A postal industry complaint service system based on multi-modal fusion

Abstract

The application discloses a postal industry complaint traffic service system based on multi-modal fusion, comprising: a multi-modal complaint feature acquisition module, generating multi-modal complaint features; a knowledge service graph and path template construction module, constructing a knowledge service graph and a path template; a probability distribution modeling module, constructing three types of probability distributions and introducing complaint priority and service capacity; a hierarchical cost and constraint construction module, constructing hierarchical cost and constraint; a hierarchical three-edge Monge-Kantorovich optimal transport solution module, solving a three-edge optimal transport model to obtain a hierarchical transport plan; an incremental and feedback update module, updating the hierarchical transport plan based on new complaints and processing results; a decision generation and traffic service control module, outputting an explanation path and a decision result. The application unifies complaint priority, path and resource capacity by hierarchical three-edge optimal transport, and improves the explainability of scheduling and decision-making.

Description

Technical Field

[0001] This invention relates to the field of intelligent call dispatching technology, and in particular to a postal industry complaint call service system based on multimodal fusion. Background Technology

[0002] The current scale of postal services and the volume of complaints continue to grow. Postal complaint call service systems generally rely on a combination of human agents and voice interaction platforms for acceptance, triage, and recording. Existing systems typically rely solely on call audio and speech-to-text transcription, using keyword matching, simple rule engines, and fixed menu trees for business identification and flow. This approach limits the utilization of the four types of structured information contained in complaint data: waybill tracking, time points, delivery type, and customer level. Acoustic features from calls are generally only used for basic speech recognition and are difficult to model in a unified manner with textual and business features. This results in the incomplete utilization of multimodal complaint information, a high dependence on agent experience in call allocation, and a lack of refined characterization of the complaint processing chain.

[0003] In the field of intelligent customer service, existing solutions attempt to introduce business knowledge bases, knowledge graphs, and retrieval question-answering technologies to abstract postal complaint rules into nodes and relationships. Through intent recognition and slot filling, user expressions are mapped to business terms and processing paths. Common practices often employ rule-based process engines, classification models, and recommendation models, selecting paths and assigning agents based on text similarity, word vector distance, and historical frequency. However, these approaches lack unified probabilistic modeling of multimodal complaint characteristics, complaint business data, and service resource status. Service resource scheduling typically relies on simple allocation based on queue length, skill tags, and preset priorities. It is difficult to integrate complaint priority, service capacity, and business process structure into a unified framework for joint optimization, making it challenging to guarantee call load balancing and prioritized response to critical complaints in complex scenarios.

[0004] In the areas of multimodal alignment, recommendation matching, and resource allocation, the Monge-Kantorovich optimal transport model and related optimal transport algorithms have been applied to image matching, embedding alignment, and recommendation scoring fusion scenarios. However, existing solutions for postal service complaint handling typically only model transport between two-dimensional distributions, rarely simultaneously depicting the hierarchical coupling relationships between complaint samples, path templates, and service resources. They also fail to fully integrate structural constraints in the knowledge service graph with path feasibility constraints in the business process. Existing systems lack a trilateral optimal transport modeling mechanism based on three probability distributions and hierarchical cost structures. They lack a mechanism for incremental and feedback updates during the operation of complaint handling services, combining newly added multimodal complaint features and complaint processing results. Furthermore, they lack a unified method to generate traceable and interpretable paths from hierarchical transport plans and drive complaint processing decisions, making it difficult to achieve refined call scheduling and service resource allocation while maintaining business semantic consistency.

[0005] Therefore, how to provide a postal industry complaint call service system based on multimodal fusion is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0006] One objective of this invention is to propose a postal industry complaint call service system based on multimodal fusion. This invention constructs multimodal complaint features that integrate call voice, speech-to-text, and complaint business data, establishes a knowledge service graph and path template set, and designs a joint modeling and incremental feedback solution mechanism with three types of probability distributions, hierarchical cost structure, graph structure constraints, and path feasibility constraints. This enables refined call allocation from complaint samples to service resource nodes and interpretable complaint processing decision output, with advantages such as high accuracy in complaint priority identification, reasonable service resource allocation, and traceable processing paths.

[0007] A postal service complaint handling system based on multimodal fusion according to an embodiment of the present invention includes:

[0008] The multimodal appeal feature acquisition module is used to acquire and align call voice, speech-to-text, and appeal business data, extract acoustic features, text features, and business features, and generate multimodal appeal features.

[0009] The knowledge service graph and path template construction module is used to construct a knowledge service graph containing business knowledge nodes and service resource nodes based on postal complaint business rule data and service resource configuration data, and to search and generate a set of path templates on the knowledge service graph;

[0010] The probability distribution modeling module is used to construct a first probability distribution, a second probability distribution, and a third probability distribution based on multimodal appeal features, path template sets, and service resource nodes, respectively. Appeal priority weights are set in the first probability distribution, and service capacity parameters are set in the third probability distribution.

[0011] The hierarchical cost and constraint construction module is used to construct a hierarchical cost structure based on three types of probability distributions and knowledge service graphs, which describes the mapping relationship from multimodal appeal features to path templates and then to service resource nodes, and to generate graph structure constraints and path feasibility constraints.

[0012] The hierarchical trilateral Monge-Kantorovich optimal transport solution module is used to establish a hierarchical trilateral Monge-Kantorovich optimal transport model based on three types of probability distributions, hierarchical cost structure, graph structure constraints and path feasibility constraints, and obtain the hierarchical transport plan by solving the model.

[0013] The incremental and feedback update module is used to update the first probability distribution, the second probability distribution, the third probability distribution, and the hierarchical cost structure based on the newly added multimodal appeal features and appeal processing results during the operation of the appeal call service. It performs incremental and feedback solutions on the hierarchical trilateral Monge-Kantorovich optimal transmission model to generate an updated hierarchical transmission plan.

[0014] The decision generation and call service control module is used to determine the target path template and target service resource node for each multimodal appeal feature based on the updated hierarchical transport plan, construct the interpretation path, and output the appeal processing decision result.

[0015] Optionally, modules can be integrated using the following methods:

[0016] Acquire call voice recordings, speech-to-text transcripts, and appeal data to generate multimodal appeal features;

[0017] Construct a knowledge service graph containing business knowledge nodes and service resource nodes, and generate a set of path templates;

[0018] Based on multimodal appeal features, path template set and service resource nodes, a first probability distribution, a second probability distribution and a third probability distribution are constructed respectively. Appeal priority weights are set in the first probability distribution and service capacity parameters are set in the third probability distribution.

[0019] A hierarchical cost structure is constructed based on three types of probability distributions and knowledge service graphs, and graph structure constraints and path feasibility constraints are generated.

[0020] Based on three types of probability distributions, hierarchical cost structure, graph structure constraints, and path feasibility constraints, a hierarchical trilateral Monge-Kantorovich optimal transport model is established, and the hierarchical transport plan is obtained by solving it.

[0021] During the appeal call service process, based on the newly added multimodal appeal features and appeal processing results, the hierarchical trilateral Monge-Kantorovich optimal transmission model is incrementally and feedback-updated to generate an updated hierarchical transmission plan.

[0022] Based on the updated hierarchical transportation plan, a target path template and target service resource node are determined for each multimodal appeal feature, an explanatory path is generated, and the appeal processing decision result is output.

[0023] Optionally, the generation of the multimodal appeal features specifically includes:

[0024] Collect call voice recordings with call identifiers and record the call start time and call end time. Store the recordings in segments according to the call identifiers to form a call voice sequence. The call voice sequence is organized using the call identifiers and time order as indexes.

[0025] The call audio sequence is subjected to denoising, pre-emphasis, framing and windowing to obtain an audio frame sequence, which maintains the correspondence with the call identifier and time sequence.

[0026] The speech frame sequence is input into the speech recognition model, which generates a character output sequence and concatenates it into speech-to-text in chronological order. The speech recognition model is an acoustic and language joint modeling structure based on deep neural networks, which is used to map the speech frame sequence into a character probability distribution and decode it into a text sequence.

[0027] The complaint business data is obtained from the business system. The complaint business data includes customer identifier, delivery business type, waybill identifier, node time information and complaint type identifier. The complaint business data is organized with waybill identifier and node time information as primary key fields.

[0028] Based on call identifiers, waybill identifiers, and time alignment rules, the speech-to-text is associated with the appeal business data to generate a call-level appeal record. The call-level appeal record includes a speech-to-text fragment associated with the same call identifier and appeal business data fields.

[0029] Acoustic features, text features, and business features are extracted from call-level appeal records. The acoustic features, text features, and business features are combined with call identifiers in chronological order to construct a call-level multimodal feature vector. The call-level multimodal feature vector is used as the input of the multimodal appeal features to the first probability distribution.

[0030] Optionally, the generation of the path template set specifically includes:

[0031] Obtain postal complaint business rule data and service resource configuration data for constructing a knowledge service graph. The postal complaint business rule data includes complaint category code, business clause code, responsible entity code, and handling stage code. The service resource configuration data includes service resource code, service group code, and service capability parameters.

[0032] A set of business knowledge nodes is constructed based on the postal complaint business rules data. In the set of business knowledge nodes, complaint category nodes, business clause nodes, responsible entity nodes and handling link nodes are set, and each business knowledge node is assigned a code and attribute information.

[0033] A service resource node set is constructed based on the service resource configuration data. Service resource nodes are set in the service resource node set, and the service resource code, service group code and service capability parameters are written into the attribute information of the corresponding service resource node.

[0034] Based on the postal complaint business rules data, a set of rule-related edges is generated in the set of business knowledge nodes. Based on the service resource configuration data, a set of service-related edges is generated between the set of business knowledge nodes and the set of service resource nodes. The set of rule-related edges establishes connections between corresponding business knowledge nodes according to the order of complaint category code, business clause code, responsible entity code, and handling stage code in the postal complaint business rules. The set of service-related edges establishes connections between corresponding handling stage nodes and service resource nodes according to the handling stage code and service resource code in the service resource configuration data.

[0035] A knowledge service graph is constructed using a set of business knowledge nodes, a set of service resource nodes, a set of rule-related edges, and a set of service-related edges. Path search is performed in the knowledge service graph starting from the appeal category node and ending at the service resource node. Based on preset node access order constraints and path length constraints, a node sequence is obtained from the appeal category node through the business clause node, the responsible entity node, and the handling link node to the service resource node. Each node sequence is defined as a path template, and the path templates are combined into a path template set.

[0036] Optionally, the construction of the first probability distribution, the second probability distribution, and the third probability distribution specifically includes:

[0037] Based on the multimodal appeal features, a unique index is assigned to each multimodal appeal feature. A correspondence is established between the unique index and the call identifier and time interval. An appeal sample index set is constructed. A first probability distribution is set on the appeal sample index set. The first probability distribution assigns an initial probability value to each multimodal appeal feature in units of the unique index.

[0038] A unique index is assigned to each path template based on the path template set. A correspondence is established between the unique index and the appeal category node, business clause node, responsible entity node and handling link node contained in the path template. A path template index set is constructed. A second probability distribution is set on the path template index set. The second probability distribution assigns an initial probability value to each path template in units of unique index.

[0039] A unique index is assigned to each service resource node based on the service resource node. A correspondence is established between the unique index and the service resource code and the service group code. A service resource index set is constructed. A third probability distribution is set on the service resource index set. The third probability distribution assigns an initial probability value to each service resource node in units of the unique index.

[0040] The appeal priority weight is calculated based on the appeal business data associated with the multimodal appeal characteristics. The appeal priority weight is written into the first probability distribution and used as a weighting coefficient to adjust the initial probability value in the first probability distribution. The appeal business data includes the urgency of timeliness, amount, customer level and number of complaints.

[0041] The service capacity parameters are calculated based on the service resource configuration data, written into a third probability distribution, and used as weighting coefficients to adjust the initial probability values ​​in the third probability distribution. The service resource configuration data includes the number of available seats, the number of allowed concurrent calls, and the target service duration. The calculation of the service capacity parameters includes: constructing a capacity scoring function based on the number of available seats, the number of allowed concurrent calls, and the target service duration; generating a capacity score for each service resource node; and normalizing the capacity scores to obtain the service capacity parameters.

[0042] Optionally, the construction of the hierarchical cost structure and the generation of graph structure constraints and path feasibility constraints specifically include:

[0043] The first-level cost matrix is ​​constructed based on the set of appeal sample indexes and the set of path template indexes and written into the hierarchical cost structure. The first-level cost matrix is ​​used to describe the cost relationship between multimodal appeal features and path templates. The first-level cost matrix uses the appeal sample index as the row index and the path template index as the column index. Each cost element is calculated based on the semantic differences, business field differences, time location differences, and probability values ​​in the first probability distribution and the second probability distribution between multimodal appeal features and path templates.

[0044] A second-layer cost matrix is ​​constructed based on the path template index set and the service resource index set and written into the hierarchical cost structure. The second-layer cost matrix is ​​used to describe the cost relationship between the path template and the service resource node. The second-layer cost matrix uses the path template index as the row index and the service resource index as the column index. Each cost element is calculated based on the differences in service capabilities, load levels, target service durations, and probability values ​​in the second and third probability distributions between the path template and the service resource node.

[0045] In the hierarchical cost structure, the index range and hierarchical relationship between the first-level cost matrix and the second-level cost matrix are registered, so that the first-level cost matrix and the second-level cost matrix establish a corresponding mapping relationship in the hierarchical cost structure according to the appeal sample index set, the path template index set and the service resource index set;

[0046] Based on the set of business knowledge nodes, the set of service resource nodes, the set of rule association edges, and the set of service association edges in the knowledge service graph, graph structure constraints are generated. The graph structure constraints stipulate that the combination of appeal sample index and path template index with valid values ​​is allowed in the first-level cost matrix, and the combination of path template index and service resource index with valid values ​​is allowed in the second-level cost matrix. The cost elements corresponding to the index combinations that do not satisfy the connection relationship of the knowledge service graph are marked as invalid in a preset way.

[0047] The path feasibility constraints are generated based on the path template set. The path feasibility constraints are based on the order of the appeal category node, business clause node, responsible entity node, handling link node and service resource node in each path template. For the index of the path template node combination in the first-level cost matrix and the second-level cost matrix, the corresponding cost element is reserved as a candidate value. For the index of the path template node combination that does not belong to any path template node combination, the corresponding cost element is marked as a prohibited value according to a preset method.

[0048] Optionally, the establishment of the hierarchical trilateral Monge-Kantorovich optimal transmission model specifically includes:

[0049] A first-level transport relationship is established between the first probability distribution and the second probability distribution, and a second-level transport relationship is established between the second probability distribution and the third probability distribution. The first-level transport relationship and the second-level transport relationship are set as the transport variable set of the hierarchical trilateral Monge-Kantorovich optimal transport model. The first-level transport relationship is used to represent the transport volume allocation from the appeal sample index to the path template index, and the second-level transport relationship is used to represent the transport volume allocation from the path template index to the service resource index. The hierarchical trilateral Monge-Kantorovich optimal transport model is a constrained optimization model that establishes two-level transport relationships on three types of probability distributions and takes the hierarchical cost structure as the objective.

[0050] The allocation of the first-layer transport relationship in the appeal sample dimension is limited by the first probability distribution, the balance between the first-layer and second-layer transport relationships in the path template dimension is limited by the second probability distribution, and the allocation of the second-layer transport relationship in the service resource dimension is limited by the third probability distribution, so that the sum of the transport volume associated with each appeal sample index is equal to the corresponding probability value in the first probability distribution, the sum of the first-layer transport volume associated with each path template index and the sum of the second-layer transport volume are equal to the corresponding probability value in the second probability distribution, and the sum of the transport volume associated with each service resource index is equal to the corresponding probability value in the third probability distribution.

[0051] The objective function of the hierarchical trilateral Monge-Kantorovich optimal transport model is constructed based on the hierarchical cost structure. The first-layer cost structure part is weighted and summed with the corresponding elements in the first-layer transport relation, and the second-layer cost structure part is weighted and summed with the corresponding elements in the second-layer transport relation. The two weighted sums are then combined into the total cost.

[0052] The graph structure constraint and path feasibility constraint are incorporated into the hierarchical trilateral Monge-Kantorovich optimal transport model. For the combination of appeal sample index and path template index and the combination of path template index and service resource index that do not satisfy the graph structure constraint and path feasibility constraint, the transport volume is specified to be zero. For the index combination that satisfies the graph structure constraint and path feasibility constraint, the transport volume is specified to be a non-negative real number.

[0053] The numerical solution process is performed under the joint constraints of total cost, first probability distribution, second probability distribution, third probability distribution, graph structure constraint and path feasibility constraint to obtain the first-level transportation plan corresponding to the first-level transportation relationship and the second-level transportation plan corresponding to the second-level transportation relationship. The first-level transportation plan and the second-level transportation plan are then combined to form a hierarchical transportation plan.

[0054] Optionally, the generation of the updated hierarchical transportation plan specifically includes:

[0055] During the operation of the appeal call service, new multimodal appeal features are received and the appeal sample index set and the first probability distribution are expanded. The expansion includes: assigning an appeal sample index to the new multimodal appeal features, adding the appeal sample index to the appeal sample index set, and setting an initial probability value and appeal priority weight in the first probability distribution based on the associated appeal business data.

[0056] The service resource nodes and the third probability distribution are updated based on the current service resource usage status and service resource configuration data. The update includes: writing the latest service capability parameters and service capacity parameters to the service resource nodes, and adjusting the probability value of the corresponding service resource index in the third probability distribution.

[0057] The previous batch of hierarchical transport plans is used as the initial transport relationship of the hierarchical trilateral Monge-Kantorovich optimal transport model. Incremental solutions are performed on the expanded set of complaint sample indexes and the updated set of service resource indexes. The incremental solutions are limited to updating the transport volume only for index combinations that are associated with the new complaint sample indexes and changes in service capacity parameters. The transport volume of the previous batch is retained for the other index combinations.

[0058] After the appeal is processed, feedback statistics are constructed based on the appeal processing results, and the hierarchical cost structure and three types of probability distributions are adjusted. The appeal processing results include processing time, number of transfers, customer satisfaction index and duplicate appeal marker. The feedback statistics are used to correct the cost parameters associated with the path template index and service resource index, as well as the probability values ​​in the first probability distribution, the second probability distribution and the third probability distribution.

[0059] Under the constraints of the incrementally updated first, second, and third probability distributions, the updated hierarchical cost structure, graph structure constraints, and path feasibility constraints, the hierarchical trilateral Monge-Kantorovich optimal transport model is re-solved to obtain the updated first-layer transport plan and the updated second-layer transport plan. The updated first-layer transport plan and the updated second-layer transport plan are then combined to form the updated hierarchical transport plan.

[0060] Optionally, the generation of the appeal processing decision result specifically includes:

[0061] Based on the updated hierarchical transportation plan, for each multimodal appeal feature, a target path template index is selected in the first-level transportation plan, and a target service resource index is selected in the second-level transportation plan. The selection is based on the transportation volume associated with each path template index and each service resource index, as well as a preset threshold rule.

[0062] The corresponding path template is retrieved from the path template set according to the target path template index, and an explanatory path is constructed in combination with the target service resource index. The explanatory path records the appeal category node, business terms node, responsible entity node, handling process node and service resource node in node order.

[0063] Based on the interpretation path, an appeal handling decision result is generated. The appeal handling decision result includes at least the appeal category determination result, the business terms matching result, the responsible entity identification result, the handling process sequence, and the service resource allocation result.

[0064] The explanation path and appeal processing decision results are written into the business records of the appeal call service system, and the appeal processing decision results are output to the call service terminal according to the target service resource index.

[0065] The beneficial effects of this invention are:

[0066] This invention constructs multimodal appeal features that integrate call voice, speech-to-text, and appeal business data, and organizes appeal category nodes, business clause nodes, responsible entity nodes, handling process nodes, and service resource nodes in a unified knowledge service graph. This achieves integrated modeling of appeal semantic information and business structure information. Compared with traditional appeal call systems that rely solely on text matching and rule-based traffic routing, this invention can more accurately distinguish appeal scenarios and business requests of different priorities, reducing reliance on manual intervention and mis-routing.

[0067] This invention introduces a first probability distribution, a second probability distribution, and a third probability distribution. The first probability distribution sets the appeal priority weight, and the third probability distribution sets the service capacity parameter. By combining the hierarchical cost structure, graph structure constraints, and path feasibility constraints, a hierarchical three-sided Monge-Kantorovich optimal transmission model is constructed. By solving the hierarchical transmission plan, the urgency of the appeal, the business rule path, and the service resource load status are considered simultaneously within the same framework, achieving more refined traffic allocation and service resource assignment. Compared with a simple queue allocation strategy, it can improve the response speed of high-priority appeals and balance the workload of different service groups.

[0068] This invention incrementally updates and adjusts the hierarchical trilateral Monge-Kantorovich optimal transmission model based on newly added multimodal appeal features and appeal processing results during the operation of the appeal call service. It also uses the updated hierarchical transmission plan to generate an interpretation path and appeal processing decision result for each multimodal appeal feature, enabling call allocation and path selection to be continuously corrected according to business statistics and service resource status. At the same time, it provides a clear decision link for each appeal, including the appeal category determination result, business clause matching result, responsible party identification result, and handling process sequence, making it easier for operators to trace the decision basis, optimize business rules, and carry out refined management. Attached Figure Description

[0069] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0070] Figure 1 The flowchart below shows a postal industry complaint call service system based on multimodal fusion proposed in this invention.

[0071] Figure 2 This is a schematic diagram of hierarchical trilateral Monge-Kantorovich optimal transmission modeling for a postal industry complaint call service system based on multimodal fusion proposed in this invention. Detailed Implementation

[0072] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0073] refer to Figure 1-2 A postal service complaint handling system based on multimodal fusion includes:

[0074] The multimodal appeal feature acquisition module is used to acquire and align call voice, speech-to-text, and appeal business data, extract acoustic features, text features, and business features, and generate multimodal appeal features.

[0075] The knowledge service graph and path template construction module is used to construct a knowledge service graph containing business knowledge nodes and service resource nodes based on postal complaint business rule data and service resource configuration data, and to search and generate a set of path templates on the knowledge service graph;

[0076] The probability distribution modeling module is used to construct a first probability distribution, a second probability distribution, and a third probability distribution based on multimodal appeal features, path template sets, and service resource nodes, respectively. Appeal priority weights are set in the first probability distribution, and service capacity parameters are set in the third probability distribution.

[0077] The hierarchical cost and constraint construction module is used to construct a hierarchical cost structure based on three types of probability distributions and knowledge service graphs, which describes the mapping relationship from multimodal appeal features to path templates and then to service resource nodes, and to generate graph structure constraints and path feasibility constraints.

[0078] The hierarchical trilateral Monge-Kantorovich optimal transport solution module is used to establish a hierarchical trilateral Monge-Kantorovich optimal transport model based on three types of probability distributions, hierarchical cost structure, graph structure constraints and path feasibility constraints, and obtain the hierarchical transport plan by solving the model.

[0079] The incremental and feedback update module is used to update the first probability distribution, the second probability distribution, the third probability distribution, and the hierarchical cost structure based on the newly added multimodal appeal features and appeal processing results during the operation of the appeal call service. It performs incremental and feedback solutions on the hierarchical trilateral Monge-Kantorovich optimal transmission model to generate an updated hierarchical transmission plan.

[0080] The decision generation and call service control module is used to determine the target path template and target service resource node for each multimodal appeal feature based on the updated hierarchical transport plan, construct the interpretation path, and output the appeal processing decision result.

[0081] In this embodiment, the modules are interconnected using the following method:

[0082] Acquire call voice recordings, speech-to-text transcripts, and appeal data to generate multimodal appeal features;

[0083] Construct a knowledge service graph containing business knowledge nodes and service resource nodes, and generate a set of path templates;

[0084] Based on multimodal appeal features, path template set and service resource nodes, a first probability distribution, a second probability distribution and a third probability distribution are constructed respectively. Appeal priority weights are set in the first probability distribution and service capacity parameters are set in the third probability distribution.

[0085] A hierarchical cost structure is constructed based on three types of probability distributions and knowledge service graphs, and graph structure constraints and path feasibility constraints are generated.

[0086] Based on three types of probability distributions, hierarchical cost structure, graph structure constraints, and path feasibility constraints, a hierarchical trilateral Monge-Kantorovich optimal transport model is established, and the hierarchical transport plan is obtained by solving it.

[0087] During the appeal call service process, based on the newly added multimodal appeal features and appeal processing results, the hierarchical trilateral Monge-Kantorovich optimal transmission model is incrementally and feedback-updated to generate an updated hierarchical transmission plan.

[0088] Based on the updated hierarchical transportation plan, a target path template and target service resource node are determined for each multimodal appeal feature, an explanatory path is generated, and the appeal processing decision result is output.

[0089] In this embodiment, the generation of the multimodal appeal features specifically includes:

[0090] Collect call voice recordings with call identifiers and record the call start time and call end time. Store the recordings in segments according to the call identifiers to form a call voice sequence. The call voice sequence is organized using the call identifiers and time order as indexes.

[0091] The call audio sequence is subjected to denoising, pre-emphasis, framing and windowing to obtain an audio frame sequence, which maintains the correspondence with the call identifier and time sequence.

[0092] The speech frame sequence is input into the speech recognition model, which generates a character output sequence and concatenates it into speech-to-text in chronological order. The speech recognition model is an acoustic and language joint modeling structure based on deep neural networks, which is used to map the speech frame sequence into a character probability distribution and decode it into a text sequence.

[0093] The complaint business data is obtained from the business system. The complaint business data includes customer identifier, delivery business type, waybill identifier, node time information and complaint type identifier. The complaint business data is organized with waybill identifier and node time information as primary key fields.

[0094] Based on call identifiers, waybill identifiers, and time alignment rules, the speech-to-text is associated with the appeal business data to generate a call-level appeal record. The call-level appeal record includes a speech-to-text fragment associated with the same call identifier and appeal business data fields.

[0095] Acoustic features, text features, and business features are extracted from call-level appeal records. The acoustic features, text features, and business features are combined with call identifiers in chronological order to construct a call-level multimodal feature vector. The call-level multimodal feature vector is used as the input of the multimodal appeal features to the first probability distribution.

[0096] In this embodiment, the generation of the path template set specifically includes:

[0097] Obtain postal complaint business rule data and service resource configuration data for constructing a knowledge service graph. The postal complaint business rule data includes complaint category code, business clause code, responsible entity code, and handling stage code. The service resource configuration data includes service resource code, service group code, and service capability parameters.

[0098] A set of business knowledge nodes is constructed based on the postal complaint business rules data. In the set of business knowledge nodes, complaint category nodes, business clause nodes, responsible entity nodes and handling link nodes are set, and each business knowledge node is assigned a code and attribute information.

[0099] A service resource node set is constructed based on the service resource configuration data. Service resource nodes are set in the service resource node set, and the service resource code, service group code and service capability parameters are written into the attribute information of the corresponding service resource node.

[0100] Based on the postal complaint business rules data, a set of rule-related edges is generated in the set of business knowledge nodes. Based on the service resource configuration data, a set of service-related edges is generated between the set of business knowledge nodes and the set of service resource nodes. The set of rule-related edges establishes connections between corresponding business knowledge nodes according to the order of complaint category code, business clause code, responsible entity code, and handling stage code in the postal complaint business rules. The set of service-related edges establishes connections between corresponding handling stage nodes and service resource nodes according to the handling stage code and service resource code in the service resource configuration data.

[0101] A knowledge service graph is constructed using a set of business knowledge nodes, a set of service resource nodes, a set of rule-related edges, and a set of service-related edges. Path search is performed in the knowledge service graph starting from the appeal category node and ending at the service resource node. Based on preset node access order constraints and path length constraints, a node sequence is obtained from the appeal category node through the business clause node, the responsible entity node, and the handling link node to the service resource node. Each node sequence is defined as a path template, and the path templates are combined into a path template set.

[0102] In this embodiment, the construction of the first probability distribution, the second probability distribution, and the third probability distribution specifically includes:

[0103] Based on the multimodal appeal features, a unique index is assigned to each multimodal appeal feature. A correspondence is established between the unique index and the call identifier and time interval. An appeal sample index set is constructed. A first probability distribution is set on the appeal sample index set. The first probability distribution assigns an initial probability value to each multimodal appeal feature in units of the unique index.

[0104] A unique index is assigned to each path template based on the path template set. A correspondence is established between the unique index and the appeal category node, business clause node, responsible entity node and handling link node contained in the path template. A path template index set is constructed. A second probability distribution is set on the path template index set. The second probability distribution assigns an initial probability value to each path template in units of unique index.

[0105] A unique index is assigned to each service resource node based on the service resource node. A correspondence is established between the unique index and the service resource code and the service group code. A service resource index set is constructed. A third probability distribution is set on the service resource index set. The third probability distribution assigns an initial probability value to each service resource node in units of the unique index.

[0106] The appeal priority weight is calculated based on the appeal business data associated with the multimodal appeal characteristics. The appeal priority weight is written into the first probability distribution and used as a weighting coefficient to adjust the initial probability value in the first probability distribution. The appeal business data includes the urgency of timeliness, amount, customer level and number of complaints.

[0107] The service capacity parameters are calculated based on the service resource configuration data, written into a third probability distribution, and used as weighting coefficients to adjust the initial probability values ​​in the third probability distribution. The service resource configuration data includes the number of available seats, the number of allowed concurrent calls, and the target service duration. The calculation of the service capacity parameters includes: constructing a capacity scoring function based on the number of available seats, the number of allowed concurrent calls, and the target service duration; generating a capacity score for each service resource node; and normalizing the capacity scores to obtain the service capacity parameters.

[0108] In this embodiment, the construction of the hierarchical cost structure and the generation of graph structure constraints and path feasibility constraints specifically include:

[0109] The first-level cost matrix is ​​constructed based on the set of appeal sample indexes and the set of path template indexes and written into the hierarchical cost structure. The first-level cost matrix is ​​used to describe the cost relationship between multimodal appeal features and path templates. The first-level cost matrix uses the appeal sample index as the row index and the path template index as the column index. Each cost element is calculated based on the semantic differences, business field differences, time location differences, and probability values ​​in the first probability distribution and the second probability distribution between multimodal appeal features and path templates.

[0110] A second-layer cost matrix is ​​constructed based on the path template index set and the service resource index set and written into the hierarchical cost structure. The second-layer cost matrix is ​​used to describe the cost relationship between the path template and the service resource node. The second-layer cost matrix uses the path template index as the row index and the service resource index as the column index. Each cost element is calculated based on the differences in service capabilities, load levels, target service durations, and probability values ​​in the second and third probability distributions between the path template and the service resource node.

[0111] In the hierarchical cost structure, the index range and hierarchical relationship between the first-level cost matrix and the second-level cost matrix are registered, so that the first-level cost matrix and the second-level cost matrix establish a corresponding mapping relationship in the hierarchical cost structure according to the appeal sample index set, the path template index set and the service resource index set;

[0112] Based on the set of business knowledge nodes, the set of service resource nodes, the set of rule association edges, and the set of service association edges in the knowledge service graph, graph structure constraints are generated. The graph structure constraints stipulate that the combination of appeal sample index and path template index with valid values ​​is allowed in the first-level cost matrix, and the combination of path template index and service resource index with valid values ​​is allowed in the second-level cost matrix. The cost elements corresponding to the index combinations that do not satisfy the connection relationship of the knowledge service graph are marked as invalid in a preset way.

[0113] The path feasibility constraints are generated based on the path template set. The path feasibility constraints are based on the order of the appeal category node, business clause node, responsible entity node, handling link node and service resource node in each path template. For the index of the path template node combination in the first-level cost matrix and the second-level cost matrix, the corresponding cost element is reserved as a candidate value. For the index of the path template node combination that does not belong to any path template node combination, the corresponding cost element is marked as a prohibited value according to a preset method.

[0114] In this embodiment, the establishment of the hierarchical trilateral Monge-Kantorovich optimal transmission model specifically includes:

[0115] A first-level transport relationship is established between the first probability distribution and the second probability distribution, and a second-level transport relationship is established between the second probability distribution and the third probability distribution. The first-level transport relationship and the second-level transport relationship are set as the transport variable set of the hierarchical trilateral Monge-Kantorovich optimal transport model. The first-level transport relationship is used to represent the transport volume allocation from the appeal sample index to the path template index, and the second-level transport relationship is used to represent the transport volume allocation from the path template index to the service resource index. The hierarchical trilateral Monge-Kantorovich optimal transport model is a constrained optimization model that establishes two-level transport relationships on three types of probability distributions and takes the hierarchical cost structure as the objective.

[0116] The allocation of the first-layer transport relationship in the appeal sample dimension is limited by the first probability distribution, the balance between the first-layer and second-layer transport relationships in the path template dimension is limited by the second probability distribution, and the allocation of the second-layer transport relationship in the service resource dimension is limited by the third probability distribution, so that the sum of the transport volume associated with each appeal sample index is equal to the corresponding probability value in the first probability distribution, the sum of the first-layer transport volume associated with each path template index and the sum of the second-layer transport volume are equal to the corresponding probability value in the second probability distribution, and the sum of the transport volume associated with each service resource index is equal to the corresponding probability value in the third probability distribution.

[0117] The objective function of the hierarchical trilateral Monge-Kantorovich optimal transport model is constructed based on the hierarchical cost structure. The first-layer cost structure part is weighted and summed with the corresponding elements in the first-layer transport relation, and the second-layer cost structure part is weighted and summed with the corresponding elements in the second-layer transport relation. The two weighted sums are then combined into the total cost.

[0118] The graph structure constraint and path feasibility constraint are incorporated into the hierarchical trilateral Monge-Kantorovich optimal transport model. For the combination of appeal sample index and path template index and the combination of path template index and service resource index that do not satisfy the graph structure constraint and path feasibility constraint, the transport volume is specified to be zero. For the index combination that satisfies the graph structure constraint and path feasibility constraint, the transport volume is specified to be a non-negative real number.

[0119] The numerical solution process is performed under the joint constraints of total cost, first probability distribution, second probability distribution, third probability distribution, graph structure constraint and path feasibility constraint to obtain the first-level transportation plan corresponding to the first-level transportation relationship and the second-level transportation plan corresponding to the second-level transportation relationship. The first-level transportation plan and the second-level transportation plan are then combined to form a hierarchical transportation plan.

[0120] This invention explicitly defines first-layer and second-layer transport relationships among three types of probability distributions and uses them as a unified set of transport variables. It introduces a constrained optimization model with a hierarchical cost structure as the objective, applies quality conservation constraints to the appeal sample dimension, path template dimension, and service resource dimension, and incorporates graph structure constraints and path feasibility constraints into the model. It forces the transport volume to be zero for invalid index combinations and restricts valid index combinations to non-negative real numbers. Numerical solutions are performed under the joint constraints of total cost, three types of probability distributions, and dual constraints, resulting in a two-layer transport plan that simultaneously considers appeal distribution, path structure, and resource capacity. This plan is then combined into a hierarchical transport plan, thereby achieving fine-grained matching and joint optimization of the entire link from appeal sample to path template to service resource node within a unified mathematical framework.

[0121] In this embodiment, the generation of the updated hierarchical transportation plan specifically includes:

[0122] During the operation of the appeal call service, new multimodal appeal features are received and the appeal sample index set and the first probability distribution are expanded. The expansion includes: assigning an appeal sample index to the new multimodal appeal features, adding the appeal sample index to the appeal sample index set, and setting an initial probability value and appeal priority weight in the first probability distribution based on the associated appeal business data.

[0123] The service resource nodes and the third probability distribution are updated based on the current service resource usage status and service resource configuration data. The update includes: writing the latest service capability parameters and service capacity parameters to the service resource nodes, and adjusting the probability value of the corresponding service resource index in the third probability distribution.

[0124] The previous batch of hierarchical transport plans is used as the initial transport relationship of the hierarchical trilateral Monge-Kantorovich optimal transport model. Incremental solutions are performed on the expanded set of complaint sample indexes and the updated set of service resource indexes. The incremental solutions are limited to updating the transport volume only for index combinations that are associated with the new complaint sample indexes and changes in service capacity parameters. The transport volume of the previous batch is retained for the other index combinations.

[0125] After the appeal is processed, feedback statistics are constructed based on the appeal processing results, and the hierarchical cost structure and three types of probability distributions are adjusted. The appeal processing results include processing time, number of transfers, customer satisfaction index and duplicate appeal marker. The feedback statistics are used to correct the cost parameters associated with the path template index and service resource index, as well as the probability values ​​in the first probability distribution, the second probability distribution and the third probability distribution.

[0126] Under the constraints of the incrementally updated first, second, and third probability distributions, the updated hierarchical cost structure, graph structure constraints, and path feasibility constraints, the hierarchical trilateral Monge-Kantorovich optimal transport model is re-solved to obtain the updated first-layer transport plan and the updated second-layer transport plan. The updated first-layer transport plan and the updated second-layer transport plan are then combined to form the updated hierarchical transport plan.

[0127] This invention continuously receives new multimodal appeal features and expands the appeal sample index set and the first probability distribution during operation. Simultaneously, it dynamically updates the service resource nodes and the third probability distribution based on the service resource usage status. Using the previous batch of hierarchical transportation plans as the initial transportation relationship, incremental solutions are performed only on the index combinations associated with the new appeal sample index and changes in service capacity parameters. After the appeal is processed, feedback statistics are constructed based on processing time, number of transfers, customer satisfaction indicators, and duplicate appeal markers to correct the hierarchical cost structure and the three types of probability distributions. Finally, under the combined constraints of the incrementally updated probability distribution, the updated hierarchical cost structure, graph structure constraints, and path feasibility constraints, the hierarchical trilateral Monge-Kantorovich optimal transmission model is re-solved to form an updated hierarchical transportation plan. This achieves real-time reflection and dynamic response of changes in appeal distribution and resource status in the mathematical model, ensuring that the matching results remain stable and consistent and have real-time adaptability during continuous updates.

[0128] In this embodiment, the generation of the appeal processing decision result specifically includes:

[0129] Based on the updated hierarchical transportation plan, for each multimodal appeal feature, a target path template index is selected in the first-level transportation plan, and a target service resource index is selected in the second-level transportation plan. The selection is based on the transportation volume associated with each path template index and each service resource index, as well as a preset threshold rule.

[0130] The corresponding path template is retrieved from the path template set according to the target path template index, and an explanatory path is constructed in combination with the target service resource index. The explanatory path records the appeal category node, business terms node, responsible entity node, handling process node and service resource node in node order.

[0131] Based on the interpretation path, an appeal handling decision result is generated. The appeal handling decision result includes at least the appeal category determination result, the business terms matching result, the responsible entity identification result, the handling process sequence, and the service resource allocation result.

[0132] The explanation path and appeal processing decision results are written into the business records of the appeal call service system, and the appeal processing decision results are output to the call service terminal according to the target service resource index.

[0133] Example 1:

[0134] To verify the feasibility of this invention in practice, it was applied to a postal complaint call center. By integrating with the existing voice platform and business system, a continuous stream of complaint calls was scheduled online. The comparison was with a traditional complaint call system, which relies solely on speech-to-text and simple complaint type fields. It uses fixed call allocation rules and human experience for complaint routing and transfer, without utilizing multimodal complaint features, constructing a knowledge service graph, or employing an optimal transmission model. It only performs coarse-grained allocation based on queue length, skill tags, and human priority fields. High-priority complaints have long queuing times, path selection depends on agent subjective judgment, the identification of responsible parties is unstable, and service resource load varies significantly.

[0135] After deploying this invention, the system first uses a multimodal appeal feature acquisition module to simultaneously access call voice, speech-to-text, and appeal business data. Acoustic features, text features, and business features are obtained through speech recognition and feature extraction, with the call-level multimodal feature vector used as the input for multimodal appeal features. A knowledge service graph and path template construction module constructs business knowledge nodes and service resource nodes based on postal appeal business rule data and service resource configuration data, generating a path template set. A probability distribution modeling module establishes three types of probability distributions on the multimodal appeal features, path template set, and service resource nodes, introducing appeal priority weights and service capacity parameters. A hierarchical cost and constraint construction module combines the knowledge service graph to generate a hierarchical cost structure, graph structure constraints, and path feasibility constraints. A hierarchical three-sided Monge-Kantorovich optimal transmission solution module solves the hierarchical transport plan on the three types of probability distributions and the hierarchical cost structure. An incremental and feedback update module then updates the probability distribution and hierarchical cost structure based on newly added multimodal appeal features and appeal processing results. The decision generation and call service control module determines the target path template and target service resource node for each multimodal appeal feature based on the updated hierarchical transport plan, and outputs the explanation path and appeal processing decision results.

[0136] To quantify the comparison results, tens of thousands of appeal calls were extracted from a continuous period of operational data. Average access waiting time, waiting time for urgent appeals, first-time processing completion rate, repeat appeal rate, consistency of path and business rules, and customer satisfaction scores were statistically analyzed under both the traditional appeal call system and the system of this invention. The results, after cleaning and summarizing, are shown in Table 1.

[0137] Table 1. Comparison of the effects of traditional complaint handling systems and the system of this invention.

[0138] Indicator Items Traditional complaint call system This invention system Average access wait time (seconds) 52 30 Time-sensitive appeal waiting time (seconds) 49 18 First-time processing completion rate (%) 81.5 91.0 Repeat appeal rate (%) 12.3 7.0 Consistency between path and business rules (%) 87.0 96.0 Average customer satisfaction score (out of 10) 8.0 9.1

[0139] The comparison shows that, under the same complaint volume and resource conditions, this invention, through joint modeling of multimodal complaint features, knowledge service graphs, three types of probability distributions and hierarchical trilateral Monge-Kantorovich optimal transmission model, significantly reduces the waiting time for high-priority complaints, increases the first-time processing completion rate, decreases the rate of repeated complaints, improves the consistency between paths and business rules, and raises customer satisfaction scores, thus verifying the application value of this invention in the postal industry complaint call service scenario.

[0140] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A postal service complaint handling system based on multimodal fusion, characterized in that, include: The multimodal appeal feature acquisition module is used to acquire and align call voice, speech-to-text, and appeal business data, extract acoustic features, text features, and business features, and generate multimodal appeal features. The knowledge service graph and path template construction module is used to construct a knowledge service graph containing business knowledge nodes and service resource nodes based on postal complaint business rule data and service resource configuration data, and to search and generate a set of path templates on the knowledge service graph; The probability distribution modeling module is used to construct a first probability distribution, a second probability distribution, and a third probability distribution based on multimodal appeal features, path template sets, and service resource nodes, respectively. Appeal priority weights are set in the first probability distribution, and service capacity parameters are set in the third probability distribution. The construction of the first, second, and third probability distributions specifically includes: Based on the multimodal appeal features, a unique index is assigned to each multimodal appeal feature. A correspondence is established between the unique index and the call identifier and time interval. An appeal sample index set is constructed, and a first probability distribution is set on the appeal sample index set. A unique index is assigned to each path template based on the path template set. A correspondence is established between the unique index and the appeal category node, business clause node, responsible entity node and handling stage node contained in the path template. A path template index set is constructed, and a second probability distribution is set on the path template index set. A unique index is assigned to each service resource node based on the service resource node, and a correspondence is established between the unique index and the service resource code and service group code. A service resource index set is constructed, and a third probability distribution is set on the service resource index set. The appeal priority weight is calculated based on the appeal business data associated with the multimodal appeal features. The appeal priority weight is written into the first probability distribution and used as a weighting coefficient to adjust the initial probability value in the first probability distribution. Calculate the service capacity parameter based on the service resource configuration data, write the service capacity parameter into the third probability distribution, and use the service capacity parameter as a weighting coefficient to adjust the initial probability value in the third probability distribution; The hierarchical cost and constraint construction module is used to construct a hierarchical cost structure based on three types of probability distributions and knowledge service graphs, which describes the mapping relationship from multimodal appeal features to path templates and then to service resource nodes, and to generate graph structure constraints and path feasibility constraints. The hierarchical trilateral Monge-Kantorovich optimal transport solution module is used to establish a hierarchical trilateral Monge-Kantorovich optimal transport model based on three types of probability distributions, hierarchical cost structure, graph structure constraints and path feasibility constraints, and obtain the hierarchical transport plan by solving the model. The incremental and feedback update module is used to update the first probability distribution, the second probability distribution, the third probability distribution, and the hierarchical cost structure based on the newly added multimodal appeal features and appeal processing results during the operation of the appeal call service. It performs incremental and feedback solutions on the hierarchical trilateral Monge-Kantorovich optimal transmission model to generate an updated hierarchical transmission plan. The decision generation and call service control module is used to determine the target path template and target service resource node for each multimodal appeal feature based on the updated hierarchical transport plan, construct the interpretation path, and output the appeal processing decision result.

2. The postal service complaint handling system based on multimodal fusion according to claim 1, characterized in that, The generation of the multimodal appeal features specifically includes: Collect call audio with call identifiers and record the call start time and call end time. Store the audio in segments according to the call identifiers to form a call audio sequence. The call audio sequence is processed by denoising, pre-emphasis, framing, and windowing to obtain an audio frame sequence. Input the speech frame sequence into the speech recognition model, generate a character output sequence, and concatenate them in chronological order to form speech-to-text; The complaint business data is obtained from the business system. The complaint business data includes customer identifier, delivery business type, waybill identifier, node time information and complaint type identifier. Based on call identifiers, waybill identifiers, and time alignment rules, the speech-to-text is associated with the appeal business data to generate a call-level appeal record. The call-level appeal record includes a speech-to-text fragment associated with the same call identifier and appeal business data fields. Acoustic features, text features, and business features are extracted from call-level appeal records. The acoustic features, text features, and business features are combined with call identifiers in chronological order to construct a call-level multimodal feature vector.

3. The postal service complaint handling system based on multimodal fusion according to claim 1, characterized in that, The generation of the path template set specifically includes: Obtain postal complaint business rule data and service resource configuration data for constructing the knowledge service graph; A set of business knowledge nodes is constructed based on the postal complaint business rules data. In the set of business knowledge nodes, complaint category nodes, business clause nodes, responsible entity nodes and handling link nodes are set, and each business knowledge node is assigned a code and attribute information. A service resource node set is constructed based on the service resource configuration data. Service resource nodes are set in the service resource node set, and the service resource code, service group code and service capability parameters are written into the attribute information of the corresponding service resource node. Based on the postal complaint business rules data, a set of rule-related edges is generated in the business knowledge node set, and a set of service-related edges is generated between the business knowledge node set and the service resource node set based on the service resource configuration data; A knowledge service graph is constructed using a set of business knowledge nodes, a set of service resource nodes, a set of rule-related edges, and a set of service-related edges. Path search is performed in the knowledge service graph starting from the appeal category node and ending at the service resource node. Based on preset node access order constraints and path length constraints, a node sequence is obtained from the appeal category node through the business clause node, the responsible entity node, and the handling link node to the service resource node. Each node sequence is defined as a path template, and the path templates are combined into a path template set.

4. The postal service complaint handling system based on multimodal fusion according to claim 1, characterized in that, The construction of the hierarchical cost structure and the generation of graph structure constraints and path feasibility constraints specifically include: The first-level cost matrix is ​​constructed based on the set of appeal sample indexes and the set of path template indexes and written into the hierarchical cost structure. The first-level cost matrix is ​​used to characterize the cost relationship between multimodal appeal features and path templates. A second-level cost matrix is ​​constructed based on the path template index set and the service resource index set and written into the hierarchical cost structure. The second-level cost matrix is ​​used to represent the cost relationship between the path template and the service resource node. In the hierarchical cost structure, the index range and hierarchical relationship of the first-level cost matrix and the second-level cost matrix are set to make the hierarchical cost structure consistent with the first probability distribution, the second probability distribution and the third probability distribution; Based on the set of business knowledge nodes, service resource nodes, rule association edges, and service association edges in the knowledge service graph, graph structure constraints are generated. The graph structure constraints limit the index combinations that can take valid values ​​in the first-level cost matrix and the second-level cost matrix. Path feasibility constraints are generated based on the path template set. The path feasibility constraints limit the index combinations in the first-level cost matrix and the second-level cost matrix that conform to the order of the path template nodes to be feasible combinations.

5. A postal service complaint handling system based on multimodal fusion according to claim 1, characterized in that, The establishment of the hierarchical trilateral Monge-Kantorovich optimal transmission model specifically includes: A first-level transport relationship is established between the first probability distribution and the second probability distribution, and a second-level transport relationship is established between the second probability distribution and the third probability distribution. The first-level transport relationship and the second-level transport relationship are set as the transport variable set of the hierarchical trilateral Monge-Kantorovich optimal transport model. The allocation of the first-layer transport relationship in the appeal sample dimension is limited by the first probability distribution, the balance between the first-layer and second-layer transport relationships in the path template dimension is limited by the second probability distribution, and the allocation of the second-layer transport relationship in the service resource dimension is limited by the third probability distribution, so that the sum of the transport volume associated with each appeal sample index is equal to the corresponding probability value in the first probability distribution, the sum of the first-layer transport volume associated with each path template index and the sum of the second-layer transport volume are equal to the corresponding probability value in the second probability distribution, and the sum of the transport volume associated with each service resource index is equal to the corresponding probability value in the third probability distribution. The objective function of the hierarchical trilateral Monge-Kantorovich optimal transport model is constructed based on the hierarchical cost structure. The first-layer cost structure part is weighted and summed with the corresponding elements in the first-layer transport relation, and the second-layer cost structure part is weighted and summed with the corresponding elements in the second-layer transport relation. The two weighted sums are then combined into the total cost. The graph structure constraint and path feasibility constraint are incorporated into the hierarchical trilateral Monge-Kantorovich optimal transport model. For the combination of appeal sample index and path template index and the combination of path template index and service resource index that do not satisfy the graph structure constraint and path feasibility constraint, the transport volume is specified to be zero. For the index combination that satisfies the graph structure constraint and path feasibility constraint, the transport volume is specified to be a non-negative real number. The numerical solution process is performed under the joint constraints of total cost, first probability distribution, second probability distribution, third probability distribution, graph structure constraint and path feasibility constraint to obtain the first-level transportation plan corresponding to the first-level transportation relationship and the second-level transportation plan corresponding to the second-level transportation relationship. The first-level transportation plan and the second-level transportation plan are then combined to form a hierarchical transportation plan.

6. A postal service complaint handling system based on multimodal fusion according to claim 1, characterized in that, The generation of the updated hierarchical transportation plan specifically includes: During the operation of the appeal call service, new multimodal appeal features are received and the appeal sample index set and the first probability distribution are expanded; Update the service resource nodes and the third probability distribution based on the current service resource usage status and service resource configuration data; The previous batch of hierarchical transport plans is used as the initial transport relationship for the hierarchical trilateral Monge-Kantorovich optimal transport model. Incremental solutions are then performed on the expanded set of appeal sample indexes and the updated set of service resource indexes. After completing the appeal processing, feedback statistics are constructed based on the appeal processing results, and the hierarchical cost structure and three types of probability distributions are adjusted. Under the constraints of the incrementally updated first, second, and third probability distributions, the updated hierarchical cost structure, graph structure constraints, and path feasibility constraints, the hierarchical trilateral Monge-Kantorovich optimal transport model is re-solved to obtain the updated first-layer transport plan and the updated second-layer transport plan. The updated first-layer transport plan and the updated second-layer transport plan are then combined to form the updated hierarchical transport plan.

7. A postal service complaint handling system based on multimodal fusion according to claim 1, characterized in that, The generation of the appeal processing decision result specifically includes: Based on the updated hierarchical transportation plan, for each multimodal appeal feature, a target path template index is selected in the first-level transportation plan, and a target service resource index is selected in the second-level transportation plan; The corresponding path template is retrieved from the path template set based on the target path template index, and the interpretation path is constructed in combination with the target service resource index. Based on the interpretation path, an appeal handling decision result is generated. The appeal handling decision result includes at least the appeal category determination result, the business terms matching result, the responsible entity identification result, the handling process sequence, and the service resource allocation result. The explanation path and appeal processing decision results are written into the business records of the appeal call service system, and the appeal processing decision results are output to the call service terminal according to the target service resource index.

Images