Intelligent distribution and collaborative disposal method and system for urban events based on artificial intelligence

Abstract

The present application relates to a city event intelligent distribution and collaborative disposal method and system based on artificial intelligence, which uses the original city event text and attribute data output by the AI module, extracts multi-granularity semantic features and fuses behavior-object-scene triple chain information, monitors the distribution change of event semantics in each preset subspace, and generates multi-dimensional semantic drift monitoring indicators; if a drift signal is detected, anchor point labels are screened and a semantic closeness topology is constructed in combination with historical data, the identification and verification of the to-be-repaired edge are realized, the edge connection effectiveness is quantified through a semantic verification model, the label topology structure is dynamically soft-reconstructed, and semantic bridging nodes with a to-be-verified state are injected, and then the event classification rule library is incrementally updated; finally, the solidification and system evolution of the label relationship are completed in cooperation with the artificial review process. The method can significantly improve the response speed and semantic adaptability of the city event classification model to new events, and realize the self-evolution and closed-loop management of the label system.

Description

Technical Field

[0001] This invention relates to the field of intelligent allocation and semantic classification management technology for urban events, and in particular to a method and system for intelligent allocation and collaborative handling of urban events based on artificial intelligence. Background Technology

[0002] Currently, in the field of intelligent urban event allocation and semantic classification management, most mainstream technologies rely on deep neural networks or statistical learning models trained on historical samples to achieve automatic event identification and label classification. Specifically, this typically involves collecting existing urban event text descriptions, structured attributes, and manual classification results to construct an annotated corpus. A classifier is then trained using methods such as convolutional neural networks, long short-term memory networks (LSTM), and graph neural networks to automatically classify and label newly reported events. In recent years, some systems have introduced model architectures such as transfer learning, active learning, and semantic embedding, as well as supporting tools such as knowledge graphs and label system management, attempting to improve the accuracy and automation of intelligent urban event allocation. The mainstream development trend is characterized by: on the one hand, continuously optimizing the overall recognition capability of event classification algorithms driven by big data in urban governance; and on the other hand, emphasizing the manageability of label structures and exploring the standardization of labels and the integration of multi-level linkage label relationship systems. For sudden, novel, or combined urban events, some algorithms enhance model generalization capabilities through multimodal data fusion, balance sampling, and expert labeling.

[0003] Representative technical solutions mainly include urban event text classification based on end-to-end semantic segmentation, label tree structure updating driven by expert knowledge, event allocation and scheduling relying on reinforcement learning processes, and label evolution mechanisms that combine clustering algorithms with manual induction. These solutions can typically achieve efficient identification and basic classification of common event types, are suitable for scenarios where the event description distribution is consistent with historical data, and have good adaptability in daily urban governance, high-frequency event flow, and rule-based problem allocation.

[0004] However, existing technologies have the following prominent shortcomings: First, current AI classification models generally rely on historical labeled datasets for centralized training, and their generalization ability is severely limited by the distribution of training data. When sudden new events, composite or cross-type problems with significantly different semantic expressions appear in urban governance scenarios, the models often fail to identify and classify them in a timely and accurate manner, leading to delays in allocation, increased frequency of manual intervention, and affecting the overall intelligence level of the system.

[0005] Second, most existing labeling systems adopt static or semi-static tree structures, lacking adaptive evolution capabilities and failing to effectively respond to the dynamic ambiguity of event type boundaries and the need for timely soft linkage between label definitions and business feedback in urban practice scenarios. Some "label expansion" solutions require the introduction of a large number of external experts for review or retraining of large models, which is not only cumbersome and slow to respond, but also makes it difficult to continuously solidify the label correction process in a closed loop, affecting the traceability of the semantic system and the efficiency of business implementation.

[0006] Third, current mainstream methods for addressing needs such as defining label boundaries, dynamically identifying ambiguous label relationships, and automatically monitoring semantic drift largely rely on clustering or statistical distribution distance indicators. However, these methods lack full utilization of the multi-layered descriptions and collaborative feedback mechanisms of complex real-world urban governance events, failing to achieve true "self-healing of the label system" and flexible extension of semantic boundaries based on the actual data flow of urban event handling. Meanwhile, some automated methods introduce external large language models or open label generation mechanisms, making it difficult to guarantee the authority and business interpretability of their labels, thus impacting the business trust foundation of intelligent systems.

[0007] With increasingly refined urban governance, continuously evolving urban event types, and intensifying semantic boundary changes, there is an urgent need for a dynamic event classification method that does not rely on new labeled samples or large-scale model retraining, can perceive semantic drift in real time, and can automatically achieve self-repair of the label system architecture and flexible extension of semantic boundaries. This method must ensure lightweight real-time system response while also considering the self-evolution of the label system and the traceability of business closed loops, fundamentally improving the ability of AI-driven urban event distribution platforms to identify and manage unknown, novel, and ambiguous events. This is precisely the innovative starting point and core technological breakthrough direction of this technical solution. Summary of the Invention

[0008] This application provides an artificial intelligence-based method and system for intelligent allocation and collaborative handling of urban events, aiming to solve one of the problems or issues of the existing technology mentioned in the background.

[0009] The method and system for intelligent allocation and collaborative handling of urban events based on artificial intelligence provided in this application include: On the one hand, AI-based intelligent allocation and collaborative handling methods for urban events specifically include: Obtain urban event text and attribute data, extract phrase embedding vectors from the text and parse dependency paths to construct triple chains, and generate event semantic feature sets.

[0010] Based on the distribution of the event semantic feature set in the preset semantic subspace, the offset of the current batch of events relative to the cluster centers of each semantic subspace, the rate of change of edge sample density, and the number of cross-subspace bridging samples are calculated to generate a drift index set.

[0011] Determine whether any indicator in the drift indicator set exceeds a dynamic threshold. If it does, trigger a semantic drift signal and locate several anchor tags that are semantically closest to the drift event from the tag topology map based on the signal, and generate an anchor tag set.

[0012] A semantic proximity matrix is ​​constructed using the anchor tag set. Starting from the anchor node, a limited step-size diffusion is performed along the graph edge to identify the set of edges to be evaluated in the semantically ambiguous zone and generate a list of edges to be repaired.

[0013] The anchor point label definition text corresponding to each edge in the list of edges to be repaired and the drift event examples are input into the semantic verification model. By comparing the consistency of verb role filling, spatial constraint compatibility and the similarity of processing timeliness trend, an edge validity score sequence is generated.

[0014] Perform topological reconstruction based on the scoring sequence to generate an evolved tag topology map.

[0015] The city event classification rule base is updated based on the evolved tag topology map. Temporary semantic bridging nodes are incorporated into the real-time classification matching logic to generate updated classification rule data instances for event allocation and collaborative handling.

[0016] All repair operations of the evolved tag topology map are recorded in the tag evolution log and a manual review process is initiated. After receiving the manual confirmation instruction, the temporary semantic bridge nodes with the status to be verified are solidified into formal tag relationships, generating the final urban event tag system that completes the closed-loop management of the entire life cycle.

[0017] On the other hand, the AI-based intelligent allocation and collaborative handling system for urban events specifically includes: The semantic feature module is used to acquire urban event text and attribute data, extract phrase embedding vectors from the text and parse dependency paths to construct triple chains, and generate event semantic feature sets. The drift index module is used to calculate the offset of the current batch of events relative to the cluster centers of each semantic subspace, the rate of change of edge sample density, and the number of cross-subspace bridging samples based on the distribution state of the event semantic feature set in the preset semantic subspace, and to generate a drift index set. Anchor tag module is used to determine whether any index in the drift index set exceeds a dynamic threshold. If it does, a semantic drift signal is triggered, and the anchor tag module is used to locate the anchor tags that are closest to the semantic distance of the drift event from the tag topology map based on the signal, and generate an anchor tag set. The edge-to-be-repaired list generation module is used to construct a semantic proximity matrix using the anchor tag set, perform limited step-size diffusion along the graph edge starting from the anchor node, identify the set of edges to be evaluated in the semantically ambiguous zone, and generate an edge-to-be-repaired list. The edge validity scoring module is used to input the anchor point label definition text corresponding to each edge in the list of edges to be repaired and the drift event examples into the semantic verification model. By comparing the consistency of verb role filling, spatial constraint compatibility and similarity of processing timeliness trend, an edge validity scoring sequence is generated. The topology reconstruction module is used to perform topology reconstruction based on the scoring sequence to generate an evolved label topology map; The classification rule update module is used to update the city event classification rule library based on the evolved label topology map, incorporate temporary semantic bridging nodes into the real-time classification matching logic, and generate updated classification rule data instances for event allocation and collaborative handling.

[0018] The AI-based intelligent allocation and collaborative handling method and system for urban events provided in this application have the following beneficial effects: (1) By constructing a dynamic label topology with semantic evolution capabilities, the adaptability of the event classification system to the continuous changes in urban governance scenarios is significantly improved. Traditional event classification methods generally adopt static label systems, which are difficult to cope with the emergence of new event types and the semantic shift of existing events (such as "shared bicycles parked randomly" gradually evolving into "electric bicycles encroaching on blind paths"), resulting in the decay of sorting accuracy over time and frequent manual intervention. This solution introduces a semantic drift perception unit in the event sorting module, abandoning the traditional detection mechanism based on statistical distribution differences, and instead starting from a multi-granular semantic space, comprehensively analyzing the cluster center shift, edge density changes, and cross-class bridging sample growth trend of events in specific semantic subspaces, to achieve early sensitive capture of semantic evolution signals. This mechanism can accurately identify the label boundaries and relationships that need to be adjusted without relying on full data relabeling or large model fine-tuning, thereby effectively overcoming the classification lag problem caused by label rigidity in existing technologies, and greatly improving the classification stability and business fit of the system under long-term operation.

[0019] (2) A restricted graph repair mechanism guided by a semantic intimacy matrix is ​​proposed, realizing a lightweight, interpretable, and closed-loop autonomous process for optimizing the label system architecture. Addressing the high costs and uncontrollable risks associated with traditional methods that heavily rely on expert participation or external language model generation during label updates, this solution innovatively utilizes real feedback data accumulated in the urban governance closed loop—including the frequency of historical collaborations, the semantic co-occurrence intensity of citizen complaints, and correction records from handling departments—to construct a dynamic semantic intimacy matrix. Based on this matrix, local graph diffusion is initiated to lock the key edge set in the semantic ambiguity zone. Furthermore, by deploying a semantic verification model, the effectiveness of potential connections can be evaluated based solely on anchor label definitions and a small number of drifting sample texts. Fine-grained judgments are made by combining dimensions such as verb role consistency, spatial logic compatibility, and handling feedback trends, avoiding semantic distortion caused by blind merging or breakage. For inefficient connections, a soft pruning strategy is adopted to retain historical traces; for emerging reasonable associations, temporary bridging nodes with a "to be verified" state are injected and observation tasks are automatically triggered, ensuring both the flexibility of topology evolution and maintaining the system's controllability and audit traceability. The entire repair process does not require retraining complex models or introducing external large language models, which significantly reduces computational overhead and deployment threshold, while enhancing business personnel's understanding and trust in the tag change logic.

[0020] (3) The constructed adaptive label evolution path forms a semantic infrastructure closed loop driven by real governance feedback, effectively supporting the sustainable evolution of the intelligent allocation system in complex and dynamic environments. Unlike the traditional paradigm that relies on one-time clustering or offline expert induction, this solution embeds the evolution of the label system into the entire life cycle management process of urban events, enabling the classification logic to evolve synchronously with the adjustment of law enforcement standards, the migration of public demands, and the emergence of new illegal behaviors. The label evolution log fully records each repair operation and its basis, supporting one-click solidification into formal knowledge after manual review, balancing automation efficiency and human supervision authority. This closed-loop mechanism of "perception-evaluation-repair-verification" not only improves the relevance and coverage of event classification, but also enhances the system's robustness and self-repair capability against external disturbances. The resulting dynamic label topology not only serves the current allocation decision, but also serves as an expandable knowledge base to support advanced applications such as subsequent trend prediction, cross-regional experience transfer, and policy simulation, providing the urban governance system with intelligent basic support that combines real-time response capability, semantic expression depth, and continuous learning potential.

[0021] The aforementioned technologies work together to construct an event classification system that is responsive, lightweight, autonomously evolving, and business-reliable. This system overcomes the multiple bottlenecks of traditional static classification frameworks in terms of dynamic adaptability, maintenance costs, and interpretability, achieving a paradigm shift from "passive mapping" to "active evolution" and providing a solid and reliable semantic cognitive foundation for intelligent urban governance. Attached Figure Description

[0022] Figure 1 This is the main flowchart of an AI-based intelligent allocation and collaborative handling method for urban events.

[0023] Figure 2 This is a sub-flowchart of an AI-based intelligent allocation and collaborative handling method for urban events.

[0024] Figure 3 This is another sub-flowchart of the AI-based intelligent allocation and collaborative handling method for urban events. Detailed Implementation

[0025] Embodiments of the present invention are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0026] The following disclosure provides many different embodiments or examples for implementing different structures of the invention. To simplify the disclosure, specific examples of components and arrangements are described below. Of course, these are merely examples and are not intended to limit the invention. Furthermore, reference numerals and / or letters may be repeated in different examples; such repetition is for simplification and clarity and does not in itself indicate a relationship between the various embodiments and / or arrangements discussed.

[0027] like Figure 1 As shown, this application provides a method and system for intelligent allocation and collaborative handling of urban events based on artificial intelligence, specifically including: S1: Obtain urban event text and attribute data, extract phrase embedding vectors from the text and parse dependency paths to construct triple chains, and generate event semantic feature sets.

[0028] S2: Based on the distribution of the event semantic feature set in the preset semantic subspace, calculate the offset of the current batch of events relative to the cluster center of each semantic subspace, the rate of change of edge sample density, and the number of cross-subspace bridging samples, and generate a drift index set.

[0029] S3: Determine whether any indicator in the drift indicator set exceeds the dynamic threshold. If it does, trigger the semantic drift signal and locate the anchor tag closest to the semantic distance of the drift event in the tag topology map based on the signal, and generate the anchor tag set.

[0030] S4: Construct a semantic proximity matrix using the anchor tag set, and perform restricted step-size diffusion along the graph edges starting from the anchor nodes to identify the set of edges to be evaluated in the semantically ambiguous zone and generate a list of edges to be repaired.

[0031] S5: Input the anchor point label definition text corresponding to each edge in the list of edges to be repaired and the drift event examples into the semantic verification model. By comparing the consistency of verb role filling, spatial constraint compatibility and the similarity of processing timeliness trend, generate an edge validity score sequence.

[0032] S6: Perform topological reconstruction based on the scoring sequence to generate an evolved tag topological map.

[0033] S7: Update the city event classification rule base based on the evolved label topology graph, incorporate temporary semantic bridging nodes into the real-time classification matching logic, and generate updated classification rule data instances for event allocation and collaborative handling.

[0034] Step S1: Obtain city event text and attribute data, extract phrase embedding vectors from the text, and parse dependency paths to construct triple chains, generating an event semantic feature set. Specifically, this includes: S1.1: Obtain the urban event text and attribute data including time, location and subject attributes output by the AI ​​recognition module in the urban event intelligent distribution scenario, perform word segmentation, cleaning and stop word filtering preprocessing on the urban event text to generate a standardized urban event text corpus.

[0035] Parallel data reception processing is performed on the urban event text and attribute data containing time, location, and subject attributes continuously output by the AI ​​recognition module in the urban event intelligent distribution scenario. The text data and attribute data are cached in a high-reliability data channel to ensure the integrity and time synchronization of the input signal.

[0036] The text of urban events is loaded into the text preprocessing engine, and the whole-domain urban governance scenario dictionary is called to perform word segmentation based on the maximum matching principle. During the segmentation process, the missing subject-type words are supplemented by referring to the subject attributes in the structured attributes to enhance the semantic consistency between the text and the attributes.

[0037] The word segmentation results are cleaned to filter out noisy words containing non-governance scenario key information, including advertising words, redundant prompts and unstructured filler tags, to ensure that the corpus content input to the subsequent model focuses on the core semantics of governance events.

[0038] Identifying and removing stop words based on a stop word lexicon, performing partition matching of the stop word lexicon by governance scenario category and executing conditional removal logic, cleaning high-frequency but irrelevant function words appearing in attributes such as time, location, and subject, while retaining positioning words that contribute to the construction of spatial relationships.

[0039] Performing encoding standardization operations on the vocabulary set after cleaning and stop word filtering, uniformly converting multilingual mixed vocabularies into a predefined coding system in the urban governance field, and using a character mapping table and a semantic label comparison table to ensure the semantic consistency and retrievability of text corpora in multi-source input scenarios.

[0040] Through the above-mentioned chained processing method, the urban event text and attribute data input in the previous step are converted into a standardized urban event text corpus that conforms to the input format of the pre-trained embedding model and the dependency parsing model, so as to achieve the expected corpus quality improvement and semantic consistency enhancement.

[0041] For example, in a city's urban governance platform, for the original urban event text "there are vendors occupying the road at the school gate" output by the AI recognition module and the attribute data {Time: 2024-07-10 08:30, Location: "a certain city's Seventh Primary School", Subject: "vendors"}, the maximum matching window length is set to 5 characters during word segmentation to obtain the vocabulary set ["school gate", "vendor", "road occupation"], and the subject attribute "vendor" is directly bound to the relevant vocabulary through the attribute supplement logic to ensure semantic consistency; in the cleaning processing stage, the prompt word "there are" that does not affect the governance semantics is removed, and the core phrase is retained; in the stop word filtering stage, the stop word lexicon is used to remove high-frequency irrelevant function words, while retaining "school gate" related to spatial relationships; in the encoding standardization stage, "road occupation" is mapped to the governance code C321, "school gate" is mapped to the spatial code S127, and "vendor" is mapped to the subject code P512, forming a standardized corpus item [C321, S127, P512], where all codes conform to the domain preset semantic label system. This standardized corpus significantly improves the context matching accuracy of the embedding model in subsequent phrase-level embedding vector extraction, and enhances the stability of predicate-argument recognition in the dependency parsing stage, so that the construction performance of the event semantic feature set is greatly improved.

[0042] S1.2: Based on the standardized urban event text corpus, call the pre-trained urban governance domain-specific phrase embedding model to perform local context window sliding calculation, so as to extract a set of phrase-level embedding vectors that contain semantic information of specific governance scenarios.

[0043] Based on a standardized urban event text corpus, the parameter matrix of a phrase embedding model for urban governance scenarios optimized with domain knowledge is loaded, and the window length and step size configuration of the local context window sliding calculation are initialized to ensure that the embedding results are sufficiently sensitive to changes in context.

[0044] Each urban event text in the standardized text corpus is sliced ​​sequentially according to the window length to generate a set of continuous phrase fragments covering all words, and core words and related words related to a specific governance scenario are retained in each phrase fragment.

[0045] The embedding model is called to perform vectorization mapping for each phrase fragment, converting the phrase semantics into numerical vectors through the model weight matrix, and retaining multi-dimensional semantic components in the embedding result to characterize the scene semantic information.

[0046] During the embedding computation process, the semantic bias of the embedding vector is corrected by using a vocabulary of events in the field of urban governance, so that the vector components of verbs and object words that are highly relevant to governance business are highlighted in the projection computation.

[0047] The embedding vectors of all phrases are indexed and bound according to the order of the city event texts in the corpus to form a set of phrase-level embedding vectors. Phrase position identifiers and context coverage descriptions are added to the set to support index matching in the subsequent feature fusion and syntactic parsing stages.

[0048] Through the aforementioned local context window sliding calculation and embedding mapping processing, the standardized urban event text corpus from the previous step is transformed into a set of phrase-level embedding vectors containing semantic information of specific governance scenarios, thereby achieving an accurate mapping of event semantic representation from the original text to the vector space.

[0049] For example, in a city's illegal parking management scenario, a standardized urban event text corpus contains entries such as "vehicles parked at the intersection of Dongfeng Street obstructing traffic." The window length is set to 5 words, and the step size is 2 words, generating phrase fragments such as "parked at the intersection of Dongfeng Street," "vehicles parked at the intersection of Dongfeng Street," and "vehicles parked at the intersection obstructing traffic." A dedicated urban management embedding model is used, employing 200-dimensional semantic space embedding to map "parked at the intersection of Dongfeng Street" into a numerical vector. Semantic bias weights are added to dimensions 45 to 60 of the vector to highlight the governance relevance between "parking" and "obstructing traffic." The embedding model uses a local context window sliding formula: ,in For phrase embedding vectors, Let i be the word vector of the i-th word within the window. The value represents the window length. In this scenario, the embedded vector set outputs a 200-dimensional numerical vector containing 120 phrases. Each vector is appended with the phrase's starting word number and coverage information. After manual verification, this set can significantly improve the accuracy of subsequent dependency parsing in identifying the logical roles of events.

[0050] S1.3: Utilize a standardized urban event text corpus to execute a dependency parsing algorithm to identify the grammatical dependency paths between predicate verbs and their argument components in the text, in order to construct behavior-object-scene triplet chain data that reflects the inherent logical structure of the event.

[0051] S1.4: Based on the phrase-level embedded vector set and the behavior-object-scene triplet chain data, perform multi-granularity semantic feature fusion mapping operation to align and encapsulate the vector space representation with the structured logical relationship to generate an event semantic feature set.

[0052] When performing feature fusion mapping on the phrase-level embedding vector set and the behavior-object-scene triple chain data, the phrase-level embedding vector set is first used as the input of the continuous vector space. The semantic alignment matrix is ​​used to calculate the matching relationship of each phrase embedding vector in the triple chain corresponding to the behavior predicate, object noun and scene adverbial, forming a matching vector group containing mapping weights.

[0053] A one-to-one association is established between the matching vector group and the triple chain data. The contribution coefficients of different semantic granularities in the fusion process are adjusted through weight normalization operation to ensure the scale consistency of phrase semantic information and structured logical relationship in a unified space.

[0054] Multi-source heterogeneous feature concatenation is performed on the normalized vector, combining the contextual semantic features of phrase embedding with the logical path features of triple chain in node order to form a composite feature vector. Feature sparsity processing is used to eliminate low-weight noise components and improve the computational efficiency of the feature set.

[0055] By employing feature principal component projection based on the covariance matrix, the dimensionality reduction operation is performed on the concatenated composite feature vector, retaining high-contribution components that can simultaneously reflect the semantic and logical structure correlation of the text, and generating a fusion feature representation with stable metric properties in the vector space.

[0056] By constructing a data encapsulation object of multi-granular semantic feature set through feature fusion mapping results, the dimensionality-reduced vector space representation and triple logical structure are simultaneously bound to form a unified input format that can be directly used by downstream semantic subspace mapping and drift detection.

[0057] By performing multi-granularity semantic feature fusion mapping operations, the phrase-level embedding vectors and triplet chain data obtained in the previous step are transformed into an event semantic feature set with dual representations of contextual semantics and behavioral logic, achieving the expected technical effect of multi-dimensional feature alignment and unified encapsulation.

[0058] For example, in the description of the urban event "setting up a barbecue stall under the bridge," the phrase embedding vector set is derived from the sliding window calculation of a pre-trained embedding model on a standardized text corpus. The vector dimension is 128, and the behavior-object-scene triplet chain data is {behavior: setting up, object: barbecue stall, scene: under the bridge}. The mapping weights are calculated using a semantic alignment matrix, where the alignment score between the behavior predicate and the phrase embedding is 0.85, the alignment score between the object noun and the scene adverbial is 0.92, and the alignment score between the scene adverbial is 0.88. After normalization, the weights are 0.34, 0.36, and 0.30, respectively. These alignment weights are bound to the corresponding triplet nodes, and composite feature concatenation is performed to obtain a combined vector with a dimension of 384. After sparsification, noisy features with weights below 0.1 are removed, leaving 310 effective features. After calculating the covariance matrix, the top 20 principal components with the highest contribution rates are selected. The important components retained contain both high-weight components of phrase semantics and high-frequency components of triplet paths, forming the final fused vector. The vector maintains a stable metric when subsequently projected into the preset semantic subspace. Verification shows that the fused vector significantly improves the reliability of semantic matching for novel events in the joint classification of street vending and scene spatial constraints.

[0059] Step S2: Based on the distribution of the event semantic feature set in the preset semantic subspace, calculate the offset of the current batch of events relative to the cluster centers of each semantic subspace, the rate of change of edge sample density, and the number of cross-subspace bridging samples, generating a drift index set. Specifically, this includes: S2.1: Obtain phrase-level embedding vectors and behavior-object-scene triple chain data from the event semantic feature set. Using the predefined semantic subspace mapping rules in the field of urban governance, project the phrase-level embedding vectors onto the coordinate system of specific semantic subspaces such as street vendors, illegal parking, and exposed garbage to generate a set of projected semantic feature vectors with subspace affiliation identifiers.

[0060] For the input event semantic feature set, phrase-level embedding vectors and behavior-object-scene triplet chains are loaded. A specialized subspace mapping rule library for urban governance is invoked. Using semantic keywords, triplet pattern templates, and contextual weights for specific governance scenarios as matching criteria, the phrase embedding vectors are traversed one by one, aligning with the scene nodes of their associated triplet chains. The distribution coordinate indices of the embedding vectors in the semantic space are then parsed. The matched semantic elements are assigned unique attribution identifiers to preset semantic subspaces such as street vending, illegal parking, and exposed garbage. Based on these identifiers, the embedding vectors are reprojected onto the coordinate system of the corresponding subspace, achieving repositioning of the semantic vectors on different subspace coordinate axes. Coordinate normalization is performed on the reprojected semantic feature vectors. Linear scaling and centering eliminate dimensional differences between different subspace coordinate systems, ensuring a unified measurement standard in subsequent cluster center offset calculations. By combining normalized vectors with combined labels for behaviors and objects in scene triplet chains, a composite projected vector data object is constructed, containing spatial identifiers, scene labels, and normalized semantic coordinates. This object is then bound to subspace attribution identifiers to generate a set of projected semantic feature vectors that can directly participate in drift monitoring calculations. Through this mapping and projection processing, the multi-granularity event semantic feature set generated in the previous step is transformed into a projected feature set with subspace attribution attributes, achieving a structured description and computable expression of the semantic distribution state of novel events.

[0061] S2.2: Based on the set of projected semantic feature vectors, the incremental clustering center update algorithm is called to calculate the real-time centroid coordinates of the current batch of events in each semantic subspace, and the real-time centroid coordinates are compared with the historical baseline centroid coordinates by Euclidean distance to generate a subspace clustering center offset sequence that represents the degree of deviation of the core semantic concepts of the events.

[0062] Based on each vector object in the projection semantic feature vector set, it is grouped into corresponding semantic subspace data subsets such as street vendors, illegal parking, and exposed garbage, according to the semantic subspace partitioning strategy predefined in the field of urban governance.

[0063] For each subset of semantic subspace data, an incremental cluster center update algorithm is invoked. The update vector is calculated based on the distance relationship between the newly added vector of the current batch event and the historical cluster center vector. The real-time centroid coordinates of the subspace are updated using a vector weighted average strategy, and the weight coefficients are determined based on the number of samples and the time decay factor.

[0064] Perform Euclidean distance comparison between the updated real-time centroid coordinates and the historical baseline centroid coordinates. Summarize the cluster center offsets of each semantic subspace according to the time window to form an offset sequence. Each element in the offset sequence corresponds to the drift index of the core concept of a semantic subspace.

[0065] Multidimensional encapsulation is performed on the offset sequence to form a structured subspace cluster center offset sequence data object, which serves as a component of the drift index set.

[0066] By comparing cluster centers and processing sequences, the projected semantic feature vector set from the previous step is transformed into a quantifiable deviation index of the core semantic concepts of events, thereby achieving a precise numerical characterization of the distribution shift characteristics of novel event representations.

[0067] S2.3: Using edge sample data located in the neighborhood of the decision boundary of each semantic subspace in the projection semantic feature vector set, perform kernel density estimation processing to obtain the edge probability density distribution value of the current time window, and perform difference calculation between the edge probability density distribution value and the historical edge probability density distribution value of the previous time window to generate an edge sample density change rate index that characterizes the change of the aggregation trend of new events at the classification boundary.

[0068] The set of semantic feature vectors projected onto a specific semantic subspace coordinate system is subjected to decision boundary neighborhood filtering. Edge sample data points located near the discriminative hyperplane of each semantic subspace are extracted to ensure that the data points have a feature distribution state with low classification confidence and distance values ​​from the boundary within a preset neighborhood radius in the high-dimensional semantic space.

[0069] The kernel density estimation algorithm is called on the extracted edge sample data. Based on the Gaussian kernel function commonly used in the field of urban governance, the bandwidth parameter is set as a dynamic bandwidth coefficient that depends on the variance of the historical event distribution. The probability density of the edge sample data distribution within the current time window is estimated to generate edge probability density distribution values.

[0070] The probability distribution of marginal samples is calculated using the following kernel density estimation formula: in, This represents the number of edge samples in the current time window. For bandwidth parameters, For Gaussian kernel function, For the current density point to be estimated, Let be the feature location vector of the i-th edge sample.

[0071] The edge probability density distribution value obtained in the current time window is compared with the historical edge probability density distribution value in the previous time window to form the edge sample density change rate vector. The difference formula is as follows: in, This represents the marginal probability density distribution value of the current time window. This represents the marginal probability density distribution value of the previous time window. The change rate is denoted as .

[0072] Normalization is performed on the rate of change vector to ensure that the rate of change indices of different subspaces are within a uniform numerical range, which facilitates threshold comparison and drift determination.

[0073] Through the kernel density estimation and difference calculation methods described above, the set of projected semantic feature vectors from the previous step is transformed into an index representing the rate of change in edge sample density that characterizes the clustering trend of novel events at the classification boundary, thereby achieving accurate quantification of the dynamics of event clustering across time window boundaries.

[0074] S2.4: For behavior-object-scene triplet chain data, apply cross-subspace semantic association mining logic to identify bridging samples that simultaneously possess two or more different semantic subspace feature attributes, count the number of bridging samples per unit time and normalize it to a relative ratio, and generate a statistical value of the number of cross-subspace bridging samples that represents the degree of semantic boundary blurring between different event types.

[0075] For behavior-object-scene triplet chain data, the cross-subspace semantic association mining module is invoked to load a multi-subspace feature index table in the urban governance field. Based on this index table, semantic subspace attribution labels corresponding to the behavior verb, object noun, and scene location identifier are detected in each triplet chain. Subspace attribution combination judgment is performed on the detection results, and triplet chain instances with two or more different semantic subspace attribution labels are selected and marked as bridging sample candidates. These bridging sample candidates are then imported into feature conflict resolution logic to remove instances that cannot be attributed across subspaces due to mutually exclusive scene constraints, forming a valid bridging sample set. The valid bridging sample set is divided into unit time windows based on the timestamp field. The number of bridging samples within each time window is counted, and a normalization function is used to convert the statistical values ​​into relative ratios. The normalization process uses the following formula: in, For time window Number of internal bridging samples R represents the total number of bridging samples across all time windows within the monitoring period, and R is the normalized relative ratio. The normalized relative ratio is output to the main step as a statistical value representing the degree of semantic blurring between different event types, indicating the number of cross-subspace bridging samples, for subsequent encapsulation of the drift index set.

[0076] By identifying cross-subspace features, resolving conflict conditions, and performing statistical and normalization processing within a unit time window, the structured triplet data from the previous step is transformed into a quantifiable cross-subspace bridging sample quantity index, enabling real-time quantification of the fuzzy trend of event type boundaries.

[0077] S2.5: Summarize the subspace cluster center offset sequence, edge sample density change rate index, and cross-subspace bridging sample quantity statistics. Construct a structured data object according to the preset multi-dimensional index encapsulation protocol to generate a drift index set for comprehensively characterizing the distribution shift characteristics of novel event representations. This serves as the direct input basis for subsequent judgment on whether semantic drift signals are triggered.

[0078] It should be noted that the dynamic threshold is determined using an adaptive calculation method based on the statistical distribution of historical drift index sets. Specifically, the system continuously records the historical numerical sequence of drift indices for each semantic subspace using a sliding time window, and calculates the mean value of each of the three indices—offset, edge sample density change rate, and cross-subspace bridging sample ratio—within the sliding window. with standard deviation and in accordance with Each generates its corresponding dynamic threshold. The sensitivity coefficient is configurable, with a value range of [1.5, 3.0]. The sliding window size and sensitivity coefficient can be adjusted according to the business sensitivity requirements of the urban governance scenario. When the real-time value of a certain drift indicator exceeds its corresponding dynamic threshold, the system determines that a semantic drift signal is triggered.

[0079] like Figure 2 As shown, step S3: Determine whether any indicator in the drift indicator set exceeds the dynamic threshold. If it does, a semantic drift signal is triggered, and based on the signal, the anchor tags closest to the semantic distance of the drift event are located from the tag topology map, generating an anchor tag set. Specifically, this includes: S3.1: Obtain the drift index set and the drift event semantic feature set that triggers the semantic drift signal. Use the predefined vector space mapping rules to project the drift event semantic feature set to the same high-dimensional semantic vector space where the label topology map is located, so as to generate a drift event projection vector object with a comparable measurement basis.

[0080] Upon receiving the drift index set and the corresponding triggered drift event semantic feature set, the multi-dimensional quantification information encapsulated in the drift index set is used as input conditions, and the phrase embedding vectors and behavior-object-scene triplet chains in the drift event semantic feature set are used as core projection objects. Predefined vector space mapping rules are invoked, and based on the basis vector configuration of the high-dimensional semantic coordinate system used by the tag topology graph, a high-dimensional projection transformation is performed on the phrase embedding vectors in the drift event semantic feature set to ensure dimensional consistency and preserve semantic dependency structure. For the behavior-object-scene triplet chain, the triples are decomposed into three vector components according to the rules: predicate, object, and environment element. During the projection process, the coordinate components of the corresponding tag topology graph node weight matrix are bound to achieve integrated structural and semantic mapping. For parameters involving normalization processing in the projection transformation, a high-dimensional cosine normalization method is used. After projection, the phrase embedding vectors and triplet projection vectors are fused at multiple granularities, and a component-level weighted summation operation is performed to generate a single drift event projection vector object. This multi-step mapping and fusion process transforms the semantic feature data of the drift event from the previous step into a high-dimensional projection vector that can be directly compared with the node vectors of the label topology graph, thus establishing a basis for consistent comparison across semantic spaces.

[0081] S3.2: Based on the drift event projection vector object, call the cosine similarity calculation algorithm to traverse all existing tag node vectors in the existing tag topology graph, calculate the cosine value of the angle between the drift event projection vector and each existing tag node vector, and generate a node similarity score sequence.

[0082] Based on the same high-dimensional semantic vector space of the drift event projection vector object and the existing tag node vectors in the tag topology graph, a cosine similarity calculation component is loaded to perform an accurate similarity measurement between the input projection vector and the target node vector.

[0083] Using the drift event projection vector object as a reference vector, the coordinate array form of each global label node vector is extracted and normalized to eliminate the difference in magnitude between different node vectors.

[0084] The core calculation unit of the cosine similarity algorithm is invoked to perform a ratio operation on the dot product of the drift event projection vector and the existing label node vector and the product of their respective norms, using a standardized formula: in Score the node similarity. and These represent the numerical values ​​of each dimension of the drift event projection vector and the existing tag node vector, respectively. For vector dimensions.

[0085] The calculated cosine similarity value is written into the corresponding position field of the node similarity score sequence, and a bidirectional mapping relationship between the node and its unique identifier is retained in the sequence to enhance the usability of subsequent filtering.

[0086] The integrity of the node similarity score sequence is verified, and score entries with missing vectors or calculation anomalies are removed to ensure that the sequence contains only valid measurement results in the subsequent candidate anchor node selection process.

[0087] By using the above processing method, the projection results of drift events are transformed into a node similarity score sequence that represents the semantic distance, thereby realizing a quantitative expression of the node correlation of the tag topology graph.

[0088] For example, in a city governance scenario, the object dimension of the drift event projection vector is 128, and the total number of label nodes in the label topology graph is 500, with each node vector also having 128 dimensions. Normalization is performed using the L2 norm. The dot product value in the cosine similarity calculation is approximately 46.5, the norm of the drift event projection vector is 7.1, and the norm of the target node is 6.8. Applying the formula to calculate the node similarity score yields a value of approximately 0.97, indicating that the node is semantically highly similar to the drift event. A full traversal is performed on the 500 nodes to obtain a score sequence. Nodes with values ​​close to 1 in the sequence are concentrated in the subspaces of street vending and illegal parking. These score results are selected as anchor point candidate nodes in the subsequent threshold truncation operation, significantly improving the accuracy of anchor point positioning in the label self-repair process.

[0089] S3.3: Perform descending sorting and threshold truncation operations based on the node similarity score sequence to select several candidate label nodes with similarity scores higher than the preset neighborhood threshold, so as to generate an initial anchor label candidate set containing potential semantic relationships.

[0090] Given the node similarity score sequence as input, a high-precision descending index structure is established for each score value in the sequence to ensure that high-scoring nodes have priority access attributes during traversal.

[0091] The sorted node similarity score sequence is loaded into the threshold truncation operation module, and the filtering logic based on dynamic neighborhood threshold adjustment is called. The relationship between each score value and the preset neighborhood threshold is compared, and the filtering boundary is adjusted in combination with the category distribution characteristics of the current semantic drift event, so as to retain nodes whose semantic similarity meets the requirements.

[0092] The candidate label node extraction operation is performed on the score subsequence after threshold filtering. The node identifiers in the score sequence are bidirectionally mapped to the label topology index table to generate the candidate label node dataset with the smallest semantic distance to the drift event.

[0093] Embed latent semantic association markers in the candidate label node dataset, and calculate the latent association strength using the local density information of the semantic vector space and the historical association count between label nodes. The association strength value is calculated using the following formula: in, The node similarity score. For local density coefficients, The number of historical associations is used to calculate the potential association strength index for each candidate node.

[0094] By binding the potential association strength index with the candidate label node dataset, an initial anchor label candidate set containing node identifiers, similarity scores, and potential association strength is generated. This processing method transforms the similarity score sequence from the previous step into candidate node data with priority ranking and association strength identifiers, thereby achieving high-precision selection of anchor labels and optimization of semantic drift response.

[0095] For example, in a drift event detection, the node similarity score sequence is 500 in length. After sorting in descending order, the high score range is [0.85, 0.92]. A dynamic neighborhood threshold of 0.84 is set. After threshold filtering, 46 nodes with scores higher than the threshold are retained. The local density coefficient, estimated using semantic space kernel density, has an average value of 1.25, with historical association counts ranging from 5 to 20. For a node with a score of 0.90, a local density coefficient of 1.25, and 8 historical association counts, the association strength value is calculated using the above formula, resulting in 0.1125. This value is significantly higher than the internal validation threshold of 0.10, thus including this node in the initial anchor label candidate set. Performing this filtering in the test scenario can shorten the anchor label drift response time to milliseconds while maintaining high stability, supporting rapid binding and topology diffusion of subsequent multi-source business feedback features.

[0096] S3.4: For each candidate label node in the initial anchor label candidate set, retrieve the historical event handling database, citizen complaint work order database, and department feedback record database respectively, and extract the corresponding historical frequency data of collaborative occurrence, citizen complaint semantic co-occurrence intensity data, and handling department feedback correction record data to generate a multi-source heterogeneous business feedback feature tuple set.

[0097] For each candidate tag node in the initial anchor tag candidate set, a targeted search is performed in the historical event handling database using the index key bound to the node's unique identifier to extract the frequency of co-occurrence of that tag node in existing event handling records. Based on the search results, the frequency sequence of co-occurrence is segmented and aggregated according to timestamps, and the number of occurrences within a unit time window is calculated to form a time series frequency feature, ensuring the timeliness and trend analyzability of the frequency data. Using the same candidate node index key, the citizen complaint work order database is accessed, and full-text semantic parsing is performed on complaint work orders involving that node tag. Keywords semantically related to the node tag are extracted, and the frequency of keyword co-occurrence is counted to construct citizen complaint semantic co-occurrence strength data. A dependency parsing algorithm is called to perform semantic dependency mapping matching between the complaint text and the node tag definition text, eliminating expressions irrelevant to the business, thereby improving the accuracy of the co-occurrence strength index. Access the feedback record database of the handling department through the candidate node index key, retrieve feedback records involving the tagged node, parse the correction opinions, handling adjustment instructions, and tag status change information in the records to form the handling department feedback correction record data, and construct a time series of correction behavior based on the chronological order of the feedback records. Encapsulate the historical frequency data of collaborative occurrences, the semantic co-occurrence intensity data of complaints, and the feedback correction record data into multi-source heterogeneous data by node, forming a multi-source heterogeneous business feedback feature tuple set.

[0098] Through the above multi-source retrieval, parsing and encapsulation processing, the initial anchor tag candidate set selected in the previous step is transformed into a feature tuple set containing quantifiable business-related indicators, which can be used as direct input for subsequent attribute enhancement encapsulation.

[0099] For example, when performing this step for the candidate tag node "street vending," the system retrieved from the historical event handling database that it had appeared in conjunction with events such as "illegal parking" and "mobile vendors" 412 and 289 times respectively in the past 365 days. The monthly average number of occurrences was 35 and 24 respectively. In the citizen complaint work order database, 678 complaint records containing the keywords "street vending" and "business" were matched, of which 534 work orders were semantically consistent with the "street vending" tag definition. The semantic co-occurrence strength value, after normalization, was 0.78. In the handling department feedback record database, 96 handling adjustment instructions containing this tag were retrieved, of which 18 records involved tag definition modification or edge weight adjustment. The resulting correction record time series showed that 12 adjustments occurred in the last 3 months. The above data was encapsulated into a feature tuple: {co-occurrence frequency: [35,24], complaint co-occurrence intensity: 0.78, corrected record time series: [12,3,3]}, which served as the direct input for subsequent attribute enhancement encapsulation. The verification results showed that this feature tuple could significantly improve the accuracy of association weight calculation in the subsequent semantic closeness matrix construction.

[0100] S3.5: The initial anchor tag candidate set is encapsulated with attribute enhancement by using the set of feature tuples from multi-source heterogeneous business feedback. The historical frequency data of co-occurrence, the semantic co-occurrence intensity data of citizen complaints, and the feedback correction record data of the handling department are bound to the corresponding candidate tag nodes to generate an anchor tag set containing complete multi-dimensional weighted information.

[0101] like Figure 3 As shown, step S4: Construct a semantic proximity matrix using the anchor tag set, and perform a limited-step diffusion along the graph edges starting from the anchor nodes to identify the set of edges to be evaluated in the semantically ambiguous zone, generating a list of edges to be repaired. Specifically, this includes: S4.1: Obtain three weighted data items from the anchor tag set: the frequency of historical co-occurrence, the semantic co-occurrence intensity of citizen complaints, and the feedback and correction records from the handling department. Normalize the three weighted data items and perform multi-source heterogeneous feature fusion operation to generate an original semantic closeness matrix that represents the degree of association between nodes.

[0102] After receiving the final anchor tag set, three weighted data items are extracted: historical frequency of co-occurrence, semantic co-occurrence intensity of citizen complaints, and correction records from handling departments. These are then loaded into the multi-source feature processing engine as numerical input fields. A normalization operator is used to scale the three weighted data items by interval ratio, mapping the original frequency, co-occurrence intensity, and correction record count to the closed interval [0,1] to eliminate differences in the units of measurement between data from different business sources. Based on the normalization result, a multi-dimensional feature tensor is constructed. In the tensor matrix, the three data items for each node are stored in a fixed column order to ensure data alignment in subsequent fusion operations. A weighted fusion algorithm is applied, performing a linear combination of the normalized frequency value, co-occurrence intensity value, and correction record value according to a preset feature importance coefficient. The semantic closeness between nodes is calculated using the following formula: in, This represents the semantic intimacy score. , , These are the weight coefficients for the three features. , , These are the normalized historical frequency of co-occurrence, the semantic co-occurrence intensity of citizen complaints, and the feedback correction record value from the handling department, respectively. The above formula is applied to the feature vectors of pairwise anchor tag combinations, iterating through and filling the corresponding positions in the two-dimensional matrix to generate the original semantic closeness numerical matrix. Through this processing method, the multi-source heterogeneous business feedback data from the previous step is transformed into a semantic closeness matrix with a unified structure and scale, which can be directly used for graph weighted modeling, achieving a quantitative representation of the degree of correlation between nodes.

[0103] S4.2: Construct a weighted directed label topology graph structure based on the original semantic closeness numerical matrix, and map the values ​​in the original semantic closeness numerical matrix to the connection weights of the graph edges to generate a weighted label topology graph instance with dynamic semantic association attributes.

[0104] Upon receiving the raw semantic affinity matrix generated by S4.1 as input, the row and column node indices and corresponding values ​​of this matrix are selected as the basic data for constructing the directed label topology graph. According to the matrix's row and column coordinate mapping rules, each row and column index is mapped to a unique node identifier in the anchor label set, ensuring the consistency and traceability of node mapping. A weight mapping function is executed on each non-zero value in the raw semantic affinity matrix to normalize the values ​​to a preset weight range, eliminating dimensional differences between data from different sources. By combining the normalized values ​​with the node mapping relationship, directed edge data records with weight attributes are constructed, and a direction field is added to each edge. The direction is determined based on the association between matrix row nodes and column nodes. Using the graph instantiation interface, node objects and edge objects are sequentially injected into the graph structure memory model according to the topological sorting rules, forming an executable weighted label topology graph instance. Dynamic semantic association attribute binding processing is performed on this instance, establishing association indexes between the weight field and dynamic attributes such as the node's historical drift frequency and boundary ambiguity to support subsequent restricted step-size diffusion algorithms for weighted path selection. Through the above processing method, the result of the previous step is transformed into a weighted label topology graph instance with dynamic semantic association attributes, rigorous structure and support for weight operation, so as to realize the high-precision local subgraph search space construction capability required for the subsequent fuzzy edge recognition process.

[0105] S4.3: Using the anchor labels in the weighted label topology graph instance as the diffusion starting point, execute the restricted step graph traversal algorithm based on the maximum hop count constraint and weight threshold filtering to generate a local subgraph search space covering the neighborhood of the anchor node.

[0106] Initialize the diffusion parameters for the anchor labels in the weighted label topology graph instance, including the maximum hop count constraint value, the minimum edge weight filtering value, and the initialization matrix of the node access mark, to ensure that the traversal range is controlled within the target neighborhood.

[0107] The restricted step-length graph traversal algorithm is invoked based on the set maximum hop count constraint value. Anchor labels are added to the queue to be processed as the traversal starting point, and the current hop count and the weight of the reached path are recorded at each level during the traversal.

[0108] During the traversal, the connection weight of each edge is compared with a preset weight threshold, and weakly related edges with weights less than the threshold are filtered out, thereby ensuring that the semantic association strength of the local search space is not lower than the set standard.

[0109] For edges that meet the weight conditions, update the corresponding node access status and add unvisited adjacent nodes to the queue to be processed. Accumulate the hop count and continue to perform a restricted step size recursive traversal until the maximum hop count constraint value is reached.

[0110] After the traversal is completed, the visited nodes and edges are constructed into a local subgraph search space structure, which includes node semantic feature vectors, edge weight parameters and path lengths between nodes, for use in subsequent semantic boundary ambiguity calculations.

[0111] By using a restricted step-length graph traversal algorithm and weighted threshold filtering, the anchor tag set from the previous step is transformed into a local subgraph search space that covers the neighborhood of the anchor nodes and has semantic association strength, thereby enabling the restricted extraction of potential semantic connection relationships.

[0112] For example, in a city event tag topology graph, the anchor tag set contains two high-weight nodes: "illegal parking" and "street vending." The maximum hop count constraint is set to 3, the minimum edge weight filtering value is 0.65, and the initial values ​​of the node access tag matrix are all 0. During the restricted step traversal, the "illegal parking" node is added to the queue as the starting point. The first hop visits adjacent nodes with a weight greater than or equal to 0.65, such as "temporary parking." The second hop visits nodes adjacent to these nodes with a weight greater than or equal to 0.65, such as "illegal parking near bus stops." The third hop visits nodes that meet the conditions, such as "illegal parking at school gates." The path length and cumulative weight are recorded in each traversal step. The filtering determination for each edge weight uses the following formula: in Scoring edge validity The connection weight of the current edge. The cumulative decay value is used for the traversal path. In this embodiment, the edge with a weight of 0.72 is calculated to have a validity score of 0.72 / (0.72+0.05)=0.935 after a cumulative decay value of 0.05. This significantly improves the semantic relevance of the local search space. The final output local subgraph contains 5 nodes and 6 edges that meet the conditions, providing an accurate and highly relevant search foundation for subsequent fuzziness calculation.

[0113] S4.4: Perform semantic boundary ambiguity calculation on all edges in the local subgraph search space. By comparing the semantic vector angle between the two ends of the edge with the historical drift frequency, filter out the set of edges that are outside the known classification boundary and whose confidence is in the uncertain interval, so as to generate the set of edges to be evaluated in the semantic ambiguity zone.

[0114] For all edges in the local subgraph search space, load the high-dimensional semantic vector representations of the nodes at both ends and the historical drift frequency records. Call the vector operation engine to perform angle calculation processing and output the radian value of the angle between the vectors of the nodes at both ends of each edge as the basis for measuring the difference in boundary distribution.

[0115] The radian value is input into the ambiguity evaluation module and differentially analyzed with a preset semantic consistency threshold to form the original boundary ambiguity index. This differential value is linearly combined with the historical drift frequency according to a weighted coefficient to generate a comprehensive ambiguity score that integrates boundary geometric differences and time evolution frequency.

[0116] A normalization function is used to standardize the comprehensive fuzziness score across intervals to ensure that the differences in dimensions between different edges are eliminated when comparing them. The standardized score is then compared with the confidence threshold of the uncertainty interval in a synchronous manner to select the set of edges whose scores fall within the uncertainty interval.

[0117] For the selected edge set, the classification boundary position determination is performed. Edges located outside the known classification boundary and whose standardized scores meet the uncertainty interval conditions are marked as the state to be evaluated, and edges with built-in confidence above the upper limit threshold are removed, forming the edge set to be evaluated in the semantic ambiguity zone.

[0118] Through the above processing method, the local subgraph search space constructed in the previous step is transformed into a data object containing the edge set to be evaluated in the semantic ambiguity zone, so as to realize the accurate identification of potential semantic connection relationship and the input preparation for subsequent repair target.

[0119] For example, in the scenario of violations of urban appearance regulations, the local subgraph search space includes nodes "exposed garbage" and "temporary construction." The angle between the vectors of the two nodes is calculated to be 0.78 radians, and the historical drift frequency is recorded as 12 times / month. Using weight coefficients of 0.6 and 0.4 to correspond to the difference in radians and the drift frequency, respectively, a linear combination formula is used to obtain a comprehensive fuzziness score of 5.23. Mapping to the interval [0,1] using a normalization function, the standardized result is 0.41, which is compared with the confidence uncertainty interval [0.35,0.55]. Since it meets the condition, the edge is marked as a state to be evaluated. A classification boundary determination is performed, confirming that the edge is outside the known boundaries of the two categories "violations of urban appearance regulations" and "construction site management," and it is ultimately included in the semantic fuzzy band set of edges to be evaluated. This process significantly improves the fuzziness score, and the output accurately identifies potential connection edges that require subsequent semantic validity verification, enhancing the system's ability to capture the semantic boundaries of novel cross-domain events.

[0120] S4.5: Extract the node pair information and context path features of the edge set to be evaluated based on the semantically ambiguous zone, and encapsulate the node pair information and context path features into a structured data object to generate a list of edges to be repaired containing potential semantic connections.

[0121] For the edge set to be evaluated that is in a semantically ambiguous zone, the input conditions include the edge set output by S4.4, its corresponding node pair information, and contextual path features, to ensure that the subsequent edge repair process has a complete semantic association foundation. For each edge in the edge set to be evaluated, node identifier parsing is performed on the node pair information, mapping the node identifier to a unique node ID in the label topology graph, and extracting the semantic vector representations of both sides to ensure data consistency. Contextual path feature extraction is performed on the parsed nodes, encapsulating the shortest path sequence, path weight sequence, and intermediate node label definition text within the local subgraph as a contextual path feature set. The node information and contextual path feature set are structurally bound, mapping the node IDs at both ends of each edge and the corresponding path features to a unified structured data template according to preset field rules, ensuring the synchronous recording of semantic and topological attributes. Indexing is performed on the structured data template, establishing a double-key index based on node ID and path length, so that the corresponding anchor label definition text and drift event sample text can be quickly retrieved in the subsequent S5 step. Through the above-mentioned structured encapsulation and indexing, the set of edges to be evaluated that are in the semantic ambiguity zone is transformed into a list of edges to be repaired that can be directly input into the semantic verification model, thereby achieving a standardized description and efficient retrieval capability for potential semantic connections.

[0122] Step S5: Input the anchor point label definition text corresponding to each edge in the list of edges to be repaired and the drift event examples into the semantic verification model. By comparing the consistency of verb role filling, spatial constraint compatibility, and similarity of processing timeliness trends, an edge validity score sequence is generated. Specifically, this includes: S5.1: Obtain the anchor label definition text and drift event sample text corresponding to each edge to be evaluated in the edge to be repaired list. Perform dependency parsing on the anchor label definition text and drift event sample text to extract the context-sensitive verb role filling structure containing the predicate verb and its argument roles, and generate a verb role filling feature set for subsequent consistency comparison.

[0123] It should be noted that the semantic verification model adopts a multi-dimensional semantic consistency verification architecture, which includes three parallel dimensions: verb role filling consistency verification, spatial constraint compatibility verification, and processing timeliness trend similarity verification. It is achieved through dependency parsing, spatial knowledge graph mapping, and historical processing time sequence fitting. Finally, the edge validity score sequence is output through weighted fusion. It does not rely on external large models or deep neural networks and has lightweight and interpretable business characteristics.

[0124] The input data consists of the anchor point label definition text and drift event sample text corresponding to each edge to be evaluated in the edge to be repaired list. Both types of text are derived from the node context information of the label topological fuzzy edge set.

[0125] The anchor tag definition text is segmented and punctuation is removed to eliminate the interference of non-semantic characters on syntactic parsing.

[0126] Based on the word segmentation results, a customized dependency parser for the field of urban governance is called to identify the grammatical dependency paths between the predicate verb, subject, object and modifiers, and each dependency relationship is marked as a role-verb pair.

[0127] Repeat the above word segmentation and cleaning process for the drift event sample text, and then execute the same dependency parsing algorithm to maintain the structural consistency of the two types of text in the dependency relation vectorization process.

[0128] The dependency paths of anchor tag text and drifting city event text are aligned using node role mapping rules to construct a verb role filling matrix. The matrix cell represents the matching status of a predicate verb with a specific argument role, and the matching status is recorded in a way that combines Boolean values ​​and weights.

[0129] A context-sensitive embedding model is used to calculate the semantic similarity between the anchor tag verb role path and the drift event verb role path, and the path matching degree is quantified by the following formula: in, and These represent the dependency path phrases of the anchor label and the drift event at the i-th character filling position, respectively. A context-sensitive embedding similarity function is used, where n is the total number of matching paths.

[0130] The similarity score is combined with the Boolean matching status to form a verb role filling feature set. Each entry in the feature set contains dependency path text, role type, matching status value and similarity weight.

[0131] By using dependency parsing and semantic similarity calculation, the text information of the topological edges to be repaired obtained in the previous step is transformed into a quantifiable verb role filling feature set, thereby achieving the feature extraction effect of subsequent consistency comparison.

[0132] For example, in the anchor tag definition text for "street vending," the predicate verb "operate" and its object "stall" are extracted, with the subject being "merchant" and the location modifier "sidewalk." In the drift event example text, the predicate verb "sell" and its object "fruit" are extracted, with the subject being "vendor" and the location modifier "under-bridge walkway." In the path alignment table output by the dependency parser, the path similarity of the predicate verb is 0.78, the path similarity of the object is 0.64, the path similarity of the subject is 0.81, the Boolean matching state is 1 for all, and the location modifier matching state is 0. According to the formula, the path matching degree Sv is 0.74. The generated verb role filling feature set contains 4 entries, which are used for the comprehensive judgment of subsequent spatial constraint compatibility and the similarity of the disposal timeliness trend. In this implementation, the feature set significantly improves the stability of the connection validity score in the semantic verification model.

[0133] S5.2: Based on the verb role filling feature set, the pre-set urban governance spatial knowledge graph is called to perform semantic mapping processing on the location adverbs in the drift event sample text, so as to judge the compatibility between the location adverbs and the preset spatial constraints in the anchor label definition text, and generate a spatial constraint compatibility judgment result representing the degree of spatial logical matching.

[0134] Based on the verb role filling feature set obtained from the S5.1 sub-step, the location adverbial structure and its semantic encoding representation contained in the drift event sample text are obtained and used as the spatial location input source. A pre-built urban governance spatial knowledge graph is invoked to map the geographical entity names, directional descriptions, and regional qualifiers in the location adverbials to standardized spatial node identifiers and attributes in the knowledge graph, thereby eliminating the ambiguity and diversity of natural language expressions. Combined with the pre-set spatial constraints in the anchor label definition text, the set of spatial nodes and boundary parameters constrained by these constraints are extracted, forming a structured spatial constraint object. A spatial compatibility calculation process is performed, utilizing the node adjacency matrix and regional hierarchy division rules in the spatial knowledge graph to determine the inclusion, intersection, or conflict relationships between the corresponding nodes of the location adverbials and the spatial constraint nodes of the anchor labels, adjusting the compatibility weights according to the relationship type. A spatial logical matching function is employed. in To score spatial constraint compatibility, To map the spatial hierarchical distance between nodes, the spatial constraint compatibility judgment result is output as the scoring input, providing a spatial logical matching basis for subsequent timeliness trend similarity measurement. Through this processing method, the verb role filling feature set from the previous step is transformed into a quantifiable spatial compatibility index, enabling accurate determination of the degree of spatial logical matching during the tag topology repair process.

[0135] For example, in a city governance scenario, the sample text of a drifting event contains the location adverb "small square under the bridge." Dependency parsing extracts the location adverb sequence and encodes it as spatial entities: node A (under the bridge) and node B (small square). After calling the city governance spatial knowledge graph, node A is mapped to subspace identifier SP_03, and node B is mapped to subspace identifier SP_07, and associated with a geographic level (street level). The anchor tag definition text has a preset spatial constraint condition of "community public area" (node ​​SP_05, street level). The spatial hierarchy distance between nodes SP_07 and SP_05 is calculated to be 2 using an adjacency matrix. Substituting this into the above formula, the spatial constraint compatibility score is 0.333. If the calculation result is above the preset compatibility threshold, the location of the drifting event is determined to be compatible with the anchor tag spatial constraint condition, and it can subsequently participate in the timeliness trend similarity analysis; if it is below the threshold, it is marked as a potential spatial conflict and connection weight decay is performed during the tag repair process. This judgment process can significantly improve the accuracy of spatial semantic matching in this application scenario and avoid misjudgment of classification rules due to the diversity of location descriptions.

[0136] S5.3: Utilize the spatial constraint compatibility judgment results to retrieve past event handling timestamp data in the historical handling record database that are similar to the current drift event example text. Perform time series trend fitting processing on the past event handling timestamp data to calculate the similarity measure value between the current drift event example text and the anchor label definition text in the handling timeliness feedback trend, and generate a handling timeliness feedback trend similarity index.

[0137] The process involves acquiring the spatial constraint compatibility judgment results and the timestamp index data of the drift event sample text. It then calls the time series retrieval interface for similar events in the historical handling record database to filter historical handling records that match in terms of time span, event type, and spatial constraints. The filtered historical handling timestamp sequences are normalized to map the handling durations of different years and quarters to a unified time scale for trend alignment. A sliding window averaging method is used to locally smooth the normalized time series, reducing the interference of high-frequency noise on trend fitting. The average handling duration within each window interval is calculated as the trend sampling point. The least squares method is used to perform linear or polynomial fitting on the smoothed handling duration sampling points to obtain the trend function expression for historical handling timeliness. The time series corresponding to the drift event sample text is similarly normalized and trend-fitted to generate the timeliness trend function for that drift event. Finally, the similarity measurement calculation module is called to construct a trend similarity index based on the sum of squared residuals of the two trend functions, as shown in the following formula: in, This is a trend similarity index. Historical disposal trends in the first The duration value of each sampling point For the drift event trend in the first The duration value of each sampling point This represents the average processing time based on historical trends. The number of sampling points is denoted by . The calculated trend similarity index is output as the trend similarity indicator for handling timeliness feedback. This indicator is then integrated with other effectiveness features to achieve a quantitative evaluation of the side to be evaluated in the dimension of handling timeliness.

[0138] By using normalization, smoothing, trend fitting, and residual measurement methods, the spatial constraint compatibility judgment results from the previous step are transformed into quantifiable similarity indicators of the timeliness feedback trend, thereby enabling the comparability and effectiveness assessment of different events in terms of timeliness characteristics.

[0139] For example, in the drift event sample text of the municipal management platform, the spatial constraint is the "under the bridge" scenario. Illegal business operations under this scenario are retrieved from the historical handling record database, resulting in 128 records within the past 3 years, with handling times ranging from 15 to 180 minutes. The handling time is normalized to a scale of 0-100 minutes by year, and average smoothing is performed using a sliding window width of 5 records to obtain the smoothed handling mean curve. A quadratic polynomial fitting is applied to obtain the historical trend function Th(i) = 0.05i² + 2.3i + 20. The current batch of drift event timeliness data is fitted to obtain the trend function Tp(i) = 0.048i² + 2.5i + 19. The trend similarity index is calculated according to the formula. With a sampling point count n of 25, the residual sum of squares is small and the index value is close to 1, indicating that the drift event highly matches the anchor label definition text in terms of handling timeliness characteristics. The output trend similarity index enhances the connection in the subsequent edge validity scoring, significantly improving the accuracy of label topology reconstruction.

[0140] S5.4: Based on the verb role filling feature set, spatial constraint compatibility judgment results, and similarity index of handling timeliness feedback trend, perform multi-dimensional weight fusion calculation to aggregate the verb role filling consistency score, spatial constraint compatibility score, and handling timeliness feedback trend similarity score into a single-dimensional connection effectiveness quantification value, and generate the original edge effectiveness score that characterizes the reliability of the edge connection to be evaluated.

[0141] Based on the verb role filling feature set, spatial constraint compatibility judgment results, and similarity index of processing timeliness feedback trend, a preset multi-dimensional weight fusion calculation protocol is selected as the execution benchmark.

[0142] The consistency score for verb role filling is normalized to ensure that the score participates in subsequent fusion calculations under a unified dimension.

[0143] The spatial constraint compatibility judgment results are corrected by weighting coefficients and transformed into a compatibility quantification value that can be directly compared with the verb role consistency score.

[0144] The similarity index of the timeliness feedback is smoothed by filtering to remove abnormal time-series fluctuations and ensure the stability of the timeliness measurement.

[0145] The normalized verb role consistency score, the corrected spatial compatibility value, and the smoothed timeliness similarity value are linearly combined according to a preset weight ratio. The fusion result is then subjected to distribution verification to check whether it falls within the valid score range and to remove abnormal edges that do not conform to the range definition.

[0146] Through the multi-dimensional weight fusion process described above, the three evaluation results from the previous step are aggregated into a single-dimensional raw score for edge connection effectiveness, thereby quantifying the reliability of the edge connection to be evaluated.

[0147] S5.5: Based on the original edge validity score, the normalized mapping function is applied to perform interval standardization on the original edge validity score to eliminate the differences in different dimensions and output the standard score that meets the preset threshold comparison requirements, thus generating the final edge validity score sequence that characterizes the edge connection validity.

[0148] Step S6: Perform topological reconstruction based on the scoring sequence to generate an evolved tag topological map. Specifically, this includes: S6.1: Obtain each edge to be evaluated and its corresponding edge validity score sequence from the list of edges to be repaired. Perform binarization judgment processing on the edge validity score sequence based on the preset semantic connection strength threshold to generate an edge classification result set containing a set of low-scoring invalid edges and a set of high-scoring potential edges.

[0149] Based on each edge to be evaluated in the list of edges to be repaired and its corresponding edge validity score sequence, all score data is loaded into an in-memory index structure. A traversal operation is performed on the score sequence, calling the preset semantic connection strength threshold parameter reading module to compare each score value with the threshold one by one. Logical binary mapping is performed on the judgment results, mapping scores higher than the threshold to Boolean true states and scores lower than or equal to the threshold to Boolean false states. The corresponding topological edge identifiers and associated node information are extracted from the Boolean true state index set and the Boolean false state index set, respectively, forming two independent data structures: a high-scoring potential edge candidate set and a low-scoring invalid edge candidate set. Buffer serialization is performed on these two candidate sets to construct an edge classification result set containing the low-scoring invalid edge set and the high-scoring potential edge set. Through the generation of this classification result set, accurate clustering of all edges to be evaluated is achieved, providing a basic classification basis for subsequent soft pruning processing and the injection of temporary semantic bridge nodes.

[0150] S6.2: Based on the set of low-scoring invalid edges in the edge classification result set, extract the connection weight parameters of each low-scoring invalid edge in the original label topology graph, and use a linear decay function to perform gradient down-adjustment calculation on the connection weight parameters to generate a soft-pruned topology edge weight matrix with weakened connection features.

[0151] For data objects identified as low-scoring invalid edges in the edge classification result set, load the corresponding edge attribute records from the original label topology graph instance, and extract the connection weight parameter value of each low-scoring invalid edge in the topology structure through index retrieval.

[0152] The extracted connection weight parameter values ​​are processed by calling the linear decay function calculation interface in the parameter calibration module. The decay coefficient is set based on the difference between the edge validity score and the preset threshold. The larger the difference, the higher the decay coefficient, so that the weight decrease is proportional to the unreliability of the edge.

[0153] When performing the decay operation, the original weight parameters and decay coefficient are input for each low-scoring invalid edge, and gradient downgrading is completed according to the following numerical mapping rules: in, These are the original connection weights. The decayed connection weights, It is a linear attenuation coefficient, and its value ranges between 0 and 1.

[0154] Configure a minimum weight protection threshold during the decay process to ensure that w' is greater than or equal to this protection value, so as to avoid the structural connection failure caused by the edge weight dropping to zero.

[0155] After performing the gradient downscaling calculation on all low-scoring invalid edges, the decayed weight parameter values ​​and corresponding edge index codes are summarized to construct a soft-pruned topology edge weight matrix with weakened connection characteristics. The matrix is ​​stored in the weight data field of the labeled topology graph and used as an input data source for subsequent topology reconstruction.

[0156] By using a linear decay function and gradient downsampling, the edge classification results from the previous step are transformed into soft-pruned topological edge weight matrix data, achieving the expected technical effect of weakening the weights of unreliable connections and optimizing the semantic connection strength in the original graph.

[0157] For example, in a city's event tag topology graph, there are 45 low-scoring invalid edges with original connection weight parameters ranging from 0.35 to 0.62. For each low-scoring invalid edge, the difference between its edge validity score and a preset threshold of 0.5 is calculated. If the difference is 0.25, then λ is set to 0.25. Substituting this into the above formula, when the original connection weight is 0.6, the attenuated connection weight is 0.45, which meets the requirements after verification with a protection threshold of 0.2. After performing gradient downscaling on all low-scoring invalid edges, the generated soft-pruned topology edge weight matrix significantly reduces the overall average weight value. The strength of the original unreliable connections in the graph is effectively weakened. During the verification phase, it was observed that the misclassification trigger frequency of these connections in the event classification logic was significantly reduced, and the classification accuracy was greatly improved without introducing additional computational overhead.

[0158] S6.3: Based on the high-scoring potential edge set in the edge classification result set, identify the anchor label pairs and drift event sample texts that have not established direct connections, and call the temporary node injection algorithm to perform semantic space interpolation processing on the anchor label pairs to generate candidate temporary semantic bridge node instances with the status identifier to be verified.

[0159] Based on the high-scoring potential edge set in the edge classification result set, the input conditions are the node pair information corresponding to the high-scoring potential edges that have passed the binarization judgment and the drift event sample text. The execution object is the anchor label pairs in the label topology graph that have not yet established a direct connection relationship. The high-scoring potential edge set is traversed and processed. For each anchor label pair that has not established a direct connection, its coordinate position in the semantic vector space is retrieved, and the node vector difference calculation module is called to obtain the semantic center vector of the node pair. The drift event sample text is mapped to the same semantic space through a phrase-level embedding model to generate the semantic vector representation of the drift event, and is fed into the temporary node injection algorithm along with the semantic center vector of the node pair. The algorithm performs semantic space interpolation processing, linearly combining the node pair and the drift event semantic vector according to the preset interpolation coefficient β. The calculation formula is as follows: Where β is the interpolation weight coefficient. For the semantic center vector of the node pair, The interpolated vector is used as the core vector representation of the candidate temporary semantic bridge node. It is assigned a state identifier to be verified and its metadata is encapsulated, including the identifier of the connecting end node, the semantic boundary ambiguity score, and timestamp information. Through this processing, the result of the previous step is transformed into a bridge node instance in the semantic space, situated between the anchor label and the drift event, enabling local extension of the label topology and rapid injection of novel event semantic relationships.

[0160] For example, when a city's governance platform processes a high-potential edge set, it retrieves the semantic center vectors of the anchor tags "street vending" and "mobile vehicle sales," which are [0.42, 0.51, 0.33] and [0.48, 0.49, 0.37], respectively. The drift event example text "night market temporary vehicle catering" is mapped to obtain a semantic vector of [0.46, 0.53, 0.36]. The semantic center vector obtained by inter-node vector difference calculation is [0.45, 0.50, 0.35]. The interpolation coefficient β=0.6 is set, and the semantic vector of the temporary bridging node is calculated by substituting it into the formula, resulting in the interpolated vector [0.452, 0.506, 0.354]. This vector serves as the core representation of the candidate temporary semantic bridging node, is marked as pending verification, and the connection end node information and ambiguity score of 0.78 are recorded. In subsequent steps, the platform injected the node into the tag topology map and bound it to an automatic subscription observation task. The verification results showed that the bridging node maintained a significantly improved classification confidence in the new event "night market temporary carriage catering" and related variants, which greatly improved the system's ability to identify cross-domain urban events and the efficiency of tag evolution.

[0161] S6.4: Based on the candidate temporary semantic bridge node instance, construct a task descriptor containing automatic subscription observation task configuration parameters, bind the task descriptor to the candidate temporary semantic bridge node instance and perform a state activation operation to generate a temporary semantic bridge node object to be verified with real-time monitoring capabilities.

[0162] Based on the structured data input from candidate temporary semantic bridge node instances, the task descriptor construction module is invoked to load a parameter template containing configuration parameters for automatically subscribing to observation tasks. The parameter template includes the semantic path identifier of the monitoring target, a set of subscription event types, a monitoring frequency setting, and a trigger response policy priority. The parameter template is bound and mapped to the unique identifier of the candidate temporary semantic bridge node instance, and a parameter injection process is executed to generate a node task descriptor object. The state manager is invoked to set the activation flag for the node task descriptor object, and subscription channel initialization is performed based on the real-time data stream access permissions of the semantic subspace in which the node is located, ensuring that the node can call the real-time semantic drift monitoring service interface of the city event intelligent distribution system. Based on the subscription channel initialization result, the dynamic data stream of the monitoring target semantic path is loaded into the node's memory cache. By verifying the conflict detection results between the trigger response policy priority in the task descriptor and the system's global execution queue, the queuing and execution status updates of the task scheduling module are completed. Through the above binding, injection, activation and initialization processes, the candidate temporary semantic bridge node instances from the previous step are transformed into temporary semantic bridge node objects to be verified that have real-time monitoring capabilities and automatic subscription observation task execution logic, thereby enabling the nodes to instantly capture and process the semantic features of new events during subsequent evolution.

[0163] For example, in a city's urban management scenario, candidate temporary semantic bridge node instances are constructed for the anchor tag pairs "mobile shoe repair stall" and "street vending" in the high-potential edge set. The monitoring target semantic path identifier is set as Path_ID_083, the subscription event type set includes "mobile vending" and "illegal occupation of sidewalks", the monitoring frequency is set to once every 5 seconds, and the trigger response strategy priority is level 2. The above parameters are bound to the node's unique identifier CN_GZ_TMP_20240518 and the activation identifier setting is executed. The subscription channel is initialized and the drift monitoring data stream of the corresponding semantic subspace is loaded into the node buffer. Through the task scheduling module, the monitoring data stream of this node is connected to the global execution queue to ensure that it can receive and analyze similar event characteristics in real time after priority conflict detection. When the node captures a semantic drift signal after performing monitoring, the system calls the drift event semantic feature comparison logic to verify the matching result of the judgment condition corresponding to the node. The output matching confidence value is calculated using the formula. The calculated improvement is significant, among which Output the probability value of the matching event for the node. The node outputs the probability value of a mismatch event. This improvement validates the node's ability to automatically subscribe to observation tasks for real-time capture of novel events and enhances classification accuracy, achieving a significant improvement in classification and matching performance in real-world scenarios.

[0164] S6.5: Based on the topological edge weight matrix after soft pruning and the temporary semantic bridging node object to be verified, the node connection relationship and edge attribute data of the existing label topological graph are fully merged and updated to generate an evolved label topological graph with local reconstruction features and support for flexible semantic extension.

[0165] Step S7: Update the city event classification rule base based on the evolved label topology graph, incorporate temporary semantic bridging nodes into the real-time classification matching logic, and generate updated classification rule data instances for event allocation and collaborative handling. Specifically, this includes: S7.1: Obtain the temporary semantic bridge node objects to be verified and their associated soft-pruned topological edge weight matrices contained in the evolved label topology graph with local reconstruction features. Use the graph structure serialization parsing algorithm to extract the semantic connection paths and weight decay coefficients between nodes, and generate a rule mapping basic dataset containing dynamic topological structure information.

[0166] In the evolved label topology graph with local reconstruction features, the data structure containing the temporary semantic bridging node objects to be verified and the soft-pruned topology edge weight matrix are retrieved as input objects. A graph structure serialization parsing algorithm based on the node index table is executed on this graph instance, mapping the global index and local node connection relationships into traversable path data. The semantic connection chains between the temporary semantic bridging node objects to be verified and each anchor label are extracted from the path information, and the edge weight values ​​and weight decay coefficients along the chains are collected. The weight decay coefficients are calculated in intervals according to a time decay model to generate a numerical sequence characterizing connection stability, which is then bound to the path node sequence to form path stability data. The above path node sequence, connection weight values, weight decay coefficients, and stability data are merged and encapsulated into a rule mapping base dataset to support subsequent classification rule inverse projection and matching logic generation. Through graph structure serialization parsing and path weight decay calculation, the evolved label topology structure from the previous step is transformed into dynamic topology structure information that can be recognized by the classification rule base, realizing the quantitative and regularized representation of the path features of temporary semantic bridging nodes.

[0167] S7.2: Based on the dynamic topological structure information in the rule mapping base dataset, the inverse projection mechanism corresponding to the multi-granularity semantic feature fusion mapping operation is invoked to transform the vector space representation of temporary semantic bridging nodes into structured judgment conditions that can be recognized by the urban event classification rule base, and generate a set of candidate classification rule entries that are compatible with the existing rule engine format.

[0168] Based on the dynamic topological structure information in the rule mapping dataset, the inverse projection mechanism corresponding to the multi-granularity semantic feature fusion mapping operation is invoked to parse the representation of the temporary semantic bridging nodes to be verified in the high-dimensional semantic vector space into core semantic factors constituting the event classification rule judgment conditions. Feature decomposition processing is performed on the vector representation of the temporary bridging nodes to extract phrase-level embedding components, behavior-object-scene triple structure components, and weight decay coefficient components. Each extracted component is mapped according to a preset inverse projection matrix to obtain the corresponding structured semantic units in the rule engine domain. For phrase-level embedding components, the semantic relevance metric between them and the description text of existing classification rule nodes is calculated, and highly relevant components are marked as rule matching trigger conditions. For behavior-object-scene triple components, the action judgment field and object constraint field in the rule are generated based on the predicate role and object attribute in the triple chain. For weight decay coefficient components, the rule execution priority level is determined by a numerical threshold segmentation method, and the rules corresponding to high weight decay coefficients are marked as low priority execution to control system resource allocation. Through the aforementioned inverse projection and component-driven structured construction operations, the full vector representation of temporary bridging nodes is transformed into structured decision conditions that can be directly invoked in the rule base. Node state identifiers and dynamic topological path information are appended to the rule descriptions, forming a set of candidate classification rule entries that adapts to the existing format of the rule engine and can be invoked for real-time matching. Through inverse feature mapping and rule structure reconstruction, the dynamic topological information from the previous step is transformed into a data structure recognizable by the rule engine, enabling the direct usability of temporary semantic bridging nodes in real-time classification.

[0169] For example, in the urban governance scenario, the evolved label topology graph with local reconstruction features includes a temporary semantic bridge node. Its high-dimensional vector representation is decomposed into phrase embedding vectors (length 256), behavior-object-scene triple chains (5 in total), and a weight decay coefficient of 0.35. The phrase embedding vectors are mapped to the word vector space within the rule engine domain via an inverse projection matrix. Cosine similarity is used for relevance calculation, and a similarity greater than 0.75 is marked as a trigger condition. The behavior-object-scene triple chains generate the action judgment field "No Stalls Allowed" and the object constraint field "Sidewalk" through mapping, and are encapsulated in JSON format. The weight decay coefficient is prioritized using a segmentation threshold of 0.5; values ​​below 0.5 are marked as low-priority execution. During the rule entry construction process, the phrase relevance metric is calculated using the following formula: in, This is a temporary bridging node phrase vector. For the text vector of the rule node, This is a vector inner product operation. The calculation result of 0.78 satisfies the triggering condition requirements. The action determination field and the object constraint field are written into the rule base to generate candidate rule entries. During the execution effect verification phase, this entry successfully identified the new event "sidewalk mobile stalls" in real-time classification and matching, and classified it into the "street vending" category, achieving the goal of rapid classification of new events and direct invocation by the rule engine.

[0170] S7.3: Utilize the candidate classification rule entry set and historical existing classification rules to perform conflict detection and priority sorting. By comparing the consistency score of verb role filling and the compatibility score of spatial constraints, dynamically adjust the weight of overlapping rules or perform logical isolation marking on mutually exclusive rules to generate an optimized classification rule sequence that is conflict-free and has priority attributes.

[0171] Based on the structured judgment conditions of the candidate classification rule entry set and the historical existing classification rules, a conflict detection algorithm is invoked to construct a global rule comparison table using the rule index. A one-to-one comparison is performed according to the hash value of the rule triggering condition and the semantic feature key, generating a conflict detection result set containing overlapping rule pairs and mutually exclusive rule pairs. For entries identified as overlapping rule pairs in the conflict detection result set, their verb role filling consistency score and spatial constraint compatibility score are extracted. Multi-dimensional differential analysis is performed, and rule pairs with score differences less than a preset tolerance are grouped into the same category and their weights are dynamically adjusted. The above weight adjustment is implemented according to a weighted average formula, specifically: in, and These are the original weight values ​​for the conflict rule pairs. and Consistency scores are filled in for each verb role. For entries identified as mutually exclusive rule pairs in the conflict detection results, logical isolation is performed based on the spatial constraint compatibility judgment results. This disconnects the rule links of the mutually exclusive pairs and establishes isolation constraint records in the rule base, ensuring that mutually exclusive rules are not triggered simultaneously by events in the same batch. All rule entries after weight adjustment and logical isolation are serialized in descending order of remaining weight values. Rule entries with equal weight values ​​are then arranged in descending order of spatial constraint compatibility scores, generating a conflict-free optimized classification rule sequence with priority attributes. Priority sorting ensures that rule entries with high confidence and high spatial compatibility are matched preferentially in real-time classification inference. Through conflict detection, dynamic weight adjustment, and logical isolation, the candidate classification rules from the previous step and historical existing classification rules are transformed into a conflict-free optimized rule sequence with a priority execution order, achieving a smooth integration of the new event judgment logic with the existing rule system.

[0172] S7.4: Perform a hot-load update operation on the real-time classification matching logic according to the optimized classification rule sequence, inject the judgment conditions corresponding to the temporary semantic bridge node with the status to be verified into the rule execution chain of the online inference engine, and generate a dynamic event classification logic instance that supports the instant capture and classification of new event expressions.

[0173] The optimized classification rule sequence is input into the real-time classification matching logic hot-loading module. The rule execution chain structure parser is invoked to map the branch node positions of the current online inference engine's rule chain, and the mapping result identifies logic slots that allow dynamic expansion. The judgment conditions corresponding to the temporary semantic bridging nodes to be verified are formatted into rule fragments, including three types of fields: semantic condition set, execution priority flag, and matching timeliness parameters. These are then serialized into structured rule description objects consistent with the rule engine interface protocol. Dependency verification is performed on the structured rule description objects, including complementarity detection of verb role features and spatial constraint features in the judgment condition fields, and conflict resolution of dependencies with upstream and downstream nodes in the existing rule chain, ensuring that the injection operation does not disrupt the existing inference logic. The inference engine's rule chain dynamic injection API is invoked to embed the verified structured rule description objects into the mapped logic slots, while simultaneously binding a state tracking flag and trigger counter for temporary node matching, so that verification data can be collected during subsequent matching processes. The rule chain state is refreshed by performing a just-in-time compilation mechanism within the engine. Newly added decision conditions are transformed into executable rule nodes, and a unique identifier for a temporary semantic bridge node is registered in the global index of the rule chain. This allows the inference engine to capture and classify corresponding new event representations as event feature sets flow in. Through this hot-loading update process, the optimized classification rule sequence from the previous step is transformed into a dynamic event classification logic instance that can be executed instantly in the online inference engine, enabling rapid capture and classification of new events.

[0174] S7.5: Based on the dynamic event classification logic instance, receive the semantic feature set of the city events flowing in in the current batch for real-time reasoning verification, calculate the improvement of classification confidence after the introduction of temporary semantic bridging nodes for new events, and activate the full running state of the logic instance if the confidence exceeds the preset acceptance threshold to generate a dynamic classification model that is officially put into use.

[0175] The runtime environment, based on a dynamic event classification logic instance, receives the semantic feature set of urban events flowing in from the current batch as input conditions for real-time inference verification. The event semantic feature set is loaded into the rule execution chain of the online inference engine, and the initial classification confidence quantification value of the new event under the current classification system is obtained by matching the judgment conditions corresponding to temporary semantic bridging nodes. Relying on a vector space matching algorithm, the historical confidence quantification values ​​in the same batch without the introduction of temporary semantic bridging nodes are compared to construct the dual-sequence difference dataset required for calculating the confidence improvement magnitude. The quantitative evaluation is achieved by calling the improvement magnitude calculation formula, as follows: ,in, To introduce the new event classification confidence of the current batch after introducing temporary semantic bridging nodes, The historical correspondence confidence level without the introduction of temporary semantic bridging nodes. The confidence level is increased by a certain margin. After calculating the increase margin, this value is compared with the system's preset acceptance threshold to generate a binary judgment signal. If the judgment signal indicates that the acceptance condition is met, a logic instance state bit toggles, switching the dynamic event classification logic instance from observation mode to full-scale operation mode, and synchronizing it to the real-time matching logic in the classification rule base. Through the above processing method, the result of the previous step is transformed into a quantifiable confidence level increase index, achieving the technical effect of dynamic verification and timely deployment of the ability to quickly identify new events.

[0176] For example, in an event distribution system in a central urban area of ​​a city, real-time inference verification is performed on the semantic feature set of a newly added batch of "under-bridge road repair" events, with a batch size of 150 event records. After loading the judgment conditions of temporary semantic bridging nodes, the dynamic event classification logic instance calculates the batch's average confidence metric value as 0.78 through vector space matching. The average value before the introduction of temporary nodes is 0.62, and the improvement value is calculated as 0.258 using the above formula. Comparing this value with the system's preset acceptance threshold of 0.20, the judgment signal is that the acceptance condition is met, triggering a state bit flip and activating the full-scale operation mode. In the operation mode, this logic instance performs inference matching verification on the subsequent real-time event stream. The recognition response time of new events is significantly shortened, and the classification accuracy remains stable during continuous operation. The final output dynamic classification model is successfully deployed in this batch scenario and has cross-scenario generalization ability.

[0177] This application also provides an AI-based intelligent allocation and collaborative handling system for urban events. This system corresponds to the aforementioned method and can implement all the technical processes described in the method section. The system includes: a semantic feature module, a drift index module, an anchor point label module, a list of edges to be repaired generation module, an edge validity scoring module, a topology reconstruction module, and a classification rule update module.

[0178] The semantic feature module is used to acquire urban event text and attribute data, extract phrase embedding vectors and parse dependency paths to construct triple chains, and generate event semantic feature sets.

[0179] The drift index module is used to calculate the cluster center offset, edge sample density change rate, and number of cross-subspace bridging samples based on the distribution of the event semantic feature set in the preset semantic subspace, and generate a drift index set.

[0180] The anchor tag module is used to determine whether the drift index exceeds the dynamic threshold. After triggering the semantic drift signal, it locates the anchor tag with the closest semantic distance from the tag topology map and generates an anchor tag set by combining historical collaboration frequency, complaint co-occurrence intensity and departmental correction records.

[0181] The edge-to-be-repaired list generation module is used to construct a semantic proximity matrix using the anchor point label set. Starting from the anchor point node, it expands along the graph edge in a restricted manner to identify the set of edges to be evaluated in the semantically ambiguous zone and generate an edge-to-be-repaired list.

[0182] The edge validity scoring module is used to input the anchor point label definition text corresponding to each edge in the list of edges to be repaired and the drift event examples into the semantic verification model, compare the consistency of verb role filling, spatial constraint compatibility and similarity of processing timeliness trend, and generate an edge validity scoring sequence.

[0183] The topology reconstruction module is used to perform weighted soft pruning on low-scoring invalid edges based on the scoring sequence, inject temporary semantic bridge nodes with a "to be verified" state into high-scoring potential edges, and generate an evolved labeled topology graph.

[0184] The classification rule update module is used to update the urban event classification rule base based on the evolved map, incorporate temporary bridging nodes into the real-time classification matching logic, and generate updated classification rule data instances for event allocation and collaborative handling.

[0185] The modules mentioned above are connected in sequence to form a closed-loop management process from semantic awareness, drift detection, topology repair to rule updates. For those skilled in the art, various other corresponding changes and modifications can be made based on the technical solutions and concepts described above, and all such changes and modifications should fall within the protection scope of the claims of this invention.

[0186] Unless otherwise defined, the technical or scientific terms used herein shall have the ordinary meaning as understood by one of ordinary skill in the art to which this application pertains. The terms “first,” “second,” “third,” and similar terms used in this patent application specification and claims do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Similarly, the terms “an” or “a” and similar terms do not indicate a quantity limitation, but rather indicate the presence of at least one. The terms “comprising” or “including” and similar terms mean that the element or object preceding “comprising” or “including” encompasses the element or object listed following “comprising” or “including” and its equivalents, and do not exclude other elements or objects. The “multiple” mentioned in the embodiments of this application refers to two or more. A and / or B indicate three possibilities: A; B; and A and B.

[0187] The above description is merely an exemplary embodiment of this application, but the scope of protection of this application is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in this application, and such modifications or substitutions should all be covered within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for intelligent allocation and collaborative handling of urban events based on artificial intelligence, characterized in that: include: Obtain urban event text and attribute data, extract phrase embedding vectors from the text and parse dependency paths to construct triple chains, and generate event semantic feature sets; Based on the distribution of the event semantic feature set in the preset semantic subspace, the offset of the current batch of events relative to the cluster center of each semantic subspace, the change rate of edge sample density, and the number of cross-subspace bridging samples are calculated to generate a drift index set. Determine whether any indicator in the drift indicator set exceeds a dynamic threshold. If it does, trigger a semantic drift signal and locate several anchor tags that are semantically closest to the drift event from the tag topology map based on the signal, and generate an anchor tag set. A semantic proximity matrix is ​​constructed using the anchor tag set. Starting from the anchor node, a restricted step-size diffusion is performed along the graph edge to identify the set of edges to be evaluated in the semantically ambiguous zone and generate a list of edges to be repaired. The anchor point label definition text corresponding to each edge in the list of edges to be repaired and the drift event examples are input into the semantic verification model. By comparing the consistency of verb role filling, spatial constraint compatibility and the similarity of processing timeliness trend, an edge validity score sequence is generated. Perform topological reconstruction based on the scoring sequence to generate an evolved label topological map; The city event classification rule base is updated based on the evolved tag topology map. Temporary semantic bridging nodes are incorporated into the real-time classification matching logic to generate updated classification rule data instances for event allocation and collaborative handling.

2. The method for intelligent allocation and collaborative handling of urban events based on artificial intelligence according to claim 1, characterized in that, The process of acquiring urban event text and attribute data, extracting phrase embedding vectors from the text, parsing dependency paths to construct triple chains, and generating an event semantic feature set includes: The text of urban events and attribute data including time, location and subject attributes are obtained from the AI ​​recognition module in the intelligent dispatching scenario of urban events. The text of urban events is preprocessed by word segmentation, cleaning and stop word filtering to generate a standardized corpus of urban event texts. Based on the standardized urban event text corpus, a pre-trained phrase embedding model for urban governance is invoked to perform local context window sliding calculation in order to extract a set of phrase-level embedding vectors containing semantic information of specific governance scenarios. The standardized urban event text corpus is used to execute a dependency parsing algorithm to identify the grammatical dependency paths between predicate verbs and their argument components in the text, so as to construct behavior-object-scene triplet chain data that reflects the internal logical structure of the event. Based on the phrase-level embedded vector set and the behavior-object-scene triplet chain data, a multi-granularity semantic feature fusion mapping operation is performed to align and encapsulate the vector space representation with the structured logical relationship, thereby generating the event semantic feature set.

3. The method for intelligent allocation and collaborative handling of urban events based on artificial intelligence according to claim 1, characterized in that, The process involves determining whether any indicator in the drift indicator set exceeds a dynamic threshold. If it does, a semantic drift signal is triggered. Based on the signal, several anchor tags closest to the semantic distance of the drift event are located from the tag topology map, generating an anchor tag set, including: Obtain the drift index set and the drift event semantic feature set that triggers the semantic drift signal. Use the predefined vector space mapping rules to project the drift event semantic feature set to the same high-dimensional semantic vector space where the tag topology map is located, so as to generate a drift event projection vector object with a comparable measurement basis. Based on the drift event projection vector object, the cosine similarity calculation algorithm is called to traverse all existing tag node vectors in the existing tag topology graph, calculate the cosine value of the angle between the drift event projection vector and each existing tag node vector, and generate a node similarity score sequence. Based on the node similarity score sequence, a descending sort and threshold truncation operation is performed to filter out a number of candidate label nodes with similarity scores higher than a preset neighborhood threshold, so as to generate an initial anchor label candidate set containing potential semantic relationships. For each candidate label node in the initial anchor label candidate set, the historical event handling database, citizen complaint work order database and department feedback record database are retrieved respectively to extract the corresponding historical frequency data of co-occurrence, citizen complaint semantic co-occurrence intensity data and handling department feedback correction record data to generate a multi-source heterogeneous business feedback feature tuple set. The initial anchor tag candidate set is encapsulated with attribute enhancement using the multi-source heterogeneous business feedback feature tuple set. Historical co-occurrence frequency data, citizen complaint semantic co-occurrence intensity data, and handling department feedback correction record data are bound to the corresponding candidate tag nodes to generate the anchor tag set containing complete multi-dimensional weighted information.

4. The method for intelligent allocation and collaborative handling of urban events based on artificial intelligence according to claim 1, characterized in that, The generated event semantic feature set includes: using phrase embedding vectors and semantic alignment weights of each node in the triple chain, normalizing and then concatenating the features, and using feature sparsity and principal component dimensionality reduction to generate a fusion feature vector that reflects the semantics and logical relationships of the text.

5. The method for intelligent allocation and collaborative handling of urban events based on artificial intelligence according to claim 1, characterized in that, The method for determining the distribution of the event semantic feature set in the preset semantic subspace includes: firstly, projecting phrase-level vectors and triple chains to multiple semantic subspaces including street vendors, illegal parking, and exposed garbage through domain keywords, triple pattern templates, and context weights, and marking their affiliation; then uniformly normalizing the vectors in each semantic subspace.

6. The method for intelligent allocation and collaborative handling of urban events based on artificial intelligence according to claim 1, characterized in that, The drift index set consists of the centroid coordinate offset (Euclidean distance) of each subspace, the kernel density estimation change rate of edge samples in the neighborhood of the decision boundary, and the bridging sample ratio across two or more subspaces. The three indices are continuously monitored, and semantic drift is determined if any one of them exceeds the threshold.

7. The method for intelligent allocation and collaborative handling of urban events based on artificial intelligence according to claim 1, characterized in that, The retrieval of the anchor tag set includes performing high-dimensional semantic space mapping and cosine similarity traversal in the tag topology graph, filtering out nodes with similarity higher than a certain threshold, and weighting them according to the node's historical collaboration frequency, complaint co-occurrence intensity, and departmental feedback.

8. The method for intelligent allocation and collaborative handling of urban events based on artificial intelligence according to claim 1, characterized in that, Historical collaboration frequency data of nodes is obtained by aggregating the occurrence of existing events by time window. The co-occurrence intensity of complaints is generated by semantic parsing of complaint work order content and dependency syntax matching, and the temporal behavior of the handling department's correction records is extracted simultaneously.

9. The method for intelligent allocation and collaborative handling of urban events based on artificial intelligence according to claim 1, characterized in that, The multi-source heterogeneous fusion of the semantic intimacy matrix adopts normalization and feature weighted fusion. Each anchor node combines the cooperative frequency, co-occurrence strength and correction record to output a weighted directed topological matrix. All edges are normalized and used as subsequent graph edge weights.

10. An AI-based intelligent allocation and collaborative handling system for urban incidents, including: The semantic feature module is used to acquire urban event text and attribute data, extract phrase embedding vectors from the text and parse dependency paths to construct triple chains, and generate event semantic feature sets. The drift index module is used to calculate the offset of the current batch of events relative to the cluster centers of each semantic subspace, the rate of change of edge sample density, and the number of cross-subspace bridging samples based on the distribution state of the event semantic feature set in the preset semantic subspace, and to generate a drift index set. Anchor tag module is used to determine whether any index in the drift index set exceeds a dynamic threshold. If it does, a semantic drift signal is triggered, and the anchor tag module is used to locate the anchor tags that are closest to the semantic distance of the drift event from the tag topology map based on the signal, and generate an anchor tag set. The edge-to-be-repaired list generation module is used to construct a semantic proximity matrix using the anchor tag set, perform limited step-size diffusion along the graph edge starting from the anchor node, identify the set of edges to be evaluated in the semantically ambiguous zone, and generate an edge-to-be-repaired list. The edge validity scoring module is used to input the anchor point label definition text corresponding to each edge in the list of edges to be repaired and the drift event examples into the semantic verification model. By comparing the consistency of verb role filling, spatial constraint compatibility and similarity of processing timeliness trend, an edge validity scoring sequence is generated. The topology reconstruction module is used to perform topology reconstruction based on the scoring sequence to generate an evolved label topology map; The classification rule update module is used to update the city event classification rule library based on the evolved label topology map, incorporate temporary semantic bridging nodes into the real-time classification matching logic, and generate updated classification rule data instances for event allocation and collaborative handling.

Images