A low-code-based data analysis dashboard automatic generation method and system
By using natural language processing and reinforcement learning, data analysis dashboards are automatically generated, solving the problem of traditional methods relying on professional developers and enabling non-professional users to generate and continuously optimize data analysis dashboards autonomously.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- CHINA NAT BUILDING MATERIALS TECH CO LTD
- Filing Date
- 2025-08-20
- Publication Date
- 2026-05-12
AI Technical Summary
Traditional data analytics dashboards rely on professional developers, are time-consuming and costly, lack fine-grained understanding of natural language intent and have low automation levels, making it difficult to meet the generation needs of non-professional users.
By using a natural language processing module for dependency parsing and entity recognition, combined with semantic indexing and reinforcement learning, a data analysis dashboard is automatically generated, achieving a precise mapping from natural language to visualization components. The layout is further optimized through graph optimization and user interaction.
It significantly reduces dashboard construction time and manpower costs, enabling non-professional users to generate and continuously optimize data analysis dashboards independently.
Smart Images
Figure CN121116271B_ABST
Abstract
Description
Technical Field
[0001] This document relates to the field of computer technology, and in particular to a method and system for automatically generating data analysis dashboards based on low-code. Background Technology
[0002] Traditional data analytics dashboard construction typically relies on professional developers manually writing SQL, configuring charts, and repeatedly adjusting the layout, which is time-consuming and costly. While low-code platforms that have emerged in recent years have lowered the barrier to entry by allowing users to drag and drop components, users still need to manually select chart types, bind data fields, and arrange the layout, making it difficult to meet the dashboard generation needs of non-professional users.
[0003] Existing technologies also have the following drawbacks: lack of fine-grained understanding of natural language intent, resulting in a large deviation between recommended components and user expectations; layout optimization relies on static rules and cannot continuously learn user interaction behavior; the data binding process requires manual specification of fields, resulting in low automation. Summary of the Invention
[0004] This specification provides one or more embodiments of a low-code-based method for automatically generating data analysis dashboards, including:
[0005] The natural language processing module performs dependency parsing and entity recognition on the input text to extract dataset identifiers and user intent vectors.
[0006] The user intent vector is mapped to a pre-constructed semantic index space, which consists of several high-dimensional vectors, each of which corresponds to a semantic description of a visualization component.
[0007] The degree of matching between the semantic description of the visualization component in the semantic index space and the user intent vector is measured based on the cosine similarity metric, and a candidate component set is obtained based on the degree of matching.
[0008] Using the candidate component set as nodes and the semantic correlation between components as edge weights, a weighted undirected graph is constructed, and the maximum spanning tree algorithm is used to generate the initial layout topology.
[0009] Based on the initial layout topology, a data source field is bound to each candidate component based on the dataset identifier. The binding is achieved by matching the field type with the pattern of the component's input signature.
[0010] Listen for user drag events and record component position offsets. Use these offsets as training samples to input into the reinforcement learning network and update the layout strategy network parameters.
[0011] The optimized layout topology is serialized into a configuration object that conforms to the target runtime specification and pushed to the client.
[0012] Furthermore, the natural language processing module is specifically used for:
[0013] A BERT-based pre-trained language model is used to generate the context embedding of the user intent vector;
[0014] An entity recognition subnetwork based on BiLSTM-CRF is used to extract dataset identifiers.
[0015] Furthermore, the construction process of the semantic index space includes:
[0016] Word2Vec training is performed on the metadata of each visualization component to obtain component-level word vectors;
[0017] The word vectors are reduced to 512 dimensions using PCA to form the high-dimensional vectors.
[0018] Furthermore, the maximum spanning tree algorithm adopts Kruskal's algorithm, and the sum of the component canvas areas is used as a constraint when merging sets.
[0019] Furthermore, the pattern matching is achieved through the following sub-steps:
[0020] Parse the JSON schema of the component input signature;
[0021] Find a subset of fields in the data source field list that satisfy the JSON Schema;
[0022] Automatically select aggregate functions and generate SQL fragments based on the statistical distribution of fields.
[0023] Furthermore, the reinforcement learning network is a near-end policy optimization network, whose state space is a graph embedding of the current layout topology, whose action space is a discretized grid of component position offsets, and whose reward function is a weighted sum of click rate and dwell time.
[0024] Furthermore, the configuration object is in JSON Patch format and is transmitted via a WebSocket channel in gzip compressed stream.
[0025] This specification provides one or more embodiments of a low-code-based automatic data analysis dashboard generation system, including:
[0026] Semantic parsing module: Used to perform dependency parsing and entity recognition on the input text through the natural language processing module, and extract dataset identifiers and user intent vectors;
[0027] Semantic indexing module: used to map the user intent vector to a pre-built semantic index space, which consists of several high-dimensional vectors, each high-dimensional vector corresponding to a semantic description of a visualization component;
[0028] Matching module: used to measure the degree of matching between the semantic description of the visualization component in the semantic index space and the user intent vector based on the cosine similarity metric, and to obtain a set of candidate components based on the degree of matching;
[0029] Layout generation module: used to construct a weighted undirected graph with the candidate component set as nodes and the semantic correlation between components as edge weights, and to generate the initial layout topology using the maximum spanning tree algorithm;
[0030] Data binding module: used to bind data source fields to each candidate component based on the dataset identifier according to the initial layout topology. The binding is achieved by matching the field type with the pattern of the component input signature.
[0031] Interaction optimization module: used to listen for user drag events and record component position offsets, and input the offsets as training samples into the reinforcement learning network to update the layout strategy network parameters;
[0032] Rendering module: Used to serialize the optimized layout topology into a configuration object that conforms to the target runtime specification and push it to the client.
[0033] This specification provides one or more embodiments of an electronic device, including:
[0034] Processor; and,
[0035] A memory is configured to store computer-executable instructions, which, when executed, cause the processor to implement the steps of the above-described low-code-based automatic data analysis dashboard generation method.
[0036] This specification provides one or more embodiments of a storage medium for storing computer-executable instructions that, when executed, implement the steps of the above-described low-code-based automatic data analysis dashboard generation method.
[0037] By employing embodiments of the present invention, accurate mapping from natural language to visualization components is achieved through semantic parsing and indexing; based on graph optimization and reinforcement learning, the layout can continuously evolve with user behavior; and end-to-end automation significantly reduces dashboard construction time and manpower costs.
[0038] The above description is merely an overview of the technical solution of the present invention. In order to better understand the technical means of the present invention and to implement it in accordance with the contents of the specification, and in order to make the above and other objects, features and advantages of the present invention more apparent and understandable, specific embodiments of the present invention are described below. Attached Figure Description
[0039] To more clearly illustrate the technical solutions in one or more embodiments of this specification or in 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 only some embodiments recorded in this specification. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0040] Figure 1 A flowchart illustrating a low-code-based method for automatically generating data analysis dashboards, provided for one or more embodiments of this specification;
[0041] Figure 2 A schematic diagram illustrating the composition of a low-code-based automatic data analysis dashboard generation system provided for one or more embodiments of this specification;
[0042] Figure 3 This is a schematic diagram of an electronic device structure provided for one or more embodiments of this specification. Detailed Implementation
[0043] To enable those skilled in the art to better understand the technical solutions in one or more embodiments of this specification, the technical solutions in one or more embodiments of this specification will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this specification, and not all of the embodiments. Based on one or more embodiments of this specification, all other embodiments obtained by those skilled in the art without creative effort should fall within the protection scope of this document.
[0044] Method Implementation Examples
[0045] According to embodiments of the present invention, a method for automatically generating data analysis dashboards based on low-code is provided. Figure 1 A flowchart illustrating a low-code-based method for automatically generating data analysis dashboards, provided for one or more embodiments of this specification, is shown below. Figure 1 As shown, the low-code-based automatic data analysis dashboard generation method according to an embodiment of the present invention specifically includes:
[0046] S1. The input text is subjected to dependency parsing and entity recognition by the natural language processing module to extract the dataset identifier and user intent vector.
[0047] The natural language processing module is specifically used for: a BERT-based pre-trained language model for generating contextual embeddings of the user intent vector; and a BiLSTM-CRF-based entity recognition sub-network for extracting dataset identifiers.
[0048] This module first uses dependency parsing to deconstruct the grammatical structure of sentences, identify the dominance relationships between words, and construct a grammatical dependency tree to understand the deep logical framework of the statement. Simultaneously, it performs entity recognition, employing a dedicated sub-network based on a BiLSTM-CRF architecture combining bidirectional long short-term memory (LSTM) and conditional random fields. The bidirectional LSTM layer captures the contextual feature sequence of the text, and the CRF layer optimizes the label transition probability, accurately labeling and extracting specific types of named entities, including dataset identifiers. Meanwhile, this module uses a BERT pre-trained language model based on the Transformer architecture to encode the complete input text: it dynamically fuses the global contextual semantics of words through a multi-layer self-attention mechanism to generate a context-sensitive embedding representation for each token; then, pooling operations compress the variable-length text sequence into a fixed-dimensional dense vector, forming a user intent vector representing the user's core query intent. The user intent vector includes explicit semantic features and the text's sentiment, intent strength, and contextual implicit information, achieving a transformation from unstructured natural language to structured, machine-processable features.
[0049] S2. Map the user intent vector to a pre-constructed semantic index space, which consists of several high-dimensional vectors, each high-dimensional vector corresponding to a semantic description of a visualization component.
[0050] The method for constructing the semantic index space is as follows: First, the metadata description text of each visualization component is trained using a Word2Vec model for distributed vector representation training. The metadata description text includes natural language descriptions such as component name, function description, and parameter definition. The semantic associations of component description words are learned through a skip word or continuous bag-of-words architecture to generate a set of component-level word vectors. Then, principal component analysis (PCA) is performed on these high-dimensional sparse word vectors to reduce dimensionality, extract the main semantic features in the direction of maximum variance, and eliminate redundant dimensions. Finally, the semantic description of each visualization component is compressed and represented as a 512-dimensional dense vector of a uniform scale. These standardized vectors constitute a semantic index space in mathematical space.
[0051] The generated user intent vector is projected onto the pre-built semantic index space. The semantic alignment between the intent and the visualization components is achieved through coordinate positioning in the vector space. The spatial relative position of the user intent vector and the vectors of each component directly quantifies the semantic matching degree.
[0052] S3. Based on the cosine similarity metric, measure the degree of matching between the semantic description of the visualization component in the semantic index space and the user intent vector, and obtain a set of candidate components based on the degree of matching.
[0053] The cosine similarity algorithm is used to accurately calculate the directional consistency coefficient between the user intent vector and the semantic vector of each visualization component, quantifying the degree of matching between the user query semantics and the component function. After calculation, a dual screening strategy is implemented: first, a preset semantic threshold is applied to filter low-relevance components, and then the top K high-matching items are dynamically selected from the qualified components in descending order of similarity to form a candidate set, ensuring that the results have both semantic accuracy and layout feasibility.
[0054] S4. Using the candidate component set as nodes and the semantic correlation between components as edge weights, construct a weighted undirected graph, and use the maximum spanning tree algorithm to generate the initial layout topology.
[0055] Using the candidate component set as nodes, a fully connected weighted undirected graph is constructed. The weight of each edge in the graph is determined by the cosine similarity of the semantic vectors of the components corresponding to the connected nodes, which is normalized and converted into a semantic association strength value in the range [0,1]. Based on this graph structure, an improved Kruskal algorithm is used to generate the maximum spanning tree: the edges are sorted in descending order of weight and processed iteratively. When attempting to merge two disconnected subtrees, the estimated total area of the canvas containing all components is calculated in real time. The merge operation is performed and the edge is added to the spanning tree only if this value does not exceed the dynamic canvas threshold.
[0056] S5. Based on the initial layout topology, bind a data source field to each candidate component based on the dataset identifier. The binding is achieved by matching the field type with the pattern of the component input signature.
[0057] The pattern matching is achieved through the following sub-steps:
[0058] Parse the JSON schema of the component input signature;
[0059] Find a subset of fields in the data source field list that satisfy the JSON Schema;
[0060] Automatically select aggregate functions and generate SQL fragments based on the statistical distribution of fields.
[0061] Based on candidate component nodes in the initial layout topology, automated data source binding is achieved. The specific method is as follows: First, the JSON Schema input signature of each component is parsed to extract structured definitions such as field names, data types, and required / optional constraints. Then, a subset of fields matching the Schema requirements is searched in the target dataset's field list, performing strict type compatibility checks. For example, when the Schema defines numeric fields, only integer / floating-point data columns are matched; time-series fields must match date-time formats. After field matching is completed, the system analyzes the statistical characteristics of the matched fields and automatically selects aggregation functions based on the field statistical distribution to ultimately synthesize component-specific SQL query fragments. This achieves a precise transformation from semantic layout topology to an executable data view, ensuring that each visualization component receives type-safe and statistically adapted data.
[0062] S6. Listen for user drag events and record component position offsets. Use the offsets as training samples to input into the reinforcement learning network and update the layout strategy network parameters.
[0063] The reinforcement learning network is a near-end policy optimization network, whose state space is a graph embedding of the current layout topology, action space is a discretized grid of component position offsets, and reward function is a weighted sum of click rate and dwell time.
[0064] By monitoring users' component dragging behavior on the visual interface in real time, the system accurately records the position offsets (Δx, Δy) and transforms them into training samples for the reinforcement learning network, driving continuous iterative optimization of the layout strategy. This reinforcement learning framework employs a proximal policy optimization (PPO) network architecture. Its state space consists of graph embedding vectors of the current layout topology, while the action space is defined as discretized grid operations: the canvas is divided into an N×N virtual grid, and component position offsets are abstracted into a finite set of actions, such as moving up / down k grids, moving left / right k grids, scaling (±r%). The policy network outputs the action probability distribution to guide the component's movement direction and magnitude. The reward function comprehensively quantifies the value of user interaction, calculated through a weighted sum of click-through rate and normalized dwell time. During training, the offsets triggered by user dragging are first discretized into specific action executions. The system then records the state-action-reward transition sequence and stores it in the experience pool. When the samples accumulate to a threshold, the PPO algorithm constrains the policy update magnitude through a pruning probability ratio objective function, maximizing accumulated rewards while maintaining training stability, and simultaneously updating the value function network to optimize state evaluation. An action masking mechanism is introduced to prevent invalid operations that cause component overlap or out-of-bounds, and a layout compactness penalty term is embedded in the reward calculation. The policy network parameters are adjusted in real time through online incremental learning, and the client periodically synchronizes the updated model. At the same time, a reward decline monitoring and historical best policy rollback mechanism are set up to ensure that the system remains robust as it continuously adapts to user space preferences.
[0065] S7. Serialize the optimized layout topology into a configuration object that conforms to the target runtime specification and push it to the client.
[0066] The configuration object is in JSON Patch format and is transmitted via a WebSocket channel in gzip compressed stream.
[0067] The beneficial effects of this invention are as follows:
[0068] By employing embodiments of the present invention, accurate mapping from natural language to visualization components is achieved through semantic parsing and indexing; based on graph optimization and reinforcement learning, the layout can continuously evolve with user behavior; and end-to-end automation significantly reduces dashboard construction time and manpower costs.
[0069] System Implementation Examples
[0070] According to embodiments of the present invention, a low-code-based automatic data analysis dashboard generation system is provided. Figure 2 A schematic diagram illustrating the composition of a low-code-based automatic data analysis dashboard generation system provided in one or more embodiments of this specification, such as... Figure 2 As shown, the low-code-based automatic data analysis dashboard generation system according to an embodiment of the present invention specifically includes:
[0071] Semantic parsing module: Used to perform dependency parsing and entity recognition on the input text through the natural language processing module, and extract dataset identifiers and user intent vectors;
[0072] Semantic indexing module: used to map the user intent vector to a pre-built semantic index space, which consists of several high-dimensional vectors, each high-dimensional vector corresponding to a semantic description of a visualization component;
[0073] Matching module: used to measure the degree of matching between the semantic description of the visualization component in the semantic index space and the user intent vector based on the cosine similarity metric, and to obtain a set of candidate components based on the degree of matching;
[0074] Layout generation module: used to construct a weighted undirected graph with the candidate component set as nodes and the semantic correlation between components as edge weights, and to generate the initial layout topology using the maximum spanning tree algorithm;
[0075] Data binding module: used to bind data source fields to each candidate component based on the dataset identifier according to the initial layout topology. The binding is achieved by matching the field type with the pattern of the component input signature.
[0076] Interaction optimization module: used to listen for user drag events and record component position offsets, and input the offsets as training samples into the reinforcement learning network to update the layout strategy network parameters;
[0077] Rendering module: Used to serialize the optimized layout topology into a configuration object that conforms to the target runtime specification and push it to the client.
[0078] The embodiments of the present invention are system embodiments corresponding to the above method embodiments. The specific operation of each module can be understood by referring to the description of the method embodiments, and will not be repeated here.
[0079] Device Example 1
[0080] This invention provides an electronic device, such as... Figure 3 As shown, it includes: a memory 30, a processor 32, and a computer program stored in the memory 30 and executable on the processor 32. When the computer program is executed by the processor 32, it performs the following method steps:
[0081] S1. Perform dependency parsing and entity recognition on the input text using the natural language processing module to extract dataset identifiers and user intent vectors;
[0082] S2. Map the user intent vector to a pre-constructed semantic index space, which consists of several high-dimensional vectors, each high-dimensional vector corresponding to a semantic description of a visualization component;
[0083] S3. Based on the cosine similarity metric, measure the degree of matching between the semantic description of the visual components in the semantic index space and the user intent vector, and obtain a set of candidate components based on the degree of matching;
[0084] S4. Using the candidate component set as nodes and the semantic correlation between components as edge weights, construct a weighted undirected graph, and use the maximum spanning tree algorithm to generate the initial layout topology;
[0085] S5. Based on the initial layout topology, bind a data source field to each candidate component based on the dataset identifier. The binding is achieved by matching the field type with the pattern of the component input signature.
[0086] S6. Listen for user drag events and record component position offsets. Use the offsets as training samples to input into the reinforcement learning network and update the layout strategy network parameters.
[0087] S7. Serialize the optimized layout topology into a configuration object that conforms to the target runtime specification and push it to the client.
[0088] Device Example 2
[0089] This invention provides a computer-readable storage medium storing an information transmission implementation program. When executed by a processor 32, the program performs the following method steps:
[0090] S1. Perform dependency parsing and entity recognition on the input text using the natural language processing module to extract dataset identifiers and user intent vectors;
[0091] S2. Map the user intent vector to a pre-constructed semantic index space, which consists of several high-dimensional vectors, each high-dimensional vector corresponding to a semantic description of a visualization component;
[0092] S3. Based on the cosine similarity metric, measure the degree of matching between the semantic description of the visual components in the semantic index space and the user intent vector, and obtain a set of candidate components based on the degree of matching;
[0093] S4. Using the candidate component set as nodes and the semantic correlation between components as edge weights, construct a weighted undirected graph, and use the maximum spanning tree algorithm to generate the initial layout topology;
[0094] S5. Based on the initial layout topology, bind a data source field to each candidate component based on the dataset identifier. The binding is achieved by matching the field type with the pattern of the component input signature.
[0095] S6. Listen for user drag events and record component position offsets. Use the offsets as training samples to input into the reinforcement learning network and update the layout strategy network parameters.
[0096] S7. Serialize the optimized layout topology into a configuration object that conforms to the target runtime specification and push it to the client.
[0097] The computer-readable storage media described in this embodiment include, but are not limited to, ROM, RAM, disk, or optical disk.
[0098] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.
Claims
1. A method for automatically generating data analysis dashboards based on low-code, characterized in that, include: The natural language processing module performs dependency parsing and entity recognition on the input text to extract dataset identifiers and user intent vectors. The user intent vector is mapped to a pre-constructed semantic index space, which consists of several high-dimensional vectors, each of which corresponds to a semantic description of a visualization component. The degree of matching between the semantic description of the visualization component in the semantic index space and the user intent vector is measured based on the cosine similarity metric, and a candidate component set is obtained based on the degree of matching. Using the candidate component set as nodes and the semantic correlation between components as edge weights, a weighted undirected graph is constructed, and the maximum spanning tree algorithm is used to generate the initial layout topology. Based on the initial layout topology, a data source field is bound to each candidate component based on the dataset identifier. The binding is achieved by matching the field type with the pattern of the component's input signature. Listen for user drag events and record component position offsets. Use these offsets as training samples to input into the reinforcement learning network and update the layout strategy network parameters. The optimized layout topology is serialized into a configuration object that conforms to the target runtime specification and pushed to the client.
2. The method according to claim 1, characterized in that, The natural language processing module is specifically used for: A BERT-based pre-trained language model is used to generate the context embedding of the user intent vector; An entity recognition subnetwork based on BiLSTM-CRF is used to extract dataset identifiers.
3. The method according to claim 1, characterized in that, The process of constructing the semantic index space includes: Word2Vec training is performed on the metadata of each visualization component to obtain component-level word vectors; The word vectors are reduced to 512 dimensions using PCA to form the high-dimensional vectors.
4. The method according to claim 1, characterized in that, The maximum spanning tree algorithm adopts Kruskal's algorithm, and the sum of the component canvas areas is used as a constraint when merging sets.
5. The method according to claim 1, characterized in that, The pattern matching is achieved through the following sub-steps: Parse the JSON schema of the component input signature; Find a subset of fields in the data source field list that satisfy the JSON Schema; Automatically select aggregate functions and generate SQL fragments based on the statistical distribution of fields.
6. The method according to claim 1, characterized in that, The reinforcement learning network is a near-end policy optimization network, whose state space is a graph embedding of the current layout topology, action space is a discretized grid of component position offsets, and reward function is a weighted sum of click rate and dwell time.
7. The method according to claim 1, characterized in that, The configuration object is in JSON Patch format and is transmitted via a WebSocket channel in gzip compressed stream.
8. A low-code-based automatic data analysis dashboard generation system, characterized in that, include: Semantic parsing module: Used to perform dependency parsing and entity recognition on the input text through the natural language processing module, and extract dataset identifiers and user intent vectors; Semantic indexing module: used to map the user intent vector to a pre-built semantic index space, which consists of several high-dimensional vectors, each high-dimensional vector corresponding to a semantic description of a visualization component; Matching module: used to measure the degree of matching between the semantic description of the visualization component in the semantic index space and the user intent vector based on the cosine similarity metric, and to obtain a set of candidate components based on the degree of matching; Layout generation module: used to construct a weighted undirected graph with the candidate component set as nodes and the semantic correlation between components as edge weights, and to generate the initial layout topology using the maximum spanning tree algorithm; Data binding module: used to bind data source fields to each candidate component based on the dataset identifier according to the initial layout topology. The binding is achieved by matching the field type with the pattern of the component input signature. Interaction optimization module: used to listen for user drag events and record component position offsets, and input the offsets as training samples into the reinforcement learning network to update the layout strategy network parameters; Rendering module: Used to serialize the optimized layout topology into a configuration object that conforms to the target runtime specification and push it to the client.
9. An electronic device, characterized in that, include: processor; as well as, A memory is configured to store computer-executable instructions, which, when executed, cause the processor to implement the steps of the low-code-based automatic data analysis dashboard generation method as described in any one of claims 1 to 7.
10. A storage medium, characterized in that, Used to store computer-executable instructions, which, when executed, implement the steps of the low-code-based automatic data analysis dashboard generation method as described in any one of claims 1 to 7.