A task information processing method based on platform development

Abstract

The application is a task information processing method based on platform development, and relates to the technical field of software platform data processing, which comprises the following steps: splicing a task field structure description vector of each task field and a corresponding value range behavior feature vector along a feature dimension to generate a fusion feature vector.In the application, the field structure description vector and the value range behavior feature vector are fused and then input into a semantic coding network to generate a semantic fingerprint vector, thereby solving the technical problem of semantic distortion caused by only relying on field name text matching or artificial pre-defined mapping table, and achieving the technical effect of accurately reflecting the real semantic relationship of the field based on the joint representation of field structure definition and actual use behavior.

Description

Technical Field

[0001] This invention relates to the field of software platform data processing technology, and in particular to a task information processing method based on platform-based development. Background Technology

[0002] In a platform-based software development environment, multiple business applications are independently developed and deployed on the same development platform, generating a large amount of task information during operation. Because these applications are developed independently by different teams, the data structures and field semantics used in their task information differ significantly. For example, a field named "priority" may have different value ranges and meanings in different applications, and a field named "deadline" may refer to an absolute time point or a relative duration. When the platform needs to perform unified retrieval, aggregation analysis, or cross-application flow of task information from various applications, it is necessary to map the heterogeneous task fields of each application into a unified data model.

[0003] Existing technologies typically rely on pre-configured field mapping tables to complete the mapping process. However, due to the continuous iteration and updates of business applications on the platform, and the constant addition of new applications, the field definitions of task information change frequently, and there are implicit semantic differences between fields. Manually maintained field mapping tables cannot promptly cover newly added or changed field definitions, leading to semantic distortion of cross-application task information after mapping, inconsistencies in aggregated task data, and consequently affecting the accuracy of platform-level task scheduling and analysis. Summary of the Invention

[0004] The purpose of this invention is to provide a task information processing method based on platform-based development in order to solve the above-mentioned problems.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: A task information processing method based on platform development includes: Obtain the task data model definition files registered by each business application on the platform, parse the task data model definition files of each business application, and for each task field, extract the field name, data type, value constraints, and hierarchical relationship between the field and other fields in the same application, encode the multidimensional attribute information into a numerical vector, and generate a task field structure description vector set for each business application. Collect task instance data generated by each business application within a preset time window. For each task field, calculate the distribution characteristics of its filled value, the co-occurrence frequency with other fields in the same application, and the correlation coefficient between fields, and generate the value range behavior feature vector of each task field. The task field structure description vector and the corresponding value domain behavior feature vector of each task field are concatenated along the feature dimension to generate a fused feature vector. After being arranged in the hierarchical traversal order of the fields in the data model, the fused feature vector is input into the pre-trained semantic encoding network to output the semantic fingerprint vector of each field. Perform the same feature extraction and encoding processing on the standard field set of the platform's unified task model to generate reference semantic fingerprint vectors for each standard field; calculate the cosine similarity between the semantic fingerprint vectors of each business application field and the reference semantic fingerprint vectors of each standard field to generate a field similarity matrix; apply the Hungarian algorithm to the field similarity matrix to obtain the optimal matching scheme with the largest global similarity sum, and generate an automatic mapping relationship from each task field to the standard field. Based on the automatic mapping relationship, for field pairs with mapping confidence higher than or equal to a preset threshold, the value range behavior feature vectors of the source field and the target standard field are obtained, the difference features between the two are calculated, and the parameters of the value range transformation function are generated through the parameter generation network based on the difference features. The source field fill value is input into the value range transformation function to obtain the normalized value, and the normalized task information is written into the platform's unified task information index to generate a cross-application unified task information view.

[0006] Preferably, the semantic encoding network adopts a Transformer encoder structure and is pre-trained on a large-scale multi-source metadata corpus using a contrastive learning approach. Semantically equivalent field pairs from different platforms are used as positive sample pairs, and semantically unrelated field pairs are used as negative sample pairs. The contrastive loss function is used for training, so that the semantic fingerprint vectors of semantically equivalent field pairs are close to each other in the vector space, while the semantic fingerprint vectors of semantically unrelated field pairs are far apart.

[0007] Preferably, before applying the pre-trained semantic coding network to the current platform, the method further includes: collecting manually confirmed field mapping relationships in the current platform as labeled data, using the fused feature vector pairs of the source field and the target standard field in the mapping relationship as training samples, and using minimizing the Euclidean distance between their semantic fingerprint vectors as the training objective to fine-tune the pre-trained semantic coding network and generate a semantic coding network adapted to the semantic environment of the current platform.

[0008] Preferably, the parameter generation network is a multilayer perceptron, whose input is the element-wise difference vector between the value domain behavior feature vector of the source field and the value domain behavior feature vector of the target standard field, and the vector obtained by concatenating the two vectors together; for numerical fields, the output layer outputs scaling factor and offset, and maps the source field fill value to normalized value through linear transformation; For enumeration type fields, the output layer outputs the matching score between each pair of enumeration values, generates an enumeration value correspondence table, and replaces the source enumeration value with the standard enumeration value according to the correspondence table; the parameter generation network is trained with the mean squared error loss function between the source field filling value of the known standard mapping field pair after transformation and the target standard field filling value as the training target.

[0009] Preferably, the preset time window is set using a sliding time window method. Each time data is collected, task instance data within a fixed time period prior to the current moment is taken for statistical analysis, so that the value domain behavior feature vector automatically reflects the evolution trend of field usage semantics over time.

[0010] Preferably, when the number of task fields in a certain business application is inconsistent with the number of standard fields, the field similarity matrix is ​​padded before the Hungarian algorithm is applied: when the number of application fields is greater than the number of standard fields, a virtual standard field column is added, and the similarity value in the virtual column is set to zero. When the number of application fields is less than the number of standard fields, a virtual application field row is added, and the similarity value in the virtual row is set to zero; the application field that matches the virtual field in the matching results is marked as no matching field, and the standard field that matches the virtual field is marked as an uncovered standard field.

[0011] Preferably, when writing the normalized task information into the platform's unified task information index, the mapping path and transformation parameters of each task information are recorded synchronously as traceability metadata; the mapping path includes the source application identifier, source field identifier, target standard field identifier, and mapping confidence value; the transformation parameters include the type identifier and specific parameter value of the domain transformation function.

[0012] Preferably, the method further includes: The task data model definition files registered by each business application are continuously monitored. Changes are detected by comparing the version identifier or content summary value of the task data model definition files. When a change is detected in the task data model definition file of a certain business application, the semantic fingerprint vector calculation, mapping relationship calculation and value domain alignment conversion are re-executed for the changed fields and their related fields in the data model hierarchy. The updated normalized task information is written into the platform's unified task information index. Fields that have not changed retain their original semantic fingerprint vectors and mapping relationships.

[0013] Preferably, the distribution features include the mean, variance, skewness, kurtosis, proportion of null values, and proportion of unique values ​​of the filled values; for enumeration type fields, the distribution features also include the probability distribution of the occurrence of each enumeration value; for text type fields, the distribution features include the statistical distribution of text length and the word frequency statistics of high-frequency terms; the co-occurrence frequency is the frequency at which the current field and other fields simultaneously have valid filled values ​​in the same task instance record; the correlation coefficient is measured by the Pearson correlation coefficient between numerical type fields and by the Cramer V coefficient between categorical type fields.

[0014] In summary, due to the adoption of the above technical solution, the beneficial effects of the present invention are: 1. This invention solves the technical problem of semantic distortion caused by relying solely on field name text matching or manually predefined mapping tables by fusing field structure description vectors and value domain behavior feature vectors and inputting them into a semantic coding network to generate semantic fingerprint vectors. It achieves the technical effect of accurately reflecting the true semantic relationship of fields based on the joint representation of field structure definition and actual usage behavior.

[0015] 2. This invention solves the technical problem of potential one-to-many mapping conflicts caused by greedy matching strategies by applying the Hungarian algorithm to solve the optimal matching on the field similarity matrix, and achieves the technical effect of ensuring a one-to-one mapping constraint between application fields and standard fields; by generating a value domain transformation function based on the difference parameterization of value domain behavioral feature vectors, it solves the technical problem of data incomparability caused by inconsistent value ranges of the same semantic field between heterogeneous applications, and achieves the technical effect of adaptively aligning the source field values ​​to the standard field value specification; by continuously monitoring the task data model definition file and incrementally triggering the mapping relationship update, it solves the technical problem of normalization processing failure caused by the lag of the mapping table in the application continuous iteration scenario, and achieves the technical effect of automatically synchronizing and updating the mapping relationship when the business application fields change. Attached Figure Description

[0016] Further details, features, and advantages of this application are disclosed in the following description of exemplary embodiments in conjunction with the accompanying drawings, in which: Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0017] Several embodiments of this application will now be described in more detail with reference to the accompanying drawings to enable those skilled in the art to implement this application. This application may be embodied in many different forms and for various purposes and should not be limited to the embodiments set forth herein. These embodiments are provided to make this application thorough and complete, and to fully convey the scope of this application to those skilled in the art. The embodiments described do not limit this application.

[0018] Unless otherwise defined, all terms used herein (including technical and scientific terms) shall have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains. It will be further understood that terms such as those defined in commonly used dictionaries shall be interpreted as having a meaning consistent with their meaning in the relevant field and / or the context of this specification, and shall not be interpreted in an idealized or overly formal sense unless expressly defined herein.

[0019] Example 1 Its specific implementation method is combined with the appendix Figure 1 Please provide a detailed explanation.

[0020] According to an embodiment of this implementation, a method for normalizing heterogeneous task information in a platform-based environment is provided. The execution entity of this normalization method is a data processing server deployed on the platform server. This data processing server is connected to the data interfaces of various business applications on the platform and maintains a standard field set of the platform's unified task model. The normalization method includes the following steps: Step 1: Parse the task data model definitions of each business application and generate a task field structure description vector set. Obtain the task data model definition files registered by each business application on the platform. Parse the task data model definition files of each business application and extract the set of task fields contained therein. For each task field, extract the field name, data type, value constraints, and hierarchical relationship between the task field and other fields within the same application. Encode the extracted multidimensional attribute information into a numerical vector to generate a task field structure description vector set for each business application.

[0021] It should be noted that the above field name encoding refers to word embedding of the text content of the field name, mapping it to a fixed-length numerical representation. The above data type encoding refers to one-hot encoding or embedding encoding of the field's data type information (such as integer, string, timestamp, enumeration, etc.). The above value constraint encoding refers to numerically representing the constraints declared in the field definition, such as the upper and lower bounds of the value range, the length of the enumeration value list, and whether null values ​​are allowed. The above hierarchical relationship encoding refers to encoding the structural information of the field in the data model's tree structure, such as its hierarchical depth, parent node type, and the number of sibling fields at the same level, into a numerical vector.

[0022] It should be noted that the aforementioned task data model definition file refers to the configuration file submitted by each business application through the platform registration interface when accessing the platform, which describes its task data structure. The format of the task data model definition file can be a JSON Schema file, an XML Schema file, or a platform-defined metadata description file.

[0023] Step 2: Collect and analyze task instance data to generate value range behavior feature vectors for each field. Data on task instances generated by various business applications within a preset recent time window is collected. For each task field, statistical analysis is performed on the actual populated values ​​of that task field in the collected task instances. The statistical analysis includes: calculating the distribution characteristics of the populated values ​​of the task field, calculating the co-occurrence frequency of the task field with other fields within the same application, and calculating the correlation coefficient between the populated values ​​of the task field and other fields. Based on the results of the above statistical analysis, a value range behavior feature vector is generated for each task field.

[0024] It should be noted that the above distribution characteristics refer to statistical measures of the actual filled values ​​for the task field, including the mean, variance, skewness, kurtosis, proportion of null values, and proportion of unique values. For enumeration type fields, the distribution characteristics also include the probability distribution of each enumeration value. For text type fields, the distribution characteristics include the statistical distribution of text length and the frequency statistics of high-frequency terms.

[0025] It should be noted that the co-occurrence frequency mentioned above refers to the frequency at which the current field and other fields simultaneously have valid fill values ​​within the same task instance record. The correlation coefficients mentioned above refer to the Pearson correlation coefficient used to measure the correlation between numeric fields, and the Cramer V coefficient used to measure the correlation between categorical fields.

[0026] In this embodiment, to obtain a value domain behavior feature vector that better reflects the current actual usage semantics of the field, the aforementioned preset recent time window is set using a sliding time window approach. That is, each time data is collected, task instance data within a fixed time period prior to the current moment is used for statistical analysis. Using a sliding time window allows the value domain behavior feature vector to automatically reflect the evolution trend of the field's usage semantics over time, avoiding the dilution of the current semantic representation by historical data.

[0027] Step 3: Integrate field structure descriptions and value range behavior features to generate semantic fingerprint vectors for each field. The task field structure description vector of each task field is concatenated and fused with the corresponding value domain behavior feature vector to generate a fused feature vector. The fused feature vectors of each field are then input into a pre-trained semantic encoding network, which outputs a semantic fingerprint vector for each field.

[0028] The aforementioned semantic encoding network employs a Transformer encoder structure. Its input is a sequence of fused feature vectors, and its output is a sequence of semantic fingerprint vectors of the same length as the input sequence. This semantic encoding network has been pre-trained on a large-scale multi-source metadata corpus, which includes task data model definition files from multiple different software platforms and their corresponding task instance statistics. Through pre-training, this semantic encoding network acquires the ability to jointly represent the structural semantics and the usage semantics of fields. The output semantic fingerprint vector simultaneously encodes both the structural semantics and the actual usage semantics of the field.

[0029] Furthermore, the pre-training phase employs a contrastive learning approach, using known semantically equivalent field pairs from different platforms as positive sample pairs and semantically irrelevant field pairs as negative sample pairs. A contrastive loss function is used to ensure that the semantic fingerprint vectors of semantically equivalent field pairs are close to each other in the vector space, while the semantic fingerprint vectors of semantically irrelevant field pairs are far apart. The Adam optimization algorithm is employed as the optimization strategy.

[0030] It should be noted that the above-mentioned splicing and fusion refers to splicing the task field structure description vector and the value domain behavior feature vector end to end along the feature dimension direction to form a fused feature vector with a dimension equal to the sum of the dimensions of the two.

[0031] It should be noted that before inputting the fused feature vectors into the semantic encoding network, the fused feature vectors of each field within the same business application are arranged according to the hierarchical traversal order of the fields in the data model, forming an ordered sequence of fused feature vectors. This serialized input method enables the self-attention computation in the semantic encoding network to capture the structural relationships between fields within the same application, thereby generating context-aware semantic fingerprint vectors.

[0032] In this embodiment, to further improve the accuracy of semantic fingerprint vectors in representing the semantics of a platform-specific domain, the following steps are included before applying the pre-trained semantic encoding network to the current platform: collecting existing, manually verified field mapping relationships in the current platform as labeled data; using the fused feature vector pairs of the source field and the target standard field in the mapping relationship as training samples; and using minimizing the distance between their semantic fingerprint vectors as the training objective to fine-tune the pre-trained semantic encoding network, thereby generating a semantic encoding network adapted to the semantic environment of the current platform. The loss function for the fine-tuning training is the mean squared error loss function, calculating the Euclidean distance between the semantic fingerprint vectors of the source field and the target standard field, and using this Euclidean distance value as the loss for gradient descent optimization.

[0033] Step 4: Calculate the optimal mapping relationship between each application field and the standard field. On the standard field set of the platform's unified task model, the same semantic encoding network as in step 3 is used to perform the same feature extraction and encoding processing as in steps 1 to 3 on each standard field in the standard field set, generating reference semantic fingerprint vectors for each standard field.

[0034] For each business application, the cosine similarity between the semantic fingerprint vector of each field in the business application and the reference semantic fingerprint vector of each standard field is calculated, and a field similarity matrix for the business application is generated. The rows of the field similarity matrix correspond to each task field of the business application, and the columns correspond to each standard field in the standard field set. Each element in the field similarity matrix is ​​the cosine similarity value between the corresponding field pairs.

[0035] The Hungarian algorithm is applied to solve the field similarity matrix for each business application to obtain the optimal matching scheme that maximizes the global similarity sum, generating an automatic mapping relationship between each task field and the standard field in that business application. The input of the Hungarian algorithm is the field similarity matrix, and the output is the one-to-one correspondence allocation result between each row and each column, which is the automatic mapping relationship between each application field and the standard field.

[0036] It should be noted that during the feature extraction process of the above standard field set, the task field structure description vector of the standard field is extracted and generated from the definition file of the platform's unified task model, and the value domain behavior feature vector of the standard field is generated from the archived historical task instance data in the platform's unified task information index through statistical analysis.

[0037] In this embodiment, when the number of task fields in a business application differs from the number of standard fields, the field similarity matrix is ​​padded before the Hungarian algorithm is applied. Specifically, when the number of application fields exceeds the number of standard fields, virtual standard field columns are added to the field similarity matrix, and the similarity values ​​in the virtual columns are set to zero. When the number of application fields is less than the number of standard fields, virtual application field rows are added to the field similarity matrix, and the similarity values ​​in the virtual rows are set to zero. In the matching results, application fields paired with virtual fields are marked as unmatched fields, and standard fields paired with virtual fields are marked as uncovered standard fields.

[0038] Step 5: Perform value range alignment transformation to generate normalized unified task information. Based on the automatic mapping relationship generated in step 4, the mapping confidence of each field pair in the automatic mapping relationship is determined. The mapping confidence is the cosine similarity value of the corresponding field pair in the field similarity matrix. Field pairs with mapping confidence below a preset threshold are marked as mappings to be confirmed. For field pairs with mapping confidence higher than or equal to the preset threshold, value range alignment conversion is performed.

[0039] The above-described value range alignment transformation process is as follows: For each pair of fields to be transformed, obtain the value range behavior feature vectors of the source field and the target standard field. Calculate the difference features between the two value range behavior feature vectors. Based on these difference features, generate parameters for the value range transformation function, including the scaling factor and offset of the numerical mapping, or a table of correspondence between enumerated values. Input the fill value of the corresponding source field from the task instance data of the business application into the parameterized value range transformation function, and output the transformed normalized value. This normalized value conforms to the semantic definition and value specification of the target standard field.

[0040] The normalized task information of each business application is written into the platform's unified task information index to generate a unified task information view across applications.

[0041] It should be noted that the above value range conversion function uses a linear transformation for numeric fields, mapping the value range of the source field to the value range of the standard field through scaling factors and offsets. For enumeration fields, a lookup table mapping method is used, generating a table of correspondences between enumeration values ​​based on the semantic fingerprint vector similarity between each enumeration value in the source field and each enumeration value in the standard field, and then replacing the source enumeration values ​​with standard enumeration values ​​according to this table. For time fields, the corresponding time format conversion or time base conversion is selected based on the difference between the time semantics of the source field (absolute time point or relative project duration) and the time semantics of the standard field.

[0042] It should be noted that the above-mentioned difference features are calculated as follows: the value range behavior feature vector of the source field is subtracted element by element from the value range behavior feature vector of the target standard field to obtain the difference vector. At the same time, the two value range behavior feature vectors are concatenated. The difference vector and the concatenated vector are input into the parameter generation network to output the parameters of the value range transformation function.

[0043] Furthermore, the parameter generation network described above is a multilayer perceptron, whose input layer receives the concatenation result of the difference vector and the concatenation vector, denoted as the difference vector. ,in The value range of the source field is the feature vector. The value range behavior of the target standard field is a feature vector. This represents the transpose, and the concatenated vector is... Then the input of the input layer is The output layer outputs the parameter vector of the range transformation function; for numeric fields, the output layer outputs the scaling factor. and offset Corresponding linear transformation ,in Populate values ​​for the source field. This represents the normalized value after transformation. For enumeration type fields, the output layer outputs the matching score between each pair of enumeration values, which is used to generate an enumeration value correspondence table. This parameter generation network uses the mean squared error loss function between the source field fill values ​​of the known standard mapping field pairs after transformation and the target standard field fill values ​​as the training objective, and is trained using the Adam optimization algorithm.

[0044] In this embodiment, when writing the normalized task information into the platform's unified task information index, the mapping path and transformation parameters of each task information are also recorded simultaneously as traceability metadata. The mapping path includes the source application identifier, source field identifier, target standard field identifier, and mapping confidence value. The transformation parameters include the type identifier and specific parameter values ​​of the domain transformation function. By recording traceability metadata, the source and transformation basis of each normalized task information can be traced during subsequent data analysis and auditing.

[0045] In this embodiment, to address data model changes caused by continuous iteration of business applications, the following steps are also included: Continuously monitoring the task data model definition files registered by each business application. When a change is detected in the task data model definition file of a certain business application, the following processing is automatically triggered: Steps 1 to 3 are re-executed for the changed fields of the business application to generate an updated semantic fingerprint vector; the mapping relationship calculation in step 4 is re-executed based on the updated semantic fingerprint vector to obtain an updated automatic mapping relationship; the value domain alignment transformation processing in step 5 is re-executed based on the updated automatic mapping relationship, and the updated normalized task information is written into the platform's unified task information index to generate an updated cross-application unified task information view. The above change detection is determined by comparing the version identifier or content summary value of the task data model definition file.

[0046] In this embodiment, during the incremental update process, only the changed fields and their associated fields in the data model hierarchy are recalculated for semantic fingerprint vectors and their mapping relationships are updated. Fields that have not changed retain their original semantic fingerprint vectors and mapping relationships. This incremental update approach reduces the overhead of full recalculation caused by local field changes in a single application.

[0047] The foregoing has only described certain exemplary embodiments of the present invention by way of illustration. Undoubtedly, those skilled in the art can modify the described embodiments in various ways without departing from the spirit and scope of the present invention. Therefore, the foregoing drawings and descriptions are illustrative in nature and should not be construed as limiting the scope of protection of the claims of the present invention.

[0048] It should be noted that, in this document, the use of relational terms such as "first" and "second" is merely for distinguishing one entity or operation from another, and does 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 the element.

[0049] It should be understood that in the various embodiments of this application, the order of the above-mentioned processes does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0050] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0051] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0052] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0053] In addition, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0054] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0055] The foregoing has only described certain exemplary embodiments of the present invention by way of illustration. Undoubtedly, those skilled in the art can modify the described embodiments in various ways without departing from the spirit and scope of the present invention. Therefore, the foregoing drawings and descriptions are illustrative in nature and should not be construed as limiting the scope of protection of the claims of the present invention.

Claims

1. A task information processing method based on platform-based development, characterized in that, include: Obtain the task data model definition files registered by each business application on the platform, parse the task data model definition files of each business application, and for each task field, extract the field name, data type, value constraints, and hierarchical relationship between the field and other fields in the same application, encode the multidimensional attribute information into a numerical vector, and generate a task field structure description vector set for each business application. Collect task instance data generated by each business application within a preset time window. For each task field, calculate the distribution characteristics of its filled value, the co-occurrence frequency with other fields in the same application, and the correlation coefficient between fields, and generate the value range behavior feature vector of each task field. The task field structure description vector and the corresponding value domain behavior feature vector of each task field are concatenated along the feature dimension to generate a fused feature vector. After being arranged in the hierarchical traversal order of the fields in the data model, the fused feature vector is input into the pre-trained semantic encoding network to output the semantic fingerprint vector of each field. Perform the same feature extraction and encoding processing on the standard field set of the platform's unified task model to generate reference semantic fingerprint vectors for each standard field; calculate the cosine similarity between the semantic fingerprint vectors of each business application field and the reference semantic fingerprint vectors of each standard field to generate a field similarity matrix. Apply the Hungarian algorithm to the field similarity matrix to obtain the optimal matching scheme with the largest global similarity sum, and generate an automatic mapping relationship from each task field to the standard field; Based on the automatic mapping relationship, for field pairs with mapping confidence higher than or equal to a preset threshold, the value range behavior feature vectors of the source field and the target standard field are obtained, the difference features between the two are calculated, and the parameters of the value range transformation function are generated through the parameter generation network based on the difference features. The source field fill value is input into the value range transformation function to obtain the normalized value, and the normalized task information is written into the platform's unified task information index to generate a cross-application unified task information view.

2. The task information processing method based on platform development according to claim 1, characterized in that, The semantic encoding network adopts a Transformer encoder structure and is pre-trained on a large-scale multi-source metadata corpus using a contrastive learning approach. Semantically equivalent field pairs from different platforms are used as positive sample pairs, and semantically unrelated field pairs are used as negative sample pairs. The contrastive loss function is used for training, which makes the semantic fingerprint vectors of semantically equivalent field pairs closer to each other in the vector space, and the semantic fingerprint vectors of semantically unrelated field pairs farther apart.

3. The task information processing method based on platform development according to claim 2, characterized in that, Before applying the pre-trained semantic coding network to the current platform, the process includes: collecting manually confirmed field mapping relationships in the current platform as labeled data, using the fused feature vector pairs of the source field and the target standard field in the mapping relationship as training samples, minimizing the Euclidean distance between their semantic fingerprint vectors as the training objective, fine-tuning the pre-trained semantic coding network, and generating a semantic coding network adapted to the semantic environment of the current platform.

4. The task information processing method based on platform development according to claim 1, characterized in that, The parameter generation network is a multilayer perceptron. Its input is the element-wise difference vector between the value domain behavior feature vector of the source field and the value domain behavior feature vector of the target standard field, as well as the vector obtained by concatenating the two vectors. For numerical fields, the output layer outputs scaling factors and offsets, and maps the source field fill values ​​to normalized values ​​through linear transformation. For enumeration type fields, the output layer outputs the matching score between each pair of enumeration values, generates an enumeration value correspondence table, and replaces the source enumeration value with the standard enumeration value according to the correspondence table; the parameter generation network is trained with the mean squared error loss function between the source field filling value of the known standard mapping field pair after transformation and the target standard field filling value as the training target.

5. The task information processing method based on platform development according to claim 1, characterized in that, The preset time window is set using a sliding time window method. Each time data is collected, task instance data within a fixed time period backward from the current moment is taken for statistical analysis, so that the value domain behavior feature vector automatically reflects the evolution trend of field usage semantics over time.

6. The task information processing method based on platform development according to claim 1, characterized in that, When the number of task fields in a business application is inconsistent with the number of standard fields, the field similarity matrix is ​​padded before the Hungarian algorithm is applied: when the number of application fields is greater than the number of standard fields, a virtual standard field column is added, and the similarity value in the virtual column is set to zero. When the number of application fields is less than the number of standard fields, a virtual application field row is added, and the similarity value in the virtual row is set to zero; the application field that matches the virtual field in the matching results is marked as no matching field, and the standard field that matches the virtual field is marked as an uncovered standard field.

7. The task information processing method based on platform development according to claim 1, characterized in that, When writing the normalized task information into the platform's unified task information index, the mapping path and transformation parameters of each task information are recorded synchronously as traceability metadata. The mapping path includes the source application identifier, source field identifier, target standard field identifier, and mapping confidence value. The transformation parameters include the type identifier and specific parameter value of the domain transformation function.

8. The task information processing method based on platform development according to claim 1, characterized in that, Also includes: The task data model definition files registered by each business application are continuously monitored. Changes are detected by comparing the version identifier or content summary value of the task data model definition files. When a change is detected in the task data model definition file of a certain business application, the semantic fingerprint vector calculation, mapping relationship calculation and value domain alignment conversion are re-executed for the changed fields and their related fields in the data model hierarchy. The updated normalized task information is written into the platform's unified task information index. Fields that have not changed retain their original semantic fingerprint vectors and mapping relationships.

9. The task information processing method based on platform development according to claim 1, characterized in that, The distribution features include the mean, variance, skewness, kurtosis, proportion of null values, and proportion of unique values ​​for the filled values; for enumeration type fields, the distribution features also include the probability distribution of each enumeration value; for text type fields, the distribution features include the statistical distribution of text length and the word frequency statistics of high-frequency terms; the co-occurrence frequency is the frequency at which the current field and other fields have valid filled values ​​in the same task instance record; the correlation coefficient is measured by the Pearson correlation coefficient between numerical type fields and by the Cramer V coefficient between categorical type fields.

Images