An intelligent dynamic business travel difference bidding management method and system based on a micro-service architecture

Abstract

The present application belongs to the technical field of business travel difference bidding management, and particularly relates to an intelligent dynamic business travel difference bidding management method and system based on a micro-service architecture. The present application uses an improved BERT model to perform semantic analysis and entity extraction on a business travel booking request through information analysis micro-service, and outputs structured itinerary information and user job level. The present application locks a matched benchmark user pool through a difference bidding benchmark calculation micro-service, splits itinerary units and generates an initial difference bidding range through cluster analysis. The present application calculates a user's individual credit coefficient through a credit rating micro-service. The present application calibrates the difference bidding through a dynamic difference bidding calibration micro-service, combines the credit coefficient to calibrate the difference bidding, and generates a compliant difference bidding interval with a compliance red line threshold as a constraint. The present application realizes dynamic fine management and control of business travel difference bidding, solves the pain points of poor adaptability and low efficiency of traditional static management and control mode, and improves the intelligent level of enterprise business travel cost management and control.

Description

Technical Field

[0001] This invention belongs to the field of business travel expense management technology, and particularly relates to an intelligent dynamic business travel expense management method and system based on microservice architecture. Background Technology

[0002] Business travel is a routine part of enterprise operations and development, and expense control is a core component of refined cost management. As enterprises continue to expand their business across regions and scenarios, the complexity of travel itineraries continues to increase, and the segmentation of user job levels and the diversification of travel scenarios are becoming increasingly prominent. The traditional static business travel expense control model, which relies on manual review and fixed threshold settings, can no longer meet the needs of enterprises for efficient and flexible travel management. Problems such as low control efficiency and poor scenario adaptability have become prominent, highlighting the urgent need for a business travel expense control solution based on intelligent technology and dynamic adjustment logic. Summary of the Invention

[0003] To address the technical problems existing in the background art described above, this invention proposes an intelligent dynamic business travel expense management method and system based on a microservice architecture.

[0004] To achieve the above objectives, the technical solution adopted by the present invention includes the following steps:

[0005] S1. Information parsing microservice: Obtain the travel booking request initiated by the user, perform semantic parsing and entity extraction on the travel booking request through a pre-trained improved BERT model, and output the structured itinerary information and user job level of this business trip;

[0006] S2, Benchmark Calculation Microservice: Based on user job level information, a benchmark user pool matching the user making the current request is locked, and a historical compliant order database of the benchmark user pool is obtained; the structured itinerary information is divided into multiple sub-itinerary units, and in the historical compliant order database of the benchmark user pool, a historical compliant order dataset matching at least one of the same sub-itinerary units as the current itinerary is retrieved; cluster analysis is performed on the sub-itineraries to generate the initial benchmark range corresponding to the entire itinerary of this business trip;

[0007] S3, Credit Rating Microservice: Obtain the user's dynamic travel credit data; combine it with historical compliance data from the travel history behavior to calculate and generate a personalized credit coefficient corresponding to this travel trip;

[0008] S4. Dynamic Target Calibration Microservice: Based on the initial target range, the system calculates and calibrates the target range using the personalized credit coefficient. With the compliance red line threshold as the upper limit constraint, it generates a compliant target range corresponding to this business travel booking request, thereby achieving control over business travel target ranges.

[0009] Preferably, step S1, which improves the BERT model, involves improving the homogeneous multi-head attention mechanism of BERT. Two parallel heterogeneous attention branches are designed to address the global long-distance dependencies in semantic parsing and the fine-grained boundary features in entity extraction. Feature fusion is achieved through dynamic gating. Specifically:

[0010] A dual-branch heterogeneous attention parallel architecture is adopted, and the input hidden features obtained after tokenization and embedding of the travel booking request are synchronously input into the mutually independent global semantic structure attention branch and local entity span attention branch.

[0011] The global semantic structure attention branch performs a linear transformation on the input hidden layer features to generate a dedicated query matrix, key matrix, and value matrix. During the self-attention calculation process, the travel scenario semantic structure bias matrix learned in the pre-training stage is injected. After layer normalization and feedforward network processing, the global semantic parsing features corresponding to each layer of the Transformer are output.

[0012] The local entity span attention branch performs a linear transformation on the input hidden layer features to generate query, key, and value matrices that are independent of the global semantic structure attention branch. During the self-attention calculation process, a local window mask matrix and an entity span enhancement matrix are simultaneously superimposed. The local window mask matrix is ​​used to limit the global propagation range of attention. The entity span enhancement matrix is ​​used to strengthen the semantic association of tokens within the same travel entity span. After layer normalization and feedforward network processing, the corresponding entity span perceived features of each layer of the Transformer are output.

[0013] For the global semantic parsing features and entity span-aware features output by each layer of the Transformer, adaptive weighted fusion of the dual-branch features of each token in the sequence is performed through learnable dynamic gating weights. The fused features are then input into the next layer of the Transformer for iterative processing. Specifically, the global semantic parsing features output by the last layer of the Transformer are input into the semantic parsing decoding module, and the entity span-aware features are input into the entity extraction decoding module.

[0014] Preferably, the specific implementation of step S2, which involves locking the pool of comparable users based on user job level information and obtaining the historical compliant order database of the comparable user pool, includes:

[0015] The job level information of the user making this request is uniquely matched with the enterprise's preset full job level system mapping table, and the unique standardized job level corresponding to the user is output.

[0016] By traversing the job level profiles of all users in the enterprise, only the job level of the users in this case is selected to obtain the initial benchmark user pool.

[0017] For all users in the initial benchmarking user pool, retrieve historical compliant orders, retain only users whose standardized job level was still at the time each historical order was generated, and remove users whose job level corresponding to historical orders is inconsistent with their current level, thus obtaining the benchmarking user pool;

[0018] Generate a database of historical compliant orders for benchmark users by analyzing the historical compliant orders of the benchmark user pool.

[0019] Preferably, the implementation of dividing the structured travel information into multiple sub-travel units in step S2 is as follows:

[0020] Extract the complete sequence of location nodes covering the entire trip from the structured trip information to obtain the location tag set;

[0021] An undirected and unweighted location node graph is constructed using the location label set as the vertex set of the graph and the trip connectivity between any two location nodes in the entire trip as the undirected edge set of the graph. The undirected edge set does not distinguish between the sequential time of the trip and the round-trip direction, and only one undirected edge is generated between any two location nodes that have a direct trip connection.

[0022] By treating any two location nodes that have a direct travel relationship as a sub-travel unit, multiple sub-travel units can be obtained.

[0023] Preferably, in step S2, the historical compliant order dataset that matches at least one identical sub-trip unit with the current trip is retrieved from the historical compliant order database of the benchmarking user pool; the implementation of cluster analysis by sub-trip to generate the initial benchmark range corresponding to the entire trip includes:

[0024] Obtain the set of location tags from the compliant order dataset, match them with the sub-trip units, and consider the location tag data of the compliant order dataset with completely identical location tags as a successful match, and obtain the travel expenses with the same location tags;

[0025] A preset clustering algorithm is used to perform cluster analysis on each effective cost sample set to obtain the core cluster of the corresponding sub-trip unit. The center value or median of the core cluster is extracted as the typical cost value of the sub-trip unit.

[0026] The average of the typical cost values ​​corresponding to all corresponding sub-trip units is used to obtain the initial differential benchmark value for all corresponding sub-trip units. The sum of the initial differential benchmark values ​​for all sub-trip units in the entire trip is used to obtain the initial differential benchmark value for the entire trip.

[0027] Based on preset time flexibility rules, the initial tolerance benchmark is set as the lower limit, and the upper limit of the initial tolerance benchmark is generated by floating upwards, ultimately generating the initial tolerance benchmark range corresponding to the entire trip.

[0028] Preferably, the time flexibility rule is to set a floating ratio threshold based on whether the travel date is a statutory holiday, a weekend, or a peak or off-peak season preset by the company, with the initial benchmark value as the lower limit and the final floating ratio threshold as the upper limit to generate the initial benchmark.

[0029] Preferably, in step S3, the user's dynamic travel credit data is obtained; and the personalized credit coefficient corresponding to this trip is calculated by combining the historical compliance data in the travel history data.

[0030] First, calculate the average error: Where n represents the number of historical compliance data points, and n is greater than or equal to 1. The user count decay factor varies over time. These are the initial and actual differential values ​​for the i-th iteration, respectively. The maximum historical standard deviation; user count decay factor. , This is the proportionality coefficient;

[0031] The average frequency error G is inversely linearly mapped to a personalized credit coefficient: .

[0032] Preferably, in step S4, the calibration is performed by multiplying the upper limit of the initial tolerance range by the personalized credit coefficient to obtain the calibrated tolerance range, and the calibrated tolerance range shall not exceed the compliance red line threshold as an upper limit constraint.

[0033] A smart dynamic business travel expense benchmark management system based on microservice architecture includes an information parsing microservice module, an expense benchmark calculation microservice module, a credit rating microservice module, and a dynamic expense benchmark calibration microservice module.

[0034] Information parsing microservice module: Obtains travel booking requests initiated by users, performs semantic parsing and entity extraction on the travel booking requests using a pre-trained improved BERT model, and outputs the structured itinerary information and user job level for this business trip;

[0035] The benchmark calculation microservice module: Based on the user's job level information, it locks the benchmark user pool that matches the user making the current request, and obtains the historical compliant order database of the benchmark user pool; it then splits the structured itinerary information into multiple sub-itinerary units, and retrieves the historical compliant order dataset that matches at least one of the same sub-itinerary units in the historical compliant order database of the benchmark user pool; it then performs cluster analysis on the sub-itineraries to generate the initial benchmark range corresponding to the entire itinerary of this business trip;

[0036] Credit rating microservice module: Obtain the user's dynamic travel credit data; combine the historical compliance data in the travel history behavior to calculate and generate a personalized credit coefficient corresponding to this travel trip;

[0037] Dynamic Target Calibration Microservice Module: Based on the initial target range, the module calculates and calibrates the target range in conjunction with the personalized credit coefficient. With the compliance red line threshold as the upper limit constraint, it generates a compliant target range corresponding to this business travel booking request, thereby achieving control over business travel target ranges.

[0038] Compared with existing technologies, the advantages and positive effects of this invention are as follows: It adopts a microservice architecture to decompose the entire process of travel expense management, achieving decoupling and reuse of four major modules: information parsing, expense benchmark calculation, credit rating, and dynamic expense calibration, flexibly adapting to diverse business travel scenarios; it designs a dual-branch heterogeneous attention-based improved BERT model, simultaneously completing global semantic parsing and fine-grained entity extraction of travel booking requests, significantly improving the parsing accuracy and efficiency of unstructured requests; it constructs a benchmark user pool with precise job level matching, generating an initial expense benchmark range that fits the actual market through sub-trip diagram structure decomposition, cluster analysis, and time elasticity rules, solving the problem of poor adaptability of fixed expense benchmark thresholds; and it introduces a dynamic credit coefficient based on users' historical compliance behavior to perform personalized calibration of expense benchmarks, with a compliance red line threshold as a safety net, achieving refined, dynamic, and intelligent expense benchmark management. Attached Figure Description

[0039] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0040] Figure 1 This is a flowchart illustrating a method and system for intelligent dynamic business travel expense management based on a microservice architecture. Detailed Implementation

[0041] To better understand the above-mentioned objectives, features, and advantages of the present invention, the present invention will be further described below in conjunction with the accompanying drawings and embodiments. It should be noted that, unless otherwise specified, the embodiments and features described in these embodiments can be combined with each other.

[0042] Numerous specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways than those described herein, and therefore the invention is not limited to the specific embodiments disclosed in the following specification.

[0043] Examples, such as Figure 1 As shown, this application provides a method and system for intelligent dynamic business travel expense management based on a microservice architecture. The method includes:

[0044] Information parsing microservice: Obtains travel booking requests initiated by users, performs semantic parsing and entity extraction on the travel booking requests using a pre-trained improved BERT model, and outputs structured itinerary information and user job level for this business trip.

[0045] Furthermore, to improve the limitations of BERT's homogeneous multi-head attention, two parallel heterogeneous attention branches are designed to address the global long-distance dependency in semantic parsing and the fine-grained boundary features in entity extraction. Feature fusion is achieved through dynamic gating. Specifically, a dual-branch heterogeneous attention parallel architecture is adopted, in which the input hidden features obtained after tokenization and embedding of the travel booking request are synchronously input into the mutually independent global semantic structure attention branch and local entity span attention branch.

[0046] The global semantic structure attention branch performs a linear transformation on the input hidden layer features to generate a dedicated query matrix, key matrix, and value matrix. During the self-attention computation, it injects the travel scenario semantic structure bias matrix learned in the pre-training stage, strengthening the strong semantic association between core booking predicates and corresponding arguments, and between modifiers and headwords in the travel booking request. After layer normalization and feedforward network processing, it outputs the global semantic parsing features corresponding to each layer of the Transformer. Specifically, the input hidden layer representation of the travel booking request after tokenization and embedding layer processing is denoted as... ,right Perform a dedicated linear transformation to generate a query matrix with semantic branches. Key matrix AND-value matrix ,in, , , , These are the learnable parameters for the semantic branches. During the self-attention calculation process, a semantic structure bias matrix for the travel scenario is injected to strengthen the strong semantic association between the core booking predicate and its corresponding argument, and between modifiers and the head word in the travel booking request. The semantic branch attention output is then calculated using the following formula: ,in, The semantic structure bias matrix for business travel scenarios. The semantic adjacency matrix of travel texts, learned during the pre-training phase, is used to represent the semantic association strength between tokens. The semantic prior influence strength coefficient is used; after the semantic branch attention output is processed by layer normalization and feedforward network, the global semantic parsing features corresponding to each layer of the Transformer are output. .

[0047] The local entity span attention branch performs a linear transformation on the input hidden layer features to generate query, key, and value matrices independent of the global semantic structure attention branch. During self-attention computation, a local window mask matrix and an entity span enhancement matrix are simultaneously superimposed. The local window mask matrix limits the global propagation range of attention and filters out global noise irrelevant to the travel entity. The entity span enhancement matrix strengthens the semantic association of tokens within the same travel entity span. After layer normalization and feedforward processing, the output is the entity span-aware feature corresponding to each layer of the Transformer. Specifically, the input hidden layer representation... Perform a dedicated linear transformation independent of the global semantic structure attention branch to generate the query matrix for the entity branch. Key matrix AND-value matrix During the self-attention calculation process, a local window mask matrix and a travel entity span enhancement matrix are simultaneously superimposed to filter out global noise unrelated to the travel entity and strengthen the semantic association of tokens within the same travel entity span. The entity branch attention output is then calculated, as shown in the following formula: ,in, For a local window mask matrix, when , Otherwise, it would be negative infinity. The preset window size is used to limit the global propagation range of attention. This is an enhancement matrix for the entity span of the travel entity. The entity branch attention output is processed through layer normalization and a feedforward network to output the entity span-aware features corresponding to each layer of the Transformer. .

[0048] For the global semantic parsing features and entity span-aware features output from each layer of the Transformer, adaptive weighted fusion of the dual-branch features of each token in the sequence is performed using learnable dynamic gating weights. The fused features are then input into the next layer of the Transformer for iterative processing. Specifically, the global semantic parsing features output from the last layer of the Transformer are input into the semantic parsing decoding module, and the entity span-aware features are input into the entity extraction decoding module. In particular, for each token in the travel text sequence, learnable dynamic gating weights are calculated based on the global semantic parsing features and entity span-aware features at the corresponding position. ,in, Let be the fusion gating weight for the i-th token, with a value ranging from 0 to 1. These are their corresponding learnable weights and biases, The sigmoid activation function is used; the dual-branch features are adaptively weighted and fused based on dynamic gating weights, and the fused features are input to the next layer of the Transformer for iterative processing. ,in, This represents the final feature after fusing the i-th token. It is the global semantic parsing feature output from the last layer of the Transformer. Input is fed into the semantic parsing and decoding module for parsing travel booking intent and standard constraint rules; entity span perception features. The input is sent to the entity extraction and decoding module, which outputs the structured itinerary information and user job level for this business trip.

[0049] The benchmark calculation microservice: Based on the user's job level information, it locks the benchmark user pool that matches the user making the current request, and obtains the historical compliant order database of the benchmark user pool; it then splits the structured itinerary information into multiple sub-itinerary units, and retrieves the historical compliant order dataset that matches at least one of the same sub-itinerary units in the historical compliant order database of the benchmark user pool; it then performs cluster analysis on the sub-itineraries to generate the initial benchmark range corresponding to the entire itinerary of this business trip.

[0050] Furthermore, based on the user's job level information, a pool of benchmark users matching the user making the current request is identified. The statistical analysis of the historical compliant order database for this benchmark user pool specifically includes: uniquely matching the job level information of the user making the current request with the company's pre-defined full-scale job level system mapping table, outputting the unique standardized job level level corresponding to the user; traversing the current job level files of all users in the company, selecting only current users whose standardized job level level is completely consistent with the current user's level, thus obtaining the initial benchmark user pool; for all users in the initial benchmark user pool, retrieving historical compliant orders, retaining only users whose standardized job level was still at the time each historical order was generated, and removing users whose historical order job level is inconsistent with their current level, thus obtaining the benchmark user pool; and statistically analyzing the historical compliant orders of the benchmark users in the benchmark user pool to generate the historical compliant order database for the benchmark user pool.

[0051] Furthermore, the implementation of splitting the structured itinerary information into multiple sub-itin units is as follows: extract the full sequence of location nodes covering the entire itinerary from the structured itinerary information to obtain a location tag set; use the location tag set as the vertex set of the graph, and use the itinerary connectivity between any two location nodes in the entire itinerary as the undirected edge set of the graph to construct an undirected and unweighted location node graph; wherein, the undirected edge set does not distinguish between the sequential time sequence and the round-trip direction of the itinerary, and only one undirected edge is generated between any two location nodes with a direct itinerary connection; regard any two location nodes with a direct itinerary connection as a sub-itinerary unit, thereby obtaining multiple sub-itinerary units.

[0052] Furthermore, in the historical compliant order database of the benchmark user pool, retrieve the historical compliant order dataset that matches at least one sub-trip unit with the current trip; the implementation of cluster analysis by sub-trip to generate the initial differential benchmark range corresponding to the entire trip includes: obtaining the location label set of the compliant order dataset, matching it with the sub-trip unit, considering the location label data of the compliant order dataset with completely consistent location labels as a successful match, and obtaining the travel expenses with the same location labels; using a preset clustering algorithm to perform cluster analysis on each effective expense sample set to obtain the core cluster of the corresponding sub-trip unit, and extracting the center value or median of the core cluster as the typical expense value of the sub-trip unit; averaging the typical expense values ​​corresponding to all corresponding sub-trip units to obtain the initial differential benchmark benchmark value of all corresponding sub-trip units for this trip, and summing the initial differential benchmark benchmark values ​​of all sub-trip units for the entire trip to obtain the initial differential benchmark benchmark value for the entire trip; based on preset time flexibility rules, using the initial differential benchmark benchmark value as the lower limit, floating upwards to generate the initial differential benchmark upper limit, and finally generating the initial differential benchmark range corresponding to the entire trip.

[0053] The time flexibility rule is to set a floating ratio threshold based on whether the travel date is a statutory holiday, a weekend, or a peak or off-peak season preset by the company. The initial benchmark value is used as the lower limit, and the initial benchmark upper limit is generated by floating upward according to the final floating ratio threshold.

[0054] Next, the credit rating microservice will be implemented: This involves obtaining the user's dynamic travel credit data; combining this data with historical compliance data from the travel history records, and calculating a personalized credit coefficient for the current trip. Specifically, the personalized credit coefficient is calculated as follows: First, the average frequency error is calculated: Where n represents the number of historical compliance data points, and n is greater than or equal to 1. The user count decay factor varies over time. These are the initial and actual differential values ​​for the i-th iteration, respectively. The maximum historical standard deviation; user count decay factor. , The proportionality coefficient is used; the average frequency error G is inversely linearly mapped to a personalized credit coefficient: .

[0055] Finally, a dynamic tolerance calibration microservice is implemented: based on the initial tolerance range, calibration is calculated using the personalized credit coefficient, with the compliance red line threshold as the upper limit constraint, generating a compliant tolerance range corresponding to this business travel booking request, thereby achieving control over business travel tolerance. Furthermore, the implementation of calculating calibration based on the initial tolerance range and the personalized credit coefficient involves multiplying the obtained upper limit of the initial tolerance range by the personalized credit coefficient to obtain the calibrated tolerance range, and the calibrated tolerance range must not exceed the compliance red line threshold as the upper limit constraint.

[0056] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments for application in other fields. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A method for intelligent dynamic business travel expense management based on a microservice architecture, characterized in that, Includes the following steps: S1. Information parsing microservice: Obtain the travel booking request initiated by the user, perform semantic parsing and entity extraction on the travel booking request through a pre-trained improved BERT model, and output the structured itinerary information and user job level of this business trip; S2, Benchmark Calculation Microservice: Based on the user's job level information, lock the benchmark user pool that matches the user making the current request, and statistically obtain the historical compliant order database of the benchmark user pool; The structured itinerary information is then divided into multiple sub-itinerary units. In the historical compliant order database of the benchmark user pool, a historical compliant order dataset that matches at least one of the same sub-itinerary units as the current itinerary is retrieved. Cluster analysis is then performed on the sub-itineraries to generate the initial benchmark range corresponding to the entire business trip. S3, Credit Rating Microservice: Obtain the user's dynamic travel credit data; Based on the historical compliance data in the aforementioned business travel history, a personalized credit score is calculated and generated for the current business trip. S4. Dynamic Target Calibration Microservice: Based on the initial target range, the system calculates and calibrates the target range using the personalized credit coefficient. With the compliance red line threshold as the upper limit constraint, it generates a compliant target range corresponding to this business travel booking request, thereby achieving control over business travel target ranges. Step S1, which improves the BERT model, involves enhancing the homogeneous multi-head attention mechanism of BERT. To address the global long-distance dependencies in semantic parsing and the fine-grained boundary features in entity extraction, two parallel heterogeneous attention branches are designed. Feature fusion is achieved through dynamic gating. Specifically: A dual-branch heterogeneous attention parallel architecture is adopted, and the input hidden features obtained after tokenization and embedding of the travel booking request are synchronously input into the mutually independent global semantic structure attention branch and local entity span attention branch. The global semantic structure attention branch performs a linear transformation on the input hidden layer features to generate a dedicated query matrix, key matrix, and value matrix. During the self-attention calculation process, the travel scenario semantic structure bias matrix learned in the pre-training stage is injected. After layer normalization and feedforward network processing, the global semantic parsing features corresponding to each layer of the Transformer are output. The local entity span attention branch performs a linear transformation on the input hidden layer features to generate query, key, and value matrices that are independent of the global semantic structure attention branch. During the self-attention calculation process, a local window mask matrix and an entity span enhancement matrix are simultaneously superimposed. The local window mask matrix is ​​used to limit the global propagation range of attention. The entity span enhancement matrix is ​​used to strengthen the semantic association of tokens within the same travel entity span. After layer normalization and feedforward network processing, the corresponding entity span perceived features of each layer of the Transformer are output. For the global semantic parsing features and entity span-aware features output by each layer of the Transformer, adaptive weighted fusion of the dual-branch features of each token in the sequence is performed through learnable dynamic gating weights. The fused features are then input into the next layer of the Transformer for iterative processing. Specifically, the global semantic parsing features output by the last layer of the Transformer are input into the semantic parsing decoding module, and the entity span-aware features are input into the entity extraction decoding module.

2. The intelligent dynamic business travel expense management method based on microservice architecture according to claim 1, characterized in that, The specific implementation of step S2, which involves locking the pool of comparable users based on user job level information and obtaining the historical compliant order database of the comparable user pool, includes: The job level information of the user making this request is uniquely matched with the enterprise's preset full job level system mapping table, and the unique standardized job level corresponding to the user is output. By traversing the job level profiles of all users in the enterprise, only the job level of the users in this case is selected to obtain the initial benchmark user pool. For all users in the initial benchmarking user pool, retrieve historical compliant orders, retain only users whose standardized job level was still at the time each historical order was generated, and remove users whose job level corresponding to historical orders is inconsistent with their current level, thus obtaining the benchmarking user pool; Generate a database of historical compliant orders for benchmark users by analyzing the historical compliant orders of the benchmark user pool.

3. The intelligent dynamic business travel expense management method based on microservice architecture according to claim 1, characterized in that, The implementation of splitting the structured trip information into multiple sub-trip units in step S2 is as follows: Extract the complete sequence of location nodes covering the entire trip from the structured trip information to obtain the location tag set; An undirected and unweighted location node graph is constructed using the location label set as the vertex set of the graph and the trip connectivity between any two location nodes in the entire trip as the undirected edge set of the graph. The undirected edge set does not distinguish between the sequential time of the trip and the round-trip direction, and only one undirected edge is generated between any two location nodes that have a direct trip connection. By treating any two location nodes that have a direct travel relationship as a sub-travel unit, multiple sub-travel units can be obtained.

4. The intelligent dynamic business travel expense management method based on microservice architecture according to claim 1, characterized in that, In step S2, the historical compliant order dataset that matches at least one identical sub-trip unit with the current trip is retrieved from the historical compliant order database of the benchmark user pool. The implementation of cluster analysis based on sub-trips to generate the initial tolerance range corresponding to the entire business trip includes: Obtain the set of location tags from the compliant order dataset, match them with the sub-trip units, and consider the location tag data of the compliant order dataset with completely identical location tags as a successful match, and obtain the travel expenses with the same location tags; A preset clustering algorithm is used to perform cluster analysis on each effective cost sample set to obtain the core cluster of the corresponding sub-trip unit. The center value or median of the core cluster is extracted as the typical cost value of the sub-trip unit. The average of the typical cost values ​​corresponding to all corresponding sub-trip units is used to obtain the initial differential benchmark value for all corresponding sub-trip units. The sum of the initial differential benchmark values ​​for all sub-trip units in the entire trip is used to obtain the initial differential benchmark value for the entire trip. Based on preset time flexibility rules, the initial tolerance benchmark is set as the lower limit, and the upper limit of the initial tolerance benchmark is generated by floating upwards, ultimately generating the initial tolerance benchmark range corresponding to the entire trip.

5. The intelligent dynamic business travel expense management method based on microservice architecture according to claim 4, characterized in that, The time flexibility rule is to set a floating ratio threshold based on whether the travel date is a statutory holiday, a weekend, or a peak or off-peak season preset by the company. The initial benchmark value is used as the lower limit, and the initial benchmark upper limit is generated by floating upward according to the final floating ratio threshold.

6. The intelligent dynamic business travel expense management method based on microservice architecture according to claim 1, characterized in that, The method for obtaining the user's dynamic travel credit data in step S3, and calculating and generating a personalized credit coefficient corresponding to the current business trip by combining the historical compliance data in the travel history behavior, is as follows: First, calculate the average error: Where n represents the number of historical compliance data points, and n is greater than or equal to 1. The user count decay factor varies over time. These are the initial and actual differential values ​​for the i-th iteration, respectively. The maximum historical standard deviation; user count decay factor. , This is the proportionality coefficient; The average frequency error G is inversely linearly mapped to a personalized credit coefficient: .

7. The intelligent dynamic business travel expense management method based on microservice architecture according to claim 1, characterized in that, In step S4, the calibration calculation based on the initial tolerance range and the personalized credit coefficient is performed by multiplying the upper limit of the initial tolerance range by the personalized credit coefficient to obtain the calibrated tolerance range. The calibrated tolerance range shall not exceed the compliance red line threshold as an upper limit constraint.

8. A smart dynamic business travel expense benchmark management system based on microservice architecture, comprising an information parsing microservice module, an expense benchmark calculation microservice module, a credit rating microservice module, and a dynamic expense benchmark calibration microservice module; Information parsing microservice module: Obtains travel booking requests initiated by users, performs semantic parsing and entity extraction on the travel booking requests using a pre-trained improved BERT model, and outputs the structured itinerary information and user job level for this business trip; The improved BERT model is an enhanced BERT homogeneous multi-head attention model. It designs two parallel heterogeneous attention branches to address global long-distance dependencies in semantic parsing and fine-grained boundary features in entity extraction. Feature fusion is achieved through dynamic gating. Specifically: A dual-branch heterogeneous attention parallel architecture is adopted, and the input hidden features obtained after tokenization and embedding of the travel booking request are synchronously input into the mutually independent global semantic structure attention branch and local entity span attention branch. The global semantic structure attention branch performs a linear transformation on the input hidden layer features to generate a dedicated query matrix, key matrix, and value matrix. During the self-attention calculation process, the travel scenario semantic structure bias matrix learned in the pre-training stage is injected. After layer normalization and feedforward network processing, the global semantic parsing features corresponding to each layer of the Transformer are output. The local entity span attention branch performs a linear transformation on the input hidden layer features to generate query, key, and value matrices that are independent of the global semantic structure attention branch. During the self-attention calculation process, a local window mask matrix and an entity span enhancement matrix are simultaneously superimposed. The local window mask matrix is ​​used to limit the global propagation range of attention. The entity span enhancement matrix is ​​used to strengthen the semantic association of tokens within the same travel entity span. After layer normalization and feedforward network processing, the corresponding entity span perceived features of each layer of the Transformer are output. For the global semantic parsing features and entity span-aware features output by each layer of the Transformer, adaptive weighted fusion of the dual-branch features of each token in the sequence is performed through learnable dynamic gating weights. The fused features are then input into the next layer of the Transformer for iterative processing. Specifically, the global semantic parsing features output by the last layer of the Transformer are input into the semantic parsing decoding module, and the entity span-aware features are input into the entity extraction decoding module. The benchmark calculation microservice module: Based on the user's job level information, it locks the benchmark user pool that matches the user making the current request, and obtains the historical compliant order database of the benchmark user pool; it then splits the structured itinerary information into multiple sub-itinerary units, and retrieves the historical compliant order dataset that matches at least one of the same sub-itinerary units in the historical compliant order database of the benchmark user pool; it then performs cluster analysis on the sub-itineraries to generate the initial benchmark range corresponding to the entire itinerary of this business trip; Credit rating microservice module: Obtain the user's dynamic travel credit data; combine the historical compliance data in the travel history behavior to calculate and generate a personalized credit coefficient corresponding to this travel trip; Dynamic Target Calibration Microservice Module: Based on the initial target range, the module calculates and calibrates the target range in conjunction with the personalized credit coefficient. With the compliance red line threshold as the upper limit constraint, it generates a compliant target range corresponding to this business travel booking request, thereby achieving control over business travel target ranges.

Images