Government affair service appeal accurate dispatching method based on artificial intelligence large model

Abstract

The application discloses a government affair service appeal accurate dispatching method based on an artificial intelligence large model, belongs to the technical field of government affair services, and specifically comprises the following steps: constructing a city affair time-space graph by fusing work order appeal nodes and engineering event nodes; associating historical appeals with the engineering event nodes and calculating time-space matching degrees; taking the original imputation confidence distribution of the historical appeals and the time-space matching degrees as inputs, training a cause-effect correction model with the final imputation department as a label; acquiring a current appeal location and an event type, searching for candidate engineering event nodes, calculating real-time time-space matching degrees; inputting the confidence distribution output by a preset classification model and the real-time time-space matching degrees into the correction model, and outputting a corrected distribution to determine a dispatching target. The application improves the dispatching decision from single text mapping to the adaptive correction level of cause-effect perception by introducing the time-space background of the engineering event behind the appeal, and reduces the single flow transfer and disposal delay.

Description

Technical Field

[0001] This invention relates to the field of government service technology, specifically to a method for accurately dispatching government service requests based on a large-scale artificial intelligence model. Background Technology

[0002] With the advancement of digital transformation in government services, the number of public complaints received daily by government hotlines and online service platforms continues to grow. How to quickly and accurately assign these massive numbers of complaints to the appropriate departments has become crucial for improving service efficiency. Traditional manual assignment is inefficient and highly subjective, and its accuracy is difficult to guarantee in scenarios involving overlapping departmental responsibilities. Therefore, intelligent dispatch systems based on natural language processing and deep learning have become the mainstream solution.

[0003] Current intelligent order dispatch methods mainly rely on two technical approaches: department classification based on text semantic understanding and rule matching based on knowledge graphs. Text semantic understanding encodes the request description through a pre-trained language model, outputs the probability distribution of each candidate department, and selects the department with the highest probability for order dispatch. Knowledge graph rule matching constructs a structured knowledge base from department function descriptions and historical records, and completes the matching through semantic similarity or logical rules. Some solutions introduce multimodal information such as voice and images to improve classification accuracy.

[0004] However, the aforementioned approaches face common technical limitations: a semantic gap exists between the surface semantics of the complaint text and the actual logic of accountability. Existing methods simplify work order dispatch to a one-way mapping from complaint text to departmental labels, with decision-making primarily based on explicit information in the complaint text, lacking awareness of the spatiotemporal context behind the complaint. The same type of complaint event may point to different responsible parties in different spatiotemporal contexts. Taking road damage as an example, the responsible party during a regular maintenance cycle is the municipal department; however, if gas repair excavation occurs on the road section within a few days and has not been handed over, the responsible party becomes the construction company. Existing systems only identify the semantic feature of "road damage" and cannot determine whether there is an engineering event on the road section and its causal relationship. As a result, work orders are dispatched to the municipal department, which, after verification, needs to cancel and reassign the order, causing cross-departmental processing and delays in handling.

[0005] Existing technical solutions are limited by the mapping paradigm of "text features - department labels". They fail to consider the location and time period of the request during the dispatch decision process and lack the ability to perceive the spatiotemporal background hidden behind the request text. This can lead to discrepancies between the dispatch result and the actual responsibility logic when handling requests that are potentially related to the engineering event. Problems such as the handling department rejecting the request, the work order being transferred across departments, and the processing time being extended have occurred. This not only reduces the efficiency of government services but also affects public satisfaction. Summary of the Invention

[0006] The purpose of this invention is to provide a method for accurately dispatching government service requests based on a large-scale artificial intelligence model, thereby solving the following technical problems:

[0007] Existing intelligent dispatching methods rely on the semantic features of the request text but fail to integrate the spatiotemporal information of engineering events scattered across external business platforms. This results in a systematic deviation between the dispatching results and the actual attribution logic when processing requests that have a potential causal relationship with engineering events.

[0008] The objective of this invention can be achieved through the following technical solutions: A method for accurately dispatching government service requests based on a large-scale artificial intelligence model includes the following steps: S1. Obtain historical work order records of government services and registration data of urban engineering events, and establish a spatiotemporal map of urban affairs containing work order request nodes and engineering event nodes based on government business management standard data. S2, associate historical work orders with each engineering event node in the urban event spatiotemporal map, and calculate the spatiotemporal matching degree between each engineering event node and the associated work order request node. S3, extract the confidence distribution of the responsible department and the final responsible department output by the preset department classification model for each historical appeal work order, and construct a training sample set based on the spatiotemporal matching degree to generate a causal effect correction model; S4, obtain the location entity and event type entity of the current government service request work order, and obtain the corresponding geofence coordinate range based on the location entity. S5. Using the geographic fence coordinate range and the entity of the appeal event type as search conditions, perform a spatiotemporal joint query in the urban event spatiotemporal map to obtain the candidate project event node corresponding to the current government appeal work order, and calculate the spatiotemporal matching degree between the current government appeal work order and the candidate project event node. S6. Input the current government service request work order into the preset department classification model to obtain the confidence distribution of the responsible department. Input the confidence distribution of the responsible department and the spatiotemporal matching degree into the causal effect correction model to obtain the final confidence distribution of the responsible department. Select the department with the highest probability as the dispatch target department of the current government service request work order.

[0009] As a further aspect of the present invention: the specific construction process of the urban event spatiotemporal map in S1 is as follows: Acquire historical service request records and urban engineering event registration data; perform entity recognition on each historical service request record, extract the request location entity, request event type entity and request receipt timestamp, retrieve the request geofence coordinate range from the preset urban geographic information database based on the request location entity, and generate the service request node; For each urban engineering event registration data, entity recognition is performed to extract the engineering event type entity, engineering event location entity, engineering event responsible entity entity, engineering event start time, engineering event end time, and engineering event status attributes. Based on the engineering event location entity, a preset urban geographic information database is searched to obtain the spatial coordinate range of the engineering event and generate engineering event nodes. The government affairs business management standard data is split into sentences to obtain a set of standard text sentences. Each standard text sentence is input into a pre-trained semantic role labeling model to identify the agent and the patient. The engineering event type entity corresponding to the agent and the appeal event type entity corresponding to the patient are extracted. A directed causal relationship edge is established from the engineering event type entity to the appeal event type entity. The engineering event type entity and appeal event type entity with a directed causal relationship edge are stored as event type causal relationship rules. In the urban event spatiotemporal graph, the work order request node and the engineering event node are created. When the request geofence coordinate range of the work order request node includes the engineering event spatial coordinate range of the engineering event node, a first connection edge is established. When the engineering event type entity of the engineering event node and the request event type entity of the work order request node satisfy the event type causal association rule, a second connection edge is established.

[0010] As a further aspect of the present invention: the specific process of associating historical request work orders in S2 is as follows: Traverse each engineering event node in the urban event spatiotemporal graph. For the currently traversed engineering event node, obtain its engineering event spatial coordinate range and the event occurrence time window composed of the start time and end time of the engineering event. In the work order request nodes of the urban event spatiotemporal map, work order request nodes that have spatial intersection between the geographical fence coordinate range of the request and the spatial coordinate range of the project event are retrieved, and the retrieved work order request nodes are used as candidate associated work order request nodes. For each candidate related work order request node, obtain its request receipt timestamp, match the request receipt timestamp with the event occurrence time window of the project event node, and if the request receipt timestamp is within the event occurrence time window, then determine the candidate related work order request node as a related work order request node.

[0011] As a further aspect of the present invention: the specific calculation process of the spatiotemporal matching degree in S2 is as follows: Calculate the spatial overlap ratio between the geographic fence coordinate range of each associated work order request node and the spatial coordinate range of the engineering event node, as the spatial matching degree; calculate the time overlap ratio between the request receiving timestamp of each associated work order request node and the event occurrence time window of the engineering event node, as the temporal matching degree; and sum the spatial matching degree and the temporal matching degree according to a preset weighting coefficient to obtain the spatiotemporal matching degree.

[0012] As a further aspect of the present invention: in step S3, the specific construction process of the preset department classification model is as follows: Obtain historical service request records, perform entity recognition on each historical service request record, extract the entity of the request event type and the content of the request text, and obtain the final responsible department for the historical service request record. Using the text of each historical appeal work order as input samples and the corresponding final responsible department as classification label, a department classification training sample set is constructed. An initial neural network model is constructed, which includes a text encoding layer and a fully connected classification layer. The text encoding layer is used to semantically encode the input appeal text content to obtain a text semantic vector. The fully connected classification layer is used to linearly map the text semantic vector and output the predicted probability distribution of each candidate responsible department. The initial neural network model is trained under supervision using the department classification training sample set. The cross-entropy loss between the predicted probability distribution output by the fully connected classification layer and the classification label corresponding to the final responsible department is used as the loss function. The model parameters of the initial neural network model are iteratively updated through the backpropagation algorithm until the loss function converges, thus obtaining the pre-trained department classification model.

[0013] As a further aspect of the present invention: the specific construction process of the causal effect correction model in S3 is as follows: Obtain the associated work order request nodes and associated project event nodes corresponding to each historical request work order, and obtain the spatiotemporal matching degree between the associated work order request node and the project event node. Input the request text content corresponding to the related work order request node into the preset department classification model to obtain the confidence distribution of the responsible department. Determine the candidate responsible department corresponding to the maximum probability in the confidence distribution of the responsible department as the first department, and obtain the final responsible department corresponding to the related work order request node. The related work order request nodes whose first department is different from the final responsible department and whose final responsible department is the same as the responsible entity of the related engineering event node are considered as valid sample nodes. Using the confidence distribution of the responsible department corresponding to the effective sample node as the first input feature and the spatiotemporal matching degree corresponding to the effective sample node as the second input feature, the confidence distribution of the responsible department corresponding to the effective sample node is probability redistributed. The probability value of the final responsible department is increased to a preset high probability threshold, and the probability values ​​of the remaining candidate responsible departments are reduced to a sum that complements the preset high probability threshold according to their original relative proportions. The redistributed confidence distribution of the responsible department is used as the target probability distribution to generate a causal effect correction training sample set. An initial causal effect correction network is constructed, comprising a feature concatenation layer, a fully connected hidden layer, and a probability output layer. The initial causal effect correction network is trained under supervision using the causal effect correction training sample set. The cross-entropy loss between the corrected departmental probability distribution output by the probability output layer and the target probability distribution is used as the loss function. The network parameters are updated through backpropagation until convergence, thus obtaining the causal effect correction model.

[0014] As a further aspect of the present invention: in step S5, the specific process for obtaining candidate project event nodes is as follows: Based on the location entity of the current government service request work order, the spatial coordinates of the location entity are obtained by querying the preset city geographic information database. A buffer zone is generated with the spatial coordinates as the center and a preset radius. The coordinate range of the buffer zone is used as the geofence coordinate range of the location entity. Using the geographic fence coordinate range as the spatial retrieval condition, the engineering event nodes whose spatial coordinate ranges intersect with the geographic fence coordinate range are retrieved in the urban event spatiotemporal map to obtain the first set of engineering event nodes. Using the current government service request work order's request event type entity as the causal retrieval condition, filter the first set of engineering event nodes in which the engineering event type entity and the request event type entity satisfy the event type causal association rule to obtain the second set of engineering event nodes. Using the current government service request work order's request receipt timestamp as the time retrieval condition, filter the second set of engineering event nodes whose event occurrence time window contains the request receipt timestamp, and use them as candidate engineering event nodes.

[0015] As a further aspect of the present invention: in step S6, the specific process for determining the target department for order dispatch is as follows: Input the text of the current government service request work order into the preset department classification model to obtain the confidence distribution of the responsible department; input the confidence distribution of the responsible department and the spatiotemporal matching degree between the current government service request work order and the candidate project event node into the causal effect correction model to obtain the corrected department probability distribution; traverse the probability values ​​of each candidate responsible department in the corrected department probability distribution, and determine the candidate responsible department corresponding to the maximum probability as the target department for dispatching the work order.

[0016] The beneficial effects of this invention are: 1) This invention extracts the location entity of each historical petition and retrieves its geofence coordinate range. It also extracts the location entity of each engineering event registration and retrieves its spatial coordinate range. When the petition's geofence coordinate range includes the engineering event's spatial coordinate range, a first connection edge is established, thus achieving a spatial association between petitions and engineering events. Simultaneously, by semantically annotating government affairs management data, it extracts the engineering event type entity corresponding to the agent and the petition event type entity corresponding to the recipient, establishing a directed causal relationship edge from the engineering event type to the petition event type. A second connection edge is established when the engineering event type entity of the engineering event node and the petition event type entity of the petition node satisfy this causal relationship rule. This dual constraint of the first and second connection edges ensures that the urban event spatiotemporal graph not only records the spatial inclusion relationship between petitions and engineering events but also solidifies the causal relationship between event types. Based on this map, when the dispatch system receives a new request, it no longer relies solely on the semantic features of the request text for isolated judgment. Instead, it can use the location where the request occurred as an anchor point to search for engineering event nodes that have spatial intersection and causal relationship with the current request within a preset spatiotemporal window. This effectively allows it to perceive the spatiotemporal background information hidden behind the request, expanding the information basis for dispatch decisions from a single textual semantic dimension to a multidimensional information space that includes spatial and causal relationships.

[0017] 2) This invention achieves adaptive adjustment of the output results of the preset department classification model by training a causal effect correction model. During the training phase, effective sample nodes are selected from historical data. These effective sample nodes satisfy the condition that the first department corresponding to the maximum probability in the original attribution department confidence distribution is different from the final attribution department, and that the final attribution department is the same as the entity responsible for the engineering event of the associated engineering event node. This selection mechanism ensures that the training samples focus on situations where there is a deviation between the original dispatch result and the actual attribution department, and that this deviation can be corrected by the entity responsible for the engineering event. The original attribution department confidence distribution of the effective sample node is used as the first input feature, and the spatiotemporal matching degree between the node and the corresponding engineering event node is used as the second input feature. The original attribution department confidence distribution is then probability-redistributed to construct a target probability distribution, i.e., increasing the probability value of the final attribution department to a preset threshold while maintaining the relative proportion of the probability values ​​of other departments. The construction method of this target probability distribution not only clarifies the correct order assignment target, but also preserves the semantic relative relationship between each candidate department in the original distribution. The causal effect correction network learns the mapping from the first input feature and the second input feature to the target probability distribution, and can capture the influence pattern of spatiotemporal matching degree on the order assignment result from historical order cancellation cases. That is, when the spatiotemporal matching degree is high, the correction network tends to shift the probability quality towards the responsible party of the engineering event, and when the spatiotemporal matching degree is low, it tends to maintain the original distribution form. This allows the correction network to adaptively determine the correction magnitude of the original attribution department confidence distribution according to the actual spatiotemporal matching degree between the current demand and the candidate engineering event nodes during the online inference stage.

[0018] 3) This invention integrates the textual semantic understanding capability of the pre-defined department classification model with the spatiotemporal background correction capability of the causal effect correction model. The pre-defined department classification model first semantically encodes the current request text content, outputting a confidence distribution of the responsible department reflecting the textual probability of attribution. Simultaneously, the system uses the geofence coordinates of the current request, the entity type of the request event, and the request reception timestamp as search criteria, performing a three-layer screening process—spatial retrieval, causal retrieval, and temporal retrieval—in the city's spatiotemporal event map to obtain candidate engineering event nodes and calculate the spatiotemporal matching degree. Subsequently, the confidence distribution of the responsible department and the spatiotemporal matching degree are input into the trained causal effect correction model. The correction model dynamically adjusts the original confidence distribution based on the spatiotemporal matching degree, outputting a corrected department probability distribution. When a matching engineering event with a valid status exists within the time window of the request location, the correction model increases the probability value of the department corresponding to the responsible entity of the engineering event, correcting the dispatch result towards the responsible entity of the engineering event; when no matching engineering event exists, the correction model basically maintains the original confidence distribution, and the dispatch result is still dominated by textual semantics. This collaborative mechanism enables the work order dispatch system to maintain its ability to accurately dispatch routine requests while significantly improving the accuracy of dispatching requests causally related to engineering events. Because the dispatch result is matched to the responsible party for the engineering event in the first stage, it effectively reduces the number of order cancellations initiated by the responsible department after on-site verification due to mismatched responsibilities. This shortens the intermediate steps in the cross-departmental flow of work orders and the overall processing time, thereby achieving simultaneous improvements in dispatch accuracy, response efficiency, and service quality. Attached Figure Description

[0019] The invention will now be further described with reference to the accompanying drawings.

[0020] Figure 1 This is a schematic diagram of a method for accurately dispatching government service requests based on a large artificial intelligence model, according to the present invention. Detailed Implementation

[0021] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0022] Please see Figure 1 As shown, this invention is a method for accurately dispatching government service requests based on a large-scale artificial intelligence model, comprising the following steps: S1. Obtain historical work order records of government services and registration data of urban engineering events, and establish a spatiotemporal map of urban affairs containing work order request nodes and engineering event nodes based on government business management standard data. Acquire historical government service request records and urban engineering event registration data. For each historical request record, perform entity recognition to extract the request location entity, request event type entity, and request receipt timestamp. For example, for the request text "Road damage at the intersection of Qianshan Road and Wangjiang West Road," extract the request location entity as "intersection of Qianshan Road and Wangjiang West Road," the request event type entity as "road damage," and the request receipt timestamp as "April 14, 2026, 10:25 AM." Based on the request location entity, retrieve the spatial coordinates of the corresponding point in the preset urban geographic information database. Generate a buffer polygon with a radius of 50 meters centered on these coordinates. Use the coordinate range of this polygon as the geographic fence coordinate range for the request, generating a request node. For each urban engineering event registration data, perform entity recognition to extract the engineering event type entity, engineering event location entity, engineering event responsible entity entity, engineering event start time, engineering event end time, and engineering event status attribute. For example, for a gas emergency repair registration record, the following entities are extracted: the project event type entity is "road excavation," the project event location entity is "north side of the intersection of Qianshan Road and Wangjiang Road," the project event responsible entity is "Hefei Gas Group Co., Ltd.", the project event start time is "9:00 AM on April 12, 2026," the project event end time is "9:00 AM on April 19, 2026," and the project event status attribute is "not handed over." Based on the project event location entity, the spatial coordinate range of the project event is obtained from a preset urban geographic information database, generating a project event node. Government business management standard data is obtained and the statements are split to obtain a set of standard text statements. Each standard text statement is input into a pre-trained semantic role labeling model to identify the agent and patient components. The project event type entity corresponding to the agent component and the appeal event type entity corresponding to the patient component are extracted. A directed causal relationship edge is established from the project event type entity to the appeal event type entity. The project event type entity and appeal event type entity pairs with directed causal relationship edges are stored as event type causal relationship rules. For example, from the standard text "The excavation unit is responsible for maintenance during the temporary backfilling period after road excavation," the agent component "road excavation" and the recipient component "road damage" are identified, and a directed causal relationship edge is established from "road excavation" to "road damage." A first connection edge is established when the geographic coordinate range of the work order request node includes the spatial coordinate range of the engineering event node; a second connection edge is established when the engineering event type entity of the engineering event node and the request event type entity of the work order request node satisfy the causal relationship rule of the event type, thus forming a spatiotemporal graph of urban events.

[0023] S2, associate historical work orders with each engineering event node in the urban event spatiotemporal map, and calculate the spatiotemporal matching degree between each engineering event node and the associated work order request node. The process iterates through each engineering event node in the city's spatiotemporal graph, obtaining its spatial coordinate range and the event occurrence time window comprised of the start and end times of the engineering event. Work order request nodes whose geographic fence coordinate range intersects with the spatial coordinate range of the engineering event are identified as candidate related work order request nodes. For example, for the Qianshan Road gas emergency repair project event node, fifteen work order request nodes whose geographic fence coordinate range intersects with its spatial coordinate range were found. Candidate related work order request nodes whose request receipt timestamp falls within the event occurrence time window are identified as related work order request nodes. The spatial matching degree is calculated as the percentage of spatial overlap between the geographic fence coordinate range of each related work order request node and the spatial coordinate range of the engineering event node, and the temporal matching degree is calculated as the percentage of temporal overlap between the request receipt timestamp of each related work order request node and the event occurrence time window of the engineering event node. For example, for a work order with a request received at "10:25 AM on April 14, 2026," this time point falls entirely within the time window of the gas emergency repair event, and the time matching degree is calculated as 1. The spatial overlap area between the geofence coordinates of this work order and the spatial coordinates of the gas emergency repair project event is 0.85%. The spatial matching degree and the time matching degree are weighted and summed according to preset weighting coefficients (e.g., each accounting for 50%), resulting in a spatiotemporal matching degree of 0.925 between the related work order request node and the project event node.

[0024] S3, extract the confidence distribution of the responsible department and the final responsible department output by the preset department classification model for each historical appeal work order, and construct a training sample set based on the spatiotemporal matching degree to generate a causal effect correction model; This process involves acquiring historical service request records, performing entity recognition on each record to extract the event type entity and the text content of the request, and identifying the ultimately responsible department for that record. Using the text content of each historical request record as input samples and the corresponding ultimately responsible department as classification labels, a departmental classification training sample set is constructed. An initial neural network model containing a text encoding layer and a fully connected classification layer is built. This model is then trained under supervision using the departmental classification training sample set. The cross-entropy loss between the predicted probability distribution output by the fully connected classification layer and the classification label corresponding to the ultimately responsible department is used as the loss function. The model parameters are iteratively updated using a backpropagation algorithm until the loss function converges, resulting in a pre-defined departmental classification model. Finally, the associated service request nodes and associated project event nodes for each historical service request are acquired, and the spatiotemporal matching degree between these nodes is determined. The request text content corresponding to the associated work order request node is input into a preset department classification model to obtain the confidence distribution of the responsible department. The candidate responsible department corresponding to the maximum probability in the confidence distribution of the responsible department is determined as the first department, and the final responsible department corresponding to the associated work order request node is obtained. For example, in the confidence distribution of the responsible department output by the preset department classification model, the probability of "Hefei Municipal Engineering Management Office" is 0.88 and the probability of "Hefei Gas Group Co., Ltd." is 0.04. The first department is determined to be "Hefei Municipal Engineering Management Office", and the final responsible department of this work order is "Hefei Gas Group Co., Ltd." Related work order request nodes whose first department and final responsible department are different, and whose final responsible department is the same as the entity responsible for the engineering event of the associated engineering event node, are taken as valid sample nodes. Using the confidence distribution of the responsible department corresponding to the effective sample node as the first input feature and the spatiotemporal matching degree corresponding to the effective sample node as the second input feature, the confidence distribution of the responsible department corresponding to the effective sample node is probabilistically redistributed. The probability value of the final responsible department is increased to a preset high probability threshold of 0.9, and the probability values ​​of the remaining candidate responsible departments are reduced to 0.1, which is complementary to the preset high probability threshold. The redistributed confidence distribution of the responsible department is used as the target probability distribution to generate a causal effect correction training sample set. An initial causal effect correction network containing a feature concatenation layer, a fully connected hidden layer, and a probability output layer is constructed. The initial causal effect correction network is trained under supervision using the causal effect correction training sample set. The cross-entropy loss between the corrected department probability distribution output by the probability output layer and the target probability distribution is used as the loss function. The network parameters are updated through backpropagation until convergence, resulting in the causal effect correction model.

[0025] S4, obtain the location entity and event type entity of the current government service request work order, and obtain the corresponding geofence coordinate range based on the location entity. Obtain the location entity and event type entity of the current government service request. Based on the location entity of the current government service request, query the preset city geographic information database to obtain the spatial coordinates of the location entity. Generate a buffer zone with a preset radius of 50 meters centered on the spatial coordinates, and use the coordinate range of the buffer zone as the geofence coordinate range corresponding to the location entity.

[0026] S5. Using the geographic fence coordinate range and the entity of the appeal event type as search conditions, perform a spatiotemporal joint query in the urban event spatiotemporal map to obtain the candidate project event node corresponding to the current government appeal work order, and calculate the spatiotemporal matching degree between the current government appeal work order and the candidate project event node. Using geofence coordinates as the spatial retrieval condition, engineering event nodes whose spatial coordinates intersect with the geofence coordinates in the urban event spatiotemporal map are retrieved to obtain the first set of engineering event nodes. Using the current government service request work order's request event type entity as the causal retrieval condition, engineering event nodes whose event type entities and request event type entities satisfy the event type causal association rule are selected from the first set of engineering event nodes to obtain the second set of engineering event nodes. Using the current government service request work order's request receipt timestamp as the temporal retrieval condition, engineering event nodes whose event occurrence time window includes the request receipt timestamp are selected from the second set of engineering event nodes as candidate engineering event nodes. For example, if a gas emergency repair engineering event node is found within the geofence coordinates corresponding to the current request location entity, and its engineering event type entity "road excavation" and request event type entity "road surface damage" satisfy the event type causal association rule, and the request receipt timestamp "April 14, 2026, 10:25 AM" falls within the time window of this engineering event, then this gas emergency repair engineering event node is identified as a candidate engineering event node. The spatiotemporal matching degree between the current government service request work order and the candidate project event node is calculated. The calculation method includes: calculating the proportion of spatial overlap between the geographic fence coordinate range of the current request and the spatial coordinate range of the project event of the candidate project event node as the spatial matching degree; calculating the proportion of time overlap between the time stamp of the current request and the time window of the event occurrence of the candidate project event node as the time matching degree; and weighting the spatial matching degree and the time matching degree according to the preset weight coefficient to obtain the spatiotemporal matching degree.

[0027] S6. Input the current government service request work order into the preset department classification model to obtain the confidence distribution of the responsible department. Input the confidence distribution of the responsible department and the spatiotemporal matching degree into the causal effect correction model to obtain the final confidence distribution of the responsible department. Select the department with the highest probability as the dispatch target department of the current government service request work order.

[0028] The text of the current government service request is input into a preset department classification model to obtain the confidence distribution of the responsible department. For example, after inputting the current request text, the confidence distribution of the responsible department is: "Hefei Municipal Engineering Management Office" probability 0.88, "Hefei Gas Group Co., Ltd." probability 0.04, and "Shushan District Urban Management Bureau" probability 0.05. The confidence distribution of the responsible department and the spatiotemporal matching degree between the current government service request and the candidate project event nodes are input into a causal effect correction model to obtain the corrected department probability distribution. For example, the correction model adjusts the original confidence distribution based on a spatiotemporal matching degree of 0.925, outputting the corrected department probability distribution as: "Hefei Gas Group Co., Ltd." probability 0.75, "Hefei Municipal Engineering Management Office" probability 0.16, and "Shushan District Urban Management Bureau" probability 0.09. The probability values ​​of each candidate responsible department in the corrected department probability distribution are traversed, and the candidate responsible department corresponding to the maximum probability is determined as the target department for dispatching the current government service request, that is, "Hefei Gas Group Co., Ltd." is determined as the target department for dispatching the request.

[0029] In a preferred embodiment of the present invention, the specific construction process of the urban event spatiotemporal map in step S1 is as follows: First, historical service request records and urban engineering event registration data are obtained. Entity identification is performed on each historical service request record. For example, the request text "potholes appear at the intersection of Qianshan Road and Wangjiang West Road" is processed to extract the entity of the request location as "intersection of Qianshan Road and Wangjiang West Road", the entity of the request event type as "road damage", and the request receipt timestamp as "April 14, 2026, 10:25". Based on the entity of the request location, the spatial coordinates of the corresponding intersection are obtained by searching the preset urban geographic information database. A buffer polygon is generated by expanding outward by a preset distance from the coordinates of the point. The vertex coordinate sequence of the polygon is used as the geographic fence coordinate range of the request. The above-extracted information is stored as a service request node. For each urban engineering event registration data, entity identification is performed. For example, a gas emergency repair excavation approval record is processed to extract the following entities: the event type entity is "road excavation", the event location entity is "north side of the intersection of Qianshan Road and Wangjiang Road", the event responsible entity is "Hefei Gas Group Co., Ltd.", the event start time is "9:00 AM on April 12, 2026", the event end time is "9:00 AM on April 19, 2026", and the event status attribute is "temporary backfilling not handed over". Based on the event location entity, the spatial coordinate range of the precise construction area of ​​the event is obtained by retrieving it from the preset urban geographic information database. The extracted information is then stored as an engineering event node. The data of government affairs management regulations is split into sentences to obtain a set of standardized text sentences. For example, sentences such as "The responsibility for road surface restoration after road excavation shall be borne by the excavating unit before being handed over to the municipal department" are extracted from the Municipal Facilities Management Regulations. Each standardized text sentence is input into a pre-trained semantic role labeling model to identify the agent as "road excavation" and the recipient as "road surface damage". The corresponding engineering event type entity is extracted as "road excavation" and the corresponding demand event type entity is extracted as "road surface damage". A directed causal relationship edge is established from the engineering event type entity "road excavation" to the demand event type entity "road surface damage". The engineering event type entity and demand event type entity with such directed causal relationship edge are stored as event type causal relationship rules. Similarly, multiple rules such as "pipeline repair leads to road surface collapse" and "community renovation leads to water outage" can also be established.After the aforementioned nodes are created, each work order request node and each project event node is traversed. When it is determined that the geographic geofence coordinates of a work order request node spatially encompass the spatial coordinates of a project event node, a first connection edge is established between the work order request node and the project event node, representing a spatial inclusion relationship between them. When it is determined that the project event type entity of a project event node and the request event type entity of a work order request node satisfy the stored event type causal association rules, a second connection edge is established between the project event node and the work order request node, representing a type-level causal association relationship between them. Through the joint constraints of the first and second connection edges, the spatiotemporal graph of urban events is constructed.

[0030] By abstracting work orders and project events into nodes containing spatiotemporal and semantic attributes, and establishing connecting edges under the dual constraints of spatial inclusion and causal relationship, an explicit relationship link is formed between the originally isolated work order text and the hidden project background. This allows the subsequent work order dispatch process, when faced with a new work order, to no longer rely solely on the semantic features of the work order text itself for isolated judgment, but instead use the location of the work order as a spatial anchor point and the type of work order as a semantic anchor point to quickly retrieve project event nodes that overlap spatially and have causal relationships in type within the graph. This supplements the work order dispatch decision with the spatiotemporal background information hidden behind the work order.

[0031] In another preferred embodiment of the present invention, the specific process of associating historical work orders in step S2 is as follows: First, iterate through each engineering event node in the urban event spatiotemporal graph. For the currently traversed engineering event node, read its spatial coordinate range attribute and the event occurrence time window attribute, which is composed of the start time and end time of the engineering event. For example, if a gas emergency repair engineering event node is read, its spatial coordinate range is a continuous polygon coordinate string in the area of ​​the carriageway north of the intersection of Qianshan Road and Wangjiang West Road, and the time window is from 9:00 on April 12, 2026 to 9:00 on April 19, 2026. Next, a spatial search is performed on all work order request nodes in the map. The geographic coordinate range of each work order request node is determined to determine whether there is a spatial geometric intersection between the geographic fence coordinate range of the request and the spatial coordinate range of the current project event node. The intersection includes cases such as complete inclusion, partial overlap, or boundary connection. Work order request nodes with spatial intersection are selected as candidate associated work order request nodes. For example, for the gas emergency repair project event node mentioned above, more than 10 work order request nodes with request locations in the vicinity of the intersection of Qianshan Road and Wangjiang West Road are obtained through spatial search, covering various request types such as road damage, loose manhole covers, and traffic congestion. Then, for each candidate related work order request node, its request receipt timestamp attribute is extracted. The request receipt timestamp is matched with the event occurrence time window of the current project event node. If the request receipt timestamp is greater than or equal to the event start time and less than or equal to the event end time, it is determined that the request receipt timestamp is within the event occurrence time window. At this time, the candidate related work order request node is officially confirmed as a related work order request node with a relationship with the current project event node. For example, if there is a road damage request in the candidate work orders with a receipt timestamp of 10:25 on April 14, 2026, this time point falls completely within the time window of the gas emergency repair event, then this work order is confirmed as a related work order request node. On the other hand, the receipt timestamp of another manhole cover loosening request is 8:00 on April 10, 2026, which is earlier than the event start time, so this work order will not be confirmed as a related work order request node.

[0032] By filtering through hard constraints of spatial intersection and temporal inclusion, irrelevant requests that are spatially close but occur outside the project event cycle can be effectively eliminated. This ensures that the spatiotemporal matching degree calculated subsequently is based on a genuine spatiotemporal co-occurrence relationship between the request and the project event. Simultaneously, this association mechanism provides a prerequisite for constructing training samples for the causal effect correction model. Only by clearly identifying which historical request work orders are associated with which project event nodes can the spatiotemporal matching degree in each pair of relationships be further calculated, and this degree can be used as the second input feature of the correction network for supervised learning. By explicitly extracting and structurally associating the spatiotemporal co-occurrence patterns of requests and project events that were originally implicit in the scattered data, the dispatch system can determine whether a new request falls within the spatiotemporal influence range of a particular project event by referring to historical patterns. This lays a data foundation for adaptive correction of subsequent dispatch decisions.

[0033] In another preferred embodiment of the present invention, the specific calculation process of the spatiotemporal matching degree in step S2 is as follows: Calculate the spatial overlap ratio between the geographic fence coordinate range of each associated work order request node and the spatial coordinate range of the engineering event node, as the spatial matching degree; calculate the time overlap ratio between the request receiving timestamp of each associated work order request node and the event occurrence time window of the engineering event node, as the temporal matching degree; and sum the spatial matching degree and the temporal matching degree according to a preset weighting coefficient to obtain the spatiotemporal matching degree.

[0034] In another preferred embodiment of the present invention, the specific construction process of the preset department classification model in step S3 is as follows: Obtain historical service request records, perform entity recognition on each historical service request record, extract the entity of the request event type and the content of the request text, and obtain the final responsible department for the historical service request record. The text of each historical appeal work order is used as the input sample, and the corresponding final responsible department is used as the classification label to construct a department classification training sample set. An initial neural network model is constructed, which includes a text encoding layer and a fully connected classification layer. The text encoding layer is used to semantically encode the input appeal text content to obtain a text semantic vector. The fully connected classification layer is used to linearly map the text semantic vector and output the predicted probability distribution of each candidate responsible department. The initial neural network model is trained under supervision using the department classification training sample set. The cross-entropy loss between the predicted probability distribution output by the fully connected classification layer and the classification label corresponding to the final responsible department is used as the loss function. The model parameters of the initial neural network model are iteratively updated through the backpropagation algorithm until the loss function converges, thus obtaining the pre-trained department classification model.

[0035] In another preferred embodiment of the present invention, the specific construction process of the causal effect correction model in step S3 is as follows: First, obtain the associated work order request nodes and the engineering event nodes associated with each historical request work order. At the same time, read the spatiotemporal matching degree value between the associated work order request node and the engineering event node that has been calculated. For example, for a road damage request work order, the spatiotemporal matching degree between its associated gas emergency repair engineering event node and the work order is calculated to be 0.925. The request text content corresponding to the related work order request node is input into a pre-trained preset department classification model. After text encoding and fully connected classification operations, the model outputs a confidence distribution vector of the responsible department. This vector contains the probability values ​​of each candidate responsible department. For example, in the output distribution, the probability value of Hefei Municipal Engineering Management Office is 0.88, the probability value of Hefei Gas Group Co., Ltd. is 0.04, the probability value of Shushan District Urban Management Bureau is 0.05, and the probability value of the local street office is 0.03. By traversing each probability value in the distribution, the candidate responsible department corresponding to the maximum probability is found and determined as the first department. In the above example, the first department is Hefei Municipal Engineering Management Office. At the same time, the final responsible department field corresponding to the related work order request node is extracted from the historical work order completion record. This field records the name of the department that ultimately assumes the responsibility for handling the work order after the actual business process. For example, the final responsible department of this work order is Hefei Gas Group Co., Ltd. Next, the effective sample node screening operation is performed. Each related work order request node is checked to see if it meets two conditions at the same time. The first condition is that the first department and the final responsible department are different, that is, the most likely department judged by the preset department classification model based on the text semantics is different from the actual department that ultimately assumes responsibility. The second condition is that the final responsible department is the same entity as the project event responsible entity of the project event node associated with the related work order request node, that is, the actual responsible department is exactly the construction or responsible unit of the project event. The related work order request nodes that meet the above two conditions are marked as effective sample nodes. The above road damage work order is included in the effective sample nodes because the first department is the Municipal Management Office and the final responsible department is the Gas Group, and the Gas Group is the responsible entity of the gas emergency repair project event. However, work orders with the same first department and final responsible department or work orders with a final responsible department that is unrelated to the responsible entity of the project event are not included in the effective sample nodes.For each valid sample node, the original confidence distribution of the responsible department output by the preset department classification model is extracted as the first input feature, and the spatiotemporal matching degree between it and the corresponding engineering event node is extracted as the second input feature. Next, a probability reassignment operation is performed on the confidence distribution of the responsible department corresponding to the valid sample node to construct the target probability distribution. Specifically, the original probability value corresponding to the final responsible department is increased to a preset high probability threshold, such as 0.9, while maintaining the relative proportion of probability values ​​among the other candidate responsible departments. The sum of the original probability values ​​of these departments is proportionally compressed into a remaining probability space complementary to the preset high probability threshold, such as 0.1. Taking a road damage work order as an example, in the original distribution, the probability value of the gas group is 0.04, the probability value of the municipal management office is 0.88, the probability value of the urban management bureau is 0.05, and the probability value of the street office is 0.03. After reassignment, the gas group becomes the final responsible department. The probability value of the responsible department is increased to 0.9. The original sum of the probabilities of the remaining three departments is 0.96. Among them, the proportion of the Municipal Administration Office is 0.88 divided by 0.96, which is approximately 0.917; the proportion of the Urban Management Bureau is 0.05 divided by 0.96, which is approximately 0.052; and the proportion of the Subdistrict Office is 0.03 divided by 0.96, which is approximately 0.031. Multiplying these three relative proportions by the remaining probability space of 0.1, we get new probability values ​​of 0.0917 for the Municipal Administration Office, 0.0052 for the Urban Management Bureau, and 0.0031 for the Subdistrict Office. The adjusted confidence distribution of the responsible department is used as the target probability distribution. This generates a sample in the causal effect correction training sample set, which consists of the original confidence distribution of the responsible department (first input feature), the spatiotemporal matching degree (second input feature), and the target probability distribution of the training label. The above operation is repeated for all valid sample nodes to complete the construction of the causal effect correction training sample set.

[0036] Subsequently, an initial causal effect correction network is constructed. Structurally, this network sequentially includes a feature concatenation layer, a fully connected hidden layer, and a probability output layer. The feature concatenation layer is responsible for concatenating the first input feature, i.e., the confidence distribution vector of the responsible department, with the second input feature, i.e., the spatiotemporal matching degree scalar, in the same dimension to form a longer feature vector. The fully connected hidden layer contains several layers of nonlinear transformation units to perform multi-layer neural computation on the concatenated vector to extract the implicit mapping law between the spatiotemporal matching degree and the probability distribution adjustment magnitude. The probability output layer is responsible for mapping the feature vector output by the hidden layer back to a dimensional space with the same number of candidate responsible departments and outputting a complete corrected department probability distribution vector through a normalization function. The initial causal effect correction network is trained under supervision using the pre-constructed causal effect correction training sample set. In each training iteration, the first and second input features of the sample are input into the network for forward propagation to obtain the corrected departmental probability distribution. The cross-entropy loss function value between the corrected departmental probability distribution and the target probability distribution of the sample is calculated. This loss value reflects the difference between the current output correction result and the ideal correction result of the network. The gradient information of the loss value is backpropagated along the network structure layer by layer through the backpropagation mechanism to update the network parameters such as the connection weights and biases in the feature splicing layer, fully connected hidden layer and probability output layer. After multiple rounds of iterative optimization, the loss function value gradually decreases and tends to stabilize without significant fluctuations. At this time, the network parameters have converged, and the trained causal effect correction model is obtained by fixing the network parameters.

[0037] While the pre-defined departmental classification model performs well on routine requests, its decision-making mechanism makes it prone to systematic work order assignment bias when handling requests causally related to engineering events. The root cause lies in the fact that the model's input information is limited to the request text itself, failing to perceive the crucial spatiotemporal context of whether an ongoing engineering event exists at the location and time of the request. Therefore, when faced with a road damage request, regardless of whether there are records of gas repair excavation on that section of road, the model will assign the work order to the municipal department based on the semantics of the text. Therefore, this invention retains only work orders where the first department differs from the final responsible department, and the final responsible department matches the entity responsible for the engineering event, as valid samples. This selection criterion precisely identifies historical cases where the pre-defined departmental classification model incorrectly assigned work orders due to a lack of spatiotemporal context information, and these errors can be corrected by introducing the entity responsible for the engineering event. It excludes routine samples assigned by the model and noisy samples incorrectly assigned for other reasons, making the learning objective of the correction network highly focused.

[0038] In constructing the target probability distribution, instead of the extreme approach of setting the final responsible department to a single high value and all others to zero, the original relative proportions between the remaining candidate responsible departments are preserved. Only the probability quality is centrally transferred from non-responsible departments to the final responsible department. This construction method clarifies the correct dispatch target direction while respecting the relative characterization of the responsibility relationships between the candidate departments in the original text semantics. This allows the correction network to understand the semantic closeness between different departments during learning, rather than simply treating all non-responsible departments the same. The trained causal effect correction network actually masters an adaptive mapping function from spatiotemporal matching degree to the probability distribution adjustment magnitude. When the spatiotemporal matching degree is high, the network tends to transfer more probability quality towards the responsible party of the engineering event; when the spatiotemporal matching degree is low, the network tends to keep the original confidence distribution basically unchanged. This adaptive mechanism ensures that the correction model will not incorrectly correct routine requests without an engineering event background. The causal effect correction model and the preset department classification model constitute a collaborative two-stage dispatch decision architecture. The preset department classification model is responsible for providing a baseline confidence distribution based on text semantics, while the causal effect correction model is responsible for dynamically fine-tuning this baseline distribution based on the spatiotemporal matching degree retrieved in real time. The combination of the two enables the dispatch system to maintain a stable dispatch capability for routine requests while also possessing an accurate attribution capability for requests related to engineering events. This effectively reduces dispatch deviations and subsequent order cancellation issues caused by information gaps, fundamentally improving the first-hit rate and overall processing efficiency of government dispatch.

[0039] In another preferred embodiment of the present invention, the specific process of obtaining candidate project event nodes in step S5 is as follows: Based on the location entity of the current government service request, the spatial coordinates of the corresponding location entity are obtained by querying the preset city geographic information database. For example, if the location entity of the current service request is the intersection of Qianshan Road and Wangjiang West Road, the city geographic information database is searched using a place name and address matching engine to obtain the latitude and longitude coordinates of the center point of the intersection as 117.25123 degrees east longitude and 31.82245 degrees north latitude. A circular or rectangular buffer area is generated by expanding outward with a preset radius from the spatial coordinates of this point. For example, a square buffer area with a side length of 100 meters is formed by expanding 50 meters in each of the four directions of east, south, west and north from the point coordinates. The sequence of polygon vertex coordinates of the geographical area covered by the buffer is used as the geofence coordinate range corresponding to the location entity of the request. Using the geofence coordinate range as the spatial retrieval condition, all engineering event nodes in the urban event spatiotemporal graph are traversed. The spatial coordinate range attribute of each engineering event node is read one by one. Spatial geometric operation functions are called to determine whether there is a spatial intersection between the spatial coordinate range of the engineering event and the geofence coordinate range. The forms of intersection include the engineering event spatial coordinate range falling completely within the geofence coordinate range, the geofence coordinate range completely containing the engineering event spatial coordinate range, and the boundary of the two partially overlapping. All engineering event nodes with spatial intersection are extracted to form the first set of engineering event nodes. For example, within the 50-meter buffer zone around the intersection of Qianshan Road and Wangjiang West Road, a gas pipeline emergency repair and excavation engineering event node, a communication optical cable laying engineering event node, and a stormwater and sewage pipe network dredging engineering event node are retrieved. These three engineering event nodes together constitute the first set of engineering event nodes. Using the current government service request work order's request event type entity as the causal retrieval condition, further filtering is performed on each engineering event node in the first engineering event node set. The engineering event type entity attributes of each engineering event node are read, and the engineering event type entity is paired with the current request event type entity. It is then checked whether the pairing of the engineering event type entity and the request event type entity exists in the previously established event type causal association rule storage table. For example, if the current request event type entity is road damage, and the engineering event type entity for gas pipeline emergency repair excavation in the first engineering event node set is road excavation, the event type is queried. If the causal association rule storage table finds that a directed causal association edge from road excavation to road surface damage already exists, then the gas pipeline emergency repair excavation project event node is retained. If the communication optical cable laying project event type entity is weak current construction, and the query storage table finds that there is no causal association rule between weak current construction and road surface damage, and the stormwater and sewage pipe network dredging project event type entity is pipe maintenance, and the query storage table finds that there is also no causal association rule between pipe maintenance and road surface damage, then the communication optical cable laying and stormwater and sewage pipe network dredging project event nodes are removed. The project event nodes retained after this screening constitute the second set of project event nodes.Using the current government service request work order's request receipt timestamp as the time retrieval condition, a final layer of filtering is performed on each project event node in the second project event node set. The start time and end time attributes of each project event node are read; these two attributes together constitute the event occurrence time window for that project event. It is then determined whether the current request receipt timestamp falls within this event occurrence time window. The judgment condition is that the request receipt timestamp is greater than or equal to the project event start time and less than or equal to the project event end time. For example, if the current request receipt timestamp is April 14, 2026 at 10:00 AM... At 25 minutes past the hour, the gas pipeline emergency repair and excavation project started at 9:00 AM on April 12, 2026, and ended at 9:00 AM on April 19, 2026. After comparison, the timestamp of the request received completely fell within the time window. Therefore, the gas pipeline emergency repair and excavation project event node was finally determined as the candidate project event node corresponding to the current government service request work order. If there are multiple project event nodes in the second project event node set that meet the time retrieval conditions, they will all be retained as candidate project event nodes. If there are no project event nodes that meet the time retrieval conditions, the current request has no candidate project event nodes.

[0040] Spatial retrieval serves as the first layer of filtering. It utilizes the geometric intersection of geofence coordinates and the spatial coordinates of engineering events to quickly define the scope of engineering events related to the location of the complaint, excluding a large number of irrelevant events and thus narrowing the subsequent search scope and improving efficiency. Causal retrieval, as the second layer of filtering, uses causal association rules extracted from government business management data to semantically filter engineering events within the spatial scope. It retains only those engineering event types that, according to management regulations, may cause the current complaint type. For example, road excavation may cause road surface damage, while low-voltage electrical construction generally does not. This layer of filtering eliminates spatially overlapping but logically unrelated engineering events, preventing invalid associations from interfering with subsequent dispatch adjustments. Temporal retrieval, as the third layer of filtering, uses the inclusion relationship between the complaint receipt timestamp and the engineering event time window to determine whether the engineering event was still within its impact period when the complaint occurred. For example, for completed and handed-over engineering events, the responsible party has changed to the routine maintenance department and no longer bears direct responsibility for subsequent complaints. Temporal retrieval eliminates the misleading influence of historical engineering events on current dispatch decisions. The synergistic effect of the three-layer screening mechanism ensures that the finally obtained candidate engineering event nodes have a high degree of co-occurrence in time and space with the current request and reasonable causality in business logic. This provides accurate and reliable input information for subsequent calculation of time and space matching degree and causal effect correction model. When faced with each new request, the system can quickly perceive whether there are ongoing engineering events in the vicinity of the request location that may trigger the request type. This provides key factual basis for the time and space background correction of subsequent order dispatching decisions. It is the core retrieval link to realize the transformation from simple text semantic order dispatching to time and space causal perception order dispatching.

[0041] In another preferred embodiment of the present invention, the specific process for determining the target department for order dispatch in step S6 is as follows: First, the text of the current government service request is input into a pre-trained department classification model. The model outputs a confidence distribution of the responsible department, which consists of the probability values ​​of each candidate responsible department. For example, if the current request text is "There is a large pothole at the intersection of Qianshan Road and Wangjiang West Road, causing severe bumps when vehicles pass by," the confidence distribution of the responsible department output by the pre-trained department classification model is as follows: Hefei Municipal Engineering Management Office (probability 0.88), Hefei Gas Group Co., Ltd. (probability 0.04), Shushan District Urban Management Bureau (probability 0.05), and the local street office (probability 0.03). At the same time, the spatiotemporal matching degree value calculated in step S5 for the current government affairs request work order is read. This spatiotemporal matching degree is obtained by weighting and summing the spatial matching degree and temporal matching degree between the current request work order and the candidate project event node according to the preset weight coefficient. For example, the spatial overlap area ratio between the current request work order and the gas emergency repair project event node is 0.85, the temporal overlap ratio is 1.0, the preset spatial weight coefficient is 0.5, the preset temporal weight coefficient is 0.5, and the weighted summation of the spatiotemporal matching degree is 0.85 multiplied by 0.5 plus 1.0 multiplied by 0.5 equals 0.925. The aforementioned confidence distribution vector of the responsible department and the spatiotemporal matching degree scalar are used as joint inputs to the trained causal effect correction model. Within the causal effect correction network, a feature concatenation layer concatenates the probability values ​​in the confidence distribution vector of the responsible department with the spatiotemporal matching degree scalar in a dimensional manner to form an extended feature vector. For example, the probability values ​​of the four departments (0.88, 0.04, 0.05, 0.03) are concatenated with the spatiotemporal matching degree of 0.925 to form a feature vector containing five values. This concatenated vector is then passed through a fully connected hidden layer. The forward propagation and nonlinear activation transformation of multi-layer neurons in the hidden layer automatically extract the implicit mapping relationship between the spatiotemporal matching degree and the probability adjustment magnitude of each candidate department. Finally, the probability output layer maps the output of the hidden layer back to four dimensions with the same number of candidate responsible departments and outputs the corrected department probability distribution through a normalized exponential function. For example, the corrected department probability distribution is: Hefei Gas Group Co., Ltd. probability 0.75, Hefei Municipal Engineering Management Office probability 0.16, Shushan District Urban Management Bureau probability 0.06, and local street office probability 0.03. The system iterates through all candidate responsible departments and their corresponding probability values ​​in this corrected department probability distribution, comparing the magnitude of each probability value to find the candidate responsible department with the highest probability value. In the above example, the candidate responsible department corresponding to the maximum probability value of 0.75 is Hefei Gas Group Co., Ltd., so this department is determined as the target department for dispatching the current government service request work order, and the system dispatches the work order to the handling terminal of Hefei Gas Group Co., Ltd.If no candidate project event node is found in the S5 step of the current government service request work order, that is, the spatiotemporal matching degree is zero or there is no candidate project event node, the causal effect correction model will keep the original attribution department confidence distribution basically unchanged. At this time, the candidate responsible department corresponding to the maximum probability is still the Hefei Municipal Engineering Management Office, and the system will send the work order to this regular attribution department.

[0042] The pre-defined department classification model is trained based on the text-to-department mapping patterns of massive historical work orders, and can stably handle dispatch tasks for various routine requests. This serves as the baseline capability and fallback guarantee for the entire dispatch system. However, this model has an inherent information perception limitation; its input channel is limited to the request text itself, and it is completely unaware of the crucial spatiotemporal background information, such as whether an engineering event occurred within the time frame of the request. Therefore, it is prone to systematic dispatch bias when dealing with requests causally related to engineering events. The introduction of the causal effect correction model is precisely to fill this information perception gap. This model learns the correction pattern of spatiotemporal matching degree on dispatch results from historical valid samples. During the online inference stage, it can adaptively determine the direction and magnitude of adjustment to the confidence distribution of the original responsible department based on the actual spatiotemporal matching degree between the current request and the candidate engineering event nodes. When the spatiotemporal matching degree is high, the probability quality is concentrated towards the responsible party of the engineering event; when the spatiotemporal matching degree is low or there are no candidate engineering event nodes, the original distribution shape is basically maintained. This dual-model collaborative architecture, combining baseline and correction, enables the dispatch system to maintain a stable dispatch capability for routine requests while significantly enhancing its ability to accurately attribute requests causally related to engineering events. From the ultimate goal of the solution, this mechanism for determining the target department for dispatch directly impacts the first stage of the work order process, correcting dispatch deviations caused by information gaps at the source. This ensures that work orders that previously required on-site verification by municipal departments before being returned and reassigned can be accurately delivered to the responsible party for the engineering event upon initial dispatch. This effectively reduces intermediate steps in the cross-departmental flow of work orders and overall processing time, achieving simultaneous improvements in dispatch accuracy, response efficiency, and service quality.

[0043] The foregoing has provided a detailed description of one embodiment of the present invention, but this description is merely a preferred embodiment and should not be construed as limiting the scope of the invention. All equivalent variations and modifications made within the scope of the claims of this invention should still fall within the patent coverage of this invention.

Claims

1. A method for accurately dispatching government service requests based on a large-scale artificial intelligence model, characterized in that, Includes the following steps: S1. Obtain historical work order records of government services and registration data of urban engineering events, and establish a spatiotemporal map of urban affairs containing work order request nodes and engineering event nodes based on government business management standard data. S2, associate historical work orders with each engineering event node in the urban event spatiotemporal map, and calculate the spatiotemporal matching degree between each engineering event node and the associated work order request node. S3, extract the confidence distribution of the responsible department and the final responsible department output by the preset department classification model for each historical appeal work order, and construct a training sample set based on the spatiotemporal matching degree to generate a causal effect correction model; S4, obtain the location entity and event type entity of the current government service request work order, and obtain the corresponding geofence coordinate range based on the location entity. S5. Using the geographic fence coordinate range and the entity of the appeal event type as search conditions, perform a spatiotemporal joint query in the urban event spatiotemporal map to obtain the candidate project event node corresponding to the current government appeal work order, and calculate the spatiotemporal matching degree between the current government appeal work order and the candidate project event node. S6. Input the current government service request work order into the preset department classification model to obtain the confidence distribution of the responsible department. Input the confidence distribution of the responsible department and the spatiotemporal matching degree into the causal effect correction model to obtain the final confidence distribution of the responsible department. Select the department with the highest probability as the dispatch target department of the current government service request work order.

2. The method for accurately dispatching government service requests based on a large-scale artificial intelligence model according to claim 1, characterized in that, In S1, the specific construction process of the urban event spatiotemporal map is as follows: Acquire historical service request records and urban engineering event registration data; perform entity recognition on each historical service request record, extract the request location entity, request event type entity and request receipt timestamp, retrieve the request geofence coordinate range from the preset urban geographic information database based on the request location entity, and generate the service request node; For each urban engineering event registration data, entity recognition is performed to extract the engineering event type entity, engineering event location entity, engineering event responsible entity entity, engineering event start time, engineering event end time, and engineering event status attributes. Based on the engineering event location entity, a preset urban geographic information database is searched to obtain the spatial coordinate range of the engineering event and generate engineering event nodes. The government affairs business management standard data is split into sentences to obtain a set of standard text sentences. Each standard text sentence is input into a pre-trained semantic role labeling model to identify the agent and the patient. The engineering event type entity corresponding to the agent and the appeal event type entity corresponding to the patient are extracted. A directed causal relationship edge is established from the engineering event type entity to the appeal event type entity. The engineering event type entity and appeal event type entity with a directed causal relationship edge are stored as event type causal relationship rules. In the urban event spatiotemporal graph, the work order request node and the engineering event node are created. When the request geofence coordinate range of the work order request node includes the engineering event spatial coordinate range of the engineering event node, a first connection edge is established. When the engineering event type entity of the engineering event node and the request event type entity of the work order request node satisfy the event type causal association rule, a second connection edge is established.

3. The method for accurately dispatching government service requests based on a large-scale artificial intelligence model according to claim 1, characterized in that, In S2, the specific process of associating historical request work orders is as follows: Traverse each engineering event node in the urban event spatiotemporal graph. For the currently traversed engineering event node, obtain its engineering event spatial coordinate range and the event occurrence time window composed of the start time and end time of the engineering event. In the work order request nodes of the urban event spatiotemporal map, work order request nodes that have spatial intersection between the geographical fence coordinate range of the request and the spatial coordinate range of the project event are retrieved, and the retrieved work order request nodes are used as candidate associated work order request nodes. For each candidate related work order request node, obtain its request receipt timestamp, match the request receipt timestamp with the event occurrence time window of the project event node, and if the request receipt timestamp is within the event occurrence time window, then determine the candidate related work order request node as a related work order request node.

4. The method for accurately dispatching government service requests based on a large-scale artificial intelligence model according to claim 3, characterized in that, In S2, the specific calculation process of the spatiotemporal matching degree is as follows: Calculate the percentage of spatial overlap between the geographic fence coordinates of each associated work order request node and the spatial coordinates of the engineering event node, and use this as the spatial matching degree. Calculate the time overlap ratio between the request receipt timestamp of each associated work order request node and the event occurrence time window of the project event node, as the time matching degree; The spatial matching degree and the temporal matching degree are weighted and summed according to a preset weighting coefficient to obtain the spatiotemporal matching degree.

5. The method for accurately dispatching government service requests based on a large-scale artificial intelligence model according to claim 1, characterized in that, In S3, the specific construction process of the preset department classification model is as follows: Obtain historical service request records, perform entity recognition on each historical service request record, extract the entity of the request event type and the content of the request text, and obtain the final responsible department for the historical service request record. The text of each historical appeal work order is used as the input sample, and the corresponding final responsible department is used as the classification label to construct a department classification training sample set. An initial neural network model is constructed, which includes a text encoding layer and a fully connected classification layer. The text encoding layer is used to semantically encode the input appeal text content to obtain a text semantic vector. The fully connected classification layer is used to linearly map the text semantic vector and output the predicted probability distribution of each candidate responsible department. The initial neural network model is trained under supervision using the department classification training sample set. The cross-entropy loss between the predicted probability distribution output by the fully connected classification layer and the classification label corresponding to the final responsible department is used as the loss function. The model parameters of the initial neural network model are iteratively updated through the backpropagation algorithm until the loss function converges, thus obtaining the pre-trained department classification model.

6. The method for accurately dispatching government service requests based on a large-scale artificial intelligence model according to claim 1, characterized in that, In S3, the specific construction process of the causal effect correction model is as follows: Obtain the associated work order request nodes and associated project event nodes corresponding to each historical request work order, and obtain the spatiotemporal matching degree between the associated work order request node and the project event node. Input the request text content corresponding to the related work order request node into the preset department classification model to obtain the confidence distribution of the responsible department. Determine the candidate responsible department corresponding to the maximum probability in the confidence distribution of the responsible department as the first department, and obtain the final responsible department corresponding to the related work order request node. The related work order request nodes whose first department is different from the final responsible department and whose final responsible department is the same as the responsible entity of the related engineering event node are considered as valid sample nodes. Using the confidence distribution of the responsible department corresponding to the effective sample node as the first input feature and the spatiotemporal matching degree corresponding to the effective sample node as the second input feature, the confidence distribution of the responsible department corresponding to the effective sample node is probability redistributed. The probability value of the final responsible department is increased to a preset high probability threshold, and the probability values ​​of the remaining candidate responsible departments are reduced to a sum that complements the preset high probability threshold according to their original relative proportions. The redistributed confidence distribution of the responsible department is used as the target probability distribution to generate a causal effect correction training sample set. An initial causal effect correction network is constructed, comprising a feature concatenation layer, a fully connected hidden layer, and a probability output layer. The initial causal effect correction network is trained under supervision using the causal effect correction training sample set. The cross-entropy loss between the corrected departmental probability distribution output by the probability output layer and the target probability distribution is used as the loss function. The network parameters are updated through backpropagation until convergence, thus obtaining the causal effect correction model.

7. The method for accurately dispatching government service requests based on a large-scale artificial intelligence model according to claim 2, characterized in that, In step S5, the specific process for obtaining candidate project event nodes is as follows: Based on the location entity of the current government service request work order, the spatial coordinates of the location entity are obtained by querying the preset city geographic information database. A buffer zone is generated with the spatial coordinates as the center and a preset radius. The coordinate range of the buffer zone is used as the geofence coordinate range of the location entity. Using the geographic fence coordinate range as the spatial retrieval condition, the engineering event nodes whose spatial coordinate ranges intersect with the geographic fence coordinate range are retrieved in the urban event spatiotemporal map to obtain the first set of engineering event nodes. Using the current government service request work order's request event type entity as the causal retrieval condition, filter the first set of engineering event nodes in which the engineering event type entity and the request event type entity satisfy the event type causal association rule to obtain the second set of engineering event nodes. Using the current government service request work order's request receipt timestamp as the time retrieval condition, filter the second set of engineering event nodes whose event occurrence time window contains the request receipt timestamp, and use them as candidate engineering event nodes.

8. The method for accurately dispatching government service requests based on a large-scale artificial intelligence model according to claim 1, characterized in that, In step S6, the specific process for determining the target department for order placement is as follows: Input the text of the current government service request work order into the preset department classification model to obtain the confidence distribution of the responsible department; input the confidence distribution of the responsible department and the spatiotemporal matching degree between the current government service request work order and the candidate project event node into the causal effect correction model to obtain the corrected department probability distribution; traverse the probability values ​​of each candidate responsible department in the corrected department probability distribution, and determine the candidate responsible department corresponding to the maximum probability as the target department for dispatching the work order.

Images