A multi-level data modeling method and system based on a large language model

Abstract

The application provides a multi-level data modeling method based on a large language model, comprising the following steps: S1. data labeling; S2. data adaptation; S3. dictionary library construction; S4. access analysis; S5. relationship extraction; S6. entity conversion; S7. standardization processing; S8. fusion processing; S9. hierarchical classification; S10. classification calculation; S11. clustering calculation; S12. step-by-step generation; S13. completion and saving. The application also provides a multi-level data modeling system based on a large language model, comprising a data labeling and training unit, a data relationship extraction unit, a data standardization and fusion unit, a data model grouping aggregation unit and a data model grading unit. The application can effectively solve the problems of incomplete data model parameter coverage and long data model modeling process when modeling data from massive multi-source heterogeneous open source knowledge data, automatically complete multi-layer grading of the model, improve the data modeling efficiency, and effectively support data feature analysis and knowledge tracing.

Description

Technical Field

[0001] This invention relates to a multi-level data modeling method and system based on a large language model, belonging to the field of data processing technology. Background Technology

[0002] The publicly available military targets on the internet are diverse, primarily including individuals, organizations, weapons and equipment, reconnaissance sensors, and fixed facilities. Furthermore, with advancements in science and technology, new weapons and equipment are constantly emerging globally, involving a wide range of platforms, including land-based, sea-based, air-based, and space-based systems. According to publicly available data, there are numerous mainstream in-service weapons and equipment, numerous military units and paramilitary organizations, and hundreds of thousands of important figures worldwide, with over ten thousand subcategories. Such a diverse range of targets not only exhibits significant differences in individual performance parameters, but also suffers from a complex and inconsistent structure in the related knowledge data available on the internet. When conducting open-source military research, systematically sorting and analyzing target parameter information from massive amounts of multi-source heterogeneous data using manual methods would be extremely labor-intensive. Traditional NLP algorithms for target identification in open-source military knowledge data have low accuracy rates, and the required training and annotation datasets for improving accuracy are very large, often resulting in high manpower costs for data annotation. Furthermore, due to differences in usage scenarios and continuous updates and iterations with technological advancements, the same type of target has given rise to various series models with unique characteristics. Therefore, scientifically grouping targets into multiple categories and levels requires extensive data statistical analysis. Summary of the Invention

[0003] To address the aforementioned technical problems, this invention provides a multi-level data modeling method and system based on a large language model. This method and system effectively solves the problems of time-consuming sorting of target data models from massive amounts of multi-source heterogeneous open-source knowledge data and the low accuracy of knowledge recognition in traditional NLP algorithms. At the same time, it uses a hierarchical clustering algorithm to automatically complete the multi-level grouping calculation of the target model, solving the problem of needing manual data statistical analysis for hierarchical design.

[0004] The present invention is achieved through the following technical solutions.

[0005] This invention provides a multi-level data modeling method based on a large language model, comprising the following steps:

[0006] S1. Data annotation: Define open-source military domain target entity and entity relationship label groups, and relationship feature classification label groups, and complete the annotation of target dataset entities and entity relationships and entity relationship feature classification annotation;

[0007] S2. Data Adaptation: Using an open-source large language model as the foundation, incremental model training is performed using the open-source military domain target annotation dataset from step S1 to fine-tune the parameters of the original large language model, forming an enhanced large language algorithm model MFLLM suitable for the open-source military domain.

[0008] S3. Construct a dictionary library: Construct an open-source standard dictionary library for military target data, including standardized data model parameter names, standardized units of measurement, and unified units of measurement for standardized data model parameters, and save it to the basic data repository;

[0009] S4. Access and Parsing: Access and parse multi-source heterogeneous knowledge data containing multiple target objects, clean and convert it into text format data, and then save the text data results to the basic data repository;

[0010] S5. Relation Extraction: The enhanced large language algorithm model MFLLM is used to extract target entities and entity relations from the text format data in step S4, and the extraction results containing head entities, tail entities, entity relations and source data statements are obtained. The extraction results are then processed into an entity relation triplet aggregation set entryArray.

[0011] S6. Entity Conversion: Convert the data structure of the entity relation triplet aggregate collection entryArray to obtain entity object data entryObject;

[0012] S7. Standardization Processing: Convert the entity object data entryObject into the three elements of entity attributes and perform data standardization processing to obtain the entity parameter model parametricModel;

[0013] S8. Fusion Processing: Use a text similarity calculation algorithm to compare entity names, entity attribute names, attribute values ​​and units of measurement for entities of the same type, and perform fusion processing on the entity parametricModel;

[0014] S9. Hierarchical Classification: Cluster the entity parameter models according to entity type and add hierarchical classification information to each entity parameter model;

[0015] S10. Classification Calculation: Use the Enhanced Large Language Algorithm Model MFLLM to classify and group the entity attributes in the entity parameter model parametricModel, and add the grouping information to the entity parameter model parametricModel to obtain the entity parameter grouping model groupModel.

[0016] S11. Clustering Calculation: Perform correlation analysis on the entity parameter grouping model groupModel, and then perform a level 1 clustering calculation on the entity parameter grouping model groupModel using a hierarchical clustering algorithm to obtain the next level clusters. After generating names for the clusters using the enhanced large language algorithm model MFLLM, add the clustering results to the hierarchical classification information of the entity parameter grouping model.

[0017] S12. Generation by level: Repeat step S11 to generate multi-level model classifications, forming a multi-level data model system with inheritance relationships;

[0018] S13. Complete the saving: Store the multi-level data model system in the business data storage or save it as a configuration file for subsequent business use.

[0019] The entity and entity relationship tag group includes target entity classification, attribute entity classification, predefined relationship types between entities, and descriptive information for each type of tag; the relationship feature classification tag group includes relationship feature types and descriptive information.

[0020] The three elements of the entity attribute are name, value, and unit of measurement.

[0021] Step S1 specifically includes the following steps:

[0022] S1.1. Select the header entity text and mark its target entity type;

[0023] S1.2. Select the tail entity text and mark its target entity type or attribute entity type;

[0024] S1.3. If the text contains a relation name, select the relation text and mark it as a custom relation type; otherwise, directly select the predefined relation type between the head entity and the tail entity.

[0025] S1.4. Select a feature type for the labeled relationship. The text data and labeled data are transformed and merged to form an open-source military domain target labeling dataset, which is then saved to the basic data repository.

[0026] In step S3, the standardized data model parameter name includes the standardized data model parameter name paramName and a set of synonymous aliases paramName'; the standardized unit of measurement includes the standardized unit of measurement measureUnit and a set of synonymous symbols measureUnit'; the standardized data model parameter unified unit of measurement corresponds one-to-one with the standardized data model parameter.

[0027] In step S4, the sources of multi-source heterogeneous knowledge data include files, databases, and interfaces; among them, file content is accessed through text reading, database fields are queried through SQL statements, and interface input parameters are processed through interface integration.

[0028] In step S5, the extraction results include head entity, tail entity, entity relationship and source data statement, forming a hash set of entity relationship triples. Then, triples with the same head entity name and head entity type are grouped and merged to obtain an aggregate set of relationship triples for each entity, entryArray.

[0029] In step S7, the data standardization process specifically includes the following steps:

[0030] S7.1. Standardization of entity attribute names: By comparing with the data standard dictionary, find the predefined data model parameter alias 'paramName' that exactly matches the entity attribute name, and replace the entity attribute name with the standardized data model parameter name 'paramName' corresponding to the alias;

[0031] S7.2. Separation of entity attribute value and unit of measurement: When an entity attribute contains an attribute value and a unit of measurement, the attribute value and the unit of measurement shall be separated.

[0032] S7.3. Standardization of Entity Attribute Measurement Units: By comparing with the data standard dictionary, find the synonym 'measureUnit' that perfectly matches the entity attribute measurement unit, and replace the entity attribute measurement unit with the standardized measurement unit 'measureUnit' corresponding to the synonym.

[0033] S7.4. Unification of Entity Attribute Measurement Units: First, unify the measurement units by comparing the standardized data model parameters in the data standard dictionary to find the unified measurement unit for the entity attribute. Then, calculate the conversion ratio between the current measurement unit and the unified measurement unit. Finally, perform proportional scaling calculation on the attribute value.

[0034] The fusion process in step S8 involves removing redundant entity parameter models and redundant entity attributes within the models, and concentrating the different entity attributes of multiple models with the same entity parameter.

[0035] In step S10, the process of obtaining the entity parameter grouping model (groupModel) specifically includes the following steps:

[0036] S10.1. Collect the entity attributes of all entity parameter models. Iterate through each entity parameter model in step S9, collect all entity attribute names in the form of a set, and then remove duplicate entity attribute names in the set to form a complete set of entity attributes;

[0037] S10.2. Classify and calculate the names of each entity attribute in the full set of entity attributes. Using the entity relation classification algorithm provided by the Enhanced Large Language Algorithm Model (MFLLM) in step S2, classify and calculate the names of each entity attribute in the full set of entity attributes. The entity attribute names and classification results are stored in the form of key-value pairs to form an entity attribute classification mapping table.

[0038] S10.3. Apply the entity attribute classification mapping table to the entity parameter model, add grouping information, and form an entity parameter grouping model: Match the entity attribute names in the entity parameter model parametricModel in step S9 with the entity attribute names in the entity attribute classification mapping table. After a successful match, add a grouping structure to the entity parameter model parametricModel. The group name is the category name. Associate each entity attribute in the entity parameter model parametricModel with the corresponding group.

[0039] This invention also provides a multi-level data modeling system based on a large language model, including a data annotation and training unit, a data relationship extraction unit, a data standardization and fusion unit, a data model grouping and aggregation unit, and a data model hierarchical unit, for implementing the multi-level data modeling method based on a large language model as described above, wherein...

[0040] The data annotation and training unit is used to define label groups such as entities and entity relationships, to annotate open-source military text knowledge data, and to incrementally train the open-source large language model using the annotated dataset to form an enhanced large language algorithm model.

[0041] The data relationship extraction unit is used to extract entities and entity relationships from open-source military domain text knowledge data as a preliminary structured basic dataset;

[0042] The data standardization and fusion unit is used to standardize and replace the attribute names, attribute values ​​and attribute units of an entity, and to fuse data of the same entities and attributes to form a standardized data parameter model.

[0043] The data model grouping and aggregation unit is used to group entity attributes in the standardized data parameter model, and aggregate the model according to the entity classification defined by entity and entity relationship label groups, adding classification and grouping information to the model to form a data grouping model;

[0044] The data model hierarchical unit is used to perform clustering calculations on the data model clusters under the last-level classification, and gradually split the data model clusters downward to abstract a data hierarchical model with multi-level inheritance relationship.

[0045] The beneficial effects of this invention are: it can effectively solve the problems of incomplete data model parameter coverage and long data modeling time when performing data modeling from massive multi-source heterogeneous open source knowledge data, automatically complete multi-level model hierarchies, improve data modeling efficiency, and effectively support data feature analysis and knowledge tracing. Attached Figure Description

[0046] Figure 1 This is a schematic diagram of module connection and data flow according to at least one embodiment of the present invention. Detailed Implementation

[0047] The technical solution of the present invention is further described below, but the scope of protection is not limited to what is described.

[0048] Example 1

[0049] This invention provides a multi-level data modeling method based on a large language model, comprising the following steps:

[0050] S1. Data annotation: Define open-source military domain target entity and entity relationship label groups, and relationship feature classification label groups, and complete the annotation of target dataset entities and entity relationships and entity relationship feature classification annotation;

[0051] S2. Data Adaptation: Using an open-source large language model as the foundation, incremental model training is performed using the open-source military domain target annotation dataset from step S1 to fine-tune the parameters of the original large language model, forming an enhanced large language algorithm model MFLLM suitable for the open-source military domain.

[0052] S3. Construct a dictionary library: Construct an open-source standard dictionary library for military target data, including standardized data model parameter names, standardized units of measurement, and unified units of measurement for standardized data model parameters, and save it to the basic data repository;

[0053] S4. Access and Parsing: Access and parse multi-source heterogeneous knowledge data containing multiple target objects, clean and convert it into text format data, and then save the text data results to the basic data repository;

[0054] S5. Relation Extraction: The enhanced large language algorithm model MFLLM is used to extract target entities and entity relations from the text format data in step S4, and the extraction results containing head entities, tail entities, entity relations and source data statements are obtained. The extraction results are then processed into an entity relation triplet aggregation set entryArray.

[0055] S6. Entity Conversion: Convert the data structure of the entity relation triplet aggregate collection entryArray to obtain entity object data entryObject;

[0056] S7. Standardization Processing: Convert the entity object data entryObject into the three elements of entity attributes and perform data standardization processing to obtain the entity parameter model parametricModel;

[0057] S8. Fusion Processing: Use a text similarity calculation algorithm to compare entity names, entity attribute names, attribute values ​​and units of measurement for entities of the same type, and perform fusion processing on the entity parametricModel;

[0058] S9. Hierarchical Classification: Cluster the entity parameter models according to entity type and add hierarchical classification information to each entity parameter model;

[0059] S10. Classification Calculation: Use the Enhanced Large Language Algorithm Model MFLLM to classify and group the entity attributes in the entity parameter model parametricModel, and add the grouping information to the entity parameter model parametricModel to obtain the entity parameter grouping model groupModel.

[0060] S11. Clustering Calculation: Perform correlation analysis on the entity parameter grouping model groupModel, and then perform a level 1 clustering calculation on the entity parameter grouping model groupModel using a hierarchical clustering algorithm to obtain the next level clusters. After generating names for the clusters using the enhanced large language algorithm model MFLLM, add the clustering results to the hierarchical classification information of the entity parameter grouping model.

[0061] S12. Generation by level: Repeat step S11 to generate multi-level model classifications, forming a multi-level data model system with inheritance relationships;

[0062] S13. Complete the saving: Store the multi-level data model system in the business data storage or save it as a configuration file for subsequent business use.

[0063] Example 2

[0064] Based on Example 1, the entity and entity relationship tag group includes target entity classification, attribute entity classification, predefined relationship types between entities, and descriptive information of various tags; the relationship feature classification tag group includes relationship feature types and descriptive information.

[0065] Furthermore, the three elements of an entity attribute are name, value, and unit of measurement.

[0066] Furthermore, in step S3, the standardized data model parameter name includes the standardized data model parameter name paramName and a set of synonymous aliases paramName'; the standardized unit of measurement includes the standardized unit of measurement measureUnit and a set of synonymous symbols measureUnit'; the standardized data model parameter unified unit of measurement corresponds one-to-one with the standardized data model parameter.

[0067] Furthermore, in step S4, the sources of multi-source heterogeneous knowledge data include files, databases, and interfaces; among them, file content is accessed through text reading, database fields are queried through SQL statements, and interface input parameters are processed through interface integration.

[0068] Furthermore, in step S5, the extraction results include head entity, tail entity, entity relationship and source data statement, forming a hash set of entity relationship triples. Then, triples with the same head entity name and head entity type are grouped and merged to obtain an aggregate set of relationship triples for each entity, entryArray.

[0069] Example 3

[0070] Based on Example 1, step S1 specifically includes the following steps:

[0071] S1.1. Select the header entity text and mark its target entity type;

[0072] S1.2. Select the tail entity text and mark its target entity type or attribute entity type;

[0073] S1.3. If the text contains a relation name, select the relation text and mark it as a custom relation type; otherwise, directly select the predefined relation type between the head entity and the tail entity.

[0074] S1.4. Select a feature type for the labeled relationship. The text data and labeled data are transformed and merged to form an open-source military domain target labeling dataset, which is then saved to the basic data repository.

[0075] Example 4

[0076] Based on Example 1, step S7, the data standardization process specifically includes the following steps:

[0077] S7.1. Standardization of entity attribute names: By comparing with the data standard dictionary, find the predefined data model parameter alias 'paramName' that exactly matches the entity attribute name, and replace the entity attribute name with the standardized data model parameter name 'paramName' corresponding to the alias;

[0078] S7.2. Separation of entity attribute value and unit of measurement: When an entity attribute contains an attribute value and a unit of measurement, the attribute value and the unit of measurement shall be separated.

[0079] S7.3. Standardization of Entity Attribute Measurement Units: By comparing with the data standard dictionary, find the synonym 'measureUnit' that perfectly matches the entity attribute measurement unit, and replace the entity attribute measurement unit with the standardized measurement unit 'measureUnit' corresponding to the synonym.

[0080] S7.4. Unification of Entity Attribute Measurement Units: First, unify the measurement units by comparing the standardized data model parameters in the data standard dictionary to find the unified measurement unit for the entity attribute. Then, calculate the conversion ratio between the current measurement unit and the unified measurement unit. Finally, perform proportional scaling calculation on the attribute value.

[0081] Example 5

[0082] Based on Example 1, the fusion process in step S8 is to remove redundant entity parameter models and redundant entity attributes within the models, and to concentrate the different entity attributes of multiple models with the same entity parameter.

[0083] Furthermore, in step S10, the process of obtaining the entity parameter grouping model (groupModel) specifically includes the following steps:

[0084] S10.1. Collect the entity attributes of all entity parameter models. Iterate through each entity parameter model in step S9, collect all entity attribute names in the form of a set, and then remove duplicate entity attribute names in the set to form a complete set of entity attributes;

[0085] S10.2. Classify and calculate the names of each entity attribute in the full set of entity attributes. Using the entity relation classification algorithm provided by the Enhanced Large Language Algorithm Model (MFLLM) in step S2, classify and calculate the names of each entity attribute in the full set of entity attributes. The entity attribute names and classification results are stored in the form of key-value pairs to form an entity attribute classification mapping table.

[0086] S10.3. Apply the entity attribute classification mapping table to the entity parameter model, add grouping information, and form an entity parameter grouping model: Match the entity attribute names in the entity parameter model parametricModel in step S9 with the entity attribute names in the entity attribute classification mapping table. After a successful match, add a grouping structure to the entity parameter model parametricModel. The group name is the category name. Associate each entity attribute in the entity parameter model parametricModel with the corresponding group.

[0087] Example 6

[0088] Based on the above embodiments, intelligent multi-level data modeling of targets such as open-source military personnel, organizations, weapons and equipment, and reconnaissance sensors based on a large language model includes the following steps:

[0089] (1) Define open-source military domain target entity and entity relationship label groups and relationship feature classification label groups to complete the target dataset entity and entity relationship annotation and entity relationship feature classification annotation. The entity and entity relationship label groups include target entity classification, attribute entity classification, predefined relationship types between entities and description information of various labels. The relationship feature classification label groups include relationship feature types and description information, and are saved to the basic data repository. When selecting open-source military domain text knowledge data containing target information for data annotation, the main process is as follows: ① Select the head entity text and annotate its target entity type; ② Select the tail entity text and annotate its target entity type or attribute entity type; ③ If the text contains a relationship name, select the relationship text and annotate it as a custom relationship type; otherwise, directly select the predefined relationship type between the head entity and the tail entity; ④ Select a feature type to which the annotated relationship belongs. The text data and the annotated data are converted and merged to form an open-source military domain target annotation dataset, which is saved to the basic data repository.

[0090] (2) Adapting the large language model to open-source military domain knowledge data. Using the open-source large language model as a foundation, incremental model training was performed using the open-source military domain target annotation dataset from step 1 to fine-tune the parameters of the original large language model, forming an enhanced large language algorithm model MFLLM (MilitaryField LLM) suitable for the open-source military domain. The MFLLM model was loaded using a dedicated algorithm runtime environment, and entity and entity relationship extraction interfaces were extended through secondary development to form the ability to extract targets from the open-source military domain. Using predefined relationship types, custom relationship types, relationship feature classification label groups, and other annotation data, a relationship classification interface was extended through secondary development to form the ability to classify target entity relationships.

[0091] (3) Construct an open-source military target data standard dictionary library and save it to the basic data repository. The data standard dictionary mainly covers common attribute parameters of targets in the open-source military field, and provides three types of data: ① Standardized data model parameter name paramName and a set of synonyms paramName'; ② Standardized unit of measurement measureUnit and a set of synonyms measureUnit'; ③ Standardized unified unit of measurement for data model parameters, with each standardized parameter corresponding to only one unified unit of measurement.

[0092] (4) Complete the access and parsing of multi-source heterogeneous knowledge data on targets such as open-source military weapons, equipment, and reconnaissance sensors. After cleaning, the data is converted into text format and saved to the basic data storage library. The data sources include files, databases, and interfaces. The open-source military knowledge data is converted into text format by reading file content through text, querying database fields through SQL statements, and processing interface input parameters through interface integration, thus constructing a text-based domain dataset.

[0093] (5) Based on the enhanced large language model MFLLM in step 2, target entities (hereinafter referred to as entities) such as weapons and equipment and reconnaissance sensors in the text domain dataset in step 4 are extracted and entity relations are extracted. The extraction results include head entities, tail entities, entity relations and source data statements, forming a hash set of entity relation triples. Then, triples with the same head entity name and head entity type are grouped and merged to obtain the relation triple aggregation set entryArray for each entity.

[0094] (6) Perform data structure transformation on the entity relation triplet aggregation set entryArray in step 5 to generate entity object data. Convert each entity relation triplet data into an entity attribute key-value object. The key-value object contains the entity attribute name (triple relation name), the entity attribute value (triple tail entity), and the source data object. In this way, the triples in the triplet aggregation set of each entity are transformed into data structures. Finally, all entity attribute key-value objects of each entity are merged to become the entity object data entryObject.

[0095] (7) First, convert the key-value object data structure of each entity attribute in the entryObject data of step 6 into an entity attribute three-element (name, value, unit of measurement) object data structure. Second, based on the data standard dictionary in step 3, perform data standardization processing on the entity attribute three-element object data to form the entity parameter model parametricModel. The data standardization process consists of the following four steps:

[0096] A. Standardization of entity attribute names. By comparing with the data standard dictionary, find the predefined data model parameter alias 'paramName' that exactly matches the entity attribute name, and replace the entity attribute name with the standardized data model parameter name 'paramName' corresponding to the alias;

[0097] B. Separating Entity Attribute Values ​​and Units of Measurement. When an entity attribute contains both an attribute value and a unit of measurement, the attribute value and unit of measurement are separated.

[0098] C. Standardization of entity attribute measurement units. By comparing with the data standard dictionary, find the synonym 'measureUnit' that perfectly matches the entity attribute measurement unit, and replace the entity attribute measurement unit with the standardized measurement unit corresponding to the synonym;

[0099] D. Unified Units of Measurement for Entity Attributes. To build a universal data model, the units of measurement for entity attribute names need to be unified. First, the unified units of measurement for the entity attribute are found by comparing them with the standardized data model parameters in the data standard dictionary. Then, the conversion ratio between the current unit of measurement and the unified unit of measurement is calculated. Finally, the attribute values ​​are scaled proportionally.

[0100] (8) Use a text similarity calculation algorithm to compare entity names, entity attribute names, attribute values ​​and units of measurement for entities of the same type. Automatically fuse the entity parameter models in step 7, remove redundant entity parameter models and redundant entity attributes within the models, and concentrate the differences in entity attributes of multiple identical entity parameter models.

[0101] (9) Use the entity type defined by the entity and entity relationship label group in step 1 as the first level classification of the model system, cluster the entity parameter model in step 8 according to the entity type, and add hierarchical classification information to each entity parameter model.

[0102] (10) Based on the entity relation classification algorithm provided by the enhanced large language model in step 2, classify and calculate the entity attributes of the entity parameter model in step 9, group the entity attributes according to the classification calculation results, and add grouping information to the entity parameter model to form the entity parameter grouping model groupModel. The entity parameter grouping model generation process consists of the following 3 steps:

[0103] A. Collect the entity attributes of all entity parameter models. Iterate through each entity parameter model in step 9, collect all entity attribute names in the form of a set, and then remove duplicate entity attribute names in the set to form a complete set of entity attributes;

[0104] B. Classify and calculate the names of each entity attribute in the full set of entity attributes. Using the entity relation classification algorithm provided by the enhanced large language model in step 2, classify and calculate the names of each entity attribute in the full set of entity attributes. The entity attribute names and classification results are stored in the form of key-value pairs to form an entity attribute classification mapping table;

[0105] C. Apply the entity attribute classification mapping table to the entity parameter model, adding grouping information to form an entity parameter grouping model. Match each entity attribute name in the entity parameter model from step 9 with the entity attribute names in the entity attribute classification mapping table. If a match is successful, add a grouping structure to the entity parameter model, with the group name being the category name, and associate each entity attribute in the model with its corresponding group.

[0106] (11) Based on the first-level model classification clustering results in step 9, the entity parameter grouping model groupModel in step 10 is subjected to correlation analysis according to different model classifications. The entity parameter grouping model is subjected to a level 1 clustering calculation by hierarchical clustering algorithm to obtain the next level cluster. The enhanced large language model in step 2 is used to generate a name for each cluster. Finally, the clustering results are added to the hierarchical classification information of the entity parameter grouping model.

[0107] (12) Based on actual business needs, repeat the processing method of step 11 to generate more levels of model classification and form a multi-level data model system with inheritance relationship.

[0108] (13) At this point, the intelligent multi-level data modeling of the target such as open-source military weapons and equipment and reconnaissance sensors is completed. The completed multi-level data model is stored in the business data storage library or saved as a configuration file for subsequent business use.

[0109] Example 7

[0110] like Figure 1 The multi-level data modeling system based on a large language model, as shown, includes a data annotation and training unit, a data relationship extraction unit, a data standardization and fusion unit, a data model grouping and aggregation unit, and a data model hierarchical unit, used to implement the above embodiments, wherein...

[0111] The data annotation and training unit is used to define label groups such as entities and entity relationships, to annotate open-source military text knowledge data, and to incrementally train the open-source large language model using the annotated dataset to form an enhanced large language algorithm model.

[0112] The data relationship extraction unit is used to extract entities and entity relationships from open-source military domain text knowledge data as a preliminary structured basic dataset;

[0113] The data standardization and fusion unit is used to standardize and replace the attribute names, attribute values ​​and attribute units of an entity, and to fuse data of the same entities and attributes to form a standardized data parameter model.

[0114] The data model grouping and aggregation unit is used to group entity attributes in the standardized data parameter model, and aggregate the model according to the entity classification defined by entity and entity relationship label groups, adding classification and grouping information to the model to form a data grouping model;

[0115] The data model hierarchical unit is used to perform clustering calculations on the data model clusters under the last-level classification, and gradually split the data model clusters downward to abstract a data hierarchical model with multi-level inheritance relationship.

[0116] As can be seen from the above, the method for multi-level data modeling of targets such as individuals, institutions, weapons, and reconnaissance sensors in the open-source military field leverages the semantic understanding and generalization capabilities of large language models. This effectively solves the problems of time-consuming task of sifting through massive amounts of multi-source heterogeneous open-source knowledge data to create target data models, and the low accuracy of traditional NLP algorithms in knowledge recognition. Furthermore, the hierarchical clustering algorithm automatically completes multi-level grouping calculations for the target model, eliminating the need for manual data statistical analysis for hierarchical design. It also supports data tracing of target attributes and effectively supports various research institutions, think tanks, universities, and media outlets in data governance and feature analysis of open-source military targets. It is readily understood that this invention is also applicable to multiple general fields beyond open-source military targets, such as instruments and equipment, automobiles, consumer electronics, and smart homes.

[0117] Example 8

[0118] Based on the above embodiments, text data annotation tools (such as BRAT) are used to create open-source military domain target entity and entity relationship label groups and relationship feature classification label groups. A batch of open-source military domain text knowledge data is imported into the annotation software as a corpus to be annotated. The defined label groups are used to annotate the text corpus by head entities, tail entities, entity relationships and relationship classifications. Finally, the annotated data is published as an open-source military domain target annotation dataset to support training. The dataset is stored using a relational database (such as MySQL).

[0119] The text annotation work includes two processes: intelligent annotation and feedback correction. First, the text corpus is automatically annotated using an open-source large language model to reduce the workload of manual annotation. Second, the annotation results are checked and corrected manually to generate a high-quality annotated dataset.

[0120] The entity and entity relationship tag groups are shown in the following example:

[0121] Entity type:

[0122] {"name":"Airborne Target","description":"Equipment that flies within the atmosphere relying on air lift"} A predefined relationship between entity type and attribute type: {"headType":"Airborne Target","endType":"Date and Time","relation":"Service Date"} Example of a relationship feature classification tag group: [{"name":"Basic Information","description":"Basic characteristics of the target, including external dimensions and weight"},{"name":"Combat Capabilities","description":"Tactical characteristics of the target, such as strike range, payload"}]

[0123] Example 9

[0124] Based on the above embodiments, using the ChatGLM3-6B large language model open-sourced by Tsinghua University as a foundation, incremental model training was performed using the open-source military domain target annotation dataset in Embodiment 8 to complete the parameter fine-tuning of the ChatGLM3-6B large language model, generating an enhanced large language algorithm model MFLLM suitable for the open-source military domain. The PyThorch library of Python was used as a dedicated runtime environment to load the MFLLM model. Through secondary development based on the Python language and encapsulation using the FastAPI library, an entity and entity relationship extraction interface was implemented, forming the ability to extract targets in the open-source military domain. Using the same method as described above, a relationship classification interface was extended through secondary development, forming the ability to classify target entity relationships.

[0125] Example 10

[0126] Based on the above embodiments, by analyzing common target parameters in the open-source military field, a data standard dictionary library is formed, which includes standardized data model parameter names, standardized units of measurement, and unified units of measurement for standardized data model parameters, and is stored using a relational database (such as MySQL).

[0127] The standardized data model parameter names include a standard name `paramName` and multiple synonymous aliases `paramName'`, forming a one-to-many relationship. The standardized units of measurement include a standard unit name `measureUnit` and multiple synonymous names `measureUnit'`, forming a one-to-many relationship. The standardized data model parameter unified units of measurement include a standard name and a standard unit, forming a one-to-one relationship.

[0128] An example of a standardized data model parameter name is as follows:

[0129] {"standard":"Service Date","alias":["Service Time","Date of Deployment","Date of Deployment"]}

[0130] An example of a standardized unit of measurement data is as follows:

[0131] {"standard":"kilometer","alias":["kilometer","kM"]}

[0132] An example of standardized data model parameters with unified units of measurement is shown below:

[0133] {"name":"Maximum Range","unit":"km"}

[0134] Example 11

[0135] Based on the above embodiments, multi-source heterogeneous knowledge data related to business, including open-source military weapons and equipment and reconnaissance sensors, are parsed into text information and saved into a relational database to form a domain dataset in text form.

[0136] The data parsing process, taking common file, database, and API data formats as examples, is as follows: For file data, a text reading tool is used to obtain the file content; for database data, SQL queries are used to access it, and relevant data fields in the query results are merged into a text message; for API data sources, API calls are used to obtain the returned data packets. Taking JSON format as an example, the data packet is first converted into a JSON object, and then relevant data fields are selected and merged into a text message. Further, a series of text cleaning and standardization methods, including removing duplicate records, filling missing values, and removing outliers, are used to preprocess the original text data, and the results are saved to a distributed database (e.g., HBase).

[0137] Example 12

[0138] Based on Example 11, the domain dataset in text form is traversed and accessed. The entity and entity relation extraction interface based on the enhanced large language model MFLLM extended in Example 9 is called to extract target entities such as weapons and equipment and reconnaissance sensors from each piece of text data, and to extract entity relations. Each entity relation triplet of the extraction result is stored in a set. After all entity relations are extracted and stored, the set of entity relation triplets is traversed and accessed. The header entity name and header entity type of each entity relation triplet are compared. Triplets with the same name and type are stored in the same set, forming multiple triplet sets grouped and aggregated by entity name. Finally, a Map data structure is constructed, using the key to store the entity name and the value to store the aggregated triplet set, forming the triplet aggregation set entryArray.

[0139] The entity relationship triple contains three core elements and one source-tracing auxiliary element: head entity name and type, relationship name, tail entity name and type, and source data statement. These are stored using an array, as shown in the following example:

[0140] [["Yak-130 Advanced Trainer Jet","Airborne Target"],"Service Date",["February 2010","Date and Time"],["Yak-130 Advanced Trainer Jet Helps Russian Air Force Train Third-Generation Fighter Pilots","According to Russian media reports, the Yak-130 advanced trainer jet officially entered service in February 2010, and the first batch of 20 aircraft has been delivered."]

[0141] "]]

[0142] The following is an example of the Map data structure for the entity relationship triplet aggregation set:

[0143] {"Yak-130 Advanced Trainer":{[["Yak-130 Advanced Trainer","Airborne Target"],"Service Date",["February 2010","Date and Time"],[(Source Data Statement)]],[["Yak-130 Advanced Trainer","Airborne Target"],"Maximum Range",["500 Miles","Space Distance"],[(Source Data Statement)]]}}

[0144] Example 13

[0145] Based on Example 12, the entity relation triplet aggregation set `entryArray` in Example 12 undergoes data structure transformation to generate entity object data. Each entity relation triplet is converted into an entity attribute key-value object. The key-value object contains a Key entity attribute name (triple relation name), a Value entity attribute value (triple object name), and a source data object. This method transforms the data structure of triples within the triplet aggregation set for each entity. Finally, all entity attribute key-value objects for each entity are merged using JSON data format. Simultaneously, the entity name and entity type are each added as an entity attribute key-value object to the JSON data, resulting in the entity object data `entryObject`.

[0146] Taking the example data in Example 12 as an example, the entity object data after data structure transformation is as follows:

[0147] {"Entity Name":"Yak-130 Advanced Trainer Jet","Entity Type":"Airborne Target","Service Date":{"value":"February 2010","source":{"article":"Yak Advanced Trainer Jet Helps…","sentence":"According to Russian media reports…"}},"Maximum Range":{"value":"500 miles","source":{"article":"Modern Weapons Issue 52","sentence":"The Yakovlev Design Bureau pointed out at the press conference…"}}}

[0148] Example 14

[0149] Based on Example 13, firstly, the key-value object data structure of each entity attribute within the entity object data entryObject in Example 13 is converted into an entity attribute three-element (name, value, unit of measurement) object data structure; secondly, based on the data standard dictionary in Example 10, the entity attribute three-element object data is standardized to form an entity parameter model parametricModel. The data standardization process consists of the following four steps:

[0150] A. Standardization of entity attribute names. If an entity attribute name is not a standardized name, the alias 'paramName' that exactly matches the entity attribute name is found by comparing it with the predefined data model parameter alias group in the data standard dictionary, and the entity attribute name is replaced with the standardized data model parameter name 'paramName' corresponding to that alias.

[0151] B. Separating Entity Attribute Values ​​and Units of Measurement. When an entity attribute contains both an attribute value and a unit of measurement, the attribute value and unit of measurement are separated.

[0152] C. Standardization of entity attribute measurement units. If the entity attribute measurement unit is not a standardized measurement unit, then by comparing the synonym group in the data standard dictionary, find the synonym symbol 'measureUnit' that completely matches the entity attribute measurement unit, and replace the entity attribute measurement unit with the standardized measurement unit 'measureUnit' corresponding to the synonym symbol;

[0153] D. Unify the unit of measurement for entity attributes. First, compare the entity attribute name with the parameter name of the unified unit of measurement in the data standard dictionary to find the unified unit of measurement for the entity attribute. Then, calculate the conversion ratio between the current unit of measurement and the unified unit of measurement. Finally, perform proportional scaling calculation on the attribute value.

[0154] The following is an example of a JSON format for the entity attribute three-element object data structure:

[0155] {"name":"Maximum Range","value":500,"unit":"miles","source":{(Source Data Object)}}

[0156] Taking the example entity object data in Example 13 as an example, the entity parameter model object data after standardization is as follows:

[0157] {{"name":"Entity Name","value":"Yak-130 Advanced Trainer Aircraft","unit":null},{"name":"Entity Type","value":"Airborne Target","unit":null},{"name":"Service Date","value":"February 2010","unit":null,"source":{(Source Data Object)}},{"name":"Maximum Range","value":804.672,"unit":"km","source":{(Source Data Object)}}}

[0158] Example 15

[0159] Based on the above embodiments, text similarity calculation algorithms (such as word2vec and SimCSE) are used to compare entity names, entity attribute names, attribute values ​​and units of measurement for entities of the same type. The entity parameter model parametricModel in Embodiment 14 is automatically fused to remove redundant entity parameter models and redundant entity attributes within the model, and to concentrate different entity attributes of the same entity parameter model.

[0160] The similarity comparison process first compares multiple entity parameter models by using the word2vec algorithm to compare entity names and find multiple sets of entity parameter model objects that may be the same entity. Then, the SimCSE algorithm compares the three elements of entity attributes of the entity parameter models. When there are one or more identical entity attributes and there are no conflicts in entity attribute values ​​and units of measurement that cannot be merged, it is determined that the two entity parameter models being compared can be automatically merged.

[0161] The entity attribute fusion process first selects the entity parameter model with a large number of entity attributes as the base entity parameter model; second, it merges the unique entity attributes of other entity parameter models into the base entity parameter model; finally, for duplicate entity attributes, it selects the entity attribute with more detailed attribute value as the fusion result, and other cases are not processed.

[0162] Example 16

[0163] Using the entity type defined by the entity and entity relationship tag group as the first-level classification of the data model system, the entity parameter model in Example 15 is clustered and grouped according to entity type, and a hierarchical classification information field is added to the object data of each entity parameter model.

[0164] Taking the example entity parameter model object data in Example 14 as an example, the model information after adding the classification information field is as follows:

[0165] {"classification":{"name":"Airborne Target","child":null},"attributes":{{"name":"Entity Name","value":"Yak-130 Advanced Trainer","unit":null},{"name":"Service Date","value":"February 2010","unit":null,"source":{(Source Data Object)}},{"name":"Maximum Range","value":804.672,"unit":"km","source":{(Source Data Object)}}}}

[0166] Example 17

[0167] Using the entity relation classification algorithm interface provided by the enhanced large language model, the entity attributes of the entity parameter model in Example 16 are classified and calculated. Entity attributes are grouped according to the classification calculation results, and a grouping information field is added to the entity parameter model to form the entity parameter grouping model (groupModel). The entity parameter grouping model generation process consists of the following three steps:

[0168] A. Collect the entity attributes of all entity parameter models. Iterate through each entity parameter model in Example 16, use a List collection data structure to collect all entity attribute names, and remove duplicate entity attribute names in the collection to form a complete set of entity attributes;

[0169] B. Classify and calculate the names of each entity attribute in the full set of entity attributes. Using the entity relationship classification algorithm interface provided by the enhanced large language model, classify and calculate the names of each entity attribute in the full set of entity attributes. Use a Map data structure to store the entity attribute name Key and the classification result Value in the form of key-value objects to form an entity attribute classification mapping table;

[0170] C. Apply the entity attribute classification mapping table to the entity parameter model, adding a grouping information field to the model to form an entity parameter grouping model. Match the names of each entity attribute in the entity parameter model in Example 16 with the entity attribute names in the entity attribute classification mapping table. If a match is successful, add a grouping structure field to the entity parameter model. The group name is the classification result name, associating each entity attribute in the model with the corresponding group.

[0171] The newly added grouping structure field in the entity parameter model includes the group name (groupName) and the list of entity attributes (attributes) it contains. Taking the example entity parameter model object data in Example 16 as an example, the corresponding entity parameter grouping model object data is as follows:

[0172] {"classification":{"name":"Airborne Target","child":null},"groups":[{"groupName":"Basic Information","attributes":[{"name":"Entity Name","value":"Yak-130 Advanced Trainer","unit":null},{"name":"Service Date","value":"February 2010","unit":null,"source":{(Source Data Object)}}]},{"groupName":"Combat Capability","attributes":[{"name":"Maximum Range","value":804.672,"unit":"km","source":{(Source Data Object)}}]}]}

[0173] Example 18

[0174] Based on the first-level model classification clustering results in Example 16, correlation analysis is performed on the entity parameter grouping model (groupModel) in Example 17 according to different model classifications. The DIANA split hierarchical clustering algorithm is used to perform a downward clustering calculation with a depth of 1 level to obtain the next-level model classification clusters. For the entity parameter grouping model name within each final-level cluster, word segmentation and counting are performed. The top three most frequent words are extracted, and an enhanced large language model is used to fuse these three words to generate a new word as the cluster name. If duplicate words are generated, an incrementing Arabic numeral is used as a suffix for differentiation. Finally, the clustering results are added to the hierarchical classification information field of the entity parameter grouping model.

[0175] Taking the example entity parameter grouping model object data in Example 17 as an example, the data model classification information field structure after adding the second-level classification information is as follows:

[0176] {"classification":{"name":"Airborne Target","child":{"name":"Trainer Aircraft","child":null}}}

Claims

1. A multi-level data modeling method based on a large language model, characterized in that: Includes the following steps: S1. Data Annotation: Define the target entity and entity relationship label group and the relationship feature classification label group in the open source military domain, and complete the annotation of the target dataset entities and entity relationships and the classification annotation of entity relationship features; S2. Data Adaptation: Using an open-source large language model as the foundation, incremental model training is performed using the open-source military domain target annotation dataset from step S1 to fine-tune the parameters of the original large language model, forming an enhanced large language algorithm model suitable for the open-source military domain. MFLLM ; S3. Construct a dictionary library: Construct an open-source standard dictionary library for military target data, including standardized data model parameter names, standardized units of measurement, and unified units of measurement for standardized data model parameters, and save it to the basic data repository; S4. Access and Parsing: Access and parse multi-source heterogeneous knowledge data containing multiple target objects, clean and convert it into text format data, and then save the text data results to the basic data repository; S5. Relation Extraction: Using an Enhanced Large Language Algorithm Model MFLLM The text-formatted data in step S4 is subjected to target entity extraction and entity relation extraction to obtain extraction results containing header entities, footer entities, entity relations, and source data statements. The extraction results are then processed into an aggregated set of entity relation triples. entryArray ; S6. Entity Transformation: Aggregate sets of entity relation triples. entryArray Perform data structure transformation to obtain entity object data. entryObject ; S7. Standardization Processing: This involves standardizing entity object data. entryObject The data is converted into the three elements of entity attributes and then standardized to obtain the entity parameter model. parametricModel ; S8. Fusion Processing: Using a text similarity calculation algorithm, compare entity names, attribute names, attribute values, and units of measurement for entities of the same type, and perform fusion processing on the entity parameter model. parametricModel Perform fusion processing; S9. Hierarchical classification: Classifying entity parameter models parametricModel Clustering is performed based on entity type, and hierarchical classification information is added to the parameter model of each entity. S10. Classification Calculation: Using an Enhanced Large Language Algorithm Model MFLLM For entity parameter model parametricModel The entity attributes are classified and grouped, and the grouping information is added to the entity parameter model. parametricModel Obtain the entity parameter grouping model groupModel ; S11. Clustering calculation: Grouping model of entity parameters groupModel Correlation analysis was performed, and entity parameters were grouped into models using a hierarchical clustering algorithm. groupModel After performing a level 1 clustering calculation to obtain the next level of clusters, the enhanced large language algorithm model is used. MFLLM After generating names for the clusters, the clustering results are added to the hierarchical classification information of the entity parameter grouping model; S12. Generation by level: Repeat step S11 to generate multi-level model classifications, forming a multi-level data model system with inheritance relationships; S13. Complete the saving: Store the multi-level data model system in the business data storage database or save it as a configuration file for subsequent business use; In step S10, the entity parameter grouping model is obtained. groupModel The process specifically includes the following steps: S10.

1. Collect the entity attributes of all entity parameter models; traverse each entity parameter model in step S9, collect all entity attribute names in the form of a set, and then remove duplicate entity attribute names in the set to form a complete set of entity attributes. S10.

2. Perform classification calculations on the names of each entity attribute within the full set of entity attributes; use the enhanced large language algorithm model from step S2. MFLLM The provided entity relationship classification algorithm performs classification calculations on the names of each entity attribute in the full set of entity attributes. The entity attribute names and classification results are stored in the form of key-value pairs, forming an entity attribute classification mapping table. S10.

3. Apply the entity attribute classification mapping table to the entity parameter model, add grouping information, and form an entity parameter grouping model: This involves modifying the entity parameter model from step S9. parametricModel The entity attribute names within the model are matched against the entity attribute classification mapping table. Upon successful matching, the entity parameter model is updated accordingly. parametricModel Add a grouping structure, where the group name is the category name, and integrate the entity parameter model. parametricModel Each entity attribute within is associated with its corresponding group.

2. The multi-level data modeling method based on a large language model as described in claim 1, characterized in that: The entity and entity relationship tag group includes target entity classification, attribute entity classification, predefined relationship types between entities, and descriptive information for each type of tag; the relationship feature classification tag group includes relationship feature types and descriptive information.

3. The multi-level data modeling method based on a large language model as described in claim 1, characterized in that: The three elements of the entity attribute are name, value, and unit of measurement.

4. The multi-level data modeling method based on a large language model as described in claim 1, characterized in that: Step S1 specifically includes the following steps: S1.

1. Select the header entity text and mark its target entity type; S1.

2. Select the tail entity text and mark its target entity type or attribute entity type; S1.

3. If the text contains a relation name, select the relation text and mark it as a custom relation type; otherwise, directly select the predefined relation type between the head entity and the tail entity. S1.

4. Select a feature type for the labeled relation, and transform and merge the text data and labeled data to form an open-source military domain target labeling dataset, which is then saved to the basic data repository.

5. The multi-level data modeling method based on a large language model as described in claim 1, characterized in that: In step S3, the standardized data model parameter name includes the standardized data model parameter name. paramName and a set of synonyms. paramName' ; Standardized units of measurement include standardized units of measurement measureUnit and a set of its synonyms measureUnit' The standardized data model parameters have a unified unit of measurement, and there is a one-to-one correspondence between the standardized data model parameters; In step S4, the sources of multi-source heterogeneous knowledge data include files, databases, and interfaces; among them, file content is accessed through text reading, database fields are queried through SQL statements, and interface input parameters are processed through interface integration.

6. The multi-level data modeling method based on a large language model as described in claim 1, characterized in that: In step S5, the extraction result includes the head entity, tail entity, entity relationship, and source data statement, forming a hash set of entity relationship triples. Then, triples with the same head entity name and head entity type are grouped and merged to obtain an aggregate set of relationship triples for each entity. entryArray .

7. The multi-level data modeling method based on a large language model as described in claim 1, characterized in that: In step S7, the data standardization process specifically includes the following steps: S7.

1. Entity Attribute Name Standardization: By comparing with the data standard dictionary, find the predefined data model parameter aliases that exactly match the entity attribute names. paramName' Replace the entity attribute name with the standardized data model parameter name corresponding to the alias. paramName ; S7.

2. Separation of entity attribute value and unit of measurement: When an entity attribute contains an attribute value and a unit of measurement, the attribute value and unit of measurement shall be separated. S7.

3. Standardization of Entity Attribute Measurement Units: By comparing with the data standard dictionary, find synonyms that perfectly match the entity attribute measurement units. measureUnit' Replace the unit of measurement for entity attributes with the standardized unit of measurement corresponding to the synonym. measureUnit ; S7.

4. Unification of Entity Attribute Measurement Units: First, by comparing the standardized data model parameters in the data standard dictionary to unify the measurement units, find the unified measurement unit for the entity attribute, then calculate the conversion ratio between the current measurement unit and the unified measurement unit, and finally perform proportional scaling calculation on the attribute value.

8. The multi-level data modeling method based on a large language model as described in claim 1, characterized in that: The fusion process in step S8 involves removing redundant entity parameter models and redundant entity attributes within the models, and concentrating the different entity attributes of multiple models with the same entity parameter.

9. A multi-level data modeling system based on a large language model, characterized in that: It includes a data annotation and training unit, a data relationship extraction unit, a data standardization and fusion unit, a data model grouping and aggregation unit, and a data model hierarchical unit, used to implement the multi-level data modeling method based on a large language model as described in any one of claims 1 to 8, wherein, The data annotation and training unit is used to define label groups such as entities and entity relationships, to annotate open-source military text knowledge data, and to incrementally train the open-source large language model using the annotated dataset to form an enhanced large language algorithm model. The data relationship extraction unit is used to extract entities and entity relationships from open-source military domain text knowledge data as a preliminary structured basic dataset; The data standardization and fusion unit is used to standardize and replace the attribute names, attribute values ​​and attribute units of an entity, and to fuse data of the same entities and attributes to form a standardized data parameter model. The data model grouping and aggregation unit is used to group entity attributes in the standardized data parameter model, and aggregate the model according to the entity classification defined by entity and entity relationship label groups, adding classification and grouping information to the model to form a data grouping model; The data model hierarchical unit is used to perform clustering calculations on the data model clusters under the last-level classification, and gradually split the data model clusters downward to abstract a data hierarchical model with multi-level inheritance relationship.

Images