A spreading trend quantitative identification system and method for time-series text

Abstract

The application relates to the technical field of natural language processing, in particular to a spreading trend quantitative identification system and method for time-series text; the method comprises the following steps: performing semantic coding on time-series text records, merging text increments with a semantic distance less than an aggregation threshold into semantic micro clusters according to a sliding time window, and recording semantic micro cluster generation time, semantic barycenter coordinates and member scale; traversing the semantic micro clusters in pairs according to the generation time, establishing a forward directed connection when the semantic barycenter distance is less than a correlation threshold and the time interval falls into a time-series correlation window, and generating a micro cluster topology correlation graph; extracting weakly connected subgraphs, aggregating the aggregation scale and calculating the diffusion rate; calculating the product of the aggregation scale and the diffusion rate to obtain a cascade trend intensity value, marking the weakly connected subgraphs with a cascade trend intensity value exceeding a perception threshold as cascade trend targets and outputting the corresponding relationship. The application realizes early perception of sheet diffusion trend by taking into account the volume and the spreading speed.

Description

Technical Field

[0001] This application relates to the field of natural language processing technology, specifically to a system and method for quantitative recognition of the spread of temporal text. Background Technology

[0002] Existing time-series text analysis systems typically categorize and statistically analyze text records within a fixed statistical period, determining whether texts require attention based on whether the number of text entries in each category exceeds a preset threshold within a single statistical period. However, in real-world scenarios, the spread of a type of text spans multiple statistical periods. In the initial stages of this spread, the number of text entries in each statistical period does not reach the threshold, and the system does not mark this type of text in any of these periods, even though it is in a state of continuous spread over time.

[0003] For example, in a telephone hotline complaint handling scenario, a sudden water supply network failure occurred in a certain area, and residents in the surrounding area reported water outages via telephone and mobile devices within 40 minutes. Using 10-minute intervals as statistical periods, the first 10 minutes saw 3 water outage complaints, the second 10 minutes saw 5, the third 10 minutes saw 4, and the fourth 10 minutes saw 6, totaling 18 complaints all reporting water outages in the same area, showing a continuous spread in terms of arrival time. However, within each 10-minute statistical period, the number of water outage complaints was only 3 to 6, not reaching the system's preset threshold, and the system did not mark the water outage complaints as matters requiring attention in the statistical results of any of the four periods. As another example, a type of complaint about parking obstruction that had existed for several months reached the threshold in a certain statistical period and was marked as a matter of concern, but the arrival of these parking obstruction complaints was more dispersed over time, and their urgency was lower than that of the rapidly spreading water outage complaints.

[0004] If the spread of text across multiple statistical periods, where the number of entries within each period is below a threshold but the overall trend is rapidly expanding, will not be detected in a timely manner, resulting in delayed discovery of large-scale, urgent issues requiring immediate response. The perception of text spread trends lags behind the actual spread process, impacting the timely handling of emergencies. Therefore, it is necessary to address the problem of perceiving large-scale, cross-period spread patterns in time-series text records where the number of entries within a single period is below a threshold.

[0005] In view of this, this application proposes a quantitative identification system and method for the spread of time-series text. Summary of the Invention

[0006] To address the problems existing in the background technology, this application provides a system and method for quantitative recognition of the spread of information in time-series text, and the specific technical solution is as follows: Firstly, this application provides a method for quantitatively identifying the spread of information in time-series text, including: Semantic encoding is performed on time-series text records. Text increments with semantic distance less than the aggregation threshold are merged into semantic micro-clusters by sliding time window. The generation time, semantic centroid coordinates, and member size of the semantic micro-clusters are recorded. Traverse the semantic micro-clusters pairwise according to their generation time, calculate the semantic centroid distance and time interval, and establish a forward directed connection when the distance is less than the association threshold and the interval falls within the temporal association window to generate a micro-cluster topology association graph. Extract the weakly connected subgraphs of the micro-cluster topology graph, summarize the size of the semantic micro-cluster members in the weakly connected subgraphs to obtain the aggregation size, and use the ratio of the number of semantic micro-cluster nodes in the weakly connected subgraphs to the value obtained by adding the minimum time granularity to the first and last generation time interval as the diffusion rate. The product of aggregation scale and diffusion rate is used to obtain the cascade situation strength value. Weakly connected subgraphs whose cascade situation strength values ​​exceed the perception threshold are marked as cascade situation targets. The correspondence between the cascade situation strength value, the semantic centroid coordinates within the target, and the member text feature codes is output.

[0007] Preferably, the semantic encoding includes: text normalization of the text in the time-series text record, wherein text normalization includes removing expressions unrelated to the text semantics, standardizing the spelling of proper nouns and merging repeated punctuation; inputting the normalized text into the semantic encoding model and mapping it into a semantic vector of fixed dimension; assigning a text feature code to each text, wherein the text feature code is obtained by transcribing the work order number, and the text feature code is saved together with the semantic vector and the acceptance time.

[0008] Preferably, the generation of the micro-cluster topology graph includes: the direction of the forward directed connection is from the semantic micro-cluster generated earlier to the semantic micro-cluster generated later; when establishing the forward directed connection, the semantic centroid distance and time interval are recorded as attributes of the forward directed connection; the micro-cluster topology graph is stored in the form of an adjacency list, and for each semantic micro-cluster node, the subsequent semantic micro-cluster nodes pointed to by the semantic micro-cluster node and the attributes of the forward directed connection are recorded.

[0009] Preferably, the forward directed connection includes: no forward directed connection is established between two semantic microclusters with the same generation time; a semantic microcluster with an earlier generation time establishes forward directed connections to multiple semantic microclusters with later generation times respectively; multiple semantic microclusters with earlier generation times establish forward directed connections to a semantic microcluster with a later generation time respectively.

[0010] Preferably, the temporal correlation window includes: the temporal correlation window is defined by a lower limit and an upper limit, the lower limit of the temporal correlation window is a sliding step size; the upper limit of the temporal correlation window is a set quantile of the time interval of the homologous semantic micro-clusters in history.

[0011] Preferably, the step of traversing semantic microclusters pairwise according to their generation time includes: when traversing semantic microclusters pairwise after sorting by generation time, comparing semantic microclusters whose generation time interval falls within the span of the temporal association window; and not calculating the semantic centroid distance between semantic microclusters whose generation time interval exceeds the upper limit of the temporal association window.

[0012] Preferably, the association threshold includes: selecting semantic micro-clusters that have been historically determined to belong to the same event, forming homologous semantic micro-cluster pairs in pairs; calculating the semantic centroid distance of each homologous semantic micro-cluster pair to obtain a set of semantic centroid distances for homologous semantic micro-cluster pairs; and taking a set quantile of the set of semantic centroid distances for homologous semantic micro-cluster pairs as the association threshold.

[0013] Preferably, the extraction of weakly connected subgraphs from the micro-cluster topological association graph includes: treating forward directed connections in the micro-cluster topological association graph as connections without distinguishing directions, obtaining connected components in the micro-cluster topological association graph, with each connected component being a weakly connected subgraph; and using a disjoint-set data structure to gradually merge semantic micro-clusters connected by forward directed connections, traversing each forward directed connection and merging the semantic micro-clusters at both ends of the connection into the same set, with the semantic micro-clusters belonging to the same set constituting a weakly connected subgraph after the traversal.

[0014] Preferably, the weakly connected subgraph includes: the interval between the first and last generation times within the weakly connected subgraph is the difference between the generation times of the earliest and latest semantic micro-clusters within the weakly connected subgraph; the minimum time granularity is the recording precision of the text acceptance time in the time-series text record; the interval between the first and last generation times and the minimum time granularity are measured in the same time unit.

[0015] Preferably, the perception threshold includes: collecting the cascaded situation strength values ​​of all historically weakly connected subgraphs to obtain a historical cascaded situation strength value set; taking a set quantile of the historical cascaded situation strength value set as the perception threshold; and re-taking the set quantile of the historical cascaded situation strength value set at a preset period to update the perception threshold.

[0016] Preferably, the correspondence includes: mapping semantic centroid coordinates to member texts with the smallest semantic distance between semantic vectors and semantic centroid coordinates, which serve as the subject representation of cascaded situation targets; the correspondence is given in the form of structured records, each record corresponding to a cascaded situation target, and the record contains cascaded situation strength value, semantic centroid coordinates of each semantic micro-cluster, and text feature codes of each member text; when there are multiple cascaded situation targets marked, the cascaded situation targets are sorted from high to low according to the cascaded situation strength value to form an order of priority of attention.

[0017] Secondly, this application also provides a system for quantitative identification of sprawling situation in time-series text, which is used to implement the aforementioned method for quantitative identification of sprawling situation in time-series text, including: a semantic micro-cluster generation module, a topological association construction module, a subgraph quantitative extraction module, and a situation target marking module; The semantic micro-cluster generation module is used to perform semantic encoding on time-series text records, merge text increments with semantic distance less than the aggregation threshold into semantic micro-clusters according to a sliding time window, and record the semantic micro-cluster generation time, semantic centroid coordinates, and member size. The topology association construction module is used to traverse semantic microclusters pairwise according to the order of generation time, calculate the semantic centroid distance and time interval, and establish a forward directed connection when the distance is less than the association threshold and the interval falls within the temporal association window to generate a microcluster topology association graph. The subgraph quantization extraction module is used to extract weakly connected subgraphs of the micro-cluster topology graph, summarize the size of semantic micro-cluster members in the weakly connected subgraph to obtain the aggregation size, and use the ratio of the number of semantic micro-cluster nodes in the weakly connected subgraph to the value obtained by adding the minimum time granularity to the first and last generation time interval as the diffusion rate. The situation target marking module is used to calculate the product of aggregation scale and diffusion rate to obtain the cascade situation strength value, mark weakly connected subgraphs whose cascade situation strength value exceeds the perception threshold as cascade situation targets, and output the correspondence between the cascade situation strength value, the semantic centroid coordinates within the target, and the member text feature codes.

[0018] The beneficial effects of this application are: This application performs semantic encoding on time-series text records and merges them into semantic micro-clusters by sliding time window increments. By merging texts in real time according to semantic distance into basic processing units carrying generation time, semantic centroid coordinates and member size, it realizes online aggregation of texts arriving one by one according to semantic meaning.

[0019] This application traverses semantic microclusters pairwise according to their generation time and establishes forward directed connections when distance and interval conditions are met. It generates a microcluster topology graph through the dual constraints of semantic centroid distance and time interval, thereby realizing the topological association of homogeneous semantic microclusters that appear across time periods along the time direction.

[0020] This application extracts the weakly connected subgraphs of the micro-cluster topology graph and calculates the aggregation scale and diffusion rate. It quantifies the spread speed by the ratio of the number of nodes to the interval between the first and last generation times plus the minimum time granularity, thus achieving a frame-based measurement of the potential event size and spread speed.

[0021] This application calculates the product of aggregation scale and diffusion rate to obtain cascade situation strength value and marks weakly connected subgraphs whose cascade situation strength value exceeds the perception threshold. By comparing the cascade situation strength value with the perception threshold, cascade situation targets are screened, thereby realizing quantitative screening and traceability of large-scale diffusion situations.

[0022] The technical solution of this application converges the scattered texts in the time-series text into quantifiable cascaded situation strength values ​​through semantic micro-clusters, topological associations and weakly connected subgraphs. It takes into account both the volume and the spread speed to perceive the spread situation of the whole area, and solves the problem that the spread situation of the time-series text that spreads across time periods and whose number in a single time period does not reach the threshold is delayed in detection. Attached Figure Description

[0023] Figure 1 This is a flowchart of a method for quantitatively identifying the spread of temporal text, as described in this application. Figure 2 This is a schematic diagram illustrating the semantic micro-cluster structure of this application; Figure 3 This is a schematic diagram illustrating the incremental incorporation within the sliding time window of this application. Figure 4 This is a schematic diagram illustrating the establishment of the forward directed connection in this application; Figure 5 This is a schematic diagram of weakly connected subgraph extraction in this application; Figure 6 This is a schematic diagram of the temporal distribution of semantic micro-clusters within a weakly connected subgraph in this application; Figure 7 This is a schematic diagram illustrating the correspondence between the cascaded situational targets in this application. Figure 8 This is a structural diagram of a temporal text propagation trend quantification and recognition system proposed in this application. Detailed Implementation

[0024] To make the above-mentioned objectives, features and advantages of this application more apparent and understandable, the specific embodiments of this application will be described in detail below with reference to the accompanying drawings.

[0025] Many specific details are set forth in the following description in order to provide a full understanding of this application. However, this application may also be implemented in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of this application. Therefore, this application is not limited to the specific embodiments disclosed below.

[0026] Secondly, the term "an embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of this application. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single embodiment or an embodiment selectively excluded from other embodiments.

[0027] Example 1 Reference Figures 1 to 7 This is the first embodiment of the present application, such as Figure 1As shown, a method for quantitative identification of the spread of time-series text is provided.

[0028] To clarify the purpose and technical solution of this application, this embodiment describes a specific implementation method for quantitatively identifying the spread of time-series text in the context of handling telephone hotline complaints. The object processed in this embodiment is the complaint text records continuously received and stored in the database during the operation of the telephone hotline. The public submits complaints through channels such as telephone, website, mobile application, and mini-program. When the work order system accepts a complaint, it records the complaint text, the acceptance time, and the work order number. All complaint text records are arranged chronologically according to the acceptance time, forming a time-series text record that extends continuously over time.

[0029] The overall processing procedure of this embodiment unfolds sequentially from steps S1 to S4. First, semantic micro-clusters are generated from time-series text records. Then, topological relationships are established between semantic micro-clusters. Subsequently, weakly connected subgraphs are extracted and quantified from the relationship structure. Finally, cascaded situation strength values ​​are calculated and cascaded situation targets are marked. The above processing procedure can be deployed and run on the computing nodes of a cloud server. It can perform incremental processing on newly added appeal texts as the sliding time window progresses, or it can perform a complete processing on all appeal texts at a fixed time each day. This embodiment mainly describes the incremental processing as the sliding time window progresses.

[0030] Step S1: Semantically encode the time-series text records, merge text increments with semantic distance less than the aggregation threshold into semantic micro-clusters by sliding time window, and record the generation time, semantic centroid coordinates and member size of the semantic micro-clusters.

[0031] Step S1 involves generating semantic micro-clusters. The acquisition of time-series text records is relatively straightforward; the work order system has already stored the request text, acceptance time, and work order number in its database. In this implementation, the already stored request text records are read as the input sequence based on the acceptance time, without further restrictions on the acquisition method. Before proceeding to subsequent processing, necessary text standardization is performed on the request text, including removing polite expressions irrelevant to the semantics of the request, standardizing the spelling of proper nouns, and merging duplicate punctuation. Text standardization aims to reduce the interference of literal differences on the subsequent semantic encoding results.

[0032] Text normalization only processes literal components that are irrelevant to the semantics of the appeal, without changing the issue reflected in the appeal itself. The semantic vector obtained by the text normalization of the appeal text in the semantic encoding stage is less affected by differences in expression habits, so that appeal texts that reflect the same issue but are expressed in different ways can still be merged into the same semantic micro-cluster.

[0033] To aggregate requests based on their meaning rather than their literal wording, semantic encoding is required for each request text. Semantic encoding employs models such as the BGE (BAAI General Embedding) series or the M3E (Moka Massive Mixed Embedding) series, mapping the standardized request text into a fixed-dimensional semantic vector. The fixed-dimensional semantic vectors output by the semantic encoding model ensure that semantic vectors corresponding to texts with similar meanings are close together in the vector space, while semantic vectors corresponding to texts with significantly different meanings are far apart, thus transforming the comparison of text meanings into the calculation of distances between semantic vectors. Simultaneously, a text feature code is assigned to each request text during semantic vector generation. This text feature code can be directly transcribed from the work order number. The text feature code, along with the semantic vector and the time of acceptance, is saved as an index for subsequent tracing back to the original request work order. For example, the character sequence of the work order number can be directly used as the text feature code.

[0034] Considering that hotline requests are primarily short texts, semantic encoding generates a semantic vector for each entire request text, rather than generating vectors word by word and then concatenating them. The semantic vector generated from the entire request text reflects the overall meaning of the request, facilitating the clustering of requests based on meaning. The dimension of the semantic vector remains fixed after the semantic encoding model is determined; the semantic vectors of different request texts reside in the same vector space, and semantic distance is calculated in the same dimension. The text feature code remains unique within the same batch of request texts. The text feature code transcribed from the work order number corresponds one-to-one with the request work order, allowing for tracing back to the specific original request work order based on the text feature code.

[0035] After obtaining the semantic vectors of the appeal texts, incremental merging of the appeal texts is performed according to a sliding time window (hereinafter referred to as the time window). A fixed-length time window and a fixed-length sliding step are pre-defined. Each time the time window slides one step along the time axis, new appeal texts falling within the current time window are processed one by one. During processing, a set of currently active semantic micro-clusters is maintained. For each new appeal text falling within the current time window, the semantic distance between the semantic vector of the new appeal text and the semantic centroid coordinates of each active semantic micro-cluster is calculated sequentially, and the active semantic micro-cluster with the smallest semantic distance is selected as a candidate semantic micro-cluster. Semantic distance is used to measure the semantic similarity between two appeal texts, and can be measured using cosine distance. The corresponding calculation formula can be expressed as: In the formula For the first Article 1 and Article 2 The semantic distance between the semantic vectors of the claim texts For the first The semantic vector obtained by semantically encoding the text of the appeal For the first The semantic vector obtained by semantically encoding the text of the appeal and To determine the sequence number of the requested text within the chronological text record. This represents the inner product of two semantic vectors. This represents the magnitude of the semantic vector.

[0036] like Figure 3 As shown, Figure 3 This is a schematic diagram illustrating the incremental incorporation within a sliding time window. Figure 3 The horizontal arrow at the bottom center represents time, and the dashed rectangle above it represents a sliding time window. The arrow above the rectangle indicates that the window slides along the time direction. The two circles inside the window represent currently active semantic micro-clusters, and the dot inside the circle represents its semantic center of gravity. The solid dot on the right side of the window represents newly added appeal text, and the arrow indicates that it has been merged into the semantic micro-cluster with the closest semantic center of gravity. The dashed circle on the left side of the window with an earlier time indicates a semantic micro-cluster that has been removed from the active set but retained. Figure 3 The identifiers include time, sliding time window, sliding, semantic micro-cluster, newly added appeal text, and merging and removing active sets while retaining them.

[0037] When the semantic distance between a candidate semantic micro-cluster and a newly added appeal text is less than the aggregation threshold, the newly added appeal text is merged into the candidate semantic micro-cluster, and the semantic centroid coordinates and member size of the candidate semantic micro-cluster are then updated. When the semantic distance between all active semantic micro-clusters and the newly added appeal text is not less than the aggregation threshold, a new semantic micro-cluster is created separately with the newly added appeal text, and the newly created semantic micro-cluster is added to the set of active semantic micro-clusters. The member size of a semantic micro-cluster is the number of member appeal texts within the semantic micro-cluster, and the generation time of the semantic micro-cluster is the time when the earliest member appeal text within the semantic micro-cluster is received. As the time window continues to slide forward, semantic micro-clusters that have not had any new appeal texts merged into them for a long time are removed from the set of active semantic micro-clusters. After being removed, they no longer participate in the semantic distance comparison of new appeal texts. The removed semantic micro-clusters, along with their generation time, semantic centroid coordinates, and member size, are retained for use in subsequent steps.

[0038] Instead of clustering all appeal texts at once, incremental merging is adopted, allowing semantic micro-clusters to gradually form as appeal texts arrive. New appeal texts only need to be compared with the currently active semantic micro-clusters, without having to compare them pairwise with all historical appeal texts. The number of comparisons required for processing varies with the number of active semantic micro-clusters, and the number of active semantic micro-clusters is less than the total number of historical appeal texts.

[0039] The semantic centroid coordinates of a semantic micro-cluster are taken as the arithmetic mean of the semantic vectors of all member appeal texts within the micro-cluster. To avoid re-accumulating the semantic vectors of all members for each newly incorporated appeal text, the semantic centroid coordinates are updated incrementally. That is, based on the existing semantic centroid coordinates, the arithmetic mean is recalculated by combining the semantic vector of the newly incorporated appeal text with the updated member size to obtain the updated semantic centroid coordinates. Each semantic micro-cluster is stored as a record containing the semantic micro-cluster identifier, generation time, semantic centroid coordinates, member size, and text feature codes of all member appeal texts, facilitating the retrieval of required attributes by semantic micro-cluster in subsequent steps.

[0040] like Figure 2 As shown, Figure 2 This is a schematic diagram of semantic micro-clusters. Figure 2 The outer rectangle represents the semantic vector space, and the solid dots inside the rectangle represent the semantic vectors obtained by semantic encoding of each appeal text. Several semantically similar semantic vectors are clustered in the same area, marked by a dashed envelope as a semantic micro-cluster. The graphic with a cross mark at the center of the envelope represents the semantic centroid of the semantic micro-cluster. In the lower right corner of the rectangle, there is another group of semantic vectors that are close to each other, marked by another dashed envelope as another semantic micro-cluster. This is used to indicate that appeal texts with different semantics are grouped into different semantic micro-clusters. Figure 2 The identifiers include the semantic vector space, the semantic vector of the appeal text, the semantic micro-cluster, the semantic centroid, and another semantic micro-cluster.

[0041] Whether a semantic micro-cluster is retained in the active semantic micro-cluster set is determined by the interval between the time when the semantic micro-cluster last incorporated a member's appeal text and the start of the current time window. Semantic micro-clusters with an interval exceeding a pre-set retention period are removed from the active semantic micro-cluster set. The retention period must be greater than the time span of arrivals within the same type of appeal set. The specific calculation method for the retention period is: the product of the time window length and a set retention multiple, where the retention multiple is an integer greater than 1. For example, if the time window length is 30 minutes and the retention multiple is 3, the retention period is the product of 30 minutes and 3, i.e., 90 minutes, ensuring that the semantic micro-clusters corresponding to a type of appeal remain active during the continuous arrival period. With the cooperation of the sliding time window and the active semantic micro-cluster set, newly added appeal texts are only compared with a limited number of active semantic micro-clusters in terms of semantic distance. Processing is carried out in real time as the appeal arrives, without having to backtrack all historical appeal texts for each processing.

[0042] The aggregation threshold is a value that needs to be preset in step S1. The value of the aggregation threshold is determined based on the statistical distribution of semantic distances between historical hotline call request texts that were manually classified as similar to other types of requests and those classified as dissimilar. The specific calculation method for the aggregation threshold is as follows: Select historically manually classified request texts as similar to other types of requests, pair them together to form similar request text pairs, calculate the semantic distance of each similar request text pair, obtain the semantic distance set of similar request text pairs, and take the set quantile of the semantic distance set of similar request text pairs as the aggregation threshold. The formula for calculating the aggregation threshold can be defined as follows: In the formula The aggregation threshold, This is a set of semantic distances between text pairs with similar claims. Represents taking a set of Quantiles The quantile parameter is a predefined parameter, ranging from 0 to 1. A larger quantile parameter value indicates that more similar claims will be incorporated into the same semantic micro-cluster. For example, the quantile parameter... A threshold of 0.9 corresponds to the 90th percentile of the semantic distance between similar appeal text pairs. This ensures that approximately 90% of similar appeal text pairs have a semantic distance less than the aggregation threshold. Under the cosine distance metric, the aggregation threshold value falls between 0.2 and 0.3. Minor adjustments are then made based on the actual merging effect. The time window length and sliding step size also need to be preset. The time window length must cover the time interval of concentrated arrival of the same batch of appeals from the same source. The specific calculation method for the time window length is as follows: statistically analyze the time span from the arrival of the first to the last appeal in each batch of historical appeals, and take the set quantile of each time span as the time window length. For example, the 90th percentile is used, ensuring that the concentrated arrival interval of most batches of appeals from the same source is covered by a single time window. The corresponding time window length is 30 minutes. The sliding step size determines the temporal resolution at which semantic microclusters are updated. The specific calculation method for the sliding step size is as follows: take the ratio of the time window length to the set number of window divisions. The number of window divisions is an integer greater than 1. For example, when the number of window divisions is 6, the sliding step size is the ratio of 30 minutes to 6, that is, 5 minutes.

[0043] The time window length and the sliding step size work together. The time window length determines the range of appeal texts covered by a single processing step, while the sliding step size determines the amount of time advancement between two adjacent processing steps. When the sliding step size is smaller than the time window length, the appeal texts covered by two adjacent processing steps partially overlap, ensuring that appeals of the same type arriving across the window boundary can still be incorporated into the same semantic micro-cluster, avoiding the splitting of the same type of appeals into different semantic micro-clusters due to window segmentation.

[0044] When the aggregation threshold is too small, semantically similar appeal texts may not be merged into the same semantic micro-cluster because their semantic distance is not less than the aggregation threshold, resulting in the splitting of the same type of appeal into multiple semantic micro-clusters. When the aggregation threshold is too large, semantically different appeal texts may be merged into the same semantic micro-cluster, leading to inconsistent meanings among the appeal texts within the same semantic micro-cluster. Determining the aggregation threshold based on the semantic distance distribution of historical appeal samples, as described above, ensures that the meanings of the appeal texts within a semantic micro-cluster remain similar, while reducing the splitting of the same type of appeal.

[0045] The processing of step S1 is illustrated using a type of heating repair request as an example. For instance, within a certain time window, three request texts reflecting insufficient heating in the same residential area arrive sequentially. The semantic distance between the semantic vectors obtained after semantic encoding of these three request texts is less than the aggregation threshold. Therefore, the three request texts are merged into the same semantic micro-cluster, with a member size of 3. The generation time is taken as the acceptance time of the earliest request text, and the semantic centroid coordinates are taken as the arithmetic mean of the semantic vectors of the three request texts. Within the same time window, another request text reflecting road flooding arrives. The semantic distance between this request text and the semantic centroid coordinates of the heating repair semantic micro-cluster is not less than the aggregation threshold, so a new semantic micro-cluster is created separately.

[0046] Through the processing in step S1, the originally scattered and numerous appeal texts are merged into a smaller number of semantic micro-clusters in real time. The fluctuations caused by a single, occasional appeal are mitigated by averaging the semantic vectors of multiple appeals within the semantic micro-cluster. Each semantic micro-cluster corresponds to a class of semantically similar appeals and carries three attributes: generation time, semantic centroid coordinates, and member size. It becomes a basic processing unit with clear meaning and controllable size, facilitating further analysis of the relationships between semantic micro-clusters on the timeline.

[0047] Step S2: Traverse the semantic micro-clusters pairwise according to the generation time, calculate the semantic centroid distance and time interval, and establish a forward directed connection when the distance is less than the association threshold and the interval falls within the temporal association window to generate a micro-cluster topology association graph.

[0048] After the semantic micro-clusters are generated, step S2 establishes a temporal forward association between them. First, all semantic micro-clusters are sorted from earliest to latest generation time. Then, each semantic micro-cluster is traversed pairwise according to its generation time. For any two semantic micro-clusters involved in the traversal, the semantic micro-cluster generated earlier and the semantic micro-cluster generated later are determined based on their generation time, and two values ​​are calculated. The first value is the semantic centroid distance between the semantic centroid coordinates of the two semantic micro-clusters, calculated using the same semantic distance metric as in step S1. The second value is the time interval between the generation times of the two semantic micro-clusters, calculated by subtracting the generation time of the earlier semantic micro-cluster from the generation time of the later semantic micro-cluster.

[0049] When the semantic centroid distance between two semantic microclusters is less than the association threshold, and the time interval between the two semantic microclusters falls within the temporal association window, a forward directed connection is established from the semantic microcluster generated earlier to the semantic microcluster generated later. A semantic centroid distance less than the association threshold indicates that the claims corresponding to the two semantic microclusters are semantically homologous or similar; a time interval falling within the temporal association window indicates that the two semantic microclusters are sequential in time, and the time interval is within the range consistent with the spread of the claim. The direction of the forward directed connection is from the semantic microcluster generated earlier to the semantic microcluster generated later, representing the direction of semantic spread of the claim along the time axis from the earlier clustering point to the later clustering point. Simultaneously with establishing the forward directed connection, the corresponding semantic centroid distance and time interval are recorded as attributes of the forward directed connection for subsequent verification and recalculation.

[0050] like Figure 4 As shown, Figure 4 This is a schematic diagram illustrating the establishment of a forward directed connection. Figure 4 The two circles on the left and right represent the semantic micro-clusters generated earlier and later, respectively, and the dots inside the circles represent the semantic centroids. The arrows pointing from the earlier semantic micro-clusters to the later semantic micro-clusters represent forward directed connections, and the semantic condition marked above the connection is that the distance between the semantic centroids is less than the association threshold. Figure 4 The horizontal arrows below represent time. The two semantic micro-clusters are mapped to two generation moments on the time axis by the vertical dashed lines. The light-colored strip between them represents the temporal correlation window, which indicates that the generation moment interval falls into the temporal correlation window. Figure 4 The identifiers include semantic microclusters generated earlier, semantic microclusters generated later, forward directed connections, semantic centroid distances less than the association threshold, and temporal association windows.

[0051] Establishing forward-directed connections between semantic microclusters, rather than directionless connections, is to preserve the propagation direction of semantic appeals along the time axis. Appeals spread from earlier clusters to later clusters; directional information reflects the temporal evolutionary order of a class of appeals. Preserving directional information allows subsequent steps to determine the sequential relationship between connected semantic microclusters along the connection direction, providing a basis for subsequent processing based on this sequential relationship. If directionless connections are used, the sequential information between semantic microclusters cannot be preserved.

[0052] Because the direction of forward directed connections is uniformly from the semantic micro-cluster generated earlier to the semantic micro-cluster generated later, and the connection direction is consistent with the order of generation, there will be no contradictory connections between semantic micro-clusters, nor will directed loops be formed. Under the condition that the connection is met, a semantic micro-cluster generated earlier can establish forward directed connections to multiple semantic micro-clusters generated later, and multiple semantic micro-clusters generated earlier can also simultaneously establish forward directed connections to a semantic micro-cluster generated later, thus reflecting the branching and convergence of the same type of claims on the timeline.

[0053] For two semantic microclusters generated at the same time, since the order of their generation cannot be distinguished, no forward directed connection is established between the two semantic microclusters, thus avoiding connections with uncertain directions. Two semantic microclusters generated at the same time but with a semantic centroid distance less than the association threshold can still establish forward directed connections with other semantic microclusters generated earlier or later, and are within the scope of mutual association through the commonly connected semantic microclusters.

[0054] Considering that the temporal association window has an upper limit on the time interval, when performing pairwise traversal of semantic microclusters sorted by generation time, only semantic microclusters whose generation time interval falls within the span of the temporal association window need to be compared. Semantic microclusters whose generation time interval exceeds the upper limit of the temporal association window are no longer calculated for semantic centroid distance. With the above restriction, the comparison range of pairwise traversal is narrowed to adjacent time periods, reducing unnecessary semantic centroid distance calculations.

[0055] The association threshold and temporal association window need to be preset. The value of the association threshold is determined based on the statistical distribution of the semantic centroid distance between semantic microclusters historically identified as belonging to the same source. The specific calculation method of the association threshold is as follows: Select semantic microclusters historically identified as belonging to the same potential appeal event, form homologous semantic microcluster pairs in pairs, calculate the semantic centroid distance of each homologous semantic microcluster pair, obtain the set of semantic centroid distances of homologous semantic microcluster pairs, and take the set quantile of the set of semantic centroid distances of homologous semantic microcluster pairs as the association threshold. The formula for calculating the association threshold can be defined as follows: In the formula For the correlation threshold, This is the set of semantic centroid distances for homologous semantic micro-clusters. Represents taking a set of Quantiles The quantile parameter is a pre-defined parameter, ranging from 0 to 1. The semantic centroid coordinates are the average of multiple claim semantic vectors within a semantic micro-cluster. While semantic micro-clusters may be semantically similar, they belong to different semantic micro-clusters, and their semantic centroid distance is generally not less than the aggregation threshold. Therefore, the association threshold is set to be similar to or slightly larger than the aggregation threshold, enabling claims that are semantically related but belong to different semantic micro-clusters to establish forward directed connections. For example, the quantile parameter... A threshold of 0.9 corresponds to the 90th quantile of the distance between the semantic centroids of homologous semantic micro-clusters. Under the cosine distance metric, the threshold value falls around 0.3. The temporal association window is defined by a lower limit and an upper limit. The lower limit is used to avoid repeatedly associating semantic micro-clusters formed by requests that arrive almost simultaneously and belong to the same batch. The upper limit is used to avoid forcibly associating semantic micro-clusters that are too far apart in time and no longer have a diffusion relationship with each other. The lower limit of the temporal association window is a sliding step size. The specific calculation method for the upper limit of the temporal association window is as follows: statistically analyze the time interval of homologous semantic micro-cluster pairs in history, and take the set quantile of the time interval of homologous semantic micro-cluster pairs as the upper limit. For example, the 90th quantile is used to ensure that the time interval of most homologous semantic micro-cluster pairs does not exceed the upper limit, and the corresponding upper limit is several hours.

[0056] The association threshold and aggregation threshold are applied to different comparison objects. The aggregation threshold is used to compare the semantic centroid coordinates of a single appeal text with the semantic centroid coordinates of a semantic micro-cluster, while the association threshold is used to compare the semantic centroid coordinates of two semantic micro-clusters. The semantic centroid coordinates are obtained by averaging the semantic vectors of multiple appeals within a semantic micro-cluster. Compared to a single appeal semantic vector, it is less affected by differences in individual appeal expressions. Therefore, when comparing the semantic centroid distance between two semantic micro-clusters, the association threshold can be slightly relaxed based on the aggregation threshold, allowing semantic micro-clusters that are semantically related but have slightly different expressions to still establish forward directed connections.

[0057] When the lower limit of the temporal association window is too large, semantic microclusters that are closely connected may not establish forward directed connections because the time interval is less than the lower limit, thus interrupting a diffusion path. When the upper limit is too large, semantic microclusters that are far apart in time and have no diffusion relationship with each other may establish forward directed connections, thus associating semantic microclusters that do not belong to the same potential appeal event together. The lower and upper limits of the temporal association window are determined according to a reasonable time span for appeal diffusion, so that semantic microclusters of the same potential appeal event are continuously connected along the time direction, and irrelevant semantic microclusters are excluded from forward directed connections.

[0058] All forward directed connections that satisfy the above conditions are aggregated, with semantic micro-clusters as nodes and forward directed connections as directed edges, to form a micro-cluster topology graph. The micro-cluster topology graph is stored in the form of an adjacency list. For each semantic micro-cluster node, the attributes of the subsequent semantic micro-cluster nodes it points to and the corresponding forward directed connection are recorded. This facilitates subsequent steps to traverse forward directed connections by node and also helps control storage usage when there are many nodes.

[0059] The semantic centroid distance and time interval recorded by the forward directed connection can be used to recalculate whether the forward directed connection is valid, and can also be used to examine the degree of semantic and temporal association between two connected semantic micro-clusters, providing auxiliary information for subsequent processing and manual interpretation.

[0060] Taking the aforementioned heating repair request as an example, the processing of step S2 is illustrated. For example, a request reflecting insufficient heating in the same community first forms a semantic micro-cluster in the morning, and then two semantic micro-clusters with similar semantic centroid coordinates are formed in the adjacent time period. The semantic centroid distance between the three semantic micro-clusters is less than the association threshold, and the time interval between the generation times of adjacent semantic micro-clusters falls into the temporal association window. Therefore, forward directed connections are established from the semantic micro-cluster generated earlier to the two semantic micro-clusters generated later, and the three semantic micro-clusters are connected in series along the time direction.

[0061] In the process of establishing forward directed connections through pairwise traversal, the calculation of semantic centroid distance follows the measurement method of semantic distance in step S1. The difference is that the comparison object changes from the semantic vector of a single appeal text to the semantic centroid coordinates of semantic micro-clusters. Maintaining the same measurement method ensures that the judgment of semantic proximity in the aggregation stage and the association stage is on the same scale.

[0062] Through step S2, semantic microclusters that originally existed independently within their respective time windows are connected into a directed topological structure based on the two conditions of semantic homology and temporal continuity. For a type of claim that gradually spreads over time, the semantic microclusters that appear successively are linked together by forward directed connections into a diffusion path, so that homologous claims that appear across time periods and whose individual member scale is not large are no longer isolated from each other, providing a basis for identifying claims that spread in clusters at the structural level.

[0063] Step S3: Extract the weakly connected subgraphs of the micro-cluster topology graph, summarize the size of the semantic micro-cluster members in the weakly connected subgraphs to obtain the aggregation size, and use the ratio of the number of semantic micro-cluster nodes in the weakly connected subgraphs to the value obtained by adding the minimum time granularity to the first and last generation time interval as the diffusion rate.

[0064] After obtaining the micro-cluster topology graph, step S3 extracts the set of interconnected semantic micro-clusters from the graph structure and quantifies the extracted set. When extracting the set of interconnected semantic micro-clusters, the forward directed connections in the micro-cluster topology graph are temporarily treated as connections without distinguishing direction. Then, the connected components in the graph are calculated, and each connected component is a weakly connected subgraph. The forward directed connections are unidirectional, pointing from the semantic micro-cluster generated earlier to the semantic micro-cluster generated later. The diffusion path is unidirectional, requiring that strong connectivity (where nodes on the path are bidirectionally reachable) does not occur in the diffusion path. Using weak connectivity to calculate the connected components allows semantic micro-clusters within the same diffusion path or the same diffusion range, connected by forward directed connections, to be grouped into the same weakly connected subgraph. When finding connected components, a disjoint-set data structure can be used to gradually merge semantic micro-clusters connected by forward directed connections. Each forward directed connection is traversed and the semantic micro-clusters at both ends are merged into the same set. After the traversal, the semantic micro-clusters belonging to the same set constitute a weakly connected subgraph.

[0065] like Figure 5 As shown, Figure 5 This is a schematic diagram of weakly connected subgraph extraction. Figure 5 In the graph, semantic micro-clusters are represented by circular nodes, and arrowed lines between nodes represent forward directed connections. Interconnected semantic micro-clusters are enclosed by a dashed line and labeled as a weakly connected subgraph. Figure 5 The text presents two sets of weakly connected subgraphs, ① and ②. There are no connecting lines between the two sets, which is used to indicate that different weakly connected subgraphs are not connected to each other. Semantic micro-clusters that are connected to each other are labeled as belonging to the same weakly connected subgraph. Figure 5 The identifiers include semantic micro-clusters, weakly connected subgraphs ①, weakly connected subgraphs ②, and semantic micro-clusters that are not connected to each other and are connected to each other are grouped into the same weakly connected subgraph.

[0066] When using a disjoint-set data structure to find connected components, each semantic micro-cluster initially forms an independent set, identified by a representative element. During the traversal of forward directed connections, for each forward directed connection, the representative element of the set to which the semantic micro-clusters at both ends of the connection belong is found. If two representative elements are different, the two sets are merged into one. To reduce the number of traversal levels when searching for representative elements, semantic micro-clusters along the way can be directly pointed to their representative elements during the search process. After all forward directed connections have been traversed, semantic micro-clusters with the same representative element are grouped into the same weakly connected subgraph.

[0067] For each weakly connected subgraph, two metrics need to be calculated: aggregation size and diffusion rate. Aggregation size is the sum of the sizes of all semantic micro-clusters within the weakly connected subgraph. The aggregation size is obtained by adding the sizes of the members of each semantic micro-cluster within the weakly connected subgraph. Aggregation size reflects the total number of claims involved in the potential claims events corresponding to the weakly connected subgraph. The larger the aggregation size, the more claims are involved in the potential claims events.

[0068] The diffusion rate is the ratio of the number of semantic cluster nodes within a weakly connected subgraph to the value obtained by adding the minimum time granularity to the first and last generation time interval within the weakly connected subgraph, where the first and last generation time interval is the difference in generation time between the earliest and latest semantic clusters within the weakly connected subgraph. The diffusion rate is calculated using the following formula: In the formula The diffusion rate of a weakly connected subgraph. This represents the number of semantic micro-cluster nodes within a weakly connected subgraph. The generation time interval between the earliest and latest semantic micro-clusters generated within a weakly connected subgraph. This is a pre-defined minimum time granularity. Adding this minimum time granularity to the denominator serves two purposes: firstly, it avoids situations where the denominator is zero when a weakly connected subgraph contains only one semantic cluster or when the earliest and latest semantic clusters were generated at the same time; secondly, it provides a unified time-based benchmark for the diffusion rate of different weakly connected subgraphs. The diffusion rate numerically corresponds to the number of newly added associated semantic clusters per unit time, used to characterize how quickly requests accumulate along the time axis.

[0069] The diffusion rate increases with the number of semantic micro-cluster nodes within the weakly connected subgraph and decreases with the increase of the time interval between the first and last generation. A large number of semantic micro-cluster nodes and a small time interval between their first and last generation indicate the rapid aggregation of semantically similar clusters within a short period, corresponding to faster aggregation of claims on the time axis. Conversely, a small number of semantic micro-cluster nodes or a large time interval between their first and last generation indicate slower aggregation of claims on the time axis. The diffusion rate represents this relationship as a numerical value, facilitating comparisons between different weakly connected subgraphs.

[0070] like Figure 6 As shown, Figure 6 This is a schematic diagram of the temporal distribution of semantic micro-clusters within a weakly connected subgraph. Figure 6 The horizontal arrows at the bottom center represent time, and the circular nodes represent semantic micro-clusters within the same weakly connected subgraph, arranged chronologically according to their generation time. The arrowed lines connecting the nodes represent forward directed connections. The leftmost node is labeled as the earliest generated semantic micro-cluster, and the rightmost node is labeled as the latest generated semantic micro-cluster. The parentheses below them indicate the interval between the first and last generation times. The density of the nodes along the time direction corresponds to the distribution of consecutive semantic micro-clusters. Figure 6The identifier includes time, the earliest generated semantic microcluster, the latest generated semantic microcluster, and the time interval between the first and last generation.

[0071] The interval between the first and last generation times is measured in the same time unit as the smallest time granularity, ensuring that the numerator and denominator of the diffusion rate calculation formula have the same dimensions. When semantic microclusters appear consecutively in the weakly connected subgraph according to their generation times, the interval between the first and last generation times is taken as the difference between the latest and earliest generation times, and the number of semantic microcluster nodes is taken as the number of semantic microclusters in the weakly connected subgraph. Both are used to obtain a value reflecting the spread speed according to the diffusion rate calculation formula.

[0072] The number of semantic micro-cluster nodes used to calculate the diffusion rate has a different meaning from the sum of member sizes used to calculate the aggregation scale. The number of semantic micro-cluster nodes represents the number of semantic micro-clusters within a weakly connected subgraph, reflecting the number of consecutive clusters on the timeline; the sum of member sizes represents the total number of all member appeal texts within the weakly connected subgraph, reflecting the total number of appeals involved. When a weakly connected subgraph contains only one semantic micro-cluster, the number of semantic micro-cluster nodes is 1, the interval between the first and last generation times is 0, and the diffusion rate is the ratio of 1 to the minimum time granularity, reflecting the situation where a cluster occurs within the minimum time granularity. The aggregation scale is equal to the member size of a single semantic micro-cluster.

[0073] A micro-cluster topological association graph typically contains multiple weakly connected subgraphs. These subgraphs are independent of each other, lacking forward directed connections, and each corresponds to a different potential appeal event. The boundary of a weakly connected subgraph is determined by the existence of forward directed connections between semantic micro-clusters. Semantic micro-clusters connected by forward directed connections are grouped into the same weakly connected subgraph, while unconnected semantic micro-clusters are grouped into different weakly connected subgraphs. The aggregation scale and diffusion rate are calculated for each weakly connected subgraph, allowing for the quantification of different potential appeal events.

[0074] The minimum time granularity needs to be preset. The minimum time granularity is determined by the recording precision of the time the request is received, ensuring that the denominator in the diffusion rate calculation formula is neither too small (losing meaning) nor too large (lowering the distinguishability between diffusion rates of different weakly connected subgraphs). For example, when the time of request acceptance is accurate to the minute, the minimum time granularity is set to 1 minute.

[0075] Taking a type of demand caused by a sudden water outage as an example, the processing of step S3 is illustrated. For example, a certain weakly connected subgraph contains 5 semantic micro-clusters reflecting the water outage in the same area. The members of the 5 semantic micro-clusters have a total of 60 demand texts. The generation time interval between the earliest and latest semantic micro-clusters is 40 minutes. The minimum time granularity is 1 minute. Then the aggregation scale is 60, and the diffusion rate is the ratio of 5 to 41, which is about 0.122.

[0076] To compare the two different scenarios of member size and diffusion rate, another example is provided. For instance, a long-standing demand for parking obstruction arrives gradually over a long period, corresponding to a weakly connected subgraph with a large number of semantic micro-clusters and a large member size, but a long interval between the first and last generation times, resulting in a low diffusion rate. Conversely, a newly emerging demand for gas odor arrives in quick succession, corresponding to a weakly connected subgraph with a small number of semantic micro-clusters and a small member size, but a short interval between the first and last generation times, resulting in a high diffusion rate. The two types of weakly connected subgraphs are distinguished by the aggregation size and diffusion rate.

[0077] In step S3, the semantic micro-clusters scattered in the micro-cluster topology graph are converged into several weakly connected subgraphs, each corresponding to a potential appeal event. Aggregation scale quantifies the total number of appeals involved in potential appeal events, while diffusion rate quantifies the speed at which potential appeal events spread over time. Using these two indicators, aggregation scale and diffusion rate, existing appeals with a large membership but slow spread can be distinguished from newly emerging appeals with a temporarily small membership but rapid spread, providing a comparable quantitative basis for subsequent situational assessment within the same framework.

[0078] Step S4: Calculate the product of aggregation scale and diffusion rate to obtain cascade situation strength value. Mark weakly connected subgraphs whose cascade situation strength value exceeds the perception threshold as cascade situation targets. Output the correspondence between cascade situation strength value, semantic centroid coordinates within the target and member text feature codes.

[0079] After obtaining the aggregation size and diffusion rate of each weakly connected subgraph, the cascade situation strength value is calculated, and the cascade situation targets are marked accordingly. The cascade situation strength value is the product of the aggregation size and the diffusion rate, calculated using the following formula: In the formula This represents the cascaded situation strength value of a weakly connected subgraph. The aggregation size is the sum of the sizes of all semantic micro-clusters within the weakly connected subgraph. The diffusion rate is the rate of diffusion of the weakly connected subgraphs obtained in step S3. The aggregation scale is multiplied by the diffusion rate so that weakly connected subgraphs with a large total amount of claims and rapid temporal spread have a higher cascading situation strength value, while weakly connected subgraphs with a small total amount of claims and slow spread correspond to a lower cascading situation strength value.

[0080] When either the aggregation size or the diffusion rate is small, the cascade situation strength value obtained by multiplying the two is also small. Weakly connected subgraphs with small member sizes and a small total number of claims have small aggregation sizes and low cascade situation strength values. When compared with the perception threshold, they are generally not marked as cascade situation targets, thus excluding scattered claims with few cumulative claims from the cascade situation targets.

[0081] The use of multiplying aggregation size by diffusion rate, rather than using aggregation size or diffusion rate alone, is to simultaneously consider both the scale and speed of spread of potential demands. If judgment is based solely on aggregation size, newly emerging demands with a small but rapidly spreading membership would be overlooked; if judgment is based solely on diffusion rate, short-term fluctuations caused by a few occasional demands could easily be misjudged as a situation requiring attention. The cascaded situation strength value obtained by multiplying aggregation size by diffusion rate unifies the scale and speed of spread of potential demands onto a single value, allowing subsequent labeling to reflect both aspects simultaneously.

[0082] The cascading situation strength value increases with both the aggregation scale and the diffusion rate. For the same aggregation scale, weakly connected subgraphs with higher diffusion rates have higher cascading situation strength values; similarly, weakly connected subgraphs with larger aggregation scales have higher cascading situation strength values ​​for the same diffusion rate. By using a product approach, the influence of both size and spread rate on the cascading situation strength value is simultaneously accounted for, ensuring that weakly connected subgraphs that have not reached a large size but are spreading rapidly are not excluded due to insufficient size.

[0083] The cascaded situation strength value of each weakly connected subgraph is compared with a perception threshold. Weakly connected subgraphs whose cascaded situation strength values ​​exceed the perception threshold are marked as cascaded situation targets. The perception threshold is determined based on the historical statistical distribution of the cascaded situation strength values ​​of all weakly connected subgraphs. The specific calculation method for the perception threshold is as follows: A set of historical cascaded situation strength values ​​is obtained by collecting all historical cascaded situation strength values ​​of weakly connected subgraphs. A predetermined quantile of this historical set of values ​​is then used as the perception threshold. The formula for calculating the perception threshold can be defined as follows: In the formula For the perception threshold, It is the set of all historically weakly connected subgraph cascaded situation strength values. Represents taking a set of Quantiles The quantile parameter is a pre-defined parameter, ranging from 0 to 1. A larger quantile parameter value indicates that the weakly connected subgraphs marked as cascaded situation targets are more concentrated in the portion with higher cascaded situation strength values. For example, the quantile parameter... The value is set to 0.95, which corresponds to the 95th percentile of the historical cascaded situation strength value, i.e. the upper 5th percentile. Adjustments are then made based on the actual number of markers and the needs of the action.

[0084] After the perception threshold is determined based on the statistical distribution of historical cascading situation strength values, as the text of the appeal accumulates, the statistical distribution of historical cascading situation strength values ​​will change. The perception threshold is re-estimated according to a pre-set period to keep the perception threshold consistent with the overall level of recent appeals and avoid the labeling results from not matching the actual situation due to the long-term use of fixed values.

[0085] For each weakly connected subgraph labeled as a cascaded situational target, the correspondence between the three is output. The first item is the cascaded situational strength value of the cascaded situational target, which indicates the priority of attention to the cascaded situational target. The higher the cascaded situational strength value, the higher the priority of attention. The second item is the semantic centroid coordinates of each semantic micro-cluster within the cascaded situational target, which is used to trace back the semantic content of the appeal corresponding to the cascaded situational target. The semantic centroid coordinates can be further mapped to the member appeal text with the smallest semantic distance between the semantic vector and the semantic centroid coordinates, serving as the appeal theme expression of the cascaded situational target, which is convenient for manual interpretation. The third item is the text feature code of each member appeal text within the cascaded situational target, which is used to trace back and locate each original appeal work order involved in the cascaded situational target. The cascaded situational strength value, semantic centroid coordinates, and text feature code correspond to each other through the affiliation relationship of the weakly connected subgraph. The cascaded situational strength value indicates which cascaded situational targets need more attention, the semantic centroid coordinates indicate the appeal theme corresponding to the cascaded situational target, and the text feature code indicates the specific appeal work order corresponding to the cascaded situational target.

[0086] When there are multiple cascaded situation targets marked, they are sorted from highest to lowest cascaded situation strength value to form a priority order, with cascaded situation targets with higher strength values ​​listed first. The output correspondence is given in the form of structured records, with each record corresponding to one cascaded situation target. The record contains the cascaded situation strength value, the semantic centroid coordinates of each semantic micro-cluster, and the text feature code of each member's request text. Based on the structured records, cascaded situation targets can be viewed by priority, and the request topic can be viewed by the semantic centroid coordinates. The original request work order can be retrieved by the text feature code.

[0087] like Figure 7 As shown, Figure 7 A schematic diagram showing the correspondence between cascaded situational targets. Figure 7 The leftmost box represents the cascaded situational objectives, connected by three arrowed lines pointing to the three boxes on the right, with corresponding labels next to the lines. The first box on the right represents the cascaded situational intensity value and indicates its priority; the second box represents the semantic centroid coordinates and indicates the corresponding appeal theme of the cascaded situational objective; the third box represents the member text feature code and indicates its corresponding original appeal work order. Figure 7 The correspondence between the three is represented by a cascaded situational target corresponding to three outputs. Figure 7 The identifier includes the cascaded situation target, cascaded situation strength value, semantic centroid coordinates, member text feature codes and corresponding relationships.

[0088] Based on the output cascading situation strength value and the order of attention, hotline managers prioritize the verification and handling of the top-ranked cascading situation targets. They also use semantic centroid coordinates to understand the theme of the complaint and retrieve relevant complaint work orders through text feature codes, thereby enabling an earlier response to complaints that are spreading rapidly and in large numbers.

[0089] Using the aforementioned example of water outage requests to illustrate the processing of step S4: The aforementioned weakly connected subgraph containing 5 semantic micro-clusters, an aggregation size of 60, and a diffusion rate of approximately 0.122 has a cascaded situation strength value of approximately 7.3, which is the product of 60 and 0.122. If the perception threshold is 5, then the weakly connected subgraph with a cascaded situation strength value of approximately 7.3 exceeds the perception threshold and is marked as a cascaded situation target. The corresponding relationship between the cascaded situation strength value of approximately 7.3, the semantic centroid coordinates of the 5 semantic micro-clusters, and the text feature codes of the 60 member request texts is output.

[0090] In step S4, potential complaints are uniformly converted into cascaded situation strength values, and cascaded situation targets that need attention are selected accordingly. The output correspondence links the cascaded situation strength values, the semantic content of the complaints, and the specific work orders involved, which facilitates subsequent verification and handling arrangements.

[0091] In the incremental processing approach that advances with the sliding time window, as new complaint texts are continuously incorporated into semantic micro-clusters and forward directed connections are established, the weakly connected subgraph is updated accordingly. The aggregation scale, diffusion rate, and cascade situation strength value are also updated, and the labeling results of the cascade situation targets are adjusted as the complaint develops. The output correspondence is layer by layer from the cascade situation strength value through the semantic centroid coordinates to the text feature code. The cascade situation strength value determines the priority of attention, the semantic centroid coordinates determine the complaint topic, and the text feature code traces back to the original complaint work order.

[0092] Combining steps S1 to S4, this implementation starts with the time-series text records of the telephone hotline. First, it groups scattered requests into semantic micro-clusters carrying their generation time, semantic centroid coordinates, and member size. Then, it connects these semantic micro-clusters into a micro-cluster topology graph based on semantic homology and temporal connection. Next, it uses a weakly connected subgraph to converge homologous semantic micro-clusters occurring across time periods into potential request events. Finally, it measures the cascading situation strength value by the product of aggregation size and diffusion rate. Even if the member size of a type of request has not accumulated to a large value within a single time window, as long as the corresponding semantic micro-clusters continue to be added along the time axis and connected to each other through forward directed connections, the diffusion rate will drive up the cascading situation strength value, thus showing a spreading trend before the total number of requests accumulates to a high level. By using a weakly connected subgraph to reconnect semantic micro-clusters scattered across different time windows and with semantic homology, it reduces the underestimation of requests spreading in clusters. The cascading situation strength value is jointly determined by aggregation size and diffusion rate, placing existing requests with larger member sizes and rapidly emerging requests within the same numerical framework, making it easier to distinguish between the two.

[0093] As can be seen from the above processing, the perception of hotline hotspots has shifted from post-event statistics based on the total number of requests to process analysis that considers both volume and spread speed. Requests that spread rapidly and in clusters can be marked before the total number of requests accumulates to a high level; requests that have existed for a long time but spread slowly, as well as short-term, sporadic requests, are distinguished by comparing the cascading trend strength value with the perception threshold. The marking results can be verified by backtracking using semantic centroid coordinates and text feature codes.

[0094] In the above process, the marking of mass-spreading demands is based on the comparison between the cascaded situation strength value and the perception threshold. The cascaded situation strength value is determined by both the aggregation scale and the diffusion rate. Demands that spread faster are included in the cascaded situation strength value through the diffusion rate, so that the spread trend can be judged before the total number of demands has accumulated to a high level. The marking no longer depends solely on the number of demands at a single moment.

[0095] It should be noted that this implementation describes a telephone hotline complaint handling scenario. The specific type of semantic encoding model, the specific measurement method of semantic distance, and the specific method of obtaining connected components can all be determined based on actual deployment conditions without deviating from the above processing procedure. The time-series text records are not limited to telephone hotline complaint texts; any text record with a time attribute, capable of semantic encoding, and exhibiting aggregation and diffusion phenomena along the time axis can be quantitatively identified for propagation trend according to the processing procedures from steps S1 to S4. The values ​​of the aggregation threshold, association threshold, time-series association window, time window length, sliding step size, and minimum time granularity can be adjusted within the range defined by the aforementioned settings, based on the arrival density and event evolution scale of the specific text record.

[0096] Example 2 Reference Figure 8 This is a second embodiment of the present application, providing a system for quantitative identification of sprawling situations in time-series text. The system includes: a semantic micro-cluster generation module, a topological association construction module, a subgraph quantization extraction module, and a situation target labeling module.

[0097] The semantic micro-cluster generation module is used to perform semantic encoding on time-series text records, merge text increments with semantic distance less than the aggregation threshold into semantic micro-clusters according to a sliding time window, and record the semantic micro-cluster generation time, semantic centroid coordinates, and member size.

[0098] The topology association construction module is used to traverse semantic microclusters pairwise according to the generation time, calculate the semantic centroid distance and time interval, and establish a forward directed connection when the distance is less than the association threshold and the interval falls within the temporal association window, thereby generating a microcluster topology association graph.

[0099] The subgraph quantization extraction module is used to extract the weakly connected subgraphs of the micro-cluster topology association graph, summarize the size of semantic micro-cluster members in the weakly connected subgraphs to obtain the aggregation size, and use the ratio of the number of semantic micro-cluster nodes in the weakly connected subgraphs to the value obtained by adding the minimum time granularity to the first and last generation time interval as the diffusion rate.

[0100] The situation target marking module is used to calculate the product of aggregation scale and diffusion rate to obtain the cascade situation strength value, mark weakly connected subgraphs whose cascade situation strength value exceeds the perception threshold as cascade situation targets, and output the correspondence between the cascade situation strength value, the semantic centroid coordinates within the target, and the member text feature codes.

[0101] In the embodiments provided in this application, it should be understood that the disclosed apparatus and methods can be implemented in other ways. The apparatus embodiments described above are merely illustrative. For example, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. Furthermore, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Additionally, the displayed or discussed mutual couplings or direct couplings or communication connections may be through some communication interfaces; indirect couplings or communication connections between devices or units may be electrical, mechanical, or other forms.

[0102] The embodiments of this application have been described above with reference to the accompanying drawings. However, this application is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments under the guidance of this application without departing from the spirit and scope of protection of the claims. All of these variations are within the protection scope of this application.

Claims

1. A method for quantitatively identifying the spread of information in time-series text, characterized in that, include: Semantic encoding is performed on time-series text records. Text increments with semantic distance less than the aggregation threshold are merged into semantic micro-clusters by sliding time window. The generation time, semantic centroid coordinates, and member size of the semantic micro-clusters are recorded. Traverse the semantic micro-clusters pairwise according to their generation time, calculate the semantic centroid distance and time interval, and establish a forward directed connection when the distance is less than the association threshold and the interval falls within the temporal association window to generate a micro-cluster topology association graph. Extract the weakly connected subgraphs of the micro-cluster topology graph, summarize the size of the semantic micro-cluster members in the weakly connected subgraphs to obtain the aggregation size, and use the ratio of the number of semantic micro-cluster nodes in the weakly connected subgraphs to the value obtained by adding the minimum time granularity to the first and last generation time interval as the diffusion rate. The product of aggregation scale and diffusion rate is used to obtain the cascade situation strength value. Weakly connected subgraphs whose cascade situation strength values ​​exceed the perception threshold are marked as cascade situation targets. The correspondence between the cascade situation strength value, the semantic centroid coordinates within the target, and the member text feature codes is output.

2. The method for quantitative identification of the spread of information in time-series text according to claim 1, characterized in that, The semantic encoding includes: Text normalization is performed on the text in the time-series text records. Text normalization includes removing expressions that are not related to the semantics of the text, standardizing the spelling of proper nouns, and merging repeated punctuation. The normalized text is input into the semantic encoding model and mapped to a fixed-dimensional semantic vector; Each text is assigned a text feature code, which is obtained by transcribing the work order number. The text feature code, semantic vector and acceptance time are saved together.

3. The method for quantitative identification of the spread of information in time-series text according to claim 1, characterized in that, The generation of the micro-cluster topology association graph includes: The direction of the forward directed connection is from the semantic micro-cluster generated earlier to the semantic micro-cluster generated later; When establishing a forward directed connection, the semantic centroid distance and time interval are recorded as attributes of the forward directed connection; The micro-cluster topology graph is stored in the form of an adjacency list. For each semantic micro-cluster node, the attributes of the subsequent semantic micro-cluster nodes pointed to by the semantic micro-cluster node and the forward directed connections are recorded.

4. The method for quantitative identification of the spread of information in time-series text according to claim 3, characterized in that, The forward directed connection includes: No forward-directed connection is established between two semantic microclusters generated at the same time; A semantic micro-cluster generated earlier establishes forward directed connections to multiple semantic micro-clusters generated later. Multiple semantic microclusters generated earlier establish forward directed connections to a semantic microcluster generated later.

5. The method for quantitative identification of the spread of information in time-series text according to claim 1, characterized in that, The timing correlation window includes: the timing correlation window is defined by a lower limit and an upper limit, and the lower limit of the timing correlation window is a sliding step size; The upper limit of the temporal correlation window is taken from the set quantile of the time interval of the historical homologous semantic micro-clusters.

6. The method for quantitative identification of the spread of information in time-series text according to claim 1, characterized in that, The semantic micro-clusters traversed pairwise according to their generation time include: When performing pairwise traversal of semantic microclusters sorted by generation time, semantic microclusters whose generation time interval falls within the span of the temporal association window are compared. Semantic centroid distance is not calculated between semantic microclusters whose generation time interval exceeds the upper limit of the temporal association window.

7. The method for quantitative identification of the spread of information in time-series text according to claim 3, characterized in that, The association thresholds include: Select semantic microclusters that have been historically identified as belonging to the same event, and pair them together to form homologous semantic microcluster pairs; Calculate the semantic centroid distance of each homologous semantic micro-cluster pair to obtain the set of semantic centroid distances of homologous semantic micro-cluster pairs; The set of quantiles of the semantic centroid distances of homologous semantic micro-clusters is taken as the association threshold.

8. The method for quantitative identification of the spread of information in time-series text according to claim 1, characterized in that, The weakly connected subgraphs extracted from the micro-cluster topological association graph include: Treat the forward directed connections in the microcluster topology graph as connections that do not distinguish directions, and find the connected components in the microcluster topology graph. Each connected component is a weakly connected subgraph. The disjoint-set data structure is used to gradually merge semantic micro-clusters connected by forward directed connections. Each forward directed connection is traversed and the semantic micro-clusters at both ends of the connection are merged into the same set. After the traversal, the semantic micro-clusters belonging to the same set form a weakly connected subgraph.

9. The method for quantitative identification of the spread of information in time-series text according to claim 8, characterized in that, The weakly connected subgraph includes: The generation time interval between the first and last generated segments within a weakly connected subgraph is the difference in generation time between the earliest and latest generated semantic micro-clusters within the weakly connected subgraph. The minimum time granularity is the recording precision of the text acceptance time in the time-series text record; The time interval between the first and last generation times and the minimum time granularity are measured in the same time unit.

10. The method for quantitative identification of the spread of information in time-series text according to claim 1, characterized in that, The sensing threshold includes: By collecting all the cascaded situation strength values ​​of weakly connected subgraphs throughout history, a set of historical cascaded situation strength values ​​is obtained. The set quantile of the historical cascaded situation strength value set is used as the perception threshold; The historical cascaded situation strength value set is re-evaluated and the perception threshold is updated by resetting the quantiles according to a pre-set period.

11. The method for quantitative identification of the propagation trend of time-series text according to claim 1, characterized in that, The correspondence includes: The semantic centroid coordinates are mapped to the member text with the smallest semantic distance between the semantic vector and the semantic centroid coordinates, which serves as the thematic representation of the cascaded situational target. The correspondence is given in the form of structured records. Each record corresponds to a cascaded situation target. The record contains the cascaded situation strength value, the semantic centroid coordinates of each semantic micro-cluster, and the text feature code of each member text. When there are multiple cascaded situation targets that are marked, they are sorted from high to low according to their cascaded situation strength values ​​to form a priority order.

12. A system for quantitatively identifying the spread of information in time-series text, used to implement the method for quantitatively identifying the spread of information in time-series text as described in any one of claims 1 to 11, characterized in that, include: The module includes a semantic micro-cluster generation module, a topological association construction module, a subgraph quantization extraction module, and a situational target labeling module. The semantic micro-cluster generation module is used to perform semantic encoding on time-series text records, merge text increments with semantic distance less than the aggregation threshold into semantic micro-clusters according to a sliding time window, and record the semantic micro-cluster generation time, semantic centroid coordinates, and member size. The topology association construction module is used to traverse semantic microclusters pairwise according to the order of generation time, calculate the semantic centroid distance and time interval, and establish a forward directed connection when the distance is less than the association threshold and the interval falls within the temporal association window to generate a microcluster topology association graph. The subgraph quantization extraction module is used to extract weakly connected subgraphs of the micro-cluster topology graph, summarize the size of semantic micro-cluster members in the weakly connected subgraph to obtain the aggregation size, and use the ratio of the number of semantic micro-cluster nodes in the weakly connected subgraph to the value obtained by adding the minimum time granularity to the first and last generation time interval as the diffusion rate. The situation target marking module is used to calculate the product of aggregation scale and diffusion rate to obtain the cascade situation strength value, mark weakly connected subgraphs whose cascade situation strength value exceeds the perception threshold as cascade situation targets, and output the correspondence between the cascade situation strength value, the semantic centroid coordinates within the target, and the member text feature codes.

Images