Metadata management risk assessment model training method and device for scientific data set

Abstract

The invention relates to a scientific data set metadata management risk assessment model training method and device, and the method comprises the steps: formulating risk assessment rules which comprise a general risk assessment rule, a specific risk assessment rule and a dynamic risk assessment rule; obtaining risk assessment task features from the metadata based on a risk assessment rule; obtaining a metadata feature matrix from the metadata based on a risk assessment rule; generating a risk assessment task tree based on a risk assessment rule and the metadata feature matrix; constructing a training sample, and performing analysis training on the training sample and the risk assessment task tree to obtain a risk assessment model; and constructing a dynamic verification set, and optimizing the risk assessment model by using the dynamic verification set. By constructing a comprehensive risk assessment framework, all-around coverage and efficient processing of potential risks in metadata management are realized, and the security and effectiveness of management are remarkably improved. The metadata is comprehensively analyzed and processed by combining the processing capabilities of the structured metadata and the unstructured metadata, so that the accuracy of risk assessment is improved.

Description

Technical Field

[0001] The present invention relates to the field of risk assessment, and more specifically, to a method and device for training a metadata management risk assessment model for a scientific data set. Background Art

[0002] Scientific data mainly includes data generated through basic research, applied research, experimental development, etc. in the fields of natural sciences, engineering sciences, etc., as well as original data and its derivative data obtained and used in scientific research activities through observation, monitoring, investigation, inspection and testing, etc. Scientific data sets are collections of scientific data organized and managed based on certain rules and standards.

[0003] With the accelerated implementation of the data-intensive scientific research paradigm, large-scale scientific data sets from different sources and channels are pouring into scientific data centers and scientific data management agencies. Metadata, as descriptive data of scientific data sets, is the basis for the rapid dissemination and high-level sharing and utilization of scientific data sets.

[0004] For scientific data centers and scientific data management institutions, metadata review and metadata dynamic management are very labor-intensive, and metadata risk review and assessment are relatively complex. Lack of professional knowledge also makes it difficult to carry out related work. In addition, traditional risk assessment methods mainly rely on predefined rules and experience, which are often unable to cope with complex and changeable scientific data metadata management needs. Summary of the invention

[0005] The present invention aims to overcome at least one defect (shortcoming) of the above-mentioned prior art and provide a method and device for training a metadata management risk assessment model for a scientific data set, so as to solve the problem that risk assessment in metadata management is not comprehensive enough.

[0006] According to a first aspect of the present application, a method for training a metadata management risk assessment model for a scientific data set is provided, the method comprising: Formulate risk assessment rules, which include general risk assessment rules, specific risk assessment rules and dynamic risk assessment rules; Obtaining a metadata feature matrix from the metadata based on risk assessment rules; Generate a risk assessment task tree based on risk assessment rules and metadata feature matrix; Building training samples based on metadata, and obtaining a risk assessment model based on the training samples and the risk assessment task tree analysis and training; A dynamic verification set is constructed based on the metadata, and the risk assessment model is optimized using the dynamic verification set.

[0007] By formulating general, specific and dynamic risk assessment rules, this method can cover different types of risks, which is conducive to improving the comprehensiveness of risk assessment. By introducing dynamic risk assessment rules, the model can automatically adjust the risk assessment strategy according to changes in actual conditions, enhancing the flexibility and adaptability of the model.

[0008] Optionally, the general risk assessment rules are extracted from a guidance document; formulating specific risk assessment rules specifically includes: Obtain specific risk assessment tasks or requirements from metadata and extract characteristic keywords for specific risk assessment tasks or requirements; Converting feature keywords into word vectors and performing cluster analysis, and establishing a specific risk assessment feature matching index based on the word vectors and cluster analysis results; Features associated with and / or similar to general risk assessment rules and relationships between features are extracted according to the specific risk assessment feature matching index to establish specific risk assessment rules.

[0009] By establishing a specific risk assessment feature matching index, the retrieval and matching of information in the risk assessment process can be accelerated, thereby improving the efficiency of risk assessment.

[0010] Optionally, formulating dynamic risk assessment rules includes: Extract feature vectors from metadata management behaviors and metadata data usage records; Performing cluster analysis on the feature vector to obtain a dynamic risk assessment prompt word set; Generate a dynamic risk assessment feature matching index based on the language model, the preset expert corpus and the prompt word set; Based on the dynamic risk assessment feature matching index, features associated with and / or similar to general risk assessment rules and relationships between features are extracted and dynamic risk assessment rules are established.

[0011] Optionally, the metadata includes structured metadata and unstructured metadata, and obtaining a metadata feature matrix from the metadata based on the risk assessment rule includes: extracting metadata ontology features from the structured metadata based on the general risk assessment rules; extracting metadata implicit features from the unstructured metadata based on the general risk assessment rule and the specific risk assessment rule; Annotating the metadata ontology features and implicit features based on the dynamic risk assessment rules; The feature association between the metadata ontology feature and the implicit feature is obtained, and a metadata feature matrix for risk assessment is constructed according to the feature association and the annotation.

[0012] By extracting and annotating the ontological and implicit features of metadata, we can more comprehensively assess the potential risks of metadata and thus more accurately identify risk points.

[0013] Optionally, the acquiring of the feature association between the metadata ontology feature and the implicit feature specifically includes: Acquire metadata features of metadata, where types of the metadata features include structured and unstructured; cluster the metadata features according to semantic similarity of the metadata features; The metadata features in each cluster are classified according to the type of metadata features, and the metadata features of the same type in the same cluster are grouped into item sets for association rule mining; Converting metadata features in the item set into semantic vectors; The association rules of metadata ontology features and implicit features are extracted from the semantic vector using an association rule algorithm.

[0014] The item sets are organized by type, which further refines the association level between features and makes the mining of association rules more targeted and effective. Converting metadata features into semantic vectors can capture the deep semantic relationship between features.

[0015] Optionally, the risk assessment task tree includes a general risk assessment task tree, a specific risk assessment task tree and a dynamic risk identification task tree; generating the risk assessment task tree based on the risk assessment rules and the metadata feature matrix includes: Construct a general risk assessment task tree based on general risk assessment rules and metadata feature matrix; Extract association rules between specific risk assessment rules and general risk assessment task tree nodes based on preset specifications, standards and practical experience, and establish a specific risk identification task tree based on the specific risk assessment rules and the associated general risk assessment task tree nodes; A dynamic risk identification task tree is established based on dynamic risk assessment rules and metadata feature matrix.

[0016] Optionally, the training samples include a first training sample set, a second training sample set, and a third training sample set, and constructing the training samples based on metadata specifically includes: Obtain metadata and full life cycle management process data of scientific data sets; the full life cycle management process data includes metadata quality review opinions and historical risk event records; Extract characteristic keywords and context sentences in metadata quality review opinions according to general risk assessment rules to obtain first review semantic features; classify the first review semantic features according to rule features, and score each type of first review semantic features through sentiment calculation to obtain a general risk score of the sample; Acquire a metadata management description according to the metadata and the full life cycle management process data, and construct a first training sample set based on the metadata management description, the first review semantic feature and the universal risk score; Extract characteristic keywords and context sentences in metadata quality review opinions according to specific risk assessment rules to obtain second review semantic features; classify the second review semantic features according to rule features, and score each type of second review semantic features through sentiment calculation to obtain a specific risk score of the sample; Constructing a second training sample set based on the metadata management description, the second review semantic feature, the general risk score and the specific risk score; According to the metadata management behavior record, the metadata state change time sequence information and the metadata management behavior characteristics are respectively obtained to obtain metadata management behavior sample data; the metadata feature association semantics in the risk event record are obtained to obtain risk event sample data; A third training sample set is constructed based on the metadata management behavior sample data, the risk event sample data, and samples in the first training sample set and the second training sample set.

[0017] Through a multi-level and multi-dimensional training sample construction process, it integrates expert review opinions, sentiment calculation scores, metadata management behaviors and risk event correlation information under general, specific and dynamic risk assessment rules, providing rich and high-quality data support for the training of risk assessment models, which helps to improve the accuracy and effectiveness of risk assessment.

[0018] Optionally, the risk assessment model is obtained by analyzing and training the training samples and the risk assessment task tree, specifically including: Analyze the first training sample based on the general risk assessment task tree to obtain a general risk assessment model; Analyzing the second training sample according to the general risk assessment model to obtain a first association rule set between specific risk features and general risk features; Analyzing the second training sample based on the specific risk assessment task tree to obtain a specific risk assessment model; Analyzing the first training sample according to the specific risk assessment model to obtain a second association rule set between the specific risk features and the general risk features; Based on the first association rule set and the second association rule set of the specific risk characteristics and the general risk characteristics, the nodes of the general risk assessment task tree are optimized, and the first training sample is analyzed using the optimized general risk assessment task tree to obtain a final general risk assessment model; The third training sample is analyzed based on the dynamic risk identification task tree to obtain a dynamic risk assessment model.

[0019] By generating association rule sets, the relationships between metadata features can be revealed and utilized, and the model can identify the dependencies between features and prioritize their combined impact during risk assessment, thereby improving the effectiveness and accuracy of the assessment model.

[0020] Optionally, the full life cycle management process data also includes user logs, system logs, and expert corpora involved in metadata management; the method also includes: Deploy risk assessment knowledge base and assessment feedback network according to specific risk assessment model and dynamic risk assessment model, including: Feedback step: establish a verification reference corpus based on user logs, system logs, and expert corpora involved in metadata management; analyze the first training sample based on the feature risk assessment model and the dynamic risk assessment model to obtain rule node data; extract classification features from the second training sample and the third training sample, use the classification features to perform hierarchical clustering analysis to obtain hierarchical cluster centers, calculate similarity between the hierarchical cluster centers and the verification reference corpora in the verification reference corpus to obtain verification corpora, and feed back the verification corpora to stakeholders of the scientific data set; the stakeholders include metadata producers, data managers, data analysts, and data consumers of the scientific data set; Knowledge extraction: Based on the opinions of stakeholders in the validation corpus and scientific data sets, metadata management risk assessment domain knowledge is extracted to build a knowledge base; Correction step: Correct the specific risk assessment rules and dynamic risk assessment rules according to the knowledge base, and optimize the training of the specific risk assessment model and dynamic risk assessment model.

[0021] By building a feedback network, knowledge base and correction mechanism, we have achieved continuous optimization and knowledge accumulation of risk assessment models. We have established verification references using user logs, system logs and expert corpora, and improved the accuracy of assessment results through multi-model analysis and hierarchical clustering verification, and provided timely feedback to relevant parties. At the same time, we have extracted domain knowledge from the verification process and feedback, built a risk assessment knowledge base, provided support for model correction and optimization, and improved the accuracy and timeliness of risk assessment.

[0022] According to a second aspect of the present application, a metadata management risk assessment model training device is provided, the device comprising: A risk assessment rule formulation module, used to formulate risk assessment rules, wherein the risk assessment rules include general risk assessment rules, specific risk assessment rules and dynamic risk assessment rules; An acquisition module, used to acquire risk assessment task features from metadata based on risk assessment rules; A risk assessment task tree generation module is used to generate a risk assessment task tree based on risk assessment rules and risk assessment task characteristics; A model training module is used to construct a training sample based on metadata, and to obtain a risk assessment model based on the training sample and the risk assessment task tree analysis and training; The optimization module is used to construct a dynamic verification set based on the metadata, and optimize the risk assessment model using the dynamic verification set.

[0023] Based on any of the above aspects, the metadata management risk assessment model training method and device for scientific data sets provided in the embodiments of the present application construct a comprehensive and intelligent risk assessment framework by formulating three types of risk assessment rules: general, specific and dynamic. General rules provide a wide range of basic risk identification, specific rules conduct refined analysis based on specific field needs, and dynamic rules ensure the flexibility and sensitivity of the assessment by adapting to new risks in real time. This multi-level and collaborative technical integration strategy achieves full coverage and efficient processing of potential risks in metadata management, significantly improving the security and effectiveness of management. At the same time, by combining the processing capabilities of structured and unstructured metadata, it is possible to comprehensively analyze and process metadata of different types and formats, thereby improving the accuracy of risk assessment. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0025] Figure 1 A schematic application scenario diagram of the metadata management risk assessment model training method for scientific datasets provided in this embodiment.

[0026] Figure 2 A flowchart of a metadata management risk assessment model training method for a scientific dataset provided in this embodiment.

[0027] Figure 3 This embodiment provides a schematic diagram of sub-steps of step S1.

[0028] Figure 4 This embodiment provides a schematic diagram of sub-steps of step S1.

[0029] Figure 5 This embodiment provides a schematic diagram of sub-steps of step S2.

[0030] Figure 6 This embodiment provides a schematic diagram of sub-steps of step S24.

[0031] Figure 7This embodiment provides a schematic diagram of sub-steps of step S3.

[0032] Figure 8 This embodiment provides a schematic diagram of sub-steps of step S4.

[0033] Fig. 9 This embodiment provides a schematic diagram of sub-steps of step S4.

[0034] Fig.10 A schematic diagram of the functional modules of the metadata management risk assessment model training device provided in this embodiment. DETAILED DESCRIPTION

[0035] The drawings of this application are only used for illustrative purposes and should not be construed as limiting the present application. In order to better illustrate the following embodiments, some parts of the drawings may be omitted, enlarged or reduced, and do not represent the size of the actual product; it is understandable to those skilled in the art that some well-known structures and their descriptions in the drawings may be omitted.

[0036] In order to enable those skilled in the art to better understand the solution of the present application, the technical solution in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work should fall within the scope of protection of the present application.

[0037] It should be noted that the terms "first", "second", etc. in the specification and claims of the present application and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchangeable where appropriate, so that the embodiments of the present application described herein can be implemented in an order other than those illustrated or described herein. In addition, the terms "including" and "having" and any of their variations are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device comprising a series of steps or units is not necessarily limited to those steps or units clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0038] In scientific data centers and scientific data management institutions, metadata review and dynamic management tasks are arduous. Traditional risk assessment methods mostly rely on fixed rules and past experience. When faced with the complex and changing needs of scientific data metadata management, their limitations and shortcomings become apparent. In order to meet this challenge, improve the efficiency of scientific data set acceptance review, and promote management and sharing levels to a new level, it is urgent to build a metadata management risk assessment model that can dynamically learn and continuously optimize.

[0039] This embodiment provides a technical solution that can solve the above-mentioned problem. The specific implementation methods of this application are described in detail below in conjunction with the accompanying drawings.

[0040] For example, a schematic diagram of an application scenario of a metadata management risk assessment model training method for a scientific data set provided in an embodiment of the present application is provided. Figure 1 As shown, the application scenario at least includes a server 100 and a terminal 200 that can communicate with the server 100, and the server 100 has a data processing function; the terminal device 200 has an image processing function and may also have a data display function.

[0041] It is understandable that the server 100 can be an independent electronic device or a cluster of multiple electronic devices; the terminal 200 can be a smart phone terminal, a personal computer, a tablet computer, a car terminal, etc., but is not limited thereto.

[0042] In one practicable manner, the server 100 and the terminal 200 may respectively execute the metadata management risk assessment model training method for a scientific data set provided in an embodiment of the present application, or, optionally, the metadata management risk assessment model training method for a scientific data set provided in an embodiment of the present application is partially executed in the server 100 and partially executed in the terminal 200.

[0043] like Figure 2 As shown, this embodiment provides a method for training a metadata management risk assessment model for a scientific data set, which may include the following steps: S1. Formulate risk assessment rules, which include general risk assessment rules, specific risk assessment rules and dynamic risk assessment rules.

[0044] Specifically, the general risk assessment rules include structured assessment rules constructed based on feature attributes, feature keywords and feature associations extracted from documents such as standards and specifications, etc. A risk point identification rule tree is constructed based on the general risk assessment rules.

[0045] General risk assessment rules can be extracted from guiding documents such as policies, standards, and work guidelines. A structured risk point identification rule tree can be constructed based on the risk assessment rule type, assessment dimension, and assessment content reflected in the document. Combined with scientific data management, scientific data metadata management actual cases, expert advice, and user feedback, a general risk feature library can be established for the corresponding risk point identification rule tree. According to the characteristic attributes and association rules in the general risk feature library, corresponding assessment conditions are formulated, and the branch logic of the risk point identification rule tree is determined by the association rules and expert knowledge in the risk feature library.

[0046] Specific risk assessment rules are formulated based on special needs such as specific project tasks, specific standards and specifications, and specific data application scenarios. For example, specific risk assessment rules can be extracted based on the processing of highly sensitive data and the management requirements of data in specific fields such as medical imaging data and biological gene data. These specific rules are not applicable to all scenarios, but are tailored for certain specific metadata management scenarios.

[0047] like Figure 3 As shown, in this embodiment, formulating a specific risk assessment rule specifically includes the following steps: A1. Obtain specific risk assessment tasks or requirements from metadata and extract characteristic keywords for specific risk assessment tasks or requirements.

[0048] During the specific implementation process, it is necessary to conduct a detailed analysis of metadata-specific risk assessment tasks or requirements, determine the specific risk areas or scenarios that need to be assessed, extract high-frequency words and important entities related to specific tasks from relevant documents, standards and specifications, and data descriptions through word frequency analysis and named entity recognition technology, and understand the role and relationship of keywords in sentences based on the results of grammatical structure analysis.

[0049] A2. Convert the feature keywords into word vectors and perform cluster analysis, and establish a specific risk assessment feature matching index based on the word vectors and cluster analysis results.

[0050] The extracted feature keywords are converted into word vectors, and the pre-trained word vector model is used to convert the feature keywords into word vectors in a high-dimensional vector space. Cluster analysis is performed on the generated word vectors to identify words or phrases with similar risk characteristics, and the words or phrases with similar risk characteristics are classified into one category to form a cluster center. The clustering results are used as index entries to construct a specific risk assessment feature matching index. The specific risk assessment feature matching index is used to quickly match risk characteristics in a specific field.

[0051] Finally, these clustering results and their cluster centers are used as index entries to construct a matching index for specific risk assessment features. This indexing system allows users or systems to quickly retrieve and match words or phrases related to specific risk features, thereby accelerating the risk assessment process and improving accuracy. By comparing the words in the text to be evaluated with the cluster centers in the index, possible risk features in the text can be quickly identified and appropriate risk management measures can be taken accordingly.

[0052] A3. Extract features and relationships between features that are associated and / or similar to general risk assessment rules based on the specific risk assessment feature matching index, and establish specific risk assessment rules.

[0053] Dynamic risk assessment rules are risk assessment rules that can adapt to changes in the metadata management environment in real time and automatically adjust according to new data, user behavior, and environmental changes. Dynamic risk assessment rules are characterized by high flexibility and the ability to dynamically identify and respond to new potential risks based on the actual management operations, usage, and external environment changes of metadata.

[0054] The sources of dynamic risk assessment rules include metadata management behavior records, data management demand changes, abnormal data interaction behaviors, etc. The dynamic risk assessment rules established based on the features extracted from the general risk assessment rules are unverified dynamic risk assessment identification rules, which are used to construct a dynamic risk assessment task tree to obtain the relationship between dynamic risks and metadata management behaviors, metadata features, general risks, and specific risks.

[0055] In this embodiment, if Figure 4 As shown, formulating dynamic risk assessment rules includes: S111. Extracting feature vectors from metadata management behaviors and metadata data usage records.

[0056] S112. Perform cluster analysis on the feature vector to obtain a dynamic risk assessment prompt word set.

[0057] The prompt word set is based on the high-frequency keywords, phrases and clustering features extracted from metadata management behaviors and scientific data usage records. These prompt words are used to search in the corpus to help identify potential risk assessment rules. Specifically, the prompt word set based on prompt words can include metadata operations such as data format updates and permission changes; data behavior trends such as increased query frequency and access to abnormal metadata; and risk indicators such as field omissions and inconsistent terms.

[0058] S113. Generate a dynamic risk assessment feature matching index based on the language model, the preset expert corpus and the prompt word set.

[0059] The preset expert corpus is a text resource library built based on expert knowledge, industry standards and practical experience in the field of scientific data management. It includes best practices, risk assessment standards, common problems and solutions related to metadata management, as well as metadata management guidelines, risk control recommendations, field-specific standards, actual cases, etc.

[0060] In this embodiment, a matching index can be generated using a large language model in combination with a preset expert corpus and a prompt word set. The core of the matching index is to find out the features associated with the existing general risk assessment rules by analyzing the keywords and feature words in the metadata management behavior and scientific data usage records. Specifically, feature vectors including permission changes, access anomalies, data format updates, etc. can be extracted from the metadata management behavior and usage records as inputs to the large language model. Through the matching index, the system can automatically compare these features with the general risk assessment rules in the expert corpus, thereby identifying the general risk assessment rules related to these features.

[0061] Based on the matching index, not only can the features directly associated with the general risk assessment rules be identified, but also other features with high semantic similarity to these features. For example, in a set of metadata management behaviors, the system detects permission change features and abnormal access frequency features, which can be associated with the data leakage risk features in the general risk assessment rules through matching indexes. Assuming that the system also identifies "data accuracy issues" that are highly correlated with these features, the system will automatically generate a dynamic risk assessment rule: when permission changes and abnormal access frequency occur at the same time, not only the possibility of data leakage should be checked, but also the problem of data accuracy should be evaluated.

[0062] S114. Extract features associated with and / or similar to general risk assessment rules and / or similar rule features and relationships between features based on the dynamic risk assessment feature matching index and establish dynamic risk assessment rules.

[0063] Association rule features refer to features, attributes or conditions that are directly related to a specific data set or metadata management task. These features can usually directly affect the decision-making process of data management. Association rule features are extracted based on known specifications, standards or practical experience, and are highly related to key indicators or requirements in specific risk assessment tasks.

[0064] Similar rule features refer to features that are similar or relevant to other rules in the general risk assessment rules through in-depth analysis of keywords, features, or attributes in a specific risk assessment task.

[0065] Similar rule features can be determined by calculation and analysis. Specifically, a clustering algorithm can be used to cluster the keywords in the feature matching index, identify similar feature words, and group the semantics and contextual relationships of the feature keywords according to the results.

[0066] Find the context words related to the feature keywords in each group and generate a related word list as the extended feature keywords.

[0067] Use association rule mining algorithms to analyze the extended feature set and find frequent item sets and strong association rules.

[0068] Obtain assessment rules similar to the current specific risk assessment task through external knowledge bases or expert consultation.

[0069] For example, a specific risk assessment task is "review of scientific dataset metadata changes", the goal of which is to ensure that each metadata change does not cause data quality issues, data consistency issues, or data usage errors. However, based on the existing specific risk assessment feature matching index, it is impossible to obtain strongly related assessment rules from general risk assessment rules.

[0070] The characteristic keywords of "metadata change review" can be analyzed, such as "metadata update", "change impact analysis", and "version control". Through semantic analysis, similar features related to these features can be discovered, including "data structure change", "data dependency analysis", "data version consistency" and other similar features.

[0071] These similar features (such as “data structure change” and “data version consistency”) are incorporated into the feature space, and general risk assessment rules are searched again based on these extended features.

[0072] Association rule mining technology is used to generate new rules by analyzing the potential associations between features.

[0073] For example, the analysis found that "after the metadata version is upgraded, if the data dependency analysis does not fully cover all downstream dependencies, it may cause data processing errors." This can generate a new risk assessment rule: after the metadata is changed, a comprehensive data dependency analysis must be performed to ensure that all dependencies are properly maintained in the new version.

[0074] S2. Obtain metadata feature matrix from metadata based on risk assessment rules.

[0075] In this embodiment, the metadata includes structured metadata and unstructured metadata.

[0076] like Figure 4 As shown, step S2 includes the following steps: S21. Extract metadata ontology features from structured metadata based on general risk assessment rules.

[0077] Extracting ontology features from structured metadata refers to extracting explicit features from the structured metadata of scientific datasets, such as extracting dataset name, data type, creation date, version number, data owner and other feature information from the dataset.

[0078] In the specific implementation process, the predefined metadata ontology model can be used to extract ontology features in structured metadata to ensure that the extracted features conform to the standard format. For example, the predefined metadata ontology model can be used to extract explicit features such as dataset name, data format, creation time, and update frequency.

[0079] S22. Extract metadata implicit features from unstructured metadata based on general risk assessment rules and specific risk assessment rules.

[0080] Extracting implicit features from unstructured metadata refers to analyzing unstructured data such as relevant documents, user feedback, and usage records of scientific data sets to identify potential implicit features. For example, key sentiment words and frequently occurring topics can be extracted from user comments, and then access logs can be analyzed to find abnormal access frequencies of certain data sets within a specific time period, thereby identifying potential risks.

[0081] S23. Annotate metadata ontological features and metadata implicit features based on dynamic risk assessment rules.

[0082] Labeling metadata ontology features and implicit features means labeling risk warnings based on dynamic risk assessment rules obtained from the dynamic management of datasets, dataset association information, and datasets of the same type and similar datasets. Analyze the management of datasets of the same type or similar datasets, including usage frequency, update frequency, and user feedback information, and generate corresponding dynamic risk rules based on the high-frequency features or outlier features of the above information.

[0083] When risk annotation rules are applied to metadata datasets, they provide suggestive annotations to warn of possible risk areas. For example, if a dataset has a surge in usage frequency within a short period of time, and this phenomenon is not common in similar datasets, it may be labeled as "abnormal usage risk."

[0084] In the risk assessment process, metadata features extracted according to general risk and specific risk rules are annotated based on dynamic risk rules in order to identify and capture potential risks brought about by dynamic changes, which are potential risks that cannot be identified by conventional rules.

[0085] Specifically, when the system identifies that most similar data sets have a high update frequency, while a certain data set has a low update frequency, it indicates that the data set may have a risk of delayed update. Therefore, the dynamic labeling rule can be used to label this feature as "low update frequency risk", thereby supplementing the risk points that are not considered by general rules and specific rules.

[0086] Similarly, in some data sets, if user feedback frequently mentions inaccurate data, but this phenomenon is not common in other data sets, then the dynamic labeling rules can be used to label this feature as "data quality abnormality risk." Labeling metadata through dynamic labeling rules can provide managers with new risk information, thereby promptly notifying possible quality problems in metadata.

[0087] S24. Obtain feature associations between metadata ontology features and implicit features, and establish a metadata feature matrix for risk assessment based on feature associations and dynamic risk annotations.

[0088] In this embodiment, if Figure 5 As shown, the feature association between metadata ontology features and implicit features is obtained, including: S241 , obtaining metadata features of metadata, where the types of metadata features include structured and unstructured; and clustering the metadata features according to semantic similarity of the metadata features.

[0089] In the specific implementation process, clustering algorithms such as K-means and hierarchical clustering can be used to divide metadata features into different clusters according to the semantic similarity between metadata features.

[0090] S242: Classify the metadata features in each cluster according to the type of the metadata features, and group the metadata features of the same type in the same cluster into an item set for association rule mining.

[0091] S243. Convert metadata features in the item set into semantic vectors.

[0092] During the specific implementation process, natural language processing technology or word embedding models can be used to convert metadata features in the item set into high-dimensional semantic vectors.

[0093] S244: extracting association rules of metadata ontology features and implicit features from the semantic vector using an association rule algorithm.

[0094] In the specific implementation process, association rule algorithms such as Apriori and FP-Growth can be used to extract association rules. After establishing feature association rules, the annotation results are integrated into the feature matrix.

[0095] The following is a description of the feature association between metadata ontology features and implicit features in conjunction with a specific scientific dataset. Taking an agricultural genomic scientific dataset as an example, there is a general risk assessment rule: each time the data usage metadata field is updated, a metadata version record description must be added. Based on this rule, features can be extracted from structured metadata: "{Original data description: Identify and analyze genetic variation in crops, help improve crop varieties, increase yield, disease resistance and environmental adaptability, and thus support precision agriculture and sustainable crop production}", "{Added data description: Now expanded to the study of drought-resistant crop varieties. By re-analyzing existing genomic data, identify genes related to drought resistance, and provide a scientific basis for the genetic improvement of drought-resistant crops}", "{Latest update date: 20240707}", "{Version number: 2024v2}".

[0096] Based on the above general risk rules, features can be extracted from unstructured metadata: "{Data usage document: This dataset contains genome sequences and related phenotypic data of various crops, mainly used to study the relationship between gene variation and crop traits. By analyzing these data, key genes related to important agricultural traits such as yield, disease resistance, and stress resistance can be identified, which is helpful for gene editing and variety improvement research. This dataset is particularly suitable for scientific research projects such as precision agriculture, breeding programs, and environmental adaptability analysis.}", "{Version record description: Update content-expand the use of agricultural genome datasets and add the application scenario of "drought-resistant crop variety research"; new data usage description-based on the existing dataset, drought resistance gene analysis is carried out to provide support for genetic improvement of drought-resistant crops; metadata field update-add the usage description of "drought-resistant crop variety research" in the "data usage" field to reflect the expanded application scope; impact scope-this extension may affect future crop improvement research, and users can adjust their research direction with reference to the new usage; update process-after review by field experts, it was decided to expand the application scope of the dataset, and the metadata record was updated accordingly. The system updates the version information and generates an update report to notify relevant users.}" Another specific risk assessment rule is: "For key R&D projects, when the 'data usage' metadata field is updated, the project results corresponding to the usage must be marked in the metadata version record description." Based on this specific risk assessment rule, features can be extracted from unstructured metadata "Project Implementation Summary Report-Project Achievement Description The genomic dataset has been verified in the improvement of drought-resistant crop varieties. The research results include the development of high-yield crops under drought conditions. The related paper "Genomic Analysis and Application of Drought-Resistant Crops" has been published; the results-related uses are in field trials conducted in multiple locations. Crops using this genomic dataset showed obvious drought resistance advantages."

[0097] According to the semantic similarity of the associated metadata features extracted from the general risk assessment rules and specific risk assessment rules and the keywords of the rule features, the association between the ontology features and the implicit features can be established. For example, the text of the "original data description" (such as "identifying and analyzing the genetic variation of crops") and the "new data description" (such as "now expanded to the research of drought-resistant crop varieties") is preprocessed and embedded using natural language processing technology to convert it into a vector representation. Then, the semantic similarity between the vector and the use description in the "data usage document" (such as "mainly used to study the relationship between genetic variation and crop traits") is calculated. According to the calculated similarity, the semantically related text features are associated and used to update the metadata task tree or risk assessment model.

[0098] S3. Generate a risk assessment task tree based on risk assessment rules and metadata feature matrix.

[0099] The metadata feature matrix is ​​the basis for building the risk assessment task tree. Different metadata features are classified according to the correlation of association rules and feature correlation. When building the risk assessment task tree, each feature is used as a task node of the risk assessment task tree.

[0100] The risk assessment task tree includes a general risk assessment task tree, a specific risk assessment task tree and a dynamic risk identification task tree.

[0101] like Figure 6 As shown, step S3 specifically includes the following steps: S31. Construct a general risk assessment task tree based on general risk assessment rules and metadata feature matrix.

[0102] Extract general risk assessment rule types and general risk assessment rule features from general risk assessment rules, extract metadata features related to each general risk from the metadata feature matrix based on the general risk assessment rule types and general risk assessment rule features, take the identified risk rule types or rule features as the root node, and the related metadata features as child nodes, analyze the dependencies between child nodes, and record these dependencies in the node attributes to form a hierarchical structure.

[0103] S32. Extract association rules between specific risk assessment rules and general risk assessment task tree nodes based on preset specifications, standards and practical experience, and establish a specific risk identification task tree based on the specific risk assessment rules and the associated general risk assessment task tree nodes.

[0104] Specific risk assessment rule features and semantic features are extracted from specific risk assessment rules, similar general risk assessment rule features are obtained based on specific risk assessment rule semantic features, and metadata features related to each risk are extracted from the metadata feature matrix based on general risk assessment rule features. Then, nodes are created for specific risk assessment rule features, and metadata features related to specific risk rules are attached to specific risk nodes. Then, the nodes in the general risk assessment task tree are analyzed, nodes related to specific risks are identified, and nodes related to specific risks are connected to specific risk nodes, thereby constructing a specific risk identification task tree. Among them, specific risk nodes include feature information and their corresponding specific risk assessment rules. In the specific implementation process, association rules between specific risk assessment rules and general risk assessment task tree nodes can be extracted based on preset specifications, standards and practical experience, and specific risk assessment rules and general risk assessment task tree nodes can be associated according to the extracted association rules.

[0105] S33. Establish a dynamic risk identification task tree based on dynamic risk assessment rules and metadata feature matrix.

[0106] Extract the annotated features from the metadata feature matrix, and initialize the nodes of the dynamic risk identification task tree based on the extracted metadata features. Each node represents a annotated metadata feature, and connect each node with the corresponding dynamic risk assessment rule. By connecting similar or related feature nodes, the initial dynamic risk identification task tree structure is formed.

[0107] For example, the low data update frequency feature is associated with the outdated data feature mentioned in the user feedback, and the association between the two nodes can be determined by the feature matching index. In specific implementation, the feature matching index can be generated using a large language model and an expert corpus.

[0108] S4. Construct training samples based on the metadata, and obtain a risk assessment model based on the training samples and the risk assessment task tree analysis and training.

[0109] In this embodiment, the training samples include a first training sample set, a second training sample set, and a third training sample set.

[0110] like Figure 7 As shown in the figure, constructing training samples based on metadata includes the following steps: S411. Obtain metadata and full life cycle management process data of scientific data sets, wherein the full life cycle management process data includes metadata quality review opinions, review corpus, and historical risk event records.

[0111] The full life cycle management process data of scientific datasets includes metadata quality review opinions, review corpus, and historical risk event records.

[0112] In the specific implementation process, the metadata field content, management behavior field type, and management record field can be used as the metadata management description of the sample.

[0113] S412. Extract characteristic keywords and context sentences in metadata quality review opinions according to general risk assessment rules to obtain first review semantic features; classify the first review semantic features according to rule features, and score each type of first review semantic features through sentiment calculation to obtain a general risk score for the sample.

[0114] S413. Obtain a metadata management description according to the metadata and the full life cycle management process data, and construct a first training sample set based on the first review semantic features and the general risk score of the metadata management description.

[0115] S414, extracting characteristic keywords and context sentences in the metadata quality review opinions according to specific risk assessment rules to obtain second review semantic features; classifying the second review semantic features according to rule features, and scoring each type of second review semantic features through sentiment calculation to obtain a specific risk score of the sample; S415, constructing a second training sample set based on the metadata management description, the second review semantic feature, the general risk score and the specific risk score; S416, respectively obtaining metadata state change time sequence information and metadata management behavior characteristics according to the metadata management behavior record to obtain metadata management behavior sample data; obtaining metadata characteristic association semantics in the risk event record to obtain risk event sample data; S417: Construct a third training sample set based on the metadata management behavior sample data, the risk event sample data, and samples in the first training sample set and the second training sample set.

[0116] In this embodiment, each sample in the sample set includes a metadata management description, an expert review semantic feature, and a risk score.

[0117] The metadata management description is the basic feature of the sample, which is used to describe the management process and status of metadata, including the metadata field content of the scientific data set, the management behavior field type, and the management record field content.

[0118] The expert review semantic features are obtained by extracting the characteristic keywords and context sentences in the metadata quality expert review opinions and classifying them according to the rule features. The expert review semantic features provide the expert review opinions and analysis for the sample, which are helpful for the subsequent evaluation of the quality and potential risks of the dataset in the management process.

[0119] The general risk score is a general risk score obtained by scoring the semantic features of the first review based on sentiment calculation. The general risk score indicates the risk level of the sample in the general risk assessment. The general risk score quantifies the risk level assessment of the sample, which helps in subsequent risk analysis and decision-making.

[0120] Similarly, the specific risk score is a general risk score obtained by scoring the semantic features of the second review based on sentiment calculation.

[0121] In the specific implementation process, when building a risk assessment model, data preparation is first carried out. Incomplete, repeated or inconsistent data in metadata management are cleaned by means of rule filtering, missing value filling and redundant deletion to ensure the accuracy of training data, and general and specific risk scores are standardized or normalized to meet the algorithm requirements. Next, feature selection and feature engineering are carried out to extract key features from metadata fields, expert review opinions and risk scores, and the data set is divided into training sets, validation sets and test sets to randomly select multiple samples to ensure the comprehensiveness of model training. In terms of algorithm selection, according to different model types, algorithms such as decision trees, random forests, logistic regression, support vector machines, recurrent neural networks (RNNs) or long short-term memory networks (LSTMs) are selected to adapt to different risk assessment tasks. When initializing the model, initial parameters and hyperparameters can be set for each algorithm, and then the training process can be started. During the model training process, the prediction results are calculated through forward propagation, and the difference between the predicted value and the true value is calculated using the loss function, and the model is optimized through the back-propagation algorithm and parameter update iteration strategy. During the model training process, batch training and cycle training are used to improve training efficiency and stability, and the early stopping strategy is used to prevent overfitting. Subsequently, the model's hyperparameters are adjusted using the validation set to optimize the model's performance, and the validation set is used to evaluate the model's performance at preset time intervals. Finally, the model's performance is evaluated using the test set, and the model's effect is measured using indicators such as accuracy, recall, and F1 score.

[0122] For example, taking the risk assessment of agricultural scientific dataset management as an example, a scientific dataset for corn variety approval includes key data on corn variety planting, yield, disease resistance, environmental adaptability, etc., and metadata fields include variety number, variety name, test site, test year, yield, disease resistance index, environmental climate conditions, etc. In the process of data management and use, there are tasks such as metadata quality control, data update, sharing management and approval compliance. In order to deal with the risk assessment of these tasks, the first general risk assessment model, specific risk assessment model and dynamic risk assessment model can be established.

[0123] Among them, the first general risk assessment model is used to assess the general risks of corn variety approval datasets, mainly including data integrity, data consistency, whether metadata fields are updated in a timely manner, and the standardization of data management behaviors. During the data cleaning process, it is necessary to remove incomplete records or entries containing outliers in the dataset. For example, delete entries that do not fill in the test year or test site. In data standardization, different metadata fields such as test site name and yield unit are standardized into a unified format. For example, the place name formats of different test sites are consistent, and the yield units used in different records are unified. Relevant metadata fields, such as variety number, variety name, yield data, and disease resistance, are selected as features for assessing general risks such as data integrity and timeliness of updates.

[0124] In the specific implementation process, the random forest algorithm can be used to train the general risk assessment model. The corn variety approval scientific data set includes a variety of metadata fields, such as variety number, test year, test site, yield, etc. These metadata features are high-dimensional data, and the random forest algorithm can effectively process this type of multi-dimensional and multi-feature data. General risk assessment usually requires data to be classified into "compliance" or "risk" categories. Random forests are good at handling classification tasks and can integrate multiple classification results through multiple decision trees to improve the accuracy of the model. Suppose the planting data of a variety in multiple test sites are inconsistent, and there are missing values ​​in the disease resistance data. Using the random forest algorithm, multiple decision trees are used to evaluate different features such as test year, yield data, and disease resistance data. Through the voting mechanism, the general risk assessment model can finally determine whether this variety has the risk of missing data or inconsistent format.

[0125] Specifically, a subset of metadata records can be randomly extracted from the training data. For example, the training data includes 2,000 metadata records, each of which contains features such as field completeness, update time, usage frequency, and number of citations. Each time, 1,500 data are randomly extracted from the data to build a decision tree.

[0126] During the splitting process of each tree, the general risk assessment model will randomly select some of the metadata features such as field completeness, usage frequency, and last update time for splitting. For example, during a split, the general risk assessment model may only assess the risk based on field completeness and usage frequency.

[0127] The split point is chosen by analyzing the contribution of specific features to metadata risk. For example, if the model finds that field completeness below 80% is a high-risk split point, then all metadata records with low completeness will go into the high-risk branch.

[0128] Use information gain or Gini coefficient to evaluate the quality of splitting. For example, the combination of completeness and frequency of use can better predict the risk of metadata loss. Continue to select appropriate features for splitting until further splitting is impossible or the depth of the tree reaches the set upper limit. For example, continue to split according to features such as "last update time" and "number of citations" until the leaf nodes of the tree represent high, medium, and low risks. When applied, if a metadata entry records the fields of corn varieties, including variety name, growth cycle, disease resistance score and other information, its field completeness is 85%, the last update time was 2 years ago, and the frequency of use is high, the random forest model analyzes these features and combines the voting results of multiple decision trees to assess its risk level as "medium risk", indicating that it may be necessary to pay attention to the update of this entry.

[0129] Specific risk assessment models are used for metadata risks in certain data interaction scenarios or specific areas. For example, when metadata involves cross-system sharing or docking with external databases, some specific risks may arise, such as synchronization failure, data conflict, compatibility issues, etc. During the review process, the accuracy, compliance and compliance of yield data and disease resistance data can be evaluated.

[0130] For example, taking the scientific data set of corn variety approval as an example, specific risks may include "inconsistent variety information", "improper handling of data sensitivity in a specific time period", "data structure does not meet requirements", etc. Keywords related to risk assessment are extracted from relevant data description documents, task books, usage records, etc. through word frequency analysis, named entity recognition (NER) and other methods, and they are converted into features that can be used in logistic regression models. The training data required for logistic regression is constructed by extracting keywords and risk-related information in metadata. For example, the categorical features of variety names are converted into dummy variables, and the approval year, format, etc. are converted into numerical features. The associated data sensitivity level can be converted into risk level through a scoring system. Logistic regression is selected as the main algorithm for model training. The logistic regression model can explain which factors play a key role in risk assessment by looking at the weights of the features. In model training, the features of each metadata entry are input into the logistic regression model. The model calculates the predicted risk probability value (between 0 and 1) based on the features, and uses maximum likelihood estimation (MLE) to optimize the model parameters. The logistic regression model determines the parameters by maximizing the likelihood function, so that the difference between the actual label and the predicted probability is minimized. During the calculation process, the iterative reweighted least squares (IRLS) algorithm is used to update the parameters.

[0131] like Figure 8 As shown, the risk assessment model obtained by the analysis and training based on the training samples and the risk assessment task tree specifically includes: S421, analyzing the first training sample based on the general risk assessment task tree to obtain a first general risk assessment model; S422, analyzing the second training sample according to the first general risk assessment model to obtain a first association rule set between specific risk features and general risk features; S423, analyzing the second training sample based on the specific risk assessment task tree to obtain a first specific risk assessment model; S424, analyzing the first training sample according to the first specific risk assessment model to obtain a second association rule set between specific risk features and general risk features; S425, based on the first association rule set and the second association rule set of the specific risk characteristics and the general risk characteristics, the nodes of the general risk assessment task tree are optimized, and the first training sample is analyzed using the optimized general risk assessment task tree to obtain a final general risk assessment model; S426. Analyze the third training sample based on the dynamic risk identification task tree to obtain a dynamic risk assessment model.

[0132] In metadata management risk assessment, the purpose of generating association rule sets is to reveal and utilize the relationships between metadata features and improve the effectiveness and accuracy of the assessment model. In metadata management, the association between certain features may lead to risks in the management process. For example, "low metadata update frequency" may be related to "metadata inconsistency risk", while "improper metadata access rights setting" may be related to "data leakage risk". By generating association rule sets, these potential risks can be identified and understood. By identifying the associations between multiple metadata features, the model is able to provide a more comprehensive risk assessment. For example, unclear metadata field definition includes if the definition of metadata fields is unclear or inconsistent, which may lead to misunderstandings or incorrect use when sharing or querying data. This risk may affect the correct interpretation and use of data. Improper metadata version control includes if the metadata version control is improper (such as not clearly marking the version or not tracking the update history), which may lead to confusion between multiple versions of metadata during use, further increasing the complexity of management and use. Through the generated association rule set, the model can identify the dependency between these two features and prioritize their combined impact in risk assessment. For example, according to the rule set, it is found that "unclear field definition" often occurs together with "improper version control". Therefore, special attention should be paid to these feature combinations during risk assessment to improve the accuracy of predicting potential risks.

[0133] Through the association rule set, the model can adjust the weights of related features in the task tree, and can also use these rules to optimize the node connections in the task tree. With the emergence of new metadata management risks and data features, the model can iteratively optimize the risk assessment model by continuously updating the association rule set. For example, if the rule set shows that "unclear metadata field definition" and "improper version control" are significantly correlated in multiple assessments, the model can increase the weight of these two features in the assessment accordingly. For another example, if a strong correlation is found between "improper permission setting" and "missing access logs", the model can connect these feature nodes in the task tree. After each new rule set is generated, the model will re-evaluate the dependencies between features and adjust the task tree structure and decision rules so that the model always adapts to the latest management environment and risk situation.

[0134] The dynamic risk assessment model can effectively identify and assess potential risks in specific areas through in-depth analysis of metadata management behaviors and risk events, combined with scientific feature extraction and task tree structure, and provide important support for metadata management. The training samples include metadata management behavior records, such as collecting user management behavior records related to metadata and obtaining metadata status change time series information. This information includes the timestamp and content of operations such as data update, modification, and deletion. The training samples also include user behavior features extracted from user management behavior records, such as update frequency, operation type, user feedback, etc., and metadata management behavior sample data is formed based on user behavior features. The training samples also include risk event feature associations. For example, by analyzing historical risk events that have occurred, the metadata feature association semantics therein are extracted to form risk event sample data. This may include feature descriptions of risk events such as data loss and data format errors. The features extracted from the metadata management behavior sample data and the risk event sample data are merged with the samples in the first training sample set and the second training sample set to form a third training sample set. The features annotated by the dynamic risk assessment rules are extracted from the metadata feature matrix. Based on the extracted metadata features, the nodes of the initial task tree are constructed, and similar or related feature nodes are connected to form a preliminary task tree structure. The dynamic risk assessment model is trained using the random forest algorithm, and the third training sample set is input to evaluate its performance in identifying potential risks in the corn variety approval scientific data set, thereby identifying and evaluating the potential risks in the data set.

[0135] S5. Collect a dynamic verification set and optimize the risk assessment model.

[0136] Use the validation set to adjust hyperparameters to optimize model performance, and regularly use the validation set to evaluate the performance of the risk assessment model. Finally, use the test set to evaluate the performance of the risk assessment model, and measure the effectiveness of the risk assessment model with indicators such as accuracy, recall, and F1 score.

[0137] In this embodiment, the full life cycle management process data also includes user logs, system logs, and expert corpora involved in metadata management.

[0138] S6. Deploy risk assessment knowledge base and assessment feedback network based on specific risk assessment model and dynamic risk assessment model.

[0139] In this embodiment, if Fig. 9 As shown, step S6 specifically includes a feedback step, a knowledge extraction step and a correction step: Feedback step: Establish a verification reference corpus based on user logs, system logs, and expert corpora involved in metadata management; analyze the first training sample based on the feature risk assessment model and the dynamic risk assessment model to obtain rule node data; extract classification features from the second training sample and the third training sample, use the classification features to perform hierarchical clustering analysis to obtain hierarchical cluster centers, calculate the similarity between the hierarchical cluster centers and the verification reference corpora in the verification reference corpus, obtain verification corpora, and feed back the verification corpora to the stakeholders of the scientific data set. The stakeholders include metadata producers, data managers, data analysts, data consumers, and other users participating in the data ecology of the scientific data set.

[0140] Knowledge extraction: Based on the opinions of stakeholders in the validation corpus and scientific data sets, metadata management risk assessment domain knowledge is extracted to build a knowledge base.

[0141] Correction step: Correct the specific risk assessment rules and dynamic risk assessment rules according to the knowledge base, and optimize the training of the specific risk assessment model and dynamic risk assessment model.

[0142] The following is a description of the technical solution of the metadata management risk assessment model training method for scientific datasets provided in this application in conjunction with a specific embodiment.

[0143] like Fig.10 As shown, the embodiment of the present application further provides a metadata management risk assessment model training device 610. Optionally, the metadata management risk assessment model training device 610 may include: A risk assessment rule formulation module 611 is used to formulate risk assessment rules, wherein the risk assessment rules include general risk assessment rules, specific risk assessment rules and dynamic risk assessment rules; An acquisition module 612, configured to acquire a metadata feature matrix from metadata based on risk assessment rules; A risk assessment task tree generation module 613 is used to generate a risk assessment task tree based on risk assessment rules and risk assessment task characteristics; Model training module 614, used to construct training samples based on metadata, and obtain a risk assessment model based on the training samples and risk assessment task tree analysis and training; The optimization module 615 is used to construct a dynamic verification set based on the metadata, and optimize the risk assessment model using the dynamic verification set.

[0144] The metadata management risk assessment model training device 610 may also include a risk assessment and knowledge management module for deploying a risk assessment knowledge base and an assessment feedback network according to a specific risk assessment model and a dynamic risk assessment model.

[0145] It can be understood that the above-mentioned device embodiments and the above-mentioned method embodiments can correspond to each other, and similar descriptions of the device embodiments can refer to the method embodiments. To avoid repetition, it will not be repeated here. A metadata management risk assessment model training device provided in an embodiment of the present application can execute a method provided in any embodiment of the present application, and has functional modules and beneficial effects corresponding to the execution method. The functional modules of the metadata management risk assessment model training device can be implemented in the form of hardware, can be implemented by instructions in the form of software, and can also be implemented by a combination of hardware and software modules.

[0146] Specifically, each step of the method embodiment of the present application can be completed by the hardware integrated logic circuit and / or software instructions in the processor, and the steps of the metadata management risk assessment model training method of the scientific data set in the embodiment of the present application can be directly embodied as a hardware coding processor for execution, or a combination of hardware and software modules in the coding processor for execution. Optionally, the software module can be located in a random access memory, a read-only memory, a programmable read-only memory, a flash memory, an electrically erasable programmable memory, a register, or other storage media. The storage medium is located in the memory, and the processor reads the information in the memory and completes the steps in the above method embodiment in combination with its hardware.

[0147] Obviously, the above embodiments of the present application are merely examples for clearly illustrating the technical solution of the present application, and are not intended to limit the specific implementation methods of the present application. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the claims of the present application shall be included in the protection scope of the claims of the present application.

Claims

1. A method for training a metadata management risk assessment model for a scientific dataset, characterized in that: The method comprises: Formulate risk assessment rules, which include general risk assessment rules, specific risk assessment rules and dynamic risk assessment rules; Obtaining a metadata feature matrix from the metadata based on risk assessment rules; Generate a risk assessment task tree based on risk assessment rules and metadata feature matrix; Building training samples based on metadata, and obtaining a risk assessment model based on the training samples and the risk assessment task tree analysis and training; A dynamic verification set is constructed based on the metadata, and the risk assessment model is optimized using the dynamic verification set.

2. The method for training a metadata management risk assessment model for a scientific dataset according to claim 1, characterized in that: The general risk assessment rules are extracted from guidance documents; The development of specific risk assessment rules specifically includes: Obtain specific risk assessment tasks or requirements from metadata and extract characteristic keywords for specific risk assessment tasks or requirements; Converting feature keywords into word vectors and performing cluster analysis, and establishing a specific risk assessment feature matching index based on the word vectors and cluster analysis results; Features associated with and / or similar to general risk assessment rules and relationships between features are extracted according to the specific risk assessment feature matching index to establish specific risk assessment rules.

3. The method for training a metadata management risk assessment model for a scientific dataset according to claim 2, characterized in that: Formulating dynamic risk assessment rules includes: Extract feature vectors from metadata management behaviors and metadata data usage records; Performing cluster analysis on the feature vector to obtain a dynamic risk assessment prompt word set; Generate a dynamic risk assessment feature matching index based on the language model, the preset expert corpus and the prompt word set; Based on the dynamic risk assessment feature matching index, features associated with and / or similar to general risk assessment rules and relationships between features are extracted and dynamic risk assessment rules are established.

4. The method for training a metadata management risk assessment model for a scientific dataset according to claim 1, characterized in that: The metadata includes structured metadata and unstructured metadata, and the step of obtaining a metadata feature matrix from the metadata based on risk assessment rules includes: extracting metadata ontology features from the structured metadata based on the general risk assessment rules; extracting metadata implicit features from the unstructured metadata based on the general risk assessment rule and the specific risk assessment rule; Annotating the metadata ontology features and implicit features based on the dynamic risk assessment rules; The feature association between the metadata ontology feature and the implicit feature is obtained, and a metadata feature matrix for risk assessment is constructed according to the feature association and the annotation.

5. The method for training a metadata management risk assessment model for a scientific dataset according to claim 4, characterized in that: The acquiring of the feature association between the metadata ontology feature and the implicit feature specifically includes: Acquire metadata features of metadata, where types of the metadata features include structured and unstructured; cluster the metadata features according to semantic similarity of the metadata features; The metadata features in each cluster are classified according to the type of metadata features, and the metadata features of the same type in the same cluster are grouped into item sets for association rule mining; Converting metadata features in the item set into semantic vectors; The association rules of metadata ontology features and implicit features are extracted from the semantic vector using an association rule algorithm.

6. A method for training a metadata management risk assessment model for a scientific dataset according to any one of claims 4-5, characterized in that: The risk assessment task tree includes a general risk assessment task tree, a specific risk assessment task tree and a dynamic risk identification task tree; The generating of the risk assessment task tree based on the risk assessment rules and the metadata feature matrix includes: Construct a general risk assessment task tree based on general risk assessment rules and metadata feature matrix; Extract association rules between specific risk assessment rules and general risk assessment task tree nodes based on preset specifications, standards and practical experience, and establish a specific risk identification task tree based on the specific risk assessment rules and the associated general risk assessment task tree nodes; A dynamic risk identification task tree is established based on dynamic risk assessment rules and metadata feature matrix.

7. A method for training a metadata management risk assessment model for a scientific dataset according to any one of claims 1 to 5, characterized in that: The training samples include a first training sample set, a second training sample set, and a third training sample set. The step of constructing the training samples based on metadata specifically includes: Obtain metadata and full life cycle management process data of scientific data sets; the full life cycle management process data includes metadata quality review opinions and historical risk event records; Extract characteristic keywords and context sentences in metadata quality review opinions according to general risk assessment rules to obtain first review semantic features; classify the first review semantic features according to rule features, and score each type of first review semantic features through sentiment calculation to obtain a general risk score of the sample; Acquire a metadata management description according to the metadata and the full life cycle management process data, and construct a first training sample set based on the metadata management description, the first review semantic feature and the universal risk score; Extract characteristic keywords and context sentences in metadata quality review opinions according to specific risk assessment rules to obtain second review semantic features; classify the second review semantic features according to rule features, and score each type of second review semantic features through sentiment calculation to obtain a specific risk score of the sample; Constructing a second training sample set based on the metadata management description, the second review semantic feature, the general risk score and the specific risk score; According to the metadata management behavior record, the metadata state change time sequence information and the metadata management behavior characteristics are respectively obtained to obtain metadata management behavior sample data; the metadata feature association semantics in the risk event record are obtained to obtain risk event sample data; A third training sample set is constructed based on the metadata management behavior sample data, the risk event sample data, and samples in the first training sample set and the second training sample set.

8. The method for training a metadata management risk assessment model for a scientific dataset according to claim 7, characterized in that: The risk assessment model is obtained by analyzing and training the training samples and the risk assessment task tree, specifically comprising: Analyze the first training sample based on the general risk assessment task tree to obtain a general risk assessment model; Analyzing the second training sample according to the general risk assessment model to obtain a first association rule set between specific risk features and general risk features; Analyzing the second training sample based on the specific risk assessment task tree to obtain a specific risk assessment model; Analyzing the first training sample according to the specific risk assessment model to obtain a second association rule set between the specific risk features and the general risk features; Based on the first association rule set and the second association rule set of the specific risk characteristics and the general risk characteristics, the nodes of the general risk assessment task tree are optimized, and the first training sample is analyzed using the optimized general risk assessment task tree to obtain a final general risk assessment model; The third training sample is analyzed based on the dynamic risk identification task tree to obtain a dynamic risk assessment model.

9. The method for training a metadata management risk assessment model for a scientific dataset according to claim 8, characterized in that: The full life management process data also includes user logs, system logs, and expert corpora involved in metadata management; The method further comprises: Deploy risk assessment knowledge base and assessment feedback network according to specific risk assessment model and dynamic risk assessment model, including: Feedback step: establish a verification reference corpus based on user logs, system logs, and expert corpora involved in metadata management; analyze the first training sample based on the feature risk assessment model and the dynamic risk assessment model to obtain rule node data; extract classification features from the second training sample and the third training sample, use the classification features to perform hierarchical clustering analysis to obtain hierarchical cluster centers, calculate similarity between the hierarchical cluster centers and the verification reference corpora in the verification reference corpus to obtain verification corpora, and feed back the verification corpora to stakeholders of the scientific data set; the stakeholders include metadata producers, data managers, data analysts, and data consumers of the scientific data set; Knowledge extraction: Based on the opinions of stakeholders in the validation corpus and scientific data sets, metadata management risk assessment domain knowledge is extracted to build a knowledge base; Correction step: Correct the specific risk assessment rules and dynamic risk assessment rules according to the knowledge base, and optimize the training of the specific risk assessment model and dynamic risk assessment model.

10. A metadata management risk assessment model training device for scientific data sets, characterized in that: The device comprises: A risk assessment rule formulation module, used to formulate risk assessment rules, wherein the risk assessment rules include general risk assessment rules, specific risk assessment rules and dynamic risk assessment rules; An acquisition module, used for acquiring a metadata feature matrix from metadata based on risk assessment rules; A risk assessment task tree generation module is used to generate a risk assessment task tree based on risk assessment rules and risk assessment task characteristics; A model training module is used to construct a training sample based on metadata, and to obtain a risk assessment model based on the training sample and the risk assessment task tree analysis and training; The optimization module is used to construct a dynamic verification set based on the metadata, and optimize the risk assessment model using the dynamic verification set.

Images