IEPL product intelligent quotation decision-making system based on multi-scene classification

Abstract

The invention discloses an IEPL product intelligent quotation decision-making system based on multi-scene classification, and relates to the technical field of telecommunication service management, the system comprises an intelligent quotation decision-making center, the intelligent quotation decision-making center is in communication connection with the following modules: an intention analysis module, a semantic analysis and scene mapping capability integration module, and an intelligent quotation decision-making module; non-structured questions of clients are accurately converted into structured business intentions, and corresponding business scenes are automatically associated; and the intelligent routing pricing module is used for integrating a routing rule base and a dynamic cost engine, automatically planning an optimal path according to the associated service scene and calculating segmentation cost. According to the invention, through identifying the unstructured customer demand and mapping the demand to the specific business scene, automatic conversion from fuzzy inquiry to an accurate scene is realized, and the optimal path is automatically planned and the segmentation cost is calculated in combination with the business scene identifier, so that the problem of omission or repeated pricing caused by manual judgment is eliminated, and the efficiency of the system is improved. The quotation generation process is rapid and accurate, and the overall quotation efficiency is greatly improved.

Description

Technical Field

[0001] This invention relates to the field of telecommunications service management technology, specifically to an intelligent pricing decision system for IEPL products based on multi-scenario classification. Background Technology

[0002] With the acceleration of global digitalization, International Ethernet Leased Lines (IEPLs) are becoming increasingly complex and diverse in their pricing scenarios as a critical infrastructure for cross-border enterprise communications. IEPL products involve various circuit types, including domestic local, domestic long-distance, cross-border, cross-border, and purely overseas circuits. The pricing structure needs to be calculated in segments based on factors such as endpoint location, routing path, and whether a point of interest (POP) is connected. In actual business, there are diverse inquiry needs, and pricing varies depending on factors such as bandwidth, landing city, and discount strategies. At the same time, the approval process and authority also differ in different scenarios, especially when it comes to fast line products, high bandwidth, or special discounts, which require a multi-level approval mechanism.

[0003] Currently, the following technical shortcomings still exist in the IEPL product pricing process: First, account managers have limited ability to recognize the intent behind questions, making it difficult to accurately understand the real business scenarios behind unstructured questions, leading to inaccurate responses or requiring manual intervention; second, the existing pricing mechanism relies on manual judgment for calculating segmented fees, especially in complex scenarios such as cross-border, transit, and purely overseas transactions, lacking a unified and automated routing and fee combination logic, which easily leads to omissions or duplications in pricing; furthermore, the judgment in the discount approval process is relatively rigid, making it difficult to dynamically match the approval path based on factors such as the customer's expected discount, product type, and city attributes, resulting in delays in the pricing process or conflicts of permissions. Overall, the existing system has certain deficiencies in scenario recognition, segmented pricing, approval linkage, and intent understanding, which restricts the improvement of pricing efficiency and customer experience. To address these issues, we propose an intelligent pricing decision system for IEPL products based on multi-scenario classification. Summary of the Invention

[0004] To solve the above technical problems, the present invention is implemented through the following technical solution: an intelligent pricing decision system for IEPL products based on multi-scenario classification, comprising an intelligent pricing decision center, wherein the intelligent pricing decision center is communicatively connected to the following modules:

[0005] The intent parsing module integrates semantic parsing and scenario mapping capabilities to accurately transform unstructured customer questions into structured business intents and automatically associate them with corresponding business scenarios. This lays the foundation for accurate pricing and achieves a seamless connection between intent and scenario.

[0006] The intelligent routing and pricing module integrates a routing rule base and a dynamic cost engine. It automatically plans the optimal path and calculates segmented costs based on the associated business scenarios, and generates cost combinations for business scenarios in an automated and standardized manner, eliminating omissions or duplicate pricing.

[0007] The strategy matching engine, based on multi-dimensional information such as authorized customer history, product attributes, and city level, intelligently matches and recommends discount ranges from the preset discount strategy matching model, providing data-driven basis for approval and improving the rationality of discount schemes and customer satisfaction.

[0008] The process decision module is used to automatically determine and assign appropriate approval roles and paths based on the discount amount and product type rules covered in the matched recommended discount range, dynamically allocate permissions, avoid process conflicts and delays, and optimize approval efficiency.

[0009] The collaborative decision-making module aggregates routing, fee, and discount approval results, generates structured and visualized customized quotations with one click, and collects interaction feedback data between account managers and customers. Through machine learning, it continuously optimizes intent recognition, scenario classification, and discount strategies, forming a closed-loop learning process to promote the continuous evolution of the system's intelligence level.

[0010] Preferably, the intent parsing module includes a semantic perception unit and a scene mapping unit;

[0011] The semantic perception unit is used to use NLP technology to parse customer questions, identify key business entities and intent features in customer questions, transform fuzzy needs into clear business query elements, output a set of query elements, improve the accuracy and automation level of initial intent recognition, accurately parse unstructured queries, and significantly improve the accuracy and automation level of initial intent recognition.

[0012] The scenario mapping unit, based on the parsed key business entities and intent features, matches specific business scenarios covering cross-border and transboundary activities through a preset multi-label scenario classification model, forming standardized business scenario identifiers, eliminating ambiguity in manual judgment, ensuring consistency of subsequent process scenarios, and providing a unique and standard business scenario context for subsequent pricing and routing.

[0013] Preferably, the semantic awareness unit performs the following steps:

[0014] It receives unstructured text queries input by account managers or customer terminals, and uses a pre-trained natural language understanding model to perform word segmentation, part-of-speech tagging, and dependency parsing on the query statements to deconstruct their grammatical structure, effectively parsing users' colloquial or non-standard expressions, and significantly improving the depth and accuracy of semantic understanding.

[0015] Based on the deconstructed grammatical structure, the named entity recognition model is invoked to accurately extract key business entities from the query text. Key business entities include at least one or more of the following: geographic endpoint city, bandwidth specification, product level, and service quality requirements. Key business parameters are automatically extracted, reducing manual extraction errors and providing reliable input for subsequent accurate quotations.

[0016] By combining the extracted key business entities, the deep business purpose of the query is identified and categorized through an intent classification model, and a structured set of query elements is output. The set of query elements clearly includes the business action type and core parameter constraints, accurately identifies the user's true intent, avoids misunderstanding bias, and realizes the automatic conversion of the query into structured business elements.

[0017] Preferably, the scene mapping unit performs the following steps:

[0018] It receives a set of structured query elements, and determines the relative positional relationship and cross-border attributes between geographic endpoint cities based on a preset geographic location and network resource topology database. It accurately defines the physical attributes of the business and avoids compliance and routing errors caused by geographic misjudgment.

[0019] Based on the relative positional relationship and cross-border attribute judgment results, the query element set is input into a pre-trained multi-label scenario classification model. The multi-label scenario classification model calculates the probability that it belongs to one or more preset standard business scenarios based on the business rule feature library. It integrates data prediction and expert rules to improve the accuracy and business fit of complex business scenario judgment.

[0020] By comparing the preset probability threshold with the calculated probability, a standardized business scenario identifier is generated and output. The business scenario identifier uniquely corresponds to a specific combination of routing and billing rules, providing an unambiguous execution context for subsequent processes. The classification result is solidified into deterministic instructions, ensuring that all subsequent automatic processing links have a unified context and a clear path.

[0021] Preferably, the intelligent routing and pricing module includes a rule-based routing unit and a cost aggregation unit;

[0022] The rule routing unit, based on a built-in multi-scenario routing rule library for cross-border and purely overseas scenarios, automatically matches compliant paths and necessary nodes according to the attributes of the business scenario, and outputs the selected end-to-end routing scheme, replacing manual path judgment, improving the accuracy and efficiency of route planning, automatically generating the optimal compliant path scheme, and greatly improving the efficiency and accuracy of route planning.

[0023] The cost aggregation unit automatically calculates and intelligently combines the costs of each segment based on the routing results by calling the built-in segmented pricing model. It supports parameterized adjustments, quickly generates and visualizes cost combinations in business scenarios, reduces human calculation errors, automates segmented pricing and cost aggregation, and effectively eliminates omissions and errors in human calculation.

[0024] Preferably, the rule routing unit performs the following steps:

[0025] The system receives the standardized business scenario identifier and, based on the business scenario identifier, retrieves a matching set of path planning constraints from a preset multi-scenario routing rule library covering cross-border and purely overseas scenarios. The set of path planning constraints includes mandatory traversal nodes, avoidance areas, and transmission protocol requirements. Through scenario-driven precise retrieval, the system achieves automatic matching of rule conditions with business requirements, avoiding errors in manual configuration.

[0026] Based on the set of path planning constraints and combined with real-time network resource status data, graph theory algorithms are used to calculate one or more optimal feasible paths that meet business requirements in the pre-constructed network topology graph. Real-time dynamic optimization ensures path feasibility and improves the efficiency and resource utilization of route planning.

[0027] The calculated feasible paths are pre-evaluated for compliance and cost. Finally, a standardized end-to-end routing scheme is selected and output. This end-to-end routing scheme clearly lists all necessary network nodes and segment divisions. Double verification ensures that the path is compliant and economical, and outputs a standard routing scheme with a clear structure that can be delivered and executed.

[0028] Preferably, the cost aggregation unit performs the following steps:

[0029] Based on the output standardized end-to-end routing scheme, the various transmission segments it contains are parsed, and the corresponding billing functions are called from the built-in segmented pricing model library according to the segment attributes. The billing function parameters include distance, bandwidth, and resource type. By matching the segmented pricing model, accurate and automated benchmark cost calculation for different network resource types is achieved, and the misuse of the billing model is prevented.

[0030] The bandwidth specifications and product level parameters extracted from the key business entities are input into the billing functions of each segment to automatically calculate the base cost of each routing segment and perform the conversion between local currency and settlement currency. Based on the automatic conversion and accumulation of real-time exchange rates, the financial accuracy of cross-border quotations and the integrity of cost composition are ensured.

[0031] Based on the preset cost combination logic, the calculated base costs for all segments are summed up, and the global cost adjustment rules corresponding to the business scenario identifier are applied to generate a structured detailed cost list as the basis for quoting. The cost is dynamically adjusted using scenario-based strategies to generate a final quotation list that is clear in structure and transparent in auditing, thereby enhancing customer trust.

[0032] Preferably, the strategy matching engine performs the following steps:

[0033] By acquiring customer identifiers from the structured query element set, querying the customer's historical transaction database, and combining business scenario identifiers and product levels, a complete feature vector for the current quote request is formed. This achieves dynamic integration of customer profiles and real-time quote scenarios, providing multi-dimensional and high-precision input features for intelligent discount matching.

[0034] The feature vector is input into a preset discount strategy matching model. This model performs multi-dimensional similarity calculation and weight evaluation between the feature vector and the rules in the strategy knowledge base, and outputs one or more applicable candidate discount strategies and their confidence levels. By combining data-driven and rule-based reasoning, it outputs candidate discount strategies with high confidence, thereby improving the rationality and reliability of discount recommendations.

[0035] Based on preset decision-making rules, candidate discount strategies are screened and integrated to generate the final recommended discount scheme and the specific policy terms matched with the scheme. This provides data support for approval decisions, automatically generates structured discount suggestions and complete supporting evidence, and significantly improves approval efficiency and decision-making transparency.

[0036] Preferably, the process decision module performs the following steps:

[0037] Receive recommended discount schemes and detailed expense lists, analyze the discount amount and product type, and determine the approval matrix version triggered by the process based on the business scenario identifier to ensure that each quotation approval strictly follows the latest internal control rules and eliminates the risk of rules being outdated or misused.

[0038] Based on the parsed parameters and the determined approval matrix, the workflow rule engine is dynamically matched to automatically determine the approval role sequence, approval path branches and approval thresholds of each node required for this quotation request, so as to realize the personalized and accurate generation of the approval path and avoid human judgment bias and process conflict.

[0039] Based on the judgment results, the corresponding electronic approval process instance is automatically generated, and the task is pushed to the to-do list of the corresponding approval role. At the same time, the quotation is locked until the process is completed, realizing the automated allocation and execution of the approval path, the fully automated flow and strong status control of the entire process, which greatly improves the approval efficiency and prevents tampering in the middle.

[0040] Preferably, the collaborative decision-making module performs the following steps:

[0041] It aggregates detailed cost lists from the intelligent routing and pricing module, recommended discount schemes from the strategy matching engine, and approval status from the process decision module. It automatically generates a structured quotation document containing routing diagrams, cost breakdowns, and terms and conditions according to a preset template. It generates standardized and trustworthy quotations with one click, significantly improving the professionalism and efficiency of business communication.

[0042] After the quotation is sent to the customer, the customer feedback results and final transaction information entered by the account manager are collected through the integration interface. Together with various intermediate decision data in the quotation process, they are structured and stored in the feedback learning database to comprehensively accumulate business decision data assets and provide a reliable basis for intelligent model optimization.

[0043] By periodically utilizing the data in the feedback learning database, the parameters of the natural language understanding model in the semantic perception unit, the multi-label scene classification model in the scene mapping unit, and the discount strategy matching model in the strategy matching engine are tuned and iterated through offline training. This forms a continuously optimized closed-loop learning mechanism, enabling the dynamic evolution of the system's intelligence level and continuously improving the accuracy of pricing and market adaptability.

[0044] This invention provides an intelligent pricing decision-making system for IEPL products based on multi-scenario classification. It has the following beneficial effects:

[0045] (i) This intelligent pricing decision system for IEPL products based on multi-scenario classification identifies unstructured customer needs and maps those needs to specific business scenarios, thereby achieving an automated conversion from fuzzy inquiries to precise scenarios. It also automatically plans the optimal path and calculates segmented costs by combining business scenario identifiers, eliminating omissions or duplicate pricing caused by manual judgment. This makes the pricing generation process fast and accurate, and significantly improves the overall pricing efficiency.

[0046] (ii) This intelligent pricing decision system for IEPL products based on multi-scenario classification is designed for complex business scenarios such as cross-border and cross-border transactions. Through the built-in multi-scenario routing rule library and segmented pricing model, it can automatically match compliant paths and billing templates according to business attributes, realize automatic conversion between local and settlement currencies and fee accumulation, and complete the final fee calculation through global fee adjustment rules, forming a structured and highly transparent detailed list, which effectively reduces manual intervention and calculation errors.

[0047] (III) This intelligent pricing decision system for IEPL products based on multi-scenario classification, based on multi-dimensional customer characteristics and scenario information, intelligently recommends discount schemes by integrating learning models and rule bases, and provides data support for the approval process. It dynamically generates standardized approval paths based on discount amount, product type and scenario identifier, realizes automated role assignment and process locking, improves the rationality of discount decisions and optimizes approval efficiency. Attached Figure Description

[0048] Figure 1 This is a schematic diagram of the workflow of an intelligent pricing decision system for IEPL products based on multi-scenario classification according to the present invention.

[0049] Figure 2 This is a data flow diagram of an IEPL product intelligent pricing decision system based on multi-scenario classification according to the present invention. Detailed Implementation

[0050] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0051] Example 1, please refer to Figure 1 , Figure 2 This invention provides a technical solution: an intelligent pricing decision-making system for IEPL products based on multi-scenario classification, including an intelligent pricing decision-making center, which is communicatively connected to the following modules:

[0052] The intent parsing module integrates semantic parsing and scenario mapping capabilities to accurately transform unstructured customer questions into structured business intents and automatically associate them with corresponding business scenarios. This lays the foundation for accurate pricing and achieves seamless connection between intent and scenario. The intent parsing module includes a semantic perception unit and a scenario mapping unit.

[0053] The semantic awareness unit utilizes NLP technology to parse customer queries, identify key business entities and intent features within them, transform vague needs into explicit business query elements, and output a set of query elements. This improves the accuracy and automation of initial intent recognition, precisely parses unstructured queries, and significantly enhances the accuracy and automation of initial intent recognition. It receives unstructured text queries input by account managers or customer terminals, and uses a pre-trained natural language understanding model to perform word segmentation, part-of-speech tagging, and dependency parsing on the query statements to deconstruct their grammatical structure. This effectively parses users' colloquial or non-standard expressions, significantly improving the depth and accuracy of semantic understanding. Based on the deconstructed grammatical structure, the named entity recognition model is invoked to accurately extract key business entities from the query text. Key business entities include at least one or more of the following: geographic endpoint city, bandwidth specification, product level, and service quality requirements. Key business parameters are automatically extracted to reduce manual extraction errors and provide reliable input for subsequent accurate pricing. Combined with the extracted key business entities, the intent classification model is used to identify and classify the deep business purpose of the query, outputting a structured set of query elements. The set of query elements clearly includes the business action type and core parameter constraints, accurately identifying the user's true intent, avoiding misunderstanding bias, and realizing the automatic conversion of the query into structured business elements.

[0054] The specific work involves: upon receiving an unstructured text query submitted by an account manager or client terminal, firstly performing semantic structure parsing and preprocessing. Specifically, a pre-trained natural language understanding model based on the Transformer architecture is used. Through a fine-tuned BERT model, sequence labeling and syntactic parsing are performed on the query statement. The maximum length of the input text is set to 128 characters; excessively long texts are truncated, and lowercase conversion and special character filtering are applied uniformly. The word segmentation process uses a word segmenter compatible with the pre-trained model, simultaneously performing part-of-speech tagging and dependency parsing to generate a dictionary containing vocabulary, part-of-speech tags, and other relevant information. A structured syntax tree (SMR) is constructed, storing the original form of words, part-of-speech tagging results, and syntactic dependency relationships between words and other words in a sentence. This provides a complete grammatical structure for entity recognition. Throughout this process, the atomicity and traceability of all processing steps are maintained to ensure the accuracy and consistency of the parsing results. Based on the constructed SMR structure, a named entity recognition model is invoked to accurately extract key business entities. This model employs a sequence labeling architecture combining bidirectional long short-term memory networks and conditional random fields, and is specifically trained for IEPL business scenarios, limiting the scope of entity recognition. Based on a predefined set of entity categories, including but not limited to: geographic endpoint entities, bandwidth specification entities, product tier entities (covering standard, enhanced, and customized versions), and service quality requirement entities (latency, jitter, availability), the identified entities undergo boundary calibration and normalization based on dependency relationships in the syntax tree, forming a structured set of parameter key-value pairs. After extracting key business entities, the extracted entity set and the semantic representation of the original query statement are input into the intent classification model. This intent classification model adopts a multi-level classification architecture. The first level distinguishes core business action types, including price inquiries, configuration queries, and other actions. The first level compares the cases; the second level determines the subcategories based on product characteristics. It uses an attention mechanism to jointly encode the contextual semantics of the query statement and the extracted entity information to generate a 256-dimensional feature vector. The softmax function is used to output the probability distribution of each preset intent category. The confidence threshold is set to 0.85. When the confidence of the highest probability intent exceeds this confidence threshold, it is determined as the final business intent. Together with the previously extracted core parameter constraints, it forms a structured set of query elements. This set is standardized and encapsulated in JSON format, including intent category identifier, entity parameter list, and processing timestamp metadata.

[0055] The scenario mapping unit, based on the parsed key business entities and intent features, matches specific business scenarios covering cross-border and transboundary operations through a pre-set multi-label scenario classification model, forming standardized business scenario identifiers. This eliminates ambiguity in manual judgment, ensures consistency in subsequent process scenarios, and provides a unique and standardized business scenario context for subsequent pricing and routing. It receives a set of structured query elements and, based on a pre-set geographical location and network resource topology database, determines the relative positional relationship and cross-border attributes between geographical endpoint cities, accurately defining the physical attributes of the business and avoiding compliance and routing errors caused by geographical misjudgments. Based on the relative positional relationship and cross-border attribute judgment results, it inputs the query element set into a pre-trained multi-label scenario classification model. The multi-label scenario classification model calculates the probability that it belongs to one or more pre-set standard business scenarios based on a business rule feature library, integrating data prediction and expert rules to improve the accuracy and business fit of complex business scenario judgments. By comparing the pre-set probability threshold with the calculated probability, it generates and outputs a standardized business scenario identifier. This business scenario identifier uniquely corresponds to a specific combination of routing and billing rules, providing an unambiguous execution context for subsequent processes. The classification results are solidified into deterministic instructions, ensuring consistent context and clear paths for all subsequent automatic processing steps.

[0056] The specific tasks are as follows: After receiving the structured query element set output by the semantic perception unit, the system parses the geographic endpoint entity parameters. Based on the built-in geographic location coding database (which contains at least the latitude and longitude coordinates, country and administrative region codes of major cities worldwide, with country codes conforming to the ISO3166-1 standard and city coordinates accurate to at least six decimal places), the system performs precise matching and positioning of the endpoints. Subsequently, based on the country / region codes of the two locations and preset geofencing rules, the system automatically executes the location relationship and cross-border attribute judgment logic: if the country / region codes of the two locations are the same and they do not belong to a specific cross-border special administrative region, ... If the code is inconsistent, it is determined to be a purely domestic scenario; if the code is inconsistent, it is determined to be a cross-border scenario; if the path involves crossing a third country or region but the endpoint is within the same country, it is determined to be a cross-border scenario. This judgment process is based on deterministic Boolean logic rules and does not rely on probability calculations. At the same time, all input parameters and logical branch paths of the judgment process are recorded to ensure the auditability of the judgment results. After completing the geographic attribute judgment, the structured query element set is input into a pre-trained multi-label scenario classification model. This model adopts a BERT-based sequence classification architecture, with an input layer dimension of 768 and a fully connected output layer dimension corresponding to 12 preset standard business scenario categories. Based on the weight parameters trained offline, forward propagation is performed on the input features, and a 12-dimensional probability vector is output through the Softmax function. Each dimension represents the probability that the input query belongs to the corresponding standard business scenario. During model prediction, a business rule feature library loaded in memory is used to constrain and post-process the initial probability distribution, ensuring that the classification results conform to the enterprise's established business logic. The business rule feature library contains over 500 rule-scenario mapping relationships defined by domain experts. The probabilities of each scenario output by the multi-label scenario classification model are compared with preset activation thresholds, which are configured hierarchically. The activation threshold for the primary scenario is 0.65, and the activation threshold for the secondary scenario is 0.35. If the probability of a scenario exceeds its corresponding threshold, the scenario is activated. Based on the combination of activated scenarios, a unique standardized business scenario identifier is generated according to a predefined mapping table. This business scenario identifier is used as a global key value and directly associated with the path constraint entries in the subsequent routing rule library and the cost calculation template in the billing model library. Finally, the business scenario identifier, along with the complete probability distribution vector, threshold judgment record, and timestamp, is encapsulated into a scenario judgment result object and output to the downstream intelligent routing and pricing module to provide a definite execution context for subsequent processes.

[0057] The intelligent routing and pricing module integrates a routing rule base and a dynamic cost engine. It automatically plans the optimal path and calculates segmented costs based on the associated business scenarios, and generates cost combinations for business scenarios in an automated and standardized manner, eliminating omissions or duplicate pricing. The intelligent routing and pricing module includes a rule routing unit and a cost aggregation unit.

[0058] The rule-based routing unit, based on a built-in multi-scenario routing rule library covering both cross-border and purely overseas scenarios, automatically matches compliant paths and mandatory nodes according to the attributes of the business scenario, and outputs the selected end-to-end routing scheme. This replaces manual path judgment, improving the accuracy and efficiency of route planning. It automatically generates the optimal compliant path scheme, significantly enhancing the efficiency and accuracy of route planning. It receives standardized business scenario identifiers and, based on these identifiers, retrieves matching path planning constraint sets from a pre-set multi-scenario routing rule library covering both cross-border and purely overseas scenarios. These constraint sets include mandatory traversal nodes, avoided areas, and transmission protocol requirements. Through scenario-driven precise retrieval, it achieves... The system automatically matches existing rules and conditions with business requirements, avoiding errors from manual configuration. Based on the set of path planning constraints and combined with real-time network resource status data, it uses graph theory algorithms to calculate one or more optimal feasible paths that meet business requirements in a pre-built network topology graph. Real-time dynamic optimization ensures path feasibility, improves the efficiency of route planning and resource utilization, and performs compliance and cost pre-evaluation and verification on the calculated feasible paths. Finally, it selects and outputs a standardized end-to-end routing scheme. This end-to-end routing scheme clearly lists all necessary network nodes and segment divisions. Double verification ensures that the path is compliant and economical, and outputs a standard routing scheme with a clear structure that can be delivered and executed.

[0059] The specific work involves: upon receiving a standardized business scenario identifier, performing a precise search from a pre-defined multi-scenario routing rule base based on that identifier. This multi-scenario routing rule base uses the business scenario identifier as the primary key and is stored in YAML format. Each rule contains at least a constraint version number, effective time, and specific constraint parameters. Then, it retrieves a matching set of path planning constraints from the multi-scenario routing rule base. This set of constraints specifically covers three dimensions: a mandatory traversal node list explicitly specifying the physical or logical nodes that the path must include, such as international gateways or designated submarine cable landing stations; a avoidance area list using GeoJSON polygon coordinates to define geographical areas or autonomous regions that the route is prohibited from traversing; and transmission protocol requirements explicitly specifying the protocols that must be used. The protocol stack, along with protocol parameters, employs an exact matching algorithm during the retrieval process. Upon successful matching, the complete set of constraints is loaded into the memory workspace, simultaneously recording the retrieval timestamp, rule version, and matching degree index (a matching rate of 100%) to ensure the integrity and traceability of constraint application. Based on the acquired path planning constraint set, the system accesses the network resource status database in real time to obtain current network status parameters. This database updates the entire network status every 30 seconds, with key parameters including node activity status, available link bandwidth, current latency, and packet loss rate. The physical network is abstracted into a weighted directed graph model, where nodes correspond to network devices and edges correspond to transmission links. An improved constraint-based shortest path first algorithm is used for calculation. The algorithm is designed... The iteration count is capped at 1000. Under the premise of satisfying the mandatory node sequence and avoiding regional constraints, the optimization objective is a comprehensive cost function (bandwidth cost 60%, latency cost 30%, reliability cost 10%). 3-5 candidate paths are computed in parallel. During the computation, the link bandwidth margin is checked in real-time to ensure sufficient reserved bandwidth for the business. Each candidate path output includes a complete node sequence, link sequence, and various performance prediction indicators. The output candidate paths undergo dual verification, including compliance verification and cost pre-assessment. In the compliance verification phase, the compliance rule engine is invoked to verify whether the path complies with regulatory policy requirements and enterprise security strategies. In the cost pre-assessment phase, based on the path's various... The segment distance (calculated based on precise latitude and longitude between nodes using the Vincenty formula) and corresponding pricing template are used to calculate the base cost. The cost fluctuation range is controlled within ±5% of the estimated value. After verification, a multi-attribute decision analysis method is used for final selection. The evaluation dimensions are weighted as follows: compliance and security score (40%), comprehensive cost score (35%), and performance prediction score (25%). After the selected path is standardized, an end-to-end routing plan document is generated. This document clearly lists all necessary nodes, segment divisions, and key performance commitment values ​​(upper limit of latency, availability not less than 99.95%). The plan is output in JSON Schema format, including metadata such as the plan ID, generation time, and verification signature.

[0060] The cost aggregation unit, based on routing results, automatically calculates and intelligently combines the costs of each segment using a built-in segmented pricing model. It supports parameterized adjustments, quickly generates and visualizes cost combinations for specific business scenarios, reduces human calculation errors, and automates segmented pricing and cost aggregation, effectively eliminating omissions and errors in manual calculations. Based on the output standardized end-to-end routing scheme, it parses the various transmission segments it contains and calls the corresponding billing functions from the built-in segmented pricing model library according to segment attributes. Billing function parameters include distance, bandwidth, and resource type. By matching the segmented pricing model, it achieves accurate and automated baseline cost calculation for different network resource types, eliminating billing model errors. To prevent misuse of pricing models, the system extracts bandwidth specifications and product level parameters from key business entities and inputs them into the billing functions of each segment. This automatically calculates the base cost for each routing segment and performs currency conversion between the local currency and the settlement currency. Based on real-time exchange rate-based automatic conversion and accumulation, the system ensures the financial accuracy and completeness of cross-border pricing. According to the preset cost combination logic, the system accumulates the base costs for all segments and applies global cost adjustment rules corresponding to the business scenario identifier to generate a structured detailed cost list as the basis for pricing. The system then dynamically adjusts costs using scenario-based strategies to generate a clear, auditable final pricing list, thereby enhancing customer trust.

[0061] The specific work involves: Based on the generated standardized end-to-end routing scheme, firstly, automatically parsing the transmission segment division results defined in the routing scheme. Each segment's attributes include key billing dimensions such as transmission type, geographical span, and the types of network resources traversed. According to the segment attributes, the corresponding billing function is precisely matched and called from the built-in segmented pricing model library. This model library pre-sets billing functions for different resource types, including domestic trunk lines, international submarine cables, cross-border terrestrial cables, and metropolitan area access. Each billing function uses distance, bandwidth, and resource type as core input parameters. During calculation, the key business entity parameters extracted by the semantic awareness unit, namely the specific bandwidth specifications and product level, are used as input values ​​and substituted into each billing function. Based on this, the calculation is automatically performed, calculating the base cost of each routing segment in local currency. After calculating the base cost of each routing segment, currency standardization is then performed. Based on the real-time exchange rate obtained from the international exchange rate service interface, the base cost of each segment, denominated in local currency, is uniformly converted to the settlement currency agreed in the contract. Exchange rate data is updated hourly, and the conversion process is accurate to four decimal places and recorded. The exchange rate version and timestamp used ensure financial traceability. Subsequently, according to the preset fee combination logic, the fees of all segments after conversion are summed. This combination logic defines clear fee collection rules to ensure no omissions or double calculations. Based on the preliminary fee total, the global fee adjustment rules corresponding to the current business scenario identifier are applied. These global fee adjustment rules are stored in the scenario fee strategy library, indexed by the business scenario identifier. They include a uniform discount coefficient based on customer level, a premium ratio for specific product level, and additional fee items required to meet service quality requirements. The relevant rules are automatically invoked to make a final adjustment to the preliminary fees, calculate the total amount including all cost components, and finally generate a structured detailed fee list. This detailed fee list lists the resource type, distance, unit price, base fee, adjustment items, and adjusted fees in detail by segment, and clearly presents the entire process of fee accumulation and adjustment, forming a complete financial basis that can be directly used for customer quotations. All calculation parameters, exchange rates, adjustment rules, and their versions are attached to the list as metadata to ensure the transparency and auditability of quotations.

[0062] The strategy matching engine, based on multi-dimensional information such as authorized customer history, product attributes, and city level, intelligently matches and recommends discount ranges from the preset discount strategy matching model, providing data-driven basis for approval, improving the rationality of discount schemes and customer satisfaction, and providing intelligent data support for discount approval.

[0063] The process decision module is used to automatically determine and assign appropriate approval roles and paths based on the discount amount and product type rules covered in the matched recommended discount range, dynamically allocate permissions, avoid process conflicts and delays, optimize approval efficiency, and dynamically assign appropriate approval roles and paths to effectively avoid process conflicts and significantly optimize approval efficiency.

[0064] The collaborative decision-making module aggregates routing, cost, and discount approval results, generates structured and visualized customized quotations with one click, and collects interaction feedback data between account managers and customers. Through machine learning, it continuously optimizes intent recognition, scenario classification, and discount strategies, forming a closed-loop learning mechanism to promote the continuous evolution of the system's intelligence level.

[0065] Example 2, as Figure 1 , Figure 2 As shown, based on Embodiment 1, the present invention provides a technical solution: the strategy matching engine performs the following steps: obtaining customer identifiers from the structured query element set, querying the customer's historical transaction database, and combining business scenario identifiers and product levels to form a complete feature vector for the current quotation request, realizing the dynamic fusion of customer profiles and real-time inquiry scenarios, providing multi-dimensional and high-precision input features for intelligent discount matching, inputting the feature vectors into a preset discount strategy matching model, which performs multi-dimensional similarity calculation and weight evaluation on the feature vectors and rules in the strategy knowledge base, outputting one or more applicable candidate discount strategies and their confidence levels, combining data-driven and rule-based reasoning to output high-confidence candidate discount strategies, improving the rationality and reliability of discount recommendations, filtering and merging candidate discount strategies according to preset decision rules, generating the final recommended discount scheme and the specific strategy terms matched by the scheme, providing data support for approval decisions, automatically generating structured discount suggestions and complete evidence, significantly improving approval efficiency and decision transparency;

[0066] The specific work involves: Based on the unique customer identifier extracted from the structured query element set, accessing the customer's historical transaction database through an encrypted secure channel. This database employs a distributed columnar storage architecture and can query transaction records from the past three years, including data such as contract amount, product type, historical discount rate, and performance evaluation. Key customer-level features are automatically extracted, including but not limited to: annual total consumption segments, cooperation duration, previous quarterly purchase growth rate, and historical quote acceptance rate. Subsequently, the customer features are vectorized and concatenated with the business scenario identifier and explicit product level parameters in the current request, forming a standardized feature vector of length 128 dimensions. This vectorization process follows enterprise data governance standards, performing maximum and minimum value normalization on continuous variables and one-hot encoding on categorical variables to ensure consistent data scale and no information loss in the input model. The generated 128-dimensional feature vector is then input in real-time into a discount strategy matching model deployed in an in-memory computing engine. This model uses an ensemble learning architecture based on gradient boosting decision trees, containing 150 decision trees, each with a maximum depth of 8. When processing the input vector, the model first compares it with over 500 predefined decision trees in the strategy knowledge base. The system employs multi-dimensional matching based on rules. Each rule is defined by domain experts and has clear activation conditions, priority weights, and associated discount ranges. The matching process calculates the cosine similarity and Euclidean distance between the feature vector and each rule condition, and performs a weighted evaluation by integrating the preset weights of the rules. The model outputs at most five candidate discount strategies. Each strategy includes a strategy number, a recommendation discount rate, a confidence score, and the key matching rule points. The confidence threshold is set at 0.65; strategies below this threshold are automatically filtered out. The system automatically processes the candidate discount strategies output by the model according to preset decision rules. The selection and fusion process involves a decision rule base that includes conflict resolution mechanisms, strategy merging logic, and risk control constraints. Candidate strategies are sorted in descending order of confidence level, and strategies that meet the merging conditions are intelligently merged to generate a structured final recommended discount scheme. This scheme clearly lists the overall discount rate, the base price (i.e., the total amount output by the cost aggregation unit), and the discounted amount, and includes detailed information on the specific policy clause number, effective date, and applicable conditions on which it is based. Finally, this scheme and all supporting data are encapsulated into a JSON object, synchronously pushed to the process decision module, and a complete decision audit log is retained.

[0067] The process decision module executes the following steps: receiving recommended discount schemes and detailed cost lists, parsing discount amounts and product types, and determining the approval matrix version triggered by the process based on business scenario identifiers to ensure that each quotation approval strictly follows the latest internal control rules, eliminating the risk of rule lag or misuse. Based on the parsed parameters and the determined approval matrix, the workflow rule engine is dynamically matched to automatically determine the approval role sequence, approval path branches, and approval thresholds of each node required for this quotation request, realizing personalized and accurate generation of approval paths, avoiding human judgment bias and process conflicts. Based on the judgment results, the corresponding electronic approval process instance is automatically generated, and the task is pushed to the to-do items of the corresponding approval role. At the same time, the quotation is locked until the process is completed, realizing the automated assignment and execution of approval paths, full-process automated flow and strong status control, greatly improving approval efficiency and preventing mid-process tampering.

[0068] The specific tasks are as follows: Receive structured discount schemes from the strategy matching engine and detailed expense lists generated by the expense aggregation unit; parse key parameters in the discount schemes, including absolute discount amounts or relative discount rates, and extract product type codes; simultaneously, based on the unique business scenario identifier carried in the expense list, perform a precise query in the enterprise's internally maintained approval matrix version control table. This approval matrix version control table uses the business scenario identifier as the primary key, associates it with a valid time range, and automatically filters out approval matrix version numbers that are in effect at the time the quote was initiated, ensuring that the approval rules followed by the process are strictly synchronized with the enterprise's latest internal control policies; based on the parsed discount amount, product type, and determined approval matrix version, call the preset workflow rule engine for automated decision-making. The workflow rule engine uses the discount amount range and business scenario attributes as input conditions, matches the logical rules defined in the matrix, and calculates and outputs the standardized approval path corresponding to this quote request. This path clearly defines the necessary approval angles. The system uses color sequences and conditional branches, and sets specific approval operation thresholds for each approval node. The entire judgment process is completed in memory. Based on the dynamically generated approval path definition, a structured electronic approval process instance is automatically instantiated in the intelligent quotation decision center. A globally unique process tracking number is assigned to this instance. Based on the enterprise organizational structure database, the tasks of each node in the process are pushed to the corresponding approver's to-do list in real time, and notifications can be sent via WeChat, DingTalk, or email. At the same time, the status of the quotation is marked as "under approval" and logically locked to prevent any modification to the core terms of the quotation amount and discount scheme before the process is completed. The processing status and timeliness of each node are monitored, and reminders are automatically sent for overdue tasks. The complete approval trajectory and opinions are recorded. The process is only completed when all necessary nodes have completed the approval or rejection operation according to the rules. The quotation is then unlocked and its status is updated, realizing automated allocation and precise execution of the entire lifecycle of the approval path.

[0069] The collaborative decision-making module executes the following steps: It aggregates detailed cost lists from the intelligent routing and pricing module, recommended discount schemes from the strategy matching engine, and approval status from the process decision-making module. Following a preset template, it automatically generates a structured quotation document containing routing diagrams, cost breakdowns, and terms and conditions. This generates a standardized and reliable quotation with a single click, significantly improving the professionalism and efficiency of business communication. After the quotation is sent to the customer, it collects customer feedback results and final transaction information entered by the account manager through an integrated interface. This, along with various intermediate decision-making data from the quotation process, is structured and stored in the feedback learning database. This comprehensively accumulates business decision-making data assets, providing a reliable basis for intelligent model optimization. Regularly, using data from the feedback learning database, it performs parameter tuning and version iteration on the natural language understanding model in the semantic perception unit, the multi-label scene classification model in the scene mapping unit, and the discount strategy matching model in the strategy matching engine through offline training. This forms a continuously optimized closed-loop learning mechanism, enabling dynamic evolution of the system's intelligence level and continuously improving quotation accuracy and market adaptability.

[0070] The specific tasks are as follows: After completing the intelligent routing pricing, discount matching, and approval process, automatically execute the quotation synthesis operation. Following a preset JSON template, aggregate the structured detailed cost list output by the intelligent routing pricing module, the standard discount scheme generated by the strategy matching engine (including the overall discount rate, discounted amount, and the basis for the strategy terms), and the final approval status provided by the process decision module. Subsequently, call the built-in chart generation engine to automatically draw an end-to-end routing diagram in SVG format based on the node sequence and geographical coordinates in the routing scheme and embed it into the document. The quotation follows enterprise document standards, and its core sections include at least: project summary, detailed routing and service level agreement, tiered cost breakdown table, and explanation of applicable discount terms. The complete approval serial number and version metadata are included. The final generated quotation document is output simultaneously in both PDF and structured JSON formats and pushed to the customer relationship management system via an encrypted channel for account managers to submit to clients. Once the quotation enters the client review stage, a feedback data collection process is initiated. When the account manager enters the client's final decision and possible transaction amounts and terms changes into the system, the information is automatically captured through a verified application programming interface. Simultaneously, all intermediate data generated throughout the entire quotation lifecycle, including but not limited to semantically parsed structured query elements, probability distribution vectors for multi-label scenario classification, candidate path sets for routing calculations and their evaluation metrics, and multiple candidate strategies for discount matching, are also collected. The confidence level and complete log of the approval process are correlated and integrated. All data is cleaned, de-identified, and standardized according to a predefined schema, and then persistently stored in a dedicated feedback learning database in time series format. This database adopts a columnar storage architecture, supporting efficient multidimensional queries and analysis, and providing a high-quality, traceable training sample set for model iteration. The offline training and iteration process of the model is initiated at fixed intervals. The training task extracts the completed quotation process data and its final business results within a statistical period from the feedback learning database as training samples. Specifically, sample pairs containing the original query text and the final confirmation business scenario are used for multi-label scenarios based on the BERT architecture. The classification model undergoes incremental training to optimize the accuracy of its 12-dimensional output vector. A gradient-boosting decision tree-based discount strategy matching model is retrained using sample pairs containing customer features, pricing context, and actual transaction discounts, adjusting the parameters of its 150 decision trees to more accurately predict acceptable discount rates. Simultaneously, the entity recognition and intent classification modules of the natural language understanding model are jointly fine-tuned using end-to-end data. Model evaluation is conducted on an independent validation set. Automated canary releases and replacements are only performed when the new model significantly outperforms the production version in key metrics (improvement of F1 score in scenario classification exceeding 0.02, and reduction of mean squared error in discount prediction exceeding 5%), forming a data-driven continuous optimization loop.

[0071] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0072] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A multi-scenario classification-based intelligent pricing decision system for IEPL products, comprising an intelligent pricing decision center, characterized in that, The intelligent pricing decision center has the following communication modules: The intent parsing module integrates semantic parsing and scenario mapping capabilities to accurately convert unstructured customer questions into structured business intents and automatically associate them with corresponding business scenarios. The intelligent routing and pricing module integrates a routing rule base and a dynamic cost engine. It automatically plans the optimal path and calculates segmented costs based on the associated business scenarios, generating cost combinations for each business scenario. The strategy matching engine intelligently matches and recommends discount ranges from a preset discount strategy matching model, based on multi-dimensional information such as authorized customer history, product attributes, and city level. The process decision module is used to automatically determine and assign appropriate approval roles and paths based on the discount amount and product type rules covered in the matched recommended discount range; The collaborative decision-making module aggregates routing, cost, and discount approval results, generates structured customized quotations with one click, and collects interaction feedback data between account managers and customers. Through machine learning, it continuously optimizes intent recognition, scenario classification, and discount strategies, forming a closed-loop learning process.

2. The intelligent pricing decision system for IEPL products based on multi-scenario classification according to claim 1, characterized in that: The intent parsing module includes a semantic perception unit and a scene mapping unit; The semantic perception unit is used to use NLP technology to parse customer questions, identify key business entities and intent features in customer questions, transform fuzzy requirements into clear business query elements, and output a set of query elements. The scenario mapping unit, based on the parsed key business entities and intent features, matches specific business scenarios covering cross-border and transboundary activities through a preset multi-label scenario classification model, forming standardized business scenario identifiers and eliminating ambiguity in manual judgment.

3. The intelligent pricing decision-making system for IEPL products based on multi-scenario classification according to claim 2, characterized in that: The semantic awareness unit performs the following steps: It receives unstructured text queries input by account managers or customer terminals, and uses a pre-trained natural language understanding model to perform word segmentation, part-of-speech tagging, and dependency parsing on the query statements to deconstruct their grammatical structure. Based on the deconstructed grammatical structure, the named entity recognition model is invoked to accurately extract key business entities from the query text. Key business entities include at least one or more of the following: geographic endpoint city, bandwidth specification, product level, and service quality requirements. By combining the extracted key business entities, the deep business purpose of the query is identified and classified through an intent classification model, and a structured set of query elements is output. The set of query elements explicitly includes the business action type and core parameter constraints.

4. The intelligent pricing decision-making system for IEPL products based on multi-scenario classification according to claim 2, characterized in that: The scene mapping unit performs the following steps: Receive a set of structured query elements and determine the relative positional relationship and cross-border attributes between geographic endpoint cities based on a preset geographic location and network resource topology database; Based on the relative positional relationship and cross-border attribute judgment results, the query element set is input into the pre-trained multi-label scenario classification model. The multi-label scenario classification model calculates the probability that it belongs to one or more preset standard business scenarios based on the business rule feature library. By comparing the preset probability threshold with the calculated probability, a standardized business scenario identifier is generated and output. The business scenario identifier uniquely corresponds to a specific combination of routing and billing rules.

5. The intelligent pricing decision-making system for IEPL products based on multi-scenario classification according to claim 2, characterized in that: The intelligent routing and pricing module includes a rule-based routing unit and a cost aggregation unit; The rule routing unit, based on a built-in multi-scenario routing rule library for cross-border and purely overseas scenarios, automatically matches compliant paths and necessary nodes according to the attributes of the business scenario, and outputs the selected end-to-end routing scheme. The cost aggregation unit automatically calculates and intelligently combines the costs of each segment based on the routing results by calling the built-in segmented pricing model. It supports parameterized adjustments and quickly generates and visualizes cost combinations for business scenarios.

6. The intelligent pricing decision-making system for IEPL products based on multi-scenario classification according to claim 5, characterized in that: The rule routing unit performs the following steps: The system receives the standardized business scenario identifier and, based on the business scenario identifier, retrieves a matching set of path planning constraints from a preset multi-scenario routing rule base covering cross-border and purely overseas scenarios. The set of path planning constraints includes mandatory traversal nodes, avoidance areas, and transmission protocol requirements. Based on the set of path planning constraints and combined with real-time network resource status data, graph theory algorithms are used to calculate one or more optimal feasible paths that meet business requirements in a pre-constructed network topology graph. The calculated feasible paths are pre-evaluated for compliance and cost, and finally a standardized end-to-end routing scheme is selected and output. This end-to-end routing scheme clearly lists all necessary network nodes and segment divisions.

7. The intelligent pricing decision-making system for IEPL products based on multi-scenario classification according to claim 5, characterized in that: The cost aggregation unit performs the following steps: Based on the output standardized end-to-end routing scheme, the various transmission segments it contains are parsed, and the corresponding billing function is called from the built-in segmented pricing model library according to the segment attributes. The billing function parameters include distance, bandwidth, and resource type. The bandwidth specifications and product level parameters extracted from the key business entities are input into the billing functions of each segment to automatically calculate the base cost of each routing segment and convert the local currency to the settlement currency. Based on the preset cost combination logic, the calculated base costs for all segments are summed up, and the global cost adjustment rules corresponding to the business scenario identifier are applied to generate a structured detailed cost list as the basis for quotation.

8. The intelligent pricing decision system for IEPL products based on multi-scenario classification according to claim 5, characterized in that: The strategy matching engine performs the following steps: Obtain customer identifiers from the structured query element set, correlate them with the customer's historical transaction database, and combine them with business scenario identifiers and product levels to form a complete feature vector for the current quote request; The feature vector is input into a preset discount strategy matching model. The discount strategy matching model performs multi-dimensional similarity calculation and weight evaluation between the feature vector and the rules in the strategy knowledge base, and outputs one or more applicable candidate discount strategies and their confidence scores. Based on preset decision-making rules, candidate discount strategies are screened and integrated to generate the final recommended discount scheme and the specific strategy terms matched with the scheme.

9. The intelligent pricing decision system for IEPL products based on multi-scenario classification according to claim 8, characterized in that: The process decision module performs the following steps: Receive recommended discount schemes and detailed expense lists, parse the discount amount and product type, and determine the approval matrix version triggered by the process based on the business scenario identifier; Based on the parsed parameters and the determined approval matrix, the workflow rule engine is dynamically matched to automatically determine the approval role sequence, approval path branches and approval thresholds of each node required for this quotation request. Based on the judgment results, the corresponding electronic approval process instance is automatically generated, and the task is pushed to the to-do list of the corresponding approval role. At the same time, the quotation is locked until the process is completed.

10. The intelligent pricing decision system for IEPL products based on multi-scenario classification according to claim 9, characterized in that: The collaborative decision-making module performs the following steps: It aggregates detailed cost lists from the intelligent routing and pricing module, recommended discount schemes from the strategy matching engine, and approval status from the process decision module, and automatically generates a structured quotation document containing route diagrams, cost breakdowns, and terms and conditions according to a preset template. After the quotation is sent to the customer, the customer feedback results and final transaction information entered by the account manager are collected through the integration interface, along with various intermediate decision data in the quotation process, and stored in the feedback learning database in a structured manner. By periodically utilizing the data in the feedback learning database, the parameters of the natural language understanding model in the semantic perception unit, the multi-label scene classification model in the scene mapping unit, and the discount strategy matching model in the strategy matching engine are tuned and iterated through offline training, forming a continuously optimized closed-loop learning mechanism.

Images