SQL cross-database conversion method and device, equipment and storage medium

Abstract

The invention relates to the technical field of artificial intelligence and the field of financial science and technology, and discloses an SQL (Structured Query Language) cross-database conversion method, which comprises the following steps of: acquiring SQL data from a plurality of database systems, performing vectorization processing, storing the SQL data into an SQL vector database, and performing fine tuning training on a preset model by utilizing the data to obtain an SQL conversion model. And receiving a to-be-converted SQL statement and a target database type, analyzing the to-be-converted SQL statement through the SQL conversion model, converting the to-be-converted SQL statement into a natural language description, and generating a target SQL statement corresponding to the target database type by using the natural language description. Through the automatic SQL conversion process, the dependence on manual intervention is reduced, the accuracy and efficiency of cross-database SQL statement conversion are remarkably improved, the method is suitable for a complex multi-database operation environment, and the consistency and reliability of data processing are ensured.

Description

Technical Field

[0001] The present invention relates to the fields of artificial intelligence technology and financial technology, and in particular to an SQL cross-database conversion method, device, equipment and storage medium. Background Art

[0002] In today's financial field, data processing and cross-platform system integration are becoming increasingly important. Financial institutions usually rely on complex database systems to manage huge amounts of customer information, transaction records, and risk management data. However, due to the differences in architecture, syntax, and operation of different database systems, cross-database SQL statement conversion and execution has become an important and challenging task.

[0003] Traditional SQL conversion methods usually rely on manual writing or simple rule matching, which is not only time-consuming but also prone to errors. Especially in financial applications involving complex queries, data analysis or real-time processing, the correct conversion of SQL statements is directly related to the accuracy of data and the reliability of the system. For example, banking systems need to seamlessly convert and execute SQL queries between multiple databases to ensure the consistency of customer information and the timeliness of transaction processing. Incorrect conversion may lead to data loss or errors, which in turn affects the quality of financial services and customer trust.

[0004] In recent years, with the advancement of artificial intelligence and natural language processing technologies, SQL conversion tools driven by large models have gradually attracted attention. These tools can automatically analyze and convert SQL statements between different database systems, reducing the need for manual intervention and improving the accuracy and efficiency of conversion. However, existing technical solutions still have certain limitations, such as insufficient processing capabilities for complex SQL statements, weak model generalization capabilities, and inconsistencies when generating cross-database SQL statements. These problems are particularly prominent in the financial field, especially in a multi-database environment. Financial institutions need a more intelligent and reliable SQL conversion solution to cope with complex financial data processing needs.

[0005] Therefore, there is an urgent need for a technical solution that can effectively solve these problems, especially in the financial environment, to achieve accurate conversion of SQL statements across databases and ensure the consistency and reliability of data processing. Summary of the invention

[0006] The main purpose of the present invention is to provide a SQL cross-database conversion method, device, equipment and storage medium, aiming to solve the technical problem in the prior art that complex SQL statements cannot be effectively processed and rely on manual intervention during the SQL statement conversion process across database systems.

[0007] To achieve the above object, the present invention provides a SQL cross-database conversion method, comprising:

[0008] Collect various types of SQL data from multiple database systems, perform vectorization on the collected SQL data, and store the processed data in the SQL vector database;

[0009] Using the data in the SQL vector database to fine-tune the preset model to obtain an SQL conversion model;

[0010] Get the SQL statement to be converted and the target database type;

[0011] Converting the SQL statement to be converted into a natural language description by using the SQL conversion model;

[0012] Based on the natural language description, a target SQL statement of a target database type is generated through the SQL conversion model.

[0013] In one embodiment, fine-tuning and training a preset model using data in the SQL vector database to obtain an SQL conversion model includes:

[0014] Freeze all pre-trained weight parameters in the preset model;

[0015] Based on the preset model, design and add one or more new network layers, and initialize the parameters of the newly added network layers;

[0016] Dividing the vectorized data in the SQL vector database according to a preset single batch sample size to generate multiple batches of training data;

[0017] Input each batch of training data into the preset model after the newly added network layer in turn, and generate the output result of the model through forward propagation;

[0018] Generate the loss value of each batch based on the output results of the forward propagation and the preset target;

[0019] Using the obtained loss value, the back propagation algorithm is used to calculate and update the gradient of only the parameters of the newly added network layer;

[0020] When the preset stop condition is met, the training is stopped to obtain the SQL conversion model.

[0021] In one embodiment, each batch of training data is sequentially input into the preset model after the newly added network layer, and the output result of the model is generated by forward propagation, including:

[0022] Introducing a MoE framework into a preset model, wherein the MoE framework includes a plurality of expert modules with the same shape but different parameters;

[0023] Input each batch of training data into the preset model after the newly added network layer in turn. The input training data passes through the newly added network layer to generate a preliminary feature representation.

[0024] Based on the feature representation, the activation weight of each expert module is determined through the configured routing module;

[0025] Selecting one or more expert modules according to the activation weights to process the input training data to generate feature representations for SQL conversion;

[0026] The feature representation processed by the expert module is passed to the final output layer of the model to obtain the output result of the model.

[0027] In one embodiment, the obtained loss value is used to calculate and update the gradient of only the parameters of the newly added network layer through the back propagation algorithm, including:

[0028] The output results after forward propagation and the generated loss values ​​are comprehensively evaluated through the reinforcement learning algorithm to generate reward scores;

[0029] Using the reward score in combination with the loss value, the magnitude and direction of the gradient are adjusted during the back propagation process;

[0030] Based on the optimized gradient, the optimizer is used to update the parameters of the newly added network layer.

[0031] In one embodiment, based on the natural language description, generating a target SQL statement of a target database type through the SQL conversion model includes:

[0032] Extract keywords, field names, table names, and conditional statements from natural language descriptions, and use syntax analysis tools or pre-trained models to parse the structure and semantics of natural language descriptions;

[0033] According to the parsing results and the target database type, select the SQL statement template that conforms to the syntax rules of the target database;

[0034] Fill the parsed keywords, field names, table names and conditional statements into the SQL statement template, and adjust the statement syntax according to the target database type;

[0035] Generate a target SQL statement of the target database type, and perform semantic consistency verification on the generated target SQL statement.

[0036] In one embodiment, after fine-tuning the preset model using the data in the SQL vector database to obtain the SQL conversion model, the method further includes:

[0037] Obtaining an original verification natural language description, inputting the original verification natural language description into an SQL conversion model, and generating an SQL statement corresponding to the verification database type through semantic parsing and structure mapping;

[0038] Re-inputting the generated SQL statement into the SQL transformation model to generate a natural language text corresponding to the original verification natural language description;

[0039] Compare the natural language text generated by the reverse conversion with the original verification natural language description item by item to identify differences in semantics and expression;

[0040] According to the comparison results, adjust and optimize the SQL conversion model;

[0041] The conversion, comparison, and optimization process is repeatedly performed, and different verification natural language descriptions are input for multiple rounds of iterative training to obtain the optimized SQL conversion model.

[0042] In one embodiment, converting the SQL statement to be converted into a natural language description by using the SQL conversion model includes:

[0043] Generate a syntax tree of the SQL statement to be converted by using a syntax analysis tool, and decompose key elements of the SQL statement to be converted therefrom, wherein the key elements include keywords, functions, table names, field names, and conditional statements;

[0044] Using the SQL conversion model, by performing semantic mapping on the parsed key elements, the key elements are converted into corresponding natural language expressions;

[0045] The generated natural language expressions are combined through natural language templates, and the sentence structure is optimized to generate a coherent and complete natural language description.

[0046] Furthermore, to achieve the above-mentioned purpose, the present invention also provides a SQL cross-database conversion device, which includes a memory, a processor, and a SQL cross-database conversion program stored in the memory and executable on the processor, and the SQL cross-database conversion program, when executed by the processor, implements the steps of the SQL cross-database conversion method described above.

[0047] Furthermore, to achieve the above-mentioned purpose, the present invention also provides a computer storage medium, on which a SQL cross-database conversion program is stored, and when the SQL cross-database conversion program is executed by a processor, the steps of the SQL cross-database conversion method described above are implemented.

[0048] Beneficial effects: The present invention relates to a SQL cross-database conversion method, which collects SQL data from multiple database systems, stores it in a SQL vector database after vectorization processing, and uses these data to fine-tune and train a preset model to obtain an SQL conversion model. The SQL statement to be converted and the target database type are received, and the SQL statement to be converted is parsed and converted into a natural language description through the SQL conversion model, and then the natural language description is used to generate a target SQL statement corresponding to the target database type. The present invention reduces the reliance on manual intervention through an automated SQL conversion process, significantly improves the accuracy and efficiency of cross-database SQL statement conversion, is suitable for complex multi-database operating environments, and ensures the consistency and reliability of data processing. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] The present invention will be further described below with reference to the accompanying drawings and embodiments, in which:

[0050] Figure 1 This is a flow chart of an embodiment of the SQL cross-database conversion method of the present invention;

[0051] Figure 2 This is a functional module diagram of a preferred embodiment of the SQL cross-database conversion device of the present invention;

[0052] Figure 3 It is a structural schematic diagram of the device hardware operating environment involved in the SQL cross-database conversion device embodiment of the present invention. DETAILED DESCRIPTION

[0053] It should be understood that the specific embodiments described herein are only used to explain the present invention, and are not used to limit the present invention.

[0054] In today's financial field, data processing and cross-platform system integration are becoming increasingly important. Financial institutions usually rely on complex database systems to manage huge amounts of customer information, transaction records, and risk management data. However, due to the differences in architecture, syntax, and operation of different database systems, cross-database SQL statement conversion and execution has become an important and challenging task.

[0055] Traditional SQL conversion methods usually rely on manual writing or simple rule matching, which is not only time-consuming but also prone to errors. Especially in financial applications involving complex queries, data analysis or real-time processing, the correct conversion of SQL statements is directly related to the accuracy of data and the reliability of the system. For example, banking systems need to seamlessly convert and execute SQL queries between multiple databases to ensure the consistency of customer information and the timeliness of transaction processing. Incorrect conversion may lead to data loss or errors, which in turn affects the quality of financial services and customer trust.

[0056] In recent years, with the advancement of artificial intelligence and natural language processing technologies, SQL conversion tools driven by large models have gradually attracted attention. These tools can automatically analyze and convert SQL statements between different database systems, reducing the need for manual intervention and improving the accuracy and efficiency of conversion. However, existing technical solutions still have certain limitations, such as insufficient processing capabilities for complex SQL statements, weak model generalization capabilities, and inconsistencies when generating cross-database SQL statements. These problems are particularly prominent in the financial field, especially in a multi-database environment. Financial institutions need a more intelligent and reliable SQL conversion solution to cope with complex financial data processing needs.

[0057] Therefore, there is an urgent need for a technical solution that can effectively solve these problems, especially in the financial environment, to achieve accurate conversion of SQL statements across databases and ensure the consistency and reliability of data processing.

[0058] See also Figure 1 , Figure 1 This is a flow chart of an embodiment of the SQL cross-database conversion method provided by the present invention. It should be noted that although the logical order is shown in the flow chart, in some cases, the steps shown or described may be performed in a different order than that shown here.

[0059] like Figure 1 As shown, the SQL cross-database conversion method proposed by the present invention includes the following steps:

[0060] S10, collecting various types of SQL data from multiple database systems, performing vectorization processing on the collected SQL data, and storing the processed data in an SQL vector database;

[0061] In this embodiment, various forms of SQL statements are obtained from different database systems (such as MySQL, PostgreSQL, Oracle, SQL Server, etc.) through interfaces or APIs. The collected SQL data covers query statements, update statements, delete statements, insert statements, and other complex SQL operations. This process requires the compatibility of database connections to ensure that SQL statements in different database systems can be collected and processed consistently.

[0062] Vectorization is the process of converting collected SQL statements into vector forms that can be understood and processed by machine learning models. This process may involve splitting SQL statements into smaller structured elements (such as keywords, table names, field names, operators, etc.), and then using embedding techniques (such as word vectors and sentence vectors) to convert these elements into numerical vector representations. This process provides structured input data for subsequent model training.

[0063] After the vectorization process is completed, the generated SQL vector data is stored in a dedicated SQL vector database. This database is used to store a large amount of processed SQL data for subsequent model training and optimization. The SQL vector database is usually an efficient data storage structure that can quickly index and retrieve the required vectorized SQL data, supporting large-scale data storage and fast access.

[0064] In one specific implementation, a specially designed distributed crawler system is used to automatically connect to multiple database systems and collect SQL data from different databases at the same time. These database systems include relational databases and NoSQL databases. The crawler system automatically identifies and collects various SQL statements by adapting to different database interfaces, ensuring the diversity and integrity of SQL data.

[0065] In the vectorization process, a method based on graph neural network (GNN) is used to represent SQL statements as graph structures, where nodes represent key elements of SQL statements (such as tables, fields, and functions), and edges represent the relationships between these elements. The graph structure is embedded through the GNN model and converted into a high-dimensional vector representation. This method can better capture the complex structure and semantic relationships of SQL statements than traditional word vector technology.

[0066] The vectorized SQL data is stored in a SQL vector database based on blockchain technology. The distributed ledger technology of blockchain ensures the security and immutability of data. Each piece of SQL vector data is recorded on the blockchain, with a timestamp and data source identifier, making the source and change records of SQL data traceable.

[0067] Through the collection and vectorization of SQL data from multiple database systems, combined with an efficient storage SQL vector database, unified processing and management of SQL data across database systems is achieved, ensuring the consistency and availability of SQL data across different database systems. Using these high-quality, structured SQL vector data, the subsequent SQL conversion model can more accurately handle complex SQL statement conversions, significantly improving the efficiency and accuracy of conversions.

[0068] S20, fine-tuning and training a preset model using the data in the SQL vector database to obtain an SQL conversion model;

[0069] In this embodiment, a pre-trained model is fine-tuned by using a large amount of SQL vector data stored in the SQL vector database to adapt it to the cross-database conversion task of SQL statements. Fine-tuning training is to further train the model with data from a specific field on the basis that the model has been pre-trained with a large amount of general data to enhance the performance of the model in this field. The SQL vector data stored in the SQL vector database provides rich training samples for fine-tuning. These samples cover SQL statements and their vector representations of different database systems, which helps the model to better understand and convert SQL statements. Through fine-tuning training, the model can enhance its performance in the cross-database SQL statement conversion task while retaining the capabilities of the original pre-trained model.

[0070] After fine-tuning training is completed, the preset model is updated to a SQL conversion model specifically for SQL conversion tasks. This model has the ability to process and convert SQL statements, and can generate SQL statements suitable for the target database based on the input SQL statements and the target database type. The SQL conversion model obtained through fine-tuning training has higher accuracy and robustness, can handle complex SQL structures, and perform accurate statement conversion between different database systems.

[0071] In one specific implementation, a stratified sampling technique is used from the SQL vector database to perform classified sampling according to the complexity of the SQL statement, the operation type (such as SELECT, UPDATE, DELETE) and the frequency of use. This method ensures that the model can learn the diversity and representativeness of various SQL statements during the fine-tuning process. Stratified sampling can not only reduce the training time, but also improve the performance of the model when processing SQL statements of different complexity.

[0072] A layer-by-layer freeze-thaw strategy is used for fine-tuning training. First, the underlying network parameters of the preset model are fixed, and only the top-level network weights are adjusted. Then, the underlying network parameters of the model are gradually unfrozen to gradually adapt the entire model to the new SQL vector data. This gradual fine-tuning strategy can effectively avoid model overfitting while retaining the model's adaptability to the original task.

[0073] Using a weighted multi-task loss function, we combine multiple tasks such as SQL conversion accuracy, execution time optimization, and memory usage optimization. Through the multi-task loss function, the fine-tuned model can not only perform well in SQL statement conversion, but also be more efficient in resource usage and adapt to different application scenarios.

[0074] By using the data in the SQL vector database for fine-tuning training, the performance of the model in the SQL statement cross-database conversion task can be greatly improved based on the preset model, which significantly improves the accuracy and efficiency of SQL conversion.

[0075] S30, obtaining the SQL statement to be converted and the target database type;

[0076] In this embodiment, the SQL statement to be converted and the corresponding target database type are obtained from the user or the system. The SQL statement to be converted usually comes from different database management systems (DBMS), such as MySQL, PostgreSQL, Oracle, SQL Server, etc., and the target database type refers to the format of the database system into which the SQL statement needs to be converted.

[0077] In the entire SQL cross-database conversion method, obtaining the SQL statements to be converted and the target database type is a key step, which determines the direction of subsequent conversion. First, the system collects the SQL statements to be converted through the API interface, user input, or automated tools, and determines the target database type based on database connection information, user specification, or system automatic detection. This information will be passed as input to the SQL conversion model to ensure that the model can generate SQL statements suitable for the target database.

[0078] In one specific implementation, an advanced database monitoring tool is used to capture the executing SQL statements in real time, and a database access pattern analyzer is used to perform context analysis on the executing SQL statements, extracting metadata such as the execution frequency and response time of the SQL statements. These metadata are attached to the SQL statements for subsequent conversion optimization. This method not only obtains the SQL statements themselves, but also provides rich context information for subsequent optimization, which can help the model better understand and handle SQL conversion tasks.

[0079] Through the database system identification module, the machine learning algorithm is used to analyze the syntax features and database-specific extension functions in the SQL statement, automatically identifying the source database system and determining the target database type. The machine learning algorithm is trained to adapt to the emerging database system and dynamically adjust the identification strategy. This method improves the flexibility and accuracy of identification and is particularly suitable for multi-database environments with complex SQL statement structures.

[0080] Through the above steps, the SQL statement to be converted can be obtained efficiently and accurately and the target database type can be determined.

[0081] S40, converting the SQL statement to be converted into a natural language description through the SQL conversion model;

[0082] In this embodiment, the input SQL statement to be converted is parsed and a corresponding natural language description is generated using a fine-tuned SQL conversion model. Natural language description refers to a clear and detailed description of specific information, operations or data in everyday human language (such as Chinese, English, etc.). Unlike professional terms such as programming languages ​​or SQL statements, the purpose of natural language description is to make information easier to understand, especially for non-technical personnel. For example, the SQL statement: SELECT name, age FROM users WHERE age>30; natural language description: "Query the name and age of all users in the users table who are older than 30."

[0083] Natural language descriptions can accurately convey the logic and semantics of SQL statements, making them easier for humans to understand and process. The model usually learns the structure, key elements, and operation logic of SQL statements through pre-training and fine-tuning, and converts them into sentences that conform to natural language expressions.

[0084] In one specific implementation, a semantic role labeling model is used to label elements in SQL statements to identify actions (such as query, update), subjects (such as table names, fields), conditions (such as filter conditions), etc. The labeling model combines the classification task of semantic roles and further improves the accuracy of labeling through a multi-layer perceptron (MLP). Each labeled semantic role is mapped to a specific natural language phrase, making the generated natural language description more readable.

[0085] The system has built-in natural language templates that cover different types of SQL operations, such as query, update, insert, etc. Based on the results of semantic role labeling, the system fills these labeled semantic fragments into the corresponding templates to generate structured natural language descriptions. The template library is regularly updated to add support for new SQL functions to ensure that the generated descriptions keep up with technological developments.

[0086] While generating natural language descriptions, the system also adds detailed explanations. Each sentence in the natural language description is followed by a short comment that explains the logic or technical background of the corresponding SQL operation. For example, when describing a complex JOIN operation, the system adds comments to explain the JOIN type (such as INNER JOIN, LEFT JOIN) and its meaning. The enhanced explainability description is suitable for scenarios such as teaching and technical document generation, helping users to understand SQL operations more deeply.

[0087] By using the SQL conversion model to convert the SQL statements to be converted into natural language descriptions, the system can present the logic and functions of SQL statements in a more intuitive and easy-to-understand way. It eliminates the user's dependence on SQL syntax, allowing the results of SQL operations to be understood and verified by non-technical users. In scenarios such as finance that require highly precise data operations, the generated natural language description can be used for automated report generation, human-computer interaction interfaces, and data audits, significantly improving the readability and interpretability of SQL conversion operations, reducing errors caused by grammatical misunderstandings, and improving the intelligence of the system and user experience.

[0088] S50: Based on the natural language description, generate a target SQL statement of a target database type through the SQL conversion model.

[0089] In this embodiment, the SQL conversion model receives the natural language description and parses the grammatical structure and semantics of the description through its internal natural language processing module. This step includes identifying key operations (such as query, update, delete), data objects (such as table names, field names), operation conditions (such as screening conditions), etc. In this process, the SQL conversion model uses the knowledge learned during its training to understand the complex expressions and potential semantics in natural language.

[0090] After parsing, the SQL conversion model maps the operation intent and structure in the natural language to the grammatical structure of the SQL statement. This mapping process is the core function of the SQL conversion model. Through the mapping rule library inside the model, the model can accurately map the elements in the natural language to the components of the SQL statement. At this time, the SQL conversion model will select the appropriate SQL template according to the target database type. These templates have been optimized for the syntax and functional characteristics of the target database.

[0091] After semantic mapping and template matching, the SQL conversion model generates preliminary SQL statements. This generation process is based on the model's understanding of natural language descriptions and is combined with the requirements of the target database type to ensure that the generated SQL statements are well structured and conform to the grammatical specifications of the target database.

[0092] The generated SQL statements need to undergo further syntax adjustments to ensure that they can be executed correctly and efficiently on the target database. The SQL conversion model adjusts and optimizes SQL statements in combination with the specific syntax rules of the target database (such as function calls, operator usage, and query optimization strategies). This process ensures that the final generated SQL statements can not only be executed, but also achieve optimal performance on the target database.

[0093] By using the SQL conversion model to convert natural language descriptions into SQL statements of the target database type, the system significantly simplifies the interaction between users and the database, reducing the complexity and error risk of manually writing SQL statements. At the same time, the SQL conversion model ensures the grammatical correctness and execution efficiency of the generated SQL statements in different database environments, thereby improving the flexibility and adaptability of database operations, especially in multi-database systems, ensuring the consistency and performance optimization of SQL statement conversion.

[0094] The present invention relates to a SQL cross-database conversion method, which collects SQL data from multiple database systems, stores them in a SQL vector database after vectorization processing, and uses these data to fine-tune and train a preset model to obtain an SQL conversion model. The SQL statement to be converted and the target database type are received, and the SQL statement to be converted is parsed and converted into a natural language description through the SQL conversion model, and then the natural language description is used to generate a target SQL statement corresponding to the target database type. The present invention reduces the dependence on manual intervention through an automated SQL conversion process, significantly improves the accuracy and efficiency of cross-database SQL statement conversion, is suitable for complex multi-database operating environments, and ensures the consistency and reliability of data processing.

[0095] In one embodiment, the above S20 includes:

[0096] S201, freezing all pre-trained weight parameters in the preset model;

[0097] S202, based on the preset model, design and add one or more new network layers, and initialize the parameters of the newly added network layers;

[0098] S203, dividing the vectorized data in the SQL vector database according to a preset single batch sample size to generate multiple batches of training data;

[0099] S204, inputting the training data of each batch into the preset model after the newly added network layer in sequence, and generating the output result of the model through forward propagation;

[0100] S205, generating a loss value for each batch based on the output result of the forward propagation and the preset target;

[0101] S206, using the obtained loss value, only the parameters of the newly added network layer are calculated and updated by the back propagation algorithm;

[0102] S207, when the preset stop condition is met, stop training and obtain the SQL conversion model.

[0103] In this embodiment, during the fine-tuning training process, the pre-trained weight parameters in the preset model remain unchanged, that is, these parameters will not be updated during the fine-tuning process. This strategy ensures that the model retains the knowledge in the original pre-training process and only adjusts the parameters of the newly added network layer to adapt to the specific requirements of the SQL conversion task. This freezing strategy can effectively avoid model overfitting while maintaining the stability of the model.

[0104] Based on the original preset model, new network layers are designed and added. These network layers are usually designed to solve specific tasks, such as adapting to the special structure of SQL statements or improving the generalization ability of the model. The parameters of the newly added network layers need to be initialized before fine-tuning training. These parameters will be adjusted and optimized according to task requirements during subsequent training.

[0105] Extract vectorized SQL data from the SQL vector database and divide the data into pre-set single batch sample sizes. Batch training can reduce computing resource usage and help the model converge faster. Each batch of data contains multiple types of SQL statement vector representations, ensuring data diversity and model robustness during training.

[0106] During the training process, the vectorized SQL data is input into the updated preset model in batches, and the model calculates the output results of each batch through forward propagation. Forward propagation is the process of calculating the output of the model. After the input data passes through each layer of the model, the output value related to the task (such as the predicted SQL statement structure) is generated.

[0107] The loss value is an indicator that measures the difference between the model output and the preset target. After each batch of training, the system calculates the error between the output and the target output based on the output of the forward propagation. This error is the loss value. The smaller the loss value, the closer the model's prediction is to the expected target.

[0108] During the back-propagation process, the loss value is used to calculate the gradient of the parameters in the model, and these gradients will guide the update of the model parameters. Since the original parameters of the preset model have been frozen, only the parameters of the newly added network layers will be updated during the training process. This process helps the model better adapt to the needs of new tasks and optimize the performance of the newly added network layers.

[0109] The training process ends when the preset stopping conditions are met, which may include the loss value reaching a certain threshold, the training reaching the set epochs, etc. After stopping the training, the obtained model is the fine-tuned SQL conversion model, which has the ability to process and convert SQL statements.

[0110] In one specific implementation, during the fine-tuning process, the system freezes all pre-trained weights of the encoder part of the preset model. These encoders can be Transformer-based encoding layers, and no adjustments are made to these layers during the fine-tuning process. This strategy ensures that the model's basic ability to process natural language is not affected, and focuses on the decoder part after fine-tuning.

[0111] A new self-attention layer is introduced in the decoder part. This layer dynamically adjusts the input weights so that the model can better capture the long-distance dependencies in SQL statements. The LeCun normal initialization method is used during initialization, which is more suitable for deep networks and ensures greater model stability in the early stages of training.

[0112] In terms of the division of training data, the system adopts a dynamic batching strategy (Dynamic Batching). Different from the traditional fixed-size batch division, this strategy dynamically adjusts the single batch sample size according to the complexity of each SQL statement. More complex SQL statements will be assigned to smaller batches to ensure that the model can fully learn complex structures.

[0113] In the forward propagation phase, the system not only calculates the output of the model, but also introduces an intermediate layer output monitoring mechanism to record the intermediate output of each network layer. This monitoring mechanism can help detect potential gradient disappearance or explosion problems during training, and interrupt and adjust when necessary.

[0114] In this embodiment, the system can effectively optimize the SQL conversion model by using the data in the SQL vector database to fine-tune the preset model, and adopt methods such as freezing some parameters, designing new network layers, and training in batches, so that it can better adapt to the conversion task of SQL statements. Freezing the pre-trained weights allows the model to focus on the optimization of specific tasks while maintaining the original knowledge, while the design and fine-tuning of the new network layer improves the model's ability to handle complex SQL structures. The precise processing of batch training and loss value calculation ensures the efficiency and stability of the training process. The SQL conversion model finally generated is not only highly accurate, but also has good generalization ability, and can efficiently perform SQL conversion operations in different database systems.

[0115] In one embodiment, the above S204 includes:

[0116] S2041, introducing a MoE framework into a preset model, wherein the MoE framework includes a plurality of expert modules having the same shape but different parameters;

[0117] S2042, inputting each batch of training data into the preset model after the newly added network layer in sequence, and the input training data passes through the newly added network layer to generate a preliminary feature representation;

[0118] S2043, based on the feature representation, determining the activation weight of each expert module through the configured routing module;

[0119] S2044, selecting one or more expert modules according to the activation weights to process the input training data to generate feature representations for SQL conversion;

[0120] S2045, passing the feature representation processed by the expert module to the final output layer of the model to obtain the output result of the model.

[0121] In this embodiment, the MoE (Mixture of Experts) framework is a mechanism that can dynamically select sub-models (i.e., expert modules) to participate in calculations. Each expert module has different parameter settings when processing input data, so that it can excel in processing certain specific input types or tasks. Introducing the MoE framework in the preset model allows the model to dynamically select the most appropriate expert module for processing based on the characteristics of the input data, thereby improving the adaptability and accuracy of the model.

[0122] After the training data is input into the preset model, it will first be processed by the newly added network layers to generate preliminary feature representations. These newly added network layers may include convolutional layers, attention mechanism layers, or other deep learning layers, which are designed to extract key features of the input data. These feature representations will serve as the basis for the selection of expert modules in subsequent steps.

[0123] The routing module is a key part of the MoE framework. It is responsible for calculating the activation weight of each expert module based on the input feature representation. The activation weight reflects the adaptability of each expert module to the current input data. The routing module usually uses a small neural network or logistic regression model to predict the activation probability of each expert module based on the input features.

[0124] Based on the calculated activation weights, the system selects one or more expert modules to process the input training data. The selection strategy can be:

[0125] Top-K selection: Select the K expert modules with the highest activation weights (for example, select the Top-2 or Top-3 expert modules) to jointly process the input data.

[0126] Probabilistic selection: According to the probability distribution of activation weights, an expert module is randomly selected for processing with a certain probability. The selected expert module will perform further feature extraction or conversion operations on the input data to generate more advanced feature representations that are closer to the requirements of the SQL conversion task. Multiple expert modules can process input data in parallel, and the final result will combine the outputs of each module.

[0127] The feature representation generated by the home module will be passed to the final output layer of the model, which may be a fully connected layer, a linear layer, or other forms of layers, and finally generate an output result for SQL conversion. The output result can be a vector representation of the SQL statement, or directly generate the SQL statement itself.

[0128] In one specific implementation, a MoE layer is added after the attention mechanism of the preset model, which contains 10 expert modules with the same shape but different parameters. Each expert module is an independent Transformer encoder, which is specifically used to process different types of SQL statement structures (such as query, update, delete, etc.). When the training data is input into the model, it first passes through a multi-head attention layer (Multi-HeadAttention Layer) to generate a preliminary feature representation of the SQL statement. The multi-head attention layer extracts multi-dimensional features from the input data by performing multiple self-attention operations on the input data, providing rich information for the selection of subsequent expert modules.

[0129] This embodiment introduces the MoE framework into the preset model, and the system can dynamically select the most suitable expert module for processing according to the characteristics of the input data. This flexibility enables the model to respond more accurately and efficiently when faced with diverse and complex SQL statements. The dynamic selection capability of the routing module further optimizes the efficiency of each expert module and ensures that the generated feature representation highly matches the requirements of SQL conversion. The final output result is not only more accurate, but the generated SQL statement can also maintain efficiency and consistency when executed on the target database.

[0130] In one embodiment, the above S206 includes:

[0131] S2061, performing a comprehensive quality evaluation on the output results after forward propagation and the generated loss value through a reinforcement learning algorithm to generate a reward score;

[0132] S2062, using the reward score in combination with the loss value, adjusting the size and direction of the gradient during the back propagation process;

[0133] S2063: Based on the optimized gradient, use the optimizer to update the parameters of the newly added network layer.

[0134] In this embodiment, during the back propagation process, a reinforcement learning algorithm is introduced to evaluate the quality of the model output. Specifically, the reinforcement learning algorithm generates a reward score by observing the output results and the corresponding loss values ​​after forward propagation. This reward score reflects the overall quality of the model output and the degree of deviation from the expected target. The introduction of reinforcement learning enables the model to not only rely on the loss value itself, but also better guide the optimization direction of the gradient through the reward mechanism, thereby improving the performance of the model when processing specific tasks.

[0135] In traditional back propagation, the size and direction of the gradient are determined only by the loss value, but after the introduction of reinforcement learning, the calculation of the gradient is also affected by the reward score. By combining the reward score with the loss value, the system can dynamically adjust the size and direction of the gradient, making the direction of the model update more optimized, especially for those tasks with higher complexity or difficulty. This gradient adjustment method combined with reinforcement learning can prevent the model from falling into the local optimum and enable the model to converge to the global optimal solution faster.

[0136] After obtaining the optimized gradient adjusted by reinforcement learning, the system will use optimizers (such as Adam, RMSProp, etc.) to update the parameters of the newly added network layer. These optimizers are responsible for applying the calculated gradients and adjusting the network parameters to minimize the loss function. Since the gradients have been adjusted by reinforcement learning, the updated parameters are more in line with the optimization goals, and the overall performance of the model is improved.

[0137] In one specific implementation, the system uses the policy gradient method as a reinforcement learning algorithm to evaluate the quality of the model output results. After each forward propagation, the policy gradient method generates a reward score based on the difference between the output of the current model and the target result. The higher the reward score, the closer the model output quality is to the target. During the back propagation process, the system combines the loss value and the reward score, and uses a weighting factor to combine the two to obtain the final gradient. This weighting factor is dynamically adjusted by the reinforcement learning algorithm to ensure that the gradient can flexibly respond to various inputs at different training stages. The generated optimized gradient will be passed to the optimizer to update the parameters of the newly added network layer. The adaptive momentum optimization algorithm (AMSGrad) is used to update the parameters of the network layer. AMSGrad can maintain a stable learning rate when processing different types of input data to prevent the model from falling into a local optimum. The updated parameters are applied to the model to ensure that the model can output higher quality results during the next forward propagation.

[0138] This embodiment introduces a reinforcement learning algorithm to comprehensively evaluate the output results and loss values ​​after forward propagation, so that the system can generate more reasonable reward scores, which not only affect the size and direction of the gradient, but also improve the generalization ability and learning efficiency of the model. By using the optimized gradient to update parameters, the model can converge more quickly when processing complex tasks and improve the final output quality.

[0139] In one embodiment, the above S50 includes:

[0140] S501, extracting keywords, field names, table names, and conditional statements from the natural language description, and using a syntax analysis tool or a pre-trained model to parse the structure and semantics of the natural language description;

[0141] S502, selecting an SQL statement template that complies with the grammatical rules of the target database according to the parsing result and the target database type;

[0142] S503, filling the parsed keywords, field names, table names and conditional statements into the SQL statement template, and adjusting the statement syntax according to the target database type;

[0143] S504: Generate a target SQL statement of the target database type, and perform semantic consistency verification on the generated target SQL statement.

[0144] In this embodiment, a syntax analysis tool or a pre-trained natural language processing model (such as BERT or GPT) is used to parse the natural language description. The goal of the analysis is to extract the core elements of the SQL statement from the description, such as keywords (such as SELECT, UPDATE, etc.), field names, table names, and conditional statements. The syntax analysis tool can be a rule-based parser or a semantic analyzer trained using a machine learning model, which can effectively identify and understand the structure and semantics in natural language.

[0145] After parsing is complete, the system selects an appropriate SQL statement template based on the extracted keywords, field names, table names, and conditional statements, combined with the specific type of the target database. Different database systems (such as MySQL, Oracle, PostgreSQL) may have different syntax and functional characteristics, so it is important to select an SQL template that is suitable for the target database. This step ensures that the generated SQL statement complies with the syntax rules of the target database and can be executed correctly on the database system.

[0146] The methods for selecting SQL statement templates include but are not limited to:

[0147] Rule-based template selection: Use a predefined set of rules to select a suitable SQL template based on the natural language description and the target database type. This method is suitable for natural language input with simple or clear semantics. First, the natural language description is parsed to extract elements such as keywords, operators, table names, and field names. Based on the extracted elements and the type of the target database, the system matches the predefined rule base and selects a SQL template that meets the conditions. For example, if the natural language description contains "find all records" and the target database is MySQL, the system will select a template containing a SELECT*FROM structure.

[0148] Template selection based on similarity matching: Select the best matching template by calculating the similarity between the natural language description and the SQL template. This method is more flexible and can adapt to a variety of natural language expressions. Convert both the natural language description and the SQL template into vector representations (for example, through TF-IDF, Word2Vec, BERT, etc.). Calculate the similarity between the natural language description and each SQL template vector using methods such as cosine similarity or Euclidean distance. Select the SQL template with the highest similarity as the final template.

[0149] Template selection based on reinforcement learning: SQL templates are selected through a reinforcement learning (RL) model. The reinforcement learning model dynamically adjusts the template selection strategy based on historical generation records and current input features. A reward mechanism is designed so that each selected template is rewarded positively after generating a correct SQL statement, and negatively rewarded for generating an incorrect SQL statement. The system gradually learns the strategy of selecting the optimal template under different input conditions by continuously and tentatively selecting SQL templates. In actual applications, the RL model selects SQL templates in real time based on natural language descriptions and target database types.

[0150] After selecting the SQL template, the system will fill the parsed elements into the corresponding positions of the template. This process includes inserting field names, table names, conditional statements, etc. into the appropriate positions in the SQL template, and adjusting the SQL statements according to the syntax characteristics of the target database. For example, different databases may have different syntax for date formats or string processing. The system will adjust the SQL statements in this step to make them fully meet the requirements of the target database.

[0151] After the template is filled and the syntax is adjusted, the system generates the final target SQL statement. In order to ensure that the generated SQL statement can correctly reflect the intention of the original natural language description, the system will perform semantic consistency verification on the SQL statement. This verification process ensures the accuracy of the conversion result by comparing the semantics of the generated SQL statement with the original natural language description. If semantic inconsistency is detected, the system may adjust the generation process to improve consistency.

[0152] This embodiment extracts key elements from natural language descriptions and generates SQL statements in combination with the grammatical rules of the target database. The system can effectively convert the user's natural language requirements into executable SQL queries. It not only ensures that the generated SQL statements accurately reflect the user's intentions, but also ensures the correctness of the conversion results through semantic consistency verification.

[0153] In one embodiment, after the above S20, the method further includes:

[0154] S208, obtaining an original verification natural language description, inputting the original verification natural language description into an SQL conversion model, and generating an SQL statement corresponding to the verification database type through semantic parsing and structure mapping;

[0155] S209, re-inputting the generated SQL statement into the SQL conversion model to generate a natural language text corresponding to the original verification natural language description;

[0156] S210, comparing the natural language text generated by the reverse conversion with the original verification natural language description item by item to identify differences in semantics and expression;

[0157] S211, adjusting and optimizing the SQL conversion model according to the comparison result;

[0158] S212, repeatedly executing the conversion, comparison and optimization process, inputting different verification natural language descriptions to perform multiple rounds of iterative training, and obtaining an optimized SQL conversion model.

[0159] In this embodiment, starting from the input natural language description, the system parses the semantics of the natural language through the SQL conversion model and maps it into SQL statements that conform to a specific database type. Semantic parsing includes extracting elements such as key operations, data objects, and conditions, while structural mapping organizes these elements into an SQL structure that conforms to the syntax of the target database.

[0160] The generated SQL statement is reverse converted and passed through the SQL conversion model again to parse it and generate the corresponding natural language text. The purpose is to verify whether the SQL statement correctly expresses the semantics and intent in the original natural language description. By reversely generating natural language text, the system can check whether there is any information loss or misunderstanding during the conversion process.

[0161] The system compares the generated natural language text with the original natural language description, and identifies possible deviations or errors in the model conversion process by analyzing the differences in semantics and expressions item by item. This process may involve detailed semantic analysis to ensure that each element (such as operation type, condition, object, etc.) is accurately understood and converted.

[0162] Based on the above comparison results, the system adjusts and optimizes the SQL conversion model. This may involve adjusting the model's parameters, updating semantic parsing rules, or optimizing mapping logic to reduce errors and inconsistencies in future conversions. The purpose of the adjustment is to improve the accuracy and robustness of the model so that it can better adapt to various natural language inputs.

[0163] The system goes through multiple rounds of iterative training, continuously inputting new natural language descriptions for conversion, comparison, and optimization. Each iteration will be fine-tuned based on the results of the previous round, allowing the model to gradually improve its accuracy and ability to handle complex inputs. Ultimately, the model will be optimized to a stable state, capable of efficiently and accurately handling various SQL conversion tasks.

[0164] In this embodiment, by adding verification, reverse generation, comparison and optimization processes after fine-tuning the SQL conversion model, the system can further improve the accuracy and robustness of the model. The repeated conversion and verification process ensures that the model can not only generate correct SQL statements, but also maintain consistency with natural language descriptions. The introduction of multiple rounds of iterative training enables the model to continuously adapt to different input types and complexities, and ultimately obtains an optimized SQL conversion model with higher accuracy and wide applicability. This continuous optimization mechanism greatly improves the reliability and execution efficiency of the system in practical applications.

[0165] In one embodiment, the above S40 includes:

[0166] S401, using a syntax analysis tool to generate a syntax tree of the SQL statement to be converted, and decomposing key elements of the SQL statement to be converted therefrom, wherein the key elements include keywords, functions, table names, field names, and conditional statements;

[0167] S402, using the SQL conversion model to perform semantic mapping on the parsed key elements, and converting the key elements into corresponding natural language expressions;

[0168] S403, combining the generated natural language expressions through natural language templates and optimizing the sentence structure to generate a coherent and complete natural language description.

[0169] In this embodiment, a syntax analysis tool is used to parse the SQL statement to be converted to generate its syntax tree. The syntax tree is a structured representation of the SQL statement, and the hierarchy and various components of the SQL statement can be clearly displayed through the syntax tree. The system extracts the key elements of the SQL statement from the syntax tree, including the operation keywords (such as SELECT, INSERT), function calls (such as SUM, AVG), table names, field names, and conditional statements (such as conditions in the WHERE clause) in the SQL statement. These elements are the basis for semantic mapping and generating natural language descriptions in the subsequent conversion process.

[0170] After obtaining the key elements of the SQL statement, the system uses the SQL conversion model to map these elements into corresponding natural language expressions. Semantic mapping converts technical SQL terms into natural language that users can understand. For example, the SELECT keyword in SQL may be mapped to "select" or "find", and the conditions of the WHERE clause may be mapped to "meet the following conditions", etc. This step ensures that each part of the SQL statement can be accurately converted into natural language, so that the final description can reflect the original intention of the SQL statement.

[0171] After completing the semantic mapping of key elements, the system combines these mapped fragments through predefined natural language templates. The template is responsible for defining the basic structure of natural language expression, such as "select fields from a table where the condition is..." During the combination process, the system also optimizes the sentence structure to ensure that the generated natural language description is fluent, coherent, and conforms to the grammatical rules of natural language. In the end, the natural language description generated by the system not only accurately expresses the meaning of the SQL statement, but also has good readability.

[0172] Example description: Assume there is a SQL statement:

[0173] SELECT name,age FROM employees WHERE department='HR'AND age>30;

[0174] Each element in the SQL statement is mapped to a natural language expression:

[0175] SELECT: "Select";

[0176] name,age:"name and age";

[0177] FROM employees: "from the employee table";

[0178] WHERE department='HR'AND age>30: "Employees whose department is Human Resources and whose age is greater than 30".

[0179] The system uses a predefined natural language template to combine these expressions into a complete natural language description. A simple template may be: "{action}{field}{data source}{condition}".

[0180] Fill in each part into the template: Action: "Select"; Field: "Name and Age"; Data Source: "From the Employee Table"; Condition: "Employees whose department is the Human Resources Department and whose age is greater than 30";

[0181] Combined sentence: "Select name and age from the employee table, where the department is human resources and the age is greater than 30."

[0182] The system optimizes the generated sentences to make them more coherent and in line with the norms of natural language. Possible optimizations include adjusting word order, adding conjunctions, modifiers, etc. For example: before optimization, the sentence structure is a bit stiff, and after optimization it may become: "Select the names and ages of employees from the employee table whose department is the human resources department and who are over 30 years old." This optimization makes the sentence more in line with the natural expression of human language, reads more smoothly, and conforms to grammatical rules.

[0183] This embodiment uses syntax analysis, semantic mapping and template combination to convert technical SQL statements into user-friendly natural language descriptions, which not only ensures that the generated descriptions accurately reflect the meaning of the SQL statements, but also makes the descriptions more fluent and understandable through natural language optimization processing.

[0184] The present invention also provides a SQL cross-database conversion device, referring to Figure 2 , Figure 2 This is a functional module diagram of a preferred embodiment of the SQL cross-database conversion device of the present invention. The SQL cross-database conversion device comprises:

[0185] The data collection and vectorization module is used to collect various types of SQL data from multiple database systems, perform vectorization on the collected SQL data, and store the processed data in the SQL vector database;

[0186] A model training and fine-tuning module, used to fine-tune the preset model using the data in the SQL vector database to obtain an SQL conversion model;

[0187] The input data acquisition module is used to obtain the SQL statement to be converted and the target database type;

[0188] A SQL statement conversion module to be converted, used to convert the SQL statement to be converted into a natural language description through the SQL conversion model;

[0189] A target SQL statement generation module is used to generate a target SQL statement of a target database type based on the natural language description and through the SQL conversion model.

[0190] The specific implementation of the SQL cross-database conversion device of the present invention is basically the same as the various embodiments of the above-mentioned SQL cross-database conversion method, and will not be repeated here.

[0191] The present invention also provides a SQL cross-database conversion device, such as Figure 3 As shown, the SQL cross-database conversion device may include: a processor 1001, such as a CPU, a communication bus 1002, a user interface 1003, a network interface 1004, and a memory 1005. Among them, the communication bus 1002 is used to realize the connection and communication between these components. The user interface 1003 may include a display screen (Display), an input unit such as a keyboard (Keyboard), and the optional user interface 1003 may also include a standard wired interface and a wireless interface. The network interface 1004 may optionally include a standard wired interface and a wireless interface (such as a WI-FI interface). The memory 1005 may be a high-speed RAM memory, or a stable memory (non-volatile memory), such as a disk memory. The memory 1005 may also be a storage device independent of the aforementioned processor 1001.

[0192] Those skilled in the art will understand that Figure 3 The hardware structure of the SQL cross-database conversion device shown in the figure does not constitute a limitation on the SQL cross-database conversion device, and may include more or less components than shown in the figure, or combine certain components, or arrange the components differently.

[0193] like Figure 3 As shown, the memory 1005 as a storage medium may include an operating system, a network communication module, a user interface module, and an SQL cross-database conversion program. The operating system is a program for managing and controlling SQL cross-database conversion equipment and software resources, and supports the operation of the network communication module, the user interface module, the SQL cross-database conversion program, and other programs or software; the network communication module is used to manage and control the network interface 1004; and the user interface module is used to manage and control the user interface 1003.

[0194] exist Figure 3In the hardware structure of the SQL cross-database conversion device shown, the network interface 1004 is mainly used to connect to the background server and communicate data with the background server; the user interface 1003 is mainly used to connect to the client and communicate data with the client; the processor 1001 can call the SQL cross-database conversion program stored in the memory 1005 and perform the same operation as the SQL cross-database conversion method.

[0195] The specific implementation of the SQL cross-database conversion device of the present invention is basically the same as the above-mentioned SQL cross-database conversion method embodiments, and will not be repeated here.

[0196] In addition, an embodiment of the present invention further provides a computer storage medium, on which a SQL cross-database conversion program is stored. When the SQL cross-database conversion program is executed by a processor, the steps of the SQL cross-database conversion method described above are implemented.

[0197] The specific implementation of the computer storage medium of the present invention is basically the same as the above-mentioned SQL cross-database conversion method embodiments, and will not be repeated here.

[0198] It should be noted that if software tools or components other than those of our company appear in the embodiments of the present application, they are only used for illustrative purposes and do not represent actual use.

Claims

1. A SQL cross-database conversion method, characterized in that: The following steps are involved: Collect various types of SQL data from multiple database systems, perform vectorization on the collected SQL data, and store the processed data in the SQL vector database; Using the data in the SQL vector database to fine-tune the preset model to obtain an SQL conversion model; Get the SQL statement to be converted and the target database type; Converting the SQL statement to be converted into a natural language description by using the SQL conversion model; Based on the natural language description, a target SQL statement of a target database type is generated through the SQL conversion model.

2. The SQL cross-database conversion method according to claim 1, characterized in that: The preset model is fine-tuned and trained using the data in the SQL vector database to obtain an SQL conversion model, including: Freeze all pre-trained weight parameters in the preset model; Based on the preset model, design and add one or more new network layers, and initialize the parameters of the newly added network layers; Dividing the vectorized data in the SQL vector database according to a preset single batch sample size to generate multiple batches of training data; Input each batch of training data into the preset model after the newly added network layer in turn, and generate the output result of the model through forward propagation; Generate the loss value for each batch based on the output results of the forward propagation and the preset target; Using the obtained loss value, the back propagation algorithm is used to calculate and update the gradient of only the parameters of the newly added network layer; When the preset stop condition is met, the training is stopped to obtain the SQL conversion model.

3. The SQL cross-database conversion method according to claim 2, characterized in that: The training data of each batch is sequentially input into the preset model after the new network layer is added, and the output results of the model are generated through forward propagation, including: Introducing a MoE framework into a preset model, wherein the MoE framework includes a plurality of expert modules with the same shape but different parameters; Input each batch of training data into the preset model after the newly added network layer in turn. The input training data passes through the newly added network layer to generate a preliminary feature representation. Based on the feature representation, the activation weight of each expert module is determined through the configured routing module; Selecting one or more expert modules according to the activation weights to process the input training data to generate feature representations for SQL conversion; The feature representation processed by the expert module is passed to the final output layer of the model to obtain the output result of the model.

4. The SQL cross-database conversion method according to claim 2, characterized in that: Using the obtained loss value, the back propagation algorithm is used to calculate and update the gradient of the parameters of the newly added network layer, including: The output results after forward propagation and the generated loss values ​​are comprehensively evaluated through the reinforcement learning algorithm to generate reward scores; Using the reward score in combination with the loss value, the magnitude and direction of the gradient are adjusted during the back propagation process; Based on the optimized gradient, the optimizer is used to update the parameters of the newly added network layer.

5. The SQL cross-database conversion method as described in claim 1, characterized in that: Based on the natural language description, generating a target SQL statement of a target database type through the SQL conversion model includes: Extract keywords, field names, table names, and conditional statements from natural language descriptions, and use syntax analysis tools or pre-trained models to parse the structure and semantics of natural language descriptions; According to the parsing results and the target database type, select the SQL statement template that complies with the grammatical rules of the target database; Fill the parsed keywords, field names, table names and conditional statements into the SQL statement template, and adjust the statement syntax according to the target database type; Generate a target SQL statement of the target database type, and perform semantic consistency verification on the generated target SQL statement.

6. The SQL cross-database conversion method according to claim 1, characterized in that: After fine-tuning the preset model using the data in the SQL vector database to obtain the SQL conversion model, the method further includes: Obtaining an original verification natural language description, inputting the original verification natural language description into an SQL conversion model, and generating an SQL statement corresponding to the verification database type through semantic parsing and structure mapping; Re-inputting the generated SQL statement into the SQL transformation model to generate a natural language text corresponding to the original verification natural language description; Compare the natural language text generated by the reverse conversion with the original verification natural language description item by item to identify differences in semantics and expression; According to the comparison results, adjust and optimize the SQL conversion model; The conversion, comparison, and optimization process is repeatedly performed, and different verification natural language descriptions are input for multiple rounds of iterative training to obtain the optimized SQL conversion model.

7. The SQL cross-database conversion method according to claim 1, characterized in that: Converting the SQL statement to be converted into a natural language description by using the SQL conversion model includes: Generate a syntax tree of the SQL statement to be converted by using a syntax analysis tool, and decompose key elements of the SQL statement to be converted therefrom, wherein the key elements include keywords, functions, table names, field names, and conditional statements; Using the SQL conversion model, by performing semantic mapping on the parsed key elements, the key elements are converted into corresponding natural language expressions; The generated natural language expressions are combined through natural language templates, and the sentence structure is optimized to generate a coherent and complete natural language description.

8. A SQL cross-database conversion device, characterized in that: The SQL cross-database conversion device comprises: The data collection and vectorization module is used to collect various types of SQL data from multiple database systems, perform vectorization on the collected SQL data, and store the processed data in the SQL vector database; A model training and fine-tuning module, used to fine-tune the preset model using the data in the SQL vector database to obtain an SQL conversion model; The input data acquisition module is used to obtain the SQL statement to be converted and the target database type; A SQL statement conversion module to be converted, used to convert the SQL statement to be converted into a natural language description through the SQL conversion model; A target SQL statement generation module is used to generate a target SQL statement of a target database type based on the natural language description and through the SQL conversion model.

9. A SQL cross-database conversion device, characterized in that: The SQL cross-database conversion device includes a memory, a processor, and a SQL cross-database conversion program stored in the memory and executable on the processor. When the SQL cross-database conversion program is executed by the processor, the steps of the SQL cross-database conversion method according to any one of claims 1 to 7 are implemented.

10. A computer storage medium, characterized in that: The storage medium stores an SQL cross-database conversion program, which, when executed by a processor, implements the steps of the SQL cross-database conversion method according to any one of claims 1 to 7.

Images