Platform trajectory data processing method and device, equipment and storage medium

Abstract

The present application relates to the technical field of intelligent logistics, and particularly relates to a platform trajectory data processing method, device and equipment and a storage medium, wherein the method first acquires a logistics trajectory data set and a preset field screening rule, performs text extraction processing on the logistics trajectory data set to obtain a plurality of business scenario description texts, then calls a pre-trained feature selection model to perform field screening processing on each business scenario description text based on the field screening rule to obtain a target field set, calls a pre-trained semantic model to perform semantic coding and template matching processing on the target field set to obtain a template field mapping table, then performs similarity calculation and screening processing based on the template field mapping table to obtain a target template set, and finally performs copywriting generation processing on the target template set to obtain a target copywriting set, and determines an initial online copywriting based on the target copywriting set, aiming to improve the efficiency of e-commerce logistics trajectory data processing.

Description

Technical Field

[0001] This invention relates to the field of intelligent logistics technology, and in particular to a platform trajectory data processing method, apparatus, equipment, and storage medium. Background Technology

[0002] The platform trajectory push system is a core support module of the e-commerce logistics chain. It is primarily responsible for pushing corresponding logistics trajectory information to multiple e-commerce platforms in real time, ensuring the smooth operation of platform logistics index statistics and related businesses. This system covers diverse business scenarios such as real-time order subscription on the platform side, standardized format packaging, logistics trajectory query, recipient identity verification, and package interception. It possesses core capabilities such as real-time ordered push, comprehensive statistics, traceability logs, anomaly warnings, and supplementary push, meeting the business needs of the entire e-commerce logistics chain.

[0003] Currently, the company's abnormal receipt and problem item handling business involves a variety of specific business scenarios, requiring frequent adjustments to the platform display text and corresponding status codes. In the existing technology, the encapsulation logic of the text and status codes is implemented by hard coding. Each iteration adjustment requires a complete development, testing and release process, which is cumbersome and time-consuming.

[0004] As e-commerce business iterates at an accelerated pace and the frequency of scenario adjustments increases year by year, hard-coded models are no longer able to adapt to the needs of rapid response, resulting in low efficiency in business adjustments and indirectly affecting the logistics service experience. It is evident that existing technologies still need to be improved and enhanced. Summary of the Invention

[0005] In order to overcome the shortcomings of the prior art, the present invention aims to provide a platform trajectory data processing method, apparatus, device and storage medium, which aims to solve the problem of low response efficiency in the existing platform trajectory push system, where business adjustments require a complete development, testing and release process. This invention also aims to improve the efficiency of e-commerce logistics trajectory data processing.

[0006] The first aspect of this invention provides a platform trajectory data processing method, comprising: acquiring a logistics trajectory dataset and preset field filtering rules; performing text extraction processing on the logistics trajectory dataset to obtain multiple business scenario description texts; calling a pre-trained feature selection model to perform field filtering processing on each of the business scenario description texts based on the field filtering rules to obtain a target field set; calling a pre-trained semantic model to perform semantic encoding and template matching processing on the target field set to obtain a template field mapping table; performing similarity calculation and filtering processing based on the template field mapping table to obtain a target template set; performing copywriting generation processing on the target template set to obtain a target copywriting set, and determining the initial online copywriting based on the target copywriting set.

[0007] Optionally, in a first implementation of the first aspect of the present invention, the step of performing text extraction processing on the logistics trajectory dataset to obtain multiple business scenario description texts includes: using a feature extraction algorithm to perform target field identification and extraction processing on the logistics trajectory dataset to obtain multiple initial business scenario fields; obtaining preset field verification rules, and performing redundant field removal processing on each of the initial business scenario fields based on the field verification rules to obtain multiple valid business scenario fields; and using a single trajectory field attribution association algorithm to perform structured splicing processing on the multiple valid business scenario fields to obtain multiple business scenario description texts.

[0008] Optionally, in a second implementation of the first aspect of the present invention, the feature selection model includes a field structured splitting module, a contribution calculation module, and a filtering module, wherein the field structured splitting module, the contribution calculation module, and the filtering module are connected sequentially; the step of calling the pre-trained feature selection model to perform field filtering processing on each of the business scenario description texts based on the field filtering rules to obtain a target field set includes: performing field splitting processing on each of the business scenario description texts based on the field structured splitting module to obtain multiple split fields; performing contribution value calculation processing on each of the split fields based on the contribution calculation module to obtain a contribution value corresponding to each split field; and performing filtering processing on the multiple split fields according to the contribution value based on the filtering module and the field filtering rules to obtain the target field set.

[0009] Optionally, in a third implementation of the first aspect of the present invention, the semantic model includes a template encoding module, a semantic encoding module, and a matching module, wherein the template encoding module and the semantic encoding module are respectively connected to the matching module; the step of calling the pre-trained semantic model to perform semantic encoding and template matching processing on the target field set to obtain a template field mapping table includes: obtaining a preset historical copywriting template library, wherein the historical copywriting template library includes multiple historical copywriting templates; performing template encoding processing on each of the historical copywriting templates based on the template encoding module to obtain multiple template semantic vectors; performing field encoding processing on the target field set based on the semantic encoding module to obtain multiple field semantic vectors; and performing one-to-one pairing processing on the multiple template semantic vectors and the multiple field semantic vectors based on the matching module to establish an association relationship between each template semantic vector and the corresponding field semantic vector to obtain the template field mapping table.

[0010] Optionally, in the fourth implementation of the first aspect of the present invention, the step of performing similarity calculation and filtering based on the template field mapping table to obtain a target template set includes: using a cosine similarity algorithm to calculate the cosine similarity between each template semantic vector and the corresponding field semantic vector based on the template field mapping table, to obtain the similarity corresponding to each historical copy template; sorting all the historical copy templates according to the corresponding similarity from high to low to obtain a template sorting result; obtaining a preset template filtering quantity, and obtaining historical copy templates matching the template filtering quantity based on the template sorting result to obtain the target template set.

[0011] Optionally, in the fifth implementation of the first aspect of the present invention, the step of performing copywriting generation processing on the target template set to obtain a target copywriting set includes: parsing the target template set using a parsing algorithm to extract multiple EL (Expression Language) placeholders; calling a pre-trained BERT (Bidirectional EncoderRepresentations from Transformers) model to classify each EL placeholder by business logic attributes, obtaining multiple business-type EL placeholders and multiple non-business-type EL placeholders, and standardizing each business-type EL placeholder to obtain standardized EL placeholders; obtaining a preset function mapping library, and matching value-taking functions for all non-business-type EL placeholders and all standardized EL placeholders based on the function mapping library to obtain the target copywriting set.

[0012] Optionally, in the sixth implementation of the first aspect of the present invention, after determining the initial online copy based on the target copy set, the method further includes: acquiring historical business push datasets, actual business scenarios, and real-time business push data of the initial online copy; using a semantic similarity algorithm to calculate the matching degree between the actual business scenario and the initial online copy based on the business push data; calling a pre-trained LSTM (Long Short-Term Memory) model to calculate a dynamic matching degree threshold based on the historical business push dataset; comparing the matching degree with the dynamic matching degree threshold; and when the matching degree is less than the dynamic matching degree threshold, re-determining a new online copy based on the target copy set.

[0013] A second aspect of the present invention provides a platform trajectory data processing device, comprising: a data acquisition module for acquiring a logistics trajectory dataset and preset field filtering rules; a text extraction module for performing text extraction processing on the logistics trajectory dataset to obtain multiple business scenario description texts; a field filtering module for calling a pre-trained feature selection model to perform field filtering processing on each of the business scenario description texts based on the field filtering rules to obtain a target field set; a matching module for calling a pre-trained semantic model to perform semantic encoding and template matching processing on the target field set to obtain a template field mapping table; a similarity calculation module for performing similarity calculation and filtering processing based on the template field mapping table to obtain a target template set; and a copywriting generation module for performing copywriting generation processing on the target template set to obtain a target copywriting set, and determining an initial online copywriting based on the target copywriting set.

[0014] A third aspect of the present invention provides a platform trajectory data processing device, the platform trajectory data processing device comprising: a memory and at least one processor, the memory storing instructions; at least one processor calling the instructions in the memory to cause the platform trajectory data processing device to perform the various steps of the platform trajectory data processing method described in any of the preceding claims.

[0015] A fourth aspect of the present invention provides a computer-readable storage medium storing instructions that, when executed by a processor, implement the steps of the platform trajectory data processing method described in any of the preceding claims.

[0016] In the technical solution of this invention, a logistics trajectory dataset and preset field filtering rules are first obtained. The logistics trajectory dataset is then processed by text extraction to obtain multiple business scenario description texts. A pre-trained feature selection model is then called to perform field filtering processing on each business scenario description text based on the field filtering rules to obtain a target field set. A pre-trained semantic model is then called to perform semantic encoding and template matching processing on the target field set to obtain a template field mapping table. Next, similarity calculation and filtering processing are performed based on the template field mapping table to obtain a target template set. Finally, the target template set is processed by copywriting generation to obtain a target copy set. The initial online copy is determined based on the target copy set. This invention aims to solve the problem of low response efficiency in existing platform trajectory push systems where business adjustments require a complete development, testing, and release process, thereby improving the efficiency of e-commerce logistics trajectory data processing. Attached Figure Description

[0017] Figure 1 This is a first flowchart of a platform trajectory data processing method provided in an embodiment of the present invention; Figure 2This is a second flowchart of a platform trajectory data processing method provided in an embodiment of the present invention; Figure 3 This is a third flowchart of the platform trajectory data processing method provided in the embodiments of the present invention; Figure 4 This is a fourth flowchart of the platform trajectory data processing method provided in the embodiments of the present invention; Figure 5 A fifth flowchart of the platform trajectory data processing method provided in the embodiments of the present invention; Figure 6 The sixth flowchart of the platform trajectory data processing method provided in the embodiments of the present invention; Figure 7 A seventh flowchart of a platform trajectory data processing method provided in an embodiment of the present invention; Figure 8 This is a schematic diagram of the structure of the platform trajectory data processing device provided in an embodiment of the present invention; Figure 9 This is a schematic diagram of the structure of the platform trajectory data processing device provided in an embodiment of the present invention. Detailed Implementation

[0018] This invention provides a platform trajectory data processing method, apparatus, device, and storage medium. In this invention, the terms "first," "second," "third," "fourth," etc. (if present)," in the specification, claims, and accompanying drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data used can be interchanged where appropriate so that the embodiments described herein can be implemented in a sequence other than that illustrated or described herein. Furthermore, the terms "comprising" or "having," and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or device that includes a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or devices.

[0019] For ease of understanding, the specific process of the embodiments of the present invention is described below. Please refer to [link / reference]. Figure 1 One embodiment of the platform trajectory data processing method in this invention includes: 101. Obtain the logistics trajectory dataset and preset field filtering rules; In this embodiment, the acquired logistics trajectory dataset covers multi-dimensional raw data generated throughout the entire express delivery process, including scan type, remarks code, remarks information, user-defined fields, time fields, and location fields. It fully carries the business description information in the e-commerce platform push scenario, accurately reflecting the core characteristics of various business scenarios such as abnormal receipts and user refusal, providing a comprehensive data foundation for subsequent feature selection and semantic processing. The preset field selection rules are based on the needs of adapting to business scenarios and integrate a pre-trained feature selection model to construct dynamic selection logic. The core of this logic is to calculate the contribution value of each field through the model, quantitatively evaluate the importance of different fields to the representation of the business scenario, and set reasonable feature importance thresholds to automatically remove redundant fields. By acquiring a comprehensive logistics trajectory dataset, it ensures that no business scenario features are missed, providing sufficient data support for subsequent model processing. The dynamic field selection rules can replace the traditional hard-coded field extraction logic, automatically retaining key fields that contribute highly to the representation of the business scenario and efficiently removing low-importance redundant information, significantly improving the accuracy and efficiency of data processing.

[0020] 102. Perform text extraction processing on the logistics trajectory dataset to obtain multiple business scenario description texts; In this embodiment, text extraction processing of the logistics trajectory dataset first involves target field identification and batch extraction. This filters out initial business scenario fields such as scan type, annotation code, annotation content, user-defined fields, and time and region, comprehensively covering key dimensions representing business situations. Then, each extracted business scenario field undergoes validity verification and redundancy removal, filtering out fields that do not substantially contribute to the representation of the business scenario and retaining valid fields with scenario differentiation value. Finally, each valid business scenario field is structurally assembled and integrated according to business logic, transforming scattered field information into coherent, standardized, and accurate business scenario description text that reflects specific business situations. This achieves centralized and standardized representation of single trajectory business information. This processing flow replaces traditional hard-coded information extraction logic, possesses stronger scenario adaptability, can flexibly address the extraction needs of different business fields, facilitates the implementation of configurable solutions, and improves the standardization, accuracy, and efficiency of overall business processing.

[0021] 103. Call the pre-trained feature selection model to perform field filtering processing on each of the business scenario description texts based on the field filtering rules to obtain the target field set; In this embodiment, the pre-trained feature selection model is built on the XGBoost (eXtreme GradientBoosting) feature selection framework. Its architecture integrates the feature extraction capabilities of gradient boosting trees with the contribution quantification features of contribution values. The XGBoost framework enables deep mining of business field features and capture of non-linear associations, and then quantifies the contribution of each field to business scenario identification. The model training process uses a large-scale e-commerce logistics business scenario corpus as training data, covering full field information and corresponding business tags for various abnormal receipt and problem package scenarios. First, a basic feature selection model is built based on the XGBoost algorithm. Through iterative training, the tree structure and parameters are optimized to achieve preliminary learning of the association between business fields and scenarios. Then, the SHAP (SHapley Additive exPlanations) algorithm is introduced for secondary optimization of the model. Using labeled business scenario data as supervision, the field contribution quantification logic is calibrated, enabling the model to accurately output the SHAP value, i.e., the contribution value, for each field. This continues until the model's accuracy and stability in evaluating field contribution reach a preset standard, completing pre-training and solidifying it into a callable model.

[0022] In this embodiment, when performing field filtering on each business scenario description text, the field filtering rules serve as the core basis. Each business scenario description text is input into the feature selection model. The model uses the XGBoost framework to parse the features of each field in the text, and combines the business association logic learned during training to calculate the SHAP value corresponding to each field one by one. This quantifies the actual contribution of each field to the identification of the current business scenario. Then, based on the SHAP value threshold standard set in the field filtering rules, fields with SHAP values ​​higher than the threshold and that make a substantial contribution to scenario differentiation are retained, while redundant fields with SHAP values ​​lower than the threshold are removed. Finally, a target field set that accurately represents the core features of the business scenario is formed. The fusion architecture of the XGBoost framework and the SHAP algorithm balances the depth of feature extraction with the accuracy of contribution evaluation, effectively improving the accuracy and efficiency of field filtering while retaining the core features of the business scenario. In addition, the feature selection model can flexibly adapt to business scenario iterations without the need for repeated development of filtering logic, meeting the needs of configurable transformation, further improving the response efficiency of business adjustments, and enhancing the flexibility and professionalism of the overall solution.

[0023] 104. Call the pre-trained semantic model to perform semantic encoding and template matching on the target field set to obtain a template field mapping table; In this embodiment, the pre-trained semantic model consists of a template encoding module, a semantic encoding module, and a matching module. It is built upon the BERT-base-Chinese (Bidirectional Encoder Representations from Transformers-base-Chinese) pre-trained model. Each semantic encoding module possesses its general semantic representation capabilities, while the matching module establishes semantic association mappings between fields and templates. During the model training phase, a large-scale e-commerce logistics corpus serves as the core data source, covering various business scenario text samples, corresponding field information, and scenario tags. First, pre-training is performed based on the BERT-base-Chinese model to enhance general semantic feature extraction capabilities, solidifying the foundation of basic semantic understanding. Then, the model is fine-tuned by iteratively optimizing model parameters and calibrating the encoding module's ability to capture semantics of logistics industry-specific terms and business scenarios. Simultaneously, the semantic association judgment logic of the matching module is optimized until the model reaches preset standards in semantic encoding accuracy, template matching recall, and precision. Finally, pre-training is completed and the model is solidified into a semantic model.

[0024] In this embodiment, when calling the pre-trained semantic model to perform semantic encoding and template matching on the target field set, the system's built-in historical copywriting template library is first loaded. This library stores multiple copywriting samples related to e-commerce logistics scenarios, providing sufficient data support for the matching process. The template encoding module performs semantic encoding on each historical copywriting template in the library, converting the template text information into a quantifiable and calculable template semantic vector, while retaining the core semantic features of the template. Simultaneously, the semantic encoding module performs encoding processing on the target field set, extracting key semantic information of the fields and converting it into corresponding field semantic vectors, thus realizing the semantic representation of the field information. Subsequently, the matching module pairs all template semantic vectors with field semantic vectors one-to-one according to preset semantic association rules, establishing a precise semantic correspondence between fields and templates, and finally generating a template field mapping table. Compared with the traditional hard-coded matching logic, the model-based processing method can flexibly cope with the iterative adjustment of copywriting templates and fields, without the need to repeatedly develop matching logic, meeting the requirements of configuration-based transformation, effectively improving business response efficiency, and enhancing the semantic adaptability and intelligence level of the overall process.

[0025] 105. Based on the template field mapping table, perform similarity calculation and filtering to obtain the target template set; In this embodiment, the template field mapping table stores the semantic vectors of the templates encoded by the semantic model and the corresponding field semantic vectors. Based on this, a cosine similarity algorithm is used to calculate the similarity between each template semantic vector and its corresponding field semantic vector. The cosine value of the angle between the vectors quantifies the semantic correlation between the two, accurately capturing the semantic fit between the template and the business field. After the calculation is completed, the historical copy templates corresponding to the template semantic vectors are arranged in descending order of their respective similarity values, forming a standardized template ranking result, so that templates with higher semantic fit are ranked higher. Subsequently, a preset number of historical copy templates are extracted based on the ranking result, and finally, a target template set is formed. The templates ranked higher are selected to ensure the quality of adaptation to the business scenario. This processing method realizes the quantitative evaluation of semantic matching through the cosine similarity algorithm. Compared with the traditional manual setting of matching rules or hard-coded logic, it significantly improves the scientificity and accuracy of template selection and effectively avoids the adaptation bias caused by subjective factors. Meanwhile, the entire process is fully automated, requiring no manual intervention in development and adjustment. It can flexibly respond to changes in scenarios such as template library updates and business field iterations, significantly improving template selection efficiency, shortening business adjustment cycles, and enhancing the overall solution's intelligence and scenario adaptability.

[0026] 106. Perform copywriting generation processing on the target template set to obtain a target copywriting set, and determine the initial online copywriting based on the target copywriting set.

[0027] In this embodiment, during the copy generation process, a parsing algorithm is first used to parse each template in the target template set, extracting all the EL placeholders contained therein to construct a complete placeholder set, laying the foundation for subsequent differentiated processing. Then, each extracted EL placeholder is classified according to its business logic attributes, accurately distinguishing between business EL placeholders and non-business EL placeholders. Business placeholders mainly involve information requiring standardized processing, such as addresses and business statuses, while non-business placeholders are basic information such as user names and order IDs that do not require additional logical processing. For the classified business EL placeholders, further standardization processing is carried out to correct deviations in the original data format, achieving standardized and unified information and ensuring the rigor and consistency of the copywriting. After completing the placeholder classification and standardization, a preset function mapping library is obtained. The mapping library is designed based on a strategy pattern and can automatically match the corresponding value function according to the specific type of the placeholder. The function extracts the corresponding information from the previously screened core trajectory data, completing the basic replacement of the placeholder. Based on the pre-defined mapping rules in the library, all non-business EL placeholders and standardized business EL placeholders are matched with corresponding value retrieval functions. These functions extract precise information from the previously screened core trajectory data, replacing each placeholder one by one. This process ultimately generates a batch of standardized, highly adaptable text, forming the target text set. The target text set provides multiple candidate texts. Selecting any candidate text from this set as the initial deployment text ensures maximum compatibility between the initial deployment text and actual business scenarios and platform specifications. Furthermore, the fully automated processing replaces traditional hard-coded logic, enhancing the intelligence and scalability of the process and guaranteeing the stability and professionalism of the logistics trajectory push service.

[0028] Please see Figure 2 In this embodiment of the invention, the text extraction processing of the logistics trajectory dataset to obtain multiple business scenario description texts includes: 201. The target field of the logistics trajectory dataset is identified and extracted using a feature extraction algorithm to obtain multiple initial business scenario fields; 202. Obtain preset field validation rules, and perform redundant field removal processing on each of the initial business scenario fields based on the field validation rules to obtain multiple valid business scenario fields; 203. A single trajectory field attribution association algorithm is used to perform structured splicing processing on multiple effective business scenario fields to obtain multiple business scenario description texts.

[0029] In this embodiment, a feature extraction algorithm is used to perform target field identification and extraction processing on the logistics trajectory dataset. This process involves batch mining core information representing business scenarios from the full volume of logistics trajectory data. A regular expression matching algorithm can be selected to identify scan type, note code, note content, user-defined fields, and structured basic fields such as time and region. At the same time, it adapts to the fixed format features and semi-structured attributes of logistics data to extract initial business scenario fields, comprehensively covering key identifier dimensions that reflect business situations, and ensuring that no core business information is omitted.

[0030] In this embodiment, the field validation rules are comprehensively set from multiple dimensions based on the actual business scenario requirements of e-commerce logistics, statistical analysis of large-scale logistics trajectory historical data, the adaptability of integrated quantitative filtering algorithms, and the standardized requirements of e-commerce platform trajectory push. These rules define the criteria for field validity judgment, algorithm application logic, and the boundaries for redundant field identification, providing a feasible execution framework for the accurate selection of initial business scenario fields. Based on these rules, redundancy removal is performed on each initial business scenario field. The validation rules can integrate multiple quantitative filtering algorithms, employing the SHAP algorithm, combined with variance thresholding and mutual information methods for auxiliary validation. The SHAP algorithm calculates the contribution of each field to business scenario identification, while the variance thresholding method removes invalid fields with extremely small variances, and the mutual information method filters fields strongly correlated with core business scenarios. This accurately filters out redundant fields such as time and region that do not substantially contribute to scenario differentiation, retaining high-value and effective business scenario fields such as scan type, note code, and note content, thus achieving quantification and precision in field selection.

[0031] In this embodiment, a single-trajectory field attribution association algorithm is used to perform structured splicing processing on the effective business scenario fields. Greedy algorithms, connected component analysis algorithms in graph theory, or sequence splicing algorithms based on business semantics can be selected. These algorithms can integrate and splice scattered effective fields according to the logistics business logic and field association relationship, transforming independent business identification information into coherent, standardized business scenario description text that can accurately represent specific business situations. This achieves centralized and standardized presentation of single-trajectory business information, providing a unified format text carrier for subsequent feature screening model input.

[0032] Please see Figure 3 In this embodiment of the invention, the feature selection model includes a field structured splitting module, a contribution calculation module, and a filtering module, which are sequentially connected. The pre-trained feature selection model is invoked to perform field filtering processing on each business scenario description text based on the field filtering rules, resulting in a target field set, including: 301. Based on the field structure splitting module, each of the business scenario description texts is split into multiple split fields. 302. Based on the contribution calculation module, the contribution value is calculated for each of the split fields to obtain the contribution value corresponding to each of the split fields; 303. Based on the filtering module and the field filtering rules, the multiple split fields are filtered according to the contribution value to obtain the target field set.

[0033] In this embodiment, the feature selection model consists of a field structured splitting module, a contribution calculation module, and a filtering module, sequentially connected according to data flow logic. The entire model is built upon a pre-trained feature importance model and integrates the XGBoost feature selection framework to achieve accurate field filtering and core dimension extraction of business scenario description text. During processing, the field structured splitting module first performs field splitting on each business scenario description text. This module can identify preset field concatenation and separation rules in the text, restoring the integrated, coherent text into independent split fields. Specifically, it extracts original field dimensions such as scan type, note code, note content, user-defined fields, time, and region from the concatenated business scenario description text, ensuring the independent representation of each field and providing a clear analytical unit for subsequent contribution value calculation. Subsequently, the contribution calculation module calculates the contribution value for each split field. This module leverages the non-linear feature mining capabilities of the XGBoost framework to capture the inherent correlation between each split field and business scenario identification. It accurately quantifies the contribution of each field to the business scenario representation using the SHAP algorithm, ultimately outputting a contribution value corresponding to each split field—the SHAP value—to objectively reflect the actual value of different fields in distinguishing business scenarios. Finally, the filtering module, combined with field filtering rules, performs filtering based on the contribution value of each split field. The field filtering rules pre-define clear contribution value quantification threshold standards, and the filtering module strictly follows these rules, judging each split field one by one, automatically retaining fields with contribution values ​​higher than the threshold, and simultaneously removing redundant fields with contribution values ​​lower than the threshold. This ultimately forms a target field set containing only valid fields, achieving preliminary optimization and precise purification of data dimensions. The contribution calculation module integrates the advantages of XGBoost and SHAP algorithms, ensuring both the accuracy and interpretability of contribution value calculation and adapting to the business characteristics of the logistics industry using a pre-trained model, making it more scientific than traditional manual judgment. The combination of the filtering module and field filtering rules enables the automatic removal of redundant fields, replacing the traditional hard-coded filtering logic. The rule thresholds can be flexibly adjusted to adapt to business scenario iterations.

[0034] Please see Figure 4In this embodiment of the invention, the semantic model includes a template encoding module, a semantic encoding module, and a matching module, wherein the template encoding module and the semantic encoding module are respectively connected to the matching module; the step of calling the pre-trained semantic model to perform semantic encoding and template matching processing on the target field set to obtain a template field mapping table includes: 401. Obtain a preset historical copywriting template library, which includes multiple historical copywriting templates; 402. Based on the template encoding module, perform template encoding processing on each of the historical text templates to obtain multiple template semantic vectors; 403. Based on the semantic encoding module, perform field encoding processing on the target field set to obtain multiple field semantic vectors; 404. Based on the matching module, perform one-to-one pairing processing on multiple template semantic vectors and multiple field semantic vectors to establish the association relationship between each template semantic vector and the corresponding field semantic vector, and obtain the template field mapping table.

[0035] In this embodiment, the semantic model is based on a pre-trained and fine-tuned BERT architecture, consisting of a template encoding module, a semantic encoding module, and a matching module. The template encoding module and the semantic encoding module are connected in parallel to the matching module, enabling collaborative processing of semantic encoding and association mapping between templates and fields. The model is built on a BERT-base-Chinese pre-trained model and fine-tuned and optimized using a large-scale e-commerce logistics corpus, specifically adapted to semantic association tasks between scenarios and templates, possessing accurate industry semantic representation capabilities. When calling this model for processing, the system first retrieves a pre-set historical text template library. This library stores multiple text samples related to e-commerce logistics scenarios, covering various abnormal receipt and problem package handling scenarios, providing sufficient and high-quality data support for semantic matching.

[0036] Subsequently, the template encoding module performs template encoding processing on each historical copy template in the library. Relying on this module, core semantic features are extracted, transforming the text-based historical copy templates into quantifiable semantic vectors. This preserves the semantic association between the templates and the business scenario, ensuring that semantic differences between different templates can be accurately represented through vector dimensions. Simultaneously, the semantic encoding module performs field encoding processing on the target field set, which consists of core business fields selected in the previous stage. Using the same semantic encoding logic, the module transforms the scattered field information into corresponding field semantic vectors, achieving representation of fields and templates within the same semantic space, laying the foundation for subsequent pairing processing. Finally, the matching module extracts the corresponding business feature dimensions from the template semantic vector and the field semantic vector. It precisely aligns the feature dimensions representing a core business attribute in the field semantic vector with the feature dimensions representing the same business attribute in the template semantic vector. This achieves one-to-one pairing of all template semantic vectors and field semantic vectors. For example, the feature dimension corresponding to "remark code = 001" in the field semantic vector is only paired with the feature dimension involving the "user rejects" remark code in the template semantic vector. This establishes the association between each field semantic vector and the corresponding template semantic vector, forming a structured template field mapping table. The generation of the template field mapping table realizes the structured accumulation of semantic association between fields and templates, replacing the traditional hard-coded semantic matching logic. This not only lays a solid foundation for subsequent similarity measurement calculations but also has the characteristics of flexible adaptation to business iterations. It can respond to new scenarios and new field requirements through template library updates and model fine-tuning, which fits the core goal of copywriting configuration transformation, significantly improves the accuracy and efficiency of semantic matching, and strengthens the intelligence and scalability of the overall process.

[0037] Please see Figure 5 In this embodiment of the invention, the step of performing similarity calculation and filtering based on the template field mapping table to obtain the target template set includes: 501. Using the cosine similarity algorithm, calculate the cosine similarity between each template semantic vector and the corresponding field semantic vector based on the template field mapping table to obtain the similarity with each historical copy template; 502. Sort all the historical copywriting templates from high to low according to their corresponding similarity to obtain the template sorting results; 503. Obtain the preset number of templates to be filtered, and obtain historical copywriting templates that match the number of templates to be filtered based on the template sorting results, thereby obtaining the target template set.

[0038] In this embodiment, when performing similarity calculation and template selection based on the template field mapping table, a cosine similarity algorithm is used. This algorithm quantifies the semantic correlation between the template semantic vector and the corresponding field semantic vector by calculating the cosine value of the angle between them in the vector space, accurately capturing the semantic fit between the field and the template. Euclidean distance, Pearson correlation coefficient, or Manhattan distance algorithms can also be used as alternatives. These algorithms can all effectively quantify the similarity between semantic vectors and can be flexibly adapted according to the vector characteristics of logistics data. The cosine similarity algorithm is used to calculate the similarity value for each historical copy template in the template field mapping table, comprehensively representing the semantic fit between each template and the core business field. After calculation, all historical copy templates are sorted in descending order of their corresponding similarity values, forming a standardized template sorting result. This ensures that templates with higher semantic fit to the business scenario are ranked higher, making subsequent selection more targeted. The preset template selection quantity is then obtained. This quantity is determined comprehensively based on the actual business needs of e-commerce logistics, the implementation effect of historical template matching, the candidate requirements for subsequent copywriting generation, and the adaptation requirements of platform trajectory push. In typical logistics business scenarios, the quantity is preset to 3 based on historical adaptation effects, which provides sufficient candidate templates for subsequent copywriting generation while avoiding excessive templates that would increase the redundancy of subsequent processing. Based on this template selection quantity, templates with matching quantities are extracted from the template ranking results, prioritizing high-similarity templates ranked higher, and finally forming a target template set to provide high-quality candidate support for subsequent copywriting generation. The entire process is fully automated, replacing the traditional hard-coded template matching and selection logic. The template selection quantity can be flexibly adjusted according to business scenario iterations and template library updates without modifying the underlying code, meeting the core requirements of copywriting and status code configuration transformation, significantly improving the efficiency of template selection and the response speed of business adjustments. At the same time, the target template set formed based on high similarity selection ensures a high degree of adaptation between subsequent copywriting generation and the actual business scenarios of e-commerce logistics.

[0039] Please see Figure 6 In this embodiment of the invention, the step of performing copywriting generation processing on the target template set to obtain the target copywriting set includes: 601. Parse the target template set using a parsing algorithm to extract multiple EL placeholders; 602. Call the pre-trained BERT model to classify the business logic attributes of each EL placeholder, to obtain multiple business-type EL placeholders and multiple non-business-type EL placeholders, and perform standardization processing on each business-type EL placeholder to obtain standardized EL placeholders; 603. Obtain a preset function mapping library, and based on the function mapping library, match and retrieve values ​​for all the non-business EL placeholders and all the standardized EL placeholders to obtain the target copy set.

[0040] In this embodiment, a parsing algorithm is used to fully parse the target template set to extract EL placeholders. The parsing algorithm can be based on regular expressions, ANTLR (Another Tool for Language Recognition) syntax parsing algorithm, or XPath (XML Path Language) parsing algorithm. These algorithms can accurately identify the syntactic features and boundaries of EL placeholders in the template, and efficiently extract all EL placeholders such as user name, order ID, address, and contact number, providing a complete set of placeholders for subsequent differentiated processing. After extraction, the pre-defined BERT model is invoked to classify the business logic attributes of each EL placeholder. This model is built on BERT-base-Chinese and adopts a 12-layer Transformer encoder architecture, including a multi-head self-attention layer, a feedforward neural network layer, a layer normalization module, and a residual connection structure. It is customized and fine-tuned for the needs of e-commerce logistics scenarios, adding a business attribute classification head and an information standardization processing layer to specifically adapt to the attribute determination and business information standardization processing of EL placeholders. The model is trained on a large-scale e-commerce logistics annotation corpus, which covers a full range of EL placeholder samples for various abnormal receipt and problem package business scenarios. It includes placeholder business logic attribute annotation data, original samples and standardized annotation samples of various business information. The model is fine-tuned by combining supervised learning methods and iteratively optimizing the model parameters until the preset standards are met in terms of attribute classification accuracy and information standardization processing precision. After calculation, the BERT model accurately classifies EL placeholders into business-type and non-business-type. Among them, addresses and other information involving business logic processing are business-type EL placeholders, while user names, order IDs, and other information that do not require additional logic processing are non-business-type EL placeholders. At the same time, the BERT model performs standardization processing on each business-type EL placeholder, corrects the format deviations of the original information, achieves standardized and unified information representation, and obtains standardized EL placeholders.Subsequently, a pre-defined function mapping library is retrieved. This library is designed based on the actual business scenario requirements of e-commerce logistics, the type characteristics of EL placeholders, the field extraction logic of core logistics trajectory data, and industry-standard function design specifications. It also utilizes a strategy pattern for its architecture design, pre-setting corresponding value-selection function matching rules according to the information extraction requirements of different types of placeholders. The library covers precise mapping relationships between various placeholders and value-selection functions. Based on this mapping library, the system automatically matches corresponding value-selection functions for all non-business EL placeholders and standardized business EL placeholders. The value-selection functions extract precise information from the pre-screened core logistics trajectory data to complete placeholder replacement. Placeholders are essentially semantic spaces reserved in the copy template to carry specific business information. The copy template itself has a standardized expression framework based on business scenarios, lacking only personalized information corresponding to actual logistics trajectories. Placeholder replacement fills these gaps, transforming the general template into complete copy adapted to specific business scenarios. By flexibly adapting to multiple parsing algorithms, accurate and efficient extraction of EL placeholders is achieved. Compared with traditional fixed parsing logic, it is more adaptable to the diverse template formats in the template library.

[0041] Please see Figure 7 In this embodiment of the invention, after determining the initial online copy based on the target copy set, the method further includes: 701. Obtain the historical business push dataset, actual business scenarios, and real-time business push data of the initial launch copy; 702. Using a semantic similarity algorithm, calculate the matching degree between the actual business scenario and the initial online copy based on the business push data; 703. Call the pre-trained LSTM model to calculate the dynamic matching degree threshold based on the historical business push dataset; 704. Compare the matching degree with the dynamic matching degree threshold; 705. When the matching degree is less than the dynamic matching degree threshold, a new online copy is determined based on the target copy set.

[0042] In this embodiment, after determining the initial online copy based on the target copy set, the process is further extended to a closed-loop management of dynamic monitoring, threshold calculation, deviation judgment, and optimization adjustment, taking into account both business stability and adaptation accuracy. First, the system synchronously acquires three types of core data, including historical business push datasets, actual business scenario information, and real-time business push data of the initial online copy. Among them, the historical business push dataset includes a time series of copy push failure rates over the past 30 days, providing time series data support for subsequent dynamic threshold calculation. The real-time business push data reflects the actual effect of the initial online copy. The three work together to lay a solid data foundation for matching degree evaluation and threshold setting.

[0043] Subsequently, a semantic similarity algorithm, specifically the BERT semantic similarity algorithm, is employed. Based on real-time business push data, core semantic features of the initial online copy and the actual business scenario are extracted, and the semantic matching degree between the two is calculated to quantitatively evaluate the adaptability of the copy to the current business scenario. Simultaneously, a pre-trained LSTM model is invoked to calculate a dynamic matching degree threshold. The LSTM model, equipped with multiple LSTM units, has been pre-trained and optimized using large-scale e-commerce logistics business push time-series data, possessing extremely strong time-series trend learning and prediction capabilities. Using historical business push datasets as input, the LSTM model learns the fluctuation patterns of historical data, mines the intrinsic correlation between failure rate and matching degree, and predicts the business push failure rate trend over a future period. Combined with a preset basic threshold and prediction deviation, a dynamic matching degree threshold is calculated, which is more adaptable to the dynamic changes in business scenarios compared to a fixed threshold. The system precisely compares the calculated real-time matching score with the dynamic matching score threshold. If the real-time matching score is lower than the dynamic matching score threshold, an early warning mechanism is immediately triggered. Based on the target copy set, a new online copy is determined. Simultaneously, based on the semantic similarity analysis results, targeted optimization suggestions are pushed, recommending the template with the best fit in the target copy set to assist in quickly completing configuration adjustments. During this process, the system synchronously generates a new configuration version, recording a unique version number, changed content, and a list of key dimensions, achieving full-process version traceability and providing a basis for subsequent review and secondary optimization.

[0044] The platform trajectory data processing method in the embodiments of the present invention has been described above. The platform trajectory data processing apparatus in the embodiments of the present invention will be described below. Please refer to [link / reference]. Figure 8 One embodiment of the platform trajectory data processing device in this invention includes: Data acquisition module 801: Used to acquire logistics trajectory datasets and preset field filtering rules; Text extraction module 802: used to perform text extraction processing on the logistics trajectory dataset to obtain multiple business scenario description texts; Field filtering module 803: used to call a pre-trained feature selection model to perform field filtering processing on each of the business scenario description texts based on the field filtering rules, so as to obtain a target field set; Matching module 804: used to call a pre-trained semantic model to perform semantic encoding and template matching processing on the target field set to obtain a template field mapping table; Similarity calculation module 805: used to perform similarity calculation and filtering based on the template field mapping table to obtain the target template set; Copy generation module 806: used to perform copy generation processing on the target template set to obtain a target copy set, and determine the initial online copy based on the target copy set.

[0045] Based on the same ideas as the methods in the above embodiments, the apparatus provided in this application can implement the methods in the above embodiments.

[0046] above Figure 8 The platform trajectory data processing device in this embodiment of the invention will be described in detail from the perspective of modular functional entities. The platform trajectory data processing device in this embodiment of the invention will be described in detail from the perspective of hardware processing.

[0047] Figure 9 This is a schematic diagram of the structure of a platform trajectory data processing device 900 provided in an embodiment of the present invention. The platform trajectory data processing device 900 can vary significantly due to different configurations or performance. It may include one or more central processing units (CPUs) 910 (e.g., one or more processors) and a memory 920, and one or more storage media 930 (e.g., one or more mass storage devices) for storing application programs 933 or data 932. The memory 920 and storage media 930 can be temporary or persistent storage. The program stored in the storage media 930 may include one or more modules (not shown in the diagram), each module may include a series of instruction operations on the platform trajectory data processing device 900. Furthermore, the processor 910 may be configured to communicate with the storage media 930 and execute the series of instruction operations in the storage media 930 on the platform trajectory data processing device 900 to implement the steps of the platform trajectory data processing method provided in the above-described method embodiments.

[0048] The platform trajectory data processing device 900 may also include one or more power supplies 940, one or more wired or wireless network interfaces 950, one or more input / output interfaces 960, and / or one or more operating systems 931, such as Windows Server, Mac OS X, Unix, Linux, FreeBSD, etc. Those skilled in the art will understand that... Figure 9 The structure of the platform trajectory data processing device shown does not constitute a limitation on the platform trajectory data processing device. It may include more or fewer components than shown, or combine certain components, or have different component arrangements.

[0049] The present invention also provides a computer-readable storage medium, which can be a non-volatile computer-readable storage medium or a volatile computer-readable storage medium, wherein the computer-readable storage medium stores instructions that, when executed on a computer, cause the computer to perform the steps of the platform trajectory data processing method.

[0050] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working process of the system, device, or unit described above can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0051] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0052] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for processing platform trajectory data, characterized in that, include: Obtain the logistics trajectory dataset and preset field filtering rules; Text extraction processing is performed on the logistics trajectory dataset to obtain multiple business scenario description texts; The pre-trained feature selection model is invoked to perform field filtering processing on each of the business scenario description texts based on the field filtering rules, thereby obtaining the target field set; The pre-trained semantic model is invoked to perform semantic encoding and template matching on the target field set to obtain a template field mapping table; Similarity calculation and filtering are performed based on the template field mapping table to obtain the target template set; The target template set is processed to generate copy, resulting in a target copy set, and the initial online copy is determined based on the target copy set.

2. The platform trajectory data processing method according to claim 1, characterized in that, The text extraction process of the logistics trajectory dataset yields multiple business scenario description texts, including: The logistics trajectory dataset is processed by a feature extraction algorithm to identify and extract target fields, resulting in multiple initial business scenario fields. Obtain preset field validation rules, and perform redundant field removal processing on each of the initial business scenario fields based on the field validation rules to obtain multiple valid business scenario fields; A single trajectory field attribution association algorithm is used to perform structured splicing processing on multiple effective business scenario fields to obtain multiple business scenario description texts.

3. The platform trajectory data processing method according to claim 1, characterized in that, The feature selection model includes a field structured splitting module, a contribution calculation module, and a filtering module, which are sequentially connected. The pre-trained feature selection model is invoked to perform field filtering processing on each business scenario description text based on the field filtering rules, resulting in a target field set, including: Based on the field structured splitting module, each of the business scenario description texts is split into multiple split fields. Based on the contribution calculation module, the contribution value is calculated for each of the split fields to obtain the contribution value corresponding to each of the split fields. Based on the filtering module and the field filtering rules, the multiple split fields are filtered according to the contribution value to obtain the target field set.

4. The platform trajectory data processing method according to claim 1, characterized in that, The semantic model includes a template encoding module, a semantic encoding module, and a matching module, wherein the template encoding module and the semantic encoding module are respectively connected to the matching module; the step of calling the pre-trained semantic model to perform semantic encoding and template matching processing on the target field set to obtain a template field mapping table includes: Obtain a preset historical copywriting template library, which includes multiple historical copywriting templates; Based on the template encoding module, each of the historical text templates is subjected to template encoding processing to obtain multiple template semantic vectors; Based on the semantic encoding module, the target field set is encoded to obtain multiple field semantic vectors; Based on the matching module, a one-to-one pairing process is performed on multiple template semantic vectors and multiple field semantic vectors to establish a mapping relationship between each template semantic vector and its corresponding field semantic vector, thereby generating the template field mapping table.

5. The platform trajectory data processing method according to claim 4, characterized in that, The similarity calculation and filtering process based on the template field mapping table yields a target template set, including: The cosine similarity algorithm is used to calculate the cosine similarity between each template semantic vector and the corresponding field semantic vector based on the template field mapping table, so as to obtain the similarity with each historical copy template. Sort all the historical copywriting templates from highest to lowest similarity to obtain the template ranking results; Obtain the preset number of templates to be filtered, and based on the template sorting result, obtain historical copywriting templates that match the number of templates to be filtered, to obtain the target template set.

6. The platform trajectory data processing method according to claim 1, characterized in that, The step of performing copywriting generation processing on the target template set to obtain the target copywriting set includes: The target template set is parsed using a parsing algorithm to extract multiple EL placeholders; The pre-trained BERT model is invoked to classify the business logic attributes of each EL placeholder, resulting in multiple business-type EL placeholders and multiple non-business-type EL placeholders. Each business-type EL placeholder is then standardized to obtain standardized EL placeholders. Obtain a preset function mapping library, and based on the function mapping library, match and retrieve values ​​for all the non-business EL placeholders and all the standardized EL placeholders to obtain the target copy set.

7. The platform trajectory data processing method according to claim 1, characterized in that, After determining the initial online copy based on the target copy set, the process also includes: Obtain real-time business push data, including historical business push datasets, actual business scenarios, and the initial launch copy; A semantic similarity algorithm is used to calculate the matching degree between the actual business scenario and the initial online copy based on the business push data; The pre-trained LSTM model is invoked to calculate the dynamic matching degree threshold based on the historical business push dataset; The matching degree is compared with the dynamic matching degree threshold; When the matching degree is less than the dynamic matching degree threshold, a new online copy is determined based on the target copy set.

8. A platform trajectory data processing device, characterized in that, include: Data acquisition module: used to acquire logistics trajectory datasets and preset field filtering rules; Text extraction module: used to perform text extraction processing on the logistics trajectory dataset to obtain multiple business scenario description texts; Field filtering module: used to call the pre-trained feature selection model to perform field filtering processing on each of the business scenario description texts based on the field filtering rules, so as to obtain the target field set; Matching module: used to call a pre-trained semantic model to perform semantic encoding and template matching on the target field set to obtain a template field mapping table; Similarity calculation module: used to perform similarity calculation and filtering based on the template field mapping table to obtain the target template set; Copy generation module: used to process the target template set to generate copy, obtain the target copy set, and determine the initial online copy based on the target copy set.

9. A platform trajectory data processing device, characterized in that, The platform trajectory data processing device includes: a memory and at least one processor, wherein the memory stores instructions; At least one of the processors invokes the instructions in the memory to cause the platform trajectory data processing device to perform the steps of the platform trajectory data processing method as claimed in any one of claims 1-7.

10. A computer-readable storage medium storing instructions thereon, characterized in that, When the instructions are executed by the processor, they implement the various steps of the platform trajectory data processing method as described in any one of claims 1-7.

Images