Multi-modal SQL database interaction analysis and intelligent visual decision-making system

Abstract

The invention discloses a multi-modal SQL database interaction analysis and intelligent visual decision-making system, and relates to the field of database intelligent analysis. Comprising the steps of 1, achieving conversion from a natural language to an SQL statement through a multi-stage semantic analysis and verification mechanism, 2, automatically generating an appropriate visual map according to query result data features, and 3, conducting continuous optimization through a two-channel feedback mechanism, correcting the generated SQL statement or a visual result through display feedback, and obtaining the SQL statement or the visual result. The correction record is stored in a training pool in real time and used for incremental training of the model; interaction behavior characteristics of the user are analyzed through implicit feedback, the interaction behavior characteristics comprise query input duration, click operation position and dwell time, and information is extracted according to the interaction behavior characteristics for model optimization.

Description

Technical Field

[0001] This invention discloses a multimodal SQL database interactive analysis and intelligent visualization decision-making system, which relates to the field of intelligent database analysis. Background Technology

[0002] In today's data-driven era, enterprises have an ever-increasing demand for database data analysis and utilization. However, traditional SQL database interaction methods require users to master professional SQL language, which poses a high barrier to entry for non-technical personnel. Therefore, the combination of natural language processing and database interaction technology has become a research hotspot, but existing technologies still have shortcomings in practical applications, such as:

[0003] Semantic gap problem: Existing NLP2SQL tools perform poorly in parsing complex queries, for example, achieving less than 62% accuracy when handling queries containing nested clauses or multi-table joins. Existing tools may fail to accurately resolve relationships between concepts, thus generating incorrect SQL statements;

[0004] Lack of Visual Intelligence: Most existing solutions only render basic charts, such as simple bar charts and line charts, lacking automatic mining of deep data relationships. Current technologies struggle to conduct in-depth analysis of the relationships between dimensions, failing to provide valuable decision support.

[0005] Poor domain adaptability: Different industries have unique terminology and data characteristics, making it difficult for existing systems to be applied across different domains. Existing systems lack effective integration and utilization of domain knowledge, resulting in low accuracy when processing data. Summary of the Invention

[0006] This invention addresses the problems of existing technologies by providing a multimodal SQL database interactive analysis and intelligent visualization decision-making system. It can accurately understand the user's natural language query intent, generate correct SQL statements, realize intelligent visualization analysis of complex data, and quickly adapt to domain knowledge of different industries, providing enterprise users with efficient, convenient, and intelligent data analysis solutions.

[0007] The specific solution proposed in this invention is as follows:

[0008] This invention also provides a multimodal SQL database interactive analysis and intelligent visualization decision-making method, including:

[0009] Step 1: Implement the conversion from natural language to SQL statements through a multi-stage semantic parsing and verification mechanism:

[0010] Step 11: Based on a domain-adaptive feature decoupling method, perform multi-industry joint fine-tuning of the open-source BERT base model, and use the fine-tuned model to parse natural language query statements.

[0011] Step 12: During query parsing, a recurrent neural network (GRU) is used to dynamically adjust the semantic weight of each keyword based on the current input and previous hidden states. This identifies core query conditions and assigns them higher semantic weights to obtain the user's query intent.

[0012] Step 13: Use a sliding window to maintain the frequency of field occurrences, determine if there are any semantic risks in the query intent, and if not, generate an SQL statement based on the query intent and perform the data query; otherwise, re-obtain the user's query intent.

[0013] Step 2: Based on the characteristics of the query results data, automatically generate a suitable visualization map:

[0014] Step 21: For numerical data: Automatically calculate statistical indicators, including skewness, kurtosis, and JS divergence.

[0015] For time series data: Perform STL decomposition to break down the time series data into trend components, seasonal components, and residual components.

[0016] Step 22: Select a visualization scheme based on a decision tree algorithm using data type and dimensional features.

[0017] Step 3: Continuous optimization through a dual-channel feedback mechanism: Explicit feedback is used to correct the generated SQL statements or visualization results, and the correction records are stored in the training pool in real time for incremental training of the model; implicit feedback is used to analyze user interaction behavior characteristics, including query input duration, click location, and dwell time, and information is extracted based on these interaction behavior characteristics for model optimization.

[0018] Furthermore, the fine-tuning of step 11 of the multimodal SQL database interaction analysis and intelligent visualization decision-making method includes: pre-training with tens of millions of domain corpora containing 12 target industries, with each industry corpus having a scale of no less than 500,000 entries; in the feature normalization stage, by calculating the covariance matrix of the feature matrix, performing eigenvalue decomposition to generate a whitening matrix, and performing dimensionality reduction and normalization processing on the high-dimensional features.

[0019] Furthermore, the formula for calculating semantic weight in step 12 of the multimodal SQL database interaction analysis and intelligent visualization decision-making method is as follows:

[0020]

[0021] Where αt: the attention semantic weight of the t-th word, used to measure the importance of the word in the query statement; υ: the model parameter vector, used to perform linear mapping on the features after non-linear transformation; Wh The model parameter matrix is ​​used to perform a linear transformation on the hidden state of the GRU at time t to extract key information from that hidden state.

[0022] h t The hidden state of GRU at time t contains contextual semantic information up to the t-th word.

[0023] W q The model parameter matrix is ​​used to perform a linear transformation on the overall feature vector of the query statement to accommodate the calculation of attention weights.

[0024] q: The overall feature vector of the query statement, which comprehensively represents the semantic information of the entire query statement.

[0025] b: Bias term, used to adjust the result after linear transformation, increasing the flexibility of the model.

[0026] T: The total number of words in the query statement, i.e., the sequence length.

[0027] t ' : The index used to iterate through all words in a query statement, ranging from 1 to T.

[0028] tanh: Hyperbolic tangent activation function, which compresses features to the range [-1, 1].

[0029] Furthermore, in step 22 of the multimodal SQL database interactive analysis and intelligent visualization decision-making method, the visualization scheme is selected, including: for time series data, selecting a line chart or heatmap according to the number of dimensions; for categorical data, selecting a bar chart or Sankey diagram according to the number of categories; for numerical data, matching a scatter plot or 3D distribution map according to the dimensions, with 12 preset basic chart templates and support for user-defined extension rules.

[0030] Furthermore, step 2 of the multimodal SQL database interaction analysis and intelligent visualization decision-making method also includes step 23: using a knowledge graph embedding method to map the database and the industry terminology library to the same vector space. When a new industry term is added, the semantic similarity between the term and the existing fields is calculated using the TransE algorithm to establish a mapping relationship. The model update adopts a parameter isolation method, adjusting only the network parameters of the domain-specific layer to maintain the stability of the basic capabilities.

[0031] This invention provides a multimodal SQL database interactive analysis and intelligent visualization decision-making system, including an intelligent semantic conversion module, a visualization decision-making module, and a continuous learning module.

[0032] The intelligent semantic conversion module achieves the conversion of natural language into SQL statements through a multi-stage semantic parsing and verification mechanism:

[0033] Step 11: Based on a domain-adaptive feature decoupling method, perform multi-industry joint fine-tuning of the open-source BERT base model, and use the fine-tuned model to parse natural language query statements.

[0034] Step 12: During query parsing, a recurrent neural network (GRU) is used to dynamically adjust the semantic weight of each keyword based on the current input and previous hidden states. This identifies core query conditions and assigns them higher semantic weights to obtain the user's query intent.

[0035] Step 13: Use a sliding window to maintain the frequency of field occurrences, determine if there are any semantic risks in the query intent, and if not, generate an SQL statement based on the query intent and perform the data query; otherwise, re-obtain the user's query intent.

[0036] The visualization decision-making module automatically generates appropriate visualization maps based on the characteristics of the query result data:

[0037] Step 21: For numerical data: Automatically calculate statistical indicators, including skewness, kurtosis, and JS divergence.

[0038] For time series data: Perform STL decomposition to break down the time series data into trend components, seasonal components, and residual components.

[0039] Step 22: Select a visualization scheme based on a decision tree algorithm using data type and dimensional features.

[0040] The continuous learning module performs continuous optimization through a dual-channel feedback mechanism: explicit feedback corrects the generated SQL statements or visualization results, and the correction records are stored in the training pool in real time for incremental training of the model; implicit feedback analyzes the user's interaction behavior characteristics, including the duration of query input, the location of click operations, and the dwell time, and extracts information based on the interaction behavior characteristics to optimize the model.

[0041] Furthermore, the intelligent semantic conversion module of the multimodal SQL database interaction analysis and intelligent visualization decision-making system performs fine-tuning in step 11, including: pre-training with tens of millions of domain corpora from 12 target industries, with each industry corpus having a scale of no less than 500,000 entries; in the feature normalization stage, by calculating the covariance matrix of the feature matrix, performing eigenvalue decomposition to generate a whitening matrix, and performing dimensionality reduction and normalization processing on the high-dimensional features.

[0042] Furthermore, the formula for calculating the semantic weight in step 12 of the intelligent semantic transformation module of the multimodal SQL database interaction analysis and intelligent visualization decision-making system is as follows:

[0043]

[0044] Where αt: the attention semantic weight of the t-th word, used to measure the importance of the word in the query statement; υ: the model parameter vector, used to perform linear mapping on the features after non-linear transformation; W h The model parameter matrix is ​​used to perform a linear transformation on the hidden state of the GRU at time t to extract key information from that hidden state.

[0045] h t The hidden state of GRU at time t contains contextual semantic information up to the t-th word.

[0046] W q The model parameter matrix is ​​used to perform a linear transformation on the overall feature vector of the query statement to accommodate the calculation of attention weights.

[0047] q: The overall feature vector of the query statement, which comprehensively represents the semantic information of the entire query statement.

[0048] b: Bias term, used to adjust the result after linear transformation, increasing the flexibility of the model.

[0049] T: The total number of words in the query statement, i.e., the sequence length.

[0050] t ' : The index used to iterate through all words in a query statement, ranging from 1 to T.

[0051] tanh: Hyperbolic tangent activation function, which compresses features to the range [-1, 1].

[0052] Furthermore, the visualization decision module of the multimodal SQL database interactive analysis and intelligent visualization decision system selects visualization schemes, including: for time series data, selecting line charts or heatmaps based on the number of dimensions; for categorical data, selecting bar charts or Sankey diagrams based on the number of categories; for numerical data, matching scatter plots or 3D distribution maps by dimension, with 12 preset basic chart templates and support for user-defined extension rules.

[0053] Furthermore, the visualization decision module of the multimodal SQL database interactive analysis and intelligent visualization decision system also performs step 23: using a knowledge graph embedding method to map the database and the industry terminology library to the same vector space. When a new industry term is added, the semantic similarity between the term and the existing fields is calculated using the TransE algorithm to establish a mapping relationship. The model update adopts a parameter isolation method, adjusting only the network parameters of the domain-specific layer to maintain the stability of the basic capabilities.

[0054] The advantages of this invention are:

[0055] Precise decision support: Improves the accuracy of complex queries based on a triple semantic mapping mechanism;

[0056] Optimizing efficiency and cost: Visualized decision trees reduce the manufacturing production anomaly analysis cycle from 3 days to 4 hours, saving 8 million yuan in annual operation and maintenance costs per factory; Voice + AR interaction reduces the error rate by 68%, reducing equipment losses by 12 million yuan annually in the manufacturing industry; the adaptation cycle in the insurance field is shortened by 90% to 2.5 person-days, saving 15 million yuan annually in system upgrade costs.

[0057] Sustainable development: Supporting intelligent mapping of dialect terms, grassroots data analysis capabilities have been improved by 15 times, benefiting 30 million people; AR 3D visualization has increased the adoption rate of public opinions on urban planning by 82% and reduced the traffic accident rate by 18%; Visually impaired engineers have achieved a 300% increase in job suitability through voice interaction, promoting employment for special groups.

[0058] Continuous technological evolution: The dual-channel feedback mechanism shortens the model iteration cycle by 83%, improves the accuracy of new domain terminology recognition by 2.3 percentage points per week, and reduces the company's total cost of ownership by 180 million yuan over 5 years. Attached Figure Description

[0059] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0060] Figure 1 This is a schematic diagram of the application process of the method of the present invention.

[0061] Figure 2 This is a schematic diagram of the intelligent semantic conversion process in this invention.

[0062] Figure 3 This is a schematic diagram of the visual decision-making process in this invention.

[0063] Figure 4 This is a schematic diagram of the BERT-Whitening algorithm.

[0064] Figure 5 This is a schematic diagram of the field correlation matrix update logic. Detailed Implementation

[0065] The present invention will be further described below with reference to the accompanying drawings and specific embodiments, so that those skilled in the art can better understand and implement the present invention. However, the embodiments described are not intended to limit the present invention.

[0066] Example 1

[0067] This invention also provides a multimodal SQL database interactive analysis and intelligent visualization decision-making method, including:

[0068] Step 1: Implement the conversion from natural language to SQL statements through a multi-stage semantic parsing and verification mechanism:

[0069] Step 11: Based on the domain-adaptive feature decoupling method, the open-source BERT base model is jointly fine-tuned across multiple industries, and the fine-tuned model is used to parse natural language query statements.

[0070] Fine-tuning may include: pre-training using domain corpora of tens of millions of words from 12 target industries, with each industry corpus having no less than 500,000 words; in the feature normalization stage, calculating the covariance matrix of the feature matrix, performing eigenvalue decomposition to generate a whitening matrix, and performing dimensionality reduction and normalization processing on high-dimensional features.

[0071] Step 12: During the query parsing process, the recurrent neural network GRU dynamically adjusts the semantic weight of each keyword based on the current input and the previous hidden state, identifies the core query conditions, and assigns higher semantic weights to the core query conditions to obtain the user's query intent.

[0072] The formula for calculating semantic weight is as follows:

[0073]

[0074] Where αt: the attention semantic weight of the t-th word, used to measure the importance of the word in the query statement; υ: the model parameter vector, used to perform linear mapping on the features after non-linear transformation; W h The model parameter matrix is ​​used to perform a linear transformation on the hidden state of the GRU at time t to extract key information from that hidden state.

[0075] h t The hidden state of GRU at time t contains contextual semantic information up to the t-th word.

[0076] W q The model parameter matrix is ​​used to perform a linear transformation on the overall feature vector of the query statement to accommodate the calculation of attention weights.

[0077] q: The overall feature vector of the query statement, which comprehensively represents the semantic information of the entire query statement.

[0078] b: Bias term, used to adjust the result after linear transformation, increasing the flexibility of the model.

[0079] T: The total number of words in the query statement, i.e., the sequence length.

[0080] t ' : The index used to iterate through all words in a query statement, ranging from 1 to T.

[0081] tanh: Hyperbolic tangent activation function, which compresses features to the range [-1, 1].

[0082] Step 13: Use a sliding window to maintain the frequency of field occurrences and determine if there are semantic risks in the query intent. If there are no semantic risks, generate an SQL statement based on the query intent and perform a data query; otherwise, re-obtain the user's query intent. The window size is set to the 1000 most recent valid queries to avoid historical noise interference caused by full statistics. When a new query is submitted, if the field appears in the current SQL statement, then the query is performed; otherwise, it decays exponentially. Semantic verification threshold setting: Define a field relevance threshold. If the relevance of the fields is less than or equal to the threshold, then the association between the fields is determined to have semantic risks (such as incorrect joins between unrelated fields across tables). For example, in a bank risk control scenario, the historical co-occurrence frequency of "customer name" and "loan annual interest rate" is extremely low. If both appear in the query, the system will trigger a semantic correction prompt.

[0083] Step 2: Based on the characteristics of the query results data, automatically generate a suitable visualization map:

[0084] Step 21: For numerical data: Automatically calculate statistical indicators, including skewness, kurtosis, and JS divergence.

[0085] For time-series data: Perform STL decomposition to break down the data into trend, seasonal, and residual components. Skewness reflects the symmetry of the data distribution, kurtosis describes the steepness of the distribution, and JS divergence measures the difference between two probability distributions. These metrics provide insights into the data's distribution characteristics, informing the selection of appropriate visualization methods.

[0086] Step 22: Select a visualization scheme based on the decision tree algorithm for data type and dimensional features.

[0087] The visualization options include: for time series data, choose line charts or heatmaps based on the number of dimensions; for categorical data, choose bar charts or Sankey diagrams based on the number of categories; for numerical data, match scatter plots or 3D distribution maps by dimension. There are 12 preset basic chart templates, and users can customize extension rules.

[0088] Step 23: Using a knowledge graph embedding method, the database and the industry terminology library are mapped to the same vector space. When a new industry term is added, the semantic similarity between the term and the existing fields is calculated using the TransE algorithm to establish a mapping relationship. The model update adopts a parameter isolation method, adjusting only the network parameters of the domain-specific layer to maintain the stability of the basic capabilities.

[0089] Step 3: Continuous optimization through a dual-channel feedback mechanism: Explicit feedback is used to correct the generated SQL statements or visualization results, and correction records are stored in the training pool in real time for incremental model training; implicit feedback is used to analyze user interaction behavior characteristics, including query input duration, click location, and dwell time, and information is extracted based on these characteristics for model optimization. Domain-adaptive training methods can be used for optimization in this regard.

[0090] Domain knowledge graph embedding method: The database schema, including table structure, field definitions, and relationships, is vectorized and aligned with an industry terminology database. Through the construction and embedding of the knowledge graph, a better understanding of domain-specific terminology and data relationships is achieved, improving the accuracy of semantic parsing.

[0091] Incremental knowledge updates: Supports incremental knowledge updates without affecting existing capabilities. When new industry terms are added or the database structure changes, the new knowledge is simply integrated into the existing model without retraining the entire model, greatly improving domain adaptation efficiency.

[0092] Example 2

[0093] This invention provides a multimodal SQL database interactive analysis and intelligent visualization decision-making system, including an intelligent semantic conversion module, a visualization decision-making module, and a continuous learning module.

[0094] The intelligent semantic conversion module achieves the conversion of natural language into SQL statements through a multi-stage semantic parsing and verification mechanism:

[0095] Step 11: Based on a domain-adaptive feature decoupling method, perform multi-industry joint fine-tuning of the open-source BERT base model, and use the fine-tuned model to parse natural language query statements.

[0096] Step 12: During query parsing, a recurrent neural network (GRU) is used to dynamically adjust the semantic weight of each keyword based on the current input and previous hidden states. This identifies core query conditions and assigns them higher semantic weights to obtain the user's query intent.

[0097] Step 13: Use a sliding window to maintain the frequency of field occurrences, determine if there are any semantic risks in the query intent, and if not, generate an SQL statement based on the query intent and perform the data query; otherwise, re-obtain the user's query intent.

[0098] The visualization decision-making module automatically generates appropriate visualization maps based on the characteristics of the query result data:

[0099] Step 21: For numerical data: Automatically calculate statistical indicators, including skewness, kurtosis, and JS divergence.

[0100] For time series data: Perform STL decomposition to break down the time series data into trend components, seasonal components, and residual components.

[0101] Step 22: Select a visualization scheme based on a decision tree algorithm using data type and dimensional features.

[0102] The continuous learning module performs continuous optimization through a dual-channel feedback mechanism: explicit feedback corrects the generated SQL statements or visualization results, and the correction records are stored in the training pool in real time for incremental training of the model; implicit feedback analyzes the user's interaction behavior characteristics, including the duration of query input, the location of click operations, and the dwell time, and extracts information based on the interaction behavior characteristics to optimize the model.

[0103] The information interaction and execution process between the modules in the above system are based on the same concept as the method embodiment of the present invention, and the specific details can be found in the description in the method embodiment of the present invention, and will not be repeated here.

[0104] Similarly, the advantages of the system of the present invention are:

[0105] Precise decision support: Improves the accuracy of complex queries based on a triple semantic mapping mechanism;

[0106] Optimizing efficiency and cost: Visualized decision trees reduce the manufacturing production anomaly analysis cycle from 3 days to 4 hours, saving 8 million yuan in annual operation and maintenance costs per factory; Voice + AR interaction reduces the error rate by 68%, reducing equipment losses by 12 million yuan annually in the manufacturing industry; the adaptation cycle in the insurance field is shortened by 90% to 2.5 person-days, saving 15 million yuan annually in system upgrade costs.

[0107] Sustainable development: Supporting intelligent mapping of dialect terms, grassroots data analysis capabilities have been improved by 15 times, benefiting 30 million people; AR 3D visualization has increased the adoption rate of public opinions on urban planning by 82% and reduced the traffic accident rate by 18%; Visually impaired engineers have achieved a 300% increase in job suitability through voice interaction, promoting employment for special groups.

[0108] Continuous technological evolution: The dual-channel feedback mechanism shortens the model iteration cycle by 83%, improves the accuracy of new domain terminology recognition by 2.3 percentage points per week, and reduces the company's total cost of ownership by 180 million yuan over 5 years.

[0109] It should be noted that not all steps and modules in the above processes and system structures are mandatory; some steps or modules can be omitted as needed. The execution order of the steps is not fixed and can be adjusted as required. The system structures described in the above embodiments can be physical or logical structures. That is, some modules may be implemented by the same physical entity, or some modules may be implemented by multiple physical entities, or they may be implemented by certain components in multiple independent devices.

[0110] The embodiments described above are merely preferred embodiments for fully illustrating the present invention, and the scope of protection of the present invention is not limited thereto. Equivalent substitutions or modifications made by those skilled in the art based on the present invention are all within the scope of protection of the present invention. The scope of protection of the present invention is defined by the claims.

Claims

1. A multimodal SQL database interactive analysis and intelligent visualization decision-making method, characterized by: include: Step 1: Implement the conversion from natural language to SQL statements through a multi-stage semantic parsing and verification mechanism: Step 11: Based on a domain-adaptive feature decoupling method, perform multi-industry joint fine-tuning of the open-source BERT base model, and use the fine-tuned model to parse natural language query statements. Step 12: During query parsing, a recurrent neural network (GRU) is used to dynamically adjust the semantic weight of each keyword based on the current input and previous hidden states. This identifies core query conditions and assigns them higher semantic weights to obtain the user's query intent. Step 13: Use a sliding window to maintain the frequency of field occurrences, determine if there are any semantic risks in the query intent, and if not, generate an SQL statement based on the query intent and perform the data query; otherwise, re-obtain the user's query intent. Step 2: Based on the characteristics of the query results data, automatically generate a suitable visualization map: Step 21: For numerical data: Automatically calculate statistical indicators, including skewness, kurtosis, and JS divergence. For time series data: Perform STL decomposition to break down the time series data into trend components, seasonal components, and residual components. Step 22: Select a visualization scheme based on a decision tree algorithm using data type and dimensional features. Step 3: Continuous optimization through a dual-channel feedback mechanism: Correct the generated SQL statements or visualization results through displayed feedback, and save the correction records into the training pool in real time for incremental training of the model; By analyzing implicit feedback, we can identify user interaction behavior characteristics, including query input duration, click location, and dwell time. We can then extract information based on these characteristics to optimize the model.

2. The multimodal SQL database interactive analysis and intelligent visualization decision-making method according to claim 1, characterized in that the fine-tuning in step 11 includes: Pre-training was performed using domain corpora of tens of millions of words from 12 target industries, with each industry corpus containing no less than 500,000 words. In the feature normalization stage, the covariance matrix of the feature matrix was calculated, and eigenvalue decomposition was performed to generate a whitening matrix, thereby reducing and normalizing the high-dimensional features.

3. The multimodal SQL database interactive analysis and intelligent visualization decision-making method according to claim 1, characterized in that: The formula for calculating semantic weights in step 12 is as follows: Where αt: the attention semantic weight of the t-th word, used to measure the importance of the word in the query statement; υ: the model parameter vector, used to perform linear mapping on the features after non-linear transformation; W h The model parameter matrix is ​​used to perform a linear transformation on the hidden state of the GRU at time t to extract key information from that hidden state. h t The hidden state of GRU at time t contains contextual semantic information up to the t-th word. W q The model parameter matrix is ​​used to perform a linear transformation on the overall feature vector of the query statement to accommodate the calculation of attention weights. q: The overall feature vector of the query statement, which comprehensively represents the semantic information of the entire query statement. b: Bias term, used to adjust the result after linear transformation, increasing the flexibility of the model. T: The total number of words in the query statement, i.e., the sequence length. t ' : The index used to iterate through all words in a query statement, ranging from 1 to T. tanh: Hyperbolic tangent activation function, which compresses features to the range [-1, 1].

4. The multimodal SQL database interactive analysis and intelligent visualization decision-making method according to claim 1, characterized in that: In step 22, select a visualization scheme, including: for time series data, choose a line chart or heatmap based on the number of dimensions; for categorical data, choose a bar chart or Sankey diagram based on the number of categories; for numerical data, match a scatter plot or 3D distribution map by dimension. There are 12 preset basic chart templates, and users can customize extension rules.

5. The multimodal SQL database interactive analysis and intelligent visualization decision-making method according to claim 1, characterized in that: Step 2 also includes step 23: using a knowledge graph embedding method to map the database and the industry terminology library to the same vector space. When a new industry term is added, the semantic similarity between the term and the existing fields is calculated using the TransE algorithm to establish a mapping relationship. The model update adopts a parameter isolation method, adjusting only the network parameters of the domain-specific layer to maintain the stability of the basic capabilities.

6. A multimodal SQL database interactive analysis and intelligent visualization decision-making system, characterized by: It includes an intelligent semantic conversion module, a visual decision-making module, and a continuous learning module. The intelligent semantic conversion module achieves the conversion of natural language into SQL statements through a multi-stage semantic parsing and verification mechanism: Step 11: Based on a domain-adaptive feature decoupling method, perform multi-industry joint fine-tuning of the open-source BERT base model, and use the fine-tuned model to parse natural language query statements. Step 12: During query parsing, a recurrent neural network (GRU) is used to dynamically adjust the semantic weight of each keyword based on the current input and previous hidden states. This identifies core query conditions and assigns them higher semantic weights to obtain the user's query intent. Step 13: Use a sliding window to maintain the frequency of field occurrences, determine if there are any semantic risks in the query intent, and if not, generate an SQL statement based on the query intent and perform the data query; otherwise, re-obtain the user's query intent. The visualization decision-making module automatically generates appropriate visualization maps based on the characteristics of the query result data: Step 21: For numerical data: Automatically calculate statistical indicators, including skewness, kurtosis, and JS divergence. For time series data: Perform STL decomposition to break down the time series data into trend components, seasonal components, and residual components. Step 22: Select a visualization scheme based on a decision tree algorithm using data type and dimensional features. The continuous learning module performs continuous optimization through a dual-channel feedback mechanism: explicit feedback corrects the generated SQL statements or visualization results, and the correction records are stored in the training pool in real time for incremental training of the model; implicit feedback analyzes the user's interaction behavior characteristics, including the duration of query input, the location of click operations, and the dwell time, and extracts information based on the interaction behavior characteristics to optimize the model.

7. The multimodal SQL database interactive analysis and intelligent visualization decision-making system according to claim 6, characterized in that: The intelligent semantic conversion module performs fine-tuning in step 11, including: pre-training with tens of millions of domain corpora from 12 target industries, with each industry corpus having a scale of no less than 500,000 entries; in the feature normalization stage, calculating the covariance matrix of the feature matrix, performing eigenvalue decomposition to generate a whitening matrix, and performing dimensionality reduction and normalization on high-dimensional features.

8. A multimodal SQL database interactive analysis and intelligent visualization decision-making system according to claim 6, characterized in that: The formula for calculating semantic weights in step 12 of the intelligent semantic transformation module is as follows: Where αt: the attention semantic weight of the t-th word, used to measure the importance of the word in the query statement; υ: the model parameter vector, used to perform linear mapping on the features after non-linear transformation; W h The model parameter matrix is ​​used to perform a linear transformation on the hidden state of the GRU at time t to extract key information from that hidden state. h t The hidden state of GRU at time t contains contextual semantic information up to the t-th word. W q The model parameter matrix is ​​used to perform a linear transformation on the overall feature vector of the query statement to accommodate the calculation of attention weights. q: The overall feature vector of the query statement, which comprehensively represents the semantic information of the entire query statement. b: Bias term, used to adjust the result after linear transformation, increasing the flexibility of the model. T: The total number of words in the query statement, i.e., the sequence length. t ' : The index used to iterate through all words in a query statement, ranging from 1 to T. tanh: Hyperbolic tangent activation function, which compresses features to the range [-1, 1].

9. The multimodal SQL database interactive analysis and intelligent visualization decision-making system according to claim 6, characterized in that it is visualized. The decision-making module selects visualization options, including: for time series data, choosing line charts or heatmaps based on the number of dimensions; for categorical data, choosing bar charts or Sankey diagrams based on the number of categories; and for numerical data, matching scatter plots or 3D distribution maps by dimension. It offers 12 preset basic chart templates and supports user-defined extension rules.

10. The multimodal SQL database interactive analysis and intelligent visualization decision-making system according to claim 6, characterized in that it is visualized. The decision module also performs step 23: using a knowledge graph embedding method, the database and the industry terminology library are mapped to the same vector space. When a new industry term is added, the semantic similarity between the term and the existing fields is calculated using the TransE algorithm to establish a mapping relationship. The model update adopts a parameter isolation method, adjusting only the network parameters of the domain-specific layer to maintain the stability of the basic capabilities.

Images