Information processing method, program, and information processing device

Abstract

Outputs response conversations that have similar characteristics to the target input conversation. [Solution] An information processing method according to one embodiment of the present disclosure extracts features of the target based on a dialogue including an input conversation by the target and a response conversation by a language model, and if it is determined that predetermined conditions are met for the extracted features, the computer is instructed to execute a process to output a feature prompt based on the features of the target at the time the predetermined conditions were met. Furthermore, when the input conversation is input, the information processing method according to one embodiment of the present disclosure outputs the features of the target along with the response conversation.

Description

Technical Field

[0001] The present technology relates to an information processing method, a program, and an information processing apparatus.

Background Art

[0002] Conventionally, a dialogue system for interacting with a user has been proposed. For example, the dialogue system described in Patent Document 1 includes a speech recognition processing unit that recognizes the content of the user's utterance, and a dialogue engine that generates an utterance to the user according to an application based on the recognition result.

Prior Art Documents

Patent Documents

[0003]

Patent Document 1

Summary of the Invention

Problems to be Solved by the Invention

[0004] However, the technology described in Patent Document 1 does not consider outputting a response dialogue whose characteristics are similar to the target input dialogue.

[0005] The present disclosure has been made in view of such circumstances, and an object thereof is to provide an information processing method or the like for outputting a feature prompt capable of constructing a pseudo-personality model that behaves similarly to a target.

Means for Solving the Problems

[0006] An information processing method according to an embodiment of the present disclosure extracts features of a target based on a dialogue including an input dialogue by the target and a response dialogue by a language model, and when it is determined that the extracted features satisfy a predetermined condition, causes a computer to execute a process of outputting a feature prompt based on the features of the target at the time when the predetermined condition is satisfied.

Effects of the Invention

[0007] In one embodiment of the information processing method of this disclosure, it is possible to output feature prompts that can construct a pseudo-personality model that exhibits similar behavior to that of the target. [Brief explanation of the drawing]

[0008] [Figure 1] This is an explanatory diagram showing an overview of the information processing system. [Figure 2] This is a block diagram showing an example configuration of a user terminal. [Figure 3] This is an explanatory diagram showing the system prompts that are input to the language model. [Figure 4] This is an explanatory diagram showing the input and output of a language model in the interaction between the target and the language model. [Figure 5] This is an explanatory diagram showing the output of feature prompts by a language model. [Figure 6] This is an explanatory diagram showing examples of dialogue and feature prompt displays on a user terminal. [Figure 7] This is an explanatory diagram showing an example of a feature prompt table. [Figure 8] This flowchart shows the feature prompt output process by the terminal processing unit of the user terminal. [Figure 9] This is an explanatory diagram showing the output of mimicked sentences by a language model. [Figure 10] This shows an example of how the imitation text is displayed on the user's terminal. [Figure 11] This flowchart shows the process of outputting imitation text by the terminal processing unit of the user terminal. [Figure 12] This is an explanatory diagram showing an overview of the information processing system according to Embodiment 2. [Figure 13] This is a block diagram showing an example of the configuration of an information processing device. [Figure 14] This is an explanatory diagram showing an example of a feature prompt table. [Figure 15] This flowchart shows an example of feature prompt output processing by the processing unit of an information processing device. [Figure 16]Shows an example of displaying a mimicked sentence on the user terminal according to Embodiment 2. [Figure 17] It is a flowchart showing the mimicked sentence output process by the processing unit of the information processing apparatus.

Mode for Carrying Out the Invention

[0009] (Embodiment 1) The language model M according to Embodiment 1 extracts the features of the target based on the dialogue including the acquired input dialogue of the target and the response dialogue by the language model M, and when the extracted features satisfy a predetermined condition, based on the extracted features, it outputs a feature prompt. Hereinafter, Embodiment 1 will be described with reference to the drawings.

[0010] FIG. 1 is an explanatory diagram showing an overview of the information processing system S. The information processing system S includes a language model server 1 and a user terminal 2. The language model server 1 is, for example, a cloud server. The user terminal 2 is a smartphone, a tablet terminal, a personal computer, or the like owned by a user who is a target for mimicking the language model M. In the present embodiment, the user terminal 2 will be described as being a personal computer. The language model server 1 and the user terminal 2 can communicate with each other via the network N. The language model server 1 stores the language model M.

[0011] In the present embodiment, the target for mimicking the language model M is a user, but it is not limited thereto. The target for mimicking the language model M is, for example, a character appearing in a work including an anime, a game, or a movie, a virtual personality such as a Vtuber or an avatar, or an AI (artificial intelligence) bot such as the AI bot "Rinna" developed by Microsoft Japan Co., Ltd., etc. It may be an AI character or the like. In the former case, the user pretends to be the character who wishes to generate the feature prompt and inputs the input dialogue. In the latter case, the response sentence of the AI bot obtained through the application in which the AI bot operates is input as the input dialogue.

[0012] The user terminal 2 transmits a system prompt indicating an instruction for the language model M and an input conversation input by the user to the language model server 1. The system prompt and the input conversation received by the language model server 1 are input into the language model M stored in the language model server 1. Hereinafter, the transmission of the system prompt or the input conversation to the language model server 1 and the input of the system prompt or the input conversation to the language model M are described as being synonymous.

[0013] The language model M outputs a response conversation for the input conversation according to the instruction indicated by the input system prompt. Also, the language model M extracts the user's characteristics from the input conversation and outputs the characteristics extracted together with the response conversation. Further, when the extracted characteristics satisfy a predetermined condition, a feature prompt based on the extracted characteristics is output to the user terminal 2. The language model server 1 transmits the response conversation, the user's characteristics, or the feature prompt output by the language model M to the user terminal 2. Hereinafter, the output of the response conversation, the user's characteristics, or the feature prompt by the language model M and the transmission of the response conversation, the user's characteristics, or the feature prompt to the user terminal 2 are described as being synonymous. The feature prompt is a text indicating the user's portrait and the characteristics of the tone, and is generated based on the user's characteristics extracted from the user's input conversation. The output feature prompt does not necessarily need to be displayed on a screen or the like in a form recognizable by the user, and may be stored in the feature prompt table 221 and used secretly as a prompt for transmission to the language model M. Also, the feature prompt may be output only when a predetermined condition is satisfied, may be output according to the user's request, or may be automatically updated at any time.

[0014] Figure 2 is a block diagram showing an example configuration of user terminal 2. User terminal 2 comprises a terminal processing unit 21, a storage unit 22, a communication unit 23, a display unit 24, and an input unit 25. The terminal processing unit 21 is composed of a CPU (Central Processing Unit), MPU (Micro Processing Unit), GPU (Graphics Processing Unit), NPU (Neural Processing Unit), or TPU (Tensor Processing Unit), and performs various control and calculation processes. The functions of user terminal 2 may be implemented by multiple devices, and may also be devices corresponding to nodes on the blockchain. Furthermore, user terminal 2 may store language model M and perform some or all of the processing executed by language model server 1.

[0015] The storage unit 22 of the user terminal 2 receives input conversation from the user and stores an application program Pa for displaying the input conversation and the response conversation output by the language model M, as well as a feature prompt table 221. The application program Pa is provided to the user terminal 2, for example, using a recording medium 22a. The recording medium 22a is, for example, portable memory. Examples of portable memory include CD-ROM, USB (Universal Serial Bus) memory, SD card, microSD card, or CompactFlash memory (registered trademark). If the recording medium 22a is portable memory, the processing element of the terminal processing unit 21 may read the program Pa from the recording medium 22a using a reading device (not shown). The terminal processing unit 21 of the user terminal 2 may also obtain the application program Pa using the Internet and store it in the storage unit 22. Details of the feature prompt table 221 will be described later.

[0016] The communication unit 23 of the user terminal 2 is a communication module or communication interface for communicating with the language model server 1 wirelessly. The terminal processing unit 21 communicates with the language model server 1 via the communication unit 23 and through the external network N.

[0017] The display unit 24 of the user terminal 2 is composed of a display and displays a screen for receiving input conversation from the user, or a screen showing a dialogue including input conversation from the user and response conversation received from the language model server 1.

[0018] The input unit 25 of the user terminal 2 is configured, for example, with a keyboard or mouse, and accepts input conversation from the user. The input unit 25 may also accept information input from input devices such as a touch panel, camera, microphone, or various sensors.

[0019] Figure 3 is an explanatory diagram showing the system prompts input to the language model M. The system prompts input to the language model M include instructions to extract user characteristics, including tone of voice, vocabulary, thoughts, values, and emotions, from the user's input conversation (for example, the sentence shown in Figure 3: "Please extract user characteristics, including tone of voice, vocabulary, thoughts, values, and emotions."). The system prompts also include instructions to output a response conversation that mirrors the user based on the user's input conversation (for example, the sentence shown in Figure 3: "Based on the user's input conversation, please output a response conversation that mimics (mirrors) the user."). The system prompts also include sentences that show rules for extracting user characteristics. The sentences that show these rules include instructions to ask the user questions about their values, frequently used vocabulary, personality, beliefs, detailed profile, or current emotions through a response conversation (for example, the sentence shown in Figure 3: "Please ask the user questions about their values, frequently used vocabulary, or current emotions through a response conversation."). The text describing the rule includes instructions to output the level of understanding of the user's characteristics (for example, the text shown in Figure 3: "When outputting the response conversation, please output your level of understanding of the user's characteristics."). The text describing the rule includes instructions to calculate the similarity between the input conversation and the response conversation (for example, the text shown in Figure 3: "Please output the similarity between the input conversation and the response conversation."). The text describing the rule includes an example of the response conversation output (for example, the text shown in Figure 3: "Please output the response conversation as follows (reflecting the characteristics extracted from the user's input conversation). Example output: That's good. I like ○○○○. And I think ○○○○ is important. Now, my next question is, do you have any favorite words or phrases that I often use?"). The text describing the rule includes the characteristic output format and an example output (for example, the text shown in Figure 3: "Please output the level of understanding, similarity, and extracted user characteristics in the following format, for example." and the format of the following 10 lines of text). The text indicating the rule includes instructions that accept corrections to the similarity score output by the user (for example, instructions that output the text "※Please correct the similarity score" as shown in Figure 3).Furthermore, the system prompt includes predetermined conditions for outputting a feature prompt (for example, the sentence shown in Figure 3: "When the similarity score entered (modified) by the user exceeds 80, or when the input conversation entered by the user contains an instruction to output a feature prompt"). The system prompt also includes an instruction to output a feature prompt when the predetermined conditions are met (for example, the sentence shown in Figure 3: "(Please output a feature prompt if the following conditions are met.)"). The trigger for generating the feature prompt may be in the form of an automatic determination or a response to a direct or indirect request such as a button press by the user, and the processing executed at that time may be a separate program or prompt specifically for generating feature prompts. Note that the language model M may be input with past conversation history or external information collected by functions such as RAG (Retrieval-Augmented Generation).

[0020] Figure 4 is an explanatory diagram showing the input and output of the language model M in a dialogue between the subject and the language model M. The language model M is composed of, for example, a GPT (Generative Pretrained Transformer). The language model M outputs a response conversation to the input conversation, or predetermined information, based on instructions indicated by pre-entered system prompts (see Figure 3).

[0021] After receiving input from a system prompt, the language model M accepts input from the user's conversation. Following the instructions contained in the system prompt, the language model M outputs the response conversation and extracted features. Figure 4 shows the first input conversation and response conversation (first dialogue) and the second input conversation and response conversation (second dialogue).

[0022] The initial input conversation includes instructions such as mirroring the user's response. The language model M outputs the mirrored response conversation and features extracted from the input conversation in the format specified by the system prompt. The output including the initial response conversation (initial output) includes, for example, a response conversation that asks about the user's values, and user features extracted from the initial input conversation. In the initial output, the language model M also outputs comprehension and similarity. Comprehension is calculated, for example, by the amount of information in the extracted features. Similarity indicates the similarity between the initial response conversation and the initial input conversation, calculated, for example, by cosine similarity or Euclidean similarity. The output of the language model M also includes a prompt for the user to rate (correct) the similarity.

[0023] The second input conversation includes, for example, the score obtained by scoring (correcting) the similarity in the first output, and the user's answers to the questions included in the first response conversation. When the second input conversation is input, the language model M outputs a response conversation that mirrors the input conversation, and features extracted from the input conversation, in the format specified by the system prompt, just as it did when the first input conversation was input. The second response conversation is generated by mirroring the first and second input conversations. The user features included in the second output are features extracted from the first and second input conversations. In other words, the language model M improves itself by outputting response conversations to input conversations and repeating the dialogue with the user, thereby increasing the accuracy of mirroring the user. The level of understanding output with the response conversation for the second and subsequent input conversations is calculated, for example, by the amount of information in the features extracted from all input conversations input so far. The similarity output with the response conversation for the second and subsequent input conversations indicates the similarity between the immediately preceding input conversation and the response conversation output for that input conversation. The similarity score may be the average of the similarity scores between each of the previously entered input conversations and the most recent response conversation, or it may be the similarity score between all of the historical input conversations and the most recent response conversation.

[0024] Figure 5 is an explanatory diagram showing the output of a feature prompt by the language model M. When the input conversation entered by the user satisfies predetermined conditions, the language model M outputs a feature prompt based on the extracted features. The predetermined conditions include, for example, that the input conversation contains a statement that the similarity of the previous response conversation by the language model M is equal to or greater than a predetermined threshold (e.g., 80 points) (the input conversation on the left in Figure 5), that there was a request from the user such as pressing a button, or that there is an instruction to output a feature prompt (the input conversation on the right in Figure 5). The language model M may automatically continue to update the latest feature prompt until the predetermined conditions are met, or it may output a feature prompt when the number of times the input conversation has been entered from the user terminal 2 exceeds a predetermined number, indicating that the predetermined conditions have been met. In addition, the terminal processing unit 21 of the user terminal 2 may determine whether the features extracted based on the input conversation entered by the user satisfy the predetermined conditions, and if it determines that the predetermined conditions are met, it may send information to the language model M instructing it to output a feature prompt. Based on pre-entered system prompts, when input conversation that meets predetermined conditions is received, the language model M outputs a feature prompt based on features extracted from multiple previous input conversations. The feature prompt includes features of the user's personality (thoughts, character, values, memory, knowledge, or mood) and features of the user's tone (tone, vocabulary, or personal pronouns). The feature prompt may also include non-verbal features (appearance, facial expressions, tone of voice, conversational tempo, actions, gestures, or behavior). In other words, the feature prompt is output based on the user's features output along with the previous response conversation. The feature prompt is a prompt to cause the language model M to output a sentence that mimics the user. That is, when a feature prompt is received, the language model M outputs a sentence that mimics the user (see Figure 6). The feature prompt only needs to include information relating to at least one item from the features of the user's personality, including thoughts, character, values, or mood, and may include information relating to multiple items.Furthermore, a feature prompt only needs to include information relating to at least one item from among the user's tone of voice, vocabulary, memory, knowledge, thoughts, personality, values, facial expressions, tone of voice, conversational tempo, actions, gestures, behavior, or tone of voice characteristics including personal pronouns. It may also include information relating to multiple items. Moreover, a feature prompt does not necessarily need to be displayed on a screen or elsewhere in a format recognizable by the user. It may be stored in the feature prompt table 221 and operated while being kept confidential as a prompt to be sent to the language model.

[0025] Figure 5 shows two examples of input conversations that satisfy predetermined conditions for language model M to output a feature prompt. The input conversation shown on the left includes a statement that the similarity of the previous response conversation is above a predetermined threshold. The input conversation shown on the right includes an instruction to output a feature prompt. When language model M receives input conversation that satisfies the predetermined conditions, it outputs a feature prompt that expresses the features extracted from previous input conversations in multiple sentences. The feature prompt may also include user features extracted from the input conversation that satisfies the conditions. Language model M may also output a feature prompt when a feature prompt generation button is pressed or when the calculated comprehension level is above a predetermined value. Furthermore, the predetermined threshold for similarity or the predetermined value for comprehension related to the predetermined conditions for language model M to output a feature prompt may be changed as appropriate by including an instruction to change the value in the system prompt or input conversation.

[0026] Figure 6 is an explanatory diagram showing an example of the display of dialogue and feature prompts in user terminal 2. The terminal processing unit 21 of user terminal 2 displays the input conversation received, the response conversation output by the language model M, and the feature prompts in chat format on the display unit 24. These outputs may also be represented by voice, video, 3D models, robot movements, etc.

[0027] The terminal processing unit 21 of the user terminal 2 displays the input conversation entered by the user in the input unit 25 on the display unit 24, and also sends the input conversation to the language model server 1 for input to the language model M. The language model M outputs a response conversation that mirrors the user's input conversation in the format instructed by the system prompt, and user features extracted from the input conversation. The terminal processing unit 21 displays the user's input conversation and the response conversation output by the language model M to the input conversation alternately, thereby displaying the dialogue including the input conversation and the response conversation in a chat format. Furthermore, if the input conversation input to the language model M satisfies predetermined conditions (i.e., the input conversation includes a statement that the similarity to the previous response conversation by the language model M is above a predetermined threshold, or includes an instruction to output a feature prompt), the language model M outputs a feature prompt that expresses the features extracted from the user's input conversation in text. In the example shown in Figure 6, the case where the input conversation includes an instruction to output a feature prompt is shown. The feature prompt is displayed next to (below) the input conversation that satisfies the predetermined conditions. The terminal processing unit 21 can store (copy) the feature prompt in the feature prompt table 221 of the storage unit 22 by accepting the selection of a copy command displayed along with the output feature prompt. In addition, as shown in Figure 6, a model generation command instructing model generation may be displayed. If a model generation command is displayed, the terminal processing unit 21 inputs the output feature prompt into the language model M and generates a language model based on the feature prompt.

[0028] Figure 7 is an explanatory diagram showing an example of a feature prompt table 221. The feature prompt table 221 stores feature prompts output by the language model M. The management items (fields) of the feature prompt table 221 include, for example, a user ID field and a feature prompt field. The user ID field stores the user ID assigned to the user from whom features have been extracted through interaction with the language model M. That is, in user terminal 2, multiple users can each input conversations and obtain their respective feature prompts through interaction with the language model M. The feature prompt field stores the feature prompts output by the language model M based on the features extracted from the input conversations of the users corresponding to each user ID. The feature prompt table 221 may also store the history of conversations between each user and the language model M. If the target is a character appearing in a work or an AI conversation bot character, the feature prompts may be associated with the character ID and stored (recorded) in the feature prompt table 221. A test environment may be provided to test the behavior of the generated feature prompts, or it may be provided as a production bot service as is. Furthermore, the feature prompt may be associated with the similarity of the response conversation immediately preceding the output of the feature prompt by the language model M to the input conversation, and may be stored (recorded) in the feature prompt table 221 along with the user ID or character ID. At this time, the feature prompt and the similarity associated with the feature prompt may be read from the feature prompt table 221 and used for tuning (retraining) the language model M. That is, the terminal processing unit 21 reads the feature prompt that was temporarily stored due to the interruption process at a later date based on the user ID. The terminal processing unit 21 then provides the read feature prompt to the language model M along with the user's new input conversation and system prompts, and may continue the regeneration of the interrupted feature prompt until the similarity reaches a predetermined value or higher.

[0029] Figure 8 is a flowchart showing the feature prompt output processing by the terminal processing unit 21 of the user terminal 2. The terminal processing unit 21 of the user terminal 2 receives input of a system prompt (S1). The terminal processing unit 21 inputs the system prompt to the language model M (sends it to the language model server 1) (S2). The terminal processing unit 21 receives input of the input conversation (S3). The terminal processing unit 21 inputs the input conversation to the language model M (sends it to the language model server 1) (S4). The terminal processing unit 21 acquires the output of the language model M (S5). The terminal processing unit 21 determines whether the output of the language model M is a feature prompt (whether it is a response conversation) (S6). If the output of the language model M is a feature prompt (S6: YES), the terminal processing unit 21 displays the acquired feature prompt on the display unit 24 (S7) and terminates the process. If the output of language model M is not a feature prompt (S6:NO), that is, if the output of language model M is a response conversation, the terminal processing unit 21 displays the acquired response conversation on the display unit 24 (S8) and returns the processing to S3. In addition, system prompts and programs for processing to output feature prompts based on the conversation history with the user and reference information may be implemented and called as a separate system independent of the system prompts that perform mirroring and similarity measurement to extract the features of the target.

[0030] Figure 9 is an explanatory diagram showing the output of mimicked text by the language model M. When the terminal processing unit 21 of the user terminal 2 receives a feature prompt output from the language model M, as well as a query specifying the role of the language model M and the content of the text to be output, it mimics the user based on the input feature prompt and outputs text with the content specified in the query.

[0031] As shown in Figure 9, the language model M receives a feature prompt indicating the user's characteristics, which is read from the memory unit 22 on the user terminal 2. The feature prompt input to the language model M includes the sentence, "You are a user with the input characteristics. Please output a sentence as if you were that user," as an instruction to output a sentence that mimics the user indicated by the feature prompt. In the example shown in Figure 9, the sentence "Please introduce yourself," which specifies the content of the sentence to be output, is also input as a query. When such a feature prompt and query are input, the language model M outputs a sentence (imitation sentence) that mimics the user's speech characteristics (tone, vocabulary, or personal pronouns) based on the user's characteristics included in the feature prompt. The language model M also reflects the characteristics of the user's personality (thoughts, personality, values, or mood) included in the feature prompt into the content of the sentence specified in the query, and outputs the imitation sentence. As a result, the imitation sentence output by the language model M will mimic the user. When generating and outputting the imitation sentence, reference information obtained from past conversation history or methods such as RAG may be reflected in addition to the feature prompt. Furthermore, the instructions for generating and outputting these feature prompts may be contained within an external application rather than a system prompt, or they may be sent in a way that is confidential to the user.

[0032] Figure 10 shows an example of how imitation text is displayed on user terminal 2. The terminal processing unit 21 of user terminal 2 displays the input feature prompts and queries, along with the imitation text, on the display unit 24, for example, in a chat format. Note that the feature prompts and queries may be input on a terminal device other than user terminal 2, and may be sent from that terminal device to the language model server 1 and input to the language model M. In this case, the terminal device may retrieve the imitation text output by the language model M from the language model server 1 and display it. The terminal processing unit 21 of user terminal 2 can send the feature prompts stored in the memory unit 22 to other terminal devices via the network N or portable memory, etc. Furthermore, the display of imitation text is not limited to screen display of text, but may also be represented using images that mimic people, audio, video, 3D avatars, stereoscopic images, dolls, robots, etc.

[0033] Figure 11 is a flowchart showing the process of outputting a mimicked text by the terminal processing unit 21 of the user terminal 2. The terminal processing unit 21 of the user terminal 2 reads a feature prompt from the storage unit 22 (S11) and accepts a query input (S12). The terminal processing unit 21 displays the read feature prompt and the input query on the display unit 24 (S13). The terminal processing unit 21 inputs the read feature prompt and the input query into the language model M (sends them to the language model server 1) (S14). Note that in S14, past user conversation history or reference information may be input into the language model M. The terminal processing unit 21 retrieves the mimicked text output by the language model M from the language model server 1 (S15). The terminal processing unit 21 displays the retrieved mimicked text on the display unit 24 (S16) and terminates the process. Some or all of this information, such as feature prompts, conversation history, reference information, queries, and mimicked text, may or may not be displayed in a way that is visible to the user.

[0034] With the above configuration and processing, it is possible to extract user characteristics and obtain feature prompts based on the extracted characteristics through a dialogue that includes the user's input conversation and the response conversation output by the language model M by mirroring the user's input conversation. Furthermore, by inputting the feature prompt into the language model M, it is possible to output a mimicked sentence that imitates the user. Note that the target of mirroring by the language model M is not limited to the user. As mentioned above, the target of mirroring by the language model M may be a conversation acted out by the user imitating the characteristics of a specific person, or it may be a fictional character or AI conversation bot character appearing in a work, etc. That is, the user terminal 2 may input, for example, a specific person's quotes, journal, profile information, or video, lines of dialogue, settings, scenarios of characters appearing in a work, input conversations entered by the user imitating a character in a work, or input conversations by an AI conversation bot character into the language model M, and the language model M may output a response conversation that mirrors the lines of a character in a work, the input conversation entered by the user imitating a character in a work, or the input conversations by an AI conversation bot character.

[0035] (Embodiment 2) Figure 12 is an explanatory diagram showing an overview of the information processing system S according to Embodiment 2. The information processing system S according to Embodiment 2 includes an information processing device 3. The information processing device 3 is, for example, a server device owned by a company that provides users with imitation text that mirrors the user using a language model M in response to user requests. The information processing device 3 inputs pre-stored system prompts and input conversations entered at the user terminal 2 to the language model M of the language model server 1. The information processing device 3 transmits the response conversations output by the language model M and the user's characteristics to the user terminal 2. The information processing device 3 also stores the characteristic prompts output by the language model M.

[0036] Figure 13 is a block diagram showing an example configuration of the information processing device 3. The information processing device 3 comprises a processing unit 31, a storage unit 32, and a communication unit 33. The information processing device 3 may have its functions implemented by multiple server devices or computers, or it may be a device corresponding to a node on a blockchain. Furthermore, the information processing device 3 may execute some or all of the processing performed by the language model server 1 or the user terminal 2.

[0037] The processing unit 31 of the information processing device 3 is composed of a CPU (Central Processing Unit), MPU (Micro Processing Unit), GPU (Graphics Processing Unit), NPU (Neural Processing Unit), or TPU (Tensor Processing Unit), and performs various control processing, calculation processing, etc. by reading and executing a program P (program product), etc., which is pre-stored in the memory unit 32.

[0038] The storage unit 32 of the information processing device 3 is, for example, a volatile memory and a non-volatile memory. The storage unit 32 stores a program P, system prompt data SP, and a feature prompt table 321. The program P may be provided to the information processing device 3 using a recording medium 32a stored in a readable format by a computer. The recording medium 32a is, for example, a portable memory. Examples of portable memory include CD-ROMs, USB memory, SD cards, microSD cards, or CompactFlash memory (registered trademark). If the recording medium 32a is a portable memory, the processing element of the processing unit 31 may read the program P from the recording medium 32a using a reading device (not shown). The read program P is written to the storage unit 32. Furthermore, the program P may be provided to the information processing device 3 by the communication unit 33 communicating with an external device.

[0039] The system prompt data SP stored in the storage unit 32 of the information processing device 3 is text data of a system prompt similar to the system prompt shown in Figure 3. In the following explanation, reading the system prompt data SP and reading the system prompt are considered synonymous. The storage unit 32 of the information processing device 3 has pre-stored the system prompt data SP. The system prompt data SP may be provided to the information processing device 3 using, for example, a recording medium 32b stored in a readable format by a computer. The recording medium 32b is, for example, portable memory. If the recording medium 32b is portable memory, the processing element of the processing unit 31 may read the program P from the recording medium 32b using a reading device (not shown). The read system prompt data SP is written to the storage unit 32. Furthermore, the system prompt data SP may also be provided to the information processing device 3 by the communication unit 33 communicating with an external device.

[0040] The communication unit 33 of the information processing device 3 is a communication module or communication interface for communicating with other devices such as the language model server 1 or user terminal 2 via wired or wireless connection, and is, for example, a wide-area wireless communication module such as LTE®, 4G, or 5G. The processing unit 31 communicates with the language model server 1 or user terminal 2 via the communication unit 33, for example, through an external network N such as the internet.

[0041] Figure 14 is an explanatory diagram showing an example of a feature prompt table 321. The record layout of the feature prompt table 321 is the same as the feature prompt table 221 (see Figure 7) stored by the user terminal 2 in Embodiment 1. The feature prompt table 321 may include information about past conversation history and reference information such as RAG. The user ID may be the ID corresponding to the user terminal 2 that received the input conversation and sent it to the information processing device 3 in order to output a feature prompt to the language model M. That is, the information processing device 3 acquires input conversations from multiple user terminals 2, and uses the language model M to acquire and store feature prompts that show the personality and speech characteristics of the user who owns each user terminal 2. The processing unit 31 of the information processing device 3 reads the feature prompt corresponding to the user ID specified in the user terminal 2 from the feature prompt table 321, and inputs the read feature prompt to the language model M, causing the language model M to output a mimicked sentence of the user.

[0042] Figure 15 is a flowchart showing an example of feature prompt output processing by the processing unit 31 of the information processing device 3. The processing unit 31 of the information processing device 3 acquires the input conversation entered at the user terminal 2 from the user terminal 2 (S21). The processing unit 31 reads the system prompt from the storage unit 32 (S22). The processing unit 31 inputs the read system prompt to the language model M (sends it to the language model server 1) (S23). The processing unit 31 inputs the input conversation to the language model M (sends it to the language model server 1) (S24). Note that in S24, past user conversation history or reference information may be input to the language model M. The processing unit 31 acquires the response conversation output by the language model M (S25). The processing unit 31 outputs the acquired response conversation to the user terminal 2 (S26). During these processes, history information, reference information, input information from microphones, cameras, sensors, etc., or variable parameters such as the current date and time and information of the logged-in user may be sent to the language model M along with the prompt and query.

[0043] The processing unit 31 calculates the similarity between the response conversation output to the user terminal 2 and the input conversation entered immediately before outputting the response conversation (S27). The processing unit 31 determines whether the calculated similarity between the input conversation and the response conversation is equal to or greater than a predetermined threshold (S28). In addition, the processing unit 31 may determine in S28 whether the similarity output by the language model M or the similarity entered (modified) by the user on the user terminal 2 is equal to or greater than a predetermined threshold.

[0044] If the similarity is above a predetermined threshold (S28: YES), or if a user requests generation, the processing unit 31 generates a feature prompt based on the user features output by the language model M along with the response conversation (S29). The processing unit 31 stores the generated feature prompt in the feature prompt table 321, associating it with the user ID of the user associated with the user terminal 2 into which the input conversation was entered (S30), and terminates processing. If the similarity is below a predetermined threshold (S28: NO), the processing unit 31 returns the process to S21. Note that in processing after the process is returned from S28, the reading of the system prompt related to S22 and the input of the system prompt to the language model M related to S23 may be skipped. Also, if the input conversation entered in S21 includes an instruction to output a feature prompt, the processing unit 31 may generate a feature prompt based on the user features output along with the response conversation output immediately before. Furthermore, feature prompts may be automatically generated, updated, or displayed whenever there is a change in the extracted feature information, even without a user generation request or the fulfillment of predetermined conditions.

[0045] Figure 16 shows an example of the display of a mimicked text on the user terminal 2 according to Embodiment 2. The terminal processing unit 21 of the user terminal 2 according to Embodiment 2 accepts input of a user ID (the user ID of the user to be mimicked by the language model M) and a query (for example, the text "Please introduce yourself." shown in Figure 16) corresponding to a feature prompt for outputting a mimicked text. Some or all of these input contents may be transmitted in a way that is not visible to the user. The processing unit 31 of the information processing device 3 reads the feature prompt corresponding to the input user ID from the feature prompt table 321, and inputs the read feature prompt and the query entered on the user terminal 2 to the language model M, thereby causing the language model M to output a mimicked text.

[0046] The terminal processing unit 21 of the user terminal 2 displays the input user ID and query, along with the imitation text, on the display unit 24, for example, in a chat format. The display of the imitation text is not limited to a text-based screen display; it may also be represented using images of people, audio, video, 3D avatars, stereoscopic images, dolls, robots, etc. The terminal processing unit 21 may also obtain a feature prompt (a feature prompt input to the language model M) corresponding to the input user ID from the information processing device 3 and display it on the same screen as the imitation text. The feature prompt may be provided to the user terminal 2 in advance from the information processing device 3 via communication or recording medium, and input to the language model M in the same way as in Embodiment 1 through operations on the user terminal 2.

[0047] Figure 17 is a flowchart showing the process of outputting a mimicked text by the processing unit 31 of the information processing device 3. The processing unit 31 of the information processing device 3 obtains the user ID and query of the user to be mimicked by the language model M input at the user terminal 2 from the user terminal 2 (S41). The processing unit 31 reads the feature prompt corresponding to the obtained user ID from the feature prompt table 321 (S42). The processing unit 31 inputs the read feature prompt and query to the language model M (S43). Note that in S43, past conversation history or reference information of the user may also be input to the language model M. The processing unit 31 obtains the mimicked text output by the language model M (S44). The processing unit 31 outputs the obtained mimicked text to the user terminal 2 (S45) and terminates the process.

[0048] With the above configuration and processing, the information processing device 3 can extract user characteristics through a dialogue that includes the user's input conversation and the response conversation output by the language model M by mirroring the user's input conversation, obtain feature prompts based on the extracted characteristics, and store feature prompts for each user. Furthermore, the information processing device 3 can output mimicked sentences that mimic the user by inputting feature prompts corresponding to the user into the language model M. In this embodiment, the processing unit 31 of the information processing device 3 may output the language model M's level of understanding of the user based on the input conversation and the response conversation. In this case, for example, when the input conversation or response conversation is input, the processing unit 31 inputs the input conversation or response conversation into a learning model that has been trained to output the feature elements (tone of voice, vocabulary, thoughts, values, or emotions, etc.) contained in the input conversation or response conversation, outputs the feature elements contained in the input conversation or response conversation, and outputs the level of understanding according to the number of elements output from the input conversation or response conversation during the dialogue. Furthermore, the processing unit 31 of the information processing device 3 may input the input conversation and the response conversation into a learning model that has been trained to output the similarity between the input conversation and the response conversation when the input conversation and the response conversation are input, and output the similarity between the input conversation and the response conversation.

[0049] The embodiments disclosed herein should be considered in all respects as illustrative and not restrictive. The technical features described in each embodiment can be combined with each other, and the scope of the present invention is intended to include all modifications within the claims and scope equivalent to the claims. Furthermore, the independent and dependent claims described in the claims can be combined with each other in any combination, regardless of the form of reference. In addition, the claims use a multi-claim format in which claims refer to two or more other claims (multi-claim format), but are not limited to this. They may also be described using a multi-claim format in which at least one multi-claim refers to another multi-claim (multi-multi-claim format). [Explanation of Symbols]

[0050] 1: Language model server 2: User terminal 21: Terminal Processing Unit 22: Storage section 221: Feature Prompt Table 22a: Recording medium 23: Communications Department 24: Display section 25: Input section 3: Information Processing Device 31: Processing Unit 32: Storage section 321: Feature Prompt Table 32a: Recording medium 32b: Recording media 33: Communications Department M: Language model N: Network P: Program Pa: App Program SP: System Prompt Data S: Information Processing System

Claims

1. Based on a dialogue that includes input conversations by the subject and response conversations by a language model, the characteristics of the subject are extracted. If it is determined that the extracted features meet predetermined conditions, a feature prompt based on the target features at the time the predetermined conditions were met will be output. An information processing method that involves having a computer perform a task.

2. When the aforementioned input conversation is received, the characteristics of the target are output along with the aforementioned response conversation. The information processing method according to claim 1.

3. When the aforementioned input conversation is received, the aforementioned response conversation is output. The similarity between the input conversation and the output response conversation is obtained. The information processing method according to claim 1 or 2.

4. If the similarity is equal to or greater than a predetermined threshold, the feature prompt is output. The information processing method according to claim 3.

5. The aforementioned characteristics include the tone of voice, vocabulary, personal pronouns, memories, knowledge, thoughts, personality, values, or mood of the subject. The information processing method according to claim 1 or 2.

6. The language model is given system prompts that instruct it to extract the aforementioned features, output the response conversation that mirrors the subject, and generate the feature prompts. The information processing method according to claim 1 or 2.

7. The system prompt includes rules for extracting the features, and instructions for asking the subject about the similarity between the input conversation and the response conversation. The information processing method according to claim 6.

8. The system prompt includes the predetermined conditions and an instruction to output the feature prompt when the predetermined conditions are met. The information processing method according to claim 6.

9. The system prompt includes an example of a question to the target and a specification of the format for outputting the features. The information processing method according to claim 6.

10. Based on the interaction between the target and the language model, the features of the target extracted are obtained, and if predetermined conditions are met, the feature prompt output based on the features is acquired. Get the user's query, By providing the acquired feature prompts and queries to the language model, a sentence is output. A program that instructs a computer to perform a process.

11. Based on a dialogue that includes input conversations by the subject and response conversations by a language model, the characteristics of the subject are extracted. If it is determined that the extracted features meet predetermined conditions, a feature prompt based on the target features at the time the predetermined conditions were met will be output. A program that instructs a computer to perform a process.

12. Based on a dialogue that includes input conversations by the subject and response conversations by a language model, the characteristics of the subject are extracted. If it is determined that the extracted features meet predetermined conditions, a feature prompt based on the target features at the time the predetermined conditions were met will be output. Processing section An information processing device equipped with the following features.

Images