A Multi-Agent Collaborative Outbound Call Interaction Content Adaptive Generation System and Method

Abstract

This invention discloses a multi-agent collaborative outbound call interactive content adaptive generation system and method. The system includes: a receiving module that receives user instructions and analyzes and summarizes them into user needs; a splitting and calling module that breaks down the user needs into multiple sub-tasks and collaboratively calls preset agents to execute the corresponding sub-tasks in sequence; wherein the preset agents include at least one of an interactive content generation agent, an interactive content adaptive optimization agent, a data analysis agent, and a work summary agent, thereby adaptively completing repetitive tasks such as interactive content writing, optimization, data analysis, and report generation, saving labor costs. Simultaneously, an interactive content generation model integrating a scenario knowledge base, scenario-supervised learning, and reward model training ensures that the generated new interactive content is both relevant to the business scenario and complies with compliance requirements, effectively improving the quality of interactive content generation.

Description

Technical Field

[0001] This invention relates to the field of speech processing, and more specifically, to a system and method for adaptive generation of outbound call interactive content through multi-agent collaboration. Background Technology

[0002] Intelligent voice robots can replace humans in conversations with users, thus reducing labor costs in telemarketing. Currently, most mainstream online telemarketing robots adopt a retrieval-based dialogue model architecture. At its core is a pre-defined dialogue flow tree. This flow tree consists of numerous nodes, each representing a specific stage of the conversation with the customer (e.g., opening remarks, product introduction, Q&A, closing the deal, closing remarks, etc.). Each node is pre-bound with one or more standard response interactions. When the telemarketing robot reaches a certain node, it selects one from the bound response interactions and outputs it to the customer. After the customer responds, the telemarketing robot first identifies the customer's intent in the response and, based on the identified intent, jumps from the current node to the next node, thus propelling the conversation to the next stage.

[0003] It is evident that the conversational communication capabilities of telemarketing outbound call robots are influenced by two factors: the quality of the interactive content and the selection of the navigation node. Regarding improving the quality of interactive content, traditional solutions require interactive content experts to manually annotate a large number (e.g., tens of thousands) of high-quality question-and-answer pairs, a process that is time-consuming and inefficient. Simultaneously, it necessitates manually writing and modifying each response from the voice robot, which is inefficient and difficult to scale. With the emergence of Large Language Models (LLMs), a simpler solution directly utilizes an external large language model to generate optimized interactive content from the original content. However, this solution still has several significant problems: (1) External LLMs lack specific business scenario knowledge and cannot fully understand the semantics of specific business scenarios. The generated interactive content cannot fit the specific business scenario and lacks appeal.

[0004] (2) External LLMs often have illusions, and the generated interactive content often has compliance issues, such as not stating their identity at the beginning of the dialogue, the promised content not matching the actual content, and the incorrect use of variables. Summary of the Invention

[0005] In view of this, the main objective of this invention is to propose a multi-agent collaborative outbound call interactive content adaptive generation system and method, in order to at least partially solve at least one of the above-mentioned technical problems.

[0006] To address the aforementioned technical problems, the first aspect of this invention proposes a multi-agent collaborative outbound call interaction content adaptive generation system, comprising: The receiving module is used to receive user instructions and analyze and summarize them into user requirements. The decomposition and invocation module is used to decompose the user request into multiple subtasks to be executed sequentially, and to coordinately invoke preset intelligent agents to execute the corresponding subtasks according to the order; wherein, the preset intelligent agents include at least one of the following intelligent agents: An interactive content generation agent is used to retrieve current scene knowledge and current node information from a scene knowledge base to construct prompt words. The interactive content generation model generates new interactive content based on the original interactive content of the current node and the prompt words. Specifically, the interactive content generation model completes optimization training through a multi-level preference optimization process, with the training phase including at least supervised learning, direct preference optimization, and group relative strategy optimization stages. In the supervised learning stage, the interactive content generation model is trained using telemarketing vertical training data and outputs historical interactive content. The compliance of the historical interactive content is then judged by a reward module to obtain positive and negative samples. Then, direct preference optimization training and group relative strategy optimization training are carried out sequentially. An interactive content adaptive optimization agent is used to adaptively optimize newly generated interactive content based on user needs, and generate optimized interactive content. A data analysis intelligent agent is used to acquire dialogue parameter change data for each node, detect whether there is abnormal data and / or data that needs optimization in the dialogue parameter change data, and identify the nodes to be optimized. The work summary agent is used to extract the original work content of user interactions within a predetermined time period and generate a work summary in a standard format that includes the predetermined content.

[0007] According to a preferred embodiment of the present invention, the scenario knowledge base stores the dialogue objectives, user preferences, resources enjoyed by the user groups, interaction content requirements, and node information of each user group in different scenarios.

[0008] According to a preferred embodiment of the present invention, the reward module breaks down compliance into multiple sub-compliance conditions, outputs a sub-score by judging the degree of compliance of historical interaction content with each sub-compliance condition, determines the compliance score of historical interaction content by combining the sub-scores, and selects positive and negative samples from historical interaction content based on the compliance score.

[0009] According to a preferred embodiment of the present invention, the receiving module includes: The real-time receiving unit is used to receive user instructions and analyze and summarize them in real time to obtain real-time user requirements. The extraction unit is used to extract historical user demands associated with the real-time user demands from the historical demand database; The summary unit is used to summarize real-time user needs and related historical user needs to obtain user requirements; Wherein: the user needs include at least one of the following: requirements, feedback, and suggestions.

[0010] According to a preferred embodiment of the present invention, when the data analysis agent detects a node to be optimized, it calls the interaction content generation agent to generate new interaction content based on the original interaction content of the node to be optimized.

[0011] To address the aforementioned technical problems, a second aspect of the present invention provides a method for adaptively generating outbound call interaction content through multi-agent collaboration, comprising: Receive user instructions and analyze and summarize them into user requirements; The user requirements are broken down into multiple sub-tasks that are executed sequentially, and a preset intelligent agent is collaboratively invoked to execute the corresponding sub-tasks according to the order; wherein the preset intelligent agent includes at least one of the following intelligent agents: An interactive content generation agent is used to retrieve current scene knowledge and current node information from a scene knowledge base to construct prompt words. The interactive content generation model generates new interactive content based on the original interactive content of the current node and the prompt words. Specifically, the interactive content generation model completes optimization training through a multi-level preference optimization process, with the training phase including at least supervised learning, direct preference optimization, and group relative strategy optimization stages. In the supervised learning stage, the interactive content generation model is trained using telemarketing vertical training data and outputs historical interactive content. The compliance of the historical interactive content is then judged by a reward module to obtain positive and negative samples. Then, direct preference optimization training and group relative strategy optimization training are carried out sequentially. An interactive content adaptive optimization agent is used to adaptively optimize newly generated interactive content based on user needs, and generate optimized interactive content. A data analysis intelligent agent is used to acquire dialogue parameter change data for each node, detect whether there is abnormal data and / or data that needs optimization in the dialogue parameter change data, and identify the nodes to be optimized. The work summary agent is used to extract the original work content of user interactions within a predetermined time period and generate a work summary in a standard format that includes the predetermined content.

[0012] According to a preferred embodiment of the present invention, the scenario knowledge base stores the dialogue objectives, user preferences, resources enjoyed by the user groups, interaction content requirements, and node information of each user group in different scenarios.

[0013] According to a preferred embodiment of the present invention, the reward module breaks down compliance into multiple sub-compliance conditions, outputs a sub-score by judging the degree of compliance of historical interaction content with each sub-compliance condition, determines the compliance score of historical interaction content by combining the sub-scores, and selects positive and negative samples from historical interaction content based on the compliance score.

[0014] According to a preferred embodiment of the present invention, receiving user instructions and analyzing and summarizing them into user needs includes: Receive user instructions and analyze and summarize them in real time to obtain real-time user needs; Extract historical user demands associated with the real-time user demands from the historical demand database; User requirements are derived by summarizing real-time user requirements and related historical user requirements; Wherein: the user needs include at least one of the following: requirements, feedback, and suggestions.

[0015] According to a preferred embodiment of the present invention, when the data analysis agent detects a node to be optimized, it calls the interaction content generation agent to generate new interaction content based on the original interaction content of the node to be optimized.

[0016] To address the aforementioned technical problems, a third aspect of the present invention provides an intelligent agent for adaptively generating outbound call interaction content, wherein the intelligent agent executes the method described in any of the preceding claims.

[0017] To address the aforementioned technical problems, a fourth aspect of the present invention provides an electronic device, comprising: A processor; and a memory storing computer-executable instructions, which, when executed, cause the processor to perform the method according to any one of the preceding claims.

[0018] To address the aforementioned technical problems, a fifth aspect of the present invention provides a computer program product, comprising a computer program that, when executed by a processor, implements the method described in any of the preceding claims.

[0019] In summary, this invention summarizes user instructions into user needs and breaks them down into multiple sequentially executed subtasks. Based on this sequence, at least one of the following intelligent agents—interaction content generation, interaction content adaptive optimization, data analysis, and work summary—is invoked to execute the corresponding subtask. This automates repetitive tasks such as writing and optimizing interaction content, data analysis, and generating work reports, saving approximately 17% of labor costs. Specifically, the interaction content generation agent invokes a scenario knowledge base and an interaction content generation model when generating new interaction content. The interaction content generation model undergoes optimization training based on a multi-level preference optimization process, including at least supervised learning, direct preference optimization, and group relative strategy optimization stages. During the supervised learning stage, the interaction content generation model is trained using telemarketing vertical training data, and historical interaction content is output. A reward module then judges the compliance of the historical interaction content to obtain positive and negative samples. On the one hand, the scenario knowledge base enhances the relevance of the generated interaction content to the business scenario. On the other hand, the interactive content generation model is integrated with the reward module. The reward module assesses the compliance of historical interactive content generated during the supervised learning phase, thereby selecting positive and negative samples to retrain and optimize the interactive content generation model, ensuring the compliance of the generated interactive content. Furthermore, the interactive content adaptive optimization agent can adaptively optimize newly generated interactive content according to user needs, meeting multi-dimensional interactive content generation requirements; the data analysis agent can detect abnormal data and / or data requiring optimization in the dialogue parameter changes at each node, identifying nodes to be optimized and promptly detecting anomalies; the work summary agent can periodically generate work summaries for convenient system access. Compared to existing technologies, this invention has at least the following beneficial effects: 1. Improve the efficiency of interactive content generation and save labor costs: Automating repetitive tasks such as interactive content writing, optimization, data analysis, and report generation can save approximately 17% of labor costs.

[0020] 2. Improve the scenario relevance and compliance of interactive content: The interactive content generation model, which integrates scenario knowledge base, scenario-supervised learning and reward model training, ensures that the generated new interactive content is both relevant to the business scenario and meets compliance requirements, effectively improving the quality of interactive content generation.

[0021] 3. Achieve closed-loop intelligent evolution: The system is integrated with online data, which can automatically discover problems and initiate optimization processes, enabling continuous and automated iteration of the telemarketing robot's capabilities and reducing absolute dependence on external manual annotation and expert experience.

[0022] 4. Possesses memory and collaborative capabilities: Through historical record mechanisms and multi-agent collaborative scheduling, the system can understand the context, provide coherent services, and flexibly handle diverse tasks such as interactive content generation, optimization, and analysis. Attached Figure Description

[0023] To make the technical problems solved by this invention, the technical means adopted, and the technical effects achieved clearer, specific embodiments of this invention will be described in detail below with reference to the accompanying drawings. However, it should be noted that the drawings described below are merely drawings of exemplary embodiments of this invention. Those skilled in the art can obtain drawings of other embodiments based on these drawings without any creative effort.

[0024] Figure 1 This is a schematic diagram of the structural framework of a multi-agent collaborative outbound call interactive content adaptive generation system provided in an embodiment of the present invention; Figure 2 This is a flowchart illustrating a method for adaptively generating outbound call interaction content through multi-agent collaboration, as provided in an embodiment of the present invention. Figure 3 This is a schematic diagram of the structure of the interactive content generation intelligent agent provided in an embodiment of the present invention; Figure 4 This is a structural block diagram of an exemplary embodiment of an electronic device according to the present invention. Detailed Implementation

[0025] Subject to the inventive concept, the structures, performance, effects or other features described in a particular embodiment may be combined in any suitable manner with one or more other embodiments.

[0026] In the description of specific embodiments, detailed descriptions of structures, performance, effects, or other features are provided to enable those skilled in the art to fully understand the embodiments. However, it is not excluded that those skilled in the art can implement the present invention under certain circumstances with technical solutions that do not contain the above-described structures, performance, effects, or other features. The figures in the accompanying drawings are merely illustrative examples and do not imply that the solutions of the present invention must include all the contents, operations, and steps shown in the figures, nor do they imply that they must be performed in the order shown in the figures.

[0027] refer to Figure 1 , Figure 1 This is a schematic diagram of the structure of a multi-agent collaborative outbound call interaction content adaptive generation system provided in an embodiment of the present invention, such as... Figure 1 As shown, the system includes: Receiver module 11 is used to receive user instructions and analyze and summarize them into user requirements; The splitting and invocation module 12 is used to split the user request into multiple sub-tasks to be executed sequentially, and to coordinately invoke a preset intelligent agent to execute the corresponding sub-tasks according to the order; wherein, the preset intelligent agent includes at least one of the following intelligent agents: An interactive content generation agent is used to retrieve current scene knowledge and current node information from a scene knowledge base to construct prompt words. The interactive content generation model generates new interactive content based on the original interactive content of the current node and the prompt words. Specifically, the interactive content generation model completes optimization training through a multi-level preference optimization process, with the training phase including at least supervised learning, direct preference optimization, and group relative strategy optimization stages. In the supervised learning stage, the interactive content generation model is trained using telemarketing vertical training data and outputs historical interactive content. The compliance of the historical interactive content is then judged by a reward module to obtain positive and negative samples. Then, direct preference optimization training and group relative strategy optimization training are carried out sequentially. An interactive content adaptive optimization agent is used to adaptively optimize newly generated interactive content based on user needs, and generate optimized interactive content. A data analysis intelligent agent is used to acquire dialogue parameter change data for each node, detect whether there is abnormal data and / or data that needs optimization in the dialogue parameter change data, and identify the nodes to be optimized. The work summary agent is used to extract the original work content of user interactions within a predetermined time period and generate a work summary in a standard format that includes the predetermined content.

[0028] In one possible implementation, the scenario knowledge base stores the dialogue goals, user preferences, resources available to the user groups, interaction content requirements, and node information for each user group in different scenarios.

[0029] The reward module breaks down compliance into multiple sub-compliance conditions. It outputs a sub-score by judging the degree of compliance of historical interaction content with each sub-compliance condition. It combines the sub-scores to determine the compliance score of historical interaction content and selects positive and negative samples from historical interaction content based on the compliance score.

[0030] In a preferred embodiment, the receiving module 11 includes: The real-time receiving unit is used to receive user instructions and analyze and summarize them in real time to obtain real-time user requirements. The extraction unit is used to extract historical user demands associated with the real-time user demands from the historical demand database; The summary unit is used to summarize real-time user needs and related historical user needs to obtain user requirements; Wherein: the user needs include at least one of the following: requirements, feedback, and suggestions.

[0031] Furthermore, when the data analysis agent detects a node to be optimized, it calls the interaction content generation agent to generate new interaction content based on the original interaction content of the node to be optimized.

[0032] based on Figure 1 The multi-agent collaborative outbound call interaction content adaptive generation system shown in this invention embodiment also provides a multi-agent collaborative outbound call interaction content adaptive generation method. This method operates offline, collecting and processing real-time outbound call interaction content data between online telemarketing robots and customers, providing feedback and optimization for the online telemarketing robot, thereby improving the online telemarketing robot's dialogue and communication capabilities, and realizing a self-optimizing cycle. Figure 2 The multi-agent collaborative outbound call interaction content adaptive generation method includes: S1. Receive user instructions and analyze and summarize them into user requirements; In this embodiment, the user can be a developer, expert, or operator in the field of generating interactive content for telemarketing outbound call robots; this invention does not impose specific limitations. The user instruction is a description of the user's needs, which can be text entered by the user through an input device (such as a keyboard, stylus, etc.) or audio entered by the user through an input device (such as a microphone). The user needs include at least one of the following: the user's requirements for the generated interactive content, the user's feedback on the generated interactive content, and the user's suggestions for the generated interactive content. As an example, the user's requirement for the generated interactive content could be to generate a brief introduction of XX product (within 20 characters); the user's feedback on the generated interactive content could be "needs improvement," "acceptable," or "very good"; and the user's suggestion for the generated interactive content could be to add a salutation.

[0033] In a preferred real-time approach, this step involves a work summary agent dynamically associating and summarizing real-time user commands, multi-round historical user requests, and / or multi-round historical user interaction information, then inputting these together into a large speech model to generate user requests. By achieving contextual coherence, the accuracy of user requests is improved. This step may include: S11. Receive user instructions and analyze and summarize them in real time to obtain real-time user needs; For example, the work summary agent can analyze and summarize real-time user needs based on preset prompts, including request keywords, feedback keywords, and / or suggestion keywords in user instructions. Preferably, the work summary agent can also store real-time user needs in a historical needs database based on their generation time (e.g., date and time). Simultaneously, it can collect user interaction information in real-time and store it in the historical needs database for easy retrieval later. Historical user needs are generated from one or more rounds of interaction between the user and the work summary agent; user interaction information includes text, voice, images, etc., sent between the user and the work summary agent in one or more rounds.

[0034] S12. Extract multi-round historical user needs and / or multi-round historical user interaction information associated with the real-time user needs from the historical demand database; In this example, the multi-round historical user requests and / or historical user interaction information associated with real-time user requests can be configured as needed. For example, the associated historical user requests and / or historical user interaction information can be: multi-round historical user requests stored within a predetermined time period (e.g., historical user requests stored within 2 days) and / or historical user interaction information; or, the multi-round user requests associated with real-time user requests can be: multi-round historical user requests in which at least one of the request keywords, feedback keywords, or suggestion keywords is the same; the multi-round user interaction information associated with real-time user requests can be multi-round historical user interaction information in which at least one of the text, voice, or image sent by the user to the work summary agent is the same or similar.

[0035] S13. Summarize the real-time user needs and the multi-round historical user needs and / or multi-round historical user interaction information associated with the real-time user needs to obtain user needs.

[0036] For example, this step can use a large language model to merge real-time user needs and related historical user needs, including requests, feedback, and / or suggestions, and remove synonyms, near-synonyms, or irrelevant words to obtain a more comprehensive set of user needs. Irrelevant words include: words with semantically opposite or unrelated words, and words with ambiguous meanings.

[0037] In this step, the work summary agent can input real-time user needs, historical user needs, and / or historical user interaction information into the large speech model to generate user needs. For example, the large speech model can merge the request-related words, feedback keywords, and / or suggestion keywords from the real-time user needs and associated multi-round historical user needs, and remove synonyms, near-synonyms, or irrelevant words to obtain a more comprehensive user need. Irrelevant words include words with semantically opposite or unrelated words, and semantically ambiguous words. And / or, the large speech model can merge the text, voice, and images from the real-time user needs and associated multi-round historical user interaction information, and remove identical or similar elements to obtain a more comprehensive user interaction information. Keywords from the user interaction information are then incorporated into the user needs, achieving multi-round dynamic association between real-time user needs and historical user needs and / or historical interaction information. This allows for a deep understanding of real-time user needs based on context, ensuring the contextual coherence and continuous optimization of the generated user needs, and improving the accuracy of subsequent subtask decomposition.

[0038] S2. The user requirements are divided into multiple sub-tasks to be executed in sequence, and a preset intelligent agent is selected to execute the corresponding sub-task according to the order. For example, taking a user need that simultaneously includes user requests, feedback, and suggestions as an example, the user request can be broken down into multiple sub-tasks based on splitting keywords. Each sub-task has a pre-configured execution priority, thus forming the execution order of the sub-tasks according to the execution priority. In one example, one splitting keyword corresponds to one sub-task, and the splitting keywords can include: generate, optimize, analyze, and summarize; wherein: generate corresponds to the interactive content generation task, optimize corresponds to the interactive content optimization task, analyze corresponds to the data analysis task, and summarize corresponds to the summary task. For example, the execution priorities of each sub-task from high to low are: data analysis task, interactive content generation task, interactive content optimization task, and summary task; or, the execution priorities of each sub-task from high to low are: interactive content generation task, interactive content optimization task, data analysis task, and summary task. User feedback can directly correspond to the interactive content optimization task, or user feedback can be broken down based on splitting keywords; this invention does not make specific limitations. User suggestions can directly correspond to the interactive content generation task, or user suggestions can be broken down based on splitting keywords; this invention does not make specific limitations.

[0039] In this embodiment, the preset intelligent agent includes at least one of an interactive content generation intelligent agent, an interactive content adaptive optimization intelligent agent, a data analysis intelligent agent, and a work summary intelligent agent, wherein: The interactive content generation agent can further optimize existing interactive content based on user feedback. Required inputs include scene name, node name, original node interactive content, generated interactive content, and user feedback. It generates and outputs optimized new interactive content by calling the interactive content generation model. Users can choose to use the generated new interactive content to update the telemarketing robot's interactive content, or to further optimize or perform data analysis on the generated new interactive content. Figure 3 As shown, the interactive content generation agent retrieves current scene knowledge and current node information from the scene knowledge base to construct prompt words. The interactive content generation model generates new interactive content based on the original interactive content of the current node and the prompt words. Specifically, the interactive content generation model completes optimization training based on a multi-level preference optimization process. The training phase includes at least supervised learning, direct preference optimization, and group relative strategy optimization stages. In the supervised learning stage, the interactive content generation model is trained using telemarketing vertical training data and outputs historical interactive content. The compliance of the historical interactive content is judged by the reward module to obtain positive and negative samples. Then, direct preference optimization training and group relative strategy optimization training are carried out sequentially. In one possible implementation, to improve the text generation quality of the interactive content generation model in specific scenarios, i.e., scenario fit, a scenario knowledge base is introduced. The scenario knowledge base stores the dialogue goals, user preferences, resources available to the user group, interactive content requirements, and node information for each user group in different scenarios. Specifically, user group resources can include material resources (e.g., coupons, red envelopes, gifts) or non-material resources (e.g., data viewing permissions, login platform permissions). Node information can include the node's current stage (e.g., opening remarks, product introduction, Q&A, transaction facilitation, closing remarks), and the original interactive content of the standard reply bound to the node. When generating interactive content, the interactive content generation agent retrieves current scenario knowledge and current node information from the scenario knowledge base to construct prompts for the interactive content generation model, thereby incorporating the user group description and node description of the current node.

[0040] To further stylize the generated interactive content and better align it with scenario requirements while enhancing the interactive effect, the prompts for the interactive content generation model can include standard telemarketing standard operating procedures (SOPs) and pre-selected high-quality interactive content. This guides the interactive content generation model to mimic the style of the high-quality interactive content to generate new interactive content, achieving stylistic rewriting. Specifically, the standard telemarketing SOP process allows the interactive content generation model to understand business logic, eliminate inherent comprehension biases, and master basic business interactive content. High-quality interactive content can provide a reference for model innovation.

[0041] To improve the compliance of interactive content generated by the interactive content generation model, this invention develops a reward module, which is used to judge the compliance of the interactive content output by the interactive content generation model and filter out positive and negative samples, so as to optimize the training of the historical interactive content output by the interactive content generation model, thereby improving the compliance of the generated interactive content.

[0042] In a preferred embodiment, the reward module is integrated with supervised learning using a scene knowledge base and telemarketing vertical training data. Prompt words are constructed based on current scene knowledge and current node information extracted from the knowledge base, as well as standard telemarketing SOP processes and pre-selected high-quality interactive content. The interactive content generation model is then optimized and trained using a multi-level preference optimization process. The training phase includes at least supervised learning, direct preference optimization, and group relative strategy optimization stages. During the supervised learning phase, telemarketing vertical training data is first fed into the interactive content generation model for training. The interactive content generation model can employ a large language model (e.g., the general-purpose large language model Deepseek-Qwen3-8B-0528). The telemarketing vertical training data input to the interactive content generation model can include: scene name, node name, original node interactive content, generated interactive content, user feedback, etc. In this embodiment, the telemarketing vertical training data can be historical dialogue data between the telemarketing robot and customers in various scenarios, collected online. High-quality interaction content from each scenario, pre-selected manually, can be used to annotate the telemarketing vertical training data (i.e., the historical dialogue data between the telemarketing robot and customers in the corresponding scenarios collected online). Supervised training is then performed using the annotated telemarketing vertical training data. Each training data point contains an "input" and a pre-given "correct output" corresponding to that input. The goal of the training is to enable the interaction content generation model to master basic business interaction content and to make correct predictions about the context of new scenarios.

[0043] After training, the interaction content generation model outputs a large amount of historical interaction content. The reward module judges the compliance of this historical interaction content to obtain positive and negative samples. Then, direct preference optimization (DPO) training and group relative strategy optimization training are performed sequentially. Specifically, the reward module can use a large language model. This module breaks down compliance into multiple sub-compliance conditions. It outputs a sub-score by judging the degree to which historical interaction content conforms to each sub-compliance condition. The compliance score of the historical interaction content is determined by combining all sub-scores, and positive and negative samples are selected from the historical interaction content based on the compliance score. In one example, sub-compliance conditions may include: format correctness, the correctness of the user's access to resource information, whether a disclaimer is included, and whether the identity of the interlocutor is included. After obtaining positive and negative samples, Direct Preference Optimization (DPO) is preferably performed based on the positive and negative samples. In this model training method, the training data exists in the form of "comparison pairs." Each set of data contains a pair of text responses to the same prompt word, clearly indicating which one is considered the better answer. By training the interactive content generation model to learn from manually labeled preference data, it understands and internalizes the quality evaluation criteria, making it easier for the model to generate high-quality text that meets expectations after training. Finally, Group Relative Policy Optimization (GRPO) is introduced. This utilizes dynamically generated data from the interactive content generation model to perform relative optimization within groups, compensating for the shortcomings of static DPO data, improving the diversity and stability of interactive content, and further enhancing the optimization effect. GRPO can dynamically generate a set of candidate interactive content for the same cue word from the interactive content generation model. It calculates the advantage function by comparing the relative rewards (rather than absolute rewards) between interactive content within the group. This dynamic comparison of data within groups reduces computational overhead and allows the interactive content generation model to more stably learn the quality differences between different outputs, thereby improving alignment and generation quality.

[0044] An interactive content adaptive optimization agent is used to adaptively optimize newly generated interactive content based on user needs, generating optimized interactive content. User needs can include at least one of the following: user feedback, user requests, and user suggestions. As long as an optimization subtask can be extracted from the aforementioned user needs, the interactive content optimization agent can be invoked to execute it. For example, the interactive content optimization agent takes the scene name, node name, original node interactive content, generated new interactive content, and user needs as input to a trained large language model, and generates and outputs the optimized interactive content.

[0045] A data analysis agent is used to acquire dialogue parameter change data for each node, detect whether there is abnormal data and / or data that needs optimization in the dialogue parameter change data, and identify nodes that need optimization. For example, the dialogue parameters can be node retention rate (i.e., node retention / node trigger count), node branch affirmation count, node trigger count, etc.; the node branch affirmation count refers to the number of customers who did not reject the request at that node. Upon receiving a data analysis task, the data analysis agent can retrieve the dialogue data of the online target telemarketing robot in the target scenario based on the robot ID and scenario ID, calculate the dialogue parameter change data for each node in the dialogue data, and input it into a detection model. The detection model detects whether there is abnormal data and / or data that needs optimization in the dialogue parameter change data. Specifically, the detection model can use a large language model, focusing on the trends of each node's retention rate and branch affirmation count, and using nodes with high trigger counts as prompts for the detection model. Guided by these prompts, the detection model identifies nodes that are abnormal and / or need optimization.

[0046] Furthermore, when the data analysis agent detects a node to be optimized, it can automatically invoke the interaction content generation agent to generate new interaction content based on the original interaction content of the node to be optimized. This new content can then be used by the user (operations personnel) to choose whether to use it, or further optimized as a subsequent user requirement. This improves the automation of the entire process and reduces manual intervention.

[0047] A work summary agent is used to extract the original work content interacting with users within a predetermined time period and generate a work summary in a standard format and with predetermined content based on the original work content. Upon receiving a summary task, the work summary agent calls a large language model to extract the original work content interacting with users within the predetermined time period from the context, and generates a work summary according to the standard format and predetermined content. For example, the predetermined content may include workload, optimization issues related to the summarized interaction content, debriefing, and reflection. For instance, if the predetermined time period is one day, the original work content interacting with users on that day is extracted to generate a daily report. Weekly reports can be generated based on the already generated daily reports; for example, the content of daily reports within the last 5 days can be summarized to obtain a weekly report. Furthermore, the work summary agent can call a storage tool to store daily and weekly reports for easy viewing or retrieval later.

[0048] The following example illustrates the working cycle of the multi-agent collaborative outbound call interactive content adaptive generation method provided by this invention: It continuously collects dialogue data, question records (such as negative examples), and user input information from the online telemarketing robot. The user input information is summarized and broken down into multiple sub-tasks executed sequentially. Corresponding agents are then invoked to execute each sub-task based on the collected dialogue data and question records. Specifically, when the data analysis agent identifies nodes requiring optimization, it automatically invokes the interactive content generation agent to generate new interactive content. After review by operations personnel, this content is updated to the online telemarketing robot, forming a self-optimizing closed loop that integrates data collection, breakdown, execution, data analysis, intelligent optimization, manual review, and online updates.

[0049] based on Figure 2 The multi-agent collaborative outbound call interaction content adaptive generation method shown in this invention also provides an intelligent agent for optimizing the outbound call interaction content of telemarketing. The intelligent agent executes the method described in any of the preceding embodiments. This intelligent agent operates offline, collecting and processing real-time dialogue data between the online telemarketing robot and customers to provide feedback and optimization to the online telemarketing robot, thereby improving the robot's conversational communication capabilities and achieving a self-optimizing cycle.

[0050] In summary, the multi-agent collaborative outbound call interactive content adaptive generation system and method provided in this invention, after the user inputs an instruction, summarizes the user's needs by associating the user's historical needs with the user's real-time needs, and schedules different intelligent agents by splitting the user's needs to meet the user's different needs for interactive content optimization. Specifically, when generating new interactive content, the interactive content generation agent calls a scenario knowledge base and an interactive content generation model. The interactive content generation model completes optimization training based on a multi-level preference optimization process that includes at least supervised learning, direct preference optimization, and group relative strategy optimization stages. In the supervised learning stage, the interactive content generation model is trained using telemarketing vertical training data and outputs historical interactive content. A reward module then judges the compliance of the historical interactive content to filter positive and negative samples. On the one hand, the scenario knowledge base improves the fit between the generated interactive content and the business scenario. On the other hand, the interactive content generation model is integrated with the reward module. The reward module judges the compliance of the historical interactive content generated in the supervised learning stage, thereby filtering out positive and negative samples to re-optimize and train the interactive content generation model, ensuring the compliance of the generated interactive content. The interactive content adaptive optimization agent uses a large language model to generate further optimized interactive content based on user feedback; the work summary agent can generate and save a weekly report of daily interactions with users based on context. The data analysis agent can obtain detailed dialogue data based on the telemarketing robot ID and scene ID, and then use a large language model to analyze node data to identify nodes that need optimization, ultimately improving the efficiency of interactive content optimization.

[0051] Those skilled in the art will understand that the modules in the above system embodiments can be distributed in the device as described, or they can be modified accordingly and distributed in one or more devices different from the above embodiments. The modules in the above embodiments can be combined into one module, or they can be further divided into multiple sub-modules.

[0052] The following describes embodiments of the electronic device of the present invention, which can be considered as implementations of the physical form of the methods and apparatus embodiments of the present invention described above. Details described in the embodiments of the electronic device of the present invention should be considered as supplements to the methods or apparatus embodiments described above; details not disclosed in the embodiments of the electronic device of the present invention can be implemented with reference to the methods or systems embodiments described above.

[0053] Figure 4 This is a structural block diagram of an exemplary embodiment of an electronic device according to the present invention. Figure 4 The electronic device shown is merely an example and should not impose any limitation on the functionality and scope of use of the embodiments of the present invention.

[0054] like Figure 4As shown, the electronic device 400 of this exemplary embodiment is manifested in the form of a general data processing device. The components of the electronic device 400 may include, but are not limited to: at least one processing unit 410, at least one storage unit 420, a bus 430 connecting different electronic device components (including storage unit 420 and processing unit 410), a display unit 440, etc.

[0055] The storage unit 420 stores a computer-readable program, which may be source code or read-only program code. The program can be executed by the processing unit 410, causing the processing unit 410 to perform the steps of various embodiments of the present invention. For example, the processing unit 410 can perform actions such as... Figure 2 The steps are shown.

[0056] Bus 430 can represent one or more of several types of bus structures, including a memory cell bus or memory cell controller, a peripheral bus, a graphics acceleration port, a processing unit, or a local bus using any of the various bus structures.

[0057] Electronic device 400 can also communicate with one or more external devices 100 (e.g., keyboard, monitor, network device, Bluetooth device, etc.), enabling users to interact with electronic device 400 via these external devices 100, and / or enabling electronic device 400 to communicate with one or more other data processing devices (e.g., router, modem, etc.). This communication can be performed via input / output (I / O) interface 450, and also via network adapter 460 to one or more networks (e.g., local area network (LAN), wide area network (WAN), and / or public network). Network adapter 460 can communicate with other modules of electronic device 400 via bus 430.

[0058] This invention also provides a computer program product, including a computer program that, when executed by a processor, implements the method described in any of the above-mentioned embodiments. Specifically: it receives user instructions and analyzes and summarizes them into user needs; it breaks down the user needs into multiple sub-tasks to be executed sequentially, and coordinates the invocation of preset intelligent agents to execute the corresponding sub-tasks according to the order; wherein the preset intelligent agent includes at least one of the following intelligent agents: an interactive content generation intelligent agent, used to retrieve current scene knowledge and current node information from a scene knowledge base to construct prompt words; and an interactive content generation model generates new interactive content based on the original interactive content of the current node and the prompt words; wherein the interactive content generation model completes optimization training based on a multi-level preference optimization process, and the training phase includes at least supervised learning, direct preference optimization, and group relative strategy optimization stages; in the supervised learning stage, vertical training data from telemarketing is used to optimize the training model. After the interactive content generation model is trained and outputs historical interactive content, the compliance of the historical interactive content is judged by the reward module to obtain positive and negative samples. Then, direct preference optimization training and group relative strategy optimization training are carried out in sequence. The interactive content adaptive optimization agent is used to adaptively optimize the newly generated interactive content according to user needs and generate optimized interactive content. The data analysis agent is used to obtain the dialogue parameter change data of each node, detect whether there is abnormal data and / or data that needs to be optimized, and identify the nodes to be optimized. The work summary agent is used to extract the original work content of user interaction within a predetermined time period and generate a work summary in a standard format that includes predetermined content.

[0059] In summary, the present invention can be implemented by methods, systems, electronic devices, or computer program products that execute computer programs. In practice, some or all of the functions of the present invention can be implemented using general-purpose data processing devices such as microprocessors or digital signal processors (DSPs).

[0060] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the present invention is not inherently related to any specific computer, virtual device, or electronic device, and various general-purpose devices can also implement the present invention. The above descriptions are merely specific embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A multi-agent collaborative outbound interaction content adaptive generation system, characterized in that, include: The receiving module is used to receive user instructions and analyze and summarize them into user requirements. The decomposition and invocation module is used to decompose the user request into multiple subtasks to be executed sequentially, and to coordinately invoke preset intelligent agents to execute the corresponding subtasks according to the order; wherein, the preset intelligent agents include at least one of the following intelligent agents: An interactive content generation agent is used to retrieve current scene knowledge and current node information from a scene knowledge base to construct prompt words. The interactive content generation model generates new interactive content based on the original interactive content of the current node and the prompt words. Specifically, the interactive content generation model completes optimization training through a multi-level preference optimization process, with the training phase including at least supervised learning, direct preference optimization, and group relative strategy optimization stages. In the supervised learning stage, the interactive content generation model is trained using telemarketing vertical training data and outputs historical interactive content. The compliance of the historical interactive content is then judged by a reward module to obtain positive and negative samples. Then, direct preference optimization training and group relative strategy optimization training are carried out sequentially. An interactive content adaptive optimization agent is used to adaptively optimize newly generated interactive content based on user needs, and generate optimized interactive content. A data analysis intelligent agent is used to acquire dialogue parameter change data for each node, detect whether there is abnormal data and / or data that needs optimization in the dialogue parameter change data, and identify the nodes to be optimized. The work summary agent is used to extract the original work content of user interactions within a predetermined time period and generate a work summary in a standard format that includes the predetermined content.

2. The system according to claim 1, characterized in that, The scenario knowledge base stores the dialogue goals, user preferences, resources available to the user group, interaction content requirements, and node information for each user group in different scenarios.

3. The system according to claim 1, characterized in that, The reward module breaks down compliance into multiple sub-compliance conditions. It outputs a sub-score by judging the degree of compliance of historical interaction content with each sub-compliance condition. It combines the sub-scores to determine the compliance score of historical interaction content and selects positive and negative samples from historical interaction content based on the compliance score.

4. The system according to claim 1, characterized in that, The receiving module includes: The real-time receiving unit is used to receive user instructions and analyze and summarize them in real time to obtain real-time user requirements. The extraction unit is used to extract historical user demands associated with the real-time user demands from the historical demand database; The summary unit is used to summarize real-time user needs and related historical user needs to obtain user requirements; Wherein: the user needs include at least one of the following: requirements, feedback, and suggestions.

5. The system according to any one of claims 1-4, characterized in that, When the data analysis agent detects a node to be optimized, it calls the interaction content generation agent to generate new interaction content based on the original interaction content of the node to be optimized.

6. A method for adaptively generating outbound call interaction content through multi-agent collaboration, characterized in that, include: Receive user instructions and analyze and summarize them into user requirements; The user requirements are broken down into multiple sub-tasks that are executed sequentially, and a preset intelligent agent is collaboratively invoked to execute the corresponding sub-tasks according to the order; wherein the preset intelligent agent includes at least one of the following intelligent agents: An interactive content generation agent is used to retrieve current scene knowledge and current node information from a scene knowledge base to construct prompt words. The interactive content generation model generates new interactive content based on the original interactive content of the current node and the prompt words. Specifically, the interactive content generation model completes optimization training through a multi-level preference optimization process, with the training phase including at least supervised learning, direct preference optimization, and group relative strategy optimization stages. In the supervised learning stage, the interactive content generation model is trained using telemarketing vertical training data and outputs historical interactive content. The compliance of the historical interactive content is then judged by a reward module to obtain positive and negative samples. Then, direct preference optimization training and group relative strategy optimization training are carried out sequentially. An interactive content adaptive optimization agent is used to adaptively optimize newly generated interactive content based on user needs, and generate optimized interactive content. A data analysis intelligent agent is used to acquire dialogue parameter change data for each node, detect whether there is abnormal data and / or data that needs optimization in the dialogue parameter change data, and identify the nodes to be optimized. The work summary agent is used to extract the original work content of user interactions within a predetermined time period and generate a work summary in a standard format that includes the predetermined content.

7. The method according to claim 6, characterized in that, The scenario knowledge base stores the dialogue goals, user preferences, resources available to the user group, interaction content requirements, and node information for each user group in different scenarios.

8. The method according to claim 6, characterized in that, The reward module breaks down compliance into multiple sub-compliance conditions. It outputs a sub-score by judging the degree of compliance of historical interaction content with each sub-compliance condition. It combines the sub-scores to determine the compliance score of historical interaction content and selects positive and negative samples from historical interaction content based on the compliance score.

9. The method according to claim 6, characterized in that, The process of receiving user instructions and analyzing and summarizing them into user needs includes: Receive user instructions and analyze and summarize them in real time to obtain real-time user needs; Extract historical user demands associated with the real-time user demands from the historical demand database; User requirements are derived by summarizing real-time user requirements and related historical user requirements; Wherein: the user needs include at least one of the following: requirements, feedback, and suggestions.

10. The method according to any one of claims 6-9, characterized in that, When the data analysis agent detects a node to be optimized, it calls the interaction content generation agent to generate new interaction content based on the original interaction content of the node to be optimized.

11. An intelligent agent, characterized in that, For adaptive generation of outbound call interaction content, the agent performs the method according to any one of claims 6 to 10.

12. An electronic device, characterized in that, include: processor; And a memory storing computer-executable instructions, which, when executed, cause the processor to perform the method according to any one of claims 6 to 10.

13. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the method of any one of claims 6 to 10.

Images