Event process scheduling method, device and storage medium for multi-agent learning

Abstract

The present application discloses a method, device and storage medium for event process scheduling based on multi-agent learning, including: receiving structured data and unstructured text of an event text set through a preset data interface; extracting text semantic features of the event text set using a preset text understanding model, and generating event text features from the extracted text semantic features through a structured data processing model; inputting the event text features into an agent combination for data processing, and scheduling process nodes according to the output results of the agent combination. By constructing a multimodal data fusion mechanism and an adaptive process scheduling system, the present application achieves comprehensive extraction of event text features and efficient collaboration between agents, significantly improving the automation level and analysis depth of interactive abnormal behavior event processing, and providing more accurate and efficient technical support for event processing.

Description

Technical Field

[0001] The present application relates to the intersection of information processing and artificial intelligence, and in particular to an event process scheduling method, device and storage medium for multi-agent learning. Background Art

[0002] Existing systems use a one-way, trigger-response communication mechanism to handle events. This mechanism restricts data exchange between functional modules to predefined, fixed interfaces, making it impossible to dynamically adjust data communication paths and resource allocation strategies based on actual event processing needs.

[0003] In this technical architecture, data interaction between modules relies entirely on preset static interfaces, which makes it impossible for the system to automatically optimize the data flow path based on the phased characteristics of event processing. Furthermore, the system lacks dynamic resource scheduling capabilities and is unable to balance the computing resource requirements between modules in real time, resulting in low overall collaborative efficiency in processing event text sets at the system level. This problem is particularly evident when dealing with major and complex events, where the system's response speed and decision-making accuracy are significantly reduced. This is essentially due to the existing architecture's inability to achieve intelligent collaboration between modules, and each functional unit is always operating in isolation.

[0004] The above content is only used to assist in understanding the technical solution of this application and does not constitute an admission that the above content is prior art. Summary of the Invention

[0005] The main purpose of this application is to provide an event process scheduling method, device and storage medium for multi-agent learning, aiming to solve the technical problems that the existing technology cannot realize the intelligent fusion and adaptive process scheduling of event multimodal data, resulting in low processing efficiency and insufficient depth of correlation analysis.

[0006] To achieve the above objectives, this application proposes a method for event flow scheduling based on multi-agent learning, which includes:

[0007] Receive structured data and unstructured text of the event text set through a preset data interface;

[0008] Using a preset text understanding model to extract text semantic features of the event text set, and generating event text features from the extracted text semantic features through a structured data processing model;

[0009] Inputting the event text features into an agent combination for data processing, and scheduling the process nodes according to the output results of the agent combination, wherein the agent combination includes an entity recognition agent, a data completion agent, and an association analysis agent;

[0010] Among them, the scheduling of the process nodes specifically includes starting the data completion process when entity recognition triggers data supplementation, entering the association analysis process after data verification is passed, and pushing manual review after the association analysis meets the standards.

[0011] In one embodiment, the entity recognition agent is provided with a deep neural network and a rule base, and the steps of inputting the event text features into the agent combination for data processing and scheduling the process nodes according to the output results of the agent combination include:

[0012] Determine the event type of the event text set based on the event text features, and obtain entity extraction rules corresponding to the event type;

[0013] Deploying the entity extraction rules in a deep neural network, and inputting the event text set into the deep neural network to extract entity outputs;

[0014] Schedule process nodes based on output entities.

[0015] In one embodiment, the data completion agent is provided with a prediction model and a multi-source interface, and the step of scheduling the process nodes according to the output entity includes:

[0016] When it is determined based on the entity that the event text set is missing, triggering a data supplement process;

[0017] Inputting the entity into the prediction model, and predicting the supplementary content of the event text set based on the output result;

[0018] A data acquisition channel is created based on the multi-source interface, and the supplementary content is acquired through the data acquisition channel.

[0019] In one embodiment, the association analysis agent is provided with a graph neural network, and the step of scheduling the process nodes according to the output entities includes:

[0020] When it is determined based on the entity that the event text set is not missing, triggering the association analysis process;

[0021] Inputting the entity into the graph neural network, and generating a multi-dimensional association result of the event text set through the graph neural network;

[0022] The multi-dimensional correlation results are pushed to manual review.

[0023] In one embodiment, the step of inputting the entity into the graph neural network and generating a multi-dimensional association result of the event text set through the graph neural network includes:

[0024] Obtaining a feature map corresponding to the entity, and generating a deep representation node of the graph neural network based on the feature map;

[0025] Comprehensively calculating the information matching degree between the event document features of the event text set and the abnormal behavior feature library through the deep representation node, wherein the information matching degree includes the semantic feature matching degree, the structural feature matching degree and the time feature matching degree;

[0026] The multi-dimensional association result is generated according to the comprehensive calculation result of the information matching degree.

[0027] In one embodiment, before the step of extracting text semantic features of the event text set using a preset text understanding model and generating event text features from the extracted text semantic features using a structured data processing model, the method further includes:

[0028] Obtaining historical event text, and generating training samples based on the historical event text;

[0029] The pre-trained Legal-BERT is incrementally trained using the training samples, and the training results of the incremental training are used as the preset text understanding model.

[0030] In one embodiment, the steps of extracting text semantic features of the event text set using a preset text understanding model and generating event text features from the extracted text semantic features using a structured data processing model include:

[0031] Determining data features of the text semantic features, and preprocessing the text semantic features according to the data features, wherein the data processing includes logarithmic normalization, data format unification, and creation of a task relationship graph structure;

[0032] The pre-processed text semantic features are concatenated to serve as the event text features.

[0033] In one embodiment, the event process scheduling method for multi-agent learning further includes:

[0034] Real-time monitoring of resource usage of each process node;

[0035] Based on the resource usage rate, selecting a resource adjustment strategy from a preset fault tolerance strategy;

[0036] The resource adjustment strategy is used to adjust the resource allocation weight of each process node.

[0037] In addition, to achieve the above-mentioned purpose, the present application also proposes an event process scheduling device for multi-agent learning, which includes: a memory, a processor, and a computer program stored on the memory and runnable on the processor, and the computer program is configured to implement the steps of the event process scheduling method for multi-agent learning as described above.

[0038] In addition, to achieve the above-mentioned purpose, the present application also proposes a storage medium, which is a computer-readable storage medium, and a computer program is stored on the storage medium. When the computer program is executed by the processor, the steps of the event process scheduling method for multi-agent learning as described above are implemented.

[0039] One or more technical solutions proposed in this application have at least the following technical effects:

[0040] The technical solution of the present application receives the structured data and unstructured text of the event text set through a preset data interface; uses a preset text understanding model to extract the text semantic features of the event text set, and generates event text features from the extracted text semantic features through a structured data processing model; inputs the event text features into an intelligent agent combination for data processing, and schedules the process nodes according to the output results of the intelligent agent combination, wherein the intelligent agent combination includes an entity recognition intelligent agent, a data completion intelligent agent, and an association analysis intelligent agent; wherein the scheduling of the process nodes specifically includes starting the data completion process when entity recognition triggers data supplementation, switching to the association analysis process after data verification is passed, and pushing manual review after the association analysis meets the standards. By constructing an intelligent multimodal data fusion mechanism and an adaptive process scheduling system, the comprehensive extraction of event text features in the event text set and the efficient collaboration between intelligent agents are achieved, which significantly improves the automation level and analysis depth of interactive abnormal behavior event processing, and provides more accurate and efficient technical support for event processing. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the present application and, together with the description, serve to explain the principles of the present application.

[0042] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, for ordinary technicians in this field, other drawings can be obtained based on these drawings without any creative work.

[0043] Figure 1 A flowchart of the first embodiment of the event process scheduling method for multi-agent learning provided in this application;

[0044] Figure 2 This is a schematic diagram of the device structure of the hardware operating environment involved in the event process scheduling method for multi-agent learning in the embodiment of the present application.

[0045] The purpose, features and advantages of this application will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. DETAILED DESCRIPTION

[0046] It should be understood that the specific embodiments described herein are merely used to explain the technical solutions of the present application and are not intended to limit the present application.

[0047] In order to better understand the technical solution of the present application, a detailed description will be given below in conjunction with the accompanying drawings and specific implementation methods.

[0048] The main solution of the embodiment of the present application is: receiving structured data and unstructured text of an event text set through a preset data interface; using a preset text understanding model to extract text semantic features of the event text set, and generating event text features from the extracted text semantic features through a structured data processing model; inputting the event text features into an intelligent agent combination for data processing, and scheduling process nodes according to the output results of the intelligent agent combination, wherein the intelligent agent combination includes an entity recognition intelligent agent, a data completion intelligent agent, and an association analysis intelligent agent; wherein the scheduling of the process nodes specifically includes starting the data completion process when entity recognition triggers data supplementation, switching to the association analysis process after data verification is passed, and pushing for manual review after the association analysis meets the standards.

[0049] Existing events suffer from a core technical bottleneck of insufficient intelligence. The fundamental flaw lies in the system's one-way "trigger-response" communication mechanism. This mechanism restricts data exchange between functional modules to predefined, fixed interfaces, making it impossible to dynamically adjust data communication paths and resource allocation strategies based on actual event processing needs. This results in low overall collaborative efficiency in processing event text sets at the system level.

[0050] This application provides a solution by constructing a multimodal data collaborative analysis architecture to eliminate information silos, using reconfigurable intelligent processing units to dynamically adapt to different event text features, and optimizing the collaborative mechanism of each link through a real-time interactive feedback network, so that the event text set processing process can intelligently adjust the node configuration according to actual needs, breaking through the limitations of traditional fixed processes, and making a technical leap from passive execution to active optimization in event processing, significantly improving the overall processing efficiency of event text sets and the technical effect of system adaptability.

[0051] Based on this, the embodiment of the present application provides an event process scheduling method for multi-agent learning, referring to Figure 1 , Figure 1This is a flow chart of the first embodiment of the event flow scheduling method for multi-agent learning of this application. In this embodiment, the event flow scheduling method for multi-agent learning includes steps S10 to S30:

[0052] Step S10, receiving structured data and unstructured text of the event text set through a preset data interface;

[0053] Step S20, extracting text semantic features of the event text set using a preset text understanding model, and generating event text features from the extracted text semantic features using a structured data processing model;

[0054] Step S30, input the event text features into the intelligent agent combination for data processing, and schedule the process nodes according to the output results of the intelligent agent combination, the intelligent agent combination includes an entity recognition intelligent agent, a data completion intelligent agent and an association analysis intelligent agent; wherein, the scheduling of the process nodes specifically includes starting the data completion process when entity recognition triggers data supplementation, entering the association analysis process after data verification is passed, and pushing manual review after the association analysis meets the standards.

[0055] In this embodiment, a full-link solution from data access to decision output is constructed to address the problems of the separation of structured data and unstructured text, rigid processes, and insufficient module collaboration in event practice. The structured text and unstructured text of the event text set are received through a preset data interface, and the event text set is processed. The preset data interface refers to a standardized communication module that supports access to multiple data formats. Specifically, a RESTful API can be combined with a message queue to ensure the synchronous reception of structured data and unstructured text. That is, a dual-channel heterogeneous data processing architecture is adopted. The RESTful API used is responsible for receiving structured JSON data such as interaction record data, while the message queue asynchronously processes the unstructured text generated by the audio transcription of the interactive voice. In addition, by designing a data verification middleware, formatted messages are automatically filtered, and a data traceability module is added to record the original data source of the structured JSON data and the audio transcription log of the interactive voice. This ensures the synchronous reception of the structured data and the unstructured text. In this embodiment, the preset data interface is also a data interface of a specified system platform, that is, a data interface created between internal systems that can obtain the event text set.

[0056] According to the event text set received by the preset data interface, the event text set is processed by the preset text understanding model to extract relevant text semantic features. The preset text understanding model refers to a deep learning model that specifically processes the semantics of relevant documents. It can be implemented by using a pre-trained language model combined with a method of fine-tuning in related fields to effectively extract deep semantic features of unstructured texts such as interactive speech. The deep semantic features specifically include multi-dimensional parameter content such as relevant field element features, event entity object association network features, relevant field context features, psychological state features, event type features, and risk level features. The deep semantic features all exist in the form of dense vectors, and the interaction of different granularity features is achieved through the model's multi-head attention mechanism. It should be clear that different event types will activate differentiated feature combinations, including related abnormal behaviors that focus on data flow features, while human dynamic abnormal behaviors pay more attention to spatiotemporal trajectory features. Specifically, this processing needs to be implemented by pre-building a preset text understanding model. Before the step of using the preset text understanding model to extract the text semantic features of the event text set and generating event text features from the extracted text semantic features through the structured data processing model, it also includes:

[0057] Obtaining historical event text, and generating training samples based on the historical event text;

[0058] The pre-trained Legal-BERT is incrementally trained using the training samples, and the training results of the incremental training are used as the preset text understanding model.

[0059] In this embodiment, it is proposed to use the pre-trained Legal-BERT as the basic model, and perform incremental training on the basic model using event text as training samples, and construct the preset text understanding model based on the incremental training results, and at the same time set the attention mechanism in the created preset text understanding model.

[0060] Legal-BERT refers to a language representation model that has been pre-trained with relevant domain corpus. This can be achieved by pre-training on a text set containing professional documents about events and related clauses. The language representation model can effectively identify professional terminology and sentence structures in event texts. Incremental training refers to optimizing domain adaptability based on a pre-trained language representation model. This can be achieved by fine-tuning training using relevant survey texts, enabling the model to learn the unique expression features of the relevant surveys. The attention mechanism refers to the dynamic weight allocation module in a neural network. This can be achieved by using a multi-head self-attention structure, which strengthens the feature representation of key information by calculating the correlation between different words.

[0061] Specifically, by loading the pre-trained parameters of Legal-BERT, a basic model framework is established, which already has the general understanding ability of event text. Further, a set of event texts containing event descriptions and evidence lists is collected, and incremental training is performed using masked language modeling tasks to enable the model to learn the entity representation patterns unique to relevant event processing scenarios. At the model architecture level, a multi-head attention module is inserted between the encoder layers. This module automatically enhances the representation strength of vocabulary related to event elements by calculating the correlation weights between word vectors. Including when processing language interaction records, the attention mechanism will increase the feature weights of key information such as the names of relevant personnel and relevant times, and reduce the interference of irrelevant descriptive sentences.

[0062] In this embodiment, through the dual-stage optimization of domain pre-training and incremental training, the model can simultaneously grasp the general features of event texts and the specific features of related investigations. Compared with the method of simply using pre-training models, the newly added attention mechanism effectively solves the problem of information dilution in long text processing, and can accurately capture key elements scattered in different positions of the text. It solves the problem of inaccurate text feature extraction in event text sets, and improves the accuracy of parsing professional content such as behavioral pattern descriptions and event entity object association network structures. In entity recognition tasks, the model can correctly distinguish the multiple meanings of related terms. At the same time, the introduction of the attention mechanism significantly enhances the model's ability to discover hidden clues, including automatically focusing on text record paragraphs containing abnormal time nodes in event text records that are several pages long.

[0063] Furthermore, it is proposed to generate event text features from the extracted text semantic features using a structured data processing model, including distinguishing the data features of the text semantic features and performing data processing based on the data features. The data processing includes logarithmic normalization, data format unification, and the creation of a task relationship graph structure. The data processing results are then spliced ​​to generate event text features. That is, the steps of generating event text features from the extracted text semantic features using a structured data processing model include:

[0064] Determining data features of the text semantic features, and preprocessing the text semantic features according to the data features, wherein the data processing includes logarithmic normalization, data format unification, and creation of a task relationship graph structure;

[0065] The pre-processed text semantic features are concatenated to serve as the event text features.

[0066] Among them, logarithmic normalization refers to the logarithmic transformation and normalization of numerical features. Specifically, it can be achieved by using the natural logarithm function combined with maximum and minimum value scaling to eliminate the numerical differences between features of different dimensions. Data format unification refers to the conversion of heterogeneous data into a unified vector format. Specifically, it can be achieved by converting text into a fixed-dimensional vector through word embedding technology, combined with standardized encoding of numerical fields, to ensure that cross-modal data is computable. The task relationship graph structure refers to the topological structure that characterizes the logical relationship between event elements. It is specifically achieved by using graph structure data modeling, with nodes representing event elements and edges representing the relationship between elements. It is used to capture implicit associations that cannot be expressed by traditional tabular data.

[0067] Specifically, logarithmic normalization is performed on the numerical features in the event text set, mapping discrete values ​​to a continuous, uniform interval. Unstructured data such as inquiry logs and monitoring reports are converted into 768-dimensional semantic vectors using a pre-trained language model. Using a graph-structured building block, communication records between relevant personnel are modeled as an adjacency matrix, and data flows are converted into weighted edge attributes. Finally, the normalized numerical values, text vectors, and graph structure features are concatenated along the feature dimension to form a composite event text feature vector containing semantic, statistical, and topological information.

[0068] Compared with existing technologies, traditional event text sets use separate feature processing, that is, the structured data and unstructured text in the event text set are processed independently in their own subsystems, resulting in inconsistent feature spaces that cannot be directly associated. However, this application achieves deep fusion of event data at the feature level by building a unified multimodal data processing framework, which not only retains the characteristics of different data types but also establishes cross-modal association channels, effectively solving the feature fragmentation problem in traditional methods.

[0069] Through the above technical solution, this application can perform correlation calculations on the interaction data in the event materials and the semantic descriptions of the interrogation records in a unified vector space, automatically identifying the correspondence between abnormal data flows of relevant personnel and contradictions in confessions, and significantly improving the integrity of the text feature representation of complex events. In the case of abnormal signal interaction behavior, this technology can jointly model structured data such as call duration and interaction records with the specialized language in the chat text, accurately identifying the correlation of behavioral characteristics in abnormal behavior patterns.

[0070] Then, the extracted deep semantic features are processed by a structured data processing model to obtain event text features that represent the event text set. The structured data processing model refers to a feature fusion module that integrates heterogeneous data. Specifically, it can be implemented by combining feature engineering with graph embedding to uniformly represent structured data such as interaction record data and text features.

[0071] Specifically, the implementation process of the structured data processing model described in this embodiment adopts a multimodal feature fusion architecture, and achieves unified representation of structured data such as interaction record data and event text features through the following technical paths:

[0072] 1) In the data alignment phase, a time alignment matrix is ​​established to time-align the interaction timestamps in the interaction records with the event descriptions in the query records. The Dynamic Time Warping (DTW) algorithm is used to process non-uniformly sampled data and generate temporal correlation weights for the "data flow-text description" relationship.

[0073] 2) In the feature encoding phase, multi-granularity encoding is implemented for structured data: interaction record values ​​are logarithmically transformed and binned to generate discretized features. Graph embedding technology is used for interaction accounts, and the topological features of account nodes are learned using the Node2Vec algorithm. Interaction frequencies are converted into Fourier coefficients to represent periodic patterns. Simultaneously, the deep semantic features in the event text are reduced to the same vector space as the structured features, and a contrastive learning loss function is used to ensure alignment of the feature distributions of the two modalities.

[0074] 3) During the feature fusion phase, a gated cross-attention mechanism was designed to enable multimodal interaction: interaction record data features were used as queries, and text features as key-value pairs. Learnable gated weights were used to dynamically adjust the contribution of each modality. Specifically, when the "interacting account" text feature was detected, the weight coefficient of the interaction feature of the associated account was automatically increased. Experiments showed that this mechanism improved the discriminability of feature representations of abnormal behavior by 41%.

[0075] 4) Unified representation output ultimately generates a composite feature vector with specialized semantics. This vector comprises three core dimensions: 1) a numerical feature layer (normalized Z-score values ​​for data volume, frequency, etc.); 2) a topological feature layer (a 128-dimensional embedding vector of the user relationship graph); and 3) a semantic association layer (cross-attention scores between textual and numerical features). This representation can be visualized using t-SNE to show the mapping between data networks and abnormal behavior.

[0076] Specifically, the text semantic features extracted by the preset text understanding model include but are not limited to relevant entity features, spatiotemporal features, relevant element features, psychological semantic features, etc. contained in the event text set, wherein the relevant entity features include the role vectors of relevant personnel and their relationship weights, the standardized coding of relevant items, and the embedded representation of abnormal behavior means; the spatiotemporal features include the probability distribution of relevant time points (including the analysis results of fuzzy time descriptions), the hierarchical coding of geographic locations (three-level geographic vectors of provinces, cities, and districts and counties), and the serialized representation of movement trajectories (path features encoded by LSTM); the relevant element features include the matching degree of abnormal behavior constituent elements (four-dimensional scores such as subjective intent / objective behavior), the relevant clause citation association matrix (cosine similarity with specific relevant clauses), and the event entity object association network integrity index (quantitative assessment of key element coverage); the psychological semantic features include the confession consistency score (semantic difference between multiple inquiries), the emotional tendency index (intensity distribution of emotional words such as anger / fear), and the language style fingerprint (frequency of first-person use, density of modifiers, etc.).

[0077] Furthermore, the event text features generated by fusing the extracted text semantic features through the structured data processing model include the following core information layers:

[0078] 1) Fusion feature layer: data-text association matrix: cross-attention weights between interaction data and event text; spatiotemporal-semantic coupling features: such as the joint distribution of "large interaction time - event statement time"; multimodal fingerprint encoding: a unified 384-dimensional feature vector (including numerical / textual / graph structure features).

[0079] 2) Event profiling layer: abnormal behavior pattern classification vectors; risk level assessment (a 5-level rating of social harm and inspection urgency); event inspection path recommendations (explainable output for evidence completion);

[0080] 3) Traceable feature layer, feature source tag (data traceability identification for each dimension); feature confidence score (reliability assessment based on data quality); dynamic update flag (version control of feature evolution over time);

[0081] For example, based on the aforementioned data analysis, in abnormal behavior analysis, the system fuses the "container number" (structured data), the "abnormal customs declaration description" (text features), and the "personnel communication records" (graph-relational data) to generate a joint feature representation, automatically outputting structured conclusions such as "87% similarity in handling methods" and "two key data gaps." This representation supports visualization, such as using heat maps to display the overlap between data flows and communication networks.

[0082] As shown above, when dynamically generating an agent combination based on generated event text features, the agent combination refers to a cluster of agents dynamically configured based on the event text features. Specifically, each agent cluster is equipped with corresponding algorithms and data structures to perform intelligent processing of entities. Specifically, a reinforcement learning algorithm can be used to implement an agent selection strategy, automatically configuring multiple languages ​​for cross-border events. Specifically, this strategy involves entity recognition agents, data completion agents, and root-link agents.

[0083] Specifically, it can also be manifested as an intelligent agent scheduling framework driven by reinforcement learning to achieve precise matching of event text features and processing capabilities, that is, the intelligent agent scheduling framework is manifested as entity intelligent recognition agents, data completion agents and association analysis agents to perform precise matching of security inspection features and processing capabilities.

[0084] Specifically, in the entity recognition stage, the text content of the event text set is input into the bidirectional long short-term memory network for sequence labeling, and the conditional random field layer performs constraint optimization on the labeling results. At the same time, the entity rule library corresponding to the event type is matched for cross-validation. When it is detected that the key entity is missing or the confidence level is lower than the preset threshold, the data supplement node is automatically triggered to activate the completion process. In the completion stage, the key attributes of the missing content are predicted by the gradient boosting decision tree, the standard database interface is called in parallel to query structured information, the external data source is obtained through the network service interface, and the multi-source data is input into the reliability verification module for consistency verification. After data completion is completed, the event entity and its relationship data are imported into the graph computing engine, and multi-dimensional similarity calculation is performed based on the deep representation nodes of the feature map. Synonymous associations are identified through semantic feature matching, topological relationships are discovered through structural feature matching, and temporal feature matching establishes a temporal logic chain.

[0085] Furthermore, the event process scheduling method for multi-agent learning also includes:

[0086] Real-time monitoring of resource usage of each process node;

[0087] Based on the resource usage rate, selecting a resource adjustment strategy from a preset fault tolerance strategy;

[0088] The resource adjustment strategy is used to adjust the resource allocation weight of each process node.

[0089] Among them, real-time monitoring of resource usage of each process node refers to the continuous collection of operating indicators such as CPU utilization, memory usage, input and output throughput, etc. through monitoring agents deployed on computing nodes. Specifically, this can be achieved by using the Prometheus monitoring system combined with a custom indicator collector. This real-time data collection provides basic data support for dynamic resource allocation. The pre-set fault-tolerant strategy refers to the establishment of a relationship mapping table that includes resource overlimit thresholds, fault type classifications, and handling priorities. Specifically, this can be achieved by loading a predefined YAML configuration file through the rule engine. This strategy enables the system to automatically trigger a standardized handling process when an anomaly is detected. Resource allocation weight adjustment refers to the dynamic allocation of computing resources through a weighted algorithm based on the current load status and historical operating data. Specifically, this can be achieved by using a resource scheduling model based on reinforcement learning. This mechanism ensures that critical business nodes have priority in obtaining computing power when resources are tight.

[0090] Specifically, the monitoring module collects CPU utilization and memory usage from each node at a frequency of seconds. When a node's CPU utilization continuously exceeds a set threshold, the fault classification module is triggered to identify the type of anomaly. If a temporary load surge is identified, the elastic expansion mechanism is activated to temporarily allocate backup computing units. If a persistent resource shortage is identified, the resource weight calculation model is invoked to reallocate the CPU quota to each node. During this process, the historical fault database continuously records the handling effects of various abnormal events, providing an optimization basis for subsequent weight adjustments.

[0091] This detection and processing achieves load balancing and rapid response to exceptions across all nodes in the event processing process, significantly reducing the risk of process blockages caused by resource competition. During peak system operation, wait times for key analysis tasks are shortened, computing resource utilization is improved, and cascading failures caused by single node failures are avoided. When the electronic evidence analysis module experiences sudden high load, the system automatically reduces resource allocation for non-urgent tasks to ensure the continued operation of core correlation analysis functions.

[0092] In this embodiment, a multimodal data collaborative analysis architecture is constructed to eliminate information islands, a reconfigurable intelligent processing unit is used to dynamically adapt to different event text features, and the collaborative mechanism of each link is optimized through a real-time interactive feedback network, so that the event text set processing process can intelligently adjust the node configuration according to actual needs, breaking through the limitations of traditional fixed processes, and making a technical leap from passive execution to active optimization of event processing, significantly improving the overall processing efficiency of event text sets and the technical effect of system adaptability.

[0093] According to the agent combination indicated in one embodiment shown above, including an entity recognition agent, a data completion agent, and an association analysis agent, wherein the entity recognition agent is provided with a deep neural network and a rule base, the event text features are input into the agent combination for data processing, and the process nodes are scheduled according to the output results of the agent combination, including:

[0094] Determine the event type of the event text set based on the event text features, and obtain entity extraction rules corresponding to the event type;

[0095] Deploying the entity extraction rules in a deep neural network, and inputting the event text set into the deep neural network to extract entity outputs;

[0096] Schedule process nodes based on output entities.

[0097] In this embodiment, the entity recognition agent, data completion agent and association analysis agent are parallel processing processes, and the process nodes are scheduled after analyzing the output results of the processing.

[0098] The entity intelligent identifier adopts a dual-channel hybrid recognition architecture, which includes a deep neural network main channel and an auxiliary channel of regular expressions and semantic rule base. The deep neural network main channel is implemented based on BERT-BiLSTM-CRF and is equipped with pre-trained weights enhanced in related fields. The auxiliary channel of regular expressions and semantic rules contains more than 50,000 entity recognition rules in related fields. To this end, when processing data, the entity intelligent identifier extracts entities layer by layer in the structured data and unstructured text of the event text set through multi-granularity entity extraction. It can also set corresponding confidence thresholds for entities at each level, including setting the first-level entity as a person related to abnormal behavior, with a confidence threshold of ≥0.9; the second-level entity as a related item / tool / place, with a confidence threshold of ≥0.8; the third-level entity as time / amount / address, etc., with a confidence threshold of ≥0.7. Specifically, it can be determined according to the event type represented by the event text set. In the process of entity recognition through the entity intelligent recognition, the intelligent agent not only identifies the classification of related matters, but also accurately maps the social common names of objects to standardized names through molecular formula matching rules, while automatically marking the fuzzy descriptions with low confidence and triggering the completion process.

[0099] In this embodiment, a hierarchical extraction architecture driven by event type is used to achieve high-precision identification and quality control of event entities. During the event type determination and entity extraction process, the event type of the event text set can be determined based on a pre-set event type determination mechanism. This event type determination mechanism is essentially implemented through a multimodal classification model and dynamic rule enhancement. Specifically, the multimodal classification model utilizes an attention-based hybrid neural network architecture, with text and data channels configured to process text features and pattern features. Specifically, an attention mechanism is incorporated into the pre-set text understanding model formed by incremental training of Legal-BERT. This allows for the extraction of text features such as "relevant processing methods" and "subject amount" from the event text set when processing documents. 1D-CNN is then used to analyze pattern features in structured data such as interaction records and communication records. Subsequently, the fusion layer formed by the gated attention mechanism dynamically adjusts the weights of each module, outputting a probability distribution of 38 related event categories, thereby deriving the event type of the event text set.

[0100] Previously, a pre-built event type-feature mapping knowledge base was used for dynamic rule enhancement. This event type-feature mapping knowledge base included at least a key trigger word list, numerical feature thresholds, and spatiotemporal feature combinations. This dynamic rule enhancement approach further improved the accuracy of event type recognition for event text sets.

[0101] Entities are then extracted from the event text set based on entity extraction rules for the domain to which the event type belongs. The entity extraction rules are essentially a domain-adapted entity extraction process, comprising a deep neural network-rule base collaborative architecture for entity extraction. This collaborative architecture is implemented by deploying the entity extraction rules within the deep neural network. After entity extraction, the quality of the extracted entities is verified through entity extraction quality control. Within the deep neural network-rule base collaborative architecture, a BiLSTM-CRF main model is used as the base data model and implemented with a domain rule engine. The BiLSTM-CRF main model comprises an input layer, a hidden layer, and an output layer. The input layer concatenates event type embedding vectors with text character-level features, while the hidden layer has a dynamic depth design that allows for different hierarchical depths depending on the event type. The output layer can be assigned domain labels based on the entity sequence processed by the hidden layer. Based on the entity sequence with assigned domain labels, rule validation is performed using mandatory rules, supplementary rules, and checksum rules (e.g., "the relevant time must not be later than the event occurrence time") within the domain rule engine. Entities that have undergone rule validation are then subjected to quality control.

[0102] The entity extraction quality control system includes a triple validation mechanism, including confidence validation, logic validation, and integrity validation. The confidence validation has dynamic thresholds, meaning different event types have corresponding thresholds. The logic validation verifies the rationality of entity relationships through a pre-established knowledge graph. The integrity validation checks the coverage of the constituent elements of abnormal behavior events. The process based on this triple validation mechanism can include identifying high-confidence entities; automatically marking integrity gaps when missing entities are found; and triggering the data completion process to retrieve relevant text reports.

[0103] The above is achieved based on the completeness of the event text set. When it is determined that the event text set is incomplete, the processing logic of the event supplement node is triggered, that is, the supplement processing is performed based on the data supplement agent. The data supplement agent can be expressed as a hierarchical supplement system, that is, multiple levels are set to perform layer-by-layer supplement processing on the data, including a prediction model layer, a data interface layer, and a verification layer. The prediction model layer includes but is not limited to three supplement paths: integrated time series prediction (ARIMA), generative adversarial network (GAN), and knowledge graph reasoning; the data interface layer connects to relevant data sources such as big data platforms and credit reporting systems; the verification layer can verify and evaluate the data based on the dual reliability evaluation of cross-validation and expert rules.

[0104] In practical applications, the low-confidence entities output by the entity recognition agent are first parsed, and the integrity gaps of the event entity object association network are evaluated based on the low-confidence entities, including spatiotemporal continuity, interaction flow, and character relationships. The evaluation results of the integrity gaps of the event entity object association network are then processed accordingly for different types of data, including but not limited to using similar case patterns to predict numerical data, using a GPT model fine-tuned in related fields to generate suggestion questions for text data, and calling a map API for trajectory interpolation for spatiotemporal data. The completed data is output along with the data traceability path, and a rationality analysis report for the completed hypothesis is generated. The specific processing process can be exemplified as follows: In abnormal behavior events, for missing intermediate account information, the agent predicts possible transition accounts through interaction flow pattern analysis, and automatically retrieves the real-name information of the matching account from the abnormal behavior system for cross-verification. That is, based on the data completion process shown above, a prediction model and a multi-source interface are provided for the data completion agent, and the step of scheduling the process nodes according to the output entity includes:

[0105] When it is determined based on the entity that the event text set is missing, triggering a data supplement process;

[0106] Inputting the entity into the prediction model, and predicting the supplementary content of the event text set based on the output result;

[0107] A data acquisition channel is created based on the multi-source interface, and the supplementary content is acquired through the data acquisition channel.

[0108] In this embodiment, a missing content detection and classification mechanism is performed on the event text set. The data pattern for missing content identification can be defined as a three-dimensional gap identification model. A hierarchical analysis architecture is used to detect missing content. Since the overall structure of the event text set can be represented by the extracted entities, the missingness of the event text set can also be determined by the entities. In the three-dimensional gap identification model, a hierarchical analysis architecture is provided, which is used to detect missing content at the structural layer, semantic layer and logical layer respectively. Each data layer can be judged by inputting the corresponding entity as the detection object. Alternatively, the text information of the event text set corresponding to the entity is used as the judgment object. Specifically, it can be set according to the confidence of the entity. Among them, the structural layer detection checks the integrity of the required fields through XML Schema verification; the semantic layer detection verifies the completeness of the abnormal behavior constituent elements based on the knowledge graph of the relevant field; the logical layer detection uses a temporal reasoning engine to check the continuity of the timeline.

[0109] In addition, the types of missing gaps in the event text set can also be identified through entities, including entity-level missing (key roles / items are not identified, confidence = 0), attribute-level missing (quantity / time and other elements are incomplete, confidence ∈ (0, threshold)) and relationship-level missing (logical links between entities are broken).

[0110] As shown above, when it is determined that there is data missing in the event text set, the data completion process is triggered, that is, the data completion agent is executed to complete the data. In response to the gap type identification shown above, an intelligent supplement strategy selection is provided, which specifically includes calling the association analysis agent to determine the internal supplement of implicit information in the event text set, connecting to a professional big data platform to retrieve external supplement of related data, and generating an inquiry outline suggestion for reference by relevant event handling personnel. The data supplement nodes shown above can be selected based on the specific situation of the incompleteness of the event type to which the actual event text set belongs.

[0111] As shown above, when determining that the event text set is missing, based on the identified missing content detection and gap type identification, a preset gradient boosting decision tree algorithm is used to predict the specific missing content. The gradient boosting decision tree is also a core predictor of the gradient boosting decision tree (GBDT), whose feature engineering includes event type embedding vectors (38 dimensions), existing entity feature matrices (256 dimensions), and similar case pattern features (extracted via KNN). Based on the missing features determined by the feature engineering, a preset dynamic weighted prediction is used to determine the missing type, including numerical missing, categorical missing, and relational missing. Specifically, a regression tree is used to predict numerical missing, a classification tree is used to predict categorical missing, and a graph propagation algorithm is used to supplement the prediction of relational missing.

[0112] Based on the determination of missing features, an adversarial network is generated with an auxiliary prediction model to complete the document paragraphs (such as asking about missing records), and a time series stacking network is used to predict the time series of missing time points, and then the event entity elements are deduced through the rule engine of knowledge graph reasoning.

[0113] Supplementing the aforementioned missing data requires a pre-defined multi-source data acquisition channel, which includes a standard database query interface and a network service call interface. The standard database query interface includes internal systems and cross-departmental data sources. Based on these internal systems, a distributed query engine can retrieve data from population information repositories, historical event records repositories, and asset registries. Cross-departmental data sources utilize blockchain-based evidence to query abnormal behavior systems and event supervision databases. The network service call interface utilizes public data APIs for map services and corporate credit reporting, while dedicated data channels are used to access operator call records and cross-border data exchange.

[0114] Furthermore, the event text set can be supplemented through a manual input interface. In the manual input interface, an inquiry outline based on the gap type can be automatically generated through intelligent assisted input, and data verification can be forced through a structured input template. Voice conversion and proofreading can also be performed on the recorded data. The manual input interface also has an expert decision support mechanism, which can realize visual comparison of prediction results, evaluate the impact of the completion plan, and finally propose event entity risk warnings.

[0115] As shown above, by setting up a design completion process tracing module, completely recording the missing detection basis, data acquisition path and verification process log, it is possible to form a complete audit chain that meets the event evidence requirements when completing the event text set.

[0116] Furthermore, in conjunction with the aforementioned intelligent entity recognition and data completion agents, an association analysis agent is used to derive multidimensional association results. Specifically, the association analysis agent utilizes a hybrid computing engine, comprising a real-time computing layer, a deep analysis layer, and a knowledge fusion layer. The real-time computing layer uses Spark GraphX ​​to process high-speed associations across tens of millions of nodes, while the deep analysis layer utilizes the GraphSAGE graph neural network to mine potential relationships. The knowledge fusion layer can integrate a large number of historical event text features from the illegal information database. The specific number of historical event text features can cover the entire stored data in the illegal information database. In practical applications, Jaccard similarity is first calculated by constructing a social relationship subgraph to identify relevant anomaly patterns such as cyclic interactions and split flows. A geographic heat map and timeline of event elements are then created, forming a multidimensional association construct. Subsequently, inference rules based on relevant domain constraints are applied to determine the output of an association matrix and a visual analysis map, implementing an intelligent deduction process. Subsequently, a feedback loop is established between the event text features and the anomaly information database based on the multidimensional association construct and intelligent recommendation results, and the association rule weights are automatically optimized for dynamic updates. Specifically, the association analysis agent is provided with a graph neural network, and the step of scheduling the process nodes according to the output entity includes:

[0117] When it is determined based on the entity that the event text set is not missing, triggering the association analysis process;

[0118] Inputting the entity into the graph neural network, and generating a multi-dimensional association result of the event text set through the graph neural network;

[0119] The multi-dimensional correlation results are pushed to manual review.

[0120] In this embodiment, when the event text set is determined to be not missing based on the entity, the association analysis process is triggered. Based on the association analysis process, the association analysis agent is executed to perform multi-dimensional data association of the event text set. A graph neural network is provided in the association analysis agent to perform data association, that is, the step of inputting the entity into the graph neural network and generating a multi-dimensional association result of the event text set through the graph neural network includes:

[0121] Obtaining a feature map corresponding to the entity, and generating a deep representation node of the graph neural network based on the feature map;

[0122] Comprehensively calculating the information matching degree between the event document features of the event text set and the abnormal behavior feature library through the deep representation node, wherein the information matching degree includes the semantic feature matching degree, the structural feature matching degree and the time feature matching degree;

[0123] The multi-dimensional association result is generated according to the comprehensive calculation result of the information matching degree.

[0124] Based on the entities identified in the event text set, a feature graph associated with the entity is obtained. The feature graph is a pre-created data structure based on multimodal feature fusion, which can realize intelligent association mining between the event text set and the historical event information library. When constructing the feature graph, it is realized through multi-source data fusion modeling and dynamic graph update mechanism. The multi-source data fusion modeling includes an entity type system and a relationship type design, and the entity type system includes a specified subject, related items, related concepts, and spatiotemporal nodes. The relationship type design includes explicit relationships and implicit relationships. The dynamic graph update mechanism triggers the incremental update of the graph related to the new event data through the event-driven architecture based on the real-time stream, and adopts a rule-based three-stage verification method (syntax verification-logical conflict-related field conflict) to verify the updated new event data to update the feature graph.

[0125] When generating the deep representation nodes of the graph neural network based on the feature graph, a hybrid graph neural network architecture and deep representation optimization are implemented based on the graph neural network shown. In the hybrid graph neural network architecture, a structural feature extraction layer, a semantic feature fusion layer and a temporal dynamic modeling layer are provided. The structural feature extraction layer uses the GraphSAGE algorithm to learn the local topological features of the nodes, and uses the attention mechanism to distinguish the weights of different edge types such as "kinship" and "data transactions"; the semantic feature fusion layer can splice the BERT embedding vector of the event text with the graph structure feature, and use the feature gating mechanism to adjust the contribution of text and non-text features; the temporal dynamic modeling layer uses T-GNN to process the edge data with timestamps to generate a time-sensitive node representation.

[0126] The deep representation optimization classifies nodes with multi-task learning objectives to distinguish normal / abnormal entities, discovers potential relationships through connection prediction, and classifies and connects events with graphs; in the types involving event text sets, the generalization ability of new abnormal behavior patterns is improved through pre-training of historical event graphs and adversarial learning enhancement modules, thereby improving the multi-dimensional relationship association of event text sets.

[0127] As shown above, after generating the deep representation nodes of the graph neural network based on the feature map, the information matching degree between the event document features of the event text set and the abnormal behavior feature library is comprehensively calculated. In essence, it is a multi-dimensional matching degree calculation. The calculation process shown is characterized by three dimensions: semantic feature matching degree, structural feature matching degree, and time feature matching degree. The matching degree calculations in different dimensions have different calculation methods. Among them, the technical method of semantic feature matching degree is to calculate the cosine similarity between the event node and the event information library node by focusing on matching the event usage description vector, the feature embedding of related items, and the citation distribution of related articles; the structural feature matching degree is obtained by analyzing subgraph isomorphism detection, graph kernel function similarity calculation and key node centrality comparison; the time feature matching degree is calculated by using dynamic time warping (DTW) to analyze the relevant time series patterns, flow cycle characteristics, and communication contact time rules.

[0128] As for the calculated information matching degree, multi-dimensional fusion is performed through the association decision engine. The semantic matching degree, structural matching degree and time matching degree set in the weighted sum model are used in the multi-dimensional fusion strategy to perform matching synthesis, and the weight of the matching degree shown is automatically adjusted according to the event type. The calculated matching degree result is output interpretably to realize the visualization of the event entity object association network and the inspection that meets the event entity requirements. The event entity object association network shown includes generating a relationship network containing a matching heat map and marking key matching paths and confidence levels, and automatically filtering event entity object associations that exceed retrospective expiration, marking weak association relationships with insufficient event entity object associations, and forming multi-dimensional association results.

[0129] It should be noted that the above examples are only used to understand this application and do not constitute a limitation on the event process scheduling method for multi-agent learning in this application. More simple transformations based on this technical concept are all within the scope of protection of this application.

[0130] The present application provides an event process scheduling device for multi-agent learning, which includes: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions that can be executed by the at least one processor, and the instructions are executed by the at least one processor so that the at least one processor can execute the event process scheduling method for multi-agent learning in the above-mentioned embodiment one.

[0131] Reference below Figure 2, which shows a schematic diagram of the structure of an event flow scheduling device suitable for implementing multi-agent learning in accordance with an embodiment of the present application. The event flow scheduling device for multi-agent learning in accordance with an embodiment of the present application may include, but is not limited to, mobile terminals such as mobile phones, laptop computers, digital broadcast receivers, PDAs (Personal Digital Assistants), PADs (Portable Application Descriptions), etc., as well as fixed terminals such as digital TVs and desktop computers. Figure 2 The event process scheduling device for multi-agent learning shown is merely an example and should not impose any limitations on the functionality and scope of use of the embodiments of the present application.

[0132] like Figure 2 As shown, the event flow scheduling device for multi-agent learning may include a processing device 1001 (e.g., a central processing unit, a graphics processing unit, etc.), which can perform various appropriate actions and processes based on a program stored in a read-only memory (ROM) 1002 or a program loaded from a storage device 1003 into a random access memory (RAM) 1004. Various programs and data required for the operation of the event flow scheduling device for multi-agent learning are also stored in RAM 1004. The processing device 1001, ROM 1002, and RAM 1004 are connected to each other via a bus 1005. An input / output (I / O) interface 1006 is also connected to the bus. Typically, the following systems can be connected to the I / O interface 1006: input devices 1007 including, for example, a touch screen, touchpad, keyboard, mouse, image sensor, microphone, accelerometer, gyroscope, etc.; output devices 1008 including, for example, a liquid crystal display (LCD), speaker, vibrator, etc.; storage devices 1003 including, for example, a magnetic tape, hard disk, etc.; and communication devices 1009. The communication device 1009 can allow the event flow scheduling device for multi-agent learning to communicate with other devices wirelessly or by wire to exchange data. Although the figure shows the event flow scheduling device for multi-agent learning with various systems, it should be understood that it is not required to implement or have all of the systems shown. More or fewer systems may be implemented or have instead.

[0133] In particular, according to the embodiments disclosed in the present application, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, the embodiments disclosed in the present application include a computer program product comprising a computer program carried on a computer-readable medium, the computer program comprising program code for executing the method shown in the flowchart. In such an embodiment, the computer program can be downloaded and installed from a network via a communication device, or installed from a storage device 1003, or installed from a ROM 1002. When the computer program is executed by the processing device 1001, the above-mentioned functions defined in the method of the embodiment disclosed in the present application are executed.

[0134] The event process scheduling device for multi-agent learning provided by the present application adopts the event process scheduling method for multi-agent learning in the above-mentioned embodiment, which can solve the technical problems of low processing efficiency and insufficient depth of correlation analysis caused by the existing intelligent fusion and adaptive process scheduling of event multimodal data. Compared with the prior art, the beneficial effects of the event process scheduling device for multi-agent learning provided by the present application are the same as the beneficial effects of the event process scheduling method for multi-agent learning provided by the above-mentioned embodiment, and the other technical features of the event process scheduling device for multi-agent learning are the same as the features disclosed in the method of the previous embodiment, which will not be repeated here.

[0135] It should be understood that the various parts disclosed in this application can be implemented using hardware, software, firmware, or a combination thereof. In the description of the above embodiments, specific features, structures, materials, or characteristics can be combined in any one or more embodiments or examples in a suitable manner.

[0136] The above description is merely a specific embodiment of the present application, but the scope of protection of the present application is not limited thereto. Any changes or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in this application should be included in the scope of protection of this application. Therefore, the scope of protection of this application should be based on the scope of protection of the claims.

[0137] The present application provides a storage medium, which is a computer-readable storage medium having computer-readable program instructions (i.e., computer programs) stored thereon, and the computer-readable program instructions are used to execute the event process scheduling method for multi-agent learning in the above-mentioned embodiment.

[0138] The computer-readable storage medium provided herein may be, for example, a USB flash drive, but is not limited to electrical, magnetic, optical, electromagnetic, infrared, or semiconductor systems, systems, or devices, or any combination thereof. More specific examples of computer-readable storage media may include, but are not limited to, an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination thereof. In this embodiment, the computer-readable storage medium may be any tangible medium that contains or stores a program that can be used by or in conjunction with an instruction execution system, system, or device. The program code contained on the computer-readable storage medium may be transmitted using any suitable medium, including, but not limited to, wires, optical cables, RF (Radio Frequency), etc., or any suitable combination thereof.

[0139] The above-mentioned computer-readable storage medium can be included in the event process scheduling device for multi-agent learning; or it can exist independently without being assembled into the event process scheduling device for multi-agent learning.

[0140] The above-mentioned computer-readable storage medium carries one or more programs. When the above-mentioned one or more programs are executed by the event process scheduling device of multi-agent learning, the event process scheduling device of multi-agent learning realizes the technical content of the embodiment of the event process scheduling method of multi-agent learning shown above.

[0141] Computer program code for performing the operations of the present application may be written in one or more programming languages, or a combination thereof, including object-oriented programming languages ​​such as Java, Smalltalk, C++, and conventional procedural programming languages ​​such as "C" or similar programming languages. The program code may be executed entirely on the user's computer, partially on the user's computer, as a stand-alone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In the case of a remote computer, the remote computer may be connected to the user's computer via any type of network, including a local area network (LAN) or a wide area network (WAN), or may be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0142] The flow charts and block diagrams in the accompanying drawings illustrate the possible architecture, functions and operations of the systems, methods and computer program products according to various embodiments of the present application. In this regard, each box in the flow chart or block diagram can represent a module, program segment or a part of code, and the module, program segment or a part of code contains one or more executable instructions for realizing the specified logical function. It should also be noted that in some alternative implementations, the functions marked in the box can also occur in a different order than that marked in the accompanying drawings. For example, two boxes represented in succession can actually be executed substantially in parallel, and they can sometimes be executed in the opposite order, depending on the functions involved. It should also be noted that each box in the block diagram and / or flow chart, and the combination of the boxes in the block diagram and / or flow chart can be implemented by a dedicated hardware-based system that performs the specified function or operation, or can be implemented by a combination of dedicated hardware and computer instructions.

[0143] The modules described in the embodiments of the present application may be implemented in software or hardware, wherein the name of a module does not necessarily limit the unit itself.

[0144] The readable storage medium provided in this application is a computer-readable storage medium, which stores computer-readable program instructions (i.e., a computer program) for executing the above-mentioned method for scheduling event flows for multi-agent learning. This computer-readable storage medium can solve the technical problems of low processing efficiency and insufficient depth of association analysis caused by existing intelligent fusion and adaptive process scheduling of event multimodal data. Compared with the prior art, the beneficial effects of the computer-readable storage medium provided in this application are the same as those of the method for scheduling event flows for multi-agent learning provided in the above-mentioned embodiments, and will not be elaborated here.

Claims

1. A multi-agent learning event process scheduling method, characterized in that: The event process scheduling method for multi-agent learning includes the following steps: Receive structured data and unstructured text from an event text set through a preset data interface, wherein a message queue asynchronously processes the unstructured text while receiving the structured data using a RESTful API. The structured data is interaction record data, and the unstructured text is generated based on audio transcription of the interaction speech. Using a preset text understanding model to extract text semantic features of the event text set, and generating event text features from the extracted text semantic features through a structured data processing model; Input the event text features into an agent combination for data processing, and schedule process nodes based on the output results of the agent combination, wherein the agent combination includes an entity recognition agent, a data completion agent, and an association analysis agent; wherein the scheduling of the process nodes specifically includes starting the data completion process when entity recognition triggers data supplementation, switching to the association analysis process after data verification is passed, and pushing manual review after the association analysis meets the standards; The entity recognition agent is provided with a deep neural network and a rule base, inputs the event text features into the agent combination for data processing, and schedules the process nodes according to the output results of the agent combination, including: Determine the event type of the event text set based on the event text features, and obtain entity extraction rules corresponding to the event type; Deploying the entity extraction rules in a deep neural network, and inputting the event text set into the deep neural network to extract entity outputs; The process nodes are scheduled according to the output entity, wherein, when the event text set is determined to be missing based on the entity, the data supplement process is triggered; the entity is input into the prediction model, and the supplementary content of the event text set is predicted based on the output result; a data acquisition channel is created based on a multi-source interface, and the supplementary content is acquired through the data acquisition channel; when the event text set is determined to be not missing based on the entity, the association analysis process is triggered; the entity is input into the graph neural network, and the multi-dimensional association results of the event text set are generated through the graph neural network; and the multi-dimensional association results are pushed to manual review.

2. The event process scheduling method for multi-agent learning according to claim 1, characterized in that: The step of inputting the entity into the graph neural network and generating a multi-dimensional association result of the event text set through the graph neural network includes: Obtaining a feature map corresponding to the entity, and generating a deep representation node of the graph neural network based on the feature map; Comprehensively calculating the information matching degree between the event document features of the event text set and the abnormal behavior feature library through the deep representation node, wherein the information matching degree includes the semantic feature matching degree, the structural feature matching degree and the time feature matching degree; The multi-dimensional association result is generated according to the comprehensive calculation result of the information matching degree.

3. The event flow scheduling method for multi-agent learning according to claim 1, characterized in that: Before the step of extracting text semantic features of the event text set using a preset text understanding model and generating event text features from the extracted text semantic features using a structured data processing model, the method further includes: Obtaining historical event text, and generating training samples based on the historical event text; The pre-trained Legal-BERT is incrementally trained using the training samples, and the training results of the incremental training are used as the preset text understanding model.

4. The event process scheduling method for multi-agent learning according to claim 1, characterized in that: The steps of extracting text semantic features of the event text set using a preset text understanding model and generating event text features from the extracted text semantic features using a structured data processing model include: Determining data features of the text semantic features, and preprocessing the text semantic features according to the data features, wherein the data processing includes logarithmic normalization, data format unification, and creation of a task relationship graph structure; The pre-processed text semantic features are concatenated to serve as the event text features.

5. The event flow scheduling method for multi-agent learning according to any one of claims 1 to 4, characterized in that: The event process scheduling method for multi-agent learning also includes: Real-time monitoring of resource usage of each process node; Based on the resource usage rate, selecting a resource adjustment strategy from a preset fault tolerance strategy; The resource adjustment strategy is used to adjust the resource allocation weight of each process node.

6. An event process scheduling device for multi-agent learning, characterized in that: The event process scheduling device for multi-agent learning stores a computer program, and when the computer program is executed by a processor, it implements the event process scheduling method for multi-agent learning described in any one of claims 1-5.

7. A storage medium, characterized in that: The storage medium stores a computer program, which, when executed by a processor, implements the event process scheduling method for multi-agent learning described in any one of claims 1-5.

Images