KR102642981B1 - Method for providing chat-bot service and user terminal therefor - Google Patents

Abstract

A method of providing a chatbot service and a user terminal for the same are disclosed. The user terminal according to the present disclosure obtains a first query about a specific conversation partner from the user, loads characteristic information related to the user, the conversation partner, or the content of the conversation between the conversation partner and the user, and issues a system prompt corresponding to the characteristic information. A user prompt and a system prompt corresponding to the first query may be generated and input into the first language model, and the output of the first language model may be provided as the conversation partner's response to the first query.

Description

Method of providing chatbot service and user terminal for the same {METHOD FOR PROVIDING CHAT-BOT SERVICE AND USER TERMINAL THEREFOR}

This disclosure relates to a method of providing a chatbot service reflecting a person's persona and a user terminal for the same.

Thanks to the computational processing power of computing devices and the development of artificial intelligence technology, chatbot services that simulate and process human conversations have emerged. For modern people who lack people to talk to, solve their questions, or share mutual understanding and empathy with, having someone they can communicate with at any time is very attractive, so chatbot services range from consulting services that handle inquiries about products or goods to virtual characters and It is being used in a variety of ways, from interactive artificial intelligence services with open topics that can form bonds.

In particular, the conversational artificial intelligence service can be said to be a very promising chatbot service in that it can communicate with a large number of users on general topics and can serve as a friend who is awake 24 hours a day. Furthermore, if conversational artificial intelligence services can provide an experience that feels like talking to a specific celebrity or character, demand could grow explosively.

However, since general-purpose artificial intelligence is required for interactive artificial intelligence services, the ability to understand the context contained in the user's input and generate various levels of responses is required. In particular, in order to provide an experience that feels like talking to a specific celebrity or character, it is necessary to faithfully reflect the persona and other characteristics of that celebrity or character in artificial intelligence.

The purpose of this disclosure is to design a prompt that allows the user to follow a conversation format using the characteristics of a specific conversation partner so that various characteristics of a specific celebrity or character can be faithfully reflected in artificial intelligence.

The technical problems to be achieved in the present disclosure are not limited to the technical problems described above, and other technical problems can be inferred from the following embodiments.

According to one aspect of the present disclosure, a method of providing a chatbot service performed on a user terminal includes obtaining a first query about a specific conversation partner from the user; loading characteristic information related to the user, the conversation partner, or content of a conversation between the conversation partner and the user; generating a system prompt corresponding to the characteristic information; inputting a user prompt and a system prompt corresponding to the first query into a first language model; and providing the output of the first language model as the conversation partner's response to the first query.

According to one embodiment, the characteristic information includes first characteristic information related to the user, and the first characteristic information may include information collected when the user signs up for membership in a chatbot service.

According to another embodiment, the characteristic information includes second characteristic information related to the conversation partner, and the second characteristic information includes persona information including conversation style data of the conversation partner and factual relationship of the conversation partner. It may also contain information. In this regard, the factual relationship information may include an embedding vector whose similarity with the first query is greater than or equal to a threshold value among a plurality of embedding vectors corresponding to the factual relationship of the conversation partner.

According to another embodiment, the characteristic information includes second characteristic information related to the conversation partner, and the second characteristic information is collected through web crawling or according to preset agreement conditions. It may be transmitted from the management server of the person you are talking to.

According to another embodiment, the characteristic information includes third characteristic information related to the conversation content, and the third characteristic information includes mid- to long-term memory information extracted and stored in the internal memory of the user terminal from the conversation content, and The conversation content may include short-term memory information stored in the random access memory (RAM) of the user terminal. In relation to this, the mid- to long-term memory information may include first mid- to long-term memory information stored in a JavaScript Object Notation (JSON) file format and second mid- to long-term memory information converted to an embedding vector and stored in a binary (BIN) file format. .

According to another embodiment, the characteristic information includes third characteristic information related to the conversation content, and the third characteristic information is mid- to long-term memory information in which at least a portion of the user's input from the conversation content is extracted and stored. And it may include short-term memory information in which the most recent N conversations (where N is a natural number) among the conversation contents are stored.

Meanwhile, according to one embodiment, the step of generating the system prompt includes inputting each element constituting the characteristic information as an argument of a predefined system prompt template; and generating the system prompt by applying a concatenation function to the output value of the system prompt template.

Meanwhile, according to one embodiment, the providing step includes, when the output of the first language model includes at least some of the prompt information defined for input to the first language model, the output of the first language model It may include filtering the prompt information included in .

Meanwhile, according to one embodiment, the chatbot service providing method may further include storing the output of the first language model in the random access memory of the user terminal.

Meanwhile, according to another embodiment, the chatbot service providing method includes inputting the first query into a second language model for extracting mid- to long-term memory information to be stored in the internal memory of the user terminal; And when the second language model extracts mid- to long-term memory information from the first query, it may further include storing the extracted mid- to long-term memory information in the internal memory.

Meanwhile, according to another embodiment, the chatbot service providing method includes: receiving a request for deletion of at least a portion of mid- to long-term memory information stored in the internal memory of the user terminal from the user; And it may further include deleting the mid- to long-term storage information corresponding to the deletion request from the internal memory.

Meanwhile, according to another embodiment, the chatbot service providing method includes providing the output of the first language model as a response from the conversation partner, and updating the number of chatbot service availability corresponding to the user's billing status. It may also include more.

According to another aspect of the present disclosure, a method of providing a chatbot service performed on a user terminal includes obtaining a first query about a specific conversation partner from the user; In addition to determining whether the user can currently use the chatbot service based on the user's billing status, if the user cannot use the chatbot service, informing that a response to the first query cannot be output. Providing a message, and when the user can use the chatbot service: loading characteristic information related to the user, the conversation partner, or the contents of a conversation between the conversation partner and the user; generating a system prompt corresponding to the characteristic information; inputting a user prompt and a system prompt corresponding to the first query into a first language model; and providing the output of the first language model as the conversation partner's response to the first query.

According to another aspect of the present disclosure, a method of providing a chatbot service performed on a user terminal includes obtaining a first query about a specific conversation partner from the user; In addition to the step of determining whether or not to input the first query based on the contents of the first query, if the first query is determined to be impossible to input, inform that a response to the first query cannot be output. providing a message that says, and when it is determined that the first query is inputable: loading characteristic information related to the user, the conversation partner, or content of a conversation between the conversation partner and the user; generating a system prompt corresponding to the characteristic information; inputting a user prompt and a system prompt corresponding to the first query into a first language model; and providing the output of the first language model as the conversation partner's response to the first query.

According to one embodiment, the step of determining whether to input the first query refers to a predefined prohibited word dataset, and when the first query includes a prohibited word included in the prohibited word dataset, the first query It may include a step of determining that input is not possible.

According to another embodiment, the step of determining whether or not to input the first query refers to a predefined sensitive data set, and, if the topic corresponding to the first query corresponds to a topic included in the sensitive data set, It may include determining that the first query cannot be entered.

According to another aspect of the present disclosure, a user terminal providing a chatbot service includes one or more processors; and a memory storing one or more instructions to be executed by the one or more processors, wherein the one or more processors, by executing the one or more instructions, obtain a first query for a specific conversation partner from a user, and the user , load characteristic information related to the conversation partner or the contents of a conversation between the conversation partner and the user, generate a system prompt corresponding to the characteristic information, and generate a user prompt and the system prompt corresponding to the first query. 1 input into a language model, and the output of the first language model may be provided as the conversation partner's response to the first query.

Another aspect of the present disclosure may provide a computer-readable recording medium on which a program for implementing a method performed by an electronic device is recorded.

Specific details of other embodiments are included in the detailed description and drawings.

According to the present disclosure, in a chatbot service for having a conversation with a virtual conversation partner modeled after a celebrity, celebrity, or popular character, the personality, habits, facts, memories related to the conversation, etc. of the celebrity, celebrity, or popular character are shared with the conversation partner. By reflecting the response, you can give the conversation partner the impression that the ego of the celebrity, celebrity, or popular character is embodied, and through this, you can provide an experience as if you are talking to a specific celebrity, celebrity, or character.

In addition, according to the present disclosure, it is possible to prevent the know-how of the chatbot service from becoming known or the quality of the chatbot service from being degraded by detecting and blocking in advance attacks that attempt to leak prompts or inject incorrect data from input queries or generated responses. You can.

Additionally, according to the present disclosure, by restricting the user's input of slang or sensitive topics, it is possible to prevent the chatbot service from being excessively used in a socially controversial way.

The effect of the invention is not limited to the effects mentioned above, and other effects not mentioned will be clearly understood by those skilled in the art from the description of the claims.

Figure 1 is a schematic configuration diagram showing the operating environment of a user terminal providing a chatbot service.
Figure 2 is a flowchart of a chatbot service provision method according to the first embodiment.
3 to 6 are flowcharts of a method for providing additional chatbot services related to the first embodiment.
Figure 7 is a flowchart of a chatbot service provision method according to a second embodiment.
Figure 8 is a flowchart of a chatbot service provision method according to a third embodiment.
Figure 9 is an example of the overall architecture for providing chatbot services.
Figure 10 is an example of a concrete architecture that outputs a conversation partner's response through a first language model.
Figures 11a and 11b are examples of screens for selecting a conversation partner in an application to which a chatbot service is applied.
Figures 12a and 12b are examples of conversation screens within an application to which the chatbot service is applied.
Figure 13 shows a block diagram of a user terminal according to one embodiment.

Hereinafter, specific embodiments will be described with reference to the drawings. The detailed description below is provided to provide a comprehensive understanding of the methods, devices and/or systems described herein. However, this is only an example, and the disclosed embodiments are not limited thereto.

In describing the embodiments, if it is determined that a detailed description of the related known technology may unnecessarily obscure the gist of the disclosed embodiments, the detailed description will be omitted. In addition, terms described below are terms defined in consideration of functions in the disclosed embodiments, and may vary depending on the intention or custom of the user or operator. Therefore, the definition should be made based on the contents throughout this specification. The terminology used in the detailed description is only for describing embodiments and should in no way be limiting. Unless explicitly stated otherwise, singular forms include plural meanings. In this description, expressions such as “comprising” or “comprising” are intended to indicate certain features, numbers, steps, operations, elements, parts or combinations thereof, and one or more than those described. It should not be construed to exclude the existence or possibility of any other characteristic, number, step, operation, element, or part or combination thereof.

The terms used in the embodiments are general terms that are currently widely used as much as possible while considering the functions in the present disclosure, but this may vary depending on the intention or precedent of a person working in the art, the emergence of new technology, etc. In addition, in certain cases, there are terms arbitrarily selected by the applicant, and in this case, the meaning will be described in detail in the relevant description. Therefore, the terms used in this disclosure should be defined based on the meaning of the term and the overall content of this disclosure, rather than simply the name of the term.

When it is said that a part "includes" a certain element throughout the specification, this means that, unless specifically stated to the contrary, it does not exclude other elements but may further include other elements. In addition, terms such as “..unit” and “..module” used in the specification refer to a unit that processes at least one function or operation, which may be implemented as hardware or software, or as a combination of hardware and software. And, unlike the example shown, specific operations may not be clearly distinguished.

The expression “at least one of a, b, and c” used throughout the specification means ‘a alone’, ‘b alone’, ‘c alone’, ‘a and b’, ‘a and c’, ‘b and c ', or 'all of a, b, and c'.

The “terminal” mentioned below may be implemented as a computer or portable terminal that can connect to a server or other terminal through a network. Here, the computer includes, for example, a laptop, desktop, laptop, etc. equipped with a web browser, and the portable terminal is, for example, a wireless communication device that guarantees portability and mobility. , all types of communication-based terminals such as IMT (International Mobile Telecommunication), CDMA (Code Division Multiple Access), W-CDMA (W-Code Division Multiple Access), and LTE (Long Term Evolution), smartphones, tablet PCs, etc. It may include a handheld-based wireless communication device.

In the following description, "transmission", "communication", "transmission", "reception" and other similar terms of a signal or information refer not only to the direct transmission of a signal or information from one component to another component. It also includes those transmitted through other components.

In particular, “transmitting” or “transmitting” a signal or information as a component indicates the final destination of the signal or information and does not mean the direct destination. This is the same for “receiving” signals or information. Also, in this specification, “related” to two or more data or information means that if one data (or information) is acquired, at least part of other data (or information) can be obtained based on it.

Additionally, terms such as first and second may be used to describe various components, but the components should not be limited by the terms. The above terms may be used for the purpose of distinguishing one component from another component.

For example, a first component may be referred to as a second component, and similarly, the second component may also be referred to as a first component without departing from the scope of the present invention.

In describing the embodiments, description of technical content that is well known in the technical field to which the present invention belongs and that is not directly related to the present invention will be omitted. This is to convey the gist of the present invention more clearly without obscuring it by omitting unnecessary explanation.

For the same reason, some components are exaggerated, omitted, or schematically shown in the accompanying drawings. Additionally, the size of each component does not entirely reflect its actual size. In each drawing, identical or corresponding components are assigned the same reference numbers.

The advantages and features of the present invention and methods for achieving them will become clear by referring to the embodiments described in detail below along with the accompanying drawings. However, the present invention is not limited to the embodiments disclosed below and may be implemented in various different forms. The present embodiments are merely provided to ensure that the disclosure of the present invention is complete and to provide common knowledge in the technical field to which the present invention pertains. It is provided to fully inform those who have the scope of the invention, and the present invention is only defined by the scope of the claims. Like reference numerals refer to like elements throughout the specification.

It will be understood that each block of the processing flow diagrams and combinations of the flow diagram diagrams may be performed by computer program instructions. These computer program instructions can be mounted on a processor of a general-purpose computer, special-purpose computer, or other programmable data processing equipment, so that the instructions performed through the processor of the computer or other programmable data processing equipment are described in the flow chart block(s). It creates the means to perform functions. These computer program instructions may also be stored in computer-usable or computer-readable memory that can be directed to a computer or other programmable data processing equipment to implement a function in a particular manner, so that the computer-usable or computer-readable memory It is also possible to produce manufactured items containing instruction means that perform the functions described in the flowchart block(s). Computer program instructions can also be mounted on a computer or other programmable data processing equipment, so that a series of operational steps are performed on the computer or other programmable data processing equipment to create a process that is executed by the computer, thereby generating a process that is executed by the computer or other programmable data processing equipment. Instructions that perform processing equipment may also provide steps for executing the functions described in the flow diagram block(s).

Additionally, each block may represent a module, segment, or portion of code that includes one or more executable instructions for executing specified logical function(s). Additionally, it should be noted that in some alternative execution examples it is possible for the functions mentioned in the blocks to occur out of order. For example, it is possible for two blocks shown in succession to be performed substantially at the same time, or it is possible for the blocks to be performed in reverse order depending on the corresponding function.

Below, with reference to the attached drawings, embodiments of the present disclosure will be described in detail so that those skilled in the art can easily practice them. However, the present disclosure may be implemented in many different forms and is not limited to the embodiments described herein. The method shown in FIGS. 2 to 12B may be performed, for example, by the user terminal 110 described with reference to FIG. 1 or FIG. 13 . That is, the user terminal 110 is an exemplary device for providing a chatbot service in the present disclosure, and may be a single device, but depending on the embodiment, it may be comprised of a plurality of detailed devices that communicate with each other through a wired or wireless network. It may be a structured system.

Figure 1 is a schematic configuration diagram showing the operating environment of a user terminal providing a chatbot service.

Figure 2 is a flowchart of a chatbot service provision method according to the first embodiment.

The user terminal 110 may obtain a first query for a specific conversation partner (S210). Next, the user terminal 110 may load characteristic information related to the user, a conversation partner, or the contents of a conversation between the conversation partner and the user (S220). Next, the user terminal 110 may generate a system prompt corresponding to the loaded characteristic information (S230). Next, the user terminal 110 may input the generated system prompt and the user prompt corresponding to the first query into the first language model (S240). Next, the user terminal 110 may provide the output of the first language model as the conversation partner's response to the first query (S250). In the present disclosure, 'user prompt' refers to information converted from the first query to fit the user prompt form to be input into the first language model, and 'system prompt' refers to information (user, conversation partner, etc.) confirmed from the first query. This refers to information converted to fit the system prompt format corresponding to the conversation content (including characteristic information related to the conversation content).

According to one embodiment, the user terminal 110 may load characteristic information from a storage space inside the user terminal 110 or from a server 120 that communicates through the network 130. The characteristic information may include at least some of first characteristic information related to the user, second characteristic information related to the conversation partner, or third characteristic information related to the content of the conversation between the conversation partner and the user.

The first characteristic information related to the user may include information collected when the user signs up for membership in the chatbot service. When signing up for the chatbot service, the user directly enters information such as name, age, date of birth, MBTI, gender, interests, and interested conversation partners according to a pre-implemented form, and the user terminal 110 selects the first of the information collected in this way. 1 You can load useful information to generate your interlocutor's response to your query. Methods for identifying information useful for generating a response among the collected information may be defined in various ways. For example, the user terminal 110 identifies a keyword corresponding to the first query, and determines whether information corresponding to the keyword may exist among the collected information. If so, you can load that information. If a user asks A, “What is your hobby?” while talking with a conversation partner, the user terminal 110 identifies the keyword “hobby,” and among the collected information, information that the user’s interest is “listening to music.” to provide a response like "My hobby is cooking. Do you like listening to music?" The first characteristic information may be stored in the storage space of the user terminal 110 or the storage space of the server 120.

The second characteristic information related to the conversation partner may include at least part of persona information including conversation style data of the conversation partner or factual relationship information of the conversation partner. The second characteristic information may be information collected through web crawling or transmitted from the management server of the conversation partner according to preset agreement conditions. At this time, the user terminal 110 may directly perform web crawling or receive information directly from the management server, but depending on the embodiment, the server 120 performs web crawling and sends the collected second characteristic information to the user terminal ( 110), or the server 120 may first receive the second characteristic information from the management server and transmit it to the user terminal 110.

In this disclosure, 'persona information' is a general term for data regarding the conversational style of the conversational partner, such as the conversationalist's tone of voice, language habits such as words or emoticons used frequently, and text input method. For example, persona information is information included in the actual conversation data set of the conversation partner (in this case, the actual conversation data set may be stored in the storage space inside the user terminal 110 in the form of an embedding vector or in a separate vector storage, For example, Pinecone), the first language model may be an artificial intelligence model learned through a few-shot learning technique using persona information as learning data. Since the actual conversation content of a celebrity or the conversation content within the official content of a virtual character (e.g., regular animation or officially released game, etc.) may be rare, learning through the few-shot learning technique can be effective.

Meanwhile, in this disclosure, the 'factual relationship information' of the interlocutor is publicly disclosed, such as profile information such as the interlocutor's gender and date of birth, information about social relationships between the interlocutor and other third parties, and the interlocutor's activity history. A general term for information about the person you are talking to. The factual relationship information is decomposed based on hyper-parameters (such as chunk size and overlap size, etc.) for which each conversation partner is preset, and then converted to an embedding vector at the user terminal (110). ) It may be stored in an internal storage space or an external vector storage. In relation to this, when loading the factual relationship information, the user terminal 110 compares the similarity with the first query among a plurality of embedding vectors corresponding to the factual relationship of the specified conversation partner and selects an embedding vector whose similarity is greater than or equal to a threshold value. It can be loaded with factual information about the person you are talking to.

The second characteristic information may be stored in the storage space of the user terminal 110 or the storage space of the server 120.

The third characteristic information related to the conversation content between the conversation partner and the user includes mid- to long-term memory information in which at least part of the user's input is extracted and stored from the conversation content, or the most recent N conversation contents (where N is a natural number) among the conversation content. At least some of the stored short-term memory information may be included. Specifically, mid- to long-term memory information may include information corresponding to a preset category (eg, pet, family, occupation, etc.) and information similar to a preset example (eg, the user hates curry). Among these, errors may occur in the process of classifying information similar to preset examples, so the user terminal 110 displays classified information (information classified as mid- to long-term memory information due to similarity to preset examples) through a user interface (User It is provided to the user through the Interface (UI), and when the user requests deletion through the UI, it can be deleted without storing it as mid- to long-term memory information. If the user terminal 110 determines that there is no content to be extracted as mid- to long-term memory information from a series of conversation contents, it may store only short-term memory information as third characteristic information. Similarly, even when loading, the user terminal 110 may load only short-term storage information if there is no content to load among the mid- to long-term storage information.

If mid- to long-term memory information and short-term memory information are classified according to where they are stored, mid- to long-term memory information may be stored in the built-in memory of the user terminal 110, and short-term memory information may be stored in the random access memory of the user terminal 110. Memory, RAM). By storing short-term memory information in RAM, when a query that is highly related to the first query is entered within a short period of time (such as when a user enters similar questions in succession), short-term memory information is quickly loaded from RAM to generate a response. Latency can be reduced. The variable N, which sets the number of conversation contents stored as short-term memory information, may be a hyperparameter set by the chatbot service provider. The chatbot service provider can set N by considering the token usage and performance of the answer of the first language model. On the other hand, mid- to long-term memory information is information that needs to be stored for a relatively long time, so it is not appropriate to store it in highly volatile RAM, and it is not suitable to store it in the server 120 as the number of users using the chatbot service increases. There is a risk that the amount of mid- to long-term memory information will rapidly increase, and there will be a significant delay in retrieving and loading mid- to long-term memory information related to conversation content with a specific conversation partner from the server 120, so it is stored in the internal memory of the user terminal 110. It may be appropriate to do so.

Meanwhile, if mid- to long-term memory information is formally classified, the mid- to long-term memory information includes first mid- to long-term memory information stored in JavaScript Object Notation (JSON) file format, or second mid- to long-term memory converted to an embedding vector and stored in binary (BIN) file format. At least some of the information may be included. For example, all mid- to long-term memory information is stored in the internal memory in a JSON file format as first mid- to long-term memory information, and among them, information that needs to be stored longer as long-term memory information is converted to an embedding vector as second mid- to long-term memory information and is converted to a BIN file. It can also be additionally saved in this format. However, in this specification, the storage format of the first mid- to long-term memory information is not limited to the JSON file format, and may be replaced with any file format stored as a dictionary in the form of a key-value. Meanwhile, when the user terminal 110 loads the second mid- to long-term memory information stored as an embedding vector, the similarity with the first query among the second mid- to long-term memory information corresponding to the specified conversation partner is compared, similar to the above-described factual relationship information. Thus, an embedding vector with a similarity value greater than or equal to a threshold can be loaded into second mid- to long-term memory information to be used to generate a response from a specified conversation partner.

According to one embodiment, in step S230, the user terminal 110 inputs each element constituting the characteristic information as an argument of a predefined system prompt template and applies a concatenation function to the output value of the system prompt template. You can create a system prompt by doing this.

According to one embodiment, in step S250, when the output of the first language model includes at least some of the prompt information defined for input to the first language model, the user terminal 110 outputs the first language model. You can filter the prompt information contained in . For example, the user terminal 110 calculates the similarity between the output of the first language model and predefined prompt information, and if the similarity is greater than a threshold, blocks the output of the first language model itself or outputs the first language model. Parts that are judged to be similar to the prompt information can be blocked or replaced with other expressions.

According to another embodiment, the user terminal 110, after step S210 and before performing step S220, predefined prompt-related characters in the obtained first query (e.g., "system", "role", which are often used in prompt attacks). ", <>, [], {}, etc.) are included, and if prompt-related characters are included, a response to the first query may not be generated. In this case, the user terminal 110 may provide a message informing that a response to the first query cannot be generated, or may provide a message requesting that a query with different content be input. Meanwhile, the database in which prompt-related characters are defined may be stored in the user terminal 110 or the server 120, but is not limited to this and may be stored in a separate storage medium.

3 to 6 are flowcharts of a method for providing additional chatbot services related to the first embodiment. For convenience of explanation, overlapping descriptions of steps identical or similar to each step constituting the method of FIG. 2 will be omitted in each drawing.

Figure 3 explains an embodiment of storing a conversation partner's response. In addition to providing the output of the first language model as the conversation partner's response to the user's first query, the user terminal 110 may store the output of the first language model in the random access memory of the user terminal 110. (S360). As described above, when storing N recent conversation contents as short-term memory information, it is necessary to store the most recent conversation (response from the recently created conversation partner) even if N is set to the minimum. In addition, the user terminal 110 may store the output of the first language model in the random access memory and at the same time delete the first stored conversation content among the N conversation content previously stored in the random access memory.

Figure 4 explains an example of extracting and storing information that needs to be stored in the mid to long term from the first query. The user terminal 110 uses the first language model to extract mid- to long-term memory information to be stored in the internal memory of the user terminal 110 from the first query, in addition to providing the conversation partner's response to the first query. The first query may be input into the second language model (S460). Thereafter, when the second language model extracts mid- to long-term memory information from the first query, the user terminal 110 may store the extracted mid- to long-term memory information in the internal memory of the user terminal 110 (S470). In other words, one query (first query) entered by the user is input to the first language model and considered when generating the conversation partner's response, while at the same time being independently input to the second language model, it contains mid- to long-term memory information to be considered in subsequent conversations. It can also be used as a source. The second language model performs the same or similar functions as described above regarding mid- to long-term memory information with reference to FIG. 2.

Figure 5 explains an embodiment of deleting mid- to long-term stored information. The user terminal 110 may receive a request from the user to delete at least some of the mid- to long-term memory information stored in the internal memory of the user terminal 110 (S560). Thereafter, the user terminal 110 may delete the mid- to long-term memory information corresponding to the deletion request from the internal memory of the user terminal 110 (S570). The user terminal 110 may inform the user of what mid- to long-term memory information is stored in the internal memory through a preset UI and may receive a deletion request from the user. For example, mid- to long-term memory information is provided in the form of a list on a specific page within the chatbot service application of the user terminal 110, and the selected mid- to long-term memory information can be deleted by selecting a specific list element and pressing the deletion request icon. . As another example, whenever the user terminal 110 stores mid- to long-term memory information, a push alarm or pop-up message is output, and the user selects a specific push alarm or pop-up message and selects a specific push alarm or pop-up message corresponding to the push alarm or pop-up message on the moving page. You may also request deletion of mid- to long-term memory information. However, the form of the UI that guides mid- to long-term memory information and receives deletion requests is not limited to the forms mentioned above.

Figure 6 explains an embodiment of updating the user's chatbot service usage status as the conversation partner's response is provided to the user. The user terminal 110 may update the chatbot service availability count corresponding to the user's billing status by providing the output of the first language model as the conversation partner's response to the user's first query (S660).

The charging system for chatbot services may be designed to limit usage time or limit the number of conversations. If the charging system is designed to limit the number of conversations (queries) that a user can input, the user terminal 110 will consider that the user has legitimately used one conversation opportunity when a response to the user's query is provided, and , the most recent ‘number of conversations that can be entered’ considering the user’s billing status can be deducted once. Setting whether or not the conversation partner's response was provided as an update standard is unfair to deduct the 'number of conversations that can be entered' even if the user did not obtain a normal response, and incorrect deduction can worsen the user experience. Considering this, renewal standards can be set in various ways within the scope that can achieve this purpose.

Figure 7 is a flowchart of a chatbot service provision method according to a second embodiment.

The user terminal 110 may obtain a first query for a specific conversation partner (S710).

Next, the user terminal 110 may determine whether the user can currently use the chatbot service based on the user's billing status (S720). For example, considering the user's billing status, the user terminal 110 may determine that the user's chatbot service is currently unavailable when the user's conversation period has expired or all available conversation numbers have been exhausted.

As a result of the determination, if the user can use the chatbot service, the user terminal 110 loads characteristic information related to the user, the conversation partner, or the content of the conversation between the conversation partner and the user (S730), and generates a system prompt corresponding to the characteristic information. (S740), the system prompt and the user prompt corresponding to the first query are input into the first language model (S750), and the output of the first language model can be provided as the conversation partner's response to the first query (S750). S760). Meanwhile, if the user cannot use the chatbot service as a result of the determination, the user terminal 110 may provide a message informing that the response to the first query cannot be output (S770).

Figure 8 is a flowchart of a chatbot service provision method according to a third embodiment.

The user terminal 110 may obtain a first query for a specific conversation partner (S810).

Next, the user terminal 110 may determine whether to input the first query based on the contents of the first query (S820).

As a result of the determination, if the first query can be input, the user terminal 110 loads characteristic information related to the user, a conversation partner, or the contents of the conversation between the user and the conversation partner (S830), and issues a system prompt corresponding to the loaded characteristic information. Generate (S840), input the system prompt and the user prompt corresponding to the first query into the first language model (S850), and provide the output of the first language model as the conversation partner's response to the first query. (S860). Meanwhile, if the determination results indicate that input of the first query is not possible, the user terminal 110 may provide a message informing that a response to the first query cannot be output (S870).

According to one embodiment, in step S820, the user terminal 110 refers to a predefined prohibited word dataset and determines that the first query cannot be input if the first query includes a prohibited word included in the prohibited word dataset. You can judge. Furthermore, the user terminal 110 determines whether the first query contains a word that is the same as the banned word in the banned word dataset by a certain number of characters or more, or corresponds to a banned word and a synonym included in the banned word dataset based on an external dictionary. If a word is included, it may be determined that the first query cannot be entered. Although not shown in FIG. 8, if the first query includes a prohibited word, the user terminal 110 may block the prohibited word or replace it with a synonym and use it as an input.

According to another embodiment, in step S820, the user terminal 110 refers to a predefined sensitive dataset and inputs the first query if the subject corresponding to the first query corresponds to a subject included in the sensitive dataset. It may be judged as impossible. For example, the user terminal 110 may identify the topic of the first query based on at least some of the keywords included in the first query and check whether the identified topic exists in the topic list included in the sensitive dataset. If the identified topic exists in the topic list included in the sensitive dataset (eg, politics, gender, class, etc.), the user terminal 110 may determine that the first query cannot be entered.

Figure 9 is an example of the overall architecture for providing chatbot services.

As shown, as the first query is input, the user terminal 110 outputs the conversation partner's response through the first language model and provides it to the user (client), and separately, the conversation partner's response is output to the user terminal. It can be stored in RAM at (110). Meanwhile, as the first query is input, the user terminal 110 checks whether mid- to long-term memory information to be extracted from the first query exists through the second language model, and if so, extracts the mid- to long-term memory information and formats it into the built-in memory in key-value format. However, information to be stored for a longer period of time can be converted to an embedding vector and stored in the internal memory.

In addition, the user terminal 110 reflects the short-term memory information stored in the RAM and the mid- to long-term memory information stored in the internal memory in the characteristic information, and then outputs the conversation partner's response through the first language model when the second query is input. This allows the updated characteristic information to be utilized. The process of updating characteristic information through short-term memory information and mid- to long-term memory information is not only performed in the relationship between the first query and the second query, and will also be performed similarly when additional queries such as the third query and fourth query are entered in the future. You can.

Figure 10 is an example of a concrete architecture that outputs a conversation partner's response through a first language model.

As shown, when a query is entered, it can be converted into a user prompt according to a predefined user prompt format. Additionally, in consideration of the query, third characteristic information including at least part of short-term memory information or mid- to long-term memory information may be loaded from the RAM or internal memory of the user terminal 110, and may be loaded from the user terminal 110 or the server 120. First characteristic information including user-related information, second characteristic information including at least part of persona information, or factual relationship information may be loaded, and the loaded characteristic information is converted into a system prompt to fit a predefined system prompt format. It can be.

The first language model can generate a response from a conversation partner selected by the user using a system prompt and a user prompt as input.

Figures 11a and 11b are examples of screens for selecting a conversation partner in an application to which a chatbot service is applied. By way of example, the corresponding screen may be provided through the output interface of the user terminal 110.

The screen in FIG. 11A is a screen that displays a profile list of conversation partners that the user has added in advance, and the user can open a chat room with desired conversation partners or exclude profiles of unwanted conversation partners from the list. In addition, users can set specific keywords for each conversation partner, and the set keywords are used as a factor to determine the characteristic information loaded when creating a response from a conversation partner, or to set the basic conversation topic when starting a conversation with a conversation partner. Can be used as an element.

The screen of FIG. 11B is a screen that displays a list of chat rooms in which the user is currently having a conversation. Each chat room list may display at least some of the conversation partner's profile picture, name, recent conversation content, or recent conversation date and time. The user can enter the chat room by selecting a specific chat room from the chat room list and continue the conversation in the entered chat room (conversation screen).

Figures 12a and 12b are examples of conversation screens within an application to which the chatbot service is applied. By way of example, the corresponding screen may be provided through the output interface of the user terminal 110.

As shown in FIG. 12A, the conversation (query) entered by the user and the responses of the conversation partner can be sorted by date and time, and the date and time of the input may be displayed around each conversation (or response). Additionally, a sound interface may be displayed around the conversation partner's response, and when the user selects the sound interface, the response may be provided in the form of voice output. The voice may be set for each conversation partner, or may be provided as a default voice.

Meanwhile, on the conversation screen, a graphic object that visually represents the conversation partner in two or three dimensions may be displayed behind the dialog box. The graphic object may change its expression or pose adaptively depending on the conversation content. Exemplarily, as shown in FIG. 12b, when words representing emotions (e.g., 'happy', 'happy', etc.) are included in the conversation content, the graphic object displays a pose set corresponding to the emotion (e.g., 'heart pose') ') can be taken.

Figure 13 shows a block diagram of a user terminal according to one embodiment.

According to one embodiment, the user terminal 110 may include a processor 111 and a memory 113. In one embodiment, the user terminal 110 may exchange data with the outside world through an input/output interface.

The processor 111 may perform at least one method described above through FIGS. 1 to 12B. The memory 113 may store information for performing at least one method described above with reference to FIGS. 1 to 12B. Memory 113 may be volatile memory or non-volatile memory.

The processor 111 can control the user terminal 110 to execute programs and provide information. The code of the program executed by the processor 111 may be stored in the memory 113.

In one embodiment, the processor 111 is coupled with the memory 113 to obtain a first query for a specific conversation partner from the user and to obtain characteristic information related to the user, the conversation partner, or the content of the conversation between the conversation partner and the user. load, generate a system prompt corresponding to the feature information, input a user prompt and a system prompt corresponding to the first query into the first language model, and output the output of the first language model as the interlocutor's response to the first query. It can be provided as

Additionally, the user terminal 110 according to one embodiment may further include an interface that can provide information to the user of the user terminal 110.

The user terminal 110 shown in FIG. 13 shows only components related to this embodiment. Accordingly, those skilled in the art can understand that other general-purpose components may be included in addition to the components shown in FIG. 13.

Devices according to the above-described embodiments include a processor, memory for storing and executing program data, permanent storage such as a disk drive, a communication port for communicating with an external device, a touch panel, keys, buttons, etc. It may include a user interface device, etc. Methods implemented as software modules or algorithms may be stored on a computer-readable recording medium as computer-readable codes or program instructions executable on the processor. Here, computer-readable recording media include magnetic storage media (e.g., ROM (read-only memory), RAM (random-access memory), floppy disk, hard disk, etc.) and optical read media (e.g., CD-ROM). ), DVD (Digital Versatile Disc), etc. The computer-readable recording medium is distributed among computer systems connected to a network, so that computer-readable code can be stored and executed in a distributed manner. The media may be readable by a computer, stored in memory, and executed by a processor.

This embodiment can be represented by functional block configurations and various processing steps. These functional blocks may be implemented in various numbers of hardware or/and software configurations that execute specific functions. For example, embodiments include integrated circuit configurations such as memory, processing, logic, look-up tables, etc. that can execute various functions under the control of one or more microprocessors or other control devices. can be hired. Similar to how the components can be implemented as software programming or software elements, the present embodiments include various algorithms implemented as combinations of data structures, processes, routines or other programming constructs, such as C, C++, Java ( It can be implemented in a programming or scripting language such as Java), assembler, etc. Functional aspects may be implemented as algorithms running on one or more processors. Additionally, this embodiment may employ conventional technologies for electronic environment setting, signal processing, message processing, and/or data processing. Terms such as “mechanism,” “element,” “means,” and “composition” can be used broadly and are not limited to mechanical and physical components. The term may include the meaning of a series of software routines in connection with a processor, etc.

The above-described embodiments are merely examples and other embodiments may be implemented within the scope of the claims described below.

Claims (20)

In the method of providing chatbot services performed on a user terminal,
Obtaining a first query for a specific conversation partner from the user;
loading characteristic information related to the user, the conversation partner, or content of a conversation between the conversation partner and the user;
generating a user prompt corresponding to the first query;
generating a system prompt corresponding to the characteristic information;
inputting the user prompt and the system prompt into a first language model, respectively;
providing an output of the first language model based on the user prompt and the system prompt as the conversation partner's response to the first query;
storing the output of the first language model in a random access memory of the user terminal;
Inputting the first query into a second language model for extracting mid- to long-term memory information to be stored in the internal memory of the user terminal; and
When the second language model extracts mid- to long-term memory information from the first query, storing the extracted mid- to long-term memory information in the internal memory.
A method of providing chatbot services, including. According to claim 1,
The above characteristic information is,
Contains first characteristic information related to the user,
The first characteristic information is,
A method of providing a chatbot service, including information collected when the user signs up for chatbot service membership. According to claim 1,
The above characteristic information is,
Contains second characteristic information related to the conversation partner,
The second characteristic information is,
A method of providing a chatbot service, including persona information including conversation style data of the conversation partner and factual relationship information of the conversation partner. According to clause 3,
The above factual information is,
A method of providing a chatbot service, including an embedding vector whose similarity to the first query is greater than or equal to a threshold value among a plurality of embedding vectors corresponding to the factual relationship of the conversation partner. According to claim 1,
The above characteristic information is,
Contains second characteristic information related to the conversation partner,
The second characteristic information is,
A method of providing a chatbot service that is collected through web crawling or transmitted from the management server of the conversation partner according to preset agreement conditions. According to claim 1,
The above characteristic information is,
Contains third characteristic information related to the conversation content,
The third characteristic information is,
A chatbot service providing method comprising mid- to long-term memory information extracted and stored in the internal memory of the user terminal among the conversation content and short-term memory information stored in the random access memory (RAM) of the user terminal among the conversation content. According to clause 6,
The mid- to long-term memory information is,
A method of providing a chatbot service, including first mid- to long-term memory information stored in a JavaScript Object Notation (JSON) file format and second mid- to long-term memory information converted to an embedding vector and stored in a binary (BIN) file format. According to claim 1,
The above characteristic information is,
Contains third characteristic information related to the conversation content,
The third characteristic information is,
A method of providing a chatbot service, including mid- to long-term memory information in which at least a portion of the user's input is extracted and stored among the conversation content, and short-term memory information in which the most recent N conversations (where N is a natural number) among the conversation content are stored. According to claim 1,
The step of creating the system prompt is,
Inputting each element constituting the characteristic information as an argument of a predefined system prompt template; and
A chatbot service providing method comprising generating the system prompt by applying a concatenation function to the output value of the system prompt template. delete delete According to claim 1,
Receiving a request for deletion of at least some of the mid- to long-term memory information stored in the internal memory of the user terminal from the user; and
A method of providing a chatbot service, further comprising deleting the mid- to long-term memory information corresponding to the deletion request from the internal memory. According to claim 1,
The steps provided above are:
When the output of the first language model includes at least some of the prompt information defined for input to the first language model, filtering the prompt information included in the output of the first language model, How to provide chatbot services. According to claim 1,
A method of providing a chatbot service, further comprising updating the number of times a chatbot service can be used corresponding to the user's billing status by providing the output of the first language model as a response from the interlocutor. In the method of providing chatbot services performed on a user terminal,
Obtaining a first query for a specific conversation partner from the user;
Based on the user's billing status, determining whether the user can currently use the chatbot service;
If the user is unable to use the chatbot service, providing a message informing that a response to the first query cannot be output;
When the above user can use the chatbot service:
generating a user prompt corresponding to the first query;
loading characteristic information related to the user, the conversation partner, or content of a conversation between the conversation partner and the user;
generating a system prompt corresponding to the characteristic information;
inputting the user prompt and the system prompt into a first language model, respectively;
providing an output of the first language model based on the user prompt and the system prompt as the conversation partner's response to the first query;
storing the output of the first language model in a random access memory of the user terminal;
Inputting the first query into a second language model for extracting mid- to long-term memory information to be stored in the internal memory of the user terminal; and
When the second language model extracts mid- to long-term memory information from the first query, storing the extracted mid- to long-term memory information in the internal memory.
A method of providing chatbot services, including. In the method of providing chatbot services performed on a user terminal,
Obtaining a first query for a specific conversation partner from the user;
determining whether to input the first query based on the contents of the first query;
If it is determined that the first query cannot be input, providing a message informing that a response to the first query cannot be output;
If the first query is determined to be inputable:
generating a user prompt corresponding to the first query;
Loading characteristic information related to the user, the conversation partner, or contents of a conversation between the conversation partner and the user;
generating a system prompt corresponding to the characteristic information;
inputting the user prompt and the system prompt into a first language model, respectively;
providing an output of the first language model based on the user prompt and the system prompt as the conversation partner's response to the first query;
storing the output of the first language model in a random access memory of the user terminal;
Inputting the first query into a second language model for extracting mid- to long-term memory information to be stored in the internal memory of the user terminal; and
When the second language model extracts mid- to long-term memory information from the first query, storing the extracted mid- to long-term memory information in the internal memory.
A method of providing chatbot services, including. According to claim 16,
The step of determining whether to input the first query is:
A method of providing a chatbot service, including the step of determining that the first query cannot be input when the first query includes a prohibited word included in the banned word dataset, with reference to a predefined banned word dataset. According to claim 16,
The step of determining whether to input the first query is:
A method of providing a chatbot service, including the step of determining that the first query cannot be entered if the topic corresponding to the first query corresponds to a topic included in the sensitive data set, with reference to a predefined sensitive data set. . A computer-readable recording medium on which a program for performing the method according to any one of claims 1 to 9 and 12 to 18 is recorded. In a user terminal that provides a chatbot service,
One or more processors; and
a memory that stores one or more instructions to be executed by the one or more processors;
The one or more processors execute the one or more instructions,
obtain a first query from the user about a specific conversation partner;
generate a user prompt corresponding to the first query,
Load characteristic information related to the user, the conversation partner, or the contents of a conversation between the conversation partner and the user,
Generate a system prompt corresponding to the characteristic information,
inputting the user prompt and the system prompt into a first language model, respectively;
providing an output of the first language model based on the user prompt and the system prompt as the conversation partner's response to the first query;
Store the output of the first language model in a random access memory of the user terminal,
Entering the first query into a second language model for extracting mid- to long-term memory information to be stored in the internal memory of the user terminal,
When the second language model extracts mid- to long-term memory information from the first query, the user terminal stores the extracted mid- to long-term memory information in the internal memory.

Images