A method and system for automated fuzzy request processing based on multi-agent systems

Abstract

This invention relates to the field of artificial intelligence technology, and particularly to an automated fuzzy appeal processing method and system based on multi-agent intelligence. This automated fuzzy appeal processing method based on multi-agent intelligence obtains the probability distribution of fuzzy information in a user's appeal through a preset size model recognition strategy; determines the information entropy of multiple words in the user's appeal based on the fuzzy information probability distribution, and identifies target fuzzy information sites by identifying words whose information entropy exceeds a preset information entropy threshold and their corresponding positions in the user's appeal; executes a disambiguation strategy on the target fuzzy information sites, and performs long-span coarse-grained task planning for the disambiguated user based on multi-agent intelligence, forming a task heterogeneity graph; constructs task execution meta-paths based on the task heterogeneity graph, and processes user appeals according to the task execution meta-paths. This achieves automated fuzzy appeal processing from fuzzy information recognition, disambiguation, and post-disambiguation task planning to final execution, as well as accurate recognition of fuzzy information.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to an automated fuzzy request processing method and system based on multi-agent intelligence. Background Technology

[0002] In existing automated complaint handling scenarios, after receiving a customer's complaint, the system first categorizes it according to the business scenario. Based on this scenario and customer characteristics, a pre-configured processing flow is matched, and the complaint is then processed according to this flow. However, when a customer's complaint is vaguely worded or the pre-defined business boundaries are ambiguous, the automated system often struggles to accurately capture the customer's pain points when understanding the complaint and matching the scenario. This results in the system failing to resolve the customer's complaint based on the pre-configured processing flow, leading to decreased customer satisfaction. Summary of the Invention

[0003] To address the aforementioned technical problems, this invention provides an automated fuzzy appeal processing method and system based on multi-agent systems.

[0004] The present invention provides an automated fuzzy request processing method based on multi-agent intelligence, comprising the following steps: acquiring user requests, obtaining the fuzzy information probability distribution of the user requests through a preset size model recognition strategy; the preset size model recognition strategy is to output a first fuzzy probability distribution and a second fuzzy probability distribution corresponding to the user requests through a fuzzy information recognition model and a large language model respectively, and to perform weighted fusion of the first fuzzy probability distribution and the second fuzzy probability distribution to obtain the fuzzy information probability distribution; wherein, the fuzzy information recognition model is obtained by training an initial neural network based on a transformer architecture using a dynamic reward sampling strategy to construct a fuzzy information recognition training set; wherein, constructing the fuzzy information recognition training set using the dynamic reward sampling strategy includes the following steps: acquiring an original sample set, the original sample set including task samples labeled with different types and levels of fuzzy information; and sorting the samples in the original sample set... Task samples are sequentially input into an initial neural network based on a transformer architecture. The initial neural network is then used to classify the samples based on their pass@k values ​​across different request samples, resulting in multiple difficulty level sample sets. Here, pass@k is the proportion of samples in which the initial neural network's answer matches the fuzzy information label corresponding to the request sample at least once in k independent repetitions, and k is a preset threshold for the number of attempts. The system checks whether the number of samples in different difficulty level sample sets decreases progressively with increasing difficulty. If the number of samples in a difficulty level does not decrease progressively, the sample set with the abnormally large number of samples is randomly sampled and adjusted until the condition that the number of samples in different difficulty level sample sets decreases progressively with increasing difficulty is met. Based on the final adjusted sample sets of each difficulty level, a fuzzy information recognition training set is constructed for training the fuzzy information recognition model.

[0005] The process of training an initial neural network based on a transformer architecture using a dynamic reward sampling strategy to construct a fuzzy information recognition training set includes the following steps: performing reinforcement learning on the initial neural network based on the transformer architecture using the fuzzy information recognition training set; the objective function of the reinforcement learning stage is:

[0006] ;

[0007] in, (·) represents the expected function; This represents the initial neural network based on the transformer architecture for recognizing each problem in the training set of fuzzy information. Total number of responses; Indicates the first Answer in the next round; This indicates that the initial neural network based on the transformer architecture is designed for the task. In parameters Time The loss from the second answer, This indicates that the initial neural network based on the transformer architecture is in the old policy. The probability distribution at time; This indicates that the initial neural network based on the transformer architecture is used in the new policy. The probability distribution at time; This indicates the distribution of the new policy parameters compared to the distribution of the old policy parameters. Divergence; These are hyperparameters used to adjust... The influence of divergence constraints;

[0008] The initial neural network that has completed the reinforcement learning is subjected to supervised learning based on the fuzzy information recognition training set; the loss function of the supervised learning is:

[0009] ;

[0010] in, To output the sequence length, For position The labeled token, Given a prior sequence of tokens In the case where the parameters of the initial neural network are Time output The probability distribution;

[0011] Based on the fuzzy information probability distribution, the information entropy of multiple word segments in the user request is determined, and the word segments with information entropy greater than a preset information entropy threshold and their corresponding positions in the user request are identified as target fuzzy information sites. A disambiguation strategy is executed on the target fuzzy information sites, and long-span coarse-grained task planning is performed on the disambiguated user request based on multi-agent systems to form a task heterogeneity graph. A task execution meta-path is constructed according to the task heterogeneity graph, and the user request is processed according to the task execution meta-path.

[0012] Preferably, the large language model outputs a second fuzzy probability distribution of user requests, comprising the following steps: pre-constructing a set of fuzzy information markers, the set of fuzzy information markers including a token mapping table corresponding to fuzzy modal words and fuzzy phrases representing uncertain semantics; inputting the user statement representation sequence corresponding to the user request into the large language model, and monitoring the large language model at each time step in processing the user statement representation sequence, predicting the probability distribution of the next token based on all token sequences before that time step; for each time step, determining the sum of the probability values ​​of tokens belonging to the set of fuzzy information tokens in the probability distribution corresponding to the user statement representation sequence, and determining the fuzzy probability value corresponding to that time step; and constructing the second fuzzy probability distribution from the fuzzy probability values ​​corresponding to each time step.

[0013] Preferably, the disambiguation strategy for the target ambiguous information site specifically includes the following steps: retrieving relevant contextual information elements about the user's request from a preset database, constructing a contextual information semantic feature vector, and performing initial disambiguation on the user request feature representation corresponding to the user's request through the contextual information semantic feature vector; if the information entropy corresponding to the user request feature representation after the initial disambiguation is greater than a preset information entropy threshold, then performing secondary disambiguation on the user request feature representation through external information elements.

[0014] Preferably, secondary disambiguation is performed on the user request feature representation using external information elements. Specific steps include: if the membership function value of the target fuzzy information point in the user request feature representation is greater than a preset fuzziness threshold, then the target fuzzy information point belongs to the target fuzzy point; based on the fuzzy information position corresponding to the target fuzzy point, joint disambiguation is performed using user information elements and context information elements; if the information entropy of the user request feature representation after joint disambiguation is less than or equal to a preset information entropy threshold, then secondary disambiguation is completed, and a fine-grained user request representation is obtained.

[0015] Preferably, based on multi-agent processing of the disambiguated user requests, a long-span coarse-to-fine granular task planning is performed to form a task heterogeneous graph. Specifically, this includes the following steps: Multiple agents perform coarse-to-fine planning on the fine-grained user request representation, obtaining a planning trajectory output by each agent consisting of multiple stage tasks; the planning trajectories output by all agents are deduplicated and integrated to obtain multiple independent coarse-grained planning trajectories; fine-grained task planning is performed on each stage task in each coarse-grained planning trajectory to form a stage planning trajectory corresponding to each stage task; the stage planning trajectories output by all agents are deduplicated and integrated to obtain multiple independent fine-grained planning trajectories; and a task heterogeneous graph is constructed using the fine-grained tasks in each independent fine-grained planning trajectory as nodes and the jump trajectories of different fine-grained tasks as edges.

[0016] Preferably, constructing a task execution metapath based on the task heterogeneous graph specifically includes the following steps: using the current task state as the root node based on the Monte Carlo tree search method, taking feasible operations in the task heterogeneous graph as the basis for branching child nodes, and exploring the child nodes according to the upper confidence interval algorithm to select the next node to be expanded; when the selected node is a leaf node that has not been fully expanded, adding new task operations as child nodes according to the association relationship of the task heterogeneous graph to form a new path; randomly simulating the execution of the new path, and after the simulation execution is completed, propagating the gains of this simulation back along the search path, updating the access count and cumulative gains of each child node on the search path, and performing multiple iterations to obtain the task execution metapath.

[0017] Preferably, after constructing the task execution metapath based on the task heterogeneous graph, the method further includes: performing task ambiguity identification on the task nodes in the task execution metapath, wherein the task ambiguity identification includes step direction fuzzy identification and resource allocation conflict identification; for the identified step direction fuzzy ambiguity, the ambiguous instructions are corrected and clarified by introducing task context information and historical execution trajectory; for the identified resource allocation conflict ambiguity, the conflicting tasks are adjusted in resource allocation using a priority scheduling method; and based on the disambiguated task nodes, an unambiguous request task execution plan is reconstructed.

[0018] To solve the above-mentioned technical problems, the present invention provides another technical solution as follows: an automated fuzzy appeal processing system based on multi-agent intelligence, used to implement the automated fuzzy appeal processing method based on multi-agent intelligence as described in any of the preceding claims, comprising:

[0019] A request acquisition and fuzzy information recognition module is used to acquire user requests and obtain the fuzzy information probability distribution of the user requests through a preset size model recognition strategy. The preset size model recognition strategy is to output a first fuzzy probability distribution and a second fuzzy probability distribution corresponding to the user requests through a fuzzy information recognition model and a large language model, respectively, and to perform weighted fusion of the first fuzzy probability distribution and the second fuzzy probability distribution to obtain the fuzzy information probability distribution. Based on the fuzzy information probability distribution, the information entropy of multiple words in the user requests is determined, and the words whose information entropy is greater than a preset information entropy threshold and their corresponding positions in the user requests are determined as target fuzzy information sites. The fuzzy information recognition model is obtained by training an initial neural network based on a transformer architecture using a dynamic reward sampling strategy to construct a fuzzy information recognition training set. The construction of the fuzzy information recognition training set using the dynamic reward sampling strategy includes the following steps: obtaining an original sample set, which includes samples of different types and levels. The task samples are labeled with fuzzy information. Each task sample in the original sample set is sequentially input into an initial neural network based on a transformer architecture. The samples are then graded by difficulty based on the pass@k value of the initial neural network on different request samples, resulting in multiple difficulty level sample sets. Here, pass@k is the proportion of samples in which the initial neural network's answer matches the fuzzy information labeling corresponding to the request sample at least once in k independent repetitions, and k is a preset threshold for multiple attempts. The number of samples in different difficulty level sample sets is checked to see if it decreases progressively with increasing difficulty level. If the number of samples in a difficulty level does not decrease progressively with increasing difficulty level, the sample set with the abnormally large number of samples is randomly sampled and adjusted until the condition that the number of samples in different difficulty level sample sets decreases progressively with increasing difficulty level is met. Based on the finally adjusted sample sets of each difficulty level, a fuzzy information recognition training set is constructed for training the fuzzy information recognition model.

[0020] The process of training an initial neural network based on a transformer architecture using a dynamic reward sampling strategy to construct a fuzzy information recognition training set includes the following steps: performing reinforcement learning on the initial neural network based on the transformer architecture using the fuzzy information recognition training set; the objective function of the reinforcement learning stage is:

[0021] ;

[0022] in, (·) represents the expected function; This represents the initial neural network based on the transformer architecture for recognizing each problem in the training set of fuzzy information. Total number of responses; Indicates the first Answer in the next round; This indicates that the initial neural network based on the transformer architecture is designed for the task. In parameters Time The loss from the second answer, This indicates that the initial neural network based on the transformer architecture is in the old policy. The probability distribution at time; This indicates that the initial neural network based on the transformer architecture is used in the new policy. The probability distribution at time; This indicates the distribution of the new policy parameters compared to the distribution of the old policy parameters. Divergence; These are hyperparameters used to adjust... The influence of divergence constraints;

[0023] The initial neural network that has completed the reinforcement learning is subjected to supervised learning based on the fuzzy information recognition training set; the loss function of the supervised learning is:

[0024] ;

[0025] in, To output the sequence length, For position The labeled token, Given a prior sequence of tokens In the case where the parameters of the initial neural network are Time output The probability distribution;

[0026] The fuzzy request disambiguation and task planning module is used to execute a disambiguation strategy on the target fuzzy information points, and perform long-span coarse-grained task planning on the disambiguated user requests based on multiple agents to form a task heterogeneous graph.

[0027] The request processing module is used to construct a task execution metapath based on the task heterogeneity graph, and process the user request along with the task execution metapath.

[0028] Compared with the prior art, the automated fuzzy appeal processing method and system based on multi-agent systems provided by the present invention has the following beneficial effects:

[0029] 1. This invention provides an automated fuzzy appeal processing method based on multi-agent systems. By constructing a collaborative analysis mechanism combining a fuzzy information recognition model and a large language model with information entropy judgment, it locates key fuzzy points from the probability distribution of fuzzy information, improving the coverage and accuracy of fuzzy information recognition. This achieves precise identification of fuzzy information in user appeals, completing a fundamental shift from qualitative judgment of whether something is fuzzy to quantitative location of where it is fuzzy, facilitating more accurate disambiguation in subsequent processing. Furthermore, through a multi-agent collaborative task planning framework, the disambiguated user appeals are transformed into structured executable plans, effectively improving the quality of decomposition of complex appeals and generation of execution schemes. Thus, this invention not only achieves automated progressive fuzzy appeal processing from fuzzy information recognition, disambiguation, and post-disambiguation task planning to final execution, but also further improves the accurate identification of fuzzy information in fuzzy appeal processing, thereby solving the problem that existing automated appeal processing systems struggle to accurately identify customer fuzzy expressions and capture fuzzy information points.

[0030] 2. The present invention also provides an automated fuzzy appeal processing system based on multiple agents, which has the same beneficial effects as the above-mentioned automated fuzzy appeal processing method based on multiple agents, and will not be elaborated here. Attached Figure Description

[0031] To more clearly illustrate the technical solutions in the embodiments of the present invention, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0032] Figure 1 This is a flowchart of the steps of an automated fuzzy appeal processing method based on multiple agents provided in the first embodiment of the present invention.

[0033] Figure 2 This is a flowchart illustrating the specific steps of steps S4041 to S4044 of an automated fuzzy appeal processing method based on multiple agents provided in the first embodiment of the present invention.

[0034] Figure 3 This is a schematic diagram of the framework for automated fuzzy appeal processing of multiple agents provided in the second embodiment of the present invention. Detailed Implementation

[0035] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0036] Currently, automated complaint processing systems face the following problems when dealing with ambiguous scenarios: (1) When the customer's statement is ambiguous, the automated processing system is unable to proactively detect whether the customer's statement is ambiguous, which leads to misjudgment of the business scenario of the customer's complaint and failure to accurately process the customer's complaint; (2) When the automated processing system detects that the customer's statement is ambiguous, it is unable to accurately locate the specific position of the ambiguous information in the customer's statement, making it difficult to build and initiate targeted proactive clarification interactions, and generating a large number of redundant interactions during the process of clarifying ambiguous information; (3) When the boundaries of the business scenario are ambiguous, when the automated processing system uses the pre-configured processing flow to process the customer's complaint, if there is ambiguous information in the processing trajectory, the processing system cannot autonomously perform task disambiguation, resulting in errors in the automated processing flow or processing failure.

[0037] Based on this, the first embodiment of the present invention provides an automated fuzzy request processing method based on multi-agent intelligence. Addressing the technical problems in the aforementioned fuzzy scenarios, this embodiment can accurately identify and precisely locate fuzzy information in customer requests, thereby performing accurate disambiguation. After disambiguation, a high-quality execution plan with a clear structure and logical consistency is generated through multi-agent collaborative planning and a task-level disambiguation mechanism. This embodiment solves the core problems faced by traditional automated systems in fuzzy scenarios, such as difficulty in detection, inaccurate positioning, redundant interactions, and failure due to ambiguity, through a complete technical closed loop of proactive discovery, precise location, efficient clarification, autonomous disambiguation, and task execution. It achieves full-process automation and intelligence from understanding fuzzy requests to reliable task execution, improving the system's processing accuracy, efficiency, and user satisfaction.

[0038] The first embodiment of this invention provides an automated fuzzy request processing method based on multi-agent intelligence. This method can be executed by an automated request processing system deployed on a server, cloud, or terminal device. It is applicable to scenarios such as intelligent customer service, IT operation and maintenance automation, and intelligent work order processing, which require parsing user natural language requests and driving automated task execution. This embodiment of the invention can be applied to the customer service field, accurately determining whether there are ambiguous expressions in user requests and performing targeted disambiguation based on the location of fuzzy information, thereby improving system interaction efficiency at the system level and enhancing user experience and satisfaction at the user level. It can also be used in the field of information security to accurately identify fuzzy and ambiguous information, improving network information security supervision.

[0039] Please see Figure 1 The first embodiment of the present invention provides an automated fuzzy appeal processing method based on multi-agent systems, comprising the following steps:

[0040] S10: Obtain user requests and obtain the fuzzy information probability distribution of the user requests through a preset size model recognition strategy; the preset size model recognition strategy is to output the first fuzzy probability distribution and the second fuzzy probability distribution corresponding to the user requests through a fuzzy information recognition model and a large language model respectively, and to perform weighted fusion of the first fuzzy probability distribution and the second fuzzy probability distribution to obtain the fuzzy information probability distribution.

[0041] Specifically, the automated request processing system can be configured to receive user requests from multiple channels, including but not limited to telephone, SMS, APP, and web pages. After receiving user requests from multiple channels, the system converts the requests into natural language representations. When a user submits a request to the automated request processing system through a channel, the system first encodes the user request into a user request feature representation through an encoding layer, then extracts key information from the user request feature representation, analyzes the probability distribution of key information representations in the user request, and determines that representations with high probability distributions of key information may contain important information that needs to be extracted from the statement representation. The key information vector extraction formula is expressed as:

[0042] ;

[0043] in, This is a dynamic sampling method. For the probability distribution of information importance, The sequence represents key information; Representing user needs and characteristics; The first in the user demand feature representation The feature representation of each location. By inputting the user request feature representation and the key information representation sequence as a whole into the fuzzy information recognition model and the large language model respectively, fuzzy type recognition and fuzzy information localization can be performed respectively.

[0044] Specifically, in this embodiment, the smaller model in the preset small-model recognition strategy is a fuzzy information recognition model specifically trained or optimized to complete a specific task, namely, recognizing semantic ambiguity in text. Its architecture and training data are highly focused on the field of fuzzy information recognition, thus enabling high-precision and high-accuracy fuzzy information recognition in this embodiment. After inputting the user request feature representation and key information representation sequence into the fuzzy information recognition model, the model can be optimized for the specific task of recognizing semantic ambiguity. It can parse the semantic structure of the input text and output a numerical sequence, namely the first fuzzy probability distribution. Each value in the first fuzzy probability distribution corresponds to a semantic unit in the input user request feature representation, such as a word, a phrase, or its context, which is judged by the fuzzy information recognition model as a probability estimate of semantic ambiguity or uncertainty.

[0045] In this embodiment, the large language model is an LLM (Large Language Model), such as the GPT series, LLaMA series, Claude series, and Gemini series. The same user request feature representation and key information representation sequence are input into a general large language model. Leveraging the powerful general language understanding and generation capabilities of this large language model, the semantic representation generated internally when processing the text is analyzed. Specifically, by examining the large language model's tendency (i.e., probability) towards the "fuzzy word set" when predicting the next word in the user request feature representation and key information representation sequence, a second fuzzy probability distribution is indirectly obtained based on general knowledge and common sense to assess the fuzziness of the statement. This second fuzzy probability distribution reflects the large language model's evaluation tendency for potential fuzzy points in the text, based on massive amounts of general knowledge.

[0046] Finally, a weighted fusion operation is performed on the first and second fuzzy probability distributions. This involves assigning weights to the first and second fuzzy probability distributions and then linearly combining them to merge them into a unified and more robust fuzzy information probability distribution. This probability distribution serves as the output of step S10, providing a comprehensive quantification of the overall and local semantic fuzziness of user requests.

[0047] More specifically, the user requests are processed using a pre-defined model identification strategy. The core of this strategy lies in the collaborative use of two models with different characteristics for analysis and judgment. By integrating the precise analysis capabilities of a dedicated fuzzy information identification model with the deep semantic understanding capabilities of a large language model, the collaborative identification of fuzzy information in user requests is achieved. This effectively overcomes the one-sidedness and limitations of a single model in fuzzy judgment. It can not only accurately capture explicit fuzzy expressions but also identify implicit semantic uncertainties, providing a comprehensive and reliable data foundation for subsequent accurate positioning and efficient disambiguation. This improves the initial understanding accuracy and robustness of the automated system for fuzzy requests.

[0048] S20: Determine the information entropy of multiple words in the user request based on the fuzzy information probability distribution, and determine the target fuzzy information position of the words whose information entropy is greater than the preset information entropy threshold and their corresponding positions in the user request.

[0049] Specifically, information entropy is a unit of measurement used in information theory to measure information uncertainty; the higher the information entropy, the higher the information uncertainty. After obtaining the probability distribution of fuzzy information, the fuzzy distribution can be used to determine the information entropy of a user's request, thus measuring the amount of information required for disambiguation of the user's request. Information Entropy for:

[0050]

[0051] in, Representing user needs and characteristics; Represent the expected function; Word segmentation for user request feature representation Feature representation; The first in the user demand feature representation The probability distribution of each word segment.

[0052] More specifically, by determining the information entropy of each word in the user's request and then comparing them, the high information entropy points representing the user's request features, i.e., the positions of fuzzy information, can be captured. These high information entropy points are also the target fuzzy information points that need to be disambiguated, thereby achieving the precise location of words, phrases, and sentences representing fuzzy information in the user's request.

[0053] S30: Execute a disambiguation strategy on the target ambiguous information sites, and perform long-span coarse-grained task planning on the disambiguated user requests based on multi-agents to form a task heterogeneous graph;

[0054] Specifically, the disambiguation strategy involves supplementing ambiguous information points in a user's request with additional information, merging them with the original request, updating the understanding of the user's intent, and forming a disambiguated user request. The purpose of implementing the disambiguation strategy is to obtain a semantically clear request expression that can be accurately interpreted by machines.

[0055] More specifically, multi-agent systems consist of multiple distinct agents, each a software module possessing independent task understanding, reasoning, and planning capabilities. Typically based on a language model, these agents can be different instances of the same model or models with different specializations, equipped with task planning toolsets and knowledge. After obtaining the disambiguated user requirements, these requirements can be input into each agent. Multiple agents then collaborate to perform long-span, coarse-grained task planning based on these requirements. This involves multiple agents coordinating a long-span analysis of the requirements, first decomposing them from top to bottom into coarse-grained stages, and then refining each coarse-grained stage into a series of specific executable steps—a fine-grained stage—realizing a hierarchical planning approach that proceeds from the overall to the local. In the task heterogeneity graph, each basic operation step is abstracted as a node, and the necessary logical order and dependencies between steps are represented as directed edges connecting nodes. This graph constitutes a structured execution network encompassing multiple feasible paths, fully defining the task steps, execution logic, and alternative solutions required to meet user requirements, providing the core processing for subsequent path selection and automated execution.

[0056] S40: Construct a task execution meta-path based on the task heterogeneity graph, and process the user request along with the task execution meta-path.

[0057] Specifically, based on the task heterogeneity graph, an optimal, linear execution sequence can be selected or searched. This coherent path from the starting node to the target node can serve as the task execution meta-path. The automated request processing system, acting as the execution engine, automatically calls the tools, service interfaces, or internal functions corresponding to each node in the path according to the node order defined in the task execution meta-path, passing necessary parameters and monitoring the execution status and return results of each step. The processing determines whether to continue to the next step, retry, or enter the exception handling process based on the execution result. Finally, the automated request processing system summarizes and integrates the results of all steps in the entire execution path, generating the final processing conclusion or output, such as confirmation information of successful operation, queried data reports, generated work order numbers, etc., and feeds it back to the user, thus completing a complete automated processing loop for the user request, from receiving the user request to identifying ambiguous information, to disambiguation and planning, and finally to execution feedback.

[0058] Understandably, by constructing a collaborative analysis mechanism combining a fuzzy information recognition model and a large language model with information entropy judgment, key fuzzy points can be located from the probability distribution of fuzzy information, improving the coverage and accuracy of fuzzy information recognition. This achieves precise identification of fuzzy information in user requests, completing a fundamental shift from qualitative judgment of whether something is fuzzy to quantitative positioning of where it is fuzzy, facilitating more accurate disambiguation in the future. Furthermore, through a multi-agent collaborative task planning framework, the disambiguated user requests are transformed into structured executable plans, effectively improving the quality of decomposition of complex requests and generation of execution schemes. Thus, this invention not only achieves automated, progressive fuzzy request processing from fuzzy information recognition, disambiguation, and post-disambiguation task planning to final execution, but also further improves the accurate identification of fuzzy information in fuzzy request processing, thereby solving the problem that existing automated request processing systems struggle to accurately identify customer fuzzy expressions and capture fuzzy information points.

[0059] Furthermore, the fuzzy information recognition model is obtained by training the initial neural network based on the transformer architecture using a fuzzy information recognition training set constructed through a dynamic reward sampling strategy; wherein, constructing the fuzzy information recognition training set through the dynamic reward sampling strategy includes the following steps:

[0060] S102: Obtain the original sample set, which includes appeal samples labeled with different types and levels of fuzzy information;

[0061] Specifically, each sample in the original sample set is a request sample, that is, a piece of natural language text simulating user input. Each sample is accompanied by fuzzy information annotations that have been manually or semi-automatically labeled. These fuzzy information annotations indicate the type and degree of fuzziness of the fuzzy information present in the sample. The types of fuzzy information include, but are not limited to, entity reference fuzziness, intent fuzziness, etc. The fuzzy information annotations can be used as the standard answer during model training.

[0062] S103: Input each task sample in the original sample set into the initial neural network based on the transformer architecture in sequence, and classify the difficulty of the samples based on the pass@k value of the initial neural network on different request samples to obtain multiple difficulty level sample sets; wherein, pass@k value is the proportion of samples in which the initial neural network has at least one answer that is consistent with the fuzzy information label corresponding to the request sample in k independent repeated responses, and k is a preset threshold for multiple number of attempts.

[0063] Specifically, in this embodiment, the dynamic reward sampling strategy is based on the dynamic reward sampling method, which classifies the sample data in the original sample set according to the difficulty of the sample task, into 1-5 difficulty levels, where the difficulty level ordinal number from smallest to largest indicates the increase in difficulty of completing the sample task.

[0064] Specifically, the initial model of the fuzzy information recognition model in this embodiment can be an initial neural network based on a transformer architecture. Each claim sample in the original sample set is sequentially input into an initial neural network with this transformer architecture. For each sample, the initial neural network is required to perform multiple, i.e., k independent inference attempts to generate a recognition result for the fuzzy information in the sample.

[0065] More specifically, when k={k1, k2, k3, k4}, the specific process of classifying the difficulty of samples based on the pass@k value of the initial neural network on different appeal samples is as follows:

[0066] S1032: Take out all the task samples in the original sample set and input them one by one into the initial neural network for knowledge extraction. If the initial neural network has at least one answer in k1 times that is the same as the label of the sample, then put the sample into the level 1 difficulty sample set. Put the samples that are different from the label in k1 times into the sample pool to be rated by the model, until the initial neural network has answered all the task samples k1 times.

[0067] S1033: Take out all the remaining samples in the model rating sample pool after step S1032, and input them one by one into the initial neural network for knowledge extraction. If the initial neural network has at least one answer in k2 responses that is the same as the label of the sample, then put the sample into the level 2 difficulty sample set. Put the samples that are different from the label in k2 responses back into the model rating sample pool until the initial neural network has performed k2 responses on all task samples.

[0068] S1034: Take out all the remaining samples in the model rating sample pool after step S1033, and input them one by one into the initial neural network for knowledge extraction. If the initial neural network has at least one answer in k3 times that is the same as the label of the sample, then put the sample into the level 3 difficulty sample set. Put the samples that are different from the label in k3 times back into the model rating sample pool until the initial neural network has answered all task samples k3 times.

[0069] S1035: Take out all the remaining samples in the model rating sample pool after step S1034, and input them one by one into the initial neural network. If the initial neural network outputs the same result as the label of the sample at least once in k4 responses, then put the sample into the 4-level model difficulty sample set. Put the samples that are different from the label in all k4 responses into the 5-level model difficulty sample set, until the initial neural network has responded to all task samples k4 times.

[0070] After step S1035, the samples from the Level 4 model difficulty sample set and the Level 4 human difficulty sample set are placed into the Level 4 difficulty sample set. The intersection of the Level 5 model difficulty sample set and the Level 5 human difficulty sample set is taken, and the task samples in the intersection are placed into the Level 5 difficulty sample set, while the other samples are placed into the Level 4 difficulty sample set. This yields Level 1, 2, 3, 4, and 5 difficulty sample sets.

[0071] S104: Verify whether the number of samples in the sample sets of different difficulty levels decreases progressively as the difficulty level increases;

[0072] S105: If there is a difficulty level where the number of samples does not decrease as the difficulty level increases, then the difficulty level sample set with abnormal sample numbers shall be randomly sampled and adjusted until the condition that the number of samples in different difficulty level sample sets decreases step by step as the difficulty level increases is met.

[0073] S106: Based on the final adjusted sample sets of each difficulty level, construct a fuzzy information recognition training set for training the fuzzy information recognition model.

[0074] Specifically, after obtaining the sample sets for difficulty levels 1, 2, 3, 4, and 5, it is checked whether the number of samples in each difficulty sample set decreases as the difficulty level increases. If the number of samples in a certain difficulty level sample set is higher than the number of samples in its lower-level difficulty sample sets, then a random sampling method is used to sample that difficulty sample set, ensuring that the number of samples in each difficulty level set decreases progressively with increasing difficulty. In this embodiment, the dynamic sampling strategy based on the dynamic reward sampling method ensures that the gradient of each batch of training samples is not zero during the backward feedback of the neural network, guaranteeing the validity of each batch of training data. In this embodiment, k1=1, k2=3, k3=5, k4=8 can be selected. This k={k1, k2, k3, k4} can also take other values, which can be determined according to the actual sampling situation.

[0075] Understandably, by introducing a dynamic reward sampling strategy based on the model's own performance pass@k value to construct the training set, and through automated difficulty grading and mandatory decreasing quantity verification, it is ensured that each training batch contains a balanced gradient of samples from easy to difficult, thereby providing the model with a continuous, stable and challenging learning signal. This not only effectively avoids the gradient vanishing problem or model overfitting caused by uneven distribution of training data difficulty, but also improves the model's generalization recognition ability and training efficiency for various fuzzy patterns from simple to complex, making it easier to obtain a fuzzy information recognition model with high fuzzy information recognition ability.

[0076] Furthermore, a training set for fuzzy information recognition is constructed using a dynamic reward sampling strategy to train the initial neural network based on the transformer architecture. This process includes the following steps:

[0077] S111: Reinforcement learning is performed on the initial neural network based on the transformer architecture using the fuzzy information recognition training set;

[0078] Specifically, a fuzzy information recognition model is obtained by using a training set to train an initial neural network based on a transformer architecture in stages. Staged training refers to performing reinforcement learning training and supervised learning training on the neural network.

[0079] Specifically, the goal of reinforcement learning is to update the initial neural network parameters through policy optimization methods, finding an optimal policy that maximizes the cumulative reward when the initial neural network interacts with the environment. This reward is calculated using the advantage function, which assesses the performance of the model's current policy compared to its average performance. The advantage function formula is as follows:

[0080]

[0081] in, This indicates the initial neural network in the policy The advantage function at that time, This indicates the initial neural network in the policy The action-state value function, also known as the reward function, is used to calculate the reward. This represents the state-value function of the initial neural network with policy π. Indicates the current state. This indicates the action taken by the current strategy.

[0082] In this embodiment, the state value function is obtained by averaging the reward of the initial neural network for all solutions to the current problem under policy π. The formula for the value function is:

[0083]

[0084] in, This indicates that under the current strategy, the initial neural network addresses the problem. of Average reward for each answer Indicates the initial neural network for the problem In the The reward earned from each answer.

[0085] Substituting the value function into the advantage function calculation formula, the advantage function is:

[0086]

[0087] in, Indicating in strategy The problem of the network In the The advantage function for the second response. Indicating in strategy The problem of the network In the The reward for answering the question for the second time. For the initial neural network in policy The following questions i of The standard deviation of the reward for each answer, It is a non-zero decimal. The initial neural network adopts a rule-based reward method for answering the knowledge extraction task. If the answer given by the initial neural network is the same as the fuzzy information label in the task sample, a reward is given; if the answer given by the initial neural network is different from the fuzzy information label in the task sample, no reward is given.

[0088] The rule-based reward function formula is expressed as:

[0089]

[0090] in, This represents the result of the initial neural network's response. This indicates the results of the labeled task; and These represent the current state and the action chosen by the initial neural network, respectively. To ensure the initial neural network outputs correct content in the required format during training, a formatted reward function is defined, expressed as:

[0091]

[0092] in, The format representing the model's response results. The format of the annotation results. and The meaning is the same as that expressed by the rule-based reward function formula. If the format of the initial neural network output is the same as the format of the labeled output, a reward is given; otherwise, no reward is given. The final reward formula combining content and format rewards is:

[0093]

[0094] in, This represents the content reward coefficient. The formatted reward coefficients are shown. The initial neural network for the task... The loss for choosing a particular action is defined as the negative logarithmic function of the advantage gained by the initial neural network from performing that action, expressed by the formula:

[0095]

[0096] in, This indicates that the initial neural network is designed for the task. In parameters Choose action The loss, This indicates the initial neural network under the current policy. At that time, in the state Choose action The probability distribution; For network to address the problem No. The advantage function of the second response.

[0097] To prevent the initial neural network from receiving excessive rewards in a single action, leading to large parameter adjustments and decreased robustness during training, a KL divergence is introduced as a constraint term into the objective function of the initial neural network. This constrains the magnitude of model parameter adjustments, ensuring that the distribution of policy parameter updates in each iteration of the initial neural network is not significantly different from the old policy. The initial neural network optimizes by maximizing the objective function.

[0098] The objective function for the reinforcement learning phase is:

[0099] ;

[0100] in, (·) represents the expected function; This represents the initial neural network based on the transformer architecture for recognizing each problem in the training set of fuzzy information. Total number of responses; Indicates the first Answer in the next round; This indicates that the initial neural network based on the transformer architecture is designed for the task. In parameters Time The loss from the second answer, This indicates that the initial neural network based on the transformer architecture is in the old policy. The probability distribution at time; This indicates that the initial neural network based on the transformer architecture is used in the new policy. The probability distribution at time; This indicates the distribution of the new policy parameters compared to the distribution of the old policy parameters. Divergence; These are hyperparameters used to adjust... The influence of divergence constraints. Each time the initial neural network updates its policy parameters, it is necessary to calculate the gradient of the initial neural network loss with respect to the initial neural network parameters, and the gradient of the initial neural network objective function with respect to the parameters. The partial derivative is expressed by the formula:

[0101]

[0102] The initial neural network parameters are updated based on the partial derivatives, as expressed by the formula:

[0103]

[0104] in, Represents the original parameters That is, the parameters before the current update. ; lr RL The learning rate is used for reinforcement learning of the model.

[0105] S112: Supervised learning is performed on the initial neural network that has completed the reinforcement learning based on the fuzzy information recognition training set;

[0106] Specifically, after completing reinforcement learning training, the initial neural network enters the supervised learning phase. During supervised learning, the initial neural network takes fuzzy recognition task data as input and outputs a response consisting of a "long thought chain + fuzzy recognition content." The initial neural network responds to the fuzzy recognition task data by measuring the probability distribution P of the model outputting the next token. θ The loss function is constructed based on the distance between the tokens and the labeled data. The loss function for this supervised learning is:

[0107] ;

[0108] in, To output the sequence length, For position The labeled token, Given a prior sequence of tokens In the case where the parameters of the initial neural network are Time output The probability distribution.

[0109] During backpropagation, the partial derivatives of the loss with respect to the initial neural network parameters are first calculated:

[0110]

[0111] The initial neural network parameters are updated based on the partial derivatives, as expressed by the formula:

[0112]

[0113] in, This indicates the parameters before the current update. ; This represents the deep learning rate of the initial neural network.

[0114] After multi-stage training, a fuzzy information recognition model is obtained. The user request feature sequence is input into the fuzzy information recognition model to obtain the first fuzzy probability distribution, expressed by the formula:

[0115]

[0116] in, To represent the characteristics of user demands, (·) represents the ambiguity rate. Word segmentation for user request feature representation The characteristics are represented.

[0117] Understandably, the two-stage training method, specifically reinforcement learning based on a training set constructed using a dynamic reward sampling strategy, effectively controls the magnitude of policy updates through the KL divergence constraint term in its objective function. This avoids training oscillations caused by excessively large single rewards, ensuring the stability and convergence of the training process. The subsequent supervised learning, using the classic cross-entropy loss function, guides the initial neural network to accurately learn the distribution of the output sequence, effectively improving the accuracy of the initial neural network in memorizing and reproducing fuzzy information patterns. This training mechanism, which connects reinforcement learning and supervised learning, enables the final fuzzy information recognition model to possess both the robustness of policy exploration and the accuracy of pattern learning. When faced with complex and varied user fuzzy expressions, it exhibits higher recognition accuracy and stronger generalization ability.

[0118] Furthermore, the large language model outputs a second fuzzy probability distribution of the user's request, including the following steps:

[0119] S121: Pre-construct a set of fuzzy information tokens, which includes a token mapping table corresponding to fuzzy modal words and fuzzy phrases representing uncertain semantics;

[0120] S122: Input the user statement representation sequence corresponding to the user request into the large language model, and monitor the large language model at each time step in processing the user statement representation sequence, and predict the probability distribution of the next token based on all token sequences before that time step;

[0121] S123: For each time step, determine the sum of the probability values ​​of tokens belonging to the set of fuzzy information token flags in the probability distribution corresponding to the user statement representation sequence, and determine the fuzzy probability value corresponding to that time step;

[0122] S124: The second fuzzy probability distribution is formed by the fuzzy probability values ​​corresponding to each time step.

[0123] Specifically, to improve the semantic recognition capability of the automated appeal processing system for fuzzy information, a Large Language Model (LLM) is used to analyze the fuzzy information. First, a set of fuzzy information markers T is established, containing fuzzy modal words such as "possibly," "probably," and "perhaps," and fuzzy phrases such as "to some extent" and "within a certain range." A mapping table of fuzzy modal words and fuzzy phrases is established as fuzzy information token markers. The token probability distribution generated by the supervised LLM hidden layer is analyzed. When the probability distribution of the fuzzy information token marker exceeds the fuzzy information threshold, or when the generation probability of the fuzzy information token marker is among the top-k of the next token, it indicates that the current appeal representation contains fuzzy information. At this point, all tokens belonging to the preset fuzzy information token marker set T are selected. Then, the probability values ​​corresponding to these selected tokens are summed, and this sum is determined as the fuzzy probability value corresponding to the current time step t. By traversing all time steps, a sequence of fuzzy probability values ​​is obtained, which constitutes the second fuzzy probability distribution P. LLM .

[0124] In this embodiment, k=5. The formula for identifying fuzzy information based on LLM is expressed as:

[0125]

[0126] in, Indicates the LLM generation time step. express The generation of tokens for time steps. Indicates time step from 1 to The generated token sequence.

[0127] Understandably, by monitoring the probability distribution generated within a large language model and focusing on predefined fuzzy semantic markers, fine-grained, probabilistic detection of implicit fuzzy information in user requests is achieved. This method does not rely on specific classification fine-tuning but leverages the inherent, powerful semantic understanding and generation capabilities of the large language model. It can effectively identify various uncertain expressions, including specific fuzzy words and phrases, thus providing a deep, semantically based probabilistic basis for the entire fuzzy information recognition system, improving the system's perception accuracy of complex and implicitly fuzzy requests.

[0128] It should be noted that obtaining the first fuzzy probability distribution of the small model and the second fuzzy probability distribution of the large model Then, the probability distributions are weighted and summed to obtain the fuzzy distribution representing the user demand features. The formula is expressed as:

[0129]

[0130] in, and This represents the weight. In this embodiment, Optionally, this weight can be adjusted according to the specific recognition scenario and other factors, and the two can also be different. This embodiment does not impose any restrictions on this.

[0131] Furthermore, a disambiguation strategy is performed on the target ambiguous information sites, specifically including the following steps:

[0132] S301: Retrieve relevant contextual information elements about the user's request from the preset database, construct a contextual information semantic feature vector, and perform initial disambiguation on the user request feature representation corresponding to the user's request through the contextual information semantic feature vector.

[0133] Specifically, for ambiguous information points in the target, an adaptive appeal disambiguation strategy is used. This strategy disambiguates the user appeal feature representation through multiple information elements, aiming to reduce the information entropy of the user appeal feature representation. This adaptive appeal disambiguation strategy first retrieves relevant contextual information elements from the database, constructs a contextual information semantic feature vector, and then uses the contextual information feature representation to disambiguate the user appeal feature representation. The formula is as follows:

[0134]

[0135] in, This represents the information entropy after context disambiguation. Representing user needs and characteristics; This represents the contextual information feature representation.

[0136] S302: If the information entropy corresponding to the user request feature representation after the initial disambiguation is greater than the preset information entropy threshold, then the user request feature representation is subjected to secondary disambiguation through external information elements.

[0137] Specifically, compare the information entropy after context disambiguation. With information entropy threshold The size; if it is less than or equal to the information entropy threshold, it means that the representation of the user's request is clear and no further disambiguation is needed; if it is greater than the information entropy threshold, it means that the representation of the user's request is still unclear and secondary disambiguation is needed by using external information elements to represent the user's request.

[0138] Understandably, the provided disambiguation strategy adopts an inside-out, layered approach, prioritizing silent disambiguation using existing internal contextual knowledge and only initiating external interactions with the user when necessary. This achieves decision-making on the disambiguation path through precise quantification of information entropy, minimizing unnecessary user interruptions and redundant interactions, and significantly improving the intelligence and efficiency of the disambiguation process. Simultaneously, using user interaction as the final disambiguation method ensures the system can handle various complex and ambiguous requests, ultimately obtaining accurate and clear user intent representations, which is more conducive to subsequent automated task planning and execution.

[0139] Furthermore, secondary disambiguation is performed on the user request feature representation using external information elements. Specific steps include:

[0140] S3022: If the membership function value of the target fuzzy information site in the user demand feature representation is greater than the preset fuzzy threshold, then the target fuzzy information site belongs to the target fuzzy site;

[0141] S3023: Based on the location of the fuzzy information corresponding to the target fuzzy point, perform joint disambiguation using user information elements and context information elements;

[0142] S3024: If the information entropy of the user request feature representation after joint disambiguation is less than or equal to the preset information entropy threshold, then the secondary disambiguation has been completed, and a fine-grained representation of the user's request has been obtained.

[0143] Specifically, based on the probability distribution of fuzzy information in the disambiguated user requests, the degree to which a target fuzzy information point belongs to fuzzy information is determined by the membership function. If the membership function value of the target fuzzy information point is greater than the fuzziness threshold, it indicates that the point information of that target fuzzy information point is fuzzy. Let the fuzziness threshold be... α Membership function The fuzzy information location of the target fuzzy point is captured based on the membership function, as shown in the formula:

[0144]

[0145] in, (·) indicates the location of fuzzy information in the representation of the user's request; This represents the characteristics of user demands.

[0146] More specifically, after determining the location of the fuzzy information, an interaction strategy is constructed based on the location of the fuzzy information according to the adaptive appeal disambiguation strategy. This strategy interacts with the user, and joint disambiguation is performed using user information elements and contextual information elements. The formula is expressed as:

[0147]

[0148] in, The feature representation of user information. Compare the information entropy of the user request representation after user information disambiguation. With information entropy threshold If the information entropy of the user request representation after user information element disambiguation is less than or equal to the information entropy threshold, it indicates that the updated user request disambiguation has been completed, and the disambiguated fine-grained user request representation is obtained. Otherwise, it is necessary to relocate the fuzzy information position using the fuzzy information function based on the updated user demand feature representation, analyze the fuzzy information position, reconstruct the interaction strategy, and disambiguate through user information elements.

[0149] Understandably, by combining fuzzy membership functions with information entropy threshold judgments, precise positioning of fuzzy information points and quantitative control of the disambiguation process are achieved. User interaction is initiated only when contextual information is insufficient, with the explicit goal of reducing information entropy. This ensures that each interaction directly addresses the most uncertain information point, thereby significantly reducing unnecessary system-user interaction rounds and improving disambiguation efficiency and user experience. Furthermore, it ultimately outputs a clear and concise representation of user needs that can be used for precise planning.

[0150] Furthermore, based on multi-agent processing of the disambiguated user requests, long-span coarse-to-fine granular task planning is performed to form a task heterogeneity graph, specifically including the following steps:

[0151] S311: By having multiple agents perform coarse-grained planning on the fine-grained user request representation, a planning trajectory consisting of multiple stages of tasks is obtained from the output of each agent.

[0152] Specifically, the automated request processing system first performs long-span task planning based on the user's fine-grained request representation. This long-span task planning refers to end-to-end task execution planning, which outlines the complete trajectory from the start of resolving the user's request to its full resolution. Firstly, the automated request processing system performs coarse-grained planning based on multiple agents for the user's fine-grained request; coarse-grained stage planning refers to the process stages required for handling the user's request, excluding the specific operations required for each stage.

[0153] The formula for coarse-grained phase planning by a single agent is expressed as:

[0154]

[0155] in, Indicates the first The coarse-grained stage planning results of each agent Indicates the first The probability distribution of an agent during coarse-grained planning. This represents the task planning at step t. Indicates the steps Task planning in the previous steps, This indicates that the user's fine-grained demands for disambiguation have been met.

[0156] It should be noted that, In c express coarse-grained , indicating coarse-grainedness.

[0157] S312: Deduplicate and integrate the planned trajectories output by all agents to obtain multiple independent coarse-grained planned trajectories;

[0158] Specifically, in order to improve the search space of the automatic request processing system's stage planning and expand the effectiveness and stability of stage planning, the automatic request processing system adopts a multi-agent parallel reasoning framework. The same user fine-grained request to be processed is distributed in parallel to m task planning agents for planning. After the m agents independently output the task planning results through task planning reasoning, the system votes on the responses of the m agents, deduplicates the planning trajectories, and retains the independent trajectories in the multi-agent task planning.

[0159] The formula for parallel programming of multiple agents is expressed as:

[0160]

[0161]

[0162] in, Indicates the first The task trajectory planned by the intelligent agent. Indicates the first The agent in the th... The probability distribution of step-by-step task planning. Indicates the first The probability distribution of task planning prior to step-by-step task planning. This indicates that the multi-agent parallel reasoning framework outputs a set of planned trajectories. For the task search space, This represents the function for removing duplicates from a trajectory.

[0163] To improve parallel inference efficiency and increase resource utilization of the multi-agent parallel inference extension system, a load balancing strategy was designed. The formula for the load balancing strategy is as follows:

[0164]

[0165] in, Indicates the first A task awaiting assignment. This indicates the workload of the intelligent agent. This indicates the processing capacity of the agent. Specifically, when the load balancer allocates tasks, it assigns the task to the agent with the lowest expected load for inference, thereby improving the overall inference performance of the system and reducing response time.

[0166] S313: Perform fine-grained task planning for each stage task in each coarse-grained planning trajectory to form the stage planning trajectory corresponding to each stage task;

[0167] Specifically, after obtaining the set of planned trajectories output by the multi-agent parallel inference framework, fine-grained task planning is performed for each stage of the coarse-grained trajectory. This fine-grained task planning refers to selecting the tools, using them in a logical order, and executing them, based on the coarse-grained trajectory task as the objective. The multi-agent parallel inference framework needs to construct a tool logic chain, based on the agent tool library, with the coarse-grained stage as the execution objective, and according to the tool description, input parameters, and output parameters, to form the fine-grained task planning trajectory.

[0168] Single agent in the first The formula for fine-grained task planning in stage j-stage is expressed as:

[0169]

[0170] in, Indicates the first Fine-grained task planning results for each agent Indicates the first The probability distribution of an agent during fine-grained task planning. Indicates the steps Timely task planning, Indicates the steps Task planning in the previous steps, This indicates that the user's granular demands for achieving disambiguation are in the first... The characteristics of a stage.

[0171] It should be noted that, In express fine-grained , indicating fine granularity.

[0172] S314: Deduplicate and integrate the stage planning trajectories output by all agents to obtain multiple independent fine-grained planning trajectories;

[0173] S315: Construct a task heterogeneous graph by using the fine-grained tasks in each independent fine-grained planning trajectory as nodes and the jump trajectories of different fine-grained tasks as edges.

[0174] Specifically, similar to the coarse-grained stage planning, the automatic request processing system employs multi-agent parallel reasoning for fine-grained task planning. The task planning for the same coarse-grained stage is distributed in parallel to m task planning agents. After each of the m agents independently performs task planning reasoning and outputs its result, a vote is taken on the responses of the m agents to deduplicate the planned trajectories, retaining only the independent trajectories from the multi-agent fine-grained task planning, forming a set of task planning trajectories. Using each independent trajectory in this set of task planning trajectories, with each fine-grained task as a node and task jump trajectories as edges, a directed acyclic heterogeneous task graph is constructed.

[0175] Understandably, by combining multi-agent parallel planning with fine-grained hierarchical deconstruction, comprehensive and robust task planning for complex user requests is achieved. Coarse-grained planning ensures the completeness and logic of the processing flow stages, while fine-grained planning fills in the specific executable actions. Multi-agent parallel reasoning and trajectory deduplication mechanisms effectively expand the exploration range of the planning space, avoid the limitations of a single planning path, and enhance the system's ability to handle requests with ambiguous boundaries or novel requirements. Furthermore, the final constructed task heterogeneous graph provides a clear and rich structured search space for subsequent optimal path exploration based on Monte Carlo tree search, facilitating the efficient and reliable execution of the entire automated processing flow.

[0176] Furthermore, constructing the task execution meta-path based on the aforementioned task heterogeneity graph specifically includes the following steps:

[0177] S401: The Monte Carlo tree search method takes the current task state as the root node, uses the feasible operations in the task heterogeneous graph as the basis for branching child nodes, and explores the child nodes according to the upper confidence interval algorithm to select the next node to be expanded.

[0178] S402: When the selected node is a leaf node that has not been fully expanded, the new task operation is added as a child node according to the association relationship of the task heterogeneous graph to form a new path;

[0179] S403: Perform random simulation execution on the new path, and after the simulation execution is completed, propagate the gains of this simulation back along the search path, update the access count and cumulative gains of each child node on the search path, and perform multiple iterations to obtain the task execution meta-path.

[0180] Specifically, the task heterogeneous graph is denoted by G(V,E), where V represents the set of heterogeneous task nodes, including tool nodes, function nodes, and data processing nodes; E represents the set of directed edges between task nodes, where each directed edge represents the logical order of task execution. The task heterogeneous graph contains all possible processing tasks and task relationships within the agent's action exploration space. The automatic request processing system explores the optimal task execution plan using an optimal trajectory exploration method based on Monte Carlo tree search. The Monte Carlo tree search method combines path exploration and dynamic decision-making to explore solutions. The optimization objectives of the solution exploration are: 1. Improve the completeness rate of request processing and ensure complete processing of requests, avoiding omissions; 2. Improve the success rate of request processing; 3. Improve the stability of request processing; 4. Reduce end-to-end processing time of requests while satisfying conditions 1, 2, and 3. Let the set of optimization objectives be... ,in, To improve the completeness of the handling of requests, To improve the success rate of handling requests, To ensure the stability of the handling of requests, The end-to-end processing time for a request can be expressed by the formula:

[0181]

[0182]

[0183]

[0184]

[0185] in, This indicates the number of requests that have been fully processed. This indicates the total number of demands. This indicates the number of requests that were successfully processed. This indicates the number of requests that were attempted to be processed. This represents the standard deviation of the success rate in handling requests. This represents the average success rate of processing requests. This represents the average end-to-end processing time for a request. The optimization objective formula for the Monte Carlo tree search method is expressed as:

[0186]

[0187]

[0188] in, The target weights for each condition are defined. The Monte Carlo tree search method uses the current task state as the root node and explores child nodes based on feasible operations in the task heterogeneous graph, such as task steps, resource calls, and state transitions. Each node records key information such as the number of visits and cumulative revenue for subsequent decision-making.

[0189] More specifically, in this embodiment, starting from the root node, the next node to be expanded is selected based on the UCT (Upper Confidence Bound for Trees) algorithm. This UCT algorithm balances selecting nodes with fewer visits during the exploration process to uncover potentially high-value paths and utilizing nodes with high cumulative returns to reinforce known optimal paths. For any node v in the heterogeneous task node set V, the formula is expressed as:

[0190]

[0191] in, The cumulative revenue of the node. This represents the cumulative number of visits to the node. For the node The parent node, To explore coefficients, the higher the UCT value of a node, the higher the probability that the node will be selected. When the selected node is a leaf node that is not fully expanded, new task operations are added as new child nodes according to the association relationships in the task heterogeneous graph, and the new path is randomly simulated for execution. After the simulation is completed, the reward of this simulation is backpropagated along the path to all ancestor nodes, updating the node's visit count and cumulative reward. The task execution metapath is obtained through multiple iterations.

[0192] Understandably, the Monte Carlo tree search method, through guided stochastic exploration and optimization in a heterogeneous task graph, can efficiently find near-optimal execution paths within a vast space of possible actions. Furthermore, the UCT algorithm balances utilizing known efficient paths with exploring new possibilities, avoiding getting trapped in local optima. Simultaneously, the multi-objective reward function ensures that the final selected meta-path achieves comprehensive optimization across multiple dimensions, including success rate, completeness, stability, and efficiency. Therefore, the Monte Carlo tree search method enables automated request processing systems to autonomously and intelligently plan high-quality execution schemes for handling complex and fuzzy requests, improving the system's decision-making capabilities and final processing efficiency.

[0193] Furthermore, after constructing the task execution meta-path based on the task heterogeneity graph, the method further includes:

[0194] S404: The automated appeal processing system analyzes whether there is task ambiguity in the execution meta-path, captures and locates task ambiguity, uses a multi-agent joint reasoning method to disambiguate ambiguous tasks, and reconstructs the appeal task execution plan.

[0195] Please see Figure 2 Step S404 includes the following steps:

[0196] S4041: Perform task ambiguity identification on the task nodes in the task execution metapath, the task ambiguity identification including step direction fuzzy identification and resource allocation conflict identification;

[0197] S4042: For the identified steps that are ambiguous, the ambiguous instructions are corrected and clarified by introducing task context information and historical execution trajectory;

[0198] S4043: For identified resource allocation conflicts and ambiguities, resource allocation for conflicting tasks is adjusted using a priority scheduling method;

[0199] S4044: Based on the disambiguated task nodes, reconstruct the execution plan for unambiguous request tasks.

[0200] Specifically, task ambiguity refers to the phenomenon in the task execution process where, due to unclear information expression, multiple possibilities in logical connections, or conflicting element matching, the automatic request processing system is unable to determine the only correct execution direction, including ambiguous step directions and resource allocation conflicts.

[0201] Ambiguous step direction refers to the ambiguity in the description or instructions of task steps, making it difficult for the agent to clearly define the specific object or method of subsequent operations. Resource allocation conflict refers to multiple task steps making demands on the same limited resource, and these demands cannot be met simultaneously in terms of time, quantity, or usage, leading to contradictions in resource allocation.

[0202] The task node set in the task execution metapath is V'. For all tasks in V', task ambiguity identification is performed. Task ambiguity identification includes step-direction fuzzy identification and resource allocation conflict identification. Step-direction fuzzy identification refers to whether there is ambiguity in the task instruction description and whether there are tasks with similar task descriptions. A fuzziness function is used. F fuzzy Capture the task description; if the fuzziness function value of a task is higher than the fuzzy membership threshold, it indicates that the task instruction description is ambiguous and requires step-by-step disambiguation.

[0203] The step-direction disambiguation refers to the precise correction and clarification of ambiguous instructions by introducing task context information and historical execution trajectories. In this embodiment, the automatic appeal processing system calls task-related scenario parameters, such as the task's domain and execution environment constraints, and combines them with the instruction descriptions of historical trajectories of similar successfully executed tasks to construct a set of instruction-execution mapping pairs. By calculating the semantic similarity between the current ambiguous instruction and the instructions in the mapping pair set, the ambiguous point information is adjusted using the standard description with the highest approximation, and fuzzy information is re-identified using the fuzziness function. If the fuzziness function score is lower than the fuzziness threshold, it indicates that the step-direction fuzziness disambiguation has been completed.

[0204] Resource allocation conflict identification refers to determining whether resource conflicts exist among tasks in the task execution metapath. For example, conflicts may arise because a previous task has not released its resource allocation, or multiple tasks simultaneously request the same resource, preventing the task using that resource from executing successfully. The resource allocation conflict identification method first simulates execution through the task metapath to construct a task-resource association matrix. ,matrix Each row represents a type of resource, and each column represents a task node. Indicates task The required resource quantity. Based on the task-resource correlation matrix, the existence of resource spatiotemporal conflicts is identified through the resource spatiotemporal conflict index, whereby the resource spatiotemporal conflict index is set as follows. The formula is expressed as:

[0205]

[0206] in, For the duration of time overlap, To meet the demand for resources that exceed the available quantity, For resources Maximum load capacity and These are the weighting coefficients. +β=1, and Representing tasks and tasks The end time point, and Representing tasks and tasks The start time point. When the resource spatiotemporal conflict index is greater than the conflict threshold, a resource conflict is considered to exist. For resource conflicts, this embodiment uses a priority scheduling method to allocate resources. The priority scheduling method sorts the tasks according to their urgency and importance coefficient, and allocates resources to conflicting tasks according to their priority. High-priority tasks are guaranteed core resource use, while low-priority tasks are adjusted to idle periods to ensure that the task meta-path can be executed smoothly and obtain the task disambiguation meta-path.

[0207] Understandably, by performing fine-grained task ambiguity detection and resolution on the task execution meta-path, the uncertainty in execution caused by ambiguity in instruction expression or resource contention is effectively resolved. Specifically, by disambiguating ambiguous step directions using historical experience, vague instructions are corrected into standard expressions, improving the clarity and executability of instructions; by disambiguating resource conflicts through priority scheduling, reasonable and conflict-free allocation of resources is achieved, ensuring the practical feasibility of the plan; thus, by targeting the disambiguation of two different types of task ambiguities, the final generated execution plan is ensured to be clear, reliable, and directly deliverable to the automated system for execution, thereby significantly improving the success rate and robustness of the automated processing flow for complex requests.

[0208] Further, in step S40: processing the user request according to the task execution meta-path specifically includes:

[0209] S42: The automated complaint processing system processes user complaints by executing planned steps and using tools based on the disambiguation meta-path tasks, collects the results of user complaint processing and feeds them back to the user, thereby automating the processing of user complaints.

[0210] Specifically, the automated complaint processing system executes tasks according to the disambiguation-based meta-path. Task execution includes tool usage and data processing. During task execution, the automated complaint processing system precisely invokes various tools based on specific task requirements. These tools include data query tools, rule engine tools, interactive response tools, and operation execution tools. Data query tools retrieve basic data such as user information and order details; rule engine tools verify whether user complaints comply with business rules such as refunds and claims; interactive response tools send supplementary information prompts to users when necessary; and operation execution tools perform specific operations, such as completing refunds or modifying package subscription status. Tool invocation is dynamically triggered by the automated complaint processing system based on the progress of the steps. After completing the task meta-path and tool invocation, the automated complaint processing system integrates and verifies the processing results of the user complaint. Process integration refers to the automated complaint processing system generating a structured complaint processing result report, which includes a brief description of the processing process, the final result, relevant evidence, and subsequent operation suggestions. Result verification refers to checking whether the processing result is consistent with the user's expected goal. Finally, the automated complaint processing system feeds back the processing results and result report to the user through preset interaction channels.

[0211] Understandably, this embodiment constructs a complete technical closed loop from fuzzy information recognition and accurate disambiguation to task planning and execution. Through the collaborative recognition strategy of large and small models, the dynamic reward sampling training mechanism, and the joint cooperation between multi-agent coarse and fine-grained task planning, it solves the problems of traditional intelligent customer service systems in fuzzy expression recognition, location positioning, and disambiguation efficiency. It can accurately capture explicit and implicit fuzzy information in user requests, so as to accurately identify, efficiently disambiguate, and deeply analyze user requests containing fuzzy expressions, achieve efficient automated processing, and provide users with more intelligent and accurate response services.

[0212] In addition, in this embodiment, the core of the product user demand analysis module is formed by the cooperation between the hierarchical multi-agent architecture, dynamic reward sampling training, and Monte Carlo tree search path planning. Compared with traditional technologies, this significantly improves the accuracy and efficiency of handling fuzzy demands.

[0213] In addition, the method in this embodiment is not only applicable to user demand analysis of intelligent customer service products, but also to multiple fields such as intelligent customer service, automated work order processing, and government service response, thus improving the versatility of this solution.

[0214] Please see Figure 3 The second embodiment of the present invention provides an automated fuzzy appeal processing system based on multiple agents, used to implement the automated fuzzy appeal processing method based on multiple agents as described in the first embodiment, comprising:

[0215] The request acquisition and fuzzy information recognition module is used to acquire user requests and obtain the fuzzy information probability distribution of the user requests through a preset size model recognition strategy. The preset size model recognition strategy is to output a first fuzzy probability distribution and a second fuzzy probability distribution corresponding to the user requests through a fuzzy information recognition model and a large language model, respectively, and to perform weighted fusion of the first fuzzy probability distribution and the second fuzzy probability distribution to obtain the fuzzy information probability distribution. Based on the fuzzy information probability distribution, the information entropy of multiple words in the user requests is determined, and the words whose information entropy is greater than a preset information entropy threshold and their corresponding positions in the user requests are used to determine target fuzzy information sites.

[0216] The fuzzy request disambiguation and task planning module is used to execute a disambiguation strategy on the target fuzzy information points, and perform long-span coarse-grained task planning on the disambiguated user requests based on multiple agents to form a task heterogeneous graph.

[0217] The request processing module is used to construct a task execution meta-path based on the task heterogeneity graph, and process the user request based on the task execution meta-path.

[0218] Understandably, the automated fuzzy appeal processing system based on multi-agent technology in this embodiment has the same beneficial effects as the aforementioned automated fuzzy appeal processing method based on multi-agent technology, and will not be elaborated upon here.

[0219] Furthermore, the request acquisition and fuzzy information recognition module includes a request acceptance module, a request encoding module, and a fuzzy information recognition module. The request acceptance module receives user requests from multiple channels, including telephone, SMS, APP, and web pages. The request encoding module converts the received user requests into natural language representations. The fuzzy information recognition module obtains the fuzzy information probability distribution of the user requests using a preset size model recognition strategy, determines the information entropy of multiple word segments in the user requests based on the fuzzy information probability distribution, and identifies target fuzzy information sites by identifying word segments with information entropy greater than a preset information entropy threshold and their corresponding positions within the user requests. The fuzzy information recognition module also identifies and locates fuzzy information using a fuzzy membership function.

[0220] The fuzzy request disambiguation and task planning module includes a fuzzy request disambiguation module, an active interaction module, and a multi-agent task planning module.

[0221] The fuzzy appeal disambiguation module is used to perform disambiguation strategies on the target fuzzy information sites. Specifically, it is used to disambiguate the fuzzy information based on the identified target fuzzy information sites through context disambiguation and external information element disambiguation methods. If the information entropy of the user appeal feature representation after disambiguation is lower than the information entropy threshold, it indicates that disambiguation has been completed.

[0222] The proactive interaction module is used to proactively interact with external information elements such as users, and collect information from external information elements to disambiguate vague requests.

[0223] The multi-agent task planning module is used to perform long-span coarse-grained and fine-grained task planning based on the disambiguated user requests from multiple agents, forming a task heterogeneity graph.

[0224] The request processing module includes a task meta-path generation module and a task execution module.

[0225] The task meta-path generation module is used to construct task execution meta-paths based on the task heterogeneity graph. Specifically, it explores task meta-paths in the task heterogeneity graph using the Monte Carlo tree search method to form task meta-path trajectories.

[0226] The task execution module is used to process the user's request according to the task execution meta-path, specifically by using tools and processing data according to the disambiguated task meta-path.

[0227] The automated fuzzy request processing system based on multi-agent intelligence also includes:

[0228] The knowledge base module is used for storing data related to business scenarios, and the stored content includes knowledge data related to the handling of requirements in various business scenarios. In this embodiment, the external knowledge base module is implemented using ElasticSearch.

[0229] The user database module is used for user data storage, which includes user profiles, user consumption information, user billing records, and user complaint records. The user profiles include user names, ages, phone numbers, and user tags such as whether the user is a caring customer. The user consumption information includes the user's order time, order details, and order amount. The user billing records include user bill details. The user complaint records include the content of the work orders for the user's past complaints, the handling solutions, and the user's approval status.

[0230] The tool library module encapsulates all the tools required by the automated complaint processing system. In this embodiment, the tools in the tool library are encapsulated using the MCP protocol.

[0231] The task fuzzy disambiguation module is used to identify task ambiguity in the task metapath trajectory, and to disambiguate the task ambiguity using the corresponding disambiguation strategy to obtain the disambiguated task metapath.

[0232] The integration and verification module is used to integrate and verify the processing flow after execution according to the disambiguation task meta-path, generate a structured request processing result report for the processing flow, and check whether the processing result is consistent with the user's expected goal.

[0233] Understandably, the system achieves end-to-end automated processing of ambiguous user requests through the combined efforts of multiple modules. The knowledge base and user database provide accurate data support, the fuzzy recognition and disambiguation module effectively parses and clarifies uncertain information, the multi-agent planning and optimization module generates robust task execution paths, and finally, the closed-loop processing of the request is completed through integrated tool calls and result verification. The entire system significantly improves the accuracy of understanding complex requests, processing efficiency, and automation, while also enhancing the user experience through structured interaction and feedback mechanisms.

[0234] In the embodiments provided by this invention, it should be understood that "B corresponding to A" means that B is associated with A, and B can be determined based on A. However, it should also be understood that determining B based on A does not mean determining B solely based on A; B can also be determined based on A and / or other information.

[0235] It should be understood that the phrase "one embodiment" or "an embodiment" throughout the specification means that a specific feature, structure, or characteristic related to the embodiment is included in at least two embodiments of the invention. Therefore, "in one embodiment" or "in an embodiment" appearing throughout the specification does not necessarily refer to the same embodiment. Furthermore, these specific features, structures, or characteristics can be combined in any suitable manner in one or more embodiments. Those skilled in the art should also recognize that the embodiments described in the specification are optional embodiments, and the actions and modules involved are not necessarily essential to the invention.

[0236] In various embodiments of the present invention, it should be understood that the sequence number of each process does not necessarily imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0237] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer systems according to various embodiments of this application. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It is particularly important to note that each block in a block diagram and / or flowchart, and combinations of blocks in block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0238] The above provides a detailed description of an automated fuzzy claim processing method and system based on multi-agent systems disclosed in the embodiments of the present invention. Specific examples have been used to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of the present invention. Therefore, the content of this specification should not be construed as a limitation of the present invention. Any modifications, equivalent substitutions, and improvements made within the principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for automated fuzzy request processing based on multi-agent systems, characterized in that: Includes the following steps: The system obtains user requests and uses a preset size model recognition strategy to obtain the fuzzy information probability distribution of the user requests. The preset size model recognition strategy is to output the first fuzzy probability distribution and the second fuzzy probability distribution corresponding to the user requests through a fuzzy information recognition model and a large language model, respectively, and to perform weighted fusion of the first fuzzy probability distribution and the second fuzzy probability distribution to obtain the fuzzy information probability distribution. The fuzzy information recognition model is obtained by training an initial neural network based on a transformer architecture using a dynamic reward sampling strategy to construct a fuzzy information recognition training set. Constructing the fuzzy information recognition training set using the dynamic reward sampling strategy includes the following steps: Obtain the original sample set, which includes task samples labeled with different types and levels of fuzzy information; Each task sample in the original sample set is sequentially input into an initial neural network based on the transformer architecture, and the difficulty of the samples is classified based on the pass@k value of the initial neural network on different request samples to obtain multiple difficulty level sample sets; where pass@k value is the proportion of samples in which the initial neural network has at least one answer that matches the fuzzy information label corresponding to the request sample in k independent repeated responses, and k is a preset threshold for multiple number of attempts. Verify whether the number of samples in the sample sets of different difficulty levels decreases progressively as the difficulty level increases; If there is a difficulty level where the number of samples does not decrease as the difficulty level increases, then the sample set of the difficulty level with the abnormal number of samples will be randomly sampled and adjusted until the condition that the number of samples in the sample sets of different difficulty levels decreases step by step as the difficulty level increases is met. Based on the final adjusted sample sets of each difficulty level, a fuzzy information recognition training set is constructed for training the fuzzy information recognition model; The initial neural network based on the transformer architecture is trained by constructing a fuzzy information recognition training set through a dynamic reward sampling strategy, specifically including the following steps: The initial neural network based on the transformer architecture is subjected to reinforcement learning using the fuzzy information recognition training set; the objective function of the reinforcement learning stage is: ； in, (·) represents the expected function; This represents the initial neural network based on the transformer architecture for recognizing each problem in the training set of fuzzy information. Total number of responses; Indicates the first Answer in the next round; This indicates that the initial neural network based on the transformer architecture is designed for the task. In parameters Time The loss from the second answer, This indicates that the initial neural network based on the transformer architecture is in the old policy. The probability distribution at time; This indicates that the initial neural network based on the transformer architecture is used in the new policy. The probability distribution at time; This indicates the distribution of the new policy parameters compared to the distribution of the old policy parameters. Divergence; These are hyperparameters used to adjust... The influence of divergence constraints; The initial neural network that has completed the reinforcement learning is subjected to supervised learning based on the fuzzy information recognition training set; the loss function of the supervised learning is: ； in, To output the sequence length, For position The labeled token, Given a prior sequence of tokens In the case where the parameters of the initial neural network are Time output The probability distribution; Based on the fuzzy information probability distribution, the information entropy of multiple words in the user request is determined, and the words whose information entropy is greater than a preset information entropy threshold and their corresponding positions in the user request are determined as target fuzzy information sites. A disambiguation strategy is executed on the target ambiguous information sites, and a long-span coarse-grained task planning is carried out on the disambiguated user requests based on multi-agents to form a task heterogeneity graph. The task execution metapath is constructed based on the task heterogeneity graph, and the user request is processed based on the task execution metapath.

2. The automated fuzzy request processing method based on multi-agent systems as described in claim 1, characterized in that: The large language model outputs a second fuzzy probability distribution of user requests, including the following steps: A set of fuzzy information tokens is pre-constructed, which includes a token mapping table corresponding to fuzzy modal words and fuzzy phrases representing uncertain semantics; The user statement sequence corresponding to the user request is input into the large language model, and the large language model is monitored at each time step when processing the user statement sequence. Based on all token sequences before that time step, the probability distribution of the next token is predicted. For each time step, the sum of the probability values ​​of tokens belonging to the set of fuzzy information token flags in the probability distribution corresponding to the user statement representation sequence is determined, and the fuzzy probability value corresponding to that time step is determined. The second fuzzy probability distribution is formed by the fuzzy probability values ​​corresponding to each time step.

3. The automated fuzzy request processing method based on multi-agent systems as described in claim 2, characterized in that: The disambiguation strategy for the target ambiguous information sites specifically includes the following steps: Retrieve relevant contextual information elements about the user's request from a preset database, construct a contextual information semantic feature vector, and perform initial disambiguation on the user request feature representation corresponding to the user's request through the contextual information semantic feature vector; If the information entropy corresponding to the user request feature representation after the initial disambiguation is greater than the preset information entropy threshold, then the user request feature representation is subjected to secondary disambiguation through external information elements.

4. The automated fuzzy request processing method based on multi-agent systems as described in claim 3, characterized in that: Secondary disambiguation is performed on the user request feature representation using external information elements. The specific steps include: If the membership function value of the target fuzzy information site in the user demand feature representation is greater than the preset fuzzy threshold, then the target fuzzy information site belongs to the target fuzzy site. Based on the location of the fuzzy information corresponding to the target fuzzy point, joint disambiguation is performed using user information elements and context information elements; If the information entropy of the user's request feature representation after joint disambiguation is less than or equal to the preset information entropy threshold, then the secondary disambiguation has been completed, and a fine-grained representation of the user's request has been obtained.

5. The automated fuzzy request processing method based on multi-agent systems as described in claim 4, characterized in that: Based on multi-agent processing of the disambiguated user requests, long-span coarse-to-fine granular task planning is performed to form a task heterogeneity graph, specifically including the following steps: By performing coarse-grained planning on the user's fine-grained request representation by multiple intelligent agents, a planning trajectory consisting of multiple stages of tasks is obtained from the output of each intelligent agent. The planned trajectories output by all agents are deduplicated and integrated to obtain multiple independent coarse-grained planned trajectories; For each stage task in each coarse-grained planning trajectory, fine-grained task planning is performed to form the stage planning trajectory corresponding to each stage task. The stage planning trajectories output by all agents are deduplicated and integrated to obtain multiple independent fine-grained planning trajectories; A task heterogeneous graph is constructed by using the fine-grained tasks in each independent fine-grained planning trajectory as nodes and the jump trajectories of different fine-grained tasks as edges.

6. The automated fuzzy request processing method based on multi-agent systems as described in claim 2, characterized in that: Constructing the task execution meta-path based on the aforementioned task heterogeneity graph specifically includes the following steps: The Monte Carlo tree search method takes the current task state as the root node, uses feasible operations in the task heterogeneous graph as the basis for branching child nodes, and explores the child nodes according to the upper confidence interval algorithm to select the next node to be expanded. When the selected node is a leaf node that has not been fully expanded, new task operations are added as child nodes according to the association relationship of the task heterogeneous graph to form a new path; The new path is randomly simulated and executed. After the simulation is completed, the gains from this simulation are propagated back along the search path to update the access count and cumulative gains of each child node on the search path. The process is repeated multiple times to obtain the task execution meta-path.

7. The automated fuzzy request processing method based on multi-agent systems as described in claim 1, characterized in that: After constructing the task execution meta-path based on the task heterogeneity graph, the method further includes: Task ambiguity identification is performed on the task nodes in the task execution metapath, and the task ambiguity identification includes step direction fuzzy identification and resource allocation conflict identification; For the identified steps that are ambiguous, the ambiguous instructions are corrected and clarified by introducing task context information and historical execution trajectory; For identified resource allocation conflicts and ambiguities, resource allocation adjustments are made to conflicting tasks using a priority scheduling method; Based on the disambiguated task nodes, a new execution plan for unambiguous requests is constructed.

8. A multi-agent-based automated fuzzy appeal processing system, used to implement the multi-agent-based automated fuzzy appeal processing method as described in any one of claims 1-7, characterized in that: include: The request acquisition and fuzzy information recognition module is used to acquire user requests and obtain the fuzzy information probability distribution of the user requests through a preset size model recognition strategy; The preset size model recognition strategy is to output the first fuzzy probability distribution and the second fuzzy probability distribution corresponding to the user's request through the fuzzy information recognition model and the large language model, respectively, and to perform weighted fusion of the first fuzzy probability distribution and the second fuzzy probability distribution to obtain the fuzzy information probability distribution; and to determine the information entropy of multiple words in the user's request based on the fuzzy information probability distribution, and to determine the target fuzzy information position of the words whose information entropy is greater than the preset information entropy threshold and their corresponding positions in the user's request. The fuzzy information recognition model is obtained by training an initial neural network based on a transformer architecture using a dynamic reward sampling strategy to construct a fuzzy information recognition training set. Constructing the fuzzy information recognition training set using the dynamic reward sampling strategy includes the following steps: Obtain the original sample set, which includes task samples labeled with different types and levels of fuzzy information; Each task sample in the original sample set is sequentially input into an initial neural network based on the transformer architecture, and the difficulty of the samples is classified based on the pass@k value of the initial neural network on different request samples to obtain multiple difficulty level sample sets; where pass@k value is the proportion of samples in which the initial neural network has at least one answer that matches the fuzzy information label corresponding to the request sample in k independent repeated responses, and k is a preset threshold for multiple number of attempts. Verify whether the number of samples in the sample sets of different difficulty levels decreases progressively as the difficulty level increases; If there is a difficulty level where the number of samples does not decrease as the difficulty level increases, then the sample set of the difficulty level with the abnormal number of samples will be randomly sampled and adjusted until the condition that the number of samples in the sample sets of different difficulty levels decreases step by step as the difficulty level increases is met. Based on the final adjusted sample sets of each difficulty level, a fuzzy information recognition training set is constructed for training the fuzzy information recognition model; The initial neural network based on the transformer architecture is trained by constructing a fuzzy information recognition training set through a dynamic reward sampling strategy, specifically including the following steps: The initial neural network based on the transformer architecture is subjected to reinforcement learning using the fuzzy information recognition training set; the objective function of the reinforcement learning stage is: ； in, (·) represents the expected function; This represents the initial neural network based on the transformer architecture for recognizing each problem in the training set of fuzzy information. Total number of responses; Indicates the first Answer in the next round; This indicates that the initial neural network based on the transformer architecture is designed for the task. In parameters Time The loss from the second answer, This indicates that the initial neural network based on the transformer architecture is in the old policy. The probability distribution at time; This indicates that the initial neural network based on the transformer architecture is used in the new policy. The probability distribution at time; This indicates the distribution of the new policy parameters compared to the distribution of the old policy parameters. Divergence; These are hyperparameters used to adjust... The influence of divergence constraints; The initial neural network that has completed the reinforcement learning is subjected to supervised learning based on the fuzzy information recognition training set; the loss function of the supervised learning is: ； in, To output the sequence length, For position The labeled token, Given a prior sequence of tokens In the case where the parameters of the initial neural network are Time output The probability distribution; The fuzzy request disambiguation and task planning module is used to execute a disambiguation strategy on the target fuzzy information points, and perform long-span coarse-grained task planning on the disambiguated user requests based on multiple agents to form a task heterogeneous graph. The request processing module is used to construct a task execution metapath based on the task heterogeneity graph, and process the user request along with the task execution metapath.

Images