Business processing method and device based on multi-agent cooperation

Abstract

The application relates to a business processing method and device based on multi-agent cooperation, and relates to the technical field of financial business processing. The method comprises the following steps: acquiring business handling content input by a user in a conversational interface; calling a first agent to perform intent recognition on the business handling content, determining a target business matter to be handled based on an intent recognition result; the first agent is a pre-trained intent recognition type agent; according to a mapping relationship between the business matter and a second agent, a target second agent corresponding to the target business matter is called, the target second agent processes the corresponding business and outputs a business processing result; the second agent is a pre-trained exclusive processing agent for a subdivided business, and different second agents correspond to processing of different business matters; and the business processing result is displayed on the conversational interface of the user. The method can realize agent modularization decoupling, improve system maintainability and expandability.

Description

Technical Field

[0001] This application relates to the field of financial business processing technology, and in particular to a business processing method, apparatus, device, readable storage medium and program product based on multi-agent collaboration. Background Technology

[0002] Currently, mobile banking and other financial service systems are accelerating the development of conversational transaction processing, which has become a key direction for banks' digital transformation. At its core, it relies on multiple platforms such as intelligent customer service, intelligent assistants, and intelligent transaction processing agents to provide users with one-stop financial services through natural language interaction, covering scenarios such as fund transfers, bill inquiries, wealth management product consultations, loan applications, and utility bill payments. In terms of application, mobile banking has embedded conversational interfaces, supporting multimodal interaction such as text and voice, attempting to replace traditional menu-based operations to improve the convenience and efficiency of transaction processing.

[0003] However, current financial business systems generally employ a monolithic architecture for their business agents, lacking an effective modular decomposition mechanism. This results in highly coupled dependencies between functional units within the agent. This restricts the maintainability, scalability, and iteration efficiency of financial business systems, making it difficult for them to respond quickly to changes in business requirements. Summary of the Invention

[0004] Therefore, it is necessary to provide a business processing method, apparatus, device, readable storage medium, and program product based on multi-agent collaboration that can achieve modular decoupling of intelligent agents and improve system maintainability and scalability, in response to the above-mentioned technical problems.

[0005] Firstly, this application provides a business processing method based on multi-agent collaboration, including:

[0006] Obtain the business processing information entered by the user in the conversational interactive interface;

[0007] The first intelligent agent is invoked to perform intent recognition on the business processing content, and the target business item to be processed is determined based on the intent recognition result; the first intelligent agent is a pre-trained intent recognition type intelligent agent;

[0008] Based on the mapping relationship between business items and second intelligent agents, the target second intelligent agent corresponding to the target business item is invoked, and the target second intelligent agent processes the corresponding business and outputs the business processing result; the second intelligent agent is a pre-trained dedicated processing intelligent agent for segmented business, and different second intelligent agents are corresponding to handle different business items;

[0009] The results of the business processing are displayed on the user's conversational interface.

[0010] In one embodiment, the step of invoking the first intelligent agent to perform intent recognition on the business processing content includes:

[0011] In the first intelligent agent

[0012] Semantic dimension feature extraction and keyword dimension feature extraction are performed on the business processing content respectively. Based on the extracted semantic dimension features and keyword dimension features, a first query representation and a second query representation are generated respectively.

[0013] Based on the first query representation, a semantic search is performed in the preset intent library to obtain the corresponding semantic search recall results; and based on the second query representation, a keyword search is performed in the preset intent library to obtain the corresponding keyword search recall results.

[0014] By combining the semantic retrieval results and the keyword retrieval results, an intent recognition result is obtained.

[0015] In one embodiment, the fusion of the semantic retrieval result and the keyword retrieval result to obtain the intent recognition result includes:

[0016] The semantic recall scores of each candidate intent in the semantic retrieval recall results are normalized to obtain the semantic normalized score;

[0017] The keyword recall scores corresponding to candidate intents in the keyword retrieval results are normalized to obtain the keyword normalization score.

[0018] Based on the semantic normalization score, the keyword normalization score, and the preset weight coefficient, the fusion score of each candidate intent is calculated through a nonlinear enhancement model;

[0019] The intent recognition result is determined based on the fusion score.

[0020] In one embodiment, the nonlinear enhancement model uses the semantic normalization score as the base score and the keyword normalization score as the enhancement factor. The higher the semantic normalization score, the weaker the enhancement effect of the keyword normalization score. Conversely, the lower the semantic normalization score and the higher the keyword normalization score, the stronger the enhancement effect.

[0021] In one embodiment, determining the intent recognition result based on the fusion score includes:

[0022] The fusion score is used as the ranking criterion for the semantic-keyword fusion recall channel to obtain the candidate intent ranking result corresponding to the channel;

[0023] The candidate intent ranking results of the semantic-keyword fusion recall channel and the candidate intent ranking results of other recall channels are input into a preset improved multi-channel ranking fusion model for unified re-ranking to obtain the comprehensive ranking result of each candidate intent.

[0024] The candidate intent ranked first in the comprehensive ranking results is selected as the intent recognition result.

[0025] In one embodiment, the step of invoking the target second intelligent agent corresponding to the target business transaction based on the mapping relationship between the business transaction and the second intelligent agent includes:

[0026] Query the mapping relationship between the business item and the second intelligent agent to determine the target second intelligent agent that matches the target business item;

[0027] Generate a target business item identifier corresponding to the target business item;

[0028] The target business item identifier, the business processing content, and the context information of the business processing content are transmitted to the target second intelligent agent to invoke the target second intelligent agent.

[0029] Secondly, this application also provides a service processing device based on multi-agent collaboration, comprising:

[0030] The acquisition module is used to acquire the business processing information entered by the user in the conversational interactive interface.

[0031] The determination module is used to invoke a first intelligent agent to perform intent recognition on the business processing content, and determine the target business item to be processed based on the intent recognition result; the first intelligent agent is a pre-trained intent recognition type intelligent agent;

[0032] The business processing module is used to call the target second intelligent agent corresponding to the target business matter according to the mapping relationship between the business matter and the second intelligent agent. The target second intelligent agent processes the corresponding business and outputs the business processing result. The second intelligent agent is a pre-trained dedicated processing intelligent agent for segmented business. Different second intelligent agents are used to process different business matters.

[0033] The feedback module is used to display the business processing results on the user's conversational interface.

[0034] Thirdly, this application also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to perform the following steps:

[0035] Obtain the business processing information entered by the user in the conversational interactive interface;

[0036] The first intelligent agent is invoked to perform intent recognition on the business processing content, and the target business item to be processed is determined based on the intent recognition result; the first intelligent agent is a pre-trained intent recognition type intelligent agent;

[0037] Based on the mapping relationship between business items and second intelligent agents, the target second intelligent agent corresponding to the target business item is invoked, and the target second intelligent agent processes the corresponding business and outputs the business processing result; the second intelligent agent is a pre-trained dedicated processing intelligent agent for segmented business, and different second intelligent agents are corresponding to handle different business items;

[0038] The results of the business processing are displayed on the user's conversational interface.

[0039] Fourthly, this application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, performs the following steps:

[0040] Obtain the business processing information entered by the user in the conversational interactive interface;

[0041] The first intelligent agent is invoked to perform intent recognition on the business processing content, and the target business item to be processed is determined based on the intent recognition result; the first intelligent agent is a pre-trained intent recognition type intelligent agent;

[0042] Based on the mapping relationship between business items and second intelligent agents, the target second intelligent agent corresponding to the target business item is invoked, and the target second intelligent agent processes the corresponding business and outputs the business processing result; the second intelligent agent is a pre-trained dedicated processing intelligent agent for segmented business, and different second intelligent agents are corresponding to handle different business items;

[0043] The results of the business processing are displayed on the user's conversational interface.

[0044] Fifthly, this application also provides a computer program product, including a computer program that, when executed by a processor, performs the following steps:

[0045] Obtain the business processing information entered by the user in the conversational interactive interface;

[0046] The first intelligent agent is invoked to perform intent recognition on the business processing content, and the target business item to be processed is determined based on the intent recognition result; the first intelligent agent is a pre-trained intent recognition type intelligent agent;

[0047] Based on the mapping relationship between business items and second intelligent agents, the target second intelligent agent corresponding to the target business item is invoked, and the target second intelligent agent processes the corresponding business and outputs the business processing result; the second intelligent agent is a pre-trained dedicated processing intelligent agent for segmented business, and different second intelligent agents are corresponding to handle different business items;

[0048] The results of the business processing are displayed on the user's conversational interface.

[0049] The aforementioned business processing methods, devices, equipment, readable storage media, and program products based on multi-agent collaboration, through a layered decoupling architecture between the first and second agents, decouple "understanding user intent" from "executing business logic," breaking the tightly coupled structure of mixed functions within a single agent. This effectively solves the technical problems of functional coupling, difficulty in expansion, and high maintenance costs inherent in monolithic architectures. Simultaneously, each second agent is independently encapsulated and evolves according to business matters; for example, the transfer and remittance agent and the loan application agent do not interfere with each other, enhancing the depth of business processing through specialized division of labor. When adding new business types, only the corresponding second agent needs to be deployed and the mapping relationship updated, without modifying the first agent or other existing second agents, supporting rapid business iteration and flexible expansion. Conversational closed-loop interaction ensures user experience, thereby comprehensively improving the maintainability, scalability, and service efficiency of financial business systems. Attached Figure Description

[0050] To more clearly illustrate the technical solutions in the embodiments of this application or related technologies, the drawings used in the description of the embodiments of this application or related technologies will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0051] Figure 1 This is an application environment diagram of a business processing method based on multi-agent collaboration in one embodiment;

[0052] Figure 2 This is a flowchart illustrating a business processing method based on multi-agent collaboration in one embodiment;

[0053] Figure 3 This is a flowchart illustrating the steps of calling a first intelligent agent to perform intent recognition on the business processing content and determining the target business item to be processed based on the intent recognition result in one embodiment.

[0054] Figure 4 This is a schematic diagram of an intent recognition process in one embodiment;

[0055] Figure 5This is a schematic diagram of the architecture of a business processing system in one embodiment;

[0056] Figure 6 This is a structural block diagram of a service processing device based on multi-agent collaboration in one embodiment;

[0057] Figure 7 This is an internal structural diagram of a computer device in one embodiment. Detailed Implementation

[0058] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0059] It should be noted that the terms "first," "second," etc., used in this application can be used to describe various elements, but these elements are not limited by these terms. These terms are only used to distinguish the first element from the second element. The terms "comprising" and "having," and any variations thereof, used in this application, are intended to cover non-exclusive inclusion. The term "multiple" used in this application refers to two or more. The term "and / or" used in this application refers to one of the embodiments, or any combination of multiple embodiments.

[0060] Mobile banking and other financial service systems are rapidly advancing their conversational transaction capabilities, which has become a core development direction for banks' digital transformation. In terms of technical implementation, these conversational transaction systems generally integrate core technologies such as Natural Language Processing (NLP), Retrieval Augmentation (RAG), and multi-agent interaction. Through intent recognition, semantic understanding, and dynamic adaptation of business processes, they achieve precise matching between user needs and financial services. In terms of business coverage, conversational transaction capabilities have gradually expanded from basic inquiry-based services to complex transaction-based services. For example, some mobile banking systems support relatively complex operations such as credit card applications and personal credit report inquiries completed through conversation.

[0061] From the user demand perspective, with the widespread adoption of mobile internet, users' demands for the convenience and personalization of financial services are constantly increasing. Conversational business processing, which eliminates the need for users to memorize complex function menus and allows for interaction anytime, anywhere, has become the preferred financial service model for an increasing number of users. However, in terms of actual application effectiveness, conversational business processing systems still face many technical bottlenecks, resulting in both the actual user experience and business processing efficiency falling short of expectations.

[0062] Specifically, the current conversational financial services process has the following problems:

[0063] I. Architecture Level:

[0064] 1. Deeply bound modules with complex coupling relationships.

[0065] For large and complex business processes, the intelligent agent is not modularized. Each business module (such as intent recognition, business process processing, data verification, and result feedback), functional component, and external dependencies are deeply bound together, creating a tightly coupled relationship where "a change in one part affects the whole." For example, in the intelligent agent for a money transfer business, the intent recognition module is directly associated with specific transfer amount verification logic and the core bank system interface call logic, making independent decomposition and reuse impossible.

[0066] 2. Lack of assembly flexibility and insufficient adaptability.

[0067] The rules for calls between intelligent agents, the combination of intelligent agents and functional components, and the docking rules between intelligent agents and external tools / interfaces are all fixed, lacking room for flexible adjustment. It is impossible to quickly assemble suitable intelligent agents and functional components for the differentiated needs of different business scenarios (such as personal loan applications, credit card bill inquiries, and utility payments); it can only rely on fixed business processes to promote conversational processing, making it difficult to adapt to the diverse and personalized business needs of users.

[0068] II. Development and Operations:

[0069] 1. Core components cannot be tested independently, making quality control difficult.

[0070] Due to the tight coupling between modules, it is impossible to perform independent unit testing on individual core components of the intelligent agent, such as the intent understanding component, business rule engine, and interface adaptation module. All testing must be carried out based on the complete intelligent agent system. This makes it difficult to discover defects in the components in advance, and they can only be exposed during full-process testing or actual business operation. This not only increases the cost of problem localization and repair, but also affects the stability of conversational business processing services.

[0071] 2. The industrial production efficiency of intelligent agents is low, and multi-team collaboration is hindered.

[0072] The chained referencing of intelligent agents creates a natural bottleneck for module decoupling and business expansion, hindering the continuous development and iteration of intelligent agents in multi-team, large-scale enterprise scenarios. Chained referencing causes modules developed by different teams to become interdependent, making parallel development difficult. Furthermore, in-depth business optimization and regression testing require simultaneous verification of all referenced modules, significantly increasing backtesting complexity and time consumption, further restricting the industrial production efficiency of intelligent agents.

[0073] 3. The iteration process is cumbersome and the maintenance cost is high.

[0074] When optimizing a specific business process or updating a component, a full overhaul, testing, and deployment of the entire intelligent agent system is required, resulting in long iteration cycles and low efficiency. Furthermore, the continuous accumulation of coupling relationships creates technical debt, constantly increasing subsequent maintenance costs and making it difficult to support the rapid iteration needs of the business.

[0075] III. Business Support Level:

[0076] 1. Lack of a unified iterative foundation makes it difficult to accumulate and solidify capabilities.

[0077] There is no standardized iterative foundation for conversational business processing intelligent agents. Each business upgrade or technical optimization requires rebuilding the calling logic, processing flow, and adaptation rules, resulting in a high proportion of repetitive development work. Due to the lack of a unified foundation, business processing capabilities cannot be continuously accumulated, and the conversational interaction experience cannot be systematically optimized, thus preventing the continuous reuse and improvement of the core capabilities of the intelligent agent.

[0078] 2. Business expansion is limited, making it difficult to support large-scale development.

[0079] The existing architecture lacks flexibility and has not established a standardized iterative foundation. When adding new business types, corresponding intelligent agent systems must be redeveloped, making it impossible to quickly reuse and expand based on existing modules. This not only prolongs the launch cycle of new businesses but also makes it difficult to support the large-scale development of conversational business processing in mobile banking, failing to meet the increasingly diversified business needs of users.

[0080] IV. Business Interaction Level:

[0081] 1. The core retrieval algorithm has poor adaptability, resulting in significant deviations in the results.

[0082] Current retrieval enhancement and recall algorithms (such as hybrid retrieval combined with RRF (Reciprocal Rank Fusion) and conventional implementations of multi-path recall combined with RRF re-ranking) have structural flaws. They struggle to effectively address scenarios where the scale of multi-path retrieval results is inconsistent or some channels have missing results, easily leading to significant ranking biases. For example, when a user inquires about "investment product returns," multi-path retrieval may return information on different types of investment products, or a certain channel may not find a matching result. Conventional RRF re-ranking mechanisms will cause the final recall results to deviate from the user's true intent.

[0083] 2. Insufficient recall accuracy affects the effectiveness of dialogue service processing.

[0084] The performance deficiencies of retrieval and recall algorithms directly lead to low recall accuracy, which is significantly lower than the theoretically optimal algorithm and mainstream optimization solutions in the industry. Low accuracy will trigger a series of dialogue interaction problems, such as inaccurate intent recognition, incorrect matching of business information, and inability to quickly locate the business processing entry point required by the user. This will lead to interruption of the dialogue process, increased business processing failure rate, and seriously affect the user's conversational business processing experience.

[0085] This application provides a business processing method based on multi-agent collaboration to solve the technical problems of module coupling, low iteration efficiency, poor scalability, insufficient adaptability of retrieval and recall algorithms, and low recall accuracy caused by the monolithic architecture of intelligent agents in existing financial business systems. It realizes the modular decomposition and dynamic arrangement of financial business intelligent agents, thereby improving the maintainability, scalability and accuracy of business processing of the system.

[0086] The multi-agent collaborative business processing method provided in this application can be applied to, for example... Figure 1 In the application environment shown, terminal 101 communicates with server 102 via a network. A data storage system can store the data that server 102 needs to process. The data storage system can be integrated onto server 102 or placed on a cloud or other network server. The user inputs business processing content through the conversational interface of the financial business system via terminal 101. Terminal 101 encapsulates the business processing content into a network request and sends it to server 102. Server 102 receives and parses the network request, obtains the business processing content, calls a first intelligent agent to perform intent recognition on the business processing content, determines the target business item to be processed based on the intent recognition result, and then, according to the mapping relationship between the business item and the second intelligent agent, calls the target second intelligent agent corresponding to the target business item. The target second intelligent agent processes the corresponding business and outputs the business processing result. Server 102 encapsulates the business processing result into a network response and returns it to terminal 101, which parses it and displays it on the conversational interface. Terminal 101 can be, but is not limited to, various personal computers, laptops, smartphones, tablets, and portable wearable devices. Portable wearable devices can include smartwatches, smart bracelets, and head-mounted devices. Head-mounted devices can be virtual reality (VR) devices, augmented reality (AR) devices, smart glasses, etc. Server 102 can be a standalone physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server providing cloud computing services.

[0087] In one exemplary embodiment, such as Figure 2 As shown, a business processing method based on multi-agent collaboration is provided, which can be applied to... Figure 1 Taking server 102 as an example, the explanation includes the following steps:

[0088] Step 201: Obtain the business processing information entered by the user in the conversational interactive interface.

[0089] For example, users input their business requests (such as "I want to transfer money to Zhang San") in natural language through the conversational interface of financial service systems such as mobile banking. This input is then collected as raw data for subsequent intent recognition and business processing.

[0090] Step 202: Invoke the first intelligent agent to perform intent recognition on the business processing content, and determine the target business item to be processed based on the intent recognition result.

[0091] The first intelligent agent is a pre-trained intention recognition intelligent agent.

[0092] In this step, a pre-trained first agent (intent recognition agent) is invoked to perform semantic understanding and intent analysis on the user's input of the business transaction. The first agent performs semantic feature extraction, keyword extraction, multi-way retrieval and fusion ranking, and outputs the business transaction identification result to determine the target business transaction to be processed by the user (such as "transfer and remittance").

[0093] Step 203: Based on the mapping relationship between business items and the second intelligent agent, call the target second intelligent agent corresponding to the target business item, and have the target second intelligent agent process the corresponding business and output the business processing result.

[0094] Among them, the second intelligent agent is a pre-trained dedicated processing intelligent agent for specific business segments, and different second intelligent agents are responsible for handling different business matters.

[0095] In this step, based on the pre-maintained business agent registration information table, the mapping relationship between business items and second agents is queried, and the target second agent corresponding to the target business item is matched and invoked. Each second agent is a dedicated processing agent for a specific business segment, and declares the business items or business processing capability types it can handle during the initialization or registration phase. For example, second agent A declares that it can handle the "transfer and remittance" business item; second agent B declares that it can handle the "balance inquiry" business item; and second agent C declares that it can handle the "branch inquiry" business item.

[0096] The target second intelligent agent receives the business processing content, executes the corresponding business logic processing, and outputs the business processing result.

[0097] Step 204: Display the business processing results on the user's conversational interface.

[0098] In this step, the business processing results (such as successful transfer confirmation, account balance information, business processing instructions, etc.) output by the target second intelligent agent are returned to the user terminal and displayed in natural language or structured form in the conversational interactive interface, thus completing the business processing loop.

[0099] In this embodiment, a layered decoupled architecture between the first and second intelligent agents decouples "understanding user intent" from "executing business logic," breaking the tightly coupled structure of mixed functions within a single intelligent agent. This effectively solves the technical problems of functional coupling, difficulty in expansion, and high maintenance costs inherent in monolithic architectures. Simultaneously, each second intelligent agent is independently encapsulated and evolves according to business matters. For example, the transfer and remittance agent and the loan application agent do not interfere with each other, enhancing the depth of business processing through specialized division of labor. When adding new business types, only the corresponding second intelligent agent needs to be deployed and the mapping relationship updated; there is no need to modify the first intelligent agent or other existing second intelligent agents, supporting rapid business iteration and flexible expansion. Conversational closed-loop interaction ensures user experience, thereby comprehensively improving the maintainability, scalability, and service efficiency of the financial business system.

[0100] In one exemplary embodiment, such as Figure 3 As shown, step 202 includes steps 301 to 303, wherein:

[0101] Step 301: In the first intelligent agent, semantic dimension feature extraction and keyword dimension feature extraction are performed on the business processing content respectively. Based on the extracted semantic dimension features and keyword dimension features, a first query representation and a second query representation are generated respectively.

[0102] In this step, the business processing content is first cleaned, and then semantic and keyword feature extraction are performed on the cleaned content to generate two types of query representations adapted to different recall channels. The first query representation can be a semantic feature vector used for semantic retrieval, and the second query representation can be a set of keywords used for keyword retrieval.

[0103] Step 302: Based on the first query representation, perform semantic retrieval in the preset intent library to obtain the corresponding semantic retrieval recall results, and based on the second query representation, perform keyword retrieval in the preset intent library to obtain the corresponding keyword retrieval recall results.

[0104] During the recall phase, a parallel multi-path recall mechanism is adopted, simultaneously executing semantic retrieval and keyword retrieval. Semantic retrieval, based on vector similarity calculation, recalls candidate intents semantically similar to the business processing content from a pre-defined intent database. Keyword retrieval, based on an inverted index and a relevance scoring mechanism, recalls candidate intents containing the user-input keywords. The two retrieval paths are executed independently and without blocking each other, fully leveraging the complementary advantages of semantic understanding and precise keyword matching in different query scenarios.

[0105] Step 303: Merge the semantic retrieval results and the keyword retrieval results to obtain the intent recognition results.

[0106] In this step, the semantic retrieval results and the keyword retrieval results are fused together to combine the candidate intent information from both channels and determine the final intent recognition result.

[0107] In this embodiment, semantic retrieval and keyword retrieval are performed in parallel, fully leveraging the advantages of different retrieval technologies. Semantic retrieval excels at understanding users' implicit intent, while keyword retrieval excels at precisely matching explicit constraints. Together, they provide a more comprehensive and stable set of candidate intents. This fusion mechanism ensures high recognition stability across different query scenarios, reducing interruptions to the dialogue process or business processing failures caused by misjudgments of intent, and improving the user's conversational business processing experience.

[0108] In an exemplary embodiment, fusing semantic retrieval results and keyword retrieval results to obtain intent recognition results may include the following steps:

[0109] The semantic recall scores of each candidate intent in the semantic retrieval recall results are normalized to obtain the semantic normalized score;

[0110] The keyword recall scores corresponding to candidate intents in the keyword retrieval results are normalized to obtain the keyword normalized scores.

[0111] Based on semantic normalization scores, keyword normalization scores, and preset weight coefficients, a nonlinear enhancement model is used to calculate the fusion score of each candidate intent.

[0112] The intent recognition result is determined based on the fusion score.

[0113] In one exemplary embodiment, the nonlinear enhancement model uses semantic normalization score as the base score and keyword normalization score as the enhancement factor. The higher the semantic normalization score, the weaker the enhancement effect of keyword normalization score. Conversely, the lower the semantic normalization score and the higher the keyword normalization score, the stronger the enhancement effect.

[0114] In this embodiment, the semantic recall scores of each candidate intent in the semantic retrieval results are normalized to eliminate the differences in the dimensions of scores from different channels, resulting in a semantically normalized score. Simultaneously, the keyword recall scores of the corresponding candidate intents in the keyword retrieval results are normalized to obtain a keyword normalized score. Then, based on the semantically normalized score, the keyword normalized score, and preset weight coefficients, a nonlinear enhancement model is used to calculate the fusion score of each candidate intent. This nonlinear enhancement model uses the semantically normalized score as the base score and the keyword normalized score as the enhancement factor. A higher semantically normalized score results in a weaker enhancement effect from the keyword normalized score, while a lower semantically normalized score combined with a higher keyword normalized score results in a stronger enhancement effect. The intent recognition result is determined based on the calculated fusion score; for example, the candidate intent with the highest fusion score is selected as the final recognition result.

[0115] Semantic retrieval relies on vector similarity for recall, and cosine similarity can be used as a similarity metric. Vector similarity calculations naturally fall within the (-1,1) or (0,1) interval, allowing for linear mapping and cropping to normalize the semantic recall score. ) interval.

[0116] Keyword retrieval relies on inverted indexes and relevance scoring mechanisms for recall. The original score is typically a positive real number, ranging from [value range missing]. To ensure the stability and controllability of keyword scores during the fusion stage, a nonlinear normalization method based on the arctangent function is adopted to map the keyword recall score to the (0,1) interval, forming the keyword normalized score K, which is calculated as follows:

[0117] ;

[0118] in, This represents the original keyword recall score for keyword retrieval. This is a smoothing parameter used to control the slope of the normalization curve. This normalization method can retain high discrimination in the low score range, allowing subtle differences between samples to be identified; it gradually saturates in the high score range, suppressing excessive interference from extremely high scores on the fusion results, and achieving a balance between discrimination and stability.

[0119] Based on the normalization of scores across different channels, the semantically normalized score is used as the base score, and the keyword normalized score is introduced as an enhancement factor to score each candidate intent. The fusion score F can be calculated as follows:

[0120] With both semantic recall and keyword recall scores normalized, a user-configured weight coefficient is introduced, and a non-linear enhancement model is used to determine the fusion score.

[0121] ;

[0122] Where F represents the fusion score corresponding to a candidate intent; S represents the semantic normalization score of the candidate intent; W represents the weight coefficient configured by the user; and K represents the keyword normalization score.

[0123] This fusion mechanism, when semantic similarity is high, reduces the decay term ( The attenuation effect is suppressed, with semantic score as the primary factor; when semantic similarity is insufficient but keyword matching is strong, the attenuation factor releases the enhancement space, and the enhancement factor ( By fully leveraging its capabilities, the fusion score is improved, thereby effectively compensating for explicit keyword constraints and enabling semantic understanding and precise matching to form a dynamic synergy in different scenarios.

[0124] In one exemplary embodiment, determining the intent recognition result based on the fusion score may include the following steps:

[0125] The fusion score is used as the ranking criterion for the semantic-keyword fusion recall channel to obtain the ranking result of the candidate intent corresponding to the channel;

[0126] The candidate intent ranking results of the semantic-keyword fusion recall channel and the candidate intent ranking results of other recall channels are input into a preset improved multi-channel ranking fusion model for unified re-ranking to obtain the comprehensive ranking result of each candidate intent.

[0127] The candidate intent ranked first in the comprehensive ranking results is selected as the intent recognition result.

[0128] In this step, the calculated fusion score is used as the ranking criterion for the semantic-keyword fusion recall channel. Candidate intents for this channel are ranked according to their fusion scores, resulting in a ranking of candidate intents for that channel. The ranking of candidate intents from the semantic-keyword fusion recall channel is then input into a pre-defined improved multi-channel ranking fusion model along with the ranking of candidate intents from other recall channels (such as historical behavior channels). This model uses result normalization, weight correction, and downgrading strategies to eliminate ranking biases caused by differences in multi-channel recall scale, channel reliability, and missing results, comprehensively ranking each candidate intent to obtain a comprehensive ranking result. The candidate intent ranked first from the comprehensive ranking result is selected as the final intent recognition result.

[0129] The improved multi-channel ranking fusion model is constructed based on an improved reciprocal ranking fusion method. This includes normalizing the recall results of each recall channel to ensure consistent effective result scale across different channels; dynamically adjusting the channel weights of the reciprocal ranking fusion based on the reliability coefficients of each recall channel; assigning penalty scores to candidates whose intents are not hit by a particular recall channel, instead of directly setting them to zero; and using the fusion score as an additional ranking factor in the unified re-ranking, ensuring that the fusion score is stably reflected in the final ranking.

[0130] The improved multi-channel ranking fusion model maintains the simplicity and robustness of the reciprocal ranking fusion algorithm while enhancing its adaptability to differences in the quality of results from multiple retrieval channels. Experimental evaluation shows that the improved multi-channel ranking fusion model improves both recall accuracy and result relevance compared to the standard reciprocal ranking fusion algorithm and conventional industry implementations.

[0131] like Figure 4 The diagram illustrates an intent recognition process in one embodiment. The intent recognition process includes three stages: preprocessing, parallel multi-path recall, and fusion weighting and improved multi-channel ranking fusion.

[0132] In the preprocessing stage, the business processing content input by the user is obtained, and the business processing content is processed by text cleaning, word segmentation, keyword extraction and semantic vector encoding to generate query representations adapted to different recall channels.

[0133] In the parallel multi-path recall stage, based on the query representation generated in the preprocessing stage, semantic retrieval and keyword retrieval are performed simultaneously to obtain semantic retrieval results and keyword retrieval results.

[0134] In the fusion stage of weighted fusion and improved multi-channel ranking, the semantic retrieval results and keyword retrieval results are fused. This includes normalizing the recall scores of each channel, calculating the fusion score through a nonlinear enhancement model, and performing unified re-ranking based on the improved multi-channel ranking fusion model, ultimately outputting the intent recognition result.

[0135] The improved multi-channel ranking fusion model is built based on an improved reciprocal ranking fusion method. Through result normalization, weight correction and degradation strategies, it eliminates the ranking bias caused by differences in multi-channel recall scale, channel reliability differences and result missingness, ensuring the stable representation of the fusion score in the final ranking and improving the accuracy and robustness of intent recognition.

[0136] In an exemplary embodiment, invoking the target second intelligent agent corresponding to the target business transaction based on the mapping relationship between business transactions and second intelligent agents may include the following steps:

[0137] Query the mapping relationship between business items and second intelligent agents, and determine the target second intelligent agent that matches the target business item;

[0138] Generate a target business item identifier corresponding to the target business item;

[0139] The target business item identifier, business processing content, and context information of the business processing content are transmitted to the target second intelligent agent to invoke the target second intelligent agent.

[0140] In this embodiment, the mapping relationship between pre-stored business items and second intelligent agents is queried to determine the target second intelligent agent that matches the target business item. A target business item identifier corresponding to the target business item is generated as a business type credential in subsequent invocation processes. The target business item identifier, business processing content, and context information of the business processing content are transmitted to the target second intelligent agent to complete the invocation of the target second intelligent agent.

[0141] Specifically, the business item identifier can be a business item identifier code. A business item identifier code corresponding to the target business item is generated according to preset business coding rules. The business item identifier code is a standardized identifier used to uniquely identify the business type corresponding to the current user's business processing content. Using this business item identifier code as a matching condition, a target second intelligent agent capable of processing the business item identifier code is retrieved from the pre-stored business intelligent agent registration information table. The processing entry point of the target second intelligent agent is dynamically located at runtime through a reflection mechanism, rather than binding to a specific business implementation through static code. Each second intelligent agent is regarded as an expert processing unit in a different business domain. The corresponding second intelligent agent is selected as needed based on the business item identifier code, realizing the scheduling control logic of "expert division of labor and on-demand invocation" in the hybrid expert model architecture. The scheduling process is driven by the business item identifier code and is completed dynamically at runtime.

[0142] After identifying the target second intelligent agent, a call request is constructed for the target second intelligent agent. Specifically, the business item identifier code, business processing content, and context information of the business processing content can be written into the call request and passed to the target second intelligent agent, thereby triggering the target second intelligent agent to execute the business processing logic corresponding to the business item identifier code. Inside the target second intelligent agent, the specific business processing branch or process node is selected according to the business item identifier code to complete the corresponding business processing.

[0143] like Figure 5 The diagram illustrates the architecture of a business processing system in one embodiment. The system employs a layered, decoupled architecture, including a user interaction layer, an intent recognition layer, an identifier mapping and reflection scheduling layer, a second intelligent agent cluster, and a result integration layer.

[0144] The user interaction layer is used to receive business processing content input by users, supports multiple interaction modes such as text, voice and multi-turn dialogue, and transmits the business processing content to the intent recognition layer.

[0145] The intent recognition layer is used to understand the intent of the business processing content, identify the target business items to be processed by the user, and transmit the recognition results to the identifier mapping and reflection scheduling layer.

[0146] The identifier mapping and reflection scheduling layer is used to generate corresponding business item identifiers based on the target business item, and to perform dynamic routing and hybrid expert model expert selection control based on the business item identifiers, thereby determining and calling the target second intelligent agent that matches the target business item from the second intelligent agent cluster.

[0147] The second-level intelligent agent cluster comprises multiple dedicated processing agents for specific business segments. Each agent corresponds to a different business transaction; for example, agent A handles money transfers, agent B handles balance inquiries, and agent C handles branch location inquiries. Each agent is independently encapsulated and evolves, receiving call requests from the identifier mapping and reflection scheduling layer as needed.

[0148] The results integration layer is used to receive the business processing results output by each second intelligent agent, integrate them, and return them to the user interaction layer for display.

[0149] The system achieves modular separation of user interaction, intent recognition, agent scheduling and business processing through a layered decoupling and dynamic scheduling mechanism. The functional boundaries of each layer are clear, and it supports independent expansion and iterative optimization.

[0150] This application draws on the design concept of "on-demand selection of expert capabilities" in Mixture of Experts (MOE) models, constructing a reflective scheduling mechanism centered on intention graph recognition at the agent architecture level, rather than expert hybrid training at the model level. A unified first agent identifies and parses user requests, mapping the identification results to standardized business item identifiers, which serve as the core basis for subsequent capability scheduling and routing.

[0151] Based on the business item identification code, the system dynamically selects and invokes the corresponding second intelligent agent, forming a processing link of "intent recognition - business routing - expert execution". Each second intelligent agent plays the role of an expert in a specific domain or process in the system, respectively undertaking specific business logic, rule processing and capability execution, thereby achieving capability decoupling and separation of responsibilities.

[0152] This type of hybrid expert model reflective architecture supports flexible combination and dynamic orchestration between second agents, between second agents and functional components, and between second agents and external tools and interfaces. It supports the independent development, testing, deployment and iteration of a single second agent. Under the premise of ensuring the system is manageable and controllable, it effectively improves the scalability, stability and business adaptability of the overall architecture, and gradually builds a foundation of agent basic capabilities for conversational business processing scenarios.

[0153] This application achieves the following technical effects:

[0154] 1. Achieve efficient decoupling of the conversational business thinking chain.

[0155] Breaking through the chain-coupled limitations of traditional intelligent agent thinking chains, we construct an independent operating architecture for business modules and backbone intent recognition capabilities. This allows for independent iteration of business logic and functional expansion without relying on adjustments to the first intelligent agent, ensuring the stability of the backbone system while significantly improving the iteration efficiency and flexibility of business scenarios.

[0156] 2. Overcome the challenges of full backtesting of intelligent agents and environment adaptation.

[0157] To address the industry pain point of traditional intelligent agents requiring full backtesting even for minor modifications, this system employs modular decomposition and independent testing mechanisms, requiring only targeted verification of modified components. Simultaneously, it optimizes cross-environment deployment consistency and adaptability, effectively avoiding deviations in business processing results caused by differences in deployment environments, thereby reducing system operation and maintenance costs and risks.

[0158] 3. Significantly improves the matching accuracy between semantic understanding and business tasks.

[0159] By leveraging the optimized retrieval enhancement and inverse ranking fusion reordering algorithm, the recall bias problem when multi-way retrieval results are inconsistent or some results are empty is solved, significantly improving the accuracy of user semantic recognition. At the same time, the ability to lock in business tasks is strengthened, realizing efficient matching between user needs and financial business processing procedures, and reducing intent deviation and process interruption in dialogue interaction.

[0160] 4. Empowering the large-scale, batch expansion of enterprise-level intelligent agent scenarios

[0161] The flexible assembly mechanism of the second agent based on reflection calls breaks the constraints of fixed calling relationships and supports the rapid combination and matching of the second agent with components, tools and interfaces. It meets the rapid construction needs of enterprise-level diversified financial business scenarios, solves the problems of high cost and long cycle of agent scenario expansion under traditional architecture, and realizes the large-scale batch deployment of business scenarios.

[0162] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages in other steps. It is understood that the steps in different embodiments can be freely combined as needed, and all non-contradictory solutions formed by such combinations are within the scope of protection of this application.

[0163] Based on the same inventive concept, this application also provides a multi-agent collaborative business processing apparatus for implementing the aforementioned multi-agent collaborative business processing method. The solution provided by this apparatus is similar to the implementation scheme described in the above method. Therefore, the specific limitations in one or more embodiments of the multi-agent collaborative business processing apparatus provided below can be found in the limitations of the multi-agent collaborative business processing method described above, and will not be repeated here.

[0164] In one exemplary embodiment, such as Figure 6 As shown, a business processing device based on multi-agent collaboration is provided, including: an acquisition module 601, a determination module 602, a business processing module 603, and a feedback module 604, wherein:

[0165] The acquisition module 601 is used to acquire the business processing content entered by the user in the conversational interactive interface;

[0166] The determination module 602 is used to call the first intelligent agent to perform intent recognition on the business processing content, and determine the target business item to be processed based on the intent recognition result; the first intelligent agent is a pre-trained intent recognition type intelligent agent;

[0167] The business processing module 603 is used to call the target second intelligent agent corresponding to the target business matter according to the mapping relationship between the business matter and the second intelligent agent. The target second intelligent agent processes the corresponding business and outputs the business processing result. The second intelligent agent is a pre-trained dedicated processing intelligent agent for subdivided business. Different second intelligent agents are used to process different business matters.

[0168] Feedback module 604 is used to display the business processing results on the user's conversational interactive interface.

[0169] In one exemplary embodiment, the determining module 602 is specifically used for:

[0170] In the first intelligent agent,

[0171] Semantic dimension feature extraction and keyword dimension feature extraction are performed on the business processing content respectively. Based on the extracted semantic dimension features and keyword dimension features, a first query representation and a second query representation are generated respectively.

[0172] Based on the first query representation, a semantic search is performed in the preset intent library to obtain the corresponding semantic search recall results; and based on the second query representation, a keyword search is performed in the preset intent library to obtain the corresponding keyword search recall results.

[0173] By integrating semantic retrieval results and keyword retrieval results, intent recognition results are obtained.

[0174] In one exemplary embodiment, the determining module 602 is specifically used for:

[0175] The semantic recall scores of each candidate intent in the semantic retrieval recall results are normalized to obtain the semantic normalized score;

[0176] The keyword recall scores corresponding to candidate intents in the keyword retrieval results are normalized to obtain the keyword normalized scores.

[0177] Based on semantic normalization scores, keyword normalization scores, and preset weight coefficients, a nonlinear enhancement model is used to calculate the fusion score of each candidate intent.

[0178] The intent recognition result is determined based on the fusion score.

[0179] In one exemplary embodiment, the nonlinear enhancement model uses semantic normalization score as the base score and keyword normalization score as the enhancement factor. The higher the semantic normalization score, the weaker the enhancement effect of keyword normalization score. Conversely, the lower the semantic normalization score and the higher the keyword normalization score, the stronger the enhancement effect.

[0180] In one exemplary embodiment, the determining module 602 is specifically used for:

[0181] The fusion score is used as the ranking criterion for the semantic-keyword fusion recall channel to obtain the ranking result of the candidate intent corresponding to the channel;

[0182] The candidate intent ranking results of the semantic-keyword fusion recall channel and the candidate intent ranking results of other recall channels are input into a preset improved multi-channel ranking fusion model for unified re-ranking to obtain the comprehensive ranking result of each candidate intent.

[0183] The candidate intent ranked first in the comprehensive ranking results is selected as the intent recognition result.

[0184] In one exemplary embodiment, the service processing module 603 is specifically used for:

[0185] Query the mapping relationship between business items and second intelligent agents, and determine the target second intelligent agent that matches the target business item;

[0186] Generate a target business item identifier corresponding to the target business item;

[0187] The target business item identifier, business processing content, and context information of the business processing content are transmitted to the target second intelligent agent to invoke the target second intelligent agent.

[0188] Each module in the aforementioned multi-agent collaborative business processing device can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the memory of a computer device as software, so that the processor can invoke and execute the operations corresponding to each module.

[0189] In one exemplary embodiment, a computer device is provided, which may be a server, and its internal structure diagram may be as follows: Figure 7 As shown, the computer device includes a processor, memory, input / output (I / O) interfaces, and a communication interface. The processor, memory, and I / O interfaces are connected via a system bus, and the communication interface is also connected to the system bus via the I / O interfaces. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and a database. The internal memory provides the environment for the operating system and computer programs stored in the non-volatile storage media. The database stores a preset intent library, mappings between business items and second intelligent agents, and user account information, transaction records, and other data involved in business processing. The I / O interfaces are used for exchanging information between the processor and external devices. The communication interface is used for communication with external terminals via a network connection. When the computer program is executed by the processor, it implements a business processing method based on multi-agent collaboration.

[0190] Those skilled in the art will understand that Figure 7The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0191] In one exemplary embodiment, a computer device is provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to perform the following steps:

[0192] Obtain the business processing information entered by the user in the conversational interactive interface;

[0193] The first intelligent agent is invoked to perform intent recognition on the business processing content, and the target business item to be processed is determined based on the intent recognition result; the first intelligent agent is a pre-trained intent recognition type intelligent agent;

[0194] Based on the mapping relationship between business items and second intelligent agents, the target second intelligent agent corresponding to the target business item is invoked. The target second intelligent agent processes the corresponding business and outputs the business processing result. The second intelligent agent is a pre-trained dedicated processing intelligent agent for segmented business. Different second intelligent agents are responsible for processing different business items.

[0195] The results of business processing are displayed on the user's conversational interface.

[0196] In one embodiment, the processor, when executing a computer program, also performs the following steps:

[0197] In the first intelligent agent,

[0198] Semantic dimension feature extraction and keyword dimension feature extraction are performed on the business processing content respectively. Based on the extracted semantic dimension features and keyword dimension features, a first query representation and a second query representation are generated respectively.

[0199] Based on the first query representation, a semantic search is performed in the preset intent library to obtain the corresponding semantic search recall results; and based on the second query representation, a keyword search is performed in the preset intent library to obtain the corresponding keyword search recall results.

[0200] By integrating semantic retrieval results and keyword retrieval results, intent recognition results are obtained.

[0201] In one embodiment, the processor, when executing a computer program, also performs the following steps:

[0202] The semantic recall scores of each candidate intent in the semantic retrieval recall results are normalized to obtain the semantic normalized score;

[0203] The keyword recall scores corresponding to candidate intents in the keyword retrieval results are normalized to obtain the keyword normalized scores.

[0204] Based on semantic normalization scores, keyword normalization scores, and preset weight coefficients, a nonlinear enhancement model is used to calculate the fusion score of each candidate intent.

[0205] The intent recognition result is determined based on the fusion score.

[0206] In one embodiment, the nonlinear enhancement model uses semantic normalization score as the base score and keyword normalization score as the enhancement factor. The higher the semantic normalization score, the weaker the enhancement effect of keyword normalization score. Conversely, the lower the semantic normalization score and the higher the keyword normalization score, the stronger the enhancement effect.

[0207] In one embodiment, the processor, when executing a computer program, also performs the following steps:

[0208] The fusion score is used as the ranking criterion for the semantic-keyword fusion recall channel to obtain the ranking result of the candidate intent corresponding to the channel;

[0209] The candidate intent ranking results of the semantic-keyword fusion recall channel and the candidate intent ranking results of other recall channels are input into a preset improved multi-channel ranking fusion model for unified re-ranking to obtain the comprehensive ranking result of each candidate intent.

[0210] The candidate intent ranked first in the comprehensive ranking results is selected as the intent recognition result.

[0211] In one embodiment, the processor, when executing a computer program, also performs the following steps:

[0212] Query the mapping relationship between business items and second intelligent agents, and determine the target second intelligent agent that matches the target business item;

[0213] Generate a target business item identifier corresponding to the target business item;

[0214] The target business item identifier, business processing content, and context information of the business processing content are transmitted to the target second intelligent agent to invoke the target second intelligent agent.

[0215] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, the computer program performing the following steps when executed by a processor:

[0216] Obtain the business processing information entered by the user in the conversational interactive interface;

[0217] The first intelligent agent is invoked to perform intent recognition on the business processing content, and the target business item to be processed is determined based on the intent recognition result; the first intelligent agent is a pre-trained intent recognition type intelligent agent;

[0218] Based on the mapping relationship between business items and second intelligent agents, the target second intelligent agent corresponding to the target business item is invoked. The target second intelligent agent processes the corresponding business and outputs the business processing result. The second intelligent agent is a pre-trained dedicated processing intelligent agent for segmented business. Different second intelligent agents are responsible for processing different business items.

[0219] The results of business processing are displayed on the user's conversational interface.

[0220] In one embodiment, when the computer program is executed by a processor, it also performs the following steps:

[0221] In the first intelligent agent,

[0222] Semantic dimension feature extraction and keyword dimension feature extraction are performed on the business processing content respectively. Based on the extracted semantic dimension features and keyword dimension features, a first query representation and a second query representation are generated respectively.

[0223] Based on the first query representation, a semantic search is performed in the preset intent library to obtain the corresponding semantic search recall results; and based on the second query representation, a keyword search is performed in the preset intent library to obtain the corresponding keyword search recall results.

[0224] By integrating semantic retrieval results and keyword retrieval results, intent recognition results are obtained.

[0225] In one embodiment, when the computer program is executed by a processor, it also performs the following steps:

[0226] The semantic recall scores of each candidate intent in the semantic retrieval recall results are normalized to obtain the semantic normalized score;

[0227] The keyword recall scores corresponding to candidate intents in the keyword retrieval results are normalized to obtain the keyword normalized scores.

[0228] Based on semantic normalization scores, keyword normalization scores, and preset weight coefficients, a nonlinear enhancement model is used to calculate the fusion score of each candidate intent.

[0229] The intent recognition result is determined based on the fusion score.

[0230] In one embodiment, the nonlinear enhancement model uses semantic normalization score as the base score and keyword normalization score as the enhancement factor. The higher the semantic normalization score, the weaker the enhancement effect of keyword normalization score. Conversely, the lower the semantic normalization score and the higher the keyword normalization score, the stronger the enhancement effect.

[0231] In one embodiment, when the computer program is executed by a processor, it also performs the following steps:

[0232] The fusion score is used as the ranking criterion for the semantic-keyword fusion recall channel to obtain the ranking result of the candidate intent corresponding to the channel;

[0233] The candidate intent ranking results of the semantic-keyword fusion recall channel and the candidate intent ranking results of other recall channels are input into a preset improved multi-channel ranking fusion model for unified re-ranking to obtain the comprehensive ranking result of each candidate intent.

[0234] The candidate intent ranked first in the comprehensive ranking results is selected as the intent recognition result.

[0235] In one embodiment, when the computer program is executed by a processor, it also performs the following steps:

[0236] Query the mapping relationship between business items and second intelligent agents, and determine the target second intelligent agent that matches the target business item;

[0237] Generate a target business item identifier corresponding to the target business item;

[0238] The target business item identifier, business processing content, and context information of the business processing content are transmitted to the target second intelligent agent to invoke the target second intelligent agent.

[0239] In one embodiment, a computer program product is provided, including a computer program that, when executed by a processor, performs the following steps:

[0240] Obtain the business processing information entered by the user in the conversational interactive interface;

[0241] The first intelligent agent is invoked to perform intent recognition on the business processing content, and the target business item to be processed is determined based on the intent recognition result; the first intelligent agent is a pre-trained intent recognition type intelligent agent;

[0242] Based on the mapping relationship between business items and second intelligent agents, the target second intelligent agent corresponding to the target business item is invoked. The target second intelligent agent processes the corresponding business and outputs the business processing result. The second intelligent agent is a pre-trained dedicated processing intelligent agent for segmented business. Different second intelligent agents are responsible for processing different business items.

[0243] The results of business processing are displayed on the user's conversational interface.

[0244] In one embodiment, when the computer program is executed by a processor, it also performs the following steps:

[0245] In the first intelligent agent,

[0246] Semantic dimension feature extraction and keyword dimension feature extraction are performed on the business processing content respectively. Based on the extracted semantic dimension features and keyword dimension features, a first query representation and a second query representation are generated respectively.

[0247] Based on the first query representation, a semantic search is performed in the preset intent library to obtain the corresponding semantic search recall results; and based on the second query representation, a keyword search is performed in the preset intent library to obtain the corresponding keyword search recall results.

[0248] By integrating semantic retrieval results and keyword retrieval results, intent recognition results are obtained.

[0249] In one embodiment, when the computer program is executed by a processor, it also performs the following steps:

[0250] The semantic recall scores of each candidate intent in the semantic retrieval recall results are normalized to obtain the semantic normalized score;

[0251] The keyword recall scores corresponding to candidate intents in the keyword retrieval results are normalized to obtain the keyword normalized scores.

[0252] Based on semantic normalization scores, keyword normalization scores, and preset weight coefficients, a nonlinear enhancement model is used to calculate the fusion score of each candidate intent.

[0253] The intent recognition result is determined based on the fusion score.

[0254] In one embodiment, the nonlinear enhancement model uses semantic normalization score as the base score and keyword normalization score as the enhancement factor. The higher the semantic normalization score, the weaker the enhancement effect of keyword normalization score. Conversely, the lower the semantic normalization score and the higher the keyword normalization score, the stronger the enhancement effect.

[0255] In one embodiment, when the computer program is executed by a processor, it also performs the following steps:

[0256] The fusion score is used as the ranking criterion for the semantic-keyword fusion recall channel to obtain the ranking result of the candidate intent corresponding to the channel;

[0257] The candidate intent ranking results of the semantic-keyword fusion recall channel and the candidate intent ranking results of other recall channels are input into a preset improved multi-channel ranking fusion model for unified re-ranking to obtain the comprehensive ranking result of each candidate intent.

[0258] The candidate intent ranked first in the comprehensive ranking results is selected as the intent recognition result.

[0259] In one embodiment, when the computer program is executed by a processor, it also performs the following steps:

[0260] Query the mapping relationship between business items and second intelligent agents, and determine the target second intelligent agent that matches the target business item;

[0261] Generate a target business item identifier corresponding to the target business item;

[0262] The target business item identifier, business processing content, and context information of the business processing content are transmitted to the target second intelligent agent to invoke the target second intelligent agent.

[0263] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of the relevant data must comply with relevant regulations.

[0264] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile memory and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM). The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, artificial intelligence (AI) processors, etc., and are not limited to these.

[0265] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this application.

[0266] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. A business processing method based on multi-agent collaboration, characterized in that, The method includes: Obtain the business processing information entered by the user in the conversational interactive interface; The first intelligent agent is invoked to perform intent recognition on the business processing content, and the target business item to be processed is determined based on the intent recognition result; the first intelligent agent is a pre-trained intent recognition type intelligent agent; Based on the mapping relationship between business items and second intelligent agents, the target second intelligent agent corresponding to the target business item is invoked, and the target second intelligent agent processes the corresponding business and outputs the business processing result; the second intelligent agent is a pre-trained dedicated processing intelligent agent for segmented business, and different second intelligent agents are corresponding to handle different business items; The results of the business processing are displayed on the user's conversational interface.

2. The method according to claim 1, characterized in that, The step of invoking the first intelligent agent to perform intent recognition on the business processing content includes: In the first intelligent agent Semantic dimension feature extraction and keyword dimension feature extraction are performed on the business processing content respectively. Based on the extracted semantic dimension features and keyword dimension features, a first query representation and a second query representation are generated respectively. Based on the first query representation, a semantic search is performed in the preset intent library to obtain the corresponding semantic search recall results; and based on the second query representation, a keyword search is performed in the preset intent library to obtain the corresponding keyword search recall results. By combining the semantic retrieval results and the keyword retrieval results, an intent recognition result is obtained.

3. The method according to claim 2, characterized in that, The intent recognition result obtained by fusing the semantic retrieval result and the keyword retrieval result includes: The semantic recall scores of each candidate intent in the semantic retrieval recall results are normalized to obtain the semantic normalized score; The keyword recall scores corresponding to candidate intents in the keyword retrieval results are normalized to obtain the keyword normalization score. Based on the semantic normalization score, the keyword normalization score, and the preset weight coefficient, the fusion score of each candidate intent is calculated through a nonlinear enhancement model; The intent recognition result is determined based on the fusion score.

4. The method according to claim 3, characterized in that, The nonlinear enhancement model uses the semantic normalization score as the base score and the keyword normalization score as the enhancement factor. The higher the semantic normalization score, the weaker the enhancement effect of the keyword normalization score. Conversely, the lower the semantic normalization score and the higher the keyword normalization score, the stronger the enhancement effect.

5. The method according to claim 3, characterized in that, Determining the intent recognition result based on the fusion score includes: The fusion score is used as the ranking criterion for the semantic-keyword fusion recall channel to obtain the candidate intent ranking result corresponding to the channel; The candidate intent ranking results of the semantic-keyword fusion recall channel and the candidate intent ranking results of other recall channels are input into a preset improved multi-channel ranking fusion model for unified rearrangement to obtain the comprehensive ranking result of each candidate intent. The candidate intent ranked first in the comprehensive ranking results is selected as the intent recognition result.

6. The method according to any one of claims 1-5, characterized in that, The step of invoking the target second intelligent agent corresponding to the target business matter based on the mapping relationship between the business matter and the second intelligent agent includes: Query the mapping relationship between the business item and the second intelligent agent to determine the target second intelligent agent that matches the target business item; Generate a target business item identifier corresponding to the target business item; The target business item identifier, the business processing content, and the context information of the business processing content are transmitted to the target second intelligent agent to invoke the target second intelligent agent.

7. A business processing device based on multi-agent collaboration, characterized in that, The device includes: The acquisition module is used to acquire the business processing information entered by the user in the conversational interactive interface. The determination module is used to invoke a first intelligent agent to perform intent recognition on the business processing content, and determine the target business item to be processed based on the intent recognition result; the first intelligent agent is a pre-trained intent recognition type intelligent agent; The business processing module is used to call the target second intelligent agent corresponding to the target business matter according to the mapping relationship between the business matter and the second intelligent agent. The target second intelligent agent processes the corresponding business and outputs the business processing result. The second intelligent agent is a pre-trained dedicated processing intelligent agent for segmented business. Different second intelligent agents are used to process different business matters. The feedback module is used to display the business processing results on the user's conversational interface.

8. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 6.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 6.

10. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 6.

Images