits the generated text data obtained by this execution to the first server; the bot of the first server modifies any other arbitrary character part of the received generated text data that was converted when transmitted to the second server back to the character before conversion, and then transmits the modified generated text data to the user terminal; and the user terminal displays the received modified generated text data in a browser or outputs it as audio. As long as this configuration can prevent the chat GPT, which is an example of the pre-trained artificial intelligence means of the second server, from memorizing information about characters with word attributes that can identify an individual, any specific implementation may be adopted. The program of the artificial intelligence system of the present invention includes a masking setting step in which the first server sets masking when there is a masking setting for a word attribute included in data from a user terminal to a bot of the first server on an administration screen of an administrator terminal capable of communicating with the first server; a masking target determination step in which, when data is input to a browser of the user terminal and transmitted to the first server, the bot of the first server receives the data and determines whether the received data contains a character of the masking word attribute; a masking conversion execution step in which, when the bot of the first server determines that the data contains a character of the masking word attribute, the character of the masking word attribute in the received data is converted to another arbitrary character of the same word attribute; a masking data transmission step in which the bot of the first server transmits the converted data to the second server; and a masking data transmission step in which the bot of the first server transmits the converted data to the second server using the pre-learned data of the second server. The specific implementation of the method includes a data generation step of using an artificial intelligence means to generate text data based on the converted data, a generated data transmission step in which the second server transmits the generated text data obtained by the execution to the first server, a reverse conversion step in which the bot of the first server changes a portion of the received generated text data, which was converted when transmitting it to the second server, back to the characters before conversion, an reverse conversion data transmission step in which the bot of the first server transmits the changed generated text data to a user terminal, and a display / audio output step in which the user terminal displays the received changed generated text data in a browser or outputs it as audio, so long as the method can prevent the chat GPT, which is an example of a pre-trained artificial intelligence means of the second server, from memorizing information about characters with word attributes that can identify an individual.

[0021] For example, the user terminal may be a desktop personal computer terminal, a notebook personal computer terminal, a smartphone terminal, a tablet terminal, or the like that sends and receives information, and may be any terminal that can be freely connected to a server via a communication network including a wide area network, such as the Internet, a local network, or a telephone line. Furthermore, the first server and the second server may each be a single server or multiple servers on a cloud. The pre-trained artificial intelligence means may be composed of an interactive model, also known as a large-scale language model, such as ChatGPT (Generative Pre-trained Transformer) (hereinafter referred to as ChatGPT), and may be configured to be installed on one server or multiple servers on the cloud. EXAMPLES

[0022] An artificial intelligence system 100 according to a first embodiment of the present invention will be described below with reference to FIGS. FIG. 1 is a diagram showing the concept of an artificial intelligence system 100 according to a first embodiment of the present invention, FIG. 2 is a chart showing an example of the operation of the artificial intelligence system 100 according to the first embodiment of the present invention, FIG. 3(A) is a diagram showing an example of masking settings on a management screen 113 of the artificial intelligence system 100 according to the first embodiment of the present invention, FIG. 3(B) is a diagram showing a state when a character to be masked is input on an artificial intelligence service screen 112 displayed by a browser 111 of a user terminal 110, and FIG. 4 is a diagram showing a masking setting of the artificial intelligence system 100 according to the first embodiment of the present invention. 5(A) is a diagram showing the state when voice input is selected on the artificial intelligence service screen displayed by the browser of the user terminal of the artificial intelligence system according to the first embodiment of the present invention, FIG. 5(B) is a diagram showing the state when voice is input on the artificial intelligence service screen displayed by the browser of the user terminal, FIG. 6(A) is a diagram showing an example of setting prohibited words PB on the management screen 113 of the artificial intelligence system 100 according to the first embodiment of the present invention, and FIG. 6(B) is a diagram showing the state when voice input is selected on the artificial intelligence service screen displayed by the browser of the user terminal, FIG. 7(A) is a diagram showing a state when a prohibited word PB is input, FIG. 7(A) is a diagram showing an example of the setting for excluding inappropriate remarks on the management screen 113 of the artificial intelligence system 100 which is the first embodiment of the present invention, FIG. 7(B) is a diagram showing an example of an exchange on the artificial intelligence service screen 112 displayed on the browser 111 of the user terminal 110, FIG. 8(A) is a diagram explaining the operation of the artificial intelligence system 100 which is the first embodiment of the present invention when a user input includes an expression UL of an inappropriate remark, and FIG. 8(B) is a diagram explaining the operation of the artificial intelligence system 100 which is the first embodiment of the present invention, FIG. 9 is a diagram illustrating the operation when an expression UL of an inappropriate remark is included in the generated text data GT by the artificial intelligence system 100 according to the first embodiment of the present invention; FIG. 10 is a diagram illustrating the operation of the main topic designation function of the artificial intelligence system 100 according to the first embodiment of the present invention; and FIG. 11 is a diagram illustrating an example of setting users available for use by the bot 121 on the management screen 113 of the artificial intelligence system 100 according to the first embodiment of the present invention.FIG. 12 is a diagram showing a state in which a log is displayed on the management screen 113 of the artificial intelligence system 100 according to the first embodiment of the present invention.

[0023] As shown in FIG. 1, the first embodiment of the present invention, an artificial intelligence system 100, includes a user terminal 110, a first server 120 having a bot 121, and a second server 130 having a pre-trained artificial intelligence means 131. The user terminal 110 is configured to display or output audio based on the information from the first server 120 . More specifically, the user terminal 110 is configured to communicate with the second server 130 via the bot 121 of the first server 120, and to interact with chat GPT, an example of the pre-trained artificial intelligence means 131 of the second server 130, via the bot 121 of the first server 120 on the artificial intelligence service screen 112 displayed on the browser 111.

[0024] Furthermore, in this embodiment, on the management screen 113 displayed on the browser 111 of the user terminal 110, which is an example of an administrator terminal capable of communicating with the first server 120, masking of word attributes contained in the data DT from the user terminal 110 to the bot 121 of the first server 120 can be freely set (see Figure 3). It is assumed that data DT including characters WD of a word attribute of masking preset in the browser 111 of the user terminal 110 is input and transmitted to the first server 120. In this case, the bot 121 of the first server 120 that receives the data determines that the received data DT contains characters WD with the masking word attribute. Then, the bot 121 of the first server 120 converts the character WD of the masked word attribute in the received data DT into another arbitrary character WD1 of the same word attribute. Then, the bot 121 of the first server 120 transmits the converted data DT1 to the second server 130.

[0025] Then, the second server 130 executes generation of text data based on the converted data DT1 using the chat GPT, which is an example of the pre-trained artificial intelligence means 131 of the second server 130. Then, the second server 130 transmits the generated text data GT obtained through the execution to the first server 120. Then, the bot 121 of the first server 120 changes the part of the received generated text data GT that was converted into another arbitrary character WD1 when it was sent to the second server 130 back to the character WD before conversion. Then, the bot 121 of the first server 120 transmits the changed generated text data GT1 to the user terminal 110. The user terminal 110 is configured to then display the received generated text data GT1 after the change on the browser 111 or output it as voice.

[0026] As a result, for example, by setting in advance masking of word attributes that can identify an individual, such as name, address, telephone number, and email address, for data DT sent from the user terminal 110 to the bot 121 of the first server 120, if the data DT entered by the user contains characters WD of these word attributes, the bot 121 of the first server 120 will convert the characters WD of these word attributes into other arbitrary characters WD1 of the same word attribute and then send the converted data DT1 to the second server 130, and text data generation will be performed based on the converted data DT1. As a result, it is possible to prevent the chat GPT, which is an example of the pre-trained artificial intelligence means 131 of the second server 130, from memorizing information about the characters WD with the word attribute that can identify an individual. In other words, the content sent to the chat GPT, which is an example of the pre-trained artificial intelligence means 131 of the second server 130, is converted into characters WD1 that cannot identify an individual. As a result, it is possible to prevent leakage of personally identifiable information to other terminals via the chat GPT, which is an example of the pre-trained artificial intelligence means 131 of the second server 130.

[0027] Furthermore, when the bot 121 of the first server 120 receives the generated text data GT from the second server 130, the portion of the received generated text data GT that was converted into another arbitrary character WD1 when it was sent to the second server 130 is changed back to the character WD before conversion, and the changed generated text data GT1 is sent to the user terminal 110. As a result, it is possible to obtain generated text data GT1 after modification that basically conforms to the intent of the command information PT, which is the prompt contained in the data DT input by the user. In other words, the characters WD1 converted by the bot 121 of the first server 120 when transmitting to the second server 130 are restored to the original characters WD when the bot 121 of the first server 120 receives from the second server 130. As a result, it is possible to avoid unintended information leakage to the second server 130 or to other terminals via the second server 130, while ensuring consistency between the instruction and command content from the user terminal 110 and the response content to the user terminal 110.

[0028] Next, an example of the operation of the artificial intelligence system 100 will be described in more detail. As shown in FIG. 2, in step S1, as a masking setting step, the first server 120 determines whether or not a masking setting operation has been performed for word attributes included in data DT sent from the user terminal 110 to the bot 121 of the first server 120 on an administration screen 113 displayed by a browser 111 of a user terminal 110, which is an example of an administrator terminal capable of communicating with the first server 120, as shown in FIG. 3(A). In the management screen 113, masking of word attributes included in the data DT sent from the user terminal 110 to the bot 121 of the first server 120 can be freely set. For example, suppose that the word attributes "name," "address," "telephone number," "My Number," and "email address" are selected and masking is set. When a masking setting operation has been performed, an HTTP request is transmitted from the user terminal 110 to the first server 120, and the first server 120 determines that a masking setting operation has been performed. If it is determined that a masking setting operation has been performed, the process proceeds to step S2. On the other hand, if it is determined that a masking setting operation has not been performed, step S1 is repeated.

[0029] In step S2, as a masking setting step, the first server 120 registers the masking setting based on an operation of the masking setting on the management screen 113 by the user terminal 110. In step S3, the first server 120 determines whether or not data DT has been input on the artificial intelligence service screen 112 of the browser 111 of the user terminal 110 and transmitted to the first server 120. That is, first server 120 determines whether or not data DT has been received from user terminal 110.

[0030] Here, when a user performs a login operation to the artificial intelligence system 100 at the user terminal 110 , the browser 111 of the user terminal 110 displays an artificial intelligence service screen 112 . As an example, the artificial intelligence service screen 112 is provided with a text input field and a display area for exchanges, and is configured like a chat talk screen. Here, the chat talk screen refers to a screen that displays exchanges of text and file data DT with one or more partners (including the bot 121) in chronological order on a single screen.

[0031] For example, as shown in FIG. 3(B), suppose that an employee list file containing characters WD with a pre-set masking word attribute is input as an example of data DT on the artificial intelligence service screen 112 of the browser 111 of the user terminal 110, and a prompt "↑Tell me who among these people has not been transferred in the last three years" is input as instruction command information PT. Here, the employee list file contains characters with the attribute "name", which is an example of the characters WD of the masking word attribute, as employee names. Then, by a send operation, a file of the employee list including the characters WD with the preset masking word attribute and a prompt "↑Tell me who here has not been transferred in the last three years" as instruction and command information PT are sent from the user terminal 110 to the first server 120. Then, first server 120 determines that data DT has been received from user terminal 110. In this way, if first server 120 determines that data DT has been received from user terminal 110, it proceeds to step S4, but if it determines that data DT has not yet been received, it repeats step S3.

[0032] As for the method of inputting data DT on the artificial intelligence service screen 112 of the browser 111 of the user terminal 110, the text data DT may be input using a keyboard, or the data DT may be input by specifying a file using a mouse or the like, or the microphone icon MC1 provided on the artificial intelligence service screen 112 may be operated to activate the microphone of the user terminal 110, and voice data input by voice may be converted into text data and input.

[0033] In step S4, as a masking target determination step, the bot 121 of the first server 120 determines whether or not the received data DT contains characters WD of the word attribute of masking. If it is determined that it is included, the process proceeds to step S5, and if it is determined that it is not included, the process proceeds to step S6. Here, the command information PT does not include anything to be masked. Since the other file, an employee list, contains the masking target, the masking process will be described using the employee list file as an example. As shown in FIG. 4, an employee list file is transmitted from a user terminal 110 to a first server 120.

[0034] Then, the bot 121 of the first server 120 that receives the data determines that the employee list file of the received data DT contains characters WD with the masking word attribute. Here, the bot 121 is configured to be able to determine the attributes of “name,” “address,” “telephone number,” “My Number,” and “email address” when scanning data DT such as text based on the database of the first server 120. More specifically, the names of employees, which are the word attribute of "name" listed in the employee directory, are detected.

[0035] In step S5, as a masking conversion execution step, the bot 121 of the first server 120 converts the characters WD of the masking word attribute in the received data DT into another arbitrary character WD1 of the same word attribute. More specifically, as shown in FIG. 4, the bot 121 of the first server 120 converts the name of an employee, which is a character WD of the masking word attribute, in the employee list file of the received data DT, into a combination of "name," which is an example of another arbitrary character WD1 of the same word attribute, and a number such as "1."

[0036] The bot 121 of the first server 120 stores a list of the conversions (a table showing the correspondence between before and after the conversion). Furthermore, when user terminal 110 transmits the employee registry file of data DT to first server 120, first server 120 converts the employee registry file of data DT into text, that is, into text data. Then, the first server 120 may transmit the masked text data (DT1) to the second server . If the chat GPT, which is an example of the pre-trained artificial intelligence means 131 of the second server 130, has a limit on the number of characters that can be input at one time, the first server 120 sends text data within the limited number of input characters.

[0037] In step S6, as a masking data transmission step, the bot 121 of the first server 120 transmits the converted data DT1 to the second server . More specifically, as shown in FIG. 4, the bot 121 of the first server 120 transmits the employee list file of the converted data DT1 to the second server 130 together with a prompt. If it is determined in step S4 that the data is not included, the bot 121 of the first server 120 simply transmits the unconverted data DT to the second server 130 in step S6.

[0038] In step S7, as a data generating step, the second server 130 executes generation of text data based on the converted data DT1 using the chat GPT, which is an example of the pre-trained artificial intelligence means 131 of the second server 130. More specifically, as shown in FIG. 4, the second server 130 uses chat GPT, which is an example of the pre-trained artificial intelligence means 131 of the second server 130, to generate text data based on the employee list file and prompts of the converted data DT1. As an example, we will select people who have not been transferred for three years for the word attributes "Name 1," "Name 2," etc., which are converted from the joining date information and past transfer history information written in the employee register file. Then, "There are 10 people: Name 1, Name 3, Name 5, Name 7, . . . " is obtained as an example of generated text data GT. If it is determined in step S4 that the data is not included, then in step S7, text data generation is simply executed based on the unconverted data DT.

[0039] In step S8, as a generated data transmitting step, the second server 130 transmits to the first server 120 the generated text data GT obtained by the execution. In step S9, as a reverse conversion step, the bot 121 of the first server 120 changes the part of the received generated text data GT that was converted into another arbitrary character WD1 when it was sent to the second server 130 back to the character WD before conversion. More specifically, as shown in FIG. 4, for example, when "Suzuki Emi" is converted to "Name 1", "Name 1" is changed back to "Suzuki Emi" (reverse conversion). Then, as an example of the generated text data GT1 after the change, "The 10 people are Emi Suzuki, Ichiro Tanaka, John Smith, Amy Marshall, . . . " is obtained. The bot 121 of the first server 120 refers to the list (pre-conversion correspondence table) saved at the time of conversion and changes it back to the characters WD before conversion. Also, if it is determined in step S4 that the content is not included, then since no conversion has been performed in the first place in step S5, no change is made in step S9.

[0040] In step S10, as a reverse conversion data transmission step, the bot 121 of the first server 120 transmits the generated text data GT1 after the change to the user terminal 110. If it is determined in step S4 that the generated text data GT is not included, then in step S10, the bot 121 of the first server 120 simply transmits the generated text data GT to the user terminal 110.

[0041] In step S11, which is a display / audio output step, the user terminal 110 displays the received generated text data GT1 after the change on the browser 111 or outputs it as audio, as shown in FIG. 3(B). If it is determined in step S4 that the generated text data GT is not included, then in step S11, the user terminal 110 simply displays the generated text data GT on the browser 111 or outputs it as voice.

[0042] As a result, as described above, for example, by setting in advance masking of word attributes that can identify an individual, such as name, address, telephone number, and email address, for data DT sent from the user terminal 110 to the bot 121 of the first server 120, if the data DT entered by the user contains characters WD of those word attributes, the bot 121 of the first server 120 will convert the characters WD of those word attributes into other arbitrary characters WD1 of the same word attribute and then send the converted data DT1 to the second server 130, and text data generation will be performed based on the converted data DT1. As a result, as shown in FIG. 4, it is possible to prevent the chat GPT, which is an example of the pre-trained artificial intelligence means 131 of the second server 130, from memorizing information about the characters WD with the word attribute that can identify an individual. In other words, the content sent to the chat GPT, which is an example of the pre-trained artificial intelligence means 131 of the second server 130, is converted into characters WD1 that cannot identify an individual. As a result, it is possible to prevent leakage of personally identifiable information to other terminals via the chat GPT, which is an example of the pre-trained artificial intelligence means 131 of the second server 130.

[0043] Furthermore, when the bot 121 of the first server 120 receives the generated text data GT from the second server 130, the portion of the received generated text data GT that was converted into another arbitrary character WD1 when it was sent to the second server 130 is changed back to the character WD before conversion, and the changed generated text data GT1 is sent to the user terminal 110. As a result, as shown in FIG. 3(B) and FIG. 4, it is possible to obtain generated text data GT1 after modification that basically conforms to the intent of the command information PT, which is the prompt contained in the data DT input by the user. In other words, the characters WD1 converted by the bot 121 of the first server 120 when transmitting to the second server 130 are restored to the original characters WD when the bot 121 of the first server 120 receives from the second server 130. As a result, it is possible to avoid unintended information leakage to the second server 130 or to other terminals via the second server 130, while ensuring consistency between the instruction and command content from the user terminal 110 and the response content to the user terminal 110.

[0044] Note that, as an example of data DT, an example of masking text data contained in an employee directory file, which is a file for a spreadsheet software with a cell-format display, has been described. However, the data DT to be masked is not limited to files for spreadsheet software. For example, when the data DT is image data such as a photograph or video, the bot 121 uses an image recognition means to determine whether or not there is a masking target, and when it determines that there is a masking target, it masks the target part and replaces it with another part having the same attributes before transmitting it to the second server 130.

[0045] The bot 121 of the first server 120 saves a list of the conversions (a table showing the correspondence between before and after the conversion). Then, the bot 121 changes (reverse-converts) the generated data received from the second server 130 by referring to the list (pre-conversion correspondence table) saved at the time of conversion to return it to its pre-conversion state, and then transmits the changed generated data to the user terminal 110. The generated data and the changed generated data are not limited to text data, but may be image data such as photographs and videos.

[0046] Furthermore, for inputting data DT on the artificial intelligence service screen 112 of the browser 111, for example, a microphone icon MC1 may be provided as shown in FIG. 5(A). Then, when the microphone icon MC1 is operated, as shown in FIG. 5(B), the user terminal 110 may display a microphone active indication mark MC2 on the artificial intelligence service screen 112, and may also activate the microphone of the user terminal 110 to convert voice data input by voice into text data for input and display. At this time, the artificial intelligence service screen 112 of the user terminal 110 displays the content (data DT, instruction command information PT) of the voice data input by voice, which has been converted into text data, allowing the user to confirm the input content (data DT, instruction command information PT).

[0047] Next, the prohibited word setting function will be explained. In order to make this function easier to understand, the explanation of the conversion and reverse conversion performed by the masking function described above will be omitted, and only the essential parts of the prohibited word setting function will be explained below. In this embodiment, as shown in FIG. 6(A), prohibited words PB can be freely set on an administration screen 113 displayed on a browser 111 of a user terminal 110, which is an example of an administrator terminal capable of communicating with the first server 120. For example, it is assumed that "confidential", "personal information", "confidential", and "credit card" are set as prohibited words PB.

[0048] Then, as shown in FIG. 6(B), assume that instruction command information PT is input as data DT including the prohibited word PB of "personal information" that is preset on the artificial intelligence service screen 112 of the browser 111 of the user terminal 110. Then, the bot 121 of the first server 120 that receives the data determines that the instruction command information PT as the received data DT contains the prohibited word PB of "personal information." The bot 121 of the first server 120 is configured not to send the received instruction command information PT to the second server 130, but to send to the user terminal 110 an example of information indicating that an unavailable word is included, such as "The text contains an unavailable word. Please try using a different expression."

[0049] As a result, for example, by presetting words such as "confidential," "personal information," and "My Number" as prohibited words PB, if the data DT entered by the user contains the prohibited word PB, the bot 121 of the first server 120 detects the prohibited word PB and stops communication with the chat GPT, which is an example of the pre-trained artificial intelligence means 131 of the second server 130, and sends information to the user terminal 110 that the data contains an unusable word. As a result, undesirable use by users can be prohibited. In other words, the transmission of information to the chat GPT, which is an example of the pre-trained artificial intelligence means 131 of the second server 130, is restricted. As a result, it is possible to prevent information leakage to other terminals via the chat GPT, which is an example of the pre-trained artificial intelligence means 131 of the second server 130. The configuration of the prohibited word setting function does not need to be premised on the configuration of the masking function.

[0050] Next, we will explain the inappropriate comment elimination function. To make this function easier to understand, we will omit the explanation of the conversion and reverse conversion that occurs through the masking function described above, and will only explain the essential parts of the inappropriate comment elimination function below. In this embodiment, as shown in FIG. 7(A), the exclusion of inappropriate comments can be freely configured on an administration screen 113 displayed on a browser 111 of a user terminal 110, which is an example of an administrator terminal capable of communicating with the first server 120. For example, let's say that the word attributes of inappropriate comments are selected and excluded, such as "racism," "gender discrimination," "address discrimination," "disrespect" that shows no respect for the other person, "abusive language," "sexual harassment," and so-called "appearance harassment," which is lookism.

[0051] Then, as shown in Figures 7(B) and 8(A), data DT including a preset expression UL for an inappropriate remark is input into the artificial intelligence service screen 112 of the browser 111 of the user terminal 110 and transmitted to the first server 120. Then, the bot 121 of the first server 120 is configured to determine that the received data DT contains an expression UL of an inappropriate remark, and not to transmit the received data DT to the second server 130, but to transmit information to the user terminal 110 indicating that the data contains an expression UL of an inappropriate remark.

[0052] As a result, when an inappropriate expression UL is included in an exchange between the user terminal 110 and the chat GPT, which is an example of the pre-trained artificial intelligence means 131, the bot 121 of the first server 120 detects the inappropriate expression UL and blocks it, and notifies the sender that the inappropriate expression UL is included. As a result, it is possible to eliminate any exchanges that contain inappropriate expressions. In other words, when a user uses an inappropriate remark expression UL, the content of the remark is not transmitted to the chat GPT, which is an example of the pre-trained artificial intelligence means 131 of the second server 130. As a result, it is possible to avoid the risk that the chat GPT, which is an example of the pre-trained artificial intelligence means 131, will learn the expression UL of inappropriate remarks. Furthermore, when a user uses an inappropriate comment UL, the bot 121 of the first server 120 returns a message to that effect to the user. As a result, the user can recognize that he or she has used an inappropriate expression UL.

[0053] On the other hand, as shown in Figure 8 (B), if the data DT transmitted from the user terminal 110 to the first server 120 does not contain an inappropriate remark expression UL, the second server 130 generates text data based on the instruction command information PT using the chat GPT, which is an example of the pre-trained artificial intelligence means 131 of the second server 130. Then, the second server 130 transmits the generated text data GT obtained by the execution to the first server 120, and the bot 121 of the first server 120 determines whether or not the received generated text data GT contains an expression UL of an inappropriate remark. Here, if it is determined that the inappropriate remark is included, the bot 121 of the first server 120 is configured to not send the received generated text data GT to the user terminal 110, but to instruct the chat GPT, which is an example of the pre-trained artificial intelligence means 131 of the second server 130, to again generate text data based on the instruction command information PT without using the inappropriate remark expression UL.

[0054] As a result, if an inappropriate remark expression UL is included in the generated text data GT generated by the chat GPT, which is an example of the pre-trained artificial intelligence means 131, the content is not transmitted to the user terminal 110. As a result, it is possible to avoid the risk that a user will use generated text data GT containing inappropriate remarks UL generated by chat GPT, which is an example of the pre-trained artificial intelligence means 131, as is. Furthermore, if the generated text data GT contains an expression UL of an inappropriate remark, an instruction command is sent from the bot 121 of the first server 120 to the chat GPT, which is an example of a pre-trained artificial intelligence means 131, to generate text data again based on the instruction command information PT without using the expression UL of the inappropriate remark. As a result, generated text data GT that does not include inappropriate remarks UL can be obtained and displayed on the user terminal 110. The configuration of the inappropriate comment removal function does not need to be premised on the configuration of the masking function.

[0055] Next, the main topic setting function will be explained. To make this function easier to understand, we will omit the explanation of the conversion and inverse conversion performed by the masking function described above, and only explain the essential parts of the main topic setting function below. In this embodiment, as shown in FIG. 9, the user terminal 110 can freely specify a main topic for the bot 121 of the first server 120. The method of specification may be, for example, by inputting "Set main topic" and information about the main topic to be specified in the text input field of the artificial intelligence service screen 112 of the browser 111. In addition, as an example of a rich menu on the artificial intelligence service screen 112 of the browser 111, a user may operate the “Set Main Topic” button 112a, and then input information about the main topic he or she wishes to specify in the text input field on the artificial intelligence service screen 112 of the browser 111. Alternatively, as shown in FIG. 9, when the cursor is moved over the exchange on the artificial intelligence service screen 112 of the browser 111, a selection icon SL is displayed, and when a star mark is selected as an example of a selection option of the selection icon SL, the content of the exchange is set as the main topic. Other selectable functions of the selection icon SL include a copy function, a voice output function, and a clipboard temporary save function.

[0056] For example, as shown in FIG. 9, "automobile industry" is specified as main topic information MT. Then, as shown in FIG. 10, when the user inputs “Tell me about the situation in America!” as an example of instruction command information PT on the artificial intelligence service screen 112 of the browser 111 while specifying the main topic, the user terminal 110 transmits the input instruction command information PT to the first server 120. Then, the bot 121 of the first server 120 adds the main topic information MT to the received instruction command information PT and transmits it to the second server 130.

[0057] Then, the second server 130 executes generation of text data based on the instruction command information PT and the main topic information MT using the chat GPT, which is an example of the pre-trained artificial intelligence means 131 of the second server 130. The second server 130 is configured to transmit the generated text data GT obtained by the execution to the first server 120.

[0058] As a result, when a main topic is being specified on the artificial intelligence service screen 112 of the browser 111, the main topic information MT is always included in the information from the bot 121 to the chat GPT, which is an example of a pre-trained artificial intelligence means 131, in the interaction between the user and the chat GPT, which is an example of a pre-trained artificial intelligence means 131, and the generated text data GT generated by the chat GPT, which is an example of a pre-trained artificial intelligence means 131, has content that is in line with the main topic. As a result, it is possible to ensure that the content of consultations, etc., is consistent (the content is consistent) in interactions between the user and chat GPT, which is an example of the pre-trained artificial intelligence means 131. This means that even if a user asks a question multiple times, the answers stay on topic. As a result, you can avoid gradually drifting away from the main topic as you continue asking questions.

[0059] When canceling the main topic setting, the user may operate the “Cancel main topic setting” button 112b in the rich menu of the artificial intelligence service screen 112 of the browser 111, or may enter, as an example, “Cancel main topic setting” in the text input field of the artificial intelligence service screen 112 of the browser 111. In addition, when the main topic is being set on the artificial intelligence service screen 112 of the browser 111, the artificial intelligence service screen 112 of the browser 111 may be configured to display a message indicating that the main topic is being set and the main topic information MT, for example, "Main topic "Automobile Industry" being set," on the artificial intelligence service screen 112 of the browser 111. This allows a user to interact with chat GPT, which is an example of a pre-trained artificial intelligence means 131, while checking that a main topic is being set and what the main topic information MT is. This means you can avoid forgetting to unsubscribe and continuing the conversation when you want to move on to the next topic. It also helps to avoid forgetting what the main topic was set to as you continue to exchange messages. The configuration of the main topic setting function does not need to be premised on the configuration of the masking function.

[0060] Furthermore, in this embodiment, as shown in Figure 11, on the management screen 113 displayed on the browser 111 of the user terminal 110, which is an example of an administrator terminal that can communicate with the first server 120, the users who can use the bot 121 of the first server 120 or the chat GPT of the second server 130 on the artificial intelligence service screen 112 of the browser 111 of the user terminal 110 can be freely set based on the user identification information US. For example, by entering personally identifying information such as the user's email address as an example of user identification information US and operating the add user button, the bot 121 of the first server 120 or the chat GPT of the second server 130 can be used.

[0061] As a result, users are identified based on the user identification information US, and users who can interact with the chat GPT, an example of the pre-trained artificial intelligence means 131 of the second server 130, via the bot 121 of the first server 120 are specified and restricted. As a result, administrators can easily set usage restrictions and permissions for Chat GPT for each employee. In addition, the configuration of the setting function for users who can use the chat GPT does not need to be premised on the configuration of the masking function.

[0062] In addition, in this embodiment, as shown in Figure 12, on an administration screen 113 displayed on a browser 111 of a user terminal 110, which is an example of an administrator terminal capable of communicating with the first server 120, historical information of interactions between the user terminal 110 and the bot 121 of the first server 120 or the chat GPT of the second server 130 on an artificial intelligence service screen 112 of the browser 111 of the user terminal 110 can be freely displayed. The history information items include the access date and time, which is transmission date and time information, the user ID, which is user identification information US, IP address information, the channel, which is chat application type information, the action, which is the type of prompt, the input text, which is the instruction command information PT entered on the artificial intelligence service screen 112 of the browser 111, and the result, which is information on whether or not a reply has been received for the generated text data GT from the chat GPT, which is an example of the pre-trained artificial intelligence means 131 of the second server 130. Then, the information is displayed for each item. This allows information regarding which user entered what content, when, where, and whether or not they obtained any results to be displayed on the management screen 113. As a result, the administrator can check whether the user's usage is appropriate or not.

[0063] The artificial intelligence system 100, which is the first embodiment of the present invention obtained in this way, comprises a user terminal 110, a first server 120 having a bot 121, and a second server 130 having a chat GPT which is an example of a pre-trained artificial intelligence means 131. The user terminal 110 communicates with the second server 130 via the bot 121 of the first server 120. In an administration screen 113 of an administrator terminal which can communicate with the first server 120, masking of the word attribute included in the data DT from the user terminal 110 to the bot 121 of the first server 120 can be freely set. When data DT including a character WD of a word attribute of masking preset on an artificial intelligence service screen 112 of a browser 111 of the user terminal 110 is input and transmitted to the first server 120, the bot 121 of the first server 120 which receives the data determines that the received data DT includes a character WD of a word attribute of masking, and converts the character WD of the word attribute of masking in the received data DT to another arbitrary character WD1 of the same word attribute, and then outputs the converted data. The second server 130 uses the chat GPT, which is an example of the pre-trained artificial intelligence means 131 of the second server 130, to generate text data based on the converted data DT1, and transmits the generated text data GT obtained by the above operation to the first server 120. The bot 121 of the first server 120 changes the part of the received generated text data GT, which is a different arbitrary character WD1 that was converted when the data was transmitted to the second server 130, back to the character WD before conversion, and then transmits the changed generated text data GT. The modified text data GT1 is sent to the user terminal 110, and the user terminal 110 displays the received modified generated text data GT1 on the artificial intelligence service screen 112 of the browser 111 or outputs it as audio. This makes it possible to avoid unintended information leakage to the chat GPT, which is an example of the pre-trained artificial intelligence means 131 of the second server 130, or to other terminals via the second server 130, and also ensures consistency between the instruction and command content from the user terminal 110 and the response content to the user terminal 110.

[0064] Furthermore, the prohibited words PB can be freely set on the management screen 113 of the administrator terminal which can communicate with the first server 120, and when data DT containing the preset prohibited words PB is input on the artificial intelligence service screen 112 of the browser 111 of the user terminal 110 and sent to the first server 120, the bot 121 of the first server 120 determines that the received data DT contains the prohibited words PB, and does not send the received data DT to the second server 130, but instead sends information to the user terminal 110 that the data contains unusable words. This makes it possible to prohibit undesirable use by users and to avoid information leakage to other terminals via the chat GPT, which is an example of the pre-trained artificial intelligence means 131 of the second server 130.

[0065] In addition, the exclusion of inappropriate comments can be freely set on the management screen 113 of the administrator terminal that can communicate with the first server 120, and when data DT including a preset expression UL of an inappropriate comment is input on the artificial intelligence service screen 112 of the browser 111 of the user terminal 110 and transmitted to the first server 120, the bot 121 of the first server 120 determines that the received data DT includes an expression UL of an inappropriate comment, does not transmit the received data DT to the second server 130, and transmits information to the user that the expression UL of an inappropriate comment is included. The second server 130 uses the chat GPT, which is an example of the pre-trained artificial intelligence means 131 of the second server 130, to generate text data based on the instruction command information PT included in the received data DT, and transmits the generated text data GT obtained by the execution to the first server 120. The bot 121 of the first server 120 determines whether the received generated text data GT contains an expression UL of an inappropriate remark, and if it determines that the expression UL is included, transmits the received generated text data GT to the user terminal 110. 0, and instructs chat GPT, an example of pre-trained artificial intelligence means 131 of the second server 130 to generate text data again based on the instruction command information PT without using the expression UL for the inappropriate remarks. This makes it possible to eliminate interactions that include the expression UL for inappropriate remarks, and if a user uses the expression UL for an inappropriate remark, it is possible to avoid the risk that chat GPT, an example of pre-trained artificial intelligence means 131, will learn the expression UL for the inappropriate remarks, and the user can recognize that they have used the expression UL for an inappropriate remarks. If the expression UL for an inappropriate remarks is included in the generated text data GT generated by chat GPT, an example of pre-trained artificial intelligence means 131, it is possible to avoid the risk that the user will use the generated text data GT containing the expression UL for the inappropriate remarks generated by chat GPT, an example of pre-trained artificial intelligence means 131, as is, and it is possible to obtain generated text data GT that does not include the expression UL for the inappropriate remarks and display it on the user terminal 110.

[0066] Furthermore, the user terminal 110 can freely specify a main topic for the bot 121 of the first server 120, and when the user inputs instruction command information PT as data DT on the artificial intelligence service screen 112 of the browser 111 while specifying the main topic and sends it to the first server 120, the bot 121 of the first server 120 adds main topic information MT to the received instruction command information PT and sends it to the second server 130, and the second server 130 generates text data based on the instruction command information PT and the main topic information MT using the chat GPT, which is an example of the pre-trained artificial intelligence means 131 of the second server 130, and sends the generated text data GT obtained by the execution to the first server 120.By this configuration, even if the user continues to ask questions multiple times, it is possible to avoid the question gradually deviating from the main topic as the question continues.

[0067] In addition, the management screen 113 of the administrator's terminal, which can communicate with the first server 120, is configured so that the users who can use the bot 121 of the first server 120 can be freely set on the artificial intelligence service screen 112 of the browser 111 of the user terminal 110 using user identification information US, so that the administrator can easily set usage restrictions and permissions for the chat GPT for each employee member.

[0068] Furthermore, the program of the artificial intelligence system 100, which is the first embodiment of the present invention, includes masking setting steps S1 and S2 in which the first server 120 sets masking when there is a masking setting for a word attribute included in data DT from the user terminal 110 to the bot 121 of the first server 120 on the management screen 113 of an administrator terminal capable of communicating with the first server 120; masking target determination steps S3 and S4 in which, when data DT is input on the artificial intelligence service screen 112 of the browser 111 of the user terminal 110 and transmitted to the first server 120, the bot 121 of the first server 120 receives the data DT and determines whether the received data DT contains a character WD of the word attribute of masking; a masking conversion execution step S5 in which, when the bot 121 of the first server 120 determines that the character WD of the word attribute of masking is included, the character WD of the word attribute of masking in the received data DT is converted into another arbitrary character WD1 of the same word attribute; and a masking data transmission step S6 in which the second server 130 transmits the converted data DT1 to the second server 130; a data generation step S7 in which the second server 130 generates text data based on the converted data DT1 using a chat GPT, which is an example of the pre-trained artificial intelligence means 131 of the second server 130; a generated data transmission step S8 in which the second server 130 transmits the generated text data GT obtained by the execution to the first server 120; an inverse conversion step S9 in which the bot 121 of the first server 120 changes the part of the received generated text data GT that was converted when it was transmitted to the second server 130 back to the character WD before conversion; an inverse conversion data transmission step S10 in which the bot 121 of the first server 120 transmits the changed generated text data GT1 to the user terminal 110; and a display / audio output step S11 in which the user terminal 110 displays the received changed generated text data GT1 on the artificial intelligence service screen 112 of the browser 111 or outputs it as audio.The effect of this is enormous, as it is possible to prevent unintended information leakage to the chat GPT, which is an example of the pre-trained artificial intelligence means 131 of the second server 130, and to prevent unintended information leakage to other terminals via the second server 130, and to ensure consistency between the instruction and command content from the user terminal 110 and the response content to the user terminal 110. EXAMPLES

[0069] Next, an artificial intelligence system 100 according to a second embodiment of the present invention will be described with reference to Figs. 13(A) and 13(B). Here, Figure 13(A) is a diagram showing an example of input in a sending message window 114, which is a message sending and receiving window displayed by a browser 111 of a user terminal 110 of an artificial intelligence system 100 which is the second embodiment of the present invention, and Figure 13(B) is a diagram showing an example of display of a receiving message window 115, which is a message sending and receiving window displayed by a browser 111 of a user terminal 110 of an artificial intelligence system 100 which is the second embodiment of the present invention. The artificial intelligence system 100 of the second embodiment is a system in which the artificial intelligence service screen 112 in the form of a chat talk screen displayed by the browser 111 of the user terminal 110 of the artificial intelligence system 100 of the first embodiment has been changed to a message sending / receiving window of an email application displayed by the browser 111. Since many elements are common to the artificial intelligence system 100 of the first embodiment, detailed explanations of the common points will be omitted.

[0070] An artificial intelligence system 100 according to an embodiment of the present invention is configured to be able to freely communicate data DT, such as text data, between a user terminal 110 and a server, which is an example of a mail server. The mail server may be configured separately from the first server 120, or may be configured in common with the first server 120. As shown in FIG. 13(A), with the address of bot 121 specified as the destination in send message window 114 displayed on browser 111 of user terminal 110, a file of an employee list containing characters WD with a masking word attribute that has been preset as an example of arbitrary data DT is attached by user operation. Furthermore, it is assumed that a prompt "Tell me who in the attached file 'Employee Directory' has not been transferred in the last three years" is input as instruction and command information PT as an example of data DT. Here, the employee list file contains characters with the attribute "name", which is an example of the characters WD of the masking word attribute, as employee names. Then, by a send operation, a file of the employee list including the characters WD with the preset masking word attribute and a prompt "↑Tell me who here has not been transferred in the last three years" as instruction and command information PT are sent from the user terminal 110 to the first server 120.

[0071] Then, similarly to the first embodiment described above, the bot 121 of the first server 120 executes data conversion using a masking function, and transmits the converted data DT1 to the second server . Furthermore, the second server 130 executes text data generation using the chat GPT, which is an example of the pre-trained artificial intelligence means 131, based on the converted data DT1, and transmits the generated text data GT to the first server 120. Then, the bot 121 of the first server 120 changes the characters WD1 converted by the masking function in the generated text data GT back to the original characters WD, and transmits the changed generated text data GT1 to the user terminal 110. Then, as shown in FIG. 13(B), the user terminal 110 is configured to display the received generated text data GT1 after the change in the received message window 115. In other words, if the bot 121 of the first server 120 determines that the data DT (including attachments) sent in the send message window 114 contains characters WD with a pre-set masking word attribute, it converts the characters WD into another character WD1 (performs masking) and sends it to the second server 130, and reverse-converts the generated text data GT received from the second server 130.

[0072] This makes it possible to obtain the same effects as in the first embodiment, the detailed description of which will be omitted. In addition, a configuration may be adopted in which a bot address insertion button 114a is displayed in the outgoing message window 114 of a user who has been granted permission to use the bot 121. When the user operates this bot address insertion button 114a, the bot address AD, which is the email address of the bot 121, is inserted into the destination field. This allows the user to set the destination to the bot 121 simply by operating the bot address insertion button 114a. As a result, users can easily interact with Chat GPT, an example of the pre-trained artificial intelligence means 131 of the second server 130.

[0073] Furthermore, in the artificial intelligence system 100 of this embodiment, it is of course possible to set prohibited words PB, similarly to the artificial intelligence system 100 of the first embodiment described above, although a detailed description thereof will be omitted. Furthermore, in the artificial intelligence system 100 of this embodiment, it is of course possible to set the exclusion of inappropriate comments in the same way as in the artificial intelligence system 100 of the first embodiment described above. A detailed explanation thereof will be omitted. Furthermore, in the artificial intelligence system 100 of this embodiment, it is of course possible to set a main topic, similar to the artificial intelligence system 100 of the first embodiment described above, although a detailed description thereof will be omitted. In addition, in the artificial intelligence system 100 of this embodiment, it is of course possible to freely set users who can use the bot 121 of the first server 120 by using the user identification information US. A detailed description thereof will be omitted. Furthermore, in the artificial intelligence system 100 of this embodiment, similar to the artificial intelligence system 100 of the first embodiment described above, it is of course possible to freely display historical information of interactions between the user terminal 110 and the bot 121 of the first server 120 on the management screen 113. A detailed explanation thereof will be omitted. [Explanation of symbols]

[0074] 100 ··· Artificial Intelligence Systems 110 User terminal 111 ··· Browser 112 ··· AI service screen 112a "Set main topic" button 112b "Cancel main topic setting" button 113: Administration screen (of administrator's terminal) 114 - Send Message Window 114a...Bot address insertion button 115 Received Message Window 120 ··· First Server 121 Bot 130 Second Server 131 ··· Pre-trained AI method (Chat GPT) DT...Data DT1: Data after conversion PT... Directive command information GT: Generated text data GT1: Generated text data after modification WD: Masking word attribute character (character before conversion) WD1: Any other character with the same word attribute (any other character converted) PB: Forbidden words UL ··· Inappropriate remarks MT ··· Main topic information US ··· User identity AD ··· Bot address MC1 ··· Microphone icon MC2: Microphone active indicator SL ··· Selection icon

Claims

1. An artificial intelligence system comprising a user terminal, a first server having a bot, and a second server having a pre-trained artificial intelligence means, the user terminal displaying or outputting audio based on the first server, The user terminal communicates with a second server via a bot of the first server; In a management screen of an administrator terminal capable of communicating with the first server, masking of word attributes included in data from a user terminal to a bot of the first server can be freely set, When data including characters of a masking word attribute preset in the browser of the user terminal is input and transmitted to a first server, a bot of the first server that receives the data determines that the received data includes characters of a masking word attribute, converts the characters of the masking word attribute in the received data into other arbitrary characters of the same word attribute, and transmits the converted data to a second server; The second server executes text data generation based on the converted data using the pre-trained artificial intelligence means of the second server, and transmits the generated text data to the first server; The bot of the first server changes the part of the received generated text data that was converted when it was transmitted to the second server back to the original character before conversion, and then transmits the changed generated text data to the user terminal; An artificial intelligence system characterized in that the user terminal is configured to display the received, modified generated text data on a browser or output it as voice.

2. A prohibited word can be freely set on a management screen of an administrator terminal capable of communicating with the first server, The artificial intelligence system described in claim 1, characterized in that when data containing predetermined prohibited words is input into the browser of the user terminal and sent to the first server, the bot of the first server determines that the received data contains prohibited words, does not send the received data to the second server, and instead sends information to the user terminal that the data contains unusable words.

3. The exclusion of inappropriate comments can be freely set on a management screen of an administrator terminal capable of communicating with the first server, When data including a preset inappropriate remark is input to a browser of the user terminal and transmitted to a first server, a bot of the first server determines that the received data includes an inappropriate remark, does not transmit the received data to a second server, and transmits information to the user terminal indicating that the data includes an inappropriate remark; On the other hand, the second server uses the pre-trained artificial intelligence means of the second server to generate text data based on the instruction and command information contained in the received data, and transmits the generated text data obtained by this execution to the first server, and the bot of the first server determines whether the received generated text data contains inappropriate remarks, and if it determines that it does, does not send the received generated text data to the user terminal, but instructs the pre-trained artificial intelligence means of the second server to generate text data based on the instruction and command information again without using the inappropriate remarks.

4. the user terminal is free to designate a main topic for the bot of the first server, and when a user inputs instruction command information as data in a browser while designating the main topic and transmits the instruction command information to the first server, the bot of the first server adds main topic information to the received instruction command information and transmits the information to the second server; The artificial intelligence system according to any one of claims 1 to 3, characterized in that the second server is configured to generate text data based on instruction command information and main topic information using a pre-trained artificial intelligence means of the second server, and transmit the generated text data obtained by the execution to the first server.

5. The artificial intelligence system described in claim 4, characterized in that the management screen of an administrator terminal capable of communicating with the first server is configured to freely set users who can use the bot of the first server in the browser of a user terminal based on user identification information.

6. A program of an artificial intelligence system that a user terminal displays or outputs by voice based on a first server, A masking setting step in which the first server sets masking when a masking setting is made for a word attribute included in data from a user terminal to a bot of the first server on a management screen of an administrator terminal capable of communicating with the first server; A masking target determination step in which, when data is input into the browser of the user terminal and transmitted to the first server, a bot of the first server receives the data and determines whether the received data contains a character having a word attribute for masking; A masking conversion execution step of converting the characters of the masking word attribute of the received data into another arbitrary character of the same word attribute when the bot of the first server determines that the characters of the masking word attribute are included; A masking data transmission step in which the bot of the first server transmits the converted data to a second server; A data generating step in which the second server generates text data based on the converted data using the pre-trained artificial intelligence means of the second server; a generated data transmission step in which the second server transmits generated text data obtained by the execution to the first server; A reverse conversion step in which the bot of the first server converts a portion of the received generated text data, which has been converted when the data is transmitted to the second server, back to the original character before conversion; a reverse conversion data transmission step in which the bot of the first server transmits the generated text data after the change to the user terminal; A program for an artificial intelligence system, comprising a display / audio output step in which the user terminal displays the received modified generated text data on a browser or outputs it as audio.

Images