Method and apparatus for providing artificial intelligence-based conference service

Abstract

According to an embodiment, provided is a computer-implemented method performed by a conference service apparatus comprising a memory, a processor, and one or more programs stored in the memory and configured to be executed by the processor, the computer-implemented method comprising the steps of: acquiring, by an agent, conversations received from conference participants who participate in an online conference room including the agent connected to a language model; selecting, by the agent, a target conversation from among the conversations, the target conversation including a conversation requiring a response using local internal data; providing, by the agent, the target conversation as input data to the language model, the language model generating, for a natural language input, a natural language response by using the local internal data; acquiring, by the agent, output data for the input data from the language model; and providing, by the agent on the basis of the output data, a conversation response to the target conversation to participant terminals of the conference participants.

Description

The Expert model 1132 may include a plurality of models trained for a plurality of respective specialized fields using specialized field data 420 corresponding to the respective specialized fields. For example, second training datasets may be prepared for each of a plurality of preset specialized fields (e.g., accounting, legal, patent, tax, medical, and finance), and a plurality of Expert models 1132 may be prepared using the respective second training datasets and trained to provide an optimized professional answer to a question based on knowledge or know-how in each specialized field. For example, the agent 112 may select, from among the plurality of Expert models 1132, an Expert model 1132 corresponding to one specialized field having the highest relevance to the second target conversation, provide the second target conversation as input data to the selected Expert model 1132, and obtain output data from the selected Expert model 1132. As another example, while the agent 112 participates in the online meeting room 10, the host terminal 200 may select, from among the plurality of Expert models 1132, an Expert model 1132 corresponding to one specialized field, and the agent 112 may participate in the online meeting room 10 using the selected Expert model 1132. The plurality of second training datasets may be obtained from public data associated with the corresponding specialized field. For example, specialized field data 420 or data intensively obtained from various external apparatuses within the scope of the corresponding specialized field may be used to prepare training data. Alternatively, the plurality of second training datasets may be obtained from local internal data 410 or specialized field data 420 associated with the corresponding specialized field. For example, a plurality of second training datasets associated with accounting may be prepared using the local internal data 410 or specialized field data 420 provided by an apparatus of a company providing accounting consulting services, and an Expert model 1132 associated with accounting may be prepared through training using the second training datasets. As an example, a meeting service may be provided by allowing an agent 112 using an Expert model 1132 associated with a particular specialized field to participate in an online meeting room 10 in response to an invitation request from a host terminal 200 or a participant terminal 300. The meeting service may also be provided through a paid transaction associated with the company according to a level of expertise, know-how, difficulty, or the like applied to the training datasets. As a result of the foregoing training, the Expert model 1132 may be capable of generating output data using specialized field data 420 upon receiving a target conversation as input. FIG. 10 is an exemplary flowchart illustrating a process in which an agent 112 according to a second embodiment obtains output data for a target conversation using specialized field data 420. Detailed descriptions of embodiments having technical features similar to the features described above are omitted. Referring to FIG. 10, in operation S1010, the agent 112 may select a second target conversation from the obtained conversation or conversations, the second target conversation comprising a conversation requiring a response using specialized field data 420. The agent 112 may select the second target conversation from the conversation or conversations based on the preset selection rule described above, and the preset selection rule may further be based on whether text related to specialized field data 420 (e.g., accounting, legal, patent, tax, medical, or finance) is included. Alternatively, the language model 113 (or a target conversation selection model) may be trained to select the second target conversation, and the agent 112 may select the second target conversation using the language model 113 (or the target conversation selection model). The foregoing training may use a training dataset prepared in consideration of characteristics of conversations using specialized field data 420. As in the embodiments described above, for example, the Expert model 1132 may be trained to select the second target conversation from the conversations and generate output data for the second target conversation using specialized field data 420. Similarly, the foregoing training may further use training datasets prepared in consideration of characteristics of conversations using specialized field data 420. In operation S1020, the agent 112 may provide the second target conversation as input data to the language model 113, and in operation S1030, the agent 112 may obtain, from the language model 113, output data for the second target conversation generated using specialized field data 420. As in the foregoing embodiments, the Expert model 1132, trained to respond to an input using specialized field data 420, may receive the second target conversation from the agent 112 and generate output data for the second target conversation using specialized field data 420. The agent 112 may provide input data including the second target conversation and the conversations to the Expert model 1132 and request the Expert model 1132 to generate output data for the second target conversation using the conversations. For example, the agent 112 may generate a prompt to instruct the Expert model 1132 to generate output data using the conversations of the meeting participants in a chat room and may obtain, from the Expert model 1132, output data generated using various matters included in the conversations and specialized field data 420. According to the foregoing embodiment, the agent 112 corresponding to an AI meeting participant may provide a professional response using specialized field data 420 for the second target conversation as an expert invited to the meeting. Accordingly, based on training in indepth specialized knowledge and know-how focused on a particular specialized field, the agent 112 may function as an expert AI capable of providing the meeting participants with consulting for problem solving. Even without a separate expert taking time to participate in the meeting, the expert AI may enable resolution of problems in a specialized field that may otherwise be difficult to resolve. Referring again to FIG. 8, the Participant model 1133 may be trained to converse with a plurality of users using public data. For example, the Participant model 1133 may be a generative AI model trained to participate flexibly in conversations among a plurality of users and may be a conversational model implemented to engage in context-appropriate conversations among multiple meeting participants exchanging conversations in real time or on a non-real-time basis. In an embodiment, the Participant model 1133 may be trained using training data including predefined reference conversations. The Participant model 1133 may be obtained by applying further training using the reference conversations to a generative AI model pretrained to converse with a user using public data. The reference conversations represent meeting conversations for training, including conversation sets involving three or more users, and may include sets of various conversations likely to occur for different organizations and situations, similar to instructional materials for multilateral conversations likely to occur during meetings. The reference conversations may further include, together with conversations involving three or more users, information about at least one of a meeting situation involving the conversations (e.g., ideation or establishment of project phases), a related organization (e.g., a company, school, military organization, or private organization), and user circumstances (e.g., a role, rank, or work experience of each user). For example, under an assumed situation involving exchanges of related conversations among three participants having the same or different ranks during a project-planning meeting at a company associated with a given industry, the reference conversations involving the three participants may be prepared from conversation samples obtained according to a predefined method (e.g., design by a designer or collection of public data). As another example, under an assumed situation involving exchanges of related conversations among five participants having the same or different ranks during a meeting for ideation concerning a given meeting topic at an educational institution, the reference conversations involving the five participants may be prepared. The Participant model 1133 may be trained through supervised learning, semisupervised learning, or self-supervised learning using training data including the plurality of reference conversations described above. For example, during training of the Participant model 1133, a self-supervision method may select an arbitrary participant in the reference conversations serving as training input data, mask a conversation of the selected participant without labeling a correct answer (i.e., training answer data), and define a task of predicting the masked conversation in consideration of the organization, situation, and the like described above. As another example, each reference conversation may include conversations corresponding to training input data and a conversation response corresponding to training answer data, and supervised learning may train the Participant model 1133 using each reference conversation as training data. Given conversations, the Participant model 1133 according to the foregoing embodiments may learn meeting conversation skills for independently determining a conversation requiring a response and responding to the determined conversation based on comprehensive consideration of the conversation content of the meeting participants, the relevant organization, the relevant situation, and the like. Accordingly, upon receiving conversations as input, the Participant model 1133 may be capable of generating output data using public data for a conversation requiring a response among the conversations. FIG. 11 is an exemplary flowchart illustrating a process in which the agent 112 according to a third embodiment obtains output data for a target conversation using public data 430. Referring to FIG. 11, in operation S1110, the agent 112 may select, from the obtained conversation or conversations, a third target conversation comprising a conversation requiring a response using public data 430. In an embodiment, the agent 112 may select the third target conversation based on the preset selection rule described above. For example, the agent 112 may select, as the third target conversation, a remaining conversation not corresponding to the first target conversation or the second target conversation. In another embodiment, the agent 112 may have the Participant model 1133 select the third target conversation and, for example, may transmit the obtained conversations to the Participant model 1133. In operation S1120, the agent 112 may provide the third target conversation as input data to the language model 113, and in operation S1130, the agent 112 may obtain, from the language model 113, output data generated using public data 430 for the third target conversation. In an embodiment, the agent 112 may perform operation S1120 together with operation S1110. For example, the agent 112 may provide the obtained conversations to the Participant model 1133 and request the Participant model 1133 to generate output data for a conversation requiring a response among the conversations. By using the Participant model 1133 trained in multilateral meeting conversation skills, the agent 112 according to the foregoing embodiment may function as a full participant autonomously participating in a context-appropriate manner without interrupting the flow of a conversation involving multiple participants in each organization and situation. The third target conversation may be a conversation requiring a response using recent public data. The recent public data may refer to recent public data not included in the public data 430. In an embodiment, the agent 112 may select, from the conversation or conversations, a third target conversation requiring a response using recent public data based on a preset selection rule further including whether text related to recent public data (e.g., recent trends, updated materials, or recent developments) is included. In an embodiment, if the agent 112 selects the third target conversation as a conversation requiring a response using recent public data, the agent 112 may obtain recent public data associated with the third target conversation and cause the Participant model 1133 to generate output data using the obtained recent public data. For example, recent public data included in the database 400 may be updated at a preset cycle to include recent data, and the agent 112 may obtain, from the database 400 using a keyword search method, a search result for recent public data associated with the third target conversation. As another example, the agent 112 may obtain, from external apparatuses using a preset data collection method (e.g., communication with particular data servers for data collection or data collection through crawling), a search result for recent public data associated with the third target conversation. The agent 112 may generate a prompt instructing the Participant model 1133 to generate output data using the obtained search result and may provide the prompt to the Participant model 1133. Alternatively, the agent 112 may cause further training of the Participant model 1133 using the obtained search result and may provide the third target conversation to the further-trained language model 113 to obtain output data using recent public data. In some embodiments, at least some of the embodiments described above may be implemented in combination, without being limited to the foregoing examples. For example, the agent 112 may select, from the obtained conversation or conversations, a target conversation requiring a response using at least some of local internal data 410, specialized field data 420, and public data 430, and may request the Expert model 1132 to generate a response to the target conversation using a search result for local internal data 410 associated with the target conversation. As another example, the language model 113 may include at least some of the Colleague model 1131, the Expert model 1132, and the Participant model 1133 or may be a hybrid model obtained by applying the training methods described above for the at least some models. In an embodiment, the language model 113 may generate the output data in at least one response form selected from the group consisting of a question, an answer, an opinion, and a summary. The agent 112 or the language model 113 may determine the response form. In an embodiment, the agent 112 may determine the response form. Specifically, the agent 112 may select a target conversation and determine a response form for the target conversation based on the target conversation. Furthermore, the agent 112 may provide the target conversation and the determined response form together to the language model 113 and request the language model 113 to generate output data for the target conversation in the determined response form. The agent 112 may determine the response form based at least partially on the preset selection rule described above. For example, the agent 112 may determine the response form based on at least one of a type of conversation, an intention of conversation, a context of preceding and subsequent conversations, a time interval of conversations, and a count of conversations during a given time period, as applied to the preset selection rule described above. For example, based on a type of the target conversation (e.g., a question or a request), the agent 112 may determine a corresponding response form (e.g., an answer to the question or an opinion regarding the request); based on an intention of the target conversation (e.g., an action proposal or information confirmation), the agent 112 may determine a corresponding response form (e.g., sharing a result of the action or sharing information); considering the context of preceding and subsequent conversations, the agent 112 may determine a response form (e.g., presenting an opinion if the conversation flow involves sharing opinions); considering the time interval or count of conversations, the agent 112 may determine a response form (e.g., sharing, as an opinion, recent information related to a meeting agenda if no question has been presented but a conversation update rate or conversation activity level is relatively slow or low); or the agent 112 may determine the response form through comprehensive consideration of at least some of these factors. In another embodiment, the language model 113 or a target conversation selection model may determine the response form. For example, the Participant model 1133 may be trained to generate output data in one response form selected from a question, an answer, an opinion, and a summary for a conversation requiring a response among conversations. Each reference conversation used to train the Participant model 1133 may include conversations in two or more conversation forms selected from a question, an answer, an opinion, and a summary, and the Participant model 1133 may be trained from each reference conversation to generate an output in a response form of a question, answer, opinion, or summary appropriate for the relevant organization and situation. The agent 112 may transmit the conversations or target conversation to the language model 113, and the language model 113 may determine a conversation requiring a response among the transmitted conversations and generate a response in an appropriate response form. As another example, the target conversation selection model may be trained to determine both a target conversation and a response form for the target conversation from conversations. In an embodiment, the agent 112 may cause output data corresponding to setting information of an AI meeting participant to be generated as a response to a target conversation. In an embodiment, when receiving an invitation request for the AI meeting participant from the host terminal 200 or participant terminal 300, the meeting-room providing unit 111 may also receive the setting information of the AI meeting participant and may set the agent 112 participating in the online meeting room 10 to participate in the meeting based on the setting information of the AI meeting participant. In an embodiment, the setting information of the AI meeting participant may include setting information for a method of determining a response form and / or for access authority and may further include setting information for at least one of a name, rank, affiliation, tone, age, gender, and language. The method of determining the response form represents a method for assigning a different weight to at least one of a question, an answer, an opinion, and a summary during determination of a response form of output data for a target conversation. For example, the method of determining the response form may include at least one of a question participation type assigning a higher weight to a question than to other response forms, an answer participation type assigning a higher weight to an answer, an opinion participation type assigning a higher weight to an opinion, and a summary participation type assigning a higher weight to a summary. For example, when selecting a target conversation, the agent 112 may select, with a relatively higher probability than other conversations, a conversation having a response form of output data corresponding to the setting information (e.g., answer-focused), such as a conversation having a type classified as a question. When determining the response form of the output data, the agent 112 may select, with a higher probability than other response forms, a response form corresponding to the setting information (e.g., answer-focused), such as an answer. As another example, the language model 113 may be trained to generate, with a relatively higher probability, output data in a response form (e.g., an answer) corresponding to the method of determining the response form for a conversation (e.g., a question) requiring a response based on consideration of the method of determining the response form (e.g., answer-focused). Furthermore, the access authority is granted to the agent 112 corresponding to the AI meeting participant and may be used to restrict the local internal data 410 used in generating a response to a target conversation to local internal data 410 assigned a security level corresponding to the access authority. In an embodiment, when a meeting host and meeting participants participating in the online meeting room 10 are determined, the access authority of the AI meeting participant participating in the online meeting room 10 may be determined according to a pre-stored security policy based on the access authority of each of the meeting host and the meeting participants. In an embodiment, the agent 112 may provide the setting information of the AI meeting participant together with the target conversation to the language model 113 and request the language model 113 to generate a response to the target conversation in consideration of the setting information. In another embodiment, language models 113 may be prepared, each individually trained to generate a response corresponding to a respective one of a plurality of combinable setting cases in the setting information of the AI meeting participant, and the agent 112 may communicate with the language model 113 corresponding to the setting information to obtain a response matching the setting information. In an embodiment, the agent 112 may obtain, from a plurality of pre-stored profile images, a profile image corresponding to the setting information of the AI meeting participant and may cause a conversation response to the target conversation to be provided together with the profile image. In an embodiment, the remaining setting information of the AI meeting participant, excluding the access authority, may be updated based on an update request from the host terminal 200 or participant terminal 300. In an embodiment, the agent 112 may provide a conversation response to a target conversation to a participant terminal 300 of a meeting participant based on the obtained output data. The agent 112 may generate a conversation response including the output data or a conversation response processed, converted, or modified from the output data and may provide the generated conversation response to the participant terminal 300 accessing the online meeting room 10. Provision of the conversation response is described below with reference to FIG. 12. FIG. 12 illustrates an example of a conversation response 30 displayed on a screen of the participant terminal 300. FIG. 12 illustrates an example of the conversation response 30 of the agent 112 provided as a chat message through a messenger-based chat user interface (UI). However, the illustrated case is merely an example. Different implementations may provide the conversation response 30 in various manners, for example, as a comment through a comment-based conversation UI (e.g., YouTube comments) or as a data file. Referring to FIG. 12, the agent 112 may provide the conversation response 30, in the form of a message from the AI meeting participant, to a participant terminal 300 of a meeting participant or a respective participant terminal 300 of each meeting participant. For example, based on the granted authority for conversation input, the agent 112 may add the conversation response 30 to the conversation thread object 13 of the online meeting room 10 in the form of a message from the AI meeting participant. Accordingly, each participant terminal 300 accessing the online meeting room 10 may display the conversation response 30 as a conversation from the AI meeting participant participating in the chat room. In an embodiment, the agent 112 may provide the conversation response 30, in the form of a message from the AI meeting participant visually distinguishable from the meeting participants, to a participant terminal 300 of a meeting participant or a respective participant terminal 300 of each meeting participant. For example, a display manner of the conversation response 30, such as a message color, speech-bubble form, character shape, or character size, may differ at least partially from a display manner for the meeting participants. As another example, a participant profile image representing the AI meeting participant may accompany the conversation response 30, and the participant profile image or its display manner may visually distinguish the AI meeting participant from the meeting participants. Accordingly, the meeting participants may readily identify a message from the AI meeting participant visually in the chat room. In an embodiment, the agent 112 may provide information about the target conversation together with the conversation response 30 to the participant terminal 300. For example, if the agent 112 selects a second message 22 from the conversations 20 as the target conversation and a conversation response 30 to the second message 22 is generated, the agent 112 may cause information identifying the second message 22 as the target conversation of the conversation response 30 (e.g., an indication identifying the conversation response 30 as a reply message to the second message 22) to be displayed together with the conversation response 30. In an embodiment, the agent 112 may provide a response form of the output data together with the conversation response 30 to the participant terminal 300. For example, the agent 112 may cause display information (e.g., text, an image, or an icon) representing one of a question, an answer, an opinion, and a summary corresponding to the response form of the conversation response 30 to be displayed together with the conversation response 30. In an embodiment, the agent 112 may provide, together with the conversation response 30 to the participant terminal 300, a response basis 40 including basis data used to obtain the output data and / or an access link to the basis data. Specifically, based on the access authority granted to each meeting participant, the agent 112 may provide the basis data and / or the access link as the response basis 40. For example, when generating the prompt, the agent 112 may generate a prompt further including an instruction to additionally provide information about the basis data used to obtain the output data and / or the access link to the basis data. Furthermore, the agent 112 may extract the basis data and / or the access link from the output data and provide, in the form of a message, the conversation response 30 included in the output data and the response basis 40 including the extracted basis data and / or the access link. In an embodiment, the agent 112 may determine whether to provide the conversation response 30 based on the access authority granted to each of the meeting participants. For example, the agent 112 may perform a double-check operation to determine whether a security level corresponding to the basis data received from the language model 113 corresponds to the access authority granted to each of the meeting participants (or the access authority granted to the online meeting room 10 based on the access authorities). If no correspondence exists, the agent 112 may prevent provision of the conversation response 30 to the participant terminals 300 of the meeting participants. As another example, the agent 112 may provide the conversation response 30 exclusively to a participant terminal 300 of a meeting participant having access authority corresponding to the security level of the basis data. In an embodiment, based on security levels corresponding to the access authorities of the meeting participants, the agent 112 may obtain, from the language model 113, a plurality of responses to the target conversation corresponding to the security levels and provide, to each respective participant terminal 300, a respective response corresponding to the access authority of the corresponding meeting participant. Specifically, if the security levels corresponding to the access authorities granted to the meeting participants differ from one another, the agent 112 may obtain, from the language model 113, a plurality of responses to the target conversation corresponding to the security levels. For example, if a first meeting participant having a higher rank and a second meeting participant having a lower rank (e.g., a manager) participate in a meeting, the agent 112 may respectively obtain a first search result for local internal data 410 assigned a first security level corresponding to the higher rank and a second search result for local internal data 410 assigned a second security level corresponding to the lower rank. The agent 112 may cause the language model 113 to generate first output data (e.g., an answer generated based on corporate materials having a relatively high security level) and second output data (e.g., an answer generated based on corporate materials having a relatively low security level) using the first search result and the second search result, respectively. The agent 112 may process the first output data and the second output data into respective conversation responses accessible to meeting participants having access authorities corresponding to the respective security levels and add the respective conversation responses to the conversation thread object 13. Accordingly, each participant terminal 300 may display the same answer or different answers to the same question according to the access authority of the corresponding meeting participant, or may omit display of some answers. In an embodiment, if the target conversation is a conversation for which a response has been individually requested by a participant terminal 300 of a specific meeting participant, the agent 112 may provide the corresponding conversation response 30 exclusively to the participant terminal 300 of the specific meeting participant. For example, an overall meeting chat room joined by the meeting participants may receive a conversation together with an individual response request directed to the AI meeting participant. After selecting the conversation as the target conversation, the agent 112 may apply security processing to the conversation response 30 to make the conversation response 30 viewable exclusively by the specific meeting participant and add the conversation response 30 to a conversation thread of the overall meeting chat room included in the conversation thread object 13. As another example, in association with the online meeting room 10, for each meeting participant, a one-to-one meeting chat room joined by the meeting participant and the AI meeting participant may be generated in response to a request from the meeting participant, and the one-to-one meeting chat room may receive a conversation from the corresponding meeting participant. The agent 112 may regard the conversation as the target conversation and add the conversation response 30 to the conversation thread of the entire one-to-one meeting chat room included in the conversation thread object 13. In an embodiment, the agent 112 may provide the conversation response 30 according to an input modality of the conversation selected as the target conversation or may provide the conversation response 30 in at least one preset output modality selected from text, image, video, and audio. For example, a conversation received in an audio input modality may be converted into text and provided to the chat room as the second message 22. If the second message 22 is selected as the target conversation, the conversation response 30 to the second message 22 may be converted into an audio output modality and provided to the chat room in the form of a voice message from the AI meeting participant or may be provided to the chat room in the form of a text message according to the preset output modality. In an embodiment, the further-training unit 114 may be implemented to perform further training using meeting information of the online meeting room 10 and / or the conversations. The further training may include pre-meeting training performed before a meeting starts and / or real-time training performed during the meeting. The further training encompasses post hoc further training of the pretrained language model 113 to enable the agent 112 to provide an optimized conversation response for the online meeting room 10. Related embodiments are described below with further reference to FIG. 13. FIG. 13 is an exemplary flowchart illustrating pre-meeting training performed by the further-training unit 114 according to an embodiment. Referring to FIG. 13, the further-training unit 114 may perform pre-meeting training using meeting information of the online meeting room 10 before the online meeting room 10 is started. The start of the online meeting room 10 may refer to arrival of a start time within the meeting duration of the online meeting room 10 or receipt of a meeting start request. If participation of the agent 112 in the online meeting room 10 is determined, the further-training unit 114 may perform the pre-meeting training. In operation S1310, the meeting-room providing unit 111 may determine whether the agent 112 participates in the online meeting room 10. For example, as described above, if a meeting request includes an invitation request for the AI meeting participant, the meeting-room providing unit 111 may grant the agent 112 the same authority to input and receive conversation as the meeting participants in the online meeting room 10 before the meeting starts. In operation S1320, after participation of the agent 112 is determined, the further-training unit 114 may obtain meeting information of the online meeting room 10. For example, the further-training unit 114 may retrieve the meeting information from the basic information object 11 and the meeting history object 14 of the online meeting room 10. In an embodiment, the meeting information of the online meeting room 10 used for the pre-meeting training may include at least one selected from the group consisting of: (i) a meeting duration; (ii) meeting content; (iii) information about the meeting participants; (iv) meeting classification information; and (v) a meeting history indicating a historical relationship between the online meeting room 10 and one or more other online meeting rooms 10 associated with the online meeting room 10. For example, the meeting information of the online meeting room 10 may include a meeting duration, meeting content, information about a meeting host, information about the meeting participants, meeting classification information, a meeting title, a meeting goal, and the like included in the basic information object 11, and may further include the meeting history object 14. In operation S1330, the further-training unit 114 may obtain external data related to the meeting information of the online meeting room 10 before the online meeting room 10 is started. For example, the further-training unit 114 may obtain the external data related to the meeting information of the online meeting room 10 from one or more external apparatuses using the data collection methods described above, such as communication with specific data servers for data collection and data collection through crawling. The obtained external data may include public data related to the meeting information (e.g., technical details), recently published data (e.g., recent industry trends), and / or specialized field data 420 (e.g., amendments to laws and regulations), among others. In operation S1340, the further-training unit 114 may perform pre-meeting training using the obtained meeting information of the online meeting room 10 and / or the obtained external data. In an embodiment, the further-training unit 114 may perform further training of the language model 113 using the obtained external data. For example, the further-training unit 114 may apply a predefined method to the obtained external data to preprocess the external data into training data and may perform pre-meeting training of the language model 113 using the preprocessed training data. In an exemplary case, before the meeting starts, the further-training unit 114 may collect recent industry trends related to a meeting topic or specialized knowledge about related technology and may further train the language model 113 through processes including observation and pattern matching of training data based on the collected information. In an embodiment facilitating further training of the language model 113, premeeting training of the language model 113 using additional information related to the meeting and collected through various channels before the meeting starts may prepare the language model 113 with a better understanding of information related to the meeting before the meeting starts. In another embodiment, the further-training unit 114 may update the database 400 to include the obtained external data. For example, the further-training unit 114 may perform data conversion, indexing, and the like on the external data related to the meeting information to organize the external data in a database form corresponding to a keyword search method and / or a semantic search method used by the agent 112 to obtain a search result. Accordingly, during a subsequent process, the agent 112 may obtain a search result from the database 400 further including the external data related to the meeting information, and the language model 113 may generate output data as a response to a target conversation based on the search result. In an embodiment in which further training of the language model 113 is difficult, this database update may support an environment enabling the language model 113 to provide a response optimized for the meeting by considering the related additional information. In another embodiment, the further-training unit 114 may store the obtained external data in the corresponding online meeting room 10 for additional use during a subsequent process in which the agent 112 provides a conversation response. For example, the online meeting room 10 may further include an additional information object (not illustrated) storing information for pre-meeting training and / or real-time training, and the further-training unit 114 may include the obtained external data in the additional information object of the online meeting room 10. During a subsequent process, the agent 112 may generate a prompt instructing the language model 113 to generate output data as a response to a target conversation by further using the external data related to the meeting information and included in the additional information object, and may provide the prompt to the language model 113. Under this arrangement, the language model 113 does not use the information stored in the additional information object for training. The information may be used exclusively for the corresponding online meeting room 10 and then deleted when the meeting ends, thereby providing an information security advantage. In another embodiment, the further-training unit 114 may perform further training of the language model 113 using the obtained meeting information. For example, based on the meeting history object 14, the further-training unit 114 may obtain a historical relationship between the corresponding meeting and other meetings, meeting information of the other meetings, and the like, and may use the obtained information for pre-meeting training to enable the language model 113 to understand the meeting history of the past meetings, meeting result reports from the past meetings, participant-specific task items described below, and the like. Accordingly, before the meeting starts, the pre-meeting training may prepare the language model 113 to retain and track the meeting history, task items, and the like of past meetings related to the corresponding meeting. Pre-meeting training according to an embodiment may additionally use conversations from meetings prior to generation of the online meeting room 10. Specifically, based on the meeting history object 14, the further-training unit 114 may obtain conversations from other meetings associated with the online meeting room 10 and perform pre-meeting training further using the conversations. For example, if performing pre-meeting training for a first online meeting room representing a current meeting, the further-training unit 114 may search the meeting history object 14 of the first online meeting room for one or more second online meeting rooms associated with the first online meeting room. The second online meeting rooms represent past meetings, and the further-training unit 114 may obtain, from the conversation thread object 13 of each second online meeting room, conversations from meetings prior to generation of the first online meeting room. In an embodiment, the further-training unit 114 may perform pre-meeting training of the language model 113 using the obtained conversations from meetings prior to generation of the online meeting room 10. Similarly, the further-training unit 114 may apply a predefined method to the obtained past conversations to preprocess the past conversations into training data and may further train the language model 113 using the preprocessed training data. Accordingly, before the meeting starts, the pre-meeting training may enable the language model 113 to account for a level of achievement of a meeting goal, results previously derived, relevant knowledge, factual relationships, and the like included in conversations exchanged during past meetings related to the corresponding meeting. In another embodiment, the further-training unit 114 may include the obtained conversations from meetings prior to generation of the online meeting room 10 in the local internal data 410. Accordingly, while the meeting proceeds, the agent 112 may obtain, from the local internal data 410, a search result further including the conversations from meetings prior to generation of the online meeting room 10 and may provide a conversation response based on the search result. In another embodiment, the further-training unit 114 may store the obtained conversations from meetings prior to generation of the online meeting room 10 in the corresponding online meeting room 10 for additional use during a subsequent process in which the agent 112 provides a conversation response. For example, as described above, the further-training unit 114 may include the obtained conversations from meetings prior to generation of the online meeting room 10 in the additional information object of the online meeting room 10. During a subsequent process, when transmitting a target conversation to the language model 113, the agent 112 may also provide the information included in the additional information object and request the language model 113 to generate output data as a response to the target conversation by further using the information. In the embodiments described above, pre-meeting training using external data related to the corresponding meeting information, the meeting history, conversations from past meetings, and the like before the meeting starts enables the agent 112 to provide, for conversations exchanged during the meeting, conversation responses more optimized for the corresponding meeting. The further-training unit 114 according to an embodiment may perform real-time training using conversations of the meeting participants while the online meeting room 10 is being provided. The real-time training may refer to training using a conversation obtained in real time through the online meeting room 10 or training using conversations obtained during a given time period of real-time operation. In an embodiment, the further-training unit 114 may perform real-time training of the language model 113 using conversations obtained through the online meeting room 10. For example, the further-training unit 114 may apply a predefined method to a conversation added to the conversation thread object 13 to preprocess the conversation into training data and may perform real-time training of the language model 113 using the preprocessed training data. The real-time training may occur before, concurrently with, or after providing a conversation response to a target conversation. Accordingly, the language model 113 may be trained to provide an appropriate response by comprehensively considering relevant knowledge, factual relationships, and the like included in real-time conversations. In another embodiment, the agent 112 may perform real-time training using the conversations. For example, when providing a target conversation to the language model 113, the agent 112 may also provide the conversations and request the language model 113 to generate output data based on further training using the conversations. Under this arrangement, further training of the language model 113 is performed, and the language model 113 may identify factual relationships, progress, and the like from conversations exchanged during the corresponding meeting and generate a response based on the identified factual relationships, progress, and the like. As another example, the agent 112 may request the language model 113 to generate output data as a response to the target conversation by further using the conversations and / or meeting information of the online meeting room 10. Under this arrangement, even if further training of the language model 113 is not performed, the language model 113 may generate a response by considering a meeting atmosphere, specialized terms, conversation circumstances, and the like reflected in the conversations on an ongoing basis in each meeting. In the embodiments described above, real-time training using conversations added in real time, shared materials, and the like while the corresponding meeting proceeds enables the agent 112 to provide a conversation response more closely optimized for the corresponding meeting. In an embodiment, information used for pre-meeting training and / or real-time training may be separately stored and managed in a data space of the corresponding online meeting room 10. When the meeting duration of the corresponding online meeting room 10 expires, use of the information may be restricted or prohibited. For example, the information may be deleted when the meeting end time arrives. In an embodiment, the agent 112 may obtain a meeting result report for the online meeting room 10 based on conversations of the meeting participants. When a meeting corresponding to the online meeting room 10 ends, the agent 112 may generate a meeting result report summarizing meeting results from the online meeting room 10 based on at least one selected from the group consisting of a basic information object 11, a content object 12, a conversation thread object 13, and a meeting history object 14 included in the online meeting room 10. The meeting may end when an end time within the meeting duration of the online meeting room 10 arrives or when the host terminal 200 or the participant terminal 300 submits a meeting termination request. In an embodiment, the agent 112 may obtain a meeting result report including a summary result for each of a plurality of preset criteria items from the conversations. For example, the agent 112 may obtain summary results by summarizing all conversations obtained through the online meeting room 10 for each criterion item and may generate the meeting result report by applying the summary results to a preset report format. The plurality of criterion items may include at least one selected from the group consisting of a meeting outline item, a meeting content item, a main conversation collection item, a task item, and a next meeting item. In an embodiment, the meeting outline item represents basic information about a meeting and may include at least one selected from the group consisting of a meeting duration (or a meeting start time and a meeting end time), a meeting space (e.g., connection information for the online meeting room 10), information about the meeting participants (e.g., names, ranks, affiliations, and the like), and a meeting goal. For example, the meeting outline item may be derived from meeting information (e.g., the meeting duration, the meeting participants, the meeting goal, and the like) included in the basic information object 11. If the meeting information omits information related to a specific item, conversation content corresponding to the specific item may be extracted from conversations included in the conversation thread object 13 to obtain the meeting outline item. In an embodiment, the meeting content item represents summary information about content discussed in the meeting and may include a main topic item related to the meeting goal and an additional topic item unrelated to the meeting goal. Each topic item may include at least one sub-item selected from the group consisting of a keyword, a topic description, a content summary of main conversations (e.g., opinions, questions, and answers), and a conclusion. In an embodiment, the main conversation collection item represents a classification result for main conversations. A main conversation may refer to a conversation associated with a relatively large number of conversations (e.g., a source of reply messages, reply-to-reply messages, and the like) or a conversation for which a relatively large number of meeting participants have selected a reaction expression (e.g., a like, a recommendation, and the like). A main conversation may correspond to a conversation related to a main topic item among the conversations or a conversation concerning an additional topic accounting for at least a preset ratio of the total number of conversations. In an embodiment, the main conversation collection item may include a conversation collection item by conversation form (e.g., by question, by opinion, or by answer) and / or a conversation collection item by associated conversation (e.g., a collection of a specific question and a series of opinions, answers, and the like related to the specific question). As an example of the conversation collection item by conversation form, conversations in a chat room may be classified into a conversation form selected from the group consisting of a question, an opinion, an answer, and another form. The agent 112 may generate, as a summary result corresponding to the conversation collection item by conversation form, a result in which the conversations classified by each conversation form are arranged in a tabular format. As an example of the conversation collection item by associated conversation, if a conversation in the chat room is associated with another conversation, mutual associations among the conversations (e.g., reply messages, reply-to-reply messages, and the like for a specific conversation) may be managed. The agent 112 may generate, as a summary result corresponding to the conversation collection item by associated conversation, a result in which mutually associated conversations are grouped and the conversations in each group are arranged by chronological order, meeting participant, or conversation form. In an embodiment, the task item represents summary information about a task assigned to one or more meeting participants among the content discussed in the meeting and may include at least one selected from the group consisting of information about the one or more meeting participants, task content, a task duration, task progress (e.g., a progress rate or a progress stage), and a task result (e.g., a description of a deliverable, a task file, and the like). The agent 112 may manage each task item through follow-up actions described below. In an embodiment, the next meeting item represents scheduled information about a next meeting and may include at least one selected from the group consisting of a next meeting schedule, a next meeting goal, and next meeting participants. If a summary result corresponding to any of the items is not obtained from the conversations, the agent 112 may add a message inquiring about the item to the conversation thread object 13 and may generate the corresponding summary result based on a response to the message from the host terminal 200 or the participant terminal 300. In an embodiment, the agent 112 may include one or more artificial intelligence models (e.g., a generative AI model optimized for summarization) trained to generate, from the conversations, a summary result for each of the plurality of preset criterion items described above. In another embodiment, the agent 112 may provide the conversations to the language model 113, request the language model 113 to generate the summary results described above from the conversations, and generate a meeting result report including the summary results obtained from the language model 113. In an embodiment, the agent 112 may provide the meeting result report to the participant terminal 300 by adding the meeting result report to the conversation thread object 13. Accordingly, the meeting participants may view the meeting result report as a message in the chat room. In another embodiment, the agent 112 may include the meeting result report in the basic information object 11. Accordingly, the meeting participants may view, as a post, the meeting result report included in the meeting information of the online meeting room 10. In an embodiment, the agent 112 may modify the content of the meeting result report based on a modification request from the host terminal 200 or the participant terminal 300 and may update the meeting result report to separately identify the content before the modification, the content after the modification, and the meeting participant who made the modification. In an embodiment, the agent 112 may perform follow-up actions associated with the online meeting room 10 even after an end time within the meeting duration of the online meeting room 10 arrives (or after a meeting termination request is received). The follow-up actions may include scheduling the next meeting item and sharing task progress of the task item, without being limited thereto. In an embodiment, the agent 112 may schedule the next meeting item based on an approval response from one of the meeting participants for the next meeting item. For example, based on summary results for a next meeting schedule, a next meeting goal, and next meeting participants as the next meeting item in a meeting result report for a first online meeting room corresponding to a current meeting, the agent 112 may generate a second online meeting room including the information as meeting information and may configure the second online meeting room to be provided, according to the next meeting schedule, to a respective participant terminal 300 of each next meeting participant. In an embodiment, the agent 112 may add, to the conversation thread object 13, a message inquiring whether to schedule the next meeting item according to a summary result corresponding to the next meeting item included in the meeting result report and may schedule the next meeting item in response to an approval response to the message from the host terminal 200 or the participant terminal 300. In an embodiment, the agent 112 may track each task item based on information regarding task progress and / or a task result received from a participant terminal 300 associated with each task item. The agent 112 may include information about each task item in the basic information object 11 based on a summary result corresponding to each task item and may track each task item. For example, during a task duration of each task item, the agent 112 may provide each participant terminal 300 with a message requesting an update of information about the task item and may update the information about the task item in response to an update request regarding task progress and / or a task result received from the corresponding participant terminal 300. The agent 112 may add the updated information about the task item to the conversation thread object 13, thereby sharing task progress or a deliverable for each task item in the chat room. In an embodiment, the agent 112 may add, to the conversation thread object 13, a message inquiring whether to track a task item according to a summary result corresponding to the task item included in the meeting result report and may track the task item in response to an approval response to the message from the host terminal 200 or a participant terminal 300 associated with the task item. When the follow-up actions described above are performed, for example, when the progress stages of the task items reach completion, the session for conducting the meeting in the online meeting room 10 may end, and the same authority to input and receive conversation as the meeting participants, granted to the agent 112 associated with the online meeting room 10, may be revoked. Under the embodiments described above, the meeting participants may conveniently view a meeting result report in which the artificial intelligence (AI) meeting participant summarizes meeting content from the conversations when the meeting ends and may also readily view task progress for task items assigned to the meeting participants after the meeting ends. In particular, in a meeting environment, such as a task force, involving multiple meetings directed toward a specific task, the agent 112 may effectively track task items of meeting participants derived from the multiple meetings. In an embodiment, the security unit 115 may perform security procedures to increase the security level according to a previously stored organization-specific security policy while the online meeting room 10 is being provided. If local internal data 410 is used to provide a conversation response, the security unit 115 may perform additional security procedures for the meeting participants and provide the conversation response based on the results. In an embodiment, the security unit 115 may request a respective participant terminal 300 of each meeting participant participating in the online meeting room 10 to provide biometric information (e.g., voice, face, fingerprint, and the like) and may perform a security procedure to verify whether the biometric information received from each participant terminal 300 corresponds to biometric information previously registered in a user account of each meeting participant. For example, the security unit 115 may perform the security procedure when a meeting participant initially accesses the online meeting room 10, at preset time intervals, or whenever the meeting participant accesses the online meeting room 10. In another embodiment, the security unit 115 may perform a security procedure to determine whether at least one of the participant terminals 300 operates based on an online meeting environment combined with an offline meeting. For example, if at least one participant terminal 300 is in an online meeting environment combined with an offline meeting, the security unit 115 may update the online meeting room 10 to have participation authority corresponding to the lowest security level. In still another embodiment, the security unit 115 may perform a security procedure to verify the identity and / or organizational membership of each meeting participant by confirming the meeting participant's identity and / or organizational membership using a one-time password granted by an administrator, among other techniques. In an embodiment, the security unit 115 may store and manage an organization-specific security policy. The security policy may include, for example, methods of setting security levels and participation authority, methods of performing the security procedures described above, and the like. Furthermore, the security unit 115 may monitor the agent 112 and / or the language model 113 based on the security policy and may store and manage, for example, monitoring results indicating whether the access authorities of the meeting participants correspond to the security levels used in providing the respective conversation responses. As described above, use of the local internal data 410 requires a substantial degree of security. In addition to a security procedure restricting use of the local internal data 410 based on the security levels described above, additional security procedures verifying a user's biometric information, a meeting environment, or the like may provide an even higher degree of security. Depending on the implementation, an additional security procedure may also be performed if specialized field data 420 is used to provide a conversation response. In an embodiment, the training unit 116 may be implemented to train the language model 113. Training by the training unit 116 may refer to pretraining the language model 113 before the meeting service apparatus 100 begins providing a meeting service. The following description omits portions overlapping with the content described above. In an embodiment, the training unit 116 may obtain a model pretrained using a large amount of text data. For example, the training unit 116 may pretrain an initial language model through a task of predicting a next token based on cross-entropy using a large amount of text data or may receive a model pretrained by an external apparatus, thereby obtaining a pretrained model (e.g., a transformer-based language model). In an embodiment, the training unit 116 may obtain a model pretrained to generate a natural language response to a natural language input and may obtain, for example, a generative AI model pretrained to converse with a user in natural language. In an embodiment, the training unit 116 may obtain a plurality of training datasets based on local internal data 410 and may obtain a Colleague model 1131 by applying fine-tuning to a pretrained model using the plurality of training datasets. For example, a plurality of first training datasets may be used during the fine-tuning, and each first training dataset may include training input data including conversations about the local internal data 410 and training ground-truth data including a response using the local internal data 410 associated with the conversations. Various algorithms known as methods for training language models or generative AI models, such as large language models (LLMs), variational autoencoders (VAEs), and generative adversarial networks (GANs), may be used for the training. In another embodiment, the training unit 116 may obtain a plurality of training datasets based on specialized field data 420 and may obtain an Expert model 1132 by applying finetuning to a pretrained model using the plurality of training datasets. For example, a plurality of second training datasets may be used during the fine-tuning, and each second training dataset may include training input data including conversations about a preset specialized field and training ground-truth data including a response based on the specialized field data 420. In another embodiment, the training unit 116 may obtain a model pretrained using public data associated with a specialized field as training data and may obtain an Expert model 1132 by applying, to the pretrained model, fine-tuning using a plurality of training datasets obtained from non-public data associated with the specialized field. For example, the training unit 116 may obtain a model pretrained through self-supervised learning using public data associated with the accounting field (e.g., international accounting standards, national accounting law, local accounting law, and the like) and may implement the Expert model 1132 by fine-tuning the pretrained model through supervised learning using a plurality of training datasets generated by preprocessing, as training data, specialized knowledge, know-how, and the like provided by an accounting consulting company (e.g., accounting-related questions and answers to the questions based on the specialized knowledge and know-how). Accordingly, intensive training using public data and non-public data regarding the specialized field, rather than a large amount of general-purpose public data, may produce the Expert model 1132 having extensive knowledge of the specialized field and expertise in solving problems. In still another embodiment, the training unit 116 may obtain a plurality of training datasets based on public data 430 and may obtain an Expert model 1132 by applying finetuning to a pretrained model using the plurality of training datasets. In an embodiment, the training unit 116 may train a Participant model 1133 using training data including predefined reference conversations. As described above, during the fine-tuning, the predefined reference conversations may be used as training data, and further training through supervised learning, semi-supervised learning, or self-supervised learning may be performed to predict a next conversation for each participant, thereby obtaining the Participant model 1133. In still another embodiment, the training unit 116 may obtain the language model 113 by performing training using two or more of the training schemes described above for the Colleague model 1131, the Expert model 1132, and the Participant model 1133. In an embodiment, the meeting-room providing unit 111 may perform a process of generating the online meeting room 10 based on a meeting request. Specifically, the meeting-room providing unit 111 may receive, from a host terminal 200, a meeting request for an online meeting room. In an embodiment, the meeting request may include meeting information including at least one of a meeting duration, meeting content, information about meeting participants, and meeting classification information. The meeting duration includes information about a duration during which a meeting proceeds in real time or on a non-real-time basis. For example, a meeting duration for a realtime meeting may include time information. As another example, a meeting duration for a nonreal-time meeting may include information about a start time and an end time or information about a start time and a time interval (e.g., 3 days, 7 days, or one month). The meeting content includes information about content to be addressed in a meeting. For example, the meeting content may include at least one of text data, image data, video data, and audio data for describing a meeting agenda or may include an electronic document containing the at least one type of data. The information about the meeting participants includes information about users eligible to participate in a meeting and may include, for example, user information (e.g., an ID, a name, a company name, and a department name) of user accounts invited to the meeting. In an embodiment, the number of users corresponding to the information about the meeting participants may be set to N or greater (N is a natural number). In an example, N may be 3 or greater. In another example, N may be 4 or greater or 2 or greater. The meeting classification information includes one or more items of classification information associated with a meeting. For example, the meeting classification information may include classification information indicating a technical field encompassing the meeting or a task within a project serving as the purpose of the meeting, among other information. In an embodiment, the meeting classification information may include classification information for at least one of a meeting host, a meeting participant, a department name, a project name, a specific work or task in a workflow, a technical field, and a past meeting conducted before the meeting. As a specific example, classification information for each of the meeting host, the meeting participant, the department name, the project name, the specific work or task in the workflow, the technical field, and the past meeting conducted before the meeting may be defined through a preset classification system distinguishing, for each item, a scope encompassing the meeting. In an embodiment, information about the classification system may be read from the memory 110 or may be generated or updated by an administrator account or a user account. In an embodiment, the classification system described above may include a plurality of levels having a hierarchical structure (e.g., a form in which an upper level encompasses a lower level) and / or a chronological structure (e.g., a form distinguishing a chronological order between a preceding level and a subsequent level). In an embodiment, the meeting information described above may further include, without being limited thereto, a meeting title concisely representing the meeting; a meeting goal indicating an outcome pursued or derivable through the meeting; or an introduction of either the meeting host or a user group (e.g., a company or department) of which the meeting host is a member. The meeting-room providing unit 111 may generate the online meeting room 10 based on the received meeting request. Generation of the online meeting room 10 may correspond to establishing a meeting. For example, the meeting-room providing unit 111 may generate a meeting identifier (e.g., a meeting ID) corresponding to the meeting request, generate the online meeting room 10 corresponding to the meeting identifier, and store and manage the online meeting room 10 in the database 400. The meeting-room providing unit 111 may perform the process of generating the online meeting room 10 in response to approval of the meeting request, without being limited thereto. For example, the meeting-room providing unit 111 may generate the online meeting room 10 in response to approval of the meeting request. As another example, upon obtaining the meeting request or video data, the meeting-room providing unit 111 may generate the online meeting room 10 and may store and manage the online meeting room 10 in the database 400 in response to approval of the meeting request or delete the online meeting room 10 in response to rejection of the meeting request. Furthermore, the database 400 may already store and manage information about a plurality of online meeting rooms 10 corresponding to a plurality of meeting identifiers and representing meetings generated before generation of the online meeting room 10 described above. The plurality of online meeting rooms 10 may be obtained and managed in the same or a similar manner as described throughout this specification. During generation of the online meeting room 10 according to an embodiment, whether the agent 112 participates in the online meeting room 10 may be determined, as described above. During receipt of the meeting request, the meeting-room providing unit 111 may obtain invitation information for participation of the agent 112 in a meeting and may cause the agent 112 to participate in the corresponding online meeting room 10 based on the obtained invitation information. For example, a meeting service application installed on the host terminal 200 may display an input menu for receiving meeting information for the meeting request and invitation information for an AI meeting participant, and the meeting-room providing unit 111 may receive a meeting request including the meeting information and the invitation information for the AI meeting participant entered by the host terminal 200 through the input menu. The invitation information described above may include whether the AI meeting participant is invited and / or setting information (e.g., a method of determining a response form, access authority, and the like). Based on the information about the meeting participants, the meeting-room providing unit 111 may check whether each meeting participant has provided a participation response to the meeting request. In an embodiment, the meeting-room providing unit 111 may transmit an invitation message including at least one of the meeting duration, the meeting content, information about the meeting host, and the information about the meeting participants to a respective participant terminal 300 of each meeting participant and may check whether a participation response corresponding to the invitation message is received from the respective participant terminal 300 of each meeting participant. If the meeting-room providing unit 111 receives an invitation acceptance response corresponding to an invitation message from a participant terminal 300, the meeting-room providing unit 111 may determine that a participation response of the corresponding meeting participant to the meeting request has been confirmed. Likewise, if the meeting-room providing unit 111 receives an invitation rejection response corresponding to the invitation message from the participant terminal 300, the meeting-room providing unit 111 may determine that a rejection response of the corresponding meeting participant to the meeting request has been confirmed. The meeting-room providing unit 111 may determine whether to approve the meeting request based on the participation responses of the respective meeting participants to the meeting request. In an embodiment, if the meeting-room providing unit 111 receives, within a preset deadline, participation responses corresponding to the invitation messages from respective participant terminals 300 of M meeting participants (M is a natural number) (e.g., M is 2 or greater) among the meeting participants, the meeting-room providing unit 111 may approve the meeting request. The meeting-room providing unit 111 may obtain video data. The video data according to an embodiment refers to data corresponding to a preset video format and may include, for example, data corresponding to any known file format used to store digital video data in a conventional computing environment. The meeting-room providing unit 111 may include the obtained video data in the content object 12 of the online meeting room 10. In an embodiment, the meeting-room providing unit 111 may obtain the video data from the host terminal 200 or a terminal (not illustrated) of another user designated by the host terminal 200. The terminal of the other user may refer to, for example, a user terminal corresponding to one of the user accounts using the meeting service, where the meeting host has granted the user account authority to upload video for the meeting. For example, the terminal of the other user may be a participant terminal 300 of one of the meeting participants or a user terminal unrelated to the meeting participants. In another embodiment, upon receiving a video generation request from the host terminal 200 (or the terminal of the other user described above), the meeting-room providing unit 111 may generate video data based on data received through interaction with the host terminal 200 (or the terminal of the other user described above). For example, the host terminal 200 may transmit the video generation request to the communication unit 120 through an installed meeting service application. In response to the video generation request, the meetingroom providing unit 111 may interact with the host terminal 200 through the meeting service application and may control the host terminal 200 to capture video. In still another embodiment, if the meeting request includes video data, the meetingroom providing unit 111 may obtain the video data from the meeting request. For example, if the meeting request or the meeting content includes video data, the meeting-room providing unit 111 may retrieve the video data from the meeting request or the meeting content. The meeting-room providing unit 111 may obtain the video data before receiving the meeting request, while receiving the meeting request, or after receiving the meeting request. Furthermore, in connection with obtaining the video data, the meeting-room providing unit 111 may provide the invitation messages described above and approve the meeting request based on the participation responses. In an embodiment, the meeting-room providing unit 111 may provide the online meeting room 10 to a respective participant terminal 300 of each meeting participant having a confirmed participation response for the online meeting room 10. For example, under the embodiments described above, the meeting-room providing unit 111 may provide the online meeting room 10 in real time or on a non-real-time basis exclusively to participant terminals 300 having participation responses confirmed as a result of the participant terminals 300 transmitting invitation acceptance responses corresponding to the invitation messages described above. In an embodiment, the meeting-room providing unit 111 may manage each conversation received from the respective meeting participants in association with one of a plurality of preset message types. The plurality of message types according to an embodiment may include at least one selected from the group consisting of a question, an answer, an opinion, a department, a participant, and a keyword. For example, the meeting-room providing unit 111 may provide, in a chat room implemented through the online meeting room 10, both a message input field for receiving a message and a selection button for selecting a message type of the message. The meeting-room providing unit 111 may receive each message together with the message type, associate each message with the corresponding message type, and store and manage each message in the corresponding conversation thread object 13. Accordingly, the meeting-room providing unit 111 may manage each message in the conversation thread object 13 of each of the plurality of online meeting rooms 10 stored in the database 400 in association with one of the plurality of message types. FIG. 14 is an exemplary flowchart illustrating a computer-implemented method performed by the meeting service apparatus 100 for providing an artificial intelligence-based meeting service. FIG. 14 may be understood with reference to all embodiments described above. In an embodiment, the meeting service apparatus 100 may perform the operations illustrated in FIG. 14. In an additional embodiment, a single entity may implement the operations illustrated in FIG. 14, for example, by performing the operations on a server. In another additional embodiment, a plurality of entities may implement the operations illustrated in FIG. 14, for example, by performing some of the operations illustrated in FIG. 14 on a first server and performing other operations on a second server (or a terminal). Referring to FIG. 14, in operation S1410, the agent 112 may obtain conversations received from meeting participants participating in the online meeting room 10 including the agent 112 connected to the language model 113. The agent 112 may be a meeting participant object participating in the online meeting room 10 together with the meeting participants and may be granted the same authority to input and receive conversation as the meeting participants. In operation S1420, the agent 112 may select a target conversation from the obtained conversations, the target conversation comprising a conversation requiring a response using local internal data 410. The agent 112 may select a target conversation from the obtained conversations using the preset selection rule, the target conversation comprising a conversation requiring a response using local internal data 410. Alternatively, the agent 112 may transmit the obtained conversations to the language model 113, and the language model 113 may select the target conversation. The language model 113 may be trained to select, from the conversations, a conversation requiring a response using local internal data 410 based on a plurality of conversations and information indicating whether a response is required. The agent 112 may select, from the conversations, one of a first target conversation requiring a response using local internal data 410, a second target conversation requiring a response using specialized field data 420, and a third target conversation requiring a response using public data 430. In operation S1430, the agent 112 may provide the target conversation as input data to the language model 113, and in operation S1440, the agent 112 may obtain output data from the language model 113. The language model 113 may generate a natural language response to a natural language input using local internal data 410. The language model 113 may generate a natural language response to a natural language input using, among the local internal data 410, local internal data assigned a security level at least partially corresponding to an access authority granted to each meeting participant. The language model 113 may be a generative AI model. The language model 113 may be a model pretrained using local internal data 410. The language model 113 may be further trained (e.g., through pre-meeting training) using meeting information of the online meeting room 10 and / or the conversations. The agent 112 may provide the target conversation to the language model 113 trained using local internal data 410, and the language model 113 may generate output data for the target conversation. The meeting information of the online meeting room 10 used for further training may include at least one selected from the group consisting of (i) a meeting duration, (ii) meeting content, (iii) information about the meeting participants, (iv) meeting classification information, and (v) a meeting history indicating a historical relationship between the online meeting room 10 and one or more other online meeting rooms 10 associated with the online meeting room 10. Alternatively, the agent 112 may obtain, from local internal data 410, a search result for local internal data 410 associated with the target conversation, generate a prompt to instruct the language model 113 to generate the output data as a response to the target conversation using the search result, provide the generated prompt to the language model 113, and obtain the output data from the language model 113. The prompt may instruct the language model 113 to generate the output data by further using the conversations and / or meeting information of the online meeting room 10. The language model 113 may be further trained (e.g., through pre-meeting training) using conversations from meetings prior to generation of the online meeting room 10. Alternatively, local internal data 410 may include conversations from meetings prior to generation of the online meeting room 10. The agent 112 may provide the target conversation and the conversations to the language model 113 and request the language model 113 to generate a response based on further training using the conversations (e.g., real-time training). The language model 113 may generate output data for the target conversation based on further training using the provided conversations (e.g., real-time training). The language model 113 may be further trained using training data including predefined reference conversations. The reference conversations may include a plurality of conversations among a plurality of users for each organization and each situation. The language model 113 may be obtained by applying further training using the reference conversations as training data to a generative AI model pretrained to converse with a user using public data. The language model 113 may be a model trained using specialized field data 420 associated with a preset specialized field. The language model 113 may be obtained by applying further training using specialized field data 420 to a generative AI model pretrained to converse with a user based on public data. The language model 113 may be obtained by finetuning, using private data associated with the specialized field, a model pretrained using public data associated with the specialized field. If the first target conversation is selected, the agent 112 may provide the first target conversation as input data to the language model 113, for example, the Colleague model 1131, and obtain, from the Colleague model 1131, output data for the first target conversation generated using local internal data 410. If the second target conversation is selected, the agent 112 may provide the second target conversation as input data to the language model 113, for example, the Expert model 1132, and obtain, from the Expert model 1132, output data for the second target conversation generated using specialized field data 420. The language model 113 may generate the output data in at least one response form selected from the group consisting of a question, an answer, an opinion, and a summary. The language model 113 may determine the response form based on the target conversation, or the agent 112 may determine the response form based on the target conversation. In operation S1450, the agent 112 may provide, based on the obtained output data, a conversation response to the target conversation to a participant terminal 300 of a meeting participant. The agent 112 may provide the conversation response to the participant terminal 300 as a message from an AI meeting participant visually distinguishable from the meeting participants. The agent 112 may determine whether to provide the conversation response based on an access authority granted to each of the meeting participants. The agent 112 may provide, together with the conversation response, basis data used to obtain the output data from among local internal data 410 and / or an access link to the basis data. If the target conversation is a conversation for which a response has been individually requested by a participant terminal 300 of one of the meeting participants, the agent 112 may provide the conversation response exclusively to the participant terminal 300. The agent 112 may obtain, from the conversations, a meeting result report including a summary result for each of a plurality of criterion items and provide the obtained meeting result report to the participant terminal 300. The plurality of criterion items may include at least one selected from the group consisting of a meeting outline item, a meeting content item, a main conversation collection item, a task item, and a next meeting item. The agent 112 may schedule the next meeting item based on an approval response from one of the meeting participants for the next meeting item. Additionally or alternatively, the agent 112 may track the task item based on information regarding task progress and / or a task result received from a participant terminal 300 associated with the task item. In an embodiment, an AI meeting participant may provide reliable responses using local internal data 410 during a meeting. Furthermore, by selectively responding to conversations requiring responses among the conversations, the AI meeting participant may naturally participate in a multiparty meeting conversation without disrupting the conversational flow of the meeting participants. Consequently, the AI meeting participant may participate in the meeting at the same level as the other meeting participants as a colleague AI well versed in local internal data 410 and capable of engaging in conversation. In another embodiment, an AI meeting participant may provide specialized consulting regarding a preset specialized field using specialized field data 420. Based on intensive training in specialized knowledge and know-how specific to the preset specialized field, the AI meeting participant may operate as an expert AI capable of providing consulting services for various situations requiring problem solving in the preset specialized field. In still another embodiment, an AI meeting participant may appropriately participate in conversations according to an organization and a situation in a meeting environment involving a plurality of meeting participants by using organization-specific and situation-specific reference conversations. Based on intensive training in conversation skills for multiparty meetings, the AI meeting participant may operate as a participatory AI autonomously participating in conversations appropriate for the organization and the situation during a meeting involving multiple participants without interrupting the conversational flow. An implementation includes a computer program stored in a non-transitory computer-readable storage medium. The computer program includes instructions for execution by one or more processors 130. Upon execution, the instructions cause the one or more processors 130 to perform a computer-implemented method including: obtaining, by the agent 112, conversations received from meeting participants participating in the online meeting room 10 including the agent 112 connected to the language model 113; selecting, by the agent 112, a target conversation from the obtained conversations, the target conversation comprising a conversation requiring a response using local internal data 410; providing, by the agent 112, the target conversation as input data to the language model 113, the language model 113 generating a natural language response to a natural language input using local internal data 410; obtaining, by the agent 112, output data from the language model 113; and providing, by the agent 112, based on the output data, a conversation response to the target conversation to a participant terminal 300. An implementation includes a non-transitory computer-readable storage medium storing a computer program. The computer program includes instructions for execution by one or more processors. Upon execution, the instructions cause the one or more processors to perform a computer-implemented method including: obtaining, by the agent 112, conversations received from meeting participants participating in the online meeting room 10 including the agent 112 connected to the language model 113; selecting, by the agent 112, a target conversation from the obtained conversations, the target conversation comprising a conversation requiring a response using local internal data 410; providing, by the agent 112, the target conversation as input data to the language model 113, the language model 113 generating a natural language response to a natural language input using local internal data 410; obtaining, by the agent 112, output data from the language model 113; and providing, by the agent 112, based on the output data, a conversation response to the target conversation to a participant terminal 300. In an implementation, a meeting service apparatus 100 including a memory 110, a processor 130, and one or more programs stored in the memory 110 and configured to be executed by the processor 130 performs a computer-implemented method. The computer-implemented method may include: receiving, from a host terminal 200, a meeting request including a meeting duration and information about meeting participants; generating, based on the meeting request, an online meeting room 10 including a conversation thread based on conversation input of the meeting participants, the online meeting room 10 including an agent 112 connected to a language model 113, and the language model 113 generating a natural language response to a natural language input using local internal data 410; controlling an online meeting to proceed with non-real-time participation by the meeting participants by providing the online meeting room 10 through a network during the meeting duration in response to a request from the respective participant terminal 300 of each meeting participant; obtaining, by the agent 112, conversations received from the meeting participants participating in the online meeting room 10; selecting, by the agent 112, a target conversation from the obtained conversations, the target conversation comprising a conversation requiring a response using local internal data 410; providing, by the agent 112, the target conversation as input data to the language model 113; obtaining, by the agent 112, output data from the language model 113; and providing, by the agent 112, based on the output data, a conversation response to the target conversation to the participant terminal 300. In another implementation, a meeting service apparatus 100 including a memory 110, a processor 130, and one or more programs stored in the memory 110 and configured to be executed by the processor 130 performs a computer-implemented method. The computer-implemented method may include: obtaining, by an agent 112, conversations received from meeting participants participating in an online meeting room 10 including the agent 112 connected to a language model 113; selecting, by the agent 112, a target conversation from the obtained conversations, the target conversation comprising a conversation requiring a response using specialized field data 420; providing, by the agent 112, the target conversation as input data to the language model 113, the language model 113 generating a natural language response to a natural language input using specialized field data 420; obtaining, by the agent 112, output data from the language model 113; and providing, by the agent 112, based on the output data, a conversation response to the target conversation to a participant terminal 300 of a meeting participant. In another implementation, a meeting service apparatus 100 including a memory 110, a processor 130, and one or more programs stored in the memory 110 and configured to be executed by the processor 130 performs a computer-implemented method. The computer-implemented method may include: obtaining, by an agent 112, conversations received from meeting participants participating in an online meeting room 10 including the agent 112 connected to a language model 113; selecting, by the agent 112, a target conversation from the obtained conversations, the target conversation comprising a conversation requiring a response using public data 430; providing, by the agent 112, the target conversation as input data to the language model 113, the language model 113 being trained to generate a natural language response to a natural language input using public data 430; obtaining, by the agent 112, output data from the language model 113; and providing, by the agent 112, based on the output data, a conversation response to the target conversation to a participant terminal 300 of a meeting participant. In another implementation, a meeting service apparatus 100 including a memory 110, a processor 130, and one or more programs stored in the memory 110 and configured to be executed by the processor 130 performs a computer-implemented method. The computer-implemented method may include: obtaining, by an agent 112, conversations received from meeting participants participating in an online meeting room 10 including the agent 112 connected to a language model 113; providing, by the agent 112, the obtained conversations as input data to the language model 113, the language model 113 being trained to generate a natural language response to a natural language input using training data including predefined reference conversations; obtaining, by the agent 112, from the language model 113, output data for a conversation requiring a response from among the conversations; and providing, by the agent 112, based on the output data, a conversation response to a participant terminal 300 of a meeting participant. In still another implementation, a meeting service apparatus 100 including a memory 110, a processor 130, and one or more programs stored in the memory 110 and configured to be executed by the processor 130 performs a method for obtaining a language model 113. The method for obtaining the language model 113 may include: obtaining a model pretrained to generate a natural language response to a natural language input; obtaining a plurality of training datasets based on local internal data 410, each training dataset including training input data including a conversation about local internal data 410 and training ground-truth data including a response using local internal data 410 associated with the conversation; and obtaining the language model 113 by applying fine-tuning to the pretrained model using the plurality of training datasets. In still another implementation, a meeting service apparatus 100 including a memory 110, a processor 130, and one or more programs stored in the memory 110 and configured to be executed by the processor 130 performs a method for obtaining a language model 113. The method for obtaining the language model 113 may include: obtaining a model pretrained to generate a natural language response to a natural language input; obtaining a plurality of training datasets based on specialized field data 420, each training dataset including training input data including a conversation about the corresponding specialized field and training ground-truth data including a response using specialized field data 420 associated with the conversation; and obtaining the language model 113 by applying fine-tuning to the pretrained model using the plurality of training datasets. In still another implementation, a meeting service apparatus 100 including a memory 110, a processor 130, and one or more programs stored in the memory 110 and configured to be executed by the processor 130 performs a method for obtaining a language model 113. The method for obtaining the language model 113 may include: obtaining a model pretrained to generate a natural language response to a natural language input; obtaining training data including predefined reference conversations; and obtaining the language model 113 by applying fine-tuning to the pretrained model using the training data. As described above, the method for obtaining the language model 113 described in this specification may train the language model 113 to generate a natural language response to a natural language input using local internal data 410. An implementation may independently use the technical feature of training the language model 113 without combining the technical feature with the technical feature of providing a conversation response using the trained language model 113. Likewise, the technical feature of providing a conversation response using the language model 113 may be used independently without being combined with the technical feature of training the language model 113. The operations illustrated above are examples. Various implementations may add, omit, or modify the order, combination, branching, functions, and performing entities of the operations without departing from the essential characteristics of the respective technical features described throughout this specification. In an embodiment, the language model in this specification may refer to any type of computer program operating based on a network function, an artificial neural network, and / or a neural network. Throughout this specification, model, neural network, network function, and neural network may be used interchangeably. In a neural network, one or more nodes interconnected through one or more links form relationships between input nodes and output nodes within the neural network. The number of nodes and links in the neural network, relationships among the nodes and links, and weight values assigned to the respective links may determine characteristics of the neural network. A neural network may include a set of one or more nodes. A subset of the nodes constituting the neural network may form a layer. A deep neural network (DNN) may refer to a neural network including a plurality of hidden layers in addition to an input layer and an output layer. A deep neural network may include a convolutional neural network (CNN), a recurrent neural network (RNN), an autoencoder, a generative adversarial network (GAN), a restricted Boltzmann machine (RBM), a deep belief network (DBN), a Q-network, a U-network, a Siamese network, a transformer, and the like. The foregoing description of deep neural networks is illustrative and does not limit the disclosure. A neural network may be trained using at least one of supervised learning, unsupervised learning, semi-supervised learning, self-supervised learning, and reinforcement learning. Training a neural network may involve applying knowledge for performing a specific operation to the neural network. A neural network may be trained to minimize output errors. Training a neural network may include repeatedly inputting training data to the neural network, calculating an error between an output of the neural network for the training data and a target, backpropagating the error of the neural network from the output layer toward the input layer to reduce the error, and updating a weight of each node of the neural network. Supervised learning uses labeled data including a ground-truth label for each training data item, whereas unsupervised learning may use unlabeled data lacking a ground-truth label for each training data item. A learning rate may determine an amount of change in each updated connection weight of a node. Calculation by the neural network for input data and backpropagation of an error may constitute a training cycle. The learning rate may vary based on the number of repetitions of the training cycle of the neural network. Furthermore, methods such as increasing the training data, regularization, dropout for deactivating some nodes, and using a batch normalization layer may be applied to prevent overfitting. In an embodiment, a language model may employ at least a portion of a transformer. The transformer may include an encoder encoding embedded data and a decoder decoding the encoded data. The transformer may receive a sequence of data and have a structure that outputs a sequence of data of a different type through encoding and decoding operations. In an embodiment, a sequence of data may be processed into a form processable by the transformer. Processing the sequence of data into the form processable by the transformer may include embedding. Expressions such as data token, embedding vector, and embedding token may refer to data embedded in a form processable by the transformer. To encode and decode a sequence of data, encoders and decoders in the transformer may perform processing using an attention algorithm. An attention algorithm may refer to an algorithm that calculates similarities to one or more keys for a given query, applies the calculated similarities to values respectively corresponding to the keys, and calculates an attention value by calculating a weighted sum of the values to which the similarities are applied. Various types of attention algorithms may be classified based on settings of a query, a key, and a value. For example, if the query, key, and value are all set identically, obtaining attention may constitute a self-attention algorithm. To process an input sequence of data in parallel, the transformer may reduce the dimensionality of an embedding vector, divide the embedding vector, and obtain attention by calculating an individual attention head for each divided embedding vector. Such processing may constitute a multi-head attention algorithm. In an embodiment, the transformer may include modules performing a plurality of multi-head self-attention algorithms or multi-head encoder-decoder algorithms. The transformer may further include additional components other than attention algorithms, such as embedding, normalization, and softmax. Methods of constructing a transformer using an attention algorithm may include the method disclosed in Vaswani et al., “Attention Is All You Need,” NIPS 2017. The method disclosed by Vaswani et al. is incorporated herein by reference. The transformer may transform a sequence of input data into a sequence of output data in various data domains, such as embedded natural language, divided image data, and audio waveforms. To convert data in various data domains into sequences of data suitable for input to the transformer, the transformer may embed the data. The transformer may process additional data representing relative positional relationships or phase relationships among the sequence of input data. Alternatively, the transformer may embed the sequence of input data by additionally incorporating vectors representing relative positional relationships or phase relationships among the input data. Examples of the relative positional relationships among the sequence of input data include word order in a natural language sentence, relative positional relationships among respective divided images, and a temporal order of divided audio waveforms, without limitation. A process of adding information representing the relative positional relationships or phase relationships among the sequence of input data may be referred to as positional encoding. In an embodiment, the language model may be a model trained through transfer learning. Transfer learning refers to a training method that includes obtaining a pretrained model for a first task by pretraining a model using a large volume of unlabeled training data through semi-supervised learning or self-supervised learning, and implementing a target model by fine-tuning the pretrained model for a second task using labeled training data through supervised learning. In an embodiment, fine-tuning may encompass transferring a pretrained model to a preset target task and subsequently training the model. Specifically, fine-tuning may refer to a method of taking a pretrained model, modifying the pretrained model for a task of generating a conversation response using the database 400, and updating the pretrained model through training starting from the weights of the pretrained model. For example, such fine-tuning may include updating parameters of the pretrained model by further training the pretrained model using a specific dataset for generating a conversation response. A person of ordinary skill in the art will understand that electronic hardware, various forms of program or design code (referred to herein as software for convenience), or combinations thereof may implement the various exemplary logical blocks, modules, processors, units, circuits, and algorithm operations described in connection with the embodiments disclosed herein. To clearly describe the interchangeability of hardware and software, the foregoing generally describes various exemplary components, blocks, modules, circuits, and operations in terms of their functionality. Whether hardware or software implements such functionality depends on a particular application and design constraints imposed on the overall system. A person of ordinary skill in the art may implement the described functionality in various manners for each particular application, and such implementation decisions remain within the scope of the present disclosure. The various embodiments presented herein may be implemented as a method, an apparatus, or an article of manufacture using standard programming and / or engineering techniques. The term article of manufacture includes a computer program, carrier, or medium accessible from any computer-readable storage apparatus. For example, computer-readable storage media include magnetic storage apparatuses (e.g., hard disks, floppy disks, and magnetic strips), optical disks (e.g., CDs and DVDs), smart cards, and flash memory apparatuses (e.g., EEPROMs, cards, sticks, and key drives), without limitation. Furthermore, the various storage media presented herein include one or more apparatuses and / or other machine-readable media for storing information. The particular order or hierarchy of operations in the presented processes is an example of an exemplary approach. Based on design priorities, the particular order or hierarchy of operations in the processes may be rearranged within the scope of the present disclosure. The appended method claims recite elements of various operations in a sample order without limiting the claimed subject matter to the presented particular order or hierarchy.

Claims

1. A computer-implemented method performed by a meeting service apparatus comprising a memory, a processor, and one or more programs stored in the memory and configured to be executed by the processor, the computer-implemented method comprising:obtaining, by an agent, conversations received from meeting participants participating in an online meeting room including the agent connected to a language model;selecting, by the agent, a target conversation from the conversations, wherein the target conversation comprises a conversation requiring a response using local internal data;providing, by the agent, the target conversation as input data to the language model, the language model generating a natural language response to a natural language input using the local internal data;obtaining, by the agent, output data for the input data from the language model; andproviding, by the agent, based on the output data, a conversation response to the target conversation to a participant terminal of one of the meeting participants.

2. The method of claim 1, wherein the agent is a meeting participant object participating in the online meeting room together with the meeting participants and is granted the same authority to input and receive conversation as the meeting participants.

3. The method of claim 2, further comprising generating the online meeting room based on a meeting request from a host terminal, wherein, in generating the online meeting room, whether the agent participates is determined.

4. The method of claim 3, further comprising:after participation of the agent is determined, obtaining, by the meeting service apparatus, meeting information of the online meeting room and, before the online meeting room is started, obtaining, by the meeting service apparatus, external data related to the meeting information; andperforming further training of the language model using the external data related to the meeting information.

5. The method of claim 1, wherein selecting the target conversation comprises selecting, by the agent, the target conversation using a preset selection rule, the preset selection rule being based on at least one selected from the group consisting of:a type of conversation;an intention of conversation;whether another meeting participant is addressed;whether an artificial intelligence (AI) meeting participant is addressed;a context of conversations;a time interval of conversations;a count of conversations during a given time period; andwhether text related to the local internal data is included.

6. The method of claim 5, wherein, in providing the target conversation as input data to the language model, the agent provides the target conversation and the conversations to the language model, and wherein, in obtaining the output data from the language model, the language model generates the output data based on further training using the conversations.

7. The method of claim 1, wherein the language model is trained, based on a plurality of conversations and whether a response is required, to select, from the plurality of conversations, a conversation requiring a response, and the agent transmits the received conversations to the language model.

8. The method of claim 1, wherein the local internal data is assigned a security level associated with a user's access authority, and the access authority is set in consideration of at least one selected from the group consisting of: rank; position; document security level; department; employment type; workplace; and affiliate.

9. The method of claim 8, wherein the language model generates a natural language response to a natural language input using, among the local internal data, local internal data having a security level corresponding to an access authority granted to each of the meeting participants.

10. The method of claim 1, wherein the local internal data comprises local internal data associated with an organization to which at least some of the meeting participants belong, and the local internal data associated with the organization is categorized based on one or more types selected from the group consisting of human resources, finance, accounting, legal, management, research, development, production and sales.

11. The method of claim 1, wherein the language model is a generative artificial intelligence (AI) model.

12. The method of claim 11, wherein the generative AI model is pretrained using the local internal data.

13. The method of claim 12, wherein the generative AI model is further trained using meeting information of the online meeting room and / or the conversations.

14. The method of claim 13, wherein the local internal data comprises conversations from meetings prior to generation of the online meeting room.

15. The method of claim 13, wherein the generative AI model is further trained using conversations from meetings prior to generation of the online meeting room.

16. The method of claim 13, wherein the meeting information of the online meeting room comprises at least one selected from the group consisting of:(i) a meeting duration;(ii) meeting content;(iii) information about the meeting participants;(iv) meeting classification information; and(v) a meeting history indicating a historical relationship between the online meeting room and one or more other online meeting rooms associated with the online meeting room.

17. The method of claim 11, wherein the generative AI model is further trained using training data comprising predefined reference conversations.

18. The method of claim 11, wherein the generative AI model is trained using specialized field data associated with a preset specialized field.

19. The method of claim 11, wherein obtaining the output data by the agent comprises:obtaining, by the agent, from the local internal data, a search result for local internal data associated with the target conversation;generating, by the agent, a prompt to instruct the generative AI model to generate the output data as a response to the target conversation using the search result; andproviding, by the agent, the prompt to the generative AI model to obtain the output data from the generative AI model.

20. The method of claim 19, wherein the prompt instructs the generative AI model to generate the output data by further using the conversations and / or meeting information of the online meeting room.

21. The method of claim 1, wherein selecting the target conversation comprises selecting, by the agent, from the conversations, one of a first target conversation requiring a response using the local internal data and a second target conversation requiring a response using specialized field data,wherein the language model includes a first model generating a natural language response to a natural language input using the local internal data and a second model trained to generate a natural language response to a natural language input using the specialized field data associated with a preset specialized field, andwherein, in providing the target conversation as the input data, the agent provides the first target conversation as the input data to the first model if the first target conversation isselected and provides the second target conversation as the input data to the second model if the second target conversation is selected.

22. The method of claim 1, wherein the language model is stored either in the memory or in an external apparatus connected to the meeting service apparatus through a network and is executed by the agent.

23. The method of claim 1, wherein the language model generates the output data in at least one response form selected from the group consisting of:a question;an answer;an opinion; anda summary.

24. The method of claim 23, wherein the at least one response form is determined either (1) by the language model based on the target conversation or (2) by the agent based on the target conversation.

25. The method of claim 1, wherein providing the conversation response comprises providing the conversation response to the participant terminal in the form of a message from an artificial intelligence (AI) meeting participant visually distinguishable from the meeting participants.

26. The method of claim 1, wherein providing the conversation response comprises determining whether to provide the conversation response based on an access authority granted to each of the meeting participants.

27. The method of claim 1, wherein providing the conversation response comprises providing, to a respective participant terminal of each of the meeting participants, the conversation response obtained using the local internal data assigned a security level corresponding to an access authority of the corresponding meeting participant.

28. The method of claim 1, wherein providing the conversation response comprises providing, together with the conversation response, basis data, among the local internal data, used to obtain the output data and / or an access link to the basis data.

29. The method of claim 1, wherein, if the target conversation is a conversation for which a response has been individually requested by the participant terminal, providing the conversation response comprises providing the conversation response exclusively to the participant terminal.

30. The method of claim 1, further comprising:obtaining, by the agent, from the conversations, a meeting result report comprising a summary result for each of a plurality of preset criteria items, the plurality of preset criteria items comprising at least one selected from the group consisting of a meeting outline item, a meeting content item, a main conversation collection item, a task item, and a next meeting item; andproviding, by the agent, the meeting result report to the participant terminal.

31. The method of claim 30, further comprising:(1) scheduling, by the agent, the next meeting item based on an approval response from one of the meeting participants for the next meeting item; and / or(2) tracking, by the agent, the task item based on information regarding task progress and / or a task result received from a participant terminal associated with the task item.

32. The method of claim 1, further comprising:before receiving the conversations, receiving, from a host terminal, a meeting request comprising a meeting duration, meeting content, and information about the meeting participants;generating, based on the meeting request, the online meeting room comprising a conversation thread based on conversation input of the meeting participants; andcontrolling an online meeting to proceed with non-real-time participation by the meeting participants by providing the online meeting room through a network in response to a request from a respective participant terminal of each of the meeting participants during the meeting duration.

33. The method of claim 32, further comprising obtaining video data related to the meeting content, wherein the online meeting room further comprises the video data.

34. A meeting service apparatus, comprising:a memory in which both a language model generating a natural language response to a natural language input using local internal data and at least one instruction are stored; anda processor configured to execute the at least one instruction stored in the memory, wherein the at least one instruction, when executed by the processor, causes the processor to perform a method comprising:obtaining, by an agent, conversations received from meeting participants participating in an online meeting room including the agent connected to the language model;selecting, by the agent, a target conversation from the conversations, the target conversation comprising a conversation requiring a response using the local internal data;providing, by the agent, the target conversation as input data to the language model;obtaining, by the agent, output data for the input data from the language model; andproviding, by the agent, based on the output data, a conversation response to the target conversation to a participant terminal of one of the meeting participants.

35. A computer program stored in a non-transitory computer-readable storage medium, the computer program comprising instructions that, when executed by one or more processors, cause the one or more processors to perform a computer-implemented method comprising:obtaining, by an agent, conversations received from meeting participants participating in an online meeting room including the agent connected to a language model;selecting, by the agent, a target conversation from the conversations, the target conversation comprising a conversation requiring a response using local internal data;providing, by the agent, the target conversation as input data to the language model, the language model generating a natural language response to a natural language input using the local internal data;obtaining, by the agent, output data for the input data from the language model; andproviding, by the agent, based on the output data, a conversation response to the target conversation to a participant terminal of one of the meeting participants.

36. A non-transitory computer-readable storage medium storing a computer program comprising instructions that, when executed by one or more processors, cause the one or more processors to perform a computer-implemented method comprising:obtaining, by an agent, conversations received from meeting participants participating in an online meeting room including the agent connected to a language model;selecting, by the agent, a target conversation from the conversations, the target conversation comprising a conversation requiring a response using local internal data;providing, by the agent, the target conversation as input data to the language model, the language model generating a natural language response to a natural language input using the local internal data;obtaining, by the agent, output data for the input data from the language model; andproviding, by the agent, based on the output data, a conversation response to the target conversation to a participant terminal of one of the meeting participants.

Images