A multi-channel public opinion appeal intelligent gathering and classification method

Abstract

The present application relates to a kind of multi-channel public opinion appeal intelligent convergence and classification method, to realize the quasi-dynamic clustering of similar appeal in large-scale historical work order data and intelligent attribution.The core scheme includes: obtaining nearly three months full amount of closed loop work order, screening high recurrence appeal instance, and generating standardized expression and vector anchor point by special word segmentation and light weight semantic encoder;Dynamic active anchor point is generated by combining channel distribution, time attenuation and stability parameters;After the coding of new work order text, the semantic gravity intensity is calculated by weighting with anchor point, and the sorting and construction of complex potential field are realized, the landing point determination and attribution disambiguation of new work order text are realized;Free node is upgraded according to homologous similar work order semantic anchor point, and anchor point topology is dynamically reorganized.The scheme improves the accuracy, timeliness and system adaptive capacity of work order attribution determination.

Description

Technical Field

[0001] This invention relates to the field of semantic processing and intelligent clustering technology for government documents, and in particular to a method for intelligent aggregation and classification of public opinions from multiple channels. Background Technology

[0002] Current intelligent government management and public service complaint handling platforms widely utilize text semantic processing and information aggregation technologies to achieve efficient classification, duplicate identification, and merging of work orders from multiple channels (such as telephone, WeChat, and APP), and assist government departments in targeted batch processing through clustering methods. Mainstream technical solutions generally employ methods from Natural Language Processing (NLP) such as semantic similarity calculation, knowledge graph-assisted classification, rule engine mapping, static topic clustering, hash indexing, multimodal data fusion, and comparative learning for work order clustering. These solutions largely rely on the following technical paths: First, calculating text similarity based on word vectors or sentence vectors (such as pre-trained models like word2vec and BERT), setting pre-determined clustering thresholds, and grouping similar work orders into the same category; second, using graph-structured databases to manage keywords and entities as nodes, relying on fixed semantic connection boundaries to unify cross-channel work orders; and third, using manually set topic terms or knowledge tags for rule-based classification, supplemented by batch text summarization to improve ambiguity resolution capabilities.

[0003] The aforementioned methods are practical for scenarios with single channels, standardized expressions, or slow-paced changes in business nodes. However, in actual government operations, facing the realities of multiple channels, heterogeneous sources, and highly inconsistent expressions, their technical limitations become increasingly apparent. First, fixed thresholds or static graph structures struggle to accommodate cross-channel semantic drift in practical applications: the same event often exhibits numerous variations across different channels due to dialects, colloquial expressions, and language styles, leading to drastic fluctuations in similarity calculations based on static semantic models and causing "synonym separation" or "mis-aggregation" problems. Second, static rules and standardized topics fail to capture the dynamic evolution of real-time social hotspots and high-frequency demands, resulting in an outdated knowledge base unable to respond to emerging issues. Furthermore, while simultaneous multimodal fusion can improve generalization capabilities, it demands high computational power, security, and data consistency, making large-scale deployment difficult; while comparative learning requires extensive manual annotation and continuous model training, increasing deployment and maintenance costs.

[0004] In real-world government service requests, the most challenging issue in dynamic clustering is semantic drift: the cluster centers (semantic anchors) gradually become semantically loosened or shifted as a large influx of new service request texts arrives. Requests originally belonging to a specific public event may be "misclassified" or lose their affiliation due to minor changes in expression, thus disrupting the "semantic consistency" of request aggregation. This phenomenon is particularly severe in situations involving cross-channel communication, switching between spoken and written language, prevalence of regional dialects, or the emergence of new hot topics, ultimately affecting the accurate depiction of the public's true needs and the response efficiency of government departments. Furthermore, existing solutions based on graph structures, topic models, or classic similarity algorithms all rely on a preset global cluster size or a fixed similarity threshold. In government service scenarios with highly adaptive data distribution and time sensitivity, this often fails to meet the needs of flexible evolution. Once new requests or hot terms appear, the model is prone to fluctuations, and even the overall knowledge structure may collapse.

[0005] In the current context, there is an urgent need to develop an intelligent clustering technology capable of adaptive navigation, automatically discovering and dynamically maintaining stable semantic anchors, efficiently capturing emerging demand types, and responding to changes in the diversity of expressions. This technology must be able to avoid the inherent defects of fixed thresholds and static cluster centers during dynamic clustering, fundamentally alleviating the problems of clustering deviation and knowledge degradation caused by semantic drift. Simultaneously, it is necessary to introduce autonomous growth and degradation mechanisms, enabling clustering results to evolve in real time with government business flows, ensuring semantic consistency of clustering and efficiency of government feedback in cross-channel, multi-source, and time-varying scenarios.

[0006] This patent addresses the aforementioned key technical challenges by introducing a cross-channel adaptive clustering strategy based on dynamic semantic anchor point migration. It aims to achieve the following goals: abandoning traditional fixed thresholds or static graph models, it uses high-frequency co-occurring appeal events as semantic anchors to capture the semantic attraction between new work orders and historical anchors in real time. This automatically drives the updating of anchor weights and timeliness, and uses the gradient of the gravitational field to achieve elastic boundary classification and cluster calibration, forming a dynamic semantic anchor point ecosystem that can adaptively grow and shrink. This innovative mechanism will fundamentally improve the semantic robustness of cross-channel appeal aggregation and the multi-scenario adaptability of government governance practices, effectively solving the persistent problems of semantic drift and misclassification that are difficult to overcome with existing technologies. It brings significant technological advancements to intelligent government work order merging and response. Summary of the Invention

[0007] This application provides a method for intelligent aggregation and classification of public opinion demands from multiple channels, aiming to solve one of the problems or issues of the existing technology mentioned in the background section.

[0008] This application provides a method for intelligent aggregation and classification of public opinions from multiple channels, specifically including:

[0009] S1: Obtain all closed-loop work order data from the public opinion appeal system for the past three months, filter out appeal instances that meet preset conditions in terms of repetition rate, handling consistency rate, and cross-channel recurrence frequency, and generate an initial set of semantic anchor points.

[0010] S2: Based on the consistency of attention distribution, natural day time span and channel matching frequency of each instance in the initial semantic anchor set in different channel texts, calculate and associate semantic stability score, time decay factor and channel adaptation weight to generate active anchor set;

[0011] S3: Obtain the newly added work order text and perform feature mapping processing through a preset semantic encoder to generate a three-dimensional spatial vector;

[0012] S4: Perform a weighted response function operation on the three-dimensional spatial vector and the dynamic attributes in the set of active anchor points in sequence to calculate the semantic attraction strength value that characterizes the association strength between the new work order and each anchor point;

[0013] S5: Select the top five anchor points according to the sorting results of the semantic gravity intensity values ​​to form a temporary gravity candidate region dataset, and set the influence radius according to the reverse order of gravity intensity with each candidate anchor point as the virtual center, and superimpose to generate a composite gravitational potential energy field for defining the belonging range.

[0014] S6: Determine the landing position of the three-dimensional spatial vector in the composite gravitational potential energy field. If the three-dimensional spatial vector falls into the potential energy valley, it is directly assigned to the corresponding anchor point cluster. If the three-dimensional spatial vector is in the transition zone of multiple potential energy overlaps, perform dependency path offset analysis to generate the target semantic anchor point to be finally assigned.

[0015] S7: For free nodes that do not fall into any of the temporary gravity candidate region datasets, monitor the number of similar work orders from the same source. If the number of similar work orders from the same source reaches the preset upgrade threshold within the preset time window, mark the free node as a new semantic anchor and trigger the anchor topology reorganization instruction.

[0016] S8: In response to the anchor point topology reorganization instruction, based on the semantic stability scores of the three anchor points in the original active anchor point set that are closest to the new semantic anchor point, the self-vector positions of the three anchor points are adjusted by weighting, and an updated set of active anchor points after local recalibration is generated.

[0017] This application provides a method for intelligent aggregation and classification of public opinions from multiple channels, which has the following beneficial effects:

[0018] (1) To address the common problems in existing government work order classification systems, such as static semantic modeling, rigid cluster boundaries, and weak adaptability to cross-channel expression variations, this solution significantly improves the robustness and continuity of intent recognition in complex request scenarios by constructing a semantic anchor point ecosystem with dynamic evolution capabilities. Traditional methods rely on preset categories or global similarity thresholds, which are prone to mismatches or missed clustering when faced with the same request drifting in different channels due to differences in language habits. The "anchor point gravity assessment" mechanism proposed in this invention abandons the coarse-grained full comparison mode and instead introduces a weighted response function that integrates timeliness decay factors, channel adaptation weights, and source matching degrees, enabling new work orders to achieve accurate attribution based on local semantic potential fields. Especially in multi-source heterogeneous media environments, this mechanism can effectively identify and stably map highly repetitive requests that have consistent core semantics but different surface expressions, greatly reducing classification oscillations caused by fluctuations in expression forms and significantly improving the system's continuous consistency performance in real business flows.

[0019] (2) Furthermore, to overcome the technical bottleneck of traditional clustering algorithms being unable to cope with the emergence of unknown categories and requiring frequent manual intervention to reconstruct the model structure, this solution innovatively designs a self-growth mechanism based on incremental learning and semantic collaborative calibration, realizing the closed-loop evolution capability of the knowledge system. For "free nodes" that fail to match existing anchors, the system does not simply mark them as noise or anomalies, but instead initiates correlation tracking within a time window: when similar expressions form a minimum-scale cluster within 24 hours, a new anchor upgrade process is automatically triggered, and neighboring anchors are simultaneously driven to fine-tune their positions in the vector space, simulating a process similar to "concept collaborative correction" in cognitive systems. This mechanism not only avoids the risk of overall topological collapse caused by the intrusion of a single new sample, but also enables the semantic space to have the ability to sustainably expand and locally reconstruct, fundamentally eliminating the dependence on a fixed number of clusters and prior distribution assumptions. At the same time, through the built-in anchor health monitoring and automatic freezing strategy, long-term dormant or unstable semantic units will be archived in an orderly manner, ensuring that the entire knowledge graph is always in a dynamic and fresh state, significantly enhancing the long-term reliability and maintenance efficiency of the system.

[0020] (3) In addition, while improving the accuracy of semantic matching, this solution fully considers computational efficiency and engineering feasibility. It abandons the reliance on high-cost technical paths such as graph databases, multimodal fusion coding, and contrastive learning training, and adopts a method that combines semantic coding and local density perception to effectively control resource consumption while ensuring response sensitivity. New work orders only need to be finely compared with the previously selected Top-5 active anchors, which greatly compresses the search space for real-time reasoning. The "semantic tension detection" mechanism based on dependency parsing serves as a key fallback logic to ensure that reasonable decisions can still be made when the potential energy of multiple candidate anchors overlaps, thereby improving the system's discriminative credibility in fuzzy boundary cases. The overall architecture does not require complex parameter tuning. All adjustment factors are derived from real disposal feedback closed-loop data, realizing the resonance between model evolution and governance practice. It has strong interpretability and good portability, and is particularly suitable for urban governance service scenarios with high-frequency iteration, dense demands, and diverse channels.

[0021] In summary, this solution, by constructing a scalable, regressive, and collaborative semantic anchor ecosystem, achieves refined modeling of the binary tension relationship between "stable demands and dynamic expressions" in the government sector. This not only significantly improves the accuracy and stability of cross-channel work order collection but also endows the system with the ability to continuously learn and self-optimize. Without increasing the operational burden, it comprehensively enhances the functional foundation of upper-layer applications such as intelligent allocation, trend early warning, and knowledge accumulation, providing solid support for building a highly resilient and adaptive urban operation perception network. Attached Figure Description

[0022] Figure 1 This is the main flowchart of a method for intelligently gathering and classifying public opinions from multiple channels;

[0023] Figure 2 This is a sub-flowchart of a method for intelligently aggregating and classifying public opinions from multiple channels;

[0024] Figure 3 This is another sub-flowchart of a method for intelligently gathering and classifying public opinions from multiple channels. Detailed Implementation

[0025] Embodiments of the present invention are described in detail below, examples of which 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 for intelligent aggregation and classification of public opinion demands from multiple channels, specifically including:

[0028] S1: Obtain all closed-loop work order data from the public opinion complaint system for the past three months, filter out complaint instances that meet preset conditions in terms of repetition rate, handling consistency rate, and cross-channel recurrence frequency, and generate an initial semantic anchor point set. The initial semantic anchor point set includes, but is not limited to, standardized expressions and their core semantic vectors.

[0029] S2: Based on the consistency of attention distribution, natural day time span and channel matching frequency of each instance in the initial semantic anchor set in different channel texts, calculate and associate semantic stability score, time decay factor and channel adaptation weight to generate an active anchor set with dynamic attributes.

[0030] S3: Obtain the newly added work order text and perform feature mapping processing through a preset semantic encoder to generate a three-dimensional spatial vector representing the semantic features of the newly added work order text, which serves as the input object for subsequent gravity assessment.

[0031] S4: Perform a weighted response function operation on the three-dimensional spatial vector and the dynamic attributes in the set of active anchor points in sequence to calculate the semantic attraction strength value that characterizes the association strength between the new work order and each anchor point.

[0032] S5: Based on the sorting results of the semantic gravity intensity values, select the top five anchor points to form a temporary gravity candidate region dataset, and set the influence radius according to the reverse order of gravity intensity with each candidate anchor point as the virtual center, and superimpose to generate a composite gravitational potential energy field for defining the belonging range.

[0033] S6: Determine the landing position of the three-dimensional spatial vector in the composite gravitational potential energy field. If the three-dimensional spatial vector falls into the potential energy valley, it is directly assigned to the corresponding anchor point cluster. If the three-dimensional spatial vector is in the transition zone of multiple potential energy overlaps, perform dependency path offset analysis to generate the target semantic anchor point to be finally assigned.

[0034] S7: For free nodes that do not fall into any of the temporary gravity candidate region datasets, monitor the number of similar work orders from the same source. If the number of similar work orders from the same source reaches the preset upgrade threshold within the preset time window, mark the free node as a new semantic anchor and trigger the anchor topology reorganization instruction.

[0035] S8: In response to the anchor point topology reorganization instruction, based on the semantic stability scores of the three anchor points in the original active anchor point set that are closest to the new semantic anchor point, the self-vector positions of the three anchor points are adjusted by weighting, and an updated set of active anchor points after local recalibration is generated.

[0036] Step S1: Obtain all closed-loop work order data from the channel public opinion appeal system for the past three months, filter out appeal instances that meet preset conditions in terms of repetition rate, handling consistency rate, and cross-channel recurrence frequency, and generate an initial semantic anchor point set containing standardized expressions and their core semantic vectors. Specifically, this includes:

[0037] S1.1: Extract and process all closed-loop work order data from the past three months to obtain a complete historical work order dataset containing the original request text, handling result tags, and channel source identifiers, which will serve as the input for subsequent filtering operations.

[0038] Obtain the original data storage area of ​​all closed-loop work orders for the past three months, call the government business data access interface to retrieve all work order records in the closed-loop state within a specified time window, and build an initial search list containing the unique identifier of the work order based on the returned result set.

[0039] For each work order record in the initial search list, a field reading operation is performed to parse and extract the original request text field as the basic input for subsequent semantic processing. At the same time, the disposal result label field is extracted for subsequent consistency rate calculation, and the channel source identifier field is extracted for cross-channel statistics.

[0040] During field parsing, the original request text is validated for encoding format and converted for character set. Non-uniform encoding formats in multi-source inputs are converted to uniform UTF-8 encoding to ensure consistency in text processing. The validity of enumeration values ​​is validated for the disposal result labels to remove records with empty values ​​or illegal labels. The matching rule is validated for the channel source identifier to ensure that the channel enumeration values ​​are within the preset business channel list.

[0041] After completing field extraction and validation, the original request text, handling result label, and channel source identifier are indexed and bound according to the unique identifier of the work order, and a full historical work order dataset with complete three-field information is constructed to provide structured input for subsequent multi-dimensional statistical calculations.

[0042] Through the above data retrieval, field extraction, format standardization, and legality verification processes, the results of the previous step are transformed into a full historical work order dataset containing the original request text, handling result tags, and channel source identifiers, thus achieving stable and usable input data for subsequent filtering.

[0043] S1.2: Perform multi-dimensional statistical calculation processing based on the full historical work order dataset, and calculate the recurrence rate, consistency rate of handling results, and cross-channel recurrence frequency index of each work order instance based on frequency statistical processing, and generate a work order feature enhancement dataset containing the above three quantitative indicators.

[0044] Based on each work order instance in the full historical work order dataset, the frequency statistics algorithm module is called to perform a counting operation on the cumulative number of times the same request text appears in the dataset. The cumulative number is divided by the total number of all historical work orders to generate the original value of the repetition rate, forming a basic indicator representing the repetition of work orders within the time window.

[0045] For the handling result label field of each work order instance in the full historical work order dataset, a handling consistency statistical matrix is ​​constructed. All handling results are aggregated according to the request text and the proportion of times that are consistent with the majority of results is counted. The handling result consistency rate numerical sequence is generated by calculating the proportion, which serves as a quantitative basis for measuring the consistency of multiple requests in the dimension of handling results.

[0046] For the channel source identifier field of each work order instance in the full historical work order dataset, cross-channel reproduction calculation processing is performed. The number of times the work order instance appears in different channels is counted and deduplicated by channel category to obtain the original value of cross-channel reproduction frequency. The cross-channel reproduction frequency index is generated by combining the total number of channels.

[0047] A multi-dimensional feature binding mechanism is adopted to incorporate the recurrence rate, consistency rate of handling results, and cross-channel recurrence frequency into the feature record corresponding to the work order instance in the order of fields, thereby generating a work order feature enhancement dataset containing three quantitative feature fields.

[0048] Through multi-dimensional statistical calculations, the full historical work order dataset from the previous step is transformed into a work order feature-enhanced dataset with recurrence rate, consistency rate of handling results, and cross-channel recurrence frequency, thereby enabling the construction of quantitative inputs for subsequent threshold filtering.

[0049] S1.3: Perform threshold filtering based on the recurrence rate, consistency rate of handling results, and frequency of cross-channel recurrence in the work order feature enhancement dataset to select a set of target request instances that simultaneously meet the preset high recurrence rate condition, high handling consistency rate condition, and cross-channel recurrence frequency greater than or equal to three times.

[0050] Based on the generated work order feature enhancement dataset, we receive multi-dimensional feature records containing three quantitative indicators: recurrence rate, consistency rate of handling results, and frequency of cross-channel recurrence, as the input object for this step.

[0051] The repetition rate indicator is processed by threshold determination. The comparison operation module is called to compare the repetition rate of each work order instance with the preset high repetition rate condition value one by one, and a Boolean judgment sequence is generated to identify the valid work orders that meet the repetition rate condition.

[0052] The consistency rate of handling results is subjected to the same-type screening process. The absolute difference calculation method is used to calculate the difference between the consistency rate of each work order and the preset consistency rate condition. The work order instances with a consistency rate lower than the condition value are removed by Boolean judgment to obtain the intermediate work order set that meets the consistency rate condition.

[0053] Frequency filtering is performed on the cross-channel recurrence frequency index. The cross-channel recurrence frequency value of each work order is compared with the integer threshold 3 using the greater than or equal to operator to generate a set of work orders that meet the cross-channel recurrence frequency threshold.

[0054] Perform a logical AND operation on the above three sets of Boolean judgment sequences, and retain the work order instances that are marked as valid in the repetition rate judgment sequence, the processing consistency rate judgment sequence, and the cross-channel reproduction judgment sequence to generate a set of target appeal instances.

[0055] By using Boolean logic AND operations and conditional filtering, the work order feature enhancement dataset from the previous step is transformed into a set of target requirement instances that meet the conditions of high repetition rate, high consistency rate, and cross-channel recurrence frequency greater than or equal to three times, thereby achieving accurate screening of clean data samples that meet the requirements of business stability and cross-channel consistency.

[0056] S1.4: Perform standardized mapping processing on the original request text in the set of target request instances, and use a special word segmenter for government affairs and semantic normalization rules to transform colloquial and dialectal expressions into standard written expressions, generating a standardized expression sequence of target requests with a unified format.

[0057] S1.5: Based on the standardized expression sequence of the target claim, perform vectorization encoding processing, call the semantic encoder to map each standardized expression into a core semantic vector of fixed dimension, and combine them to generate an initial set of semantic anchor points containing the standardized expression and its corresponding core semantic vector.

[0058] Step S2: Based on the consistency of attention distribution, natural day time span, and channel matching frequency ratio of each instance in the initial semantic anchor set across different channel texts, calculate and correlate the semantic stability score, time-effect decay factor, and channel adaptation weight to generate an active anchor set with dynamic attributes. Specifically, this includes:

[0059] S2.1: Obtain the attention distribution matrix of each instance in the initial semantic anchor set in different channel texts, perform variance analysis and entropy measurement on the attention distribution matrix in different channel texts, and generate the original data of semantic stability score representing the semantic expression robustness of the instance.

[0060] It receives an input data object from the initial set of semantic anchors, which contains a sequence of attention weight distribution vectors and their index identifiers for each anchor instance in different channels.

[0061] For the above input data objects, the attention distribution vector of each instance is reconstructed into a matrix according to the channel dimension. The multi-source attention data is arranged into a two-dimensional matrix with rows corresponding to channels and columns corresponding to word positions to ensure the structural consistency of subsequent statistical feature extraction.

[0062] The cross-channel attention consistency calculation module is invoked to perform variance analysis on the multi-source two-dimensional attention matrix of each instance, calculate the variance value of the corresponding column vector of each channel and take the mean of all columns to form a variance index vector that reflects the stability of cross-channel attention.

[0063] Based on the variance index vector, after normalizing the attention weight matrix of each instance, the information entropy value is calculated by column vector and the mean of all columns is taken to generate an entropy index vector that characterizes the degree of information balance in the multi-source attention distribution.

[0064] Using the variance index vector and entropy index vector as input parameters, a multi-index fusion weight calculation is performed, and the raw data for semantic stability score is obtained through the following formula:

[0065]

[0066] Where S is the original data of semantic stability score, V is the cross-channel variance mean index, E is the cross-channel entropy mean index, and wv and we are the variance weight coefficient and entropy weight coefficient obtained by fitting historical closed-loop data, respectively.

[0067] Through the above-mentioned multi-source matrix reconstruction, variance analysis, entropy measurement and weighted fusion processing, the initial semantic anchor attention distribution data of the previous step is transformed into a quantifiable original index of semantic stability score, so as to accurately characterize the robustness of cross-channel semantic expression of each instance.

[0068] S2.2: Create timestamps for instances corresponding to the original semantic stability score data, and use the natural day exponential decay model to perform nonlinear mapping on the time span variable to generate a numerical sequence of time decay factors that automatically reduce weight over time.

[0069] The natural day exponential decay model specifically includes a time span analysis module, a parameter configuration module, an exponential calculation module, and a numerical normalization module.

[0070] The time span parsing module creates a timestamp based on the instance corresponding to the original semantic stability score data, extracts the time difference from when the data was entered into the database to the current calculation time, and converts it into an integer time span variable in natural days to ensure that the time accuracy meets the daily processing requirements.

[0071] The parameter configuration module initializes the preset parameter set for exponential decay processing, sets the decay rate constant to control the overall decay speed, and introduces a periodic adjustment factor to adapt to the business fluctuation cycle, forming a nonlinear mapping configuration file that can be called.

[0072] The exponential calculation module performs a variable substitution operation, substituting the time span variable as an independent variable into a preset exponential mapping function, and performing exponential calculation by multiplying the natural constant base with the decay rate and the time span to calculate the original decay value that decreases non-linearly with the increase of natural days.

[0073] The numerical normalization module performs extreme value normalization processing on the calculated original attenuation value, linearly mapping it to a standardized interval between zero and one, generating a time-dependent attenuation factor numerical sequence that automatically reduces weight over time and is horizontally comparable.

[0074] Based on the original data of semantic stability score input and the creation timestamp of the corresponding instance, the timestamp parsing module is called to extract the time span variable of natural day units, ensuring that the accuracy of the time span data meets the requirements of daily-level processing.

[0075] The time span variable is input into the exponential decay model initialization routine, and the decay base and periodic adjustment factor are set to form a preset parameter set that can be used for nonlinear mapping.

[0076] In the exponential decay model, a variable substitution operation is performed, substituting the time span variable as the independent variable into the mapping function. The decay value increasing with each natural day is then calculated using the exponential operation module, in the following form:

[0077]

[0078] Where W is the time-dependent decay factor, λ is the decay rate constant, and ΔT is the number of natural days for the time span variable.

[0079] The calculated attenuation results are numerically normalized to map the attenuation factor values ​​to a standardized range between zero and one, ensuring the comparability of timeliness indicators across different time spans.

[0080] The normalized attenuation factors are assembled into a numerical sequence according to the instance order, and an index is established with the instance identifier to form a complete numerical sequence output object of the time-dependent attenuation factor.

[0081] By using the above-mentioned nonlinear exponential decay processing method, the original data of semantic stability score and instance timestamp from the previous step are transformed into a numerical sequence of time decay factor that represents the decreasing trend of time effectiveness, thereby realizing the standardized calculation of the time weight part in the dynamic attribute.

[0082] S2.3: Calculate the percentage of historical successful matches for each instance in the initial semantic anchor set across multiple channels such as telephone, SMS, APP, and WeChat. Based on frequency normalization, convert the percentage of historical successful matches for each instance across multiple channels into a probability distribution form to generate channel adaptation weight parameters that reflect cross-channel adaptability.

[0083] The system receives the initial set of semantic anchor points output from the previous sub-step S2.2 and their successful matching records in the multi-channel matching history as input. For each anchor point instance, it extracts historical successful matching data sequences for three channels: telephone, SMS, and APP, ensuring consistency of input data in both time sequence and channel dimensions. It then calls the module for calculating the percentage of successful matching in each instance across multiple channels, performing a channel-by-channel summation operation on the data sequences to generate the total number of matches for each anchor point on each channel. Finally, it performs a full-channel summation operation on the total number of matches for each anchor point to obtain the total number of cross-channel matches for that anchor point as a normalization benchmark. Finally, it divides the total number of matches for each channel by the total number of cross-channel matches, performing frequency normalization calculations to convert the number of matches into a proportional value between zero and one. This proportional value reflects the anchor point's adaptability to the corresponding channel.

[0084] The normalized ratio values ​​are mapped to the channel adaptation weight parameter space, and the data type is converted to floating-point to ensure the accuracy of subsequent weighted fusion calculations. Vector encapsulation is performed on the three types of channel weight parameters for each anchor point to generate a structured channel adaptation weight parameter object, which is then passed as output to subsequent step S2.4 for multi-dimensional attribute association encapsulation processing. Through frequency normalization and structured encapsulation, the matching frequency data from the previous step is transformed into quantitative weight parameters reflecting cross-channel adaptability, achieving accurate representation of the channel dimension in the dynamic attributes of the anchor points.

[0085] S2.4: Receive the original data of the semantic stability score, the numerical sequence of the time-effect decay factor and the channel adaptation weight parameters, perform multi-dimensional attribute association encapsulation operation, bind the three dynamic indicators to the corresponding standardized expression and its core semantic vector, and generate an active semantic anchor object with complete dynamic attributes.

[0086] S2.5: Traverse all generated active semantic anchor objects, and remove invalid nodes with stability scores below the critical value or time decay factors close to zero according to the preset health threshold filtering rules, and output the final set of active anchors that can be used for gravity assessment calculation.

[0087] It receives a collection of active semantic anchor objects with encapsulated dynamic attributes as input data source, and obtains the semantic stability score, time-effect decay factor and channel adaptation weight of each anchor object through iterative retrieval.

[0088] The health threshold filtering rule is invoked, the semantic stability score is input into the comparison operation module, and the difference is calculated with the preset stability threshold. When the difference is negative, the anchor point object is determined to be an unstable node and a removal mark is generated.

[0089] The time-degradation factor is input into the exponential threshold detection unit and compared with the zero critical threshold to calculate the remaining amount of attenuation. When the remaining amount is less than the threshold, the anchor point object is determined to be a failed node and a removal mark is generated.

[0090] Perform a logical OR operation on the set of culling tags to form a list of invalid node indices that need to be culled, and remove the corresponding node instances from the set of active semantic anchor objects.

[0091] Perform index reconstruction and vector storage refresh operations on the remaining set of anchor objects to generate the final active anchor set data structure that can be used for gravity assessment calculations.

[0092] By using health threshold filtering and elimination methods, the standardized expressions and dynamic attribute binding results from the previous step are transformed into a set of highly reliable active semantic anchors, thereby achieving high-precision and robust matching performance in the subsequent gravity assessment stage.

[0093] like Figure 2 As shown, step S3: Obtain the newly added work order text and perform feature mapping processing through a preset semantic encoder to generate a three-dimensional spatial vector representing the semantic features of the newly added work order text, which serves as the input object for subsequent gravity assessment.

[0094] In this embodiment, the specific architecture of the semantic encoder consists of a text preprocessing module, a semantic understanding module, a feature compression module, a normalization module, and a spatial mapping module.

[0095] The text preprocessing module uses a dedicated word segmenter for government affairs to perform fine-grained segmentation on new work order texts containing dialects and colloquial expressions, generating standardized word sequences carrying location encoding information.

[0096] The semantic understanding module calls a pre-trained lightweight distillation language model to perform forward inference operations on the word sequence, extracts a high-dimensional hidden layer state matrix that incorporates the context, and obtains an initial high-dimensional semantic tensor representing the deep semantic features of the work order.

[0097] The feature compression module performs linear projection dimensionality reduction on the initial high-dimensional semantic tensor using a learnable weight matrix, mapping it to a low-dimensional latent space to generate a compressed semantic vector with fixed dimensions that retains key semantic distinguishability.

[0098] The normalization module is based on Euclidean norm calculation and performs modulus normalization on each vector component of the compressed semantic vector to achieve standardized scaling and eliminate amplitude interference caused by differences in text length, and outputs a standardized semantic feature vector of unit length.

[0099] The spatial mapping module transforms the standard semantic feature vector into three-dimensional spatial coordinates and uses principal component analysis mapping to reconstruct the high-dimensional abstract features into a visualized three-dimensional spatial vector, generating a semantic mapping object for subsequent determination of the landing point of the composite gravitational potential energy field.

[0100] The workflow of the semantic encoder specifically includes:

[0101] S3.1: Obtain the newly added work order text data, and use a dedicated word segmenter for government affairs to perform fine-grained segmentation of the original text containing dialect words and colloquial expressions to generate a standardized word sequence carrying positional encoding information, which serves as the basic input unit for subsequent semantic mapping.

[0102] When acquiring new work order text data, it is imported into the text preprocessing module as the sole input for this step, ensuring that the original text contains complete dialect vocabulary, colloquial expressions, and channel-specific markers. This original text is then input into a dedicated word segmenter for government affairs, loading a multi-dimensional vocabulary resource containing a dialect dictionary, colloquial phrase library, and multi-channel semantic mapping table for matching priority decisions within the segmenter. Based on the dictionary matching results from the segmenter, fine-grained segmentation is performed, recursively segmenting the text based on the optimal matching path to break down continuous dialect phrases into the smallest parsable semantic units. Positional encoding information is appended to each segmented word. The relative positions of the sequence are preserved by adding position vectors to the word sequence. These position vectors are calculated based on the word's character index in the original text, and the position vector components are determined using the following formula:

[0103]

[0104] Where `idx` is the starting character index of the term in the original text, `start` is the starting character index of the text, and `len` is the total character length of the original text. Positional encoding is bound to the term content, and a standardized term sequence is generated through sequence encapsulation operations, ensuring that each term in the sequence has both a unified grammatical format and clear positional information. Through this processing method, the original new work order text is transformed into a standardized term sequence with semantic adaptability in the government domain and can be directly called by the subsequent semantic mapping module, achieving a structured input effect for the new work order text.

[0105] S3.2: Based on the standardized word sequence, call the pre-trained distillation language model to perform forward inference operations to extract the high-dimensional hidden layer state matrix that integrates the context, thereby obtaining the initial high-dimensional semantic tensor representing the deep semantic features of the work order.

[0106] Based on the standardized lexical sequence, when the pre-trained distilled language model is called to perform forward inference, the lexical sequence carrying positional encoding information is input into the model embedding layer to generate an embedding matrix containing the fusion results of word vectors and positional vectors.

[0107] The embedding matrix is ​​passed to the multi-layer self-attention module of the distillation model, and the multi-head attention mechanism is used to assign weights to the contextual associations of each word in the sequence with other words, and generate a multi-head attention output matrix.

[0108] The multi-head attention output matrix is ​​processed by residual connection and layer normalization to form an intermediate feature matrix that has been stabilized and scaled to ensure the numerical stability of subsequent feature extraction.

[0109] The intermediate feature matrix is ​​input into the feedforward fully connected network, and position-by-position linear mapping and nonlinear activation are performed to form the hidden layer state matrix after feature nonlinear transformation.

[0110] The hidden state matrix is ​​iterated sequentially in each layer of the network, and finally the hidden states of each layer are combined in the output layer to generate an initial high-dimensional semantic tensor that integrates global context and local semantics. Through this processing method, the standardized word sequence of the previous step is transformed into a high-dimensional tensor that can represent the deep semantic features of the newly added work order text, so as to realize the semantic feature construction for subsequent dimensionality reduction and gravity assessment.

[0111] For example, in a government service hotline platform, a newly added work order text, "The streetlight in front of my house has been broken for three days," is finely segmented using a government-specific word segmenter, generating a standardized word sequence of length 12. Each word's embedding vector has a dimension of 128, and its positional encoding vector also has a dimension of 128, resulting in a 12×256 fused embedding matrix. This embedding matrix is ​​input into a 4-layer distillation model, each layer containing 8 attention heads, with each head's key, value, and query vectors having a dimension of 64. During the multi-head attention calculation, the weight matrix generated by the model significantly highlights the importance of the "streetlight" and "broken" words to the overall semantics. After residual connections and layer normalization, the intermediate feature matrix maintains a dimension of 12×256. The feedforward network contains two levels of linear mapping: the first level maps the 256-dimensionality of each position to 512 dimensions and applies the ReLU activation function; the second level compresses it back to 256 dimensions. The outputs of each layer are stacked sequentially and fused at the output layer, ultimately resulting in an initial high-dimensional semantic tensor with a shape of 12×256. In this scenario, the fields of this high-dimensional tensor can be used to calculate a semantic stability metric, for example, through the formula:

[0112]

[0113] Where S is the stability score, n is the number of lexical units, x is the lexical attention weight value, and μ is the weight mean. The calculation result reaches a high value in this work order, indicating that the text has strong semantic consistency in different channels. It can be used to enhance the accuracy of gravity assessment and significantly improve the classification accuracy and merging efficiency of similar requests.

[0114] S3.3: Perform linear projection dimensionality reduction on the initial high-dimensional semantic tensor, and map the high-dimensional hidden layer state to a preset low-dimensional latent space through a learnable weight matrix to generate a compressed semantic vector with fixed dimensions and retaining key semantic distinguishability.

[0115] For the initial high-dimensional semantic tensor obtained through forward inference of the distilled language model, the linear projection dimensionality reduction operator is called to load the preset learnable weight matrix. This weight matrix is ​​obtained by gradient optimization on the semantic dataset in the government affairs domain during the training phase and has the mapping ability to capture the semantic discriminativeness of demands.

[0116] Each hidden layer state vector of the initial high-dimensional semantic tensor is sequentially input into the projection operator, and matrix multiplication is performed to complete the mapping from the original high-dimensional semantic space to the preset low-dimensional latent space. The dimensions of this low-dimensional space are set in advance according to business needs and computing resource constraints.

[0117] When performing matrix multiplication, a numerical stability enhancement strategy is adopted. By introducing a small-amplitude regularization term during the projection matrix multiplication process, the information collapse phenomenon in dimensionality compression is suppressed, ensuring that the vector after dimensionality reduction can still retain key semantic features.

[0118] The low-dimensional latent space vector set obtained by mapping is subjected to batch mean centering to eliminate the low-dimensional vector offset caused by the difference in the distribution of different work order text contexts, and to unify the reference coordinate system of the vectors in the latent space.

[0119] The centralized low-dimensional latent space vector is transformed into a compressed semantic vector, outputting feature data with fixed dimensions that retains key semantic distinguishability in the semantic space of government affairs demands.

[0120] By using the linear projection dimensionality reduction method described above, the initial high-dimensional semantic tensor from the previous step is transformed into a compressed semantic vector, thereby achieving the expected technical effect of reducing computational burden while maintaining the accuracy of cross-channel appeal semantic recognition.

[0121] For example, in the deployment scenario of the Ganzhou Municipal Government Service Hotline Platform, the initial high-dimensional semantic tensor (dimension 1024) generated by the distilled language model is subjected to linear projection dimensionality reduction processing. The target low-dimensional latent space dimension is set to 384, and the projection weight matrix size is 384×1024. The matrix multiplication operation is performed as follows:

[0122]

[0123] Where V is the dimensionality-reduced compressed semantic vector matrix, W is the learnable weight matrix, and H is the initial high-dimensional semantic tensor matrix. A λ-regularization term is introduced into the weight matrix, calculated as follows:

[0124]

[0125] Where I is the identity matrix, and λ is set to 0.001 for stable numerical calculation. After mapping, the mean-centering process is performed using the following formula:

[0126]

[0127] Where μ is the mean vector of each dimension of V. After outputting a 384-dimensional compressed semantic vector, in subsequent modulus normalization and 3D mapping, this vector significantly improves the semantic gravity computation performance compared to the original 1024-dimensional features, while significantly reducing computational latency, ensuring the accurate attribution of cross-channel dialectal spoken language appeals in dynamic clustering.

[0128] S3.4: Perform modulus normalization processing based on the compressed semantic vector, and use Euclidean norm calculation to standardize and scale each vector component to eliminate amplitude interference caused by text length differences and output a unit length normalized semantic feature vector.

[0129] S3.5: Perform a three-dimensional spatial coordinate transformation on the unit length of the standard semantic feature vector, and reconstruct the high-dimensional abstract features into a visualized three-dimensional spatial vector through principal component analysis mapping, so as to generate the final semantic mapping object directly used for determining the landing point of the composite gravitational potential energy field.

[0130] The unit-length canonical semantic feature vector is input to the principal component analysis calculation module, and the feature covariance matrix calculation unit is called to perform covariance operation on the vector in each dimension to capture inter-dimensional correlation.

[0131] Based on the covariance matrix, eigenvalue decomposition is performed to solve for all eigenvalues ​​and their corresponding eigenvectors. The obtained eigenvalues ​​are then sorted from largest to smallest to establish the priority direction for dimensionality reduction transformation.

[0132] A projection transformation matrix is ​​constructed for the eigenvectors corresponding to the top three eigenvalues. The normalized semantic eigenvectors are then multiplied with the projection matrix to achieve a linear mapping from the high-dimensional latent space to the three-dimensional visualization space.

[0133] During the mapping process, the three-dimensional coordinate components are numerically scaled through normalization to keep the value range on each coordinate axis consistent in order to avoid spatial distortion caused by different dimensions.

[0134] The normalized three-dimensional coordinates are encapsulated into a final semantic mapping object, ensuring that it can be directly input into the composite gravitational potential energy field landing point determination module to achieve quantitative analysis of spatial position.

[0135] By using principal component analysis mapping, the standardized semantic feature vectors from the previous step are transformed into semantic mapping objects with three-dimensional spatial positioning attributes, enabling the quantifiable and visualized representation of the semantic features of newly added work order texts in a gravitational potential field.

[0136] like Figure 3 As shown, step S4: The three-dimensional spatial vector is sequentially weighted and applied to the dynamic attributes in the active anchor point set using a weighted response function to calculate the semantic attraction strength value representing the association strength between the new work order and each anchor point. Specifically, this includes:

[0137] S4.1: Obtain the time-effect decay factor in the three-dimensional spatial vector and the set of active anchor points, and perform exponential decay update processing on the time-effect decay factor based on the natural day time span to generate a real-time time-effect weight parameter that reflects the freshness of the anchor points under the current time window.

[0138] The process retrieves the three-dimensional spatial vector corresponding to the new work order text output by step S3, as well as the time-effect decay factor parameters of each anchor point within the active anchor point set. It then calls the time span calculation module to perform a whole-day difference calculation between the current system date and the anchor point generation timestamp to obtain the natural day time span value. This natural day time span value is input into the exponential decay function model. Based on a preset time decay benchmark constant and decay period, a nonlinear update operation is performed on the original time-effect decay factor to form a preliminary decay weight value reflecting the freshness of the anchor points. A precision enhancement mechanism is used to retain decimal places and correct floating-point stability for the preliminary decay weight value, ensuring consistency in numerical comparisons across different time windows. The above update operation is performed on all active anchor points to generate a corresponding real-time time-effect weight parameter array. This array is bound to the three-dimensional spatial vector of the new work order text to support subsequent channel matching degree fusion calculations. The formula for calculating the real-time time-effect weight using the exponential decay model is:

[0139]

[0140] Where w is the updated real-time timeliness weight, w0 is the initial timeliness decay factor, λ is the decay constant, and Δt is the natural day time span. Through the above processing method, the timeliness decay factor associated with the three-dimensional spatial vector and the active anchor point set in the previous step is transformed into a freshness weight parameter that adjusts in real time over time, achieving precise quantification of the impact on the timeliness of anchor points and providing a reliable time dimension input for subsequent weighted response function calculations.

[0141] S4.2: Receive the real-time timeliness weight parameter and the channel identifier to which the newly added work order text belongs, and use the channel matching degree mapping table to perform dynamic calibration processing on the channel adaptation weight to generate a corrected channel adaptation coefficient that integrates channel preference characteristics.

[0142] The system receives the real-time timeliness weight parameters generated by S4.1 and the channel identifier of the newly added work order text. It then retrieves the weight baseline value of the channel from the historical matching records by calling the channel matching degree mapping table, forming an initial vector for channel adaptation weights. This initial vector and the real-time timeliness weight parameters are input to the weight correction calculation module. Matrix multiplication is used to synchronously scale the weight values ​​along the time decay dimension, resulting in time-corrected intermediate data for channel adaptation weights. Based on the matching relationship between the channel of the newly added work order text and the historical high-frequency channels of each active anchor point, channel preference coefficients are extracted from the mapping table. A preference calibration operation is then performed on the time-corrected intermediate data for channel adaptation weights, generating corrected channel adaptation coefficients that incorporate channel preference characteristics. A normalization operator is used to unify the scale of the corrected channel adaptation coefficients, mapping the numerical range to the closed interval [0,1] to eliminate the impact of differences in channel weight distribution on subsequent gravity calculations. Through the above processing method, the real-time timeliness weight parameters of the previous step and the channel identifier of the newly added work order text are transformed into a modified channel adaptation coefficient that can directly participate in the semantic gravity strength calculation, so as to achieve dynamic and accurate matching based on channel preference and timeliness.

[0143] S4.3: Input the three-dimensional spatial vector, the active anchor point core semantic vector and the corrected channel adaptation coefficient, and perform basic semantic distance calculation based on cosine similarity calculation to generate a basic similarity score that represents the degree of similarity of text content.

[0144] The input objects are the semantic vector of the newly added work order text obtained by three-dimensional spatial coordinate transformation, the core semantic vector in the active anchor point set, and the corrected channel adaptation coefficient that integrates channel preference characteristics.

[0145] The three-dimensional spatial vector and the core semantic vector of each active anchor point are respectively input into the cosine similarity calculation module. The vector operation unit is called to perform the dot product operation and simultaneously calculate the magnitude of each vector to form the denominator and numerator of the basic similarity calculation data.

[0146] By binding the modified channel adaptation coefficient to the vector pair of each active anchor point, a comprehensive basic similarity calculation formula is constructed, which includes the semantic feature numerator, the modulus multiplication integral denominator, and the channel coefficient weight term.

[0147] Using the standard cosine similarity expression, it is represented in MathML as:

[0148]

[0149] Where v is the three-dimensional semantic vector of the newly added work order text, u is the core semantic vector of the active anchor point, and S is the basic similarity score.

[0150] The weighted multiplication operation between the formula calculation result and the modified channel adaptation coefficient is represented in MathML as follows:

[0151]

[0152] Where C is the corrected channel adaptation coefficient, and S′ is the basic similarity score after channel weighting.

[0153] The S′ scores corresponding to all active anchor points are compiled into the basic similarity score sequence in the order of the anchor points, and used as input data for subsequent multi-dimensional weighted fusion to generate the original gravity intensity value.

[0154] By using a cosine similarity and channel-weighted fusion processing method, the semantic vector comparison results of the previous step are transformed into a quantitative text content similarity index, thereby achieving an accurate representation of the semantic matching degree between new work order texts and active anchors.

[0155] S4.4: Obtain the basic similarity score, real-time timeliness weight parameters and corrected channel adaptation coefficients, and perform nonlinear response function operation processing based on the multi-dimensional weighted fusion mechanism to generate the original gravity strength value that comprehensively reflects semantic content, timeliness and channel matching degree.

[0156] The basic similarity score, real-time timeliness weight parameters, and corrected channel adaptation coefficient are used as inputs for the multi-dimensional weighted fusion mechanism.

[0157] Construct a fusion vector model that includes semantic content components, time-sensitivity components, and channel matching components to ensure that the three components can participate in unified calculations under the condition that they are consistent in scale.

[0158] The nonlinear response function operation module is called, and the three components are input as independent variables into the weighted product core formula inside the function. Power-law transformation is used to enhance the contribution of high-weight components.

[0159] Each component in the fusion vector model is bound to a preset weight coefficient, and a weighted accumulation process is performed to generate a comprehensive response value matrix before fusion.

[0160] In the comprehensive response value matrix, a normalized exponential function is applied to each row of elements to form a smooth distribution to suppress abnormally high value interference, and the original gravity strength value that comprehensively reflects semantic content, time efficiency and channel matching degree is output.

[0161] The specific calculation form of the response function is as follows:

[0162]

[0163] Where S is the basic similarity score, T is the real-time timeliness weight parameter, C is the corrected channel adaptation coefficient, α, β, and χ are the weight coefficients corresponding to each component, and γ, δ, and η are nonlinear power exponent adjustment parameters.

[0164] By processing the nonlinear response function, the result of the previous step is transformed into a comprehensive response value matrix, thereby achieving multi-dimensional integration of the correlation strength between the new work order and each anchor point.

[0165] S4.5: Receive the original gravity intensity value, perform dimensional unification and boundary constraint processing based on normalization scaling, and generate the final semantic gravity intensity value that characterizes the association strength between the new work order and each anchor point.

[0166] Step S5: Based on the sorting results of the semantic gravity intensity values, select the top five anchor points to form a temporary gravity candidate region dataset. Then, using each candidate anchor point as a virtual center, set the influence radius in reverse gravity intensity order, and superimpose them to generate a composite gravitational potential energy field used to define the region's affiliation. Specifically, this includes:

[0167] S5.1: Obtain the normalized semantic gravity intensity values, and use the quicksort algorithm to sort the original gravity intensity values ​​corresponding to all anchors in the active anchor set in descending order to generate a ranking list of anchor gravity intensities with clear priority order, which serves as the input basis for screening highly relevant candidate objects.

[0168] The normalized semantic gravity intensity value is obtained as the input object. The core module of the fast sorting algorithm based on heap sort optimization is called to perform descending sort calculation on the semantic original gravity intensity value corresponding to each anchor in the active anchor set, forming a serialized sorting buffer.

[0169] The sorting buffer is mapped item by item to the index position table. The index position table accurately reflects the correspondence between the anchor point identifier and its original gravitational strength value, generating a sorting mapping matrix with a one-to-one mapping structure.

[0170] The original gravitational intensity value field is extracted from the sorting mapping matrix. The boundary detection mechanism is used to verify the consistency between the numerical range and the normalization standard, outliers are screened out, and the effective numerical domain of the sorting mapping matrix is ​​updated.

[0171] Based on the effective numerical domain, the gradient change rate of the original gravitational intensity value is recorded sequentially. The stability of the sorting result is verified twice by the gradient change sequence, and invalid sorting items that are shifted due to numerical jitter are removed.

[0172] The validated sorting results are encapsulated into a ranking list of anchor point gravity strengths with a clear priority order, and the ranking list is used as the input for filtering highly correlated candidate objects.

[0173] This sorting process transforms the normalized semantic gravitational intensity sequence into strictly ordered and indexable anchor ranking data, enabling precise determination of the action domain for subsequent candidate region constraints and composite gravitational potential field construction.

[0174] For example, in the cross-channel request classification task of the government service hotline platform, the obtained semantic gravity strength value length is 100, and the corresponding number of active anchor points is also 100. The initial heap size of the heap sorting optimization module is set to 32, and the unit comparison time during sorting is kept at 0.05 milliseconds. After sorting, the buffer is mapped to the index position to generate a 100×2 sorting mapping matrix, where the first column is the unique identifier of the anchor point, and the second column is the normalized original gravity strength value. The outlier detection threshold is set to remove values ​​greater than 1.0 or less than 0.0. In this example, three entries caused by data anomalies are removed. The gradient change sequence is calculated using the following formula:

[0175]

[0176] Where S i Δ is the original gravitational strength value of the i-th anchor point. i This represents the gradient change rate between adjacent rankings. After calculation and removal of entries with an absolute gradient value less than 0.0001 and a position offset greater than 1 position, the remaining ranking list has a size of 95, exhibiting strict descending order and stable priority. This list can be directly used in the candidate area screening process to quickly determine highly relevant anchor points.

[0177] S5.2: Based on the anchor point gravity strength ranking list, perform a top truncation filtering operation to extract anchor point instances and their corresponding core semantic vector coordinates and dynamic attribute parameters within a preset threshold range, generating a temporary gravity candidate region dataset containing five high-confidence matching objects to limit the scope of subsequent potential field construction. In this embodiment, the threshold range is the top five ranked anchor points.

[0178] Receive the list of anchor point gravity strength rankings generated by quicksort and the corresponding active anchor point index table, and extract the position information from the sorting results as input filtering conditions.

[0179] Perform a position truncation operation on the ranking list, using a preset top truncation threshold to limit the range of high-priority anchor points to be extracted, ensuring that the number of filtered objects is exactly five to meet the scale requirements for subsequent potential field construction.

[0180] The anchor point instance records that match the ranking list are called from the anchor point index table. The core semantic vector coordinate data bound to each anchor point instance is read as the spatial positioning basis for subsequent potential energy calculation.

[0181] The dynamic attribute parameters bound to each anchor instance are read synchronously, including semantic stability score, time-effect decay factor and channel adaptation weight, as input for subsequent calculation of influence radius and potential energy distribution.

[0182] The core semantic vectors and dynamic attribute parameters of the five extracted anchor point instances are combined and encapsulated into structured entries, and serialized and stored to form a temporary gravity candidate region dataset, ensuring that the data format of this dataset matches the input of the potential energy field construction module.

[0183] Through the above processing method, the ranking list results of the previous step are transformed into a temporary gravitational candidate region dataset containing five high-confidence matching objects and their spatial and dynamic attributes, thereby achieving the expected technical effect of limiting the scope of potential energy field construction.

[0184] S5.3: Receive each candidate anchor point in the temporary gravity candidate region dataset, perform inverse proportional mapping calculation based on its semantic gravity intensity value, convert the original gravity intensity value into an initial influence radius parameter characterizing the breadth of the action range, generate a dynamic action radius value sequence corresponding to each candidate anchor point, and establish the spatial coverage capability of a single anchor point.

[0185] The system receives each candidate anchor point from the temporary gravity candidate region dataset and calls the data parsing module to extract its semantic gravity intensity value as the basic input parameter for calculation.

[0186] The semantic gravitational intensity value is input into the inverse proportional mapping calculation component to establish a mathematical relationship model between the influence radius and the gravitational intensity. The model constructs an inverse proportional function based on the spatial coverage capability and the gravitational potential energy decay law to ensure that a high original gravitational intensity value is mapped to a smaller influence radius and a low original gravitational intensity value is mapped to a larger influence radius.

[0187] During the mapping calculation process, in order to prevent the radius from being too small due to extremely high gravity values ​​and causing misclassification, a boundary constraint adjustment factor is applied to control the radius calculation result within the preset range of minimum and maximum coverage radius.

[0188] Calculate using the inverse proportional function formula:

[0189]

[0190] Where R is the radius of influence, k is the proportionality coefficient, G is the semantic gravity intensity value, and ε is the boundary constraint adjustment factor, which is adjusted by... It can balance the sensitivity of high and low gravitational values ​​in radius mapping.

[0191] The calculation results of each candidate anchor point are stored in the dynamic action radius buffer structure, indexed by the anchor point identifier, to generate a dynamic action radius numerical sequence that corresponds one-to-one with each candidate anchor point.

[0192] Through the above processing method, the candidate anchor point data from the previous step is transformed into a dynamic action radius sequence that can be directly used for the construction of a spatial potential energy field, thereby achieving accurate quantification of the spatial coverage capability of a single anchor point.

[0193] S5.4: Based on the dynamic action radius numerical sequence and the core semantic vector coordinates of each candidate anchor point, a multidimensional Gaussian kernel function distribution model with the candidate anchor point as the center is constructed. The discrete initial influence radius parameter is transformed into a continuous spatial potential energy decay curve using the radial basis function interpolation algorithm, and an independent potential energy distribution matrix characterizing the gravitational action range of a single anchor point is generated.

[0194] Based on the dynamic action radius numerical sequence and the core semantic vector coordinates of each candidate anchor point, a high-dimensional space modeling engine is accessed. The semantic vector coordinates of each anchor point are used as the center parameter of the three-dimensional Gaussian kernel function to establish a spatial function prototype. The dynamic action radius numerical sequence is mapped from radius to kernel function scale factor, transforming the radius value into a standard deviation parameter that controls the Gaussian kernel diffusion rate, thus forming the boundary condition for spherical potential energy decay.

[0195] The radial basis function interpolation module is invoked, using the Gaussian kernel function as a basis. The discrete radius of action and the coordinates of the sphere's center are input, and the potential energy value in continuous spatial coordinates is calculated through interpolation, achieving a smooth transformation from a discrete radius to a continuous decay curve. A spatial distribution matrix is ​​constructed on the interpolation results, mapping the calculated potential energy values ​​to matrix elements according to three-dimensional coordinate indices, forming a potential energy distribution matrix characterizing the entire gravitational range of a single anchor point.

[0196] The matrix is ​​encapsulated in a structured manner, marking the center position information, standard deviation parameter, and potential energy decay curve index, outputting an independent potential energy distribution matrix object that can be used for subsequent superposition operations. Through a distribution modeling process using a multidimensional Gaussian kernel and radial basis function interpolation, the dynamic radius parameter from the previous step is transformed into a single-anchor-point potential energy distribution matrix with continuous spatial analytical capabilities, achieving a quantifiable description of the single-anchor-point gravitational range.

[0197] For example, in a multi-channel request clustering scenario for government service hotlines, the anchor point ranked first in the temporary gravity candidate region dataset is selected, with core semantic vector coordinates of (0.85, -0.32, 0.47) and a dynamic radius of action of 2.4. The sphere center coordinates are input into a high-dimensional space modeling engine, and 2.4 is mapped to the Gaussian kernel standard deviation σ, using the formula...

[0198]

[0199] In the formula, G(x) is the semantic gravitational potential energy value at any spatial point x, σ is set to 2.4, and d is the Euclidean distance between any spatial point and the center of the sphere. The discrete radius value is input into the radial basis function interpolation module to calculate the potential energy value at each point in the spatial coordinates, resulting in a continuous curve decaying outwards from the center of the sphere. When constructing the matrix, the entire spatial distribution is sampled with a 3D grid resolution of 0.1 units, resulting in a matrix with dimensions of 150×150×150, where each cell represents the potential energy value at the corresponding point. The matrix encapsulates the center coordinates, standard deviation parameter, and decay curve index. System verification shows that within this matrix, the potential energy value at a distance of 0.5 units from the center is significantly higher than that at a distance of 2.0 units, exhibiting a significant spatial gravitational gradient. This supports accurate calculation of potential energy superposition at multiple anchor points in subsequent S5.5 processes, ultimately significantly improving the stability of cluster assignment in classification decisions.

[0200] S5.5: Perform linear superposition operation on the independent potential energy distribution matrices generated by all candidate anchor points in the temporary gravity candidate region dataset, fuse the potential energy contribution values ​​of each anchor point in the three-dimensional semantic space, eliminate the isolation of a single potential field and form a composite gravitational potential energy field with global topological features, so as to output the final potential energy boundary model for determining the landing position of the new work order text.

[0201] The three-dimensional semantic space is a projection of the multi-dimensional feature space generated by a preset semantic encoder onto a three-dimensional dimension, where each dimension represents a specific semantic attribute.

[0202] Receive the independent potential energy distribution matrices in the temporary gravity candidate region dataset, and call the multi-matrix linear superposition module to perform a summation operation on each element of the individual matrix according to its position index, so as to generate an initial comprehensive matrix form of data that integrates the potential energy contribution values ​​of all candidate anchor points.

[0203] The element values ​​of each independent potential energy distribution matrix are weighted with the dynamic attribute parameters of the corresponding anchor point. The weight coefficient is determined by the product of the semantic gravitational strength value of the anchor point and the initial influence radius parameter. The weighted potential energy values ​​are superimposed through matrix multiplication to enhance the locational influence of high-confidence anchor points.

[0204] Based on the weighted superposition of the comprehensive matrix data, spatial normalization is performed, and the maximum value normalization strategy is used to map the matrix elements to the [0,1] interval to eliminate the interference of the difference in the magnitude of the potential energy values ​​of different anchor points on the global topological features.

[0205] Three-dimensional spatial interpolation is performed on the normalized composite matrix. A hybrid algorithm of bilinear interpolation and radial basis function is called to fill the sparse gap in the spatial continuity of the matrix data, generating a continuous three-dimensional composite gravitational potential energy field with smooth potential energy change characteristics.

[0206] In a continuous three-dimensional composite gravitational potential energy field, the local potential energy gradient at each key spatial location is calculated to form a potential energy boundary model that characterizes the global topological features, which is used to determine the landing point of the three-dimensional spatial vector of the newly added work order text.

[0207] By using linear superposition, weighted normalization, and spatial interpolation, the independent potential energy distribution matrix result from the previous step is transformed into a global composite gravitational potential energy field that can be used to determine the landing point, thereby achieving accurate definition of the semantic scope of newly added work order texts.

[0208] For example, in the cross-channel request clustering scenario of the government service hotline platform, the temporary gravity candidate region dataset contains 5 anchor points. Its independent potential energy distribution matrix is ​​a 100×100×100 three-dimensional array with maximum potential energy values ​​of 2.5, 3.2, 2.1, 4.0, and 1.8, corresponding to semantic gravity intensity values ​​of 0.85, 0.92, 0.76, 0.95, and 0.70, and initial influence radius parameters of 8.0, 9.5, 7.0, 10.0, and 6.5. The matrix elements are summed item by item according to their position indices to form a comprehensive matrix, and then the formula is used...

[0209]

[0210] Where w i The anchor weight coefficient is equal to the product of its standardized semantic gravity strength and the initial influence radius parameter, M. i Let be the element values ​​of the independent potential energy matrix, and n be the total number of anchor points. After calculating the comprehensive matrix, normalize it to its maximum value using the following formula:

[0211]

[0212] E norm The normalized potential energy matrix is ​​then used. After normalization, radial basis function interpolation is applied to make the data spatially continuous, forming a smooth three-dimensional potential energy structure. Verification results show that the composite gravitational potential energy field constructed in this scenario can significantly improve the attribution accuracy and clustering stability in the determination of the attribution of newly added work order texts, eliminate the attribution error caused by the isolation of a single anchor point potential field, and ensure the consistency of the attribution of claims in boundary areas.

[0213] Step S6: Determine the landing position of the three-dimensional spatial vector in the composite gravitational potential energy field. If the three-dimensional spatial vector falls into the potential energy valley, it is directly assigned to the corresponding anchor point cluster. If the three-dimensional spatial vector is in the transition zone of multiple overlapping potential energy fields, perform dependency path offset analysis to generate the target semantic anchor point to which it is ultimately assigned. Specifically, this includes:

[0214] S6.1: Obtain the potential energy distribution data of the three-dimensional spatial vector to be determined and the composite gravitational potential energy field. Based on the potential energy gradient calculation rule, perform differential processing on the local potential energy curvature at the location of the three-dimensional spatial vector to generate a regional attribute identifier that represents the location as a potential energy valley or a potential energy overlap transition zone.

[0215] The system acquires the potential energy distribution data of the three-dimensional spatial vector to be determined and the composite gravitational potential energy field. It then calls the spatial position index module to read the coordinate index value of the vector in the potential energy field, forming the mapping relationship data between the position vector and the potential energy matrix.

[0216] The mapping relationship data is input into the gradient calculation module, and the central difference operator is used to perform a partial derivative operation on the potential energy distribution function at this position. The local gradient components along the three coordinate axes of the three-dimensional semantic space are calculated respectively, and the gradient vector representing the rate of change of potential energy is generated.

[0217] Using the gradient vector as input, the curvature analysis module is called to perform second partial derivative operations on the gradient change trend, constructing a Hessian matrix containing second derivatives in three principal directions and a mixed derivative term to characterize the local curvature features of the potential energy surface at that location.

[0218] Perform eigenvalue decomposition on the Hessian matrix to extract the maximum and minimum eigenvalues, and determine whether all eigenvalues ​​are positive or negative. If all eigenvalues ​​are positive, the position is the potential energy valley. If there are eigenvalues ​​with different signs, the position is in the potential energy overlap transition zone.

[0219] The above judgment logic is transformed into a regional attribute identifier code to generate binary or enumerated identifier values, where the potential energy valley state corresponds to the preset first identifier code and the potential energy overlapping transition zone state corresponds to the second identifier code.

[0220] By combining gradient and curvature analysis, the results of the previous step are transformed into regional attribute identifiers that characterize the terrain of location potential energy, enabling accurate determination of the location type of three-dimensional spatial vector landing points, and providing clear input basis for subsequent attribution mapping or dependent path offset analysis.

[0221] S6.2: Based on the regional attribute identifier, perform branch logic judgment. If the regional attribute identifier indicates a potential energy valley, directly map the three-dimensional spatial vector to the corresponding target semantic anchor cluster to generate the initial ownership relationship record for confirmation of rights.

[0222] S6.3: If the regional attribute identifier indicates multiple potential energy overlapping transition zones, then extract the subject-verb-object structure template of the original new work order text and candidate anchor cluster corresponding to the three-dimensional spatial vector, and use the dependency parsing engine to perform syntax tree construction processing on the original new work order text to generate a syntactic dependency graph containing key verb phrases and their dependency paths.

[0223] For newly added work order texts located in multiple potential energy overlap transition zones, the original newly added work order text data corresponding to the three-dimensional spatial vector and the subject-verb-object structure template of the candidate anchor point cluster are received as input conditions.

[0224] The original newly added work order text is processed by a dedicated word segmenter for government affairs, which performs fine-grained word segmentation. The text containing dialect words, colloquial expressions and complex sentences is divided into standardized word sequence with part-of-speech and positional encoding, providing a structured basis for syntactic analysis.

[0225] The standardized lexical sequence is input into the dependency parsing engine for forward inference operations, extracting the dependency relation labels of each lexical and their parent-child node connection information in the syntax tree, and generating an initial syntactic dependency matrix.

[0226] Based on the syntactic dependency matrix, tree construction rules are executed, and lexical units are constructed into hierarchical nodes according to their dependency relationships to form a complete syntactic tree structure. At the same time, the path length and branch angle from each node to the root node are calculated for subsequent dependency path offset analysis.

[0227] All key verb phrase nodes are identified and extracted from the syntactic tree structure, and they are matched with the subject-verb-object structure templates provided by the candidate anchor clusters. The nodes that are successfully matched and their dependency path information are retained.

[0228] The extracted key verb phrases and path information are combined into a syntactic dependency graph data object, which contains verb phrase encoding, path link sequence and location information, and can support the calculation of structural difference measurement.

[0229] The syntactic dependency graph constructed by the dependency parsing engine transforms the results of the previous step into quantifiable structural difference analysis data, enabling accurate path measurement in semantic structure disambiguation scenarios.

[0230] S6.4: Based on the syntactic dependency graph and the subject-verb-object structure template of each candidate anchor cluster, perform path offset quantification calculation. By comparing the node distance of the key verb phrase in the syntactic dependency graph with the standard topological distance in the subject-verb-object structure template, generate a path offset numerical sequence that represents the degree of semantic structure difference.

[0231] Based on the syntactic dependency graph and the subject-verb-object structure templates of each candidate anchor cluster, the path offset quantification calculation module receives complete syntactic node information containing key verb phrases and dependency paths, forming an input matrix for difference analysis. The node distances corresponding to each key verb phrase are calculated between the node positions and the predefined standard topological node positions in the template, and a depth-first traversal algorithm is used to output the path length values ​​between nodes. The path length data obtained from the syntactic dependency graph is then compared with the standard path length data of the subject-verb-object structure template to generate a path difference vector, which characterizes the degree of offset between the current syntactic structure and the standard structure in the dependency chain. The path difference vector is input into the mathematical normalization processing module, which uses the following formula to calculate the normalized offset:

[0232]

[0233] Among them, P i L represents the path offset value of the i-th verb phrase. obs To observe the length of syntactic dependency paths, L std The standard template path length is used, and the ratio of the absolute value of the difference to the standard length characterizes the offset strength. The path offset values ​​are then aggregated in verb phrase sequence order to form a path offset value sequence, which serves as the standard output for structural difference quantification. This processing method transforms the syntactic dependency graph and subject-verb-object structure template from the previous step into path offset values ​​that can be used for disambiguation, achieving precise quantification of the grammatical structure differences of candidate anchor points.

[0234] S6.5: Perform a minimum value filtering operation based on the path offset value sequence, select the candidate anchor point with the smallest path offset value as the final matching object, forcibly bind the three-dimensional spatial vector to the target semantic anchor point cluster to which the final matching object belongs, and generate the final attribution instruction after completing the disambiguation process.

[0235] The input condition is the generated path offset numerical sequence, which is derived from the dependency parsing results and the structural difference calculation results of the candidate anchor cluster subject-verb-object structure template.

[0236] The minimum value determination process is performed on the path offset numerical sequence, and the anchor cluster instance with the smallest corresponding offset is marked as the highest confidence matching object.

[0237] The core semantic vector of the matched object is bound and associated with the three-dimensional spatial vector to be assigned, forming a unique attribution mapping relationship.

[0238] The binding of the matched object is encapsulated with the attribution instruction, and the target anchor cluster identifier and the three-dimensional spatial vector coordinates are combined into a structured attribution control signaling.

[0239] The structured attribution control signaling is written into the system attribution relationship record table to drive the subsequent work order classification and dispatch process.

[0240] By using minimum value filtering and binding, the path offset numerical sequence from the previous step is transformed into the final target semantic anchor identifier, achieving the expected technical effect of disambiguation and accurate attribution.

[0241] Step S7: Monitor the number of similar work orders from the same source for free nodes that do not fall into any temporary gravity candidate region dataset. If the number of similar work orders from the same source reaches the upgrade threshold within a preset time window, mark the free node as a new semantic anchor and trigger an anchor topology reorganization command. Specifically, this includes:

[0242] S7.1: Obtain the identifiers of free nodes that have not fallen into any temporary gravity candidate region dataset and their corresponding original new work order texts. Use the sliding time window algorithm to scan and filter the full stream of work orders to be processed within the preset time range of the current system time backtracking, so as to generate an initial monitoring dataset containing all potential homogeneous candidate work orders.

[0243] S7.2: Based on each potential candidate work order text in the initial monitoring dataset, call the semantic encoder to perform feature mapping processing, convert the text into a high-dimensional semantic vector, and calculate the Euclidean distance between each high-dimensional semantic vector and the core semantic vector of the free node to generate a distance metric sequence representing the degree of semantic proximity.

[0244] Based on each potential candidate work order text in the initial monitoring dataset, a semantic encoder model specifically designed for the government affairs domain is loaded, and the text is input into the model's forward computation path to generate an embedding representation matrix. The generated embedding matrix is ​​then processed by a hidden layer feature extraction module to capture and fuse context-dependent global semantic components. Feature normalization is applied after extraction to eliminate the influence of text length differences on numerical amplitude. The normalized semantic components are mapped to a high-dimensional latent space, forming a high-dimensional semantic vector with a fixed dimension, which is used as the input element for distance calculation. The core semantic vector corresponding to the detached node is obtained, and same-dimensional calibration is performed to ensure that distance calculations are performed in the same metric space as the high-dimensional semantic vector of the candidate work order. The Euclidean distance calculation unit is invoked to perform Euclidean distance operations between each high-dimensional semantic vector and the core semantic vector of the detached node. All calculated D values ​​are stored in candidate work order index order to generate a distance metric sequence representing the degree of semantic proximity. Through the above processing, the initial monitoring dataset from the previous step is transformed into a standardized distance index that can be directly used for threshold determination, achieving a quantitative representation of the semantic correlation between detached nodes and potential candidate work orders.

[0245] S7.3: Receive the distance measurement sequence and perform threshold determination processing. Mark the distance values ​​that are less than the preset homogeneity similarity threshold as valid matches. Count the cumulative number of valid matches to generate a homogeneous similar work order quantity index, which serves as a quantitative basis for determining whether a free node is eligible for upgrade.

[0246] S7.4: Perform a comparison operation based on the number of similar work orders from the same source and the preset upgrade threshold. If it is determined that the number of similar work orders from the same source reaches or exceeds the preset upgrade threshold, change the state attribute of the free node from the temporary observation state to the permanent anchor state, and generate a new anchor object carrying the new semantic anchor identifier and the core semantic vector coordinates.

[0247] It receives the number of similar work orders from the same source as input data for quantitative judgment, and simultaneously loads a constant configuration of a preset upgrade threshold as a reference benchmark for comparison calculation.

[0248] The number of similar work orders from the same source is input into the comparison calculation module, and the difference is calculated with the preset upgrade threshold to generate a difference parameter. The sign and magnitude of the difference are used to determine whether the current free node has met the upgrade conditions.

[0249] A logical decision function is used to map the difference parameter to a binary state identifier, where a positive value or zero value corresponds to a state that has reached or exceeded the upgrade threshold, and a negative value corresponds to a state that has not met the threshold, thereby clarifying the triggering conditions for changes in the state attributes of detached nodes.

[0250] If the upgrade threshold is reached or exceeded, the state management subsystem is invoked to update the state attribute field of the detached node from the temporary observation state to the permanent anchor state, and a unique new semantic anchor identifier is bound as a subsequent reference index during the state update process.

[0251] The core semantic vector coordinates of the free nodes in the high-dimensional semantic space are extracted in parallel and encapsulated into the data structure of the newly added anchor object. Together with the new semantic anchor identifier, they constitute the output package of the newly added anchor object with complete spatial features.

[0252] Through the above comparison operation and state update processing, the quantitative indicators of the previous step are transformed into trigger signals and anchor space data for the generation of new semantic anchors, thereby achieving accurate upgrade processing of free nodes.

[0253] S7.5: In response to the generation event of the newly added anchor point object, construct a structured control signaling containing the new semantic anchor point identifier, generation timestamp and trigger source information, execute protocol encapsulation operation to generate anchor point topology reconfiguration instruction, so as to drive the subsequent local position recalibration process of the original active anchor point set.

[0254] Step S8: In response to the anchor point topology reorganization instruction, based on the semantic stability scores of the three anchor points in the original active anchor point set that are semantically closest to the new semantic anchor point, the self-vector positions of these three anchor points are adjusted by weighting, generating an updated set of active anchor points after local recalibration. Specifically, this includes:

[0255] S8.1: Based on the anchor point topology reorganization instruction and the newly generated semantic anchor point, perform a high-dimensional spatial proximity retrieval operation on the original active anchor point set, calculate the Euclidean distance between the new semantic anchor point and all original anchor points, sort the original anchor point instances from smallest to largest according to the distance values, generate a subset of anchor points to be reorganized containing the identifiers of these three original anchor points and their corresponding vector coordinates, and provide clear targets for subsequent local topology adjustments.

[0256] S8.2: For each original anchor instance in the subset of anchor points to be reorganized, extract the semantic stability score value accumulated during its historical evolution, input the semantic stability score value into the normalization processing module to perform linear mapping transformation, eliminate the difference in dimensions and convert it into a weight coefficient with a value range between zero and one, generate a dynamic adjustment weight parameter corresponding one-to-one with each original anchor instance, and establish the influence level of each anchor in the reorganization process.

[0257] S8.3: Based on the dynamically adjusted weight parameters and the core semantic vector coordinates of the new semantic anchor point, a vector displacement correction model is constructed. The difference between the current vector coordinates of each original anchor point and the core semantic vector coordinates of the new semantic anchor point is calculated to obtain the initial displacement vector. Then, the initial displacement vector is multiplied by the corresponding dynamically adjusted weight parameters to generate a weighted displacement component that represents the tendency of a single original anchor point to move closer to the new semantic anchor point, thereby realizing the differentiated position fine-tuning based on credibility.

[0258] Based on the normalized dynamically adjusted weight parameters and the core semantic vector coordinates of the new semantic anchor, the system receives the current vector coordinates of each original anchor in the anchor subset to be recombined, calls the high-dimensional space difference operation module to perform coordinate difference processing, and subtracts the current vector coordinates of the original anchor from the core semantic vector coordinates of the new semantic anchor component by component to generate the initial displacement vector.

[0259] The initial displacement vector is input into the scalar multiplication execution unit, and the dynamically adjusted weight parameter corresponding to the original anchor point is called as the multiplication coefficient. The component-by-component multiplication process is performed to generate a structured weighted displacement component object, which represents the displacement trend strength of the original anchor point moving towards the new semantic anchor point.

[0260] Based on the mathematical vector displacement correction model, the displacement correction formula is constructed as follows:

[0261]

[0262] Where Δ is the weighted displacement component, w is the dynamically adjusted weight parameter, P is the coordinates of the core semantic vector of the new semantic anchor point, and Q is the current vector coordinates of the original anchor point.

[0263] The calculated weighted displacement components are stored in an intermediate buffer and used as direct input data for subsequent vector superposition updates of the original anchor point absolute position coordinates.

[0264] By using the above displacement correction processing method, the distance difference generated in the previous step is transformed into a spatial displacement component after dynamic weight modulation, thereby achieving a differentiated position fine-tuning effect based on semantic stability credibility.

[0265] S8.4: Perform vector superposition processing on the weighted displacement components generated by the three original anchor points in the anchor point subset to be recombined. Perform algebraic addition on the original vector coordinates of each original anchor point and its corresponding weighted displacement component. Update the absolute position coordinates of each original anchor point in the high-dimensional semantic space. Generate an intermediate anchor point vector set after local gravity traction correction. Complete the spatial position recalibration of the original anchor points relative to the new semantic anchor points.

[0266] The system receives the weighted displacement components and corresponding original vector coordinates of the three original anchor points in the subset of anchor points to be recombined. It then calls the vector operation module to perform algebraic addition on the weighted displacement components of each original anchor point and its current vector coordinates, forming a preliminary result vector for position correction. Following the high-dimensional semantic space coordinate system, the output of this algebraic addition is used as an intermediate position matrix. Position normalization is performed on each row of the matrix to ensure that the corrected vectors maintain a uniform dimension within a preset spatial scale. Based on the normalized intermediate position matrix and the core semantic vector coordinates of the new semantic anchor points, the relative displacement change rate between each original anchor point and the new anchor point is calculated. If the change rate exceeds a preset upper limit, restrictive scaling is performed to prevent distortion of the local topology due to large displacements. After completing the change rate calibration, the corrected vector coordinates are written into the intermediate anchor point vector set data structure, providing stable and usable anchor point position information for subsequent full set merging. By using this vector superposition and normalization, and rate of change calibration processing method, the weighted displacement components of the previous step are transformed into a set of intermediate anchor point vectors that can be directly used for semantic gravitational field updates, thereby realizing the spatial recalibration of the original anchor points relative to the new semantic anchor points and maintaining the stability of the local topological structure.

[0267] S8.5: Integrate the original anchor instances that have completed position updates and other active anchor instances that did not participate in this reorganization in the intermediate anchor vector set, perform a full set merging operation, refresh the vector storage data in the global anchor index table, and generate an updated active anchor set with the latest topology to ensure that the semantic gravity field can reflect the latest business demand distribution characteristics in real time and support subsequent work order classification tasks.

[0268] It receives the intermediate anchor point vector set after local gravity traction correction and other active anchor point instances that did not participate in this reorganization. It calls the full set merging module to perform data-level merging processing on the two types of objects, and encapsulates their respective vector coordinates, dynamic attribute parameters and anchor point identifiers into the same set structure.

[0269] A global index refresh operation is performed on the merged set. The high-dimensional semantic space index building function is called to regenerate the global anchor index table. The location information mapping of each anchor instance is established in the index table according to its updated vector coordinates. At the same time, the dynamic attribute fields recorded in the index table are updated to reflect the latest stability score, time-depletion factor and channel adaptation weight.

[0270] The updated global anchor index table is subjected to consistency verification. The data integrity verification logic is called to check the uniqueness of each anchor identifier in the collection, and the range checks are performed on the number of vector dimensions and the number of dynamic attribute fields to ensure the structural integrity and content validity of the entire collection.

[0271] The verified global anchor point index table is output to the semantic gravity field construction module, triggering the full recalculation process of the semantic gravity potential field. The latest updated anchor point spatial distribution features are then applied to subsequent work order classification and judgment tasks, realizing the dynamic reflection of the real-time business demand distribution status of the semantic gravity field.

[0272] By merging the entire set and refreshing the global index, the local position update results of the previous step are transformed into anchor point topology data that is consistent, complete and available in real time in the global semantic space, so as to achieve accurate support for the classification gravitational field of the updated active anchor point set.

[0273] For example, in the operating environment of the Ganzhou Municipal Government Service Hotline platform, the intermediate anchor point set after local gravity-driven correction contains three original anchor points with updated positions, each with a vector dimension of 384. There are 97 active anchor point instances that did not participate in the reorganization. The full set merging module integrates the three updated anchor points with the 97 non-updated anchor points to generate a full set containing 100 instances. Encapsulated fields include anchor point ID, vector coordinates, stability score, time-degradation factor, and channel adaptation weight. During the index refresh process, the KD-Tree spatial index construction method is used, with the 384-dimensional vector as spatial coordinate input to establish an index table retrieved by Euclidean distance. The stability score field, after normalization, ranges from 0 to 1. The time-degradation factor is calculated according to the natural day exponential decay model, with a decay rate of 0.92. The channel adaptation weight is normalized to a probability distribution based on cross-channel matching frequency. The index table consistency check results show that all anchor point IDs are unique, the vector dimension meets the 384-dimensional requirement, and the dynamic attribute fields are complete. During the semantic gravitational field recalculation, the gravitational potential energy value of each anchor point is weighted according to the following formula:

[0274]

[0275] Where S is the stability score, T is the time-decrease factor, and C is the channel adaptation weight. This weight, along with the anchor point position, is used to construct the Gaussian kernel of the potential field, resulting in a new gravitational field that significantly improves the stability and accuracy of clustering complex claims in real-time classification.

[0276] In one embodiment, a multi-channel public opinion intelligent aggregation and classification system is provided to execute a multi-channel public opinion intelligent aggregation and classification method described in any of the above embodiments.

[0277] This application provides a method for intelligent aggregation and classification of public opinion demands from multiple channels, which can be deployed in a computing environment consisting of terminals, servers, and a network. The terminals communicate with the servers via the network. The terminals can be, but are not limited to, various personal computers, laptops, smartphones, tablets, and portable wearable devices. The servers can be independent servers or server clusters composed of multiple servers.

[0278] In one embodiment, a computer device is provided, which may be a server. The computer device includes a processor, memory, a network interface, and a database connected via a system bus. The processor provides computing and control capabilities. The memory includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores an operating system, computer programs, and a database. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The database stores data generated during the implementation of a method for intelligent aggregation and classification of public opinion from multiple channels. The network interface communicates with external terminals via a network connection. When the computer program is executed by the processor, it implements a method for intelligent aggregation and classification of public opinion from multiple channels.

[0279] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements a method for intelligent aggregation and classification of multi-channel public opinion demands as described in any of the above embodiments.

[0280] 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.

[0281] 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.

[0282] 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 aggregation and classification of public opinion demands from multiple channels, characterized in that, include: S1: Obtain all closed-loop work order data from the channel public opinion appeal system, filter out appeal instances that meet preset conditions in terms of repetition rate, handling consistency rate, and cross-channel recurrence frequency, and generate an initial set of semantic anchor points. S2: Based on the consistency of attention distribution, natural day time span and channel matching frequency of each instance in the initial semantic anchor set in different channel texts, calculate and associate semantic stability score, time decay factor and channel adaptation weight to generate an active anchor set with dynamic attributes. S3: Obtain the newly added work order text and perform feature mapping processing through a preset semantic encoder to generate a three-dimensional space vector representing the semantic features of the newly added work order text; S4: Perform a weighted response function operation on the three-dimensional spatial vector and the dynamic attributes in the set of active anchor points in sequence to calculate the semantic attraction strength value that characterizes the association strength between the new work order and each anchor point; S5: Select the top five anchor points according to the sorting results of the semantic gravity intensity values ​​to form a temporary gravity candidate region dataset, and set the influence radius according to the reverse order of gravity intensity with each candidate anchor point as the virtual center, and superimpose to generate a composite gravitational potential energy field for defining the belonging range. S6: Determine the landing position of the three-dimensional spatial vector in the composite gravitational potential energy field. If the three-dimensional spatial vector falls into the potential energy valley, it is directly assigned to the corresponding anchor point cluster. If the three-dimensional spatial vector is in the transition zone of multiple potential energy overlaps, perform dependency path offset analysis to generate the target semantic anchor point to be finally assigned.

2. The method for intelligent aggregation and classification of public opinion demands from multiple channels as described in claim 1, characterized in that, After step S6, the following is included: S7: For free nodes that do not fall into any of the temporary gravity candidate region datasets, monitor the number of similar work orders from the same source. If the number of similar work orders from the same source reaches the preset upgrade threshold within the preset time window, mark the free node as a new semantic anchor and trigger the anchor topology reorganization instruction. S8: In response to the anchor point topology reorganization instruction, based on the semantic stability scores of the three anchor points in the original active anchor point set that are closest to the new semantic anchor point, the self-vector positions of the three anchor points are adjusted by weighting, and an updated set of active anchor points after local recalibration is generated.

3. The method for intelligent aggregation and classification of public opinion demands from multiple channels as described in claim 1, characterized in that, The initial set of semantic anchors includes standardized expressions and their core semantic vectors.

4. The method for intelligent aggregation and classification of public opinion demands from multiple channels according to claim 3, characterized in that, Step S2 specifically includes: Obtain the attention distribution matrix of each instance in the initial semantic anchor set in different channel texts, perform variance analysis and entropy measurement on the attention distribution matrix of the different channel texts, and generate the original data of semantic stability score; A timestamp is created based on the instance corresponding to the original data of the semantic stability score. The time span variable is nonlinearly mapped using the natural day exponential decay model to generate a numerical sequence of time decay factors that automatically reduce their weight over time. The percentage of successful historical matching times of each instance in the initial semantic anchor point set across multiple channels is statistically analyzed. Based on frequency normalization, the percentage of successful historical matching times of each instance across multiple channels is transformed into a probability distribution form, generating channel adaptation weight parameters that reflect cross-channel adaptability. Receive the original data of the semantic stability score, the numerical sequence of the time-effect decay factor and the channel adaptation weight parameters, perform multi-dimensional attribute association and encapsulation operation, and bind them to the corresponding standardized expression and its core semantic vector to generate an active semantic anchor object. Iterate through all the generated active semantic anchor objects, and remove invalid nodes whose stability scores are below the critical value or whose time decay factors are close to zero according to the preset health threshold filtering rules, and output the active anchor set.

5. The method for intelligent aggregation and classification of public opinion demands from multiple channels according to claim 4, characterized in that, The multiple channels include: telephone, SMS and APP.

6. The method for intelligent aggregation and classification of public opinion demands from multiple channels according to claim 3, characterized in that, Step S4 specifically includes: Obtain the time-effect decay factor from the three-dimensional spatial vector and the set of active anchor points, and perform exponential decay update processing on the time-effect decay factor based on the natural day time span to generate a real-time time-effect weight parameter that reflects the freshness of anchor points under the current time window. Receive the real-time timeliness weight parameter and the channel identifier to which the new work order text belongs, and use the channel matching degree mapping table to perform dynamic calibration processing on the channel adaptation weight to generate a corrected channel adaptation coefficient; Input the three-dimensional spatial vector, the core semantic vector of the active anchor point, and the modified channel adaptation coefficient, perform basic semantic distance calculation based on cosine similarity calculation, and generate a basic similarity score that represents the degree of similarity of text content; The basic similarity score, real-time timeliness weight parameters, and corrected channel adaptation coefficient are obtained. Based on the multi-dimensional weighted fusion mechanism, a nonlinear response function operation is performed to generate an original gravity strength value that comprehensively reflects semantic content, timeliness, and channel matching degree. The original gravity intensity value is received, and dimensional unification and boundary constraint processing are performed based on normalization scaling to generate the final semantic gravity intensity value that represents the association strength between the new work order and each anchor point.

7. The method for intelligent aggregation and classification of public opinion demands from multiple channels as described in claim 6, characterized in that, Step S5 specifically includes: The semantic gravity strength value is obtained, and the original gravity strength values ​​corresponding to all anchors in the set of active anchors are sorted in descending order using a quick sorting algorithm to generate a ranking list of anchor gravity strength with a clear priority order. Based on the anchor point gravity strength ranking list, a top truncation filtering operation is performed to extract anchor point instances whose ranking is within a preset threshold range and their corresponding core semantic vector coordinates and dynamic attribute parameters, thereby generating a temporary gravity candidate region dataset. Receive each candidate anchor point in the temporary gravity candidate region dataset, perform inverse proportional mapping calculation based on its semantic gravity intensity value, convert the original gravity intensity value into an initial influence radius parameter characterizing the breadth of the action range, and generate a dynamic action radius value sequence corresponding to each candidate anchor point. Based on the dynamic action radius numerical sequence and the core semantic vector coordinates of each candidate anchor point, a multidimensional Gaussian kernel function distribution model with the candidate anchor point as the center is constructed. The radial basis function interpolation algorithm is used to transform the discrete initial influence radius parameter into a continuous spatial potential energy decay curve, generating an independent potential energy distribution matrix. A linear superposition operation is performed on the independent potential energy distribution matrix to fuse the potential energy contribution values ​​of each anchor point in the three-dimensional semantic space, thereby forming the composite gravitational potential energy field.

8. A multi-channel public opinion intelligent aggregation and classification system, used to execute the multi-channel public opinion intelligent aggregation and classification method as described in any one of claims 1 to 7.

9. A computer device, characterized in that, include: A processor, a memory, and a computer program stored in the memory and executable on the processor; when executed by the processor, the computer program implements the steps of a method for intelligent aggregation and classification of multi-channel public opinion demands as described in any one of claims 1 to 7.

10. A storage medium, characterized in that: The storage medium stores a computer program, which, when executed by a processor, implements the steps of a method for intelligent aggregation and classification of multi-channel public opinion demands as described in any one of claims 1 to 7.

Images