A Visual Analysis System and Method for Tabular Data Based on Large Language Models

Abstract

This invention discloses an exploratory visual analysis system and method for tabular data based on a large language model, comprising: parsing raw tabular data and building a data model; automatically mining insights and generating three different forms of insights; user-led exploratory data analysis and data story generation in the interaction module; an insight engine in the calculation module finding insights with structural relationships through insight maps; and insight recommendation through semantic association in the reasoning module using a large language model. This invention can model structured tabular data, thereby performing data parsing and insight mining. Automating the mining of potential insights from tables frees users from repetitive and intensive analysis operations. By integrating human intuition and insights, the computational power of external executors, and the intelligence of a large language model, it enhances the interactivity, transparency, and interpretability of the exploratory data analysis process, thereby promoting efficient and intuitive data analysis.

Description

Technical Field

[0001] This invention relates to the field of big data visualization and human-computer interaction, specifically to a tabular data exploratory visual analysis system and method based on a large language model. Background Technology

[0002] Tabular data is a fundamental data format used to represent information across a wide range of fields, from business analytics and scientific research to social sciences and healthcare. However, the sheer volume and complexity of tabular data make understanding and exploration challenging. Effectively exploring tabular data is crucial for making informed decisions, requiring the identification of data insights such as trends and outliers. However, raw tabular data presents a raw and concrete view of information, often lacking the clarity and context needed for deeper understanding. To bridge this gap, transforming extracted data insights into narrative data stories becomes essential. Narrative visualizations offer a powerful way to connect discrete data points, creating coherent data stories to help users understand the key findings. By crafting engaging data stories, these visualizations enhance comprehension, making complex information more accessible and enabling non-specialist audiences to engage more effectively and retain data insights.

[0003] Traditional methods of creating data stories require users to manually extract insights and logically combine multiple story elements—a process that is inefficient and highly dependent on user experience and skills. Extensive research has been conducted to simplify this process. A common approach is to use rule-based algorithms to automatically extract insights from tabular data and generate data stories. However, because these data stories are automatically generated, they may not perfectly align with the user's specific analytical intent. Another approach involves connecting extracted insights to help users discover relevant information, allowing them to participate in the data story creation process. However, this approach remains inefficient because it requires users to rely on their intuition and experience to select and combine relevant insights through trial and error to ultimately form a coherent narrative.

[0004] Data scientists use exploratory data analysis to analyze and study valuable or "interesting" features in data. In this process, data scientists use graphical and visualization tools to observe data from various perspectives and progressively choose specific methods to explore the data according to their interests. For flat data, traditional exploratory data analysis often requires iterative and complex operations, such as filtering, classifying, and integrating. The manual iterative calculation process is not only time-consuming but also creates a high barrier to entry due to the need for specialized domain knowledge. This makes it difficult for ordinary people without professional training and relevant domain knowledge to use exploratory analysis. Even professional data scientists, when dealing with large amounts of data, need to consider too many potential analytical methods, which can affect the efficiency of exploratory analysis.

[0005] Therefore, decoupling users from the complex and time-consuming process of manual data exploration and understanding, and automating the mining and analysis of valuable data fragments (i.e., data insights), is a highly demanded and important task. Existing research has proposed system frameworks for automating data insight mining and established standard specifications for the types of data insights and evaluation criteria for their importance. Building on this, many studies focusing on the development of automated exploration systems aim to design and implement more intelligent recommendation systems to replace users in performing computation and reasoning during exploratory data analysis to the greatest extent possible, directly returning valuable and easily understandable exploration results to users. Generally, existing systems use automated methods to mine data insights, then organize the obtained data insights into an optimal exploration sequence based on certain rules or heuristic evaluation criteria, and finally directly output the corresponding visual representations of the exploration sequence to the user. These existing automata can output richer, more advanced visualizations, demonstrating the enormous potential of data visualization storytelling, but they all generally face key limitations:

[0006] (1) The output results are of poor quality. The data insights that make up the results are not strongly related in terms of semantics and logic, which leads to a significant reduction in the overall quality of the generated high-level data narrative.

[0007] (2) Ignoring user intent. In exploratory data analysis, the user's active exploration is the most important part, and it is often dynamic, adjusting with the feedback of the current exploration results. However, these automated systems have limited means of interaction with users. Users can only perform initial settings and obtain the final results before and after the exploration begins. They cannot intervene in the intermediate process of data insight exploration and organization, which makes these systems unable to capture the dynamic needs of users. The generated results often have a huge gap with the user's actual exploration purpose. Summary of the Invention

[0008] To address the shortcomings of existing technologies, this invention aims to propose a visualization and human-computer interaction method for widely existing structured tabular data, supporting exploratory data analysis and interactive data story generation. The emergence of large language models has brought revolutionary progress to the field of natural language processing. Within a suitable cue word framework, large language models have demonstrated outstanding application potential in scenarios where users understand their intent, infer the next action, and thus achieve user-specified goals. Therefore, the key idea of ​​this method is to fully utilize the excellent semantic understanding, text generation, and reasoning capabilities of large language models. Based on user questions about data insights given in natural language, it performs logical reasoning of the exploration path and generates textual descriptions of the semantic relationships between data insights. This allows for the natural and logical integration of users' proactive exploration needs, constructing high-quality data insight sequences that support full user participation in the construction process of data insight exploration sequences.

[0009] Specifically, the ReInAct framework is proposed for a tabular data exploration-based visual analytics system based on a large language model. This framework utilizes a reasoning module based on the large language model, an interaction module that generates human intuition through interaction, and a computational module that generates computational insights. Different formats of data insights are customized for each module: visual charts of insights are presented to the interaction module, connected insight maps are constructed for the insight computation module, and textual descriptions of insights are generated for the reasoning module. Within this framework, the insight engine searches for structure-related data insights, the large language model recommends semantically relevant data insights, and users guide the visual exploration process using natural language questions and interactive choices, ultimately generating a coherent data story.

[0010] The table data exploration and visual analysis method based on the big language model includes the following five steps: (1) parsing the original table data and building a data model; (2) automatically mining insights and generating three different forms of insights; (3) user-led exploratory data analysis and generating data stories in the interaction module; (4) the insight engine in the calculation module finds insights with structural relationships through the insight map; and (5) the big language model in the reasoning module recommends insights through semantic association.

[0011] Specifically:

[0012] (1) Parse the original table data and build a data model:

[0013] Data models provide a fundamental framework for data processing and analysis, ensuring the effective location of specific subspaces within multi-dimensional tabular data for in-depth analysis and extraction of valuable data insights. Building a data model requires clearly defining data subspaces, analytical entities, insights, and their interrelationships from coarse to fine granular levels. This solution is designed to meet two core requirements: accuracy—the ability to accurately define and calculate various types of data insights across different data ranges; and flexibility—the solution is applicable to various tabular data structures and supports flexible data exploration and insight extraction.

[0014] First, tabular data can be divided into multiple data subspaces. To support diverse data analysis, filters are set for data dimensions to precisely locate target data subsets. These filters allow users to specify specific conditions for certain dimensions, thereby filtering the data to a specific range. For dimensions without specified conditions, all possible attribute values ​​are retained to ensure data integrity. Through this flexible filtering mechanism, the system can traverse all existing subspaces and supports dynamic subspace generation, thus adapting to various analytical needs.

[0015] Building upon the data subspace, analytical entities are further defined. Analytical entities represent finer-grained partitions of the data. The system groups data by selecting certain categorical attributes and performs statistical summarization operations on numerical attributes, such as calculating maximum, minimum, or sum values. By subdividing analytical entities, users can extract meaningful information from these aggregation results, ensuring that data analysis can delve into more granular levels.

[0016] Insights are extracted by identifying and classifying patterns within each analytical entity. Insights represent key findings in data analysis, such as trends, anomalies, and other pattern information. To help users better understand insights and quickly locate key findings in data exploration, the system generates a textual description for each insight and quantifies its importance.

[0017] Finally, to construct a coherent analytical process, a network of relationships between insights was established. By establishing structural and semantic connections between insights, the system can reveal deeper logic behind the data. For example, some insights may exhibit causal relationships or thematic relevance. The system connects analytical entities through structural attributes, while semantic attributes help users understand the contextual connections between insights. The establishment of this network of insight relationships helps construct a complete data narrative, allowing users to grasp the logical flow of the data holistically.

[0018] By constructing data subspaces, analyzing entities, insights, and their relationship networks, the abstract model effectively separates complex data structures from the data analysis process. Its advantage lies in allowing users to directly and flexibly explore different subspaces interactively, quickly filtering and extracting valuable insights. Furthermore, the definitions of different insights and their relationships can be integrated into the abstract model for reuse and expansion.

[0019] (2) Automated discovery of insights and generation of three different forms of insights;

[0020] The insight mining process begins by dividing the raw tabular data into multiple analytical entities. For tabular data with multiple dimensions, all possible combinations of each dimension and its categorical values ​​are enumerated. These combinations are used to filter row data, thus defining different subspaces. Within each subspace, aggregation operations are performed based on different categorical dimensions, merging rows with the same categorical attribute to generate a new analytical entity. The numerical attributes of this entity are calculated using a predefined aggregation function. Each analytical entity represents a specific perspective of the data, allowing for more granular exploration and analysis.

[0021] Next, all data insights are extracted within each analytical entity. The system iterates through each analytical entity and extracts relevant insights based on predefined calculation rules and thresholds. Different types of insights have corresponding calculation standards, thresholds, visualization methods, and description templates; only insights that pass the preset thresholds are retained. These extracted insights generate corresponding text descriptions based on the templates, forming insight charts, and are linked into insight graphs. These three insight formats will serve as inputs for subsequent system analysis and display.

[0022] (a) Input for the interactive module: Visual charts

[0023] Insights into tabular data are categorized into three main types: shape insights, point insights, and composite insights. Point insights highlight the importance of specific data points or deviations from common patterns, such as dominance, top two performers, outliers, and prominent negative values. Shape insights focus on the overall distribution and structure of the data, helping users understand the overall trends and distribution characteristics, such as trend, skewness, kurtosis, and meanness. Composite insights describe complex patterns between different subsets of the data, including temporal correlation, linear correlation, and dependency. To effectively present these insights, a suitable visual chart format is selected for each type of insight, and the Vega-Lite specification is used to implement the corresponding visualizations. Through these intuitive visual representations, users can quickly understand the meaning of data insights, thereby facilitating further data exploration and analysis. The entire interactive module helps users more efficiently grasp the information behind the data and make corresponding decisions by inputting these insight visualizations.

[0024] (b) Input to the computation module: Insight map

[0025] Insight maps reflect the structural relationships between data subspaces, which exhibit a hierarchical structure of parallelism and inclusion, helping to reveal the structural connections between different data insights. Based on the characteristics of the subspace locators, insight relationships are categorized into sibling relationships and parent-child relationships. Sibling relationships refer to situations where two insight locators have the same length and canonical dimensions, differ in only one dimension's attribute value, and have identical attribute values ​​in all other dimensions. For example, insights located in different regions but belonging to the same brand have a sibling relationship; this relationship allows for the analysis of different data features within the same dimension. Parent-child relationships represent data refinement or generalization, exhibiting hierarchy and directionality. A parent-child relationship is formed when one subspace has one more dimension than another, and the former's dimensions encompass all of the latter's dimensions. Parent-child relationships help understand how data unfolds from general to specific. Organizing data insights based on these relationships allows for the construction of insight maps. This insight map serves as input to the computation module, supporting the module in filtering and associating different data insights based on structural relationships.

[0026] (c) Input to the inference module: Insight description

[0027] To fully leverage the reasoning capabilities of large language models, descriptions are introduced as a third form of insight representation. The purpose of descriptions is to concisely and effectively convey key information from data insights, especially when dealing with massive tabular data. This helps reduce high demands on system resources and minimizes unnecessary computational burden. By generating accurate text descriptions, the limitations of the context window can be effectively mitigated, and the risk of generating incorrect or irrelevant content can be reduced.

[0028] Tabular data often contains a wealth of information, and processing it directly is both time-consuming and resource-intensive, easily leading to inaccurate or misleading results from large language models. Generating high-quality descriptions helps filter out irrelevant or low-value information, thereby reducing the computational burden on the model and lowering the probability of misleading results. These descriptions not only simplify the expression of insights but also help users quickly grasp the core patterns in the data, avoiding the need to process the entire table.

[0029] To automate the generation of descriptions, a predefined template-based approach was employed. Each insight's description is structured as a tuple containing four elements: filtering criteria for the data subspace, insight type, insight score, and descriptive content. These elements provide contextual information about the insight, its nature and importance, and a concise explanation of data patterns. The description further clarifies key information such as trends and anomalies in the data, and details the dimensions and data operations involved in the insight.

[0030] Through this automated description generation method, the system can quickly and efficiently generate concise explanations for different types of insights, providing important input for large language model reasoning, helping users intuitively understand the core content of the data, and providing decision support.

[0031] (3) User-led exploratory data analysis and data story generation in the interactive module:

[0032] In the ReInAct framework, users are the core driving force of the entire data exploration process; their analysis and decisions influence the direction of the exploration. The interactive modules within the framework use visual charts to display insights extracted from the data, helping users identify potential patterns and trends. Users can guide their data exploration by asking questions, and the system provides personalized suggestions based on these questions and current insights, helping users make informed decisions.

[0033] Specifically, users first browse the extracted initial insights, then select an insight of interest as a starting point and ask a question. This question sets the context and direction for subsequent analysis. The system then executes a two-stage filtering process using the insight engine and related algorithms to present recommended insights relevant to the user's question, providing clear reasons for the recommendations. This transparency allows users to understand why the system recommends certain insights and how they relate to the user's original question and previously selected insights.

[0034] These recommended insights and interpretations help users better decide on their next steps in exploration. For example, users can choose to delve deeper into a specific topic or adjust their questions to explore new directions. The ReInAct framework supports iterative exploration, meaning that the results of each exploration can form the basis for the next, ensuring that users can comprehensively and deeply analyze and understand the data.

[0035] (4) In the calculation module, the insight engine finds insights with structural relationships through the insight map:

[0036] The computation module takes the constructed insight map as input and uses the insight engine to perform structure filtering based on the relationships between insights. This mechanism first identifies all insights related to the selected insight. During this process, the insight engine acts as an external executor, filtering and organizing candidate insights from the source table to generate a compressed subset of data. This step aims to remove poorly relevant information, ensuring that the received structured data contains necessary content and eliminating distracting information. During structure filtering, each relationship is grouped by label for better organization and presentation of related insights. This grouping process is automated by the insight engine, ensuring that all relevant candidate insights are presented clearly to the subsequent inference module. This mechanism enables effective organization and filtering of data insights, thereby optimizing the information processing flow and improving the system's ability to identify and analyze relevant insights.

[0037] (5) In the reasoning module, the large language model performs insightful recommendations through semantic association:

[0038] The large language model receives descriptions of grouped candidate insights as input and performs two-stage reasoning based on user-provided questions, recommending the next insight to explore from a semantic perspective. Since the insight engine has already determined the candidate insight set, effectively limiting the decision space, the large language model's reasoning mode can shift from generation to decision-making. To prevent context window overload and reduce the complexity of question analysis, the decision-making process is decomposed into two stages: group selection and insight ranking, guiding the large language model to use thought chain reasoning.

[0039] During the grouping selection phase, only grouping information relevant to the current insight and user question is provided to the large language model. This grouping information is represented by identifiers and does not include specific insight information. The system selects the group most relevant to the current insight from multiple dimensions and levels. The choice of grouping type can effectively guide the direction of subsequent exploration. For example, grouping with parent-child relationships can help expand the context of the current insight or provide more general background information, thereby supporting decision-making.

[0040] During the insight ranking phase, the large language model receives textual descriptions of all insights within the selected group. The system determines the ranking of these insights by constructing semantic relationships between the current insight and each candidate insight. These semantic relationships include logicality, data relevance, question relevance, and insight score. Based on these criteria, the system recommends the top five most relevant insights and provides the underlying principles for each recommendation, explaining the reasons for its selection and how it was derived from the current insights.

[0041] Users explore data through multiple iterations, ultimately generating a data story as output. This data story encompasses all meaningful exploration paths, possessing both a sequential, step-by-step reasoning structure and a non-linear structure that allows multiple insights to support a single analytical outcome. To this end, an extended tree structure is employed to capture these characteristics, using user-selected insights as nodes and relationships as edges. This exploration process tree visually illustrates the user's discovery process.

[0042] The method described in this invention supports modeling structured tabular data, enabling data parsing and insight mining, and can serve as a foundation for exploratory data analysis recommendations. Automated mining of potential insights from tables frees users from repetitive and intensive analytical operations. By integrating human intuition and insights, the computational power of external executors, and the intelligence of large language models, it enhances the interactivity, transparency, and interpretability of the exploratory data analysis process, thereby promoting efficient and intuitive data analysis.

[0043] Specifically, the effects of the present invention include:

[0044] Simplify the cumbersome user interaction process required to build insight connections: Step (1) above supports the system to parse complex tabular data, the insight mining algorithm in step (2) supports the automatic division of the table into subspaces and the extraction of data insights from each subspace using a rule-based method, and step (3) supports the strategic integration of user interactions at key points in the data story building process, thereby improving the overall quality and relevance of the data story narrative and achieving the invention effect.

[0045] It has the ability to establish correlations between data insights to guide exploration: Step (1) above supports the detection and utilization of the inherent attributes and features of insights, and step (2) supports the system to construct an insight relationship graph using structural correlations. This graph serves as the input for step (4) to support the insight engine in performing structural filtering, removing interfering information and organizing relevant insights, effectively limiting the decision space. This supports the shift of the reasoning mode of the large language model in step (5) from generation to decision-making, which helps to further clarify semantic relationships, guide analysts to conduct more effective visual story exploration, and thus achieve the effect of invention.

[0046] It has the ability to support different stages of data story building: Steps (1) and (2) above support the system to adapt to the scenario where the user has specific analysis goals or intentions to use their data to explore professional knowledge and experience. Steps (4) and (5) support the system to recommend insights to the user through structural and semantic associations to inspire initial analysis, break through analysis bottlenecks and provide key guidance. Step (3) supports the user to flexibly change requirements, while promoting extensive exploration and targeted, goal-driven storytelling, thereby achieving the invention effect.

[0047] Enhancing the rationality of insight recommendations: The insight description generation in step (3) above supports the efficient transmission of key pattern information in the data to the large language model. Combined with the candidate insight set organized and selected in step (4), it optimizes the reasoning process in step (5), ensuring that the relevance of the insights is logical and has sufficient basis. The final result is presented transparently in the interaction module of step (3), allowing users to view the basic principles behind the suggestions, making the system interpretable, and thus achieving the effect of invention. Attached Figure Description

[0048] Figure 1 The interactive exploratory data analysis framework in a specific embodiment of the present invention is as follows:

[0049] Figure 2 This is a table data insight and mining process in a specific embodiment of the present invention;

[0050] Figure 3 These are three forms of data insight input to different modules in specific embodiments of the present invention;

[0051] Figure 4 This invention provides specific embodiments of data insights and their visualizations.

[0052] Figure 5 This invention explores the visual mapping design of path trees in specific embodiments of the invention.

[0053] Figure 6 This is a diagram of the visual analysis system interface in a specific embodiment of the present invention. Detailed Implementation

[0054] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments.

[0055] This embodiment proposes the ReInAct framework for a tabular data exploratory visual analysis system based on a large language model. Figure 1 This invention illustrates an interactive exploratory data analysis framework comprising three modules: an inference module supported by a large language model, a user-centric interaction module, and a computation module driven by an insight engine. The extracted data insights are processed differently in each module: the insight visualization is input to the interaction module for user understanding, the constructed insight relationship graph is input to the computation module, and the textual description of the insight is input to the inference module. Within this framework, users can select insights and ask questions. The insight engine then filters structurally relevant candidate insights through the insight relationship graph. Subsequently, the large language model ranks the insights and recommends potential next steps. Users can iteratively explore and construct a data story tree recording their exploration process.

[0056] Figure 2 The flowchart illustrating the table data insight mining process in a specific embodiment of the present invention is shown below: First, the system constructs a data model by parsing the original table data, and then uses this model to traversely divide the table data into all subspaces. Next, for each subspace, it further divides it into all analytical entities. Finally, insight calculations are performed within all analytical entities. This process ensures that the system can traverse all analytical objects to extract meaningful data insights, achieving fully automated extraction of data insights.

[0057] Figure 3 This invention illustrates three forms in which data insights are input into different modules in a specific embodiment: insight visualizations are input into the interaction module so users can intuitively understand data patterns; insight relationship graphs are input into the calculation module to support the computational processing of the insight engine; and textual descriptions of insights are input into the inference module to assist the large language model in performing logical reasoning. The extracted data insights are transformed into different forms, working collaboratively within their respective modules to help users iteratively explore tabular data and build coherent data visualization stories.

[0058] Figure 4 The present invention illustrates the types of data insights and their visualizations in specific embodiments: data insights are divided into three categories: point insights, shape insights, and composite insights. These three categories are further subdivided into eleven subtypes. Each subtype represents specific features and provides different data perspectives, which are represented intuitively through their respective visualizations.

[0059] Figure 5 This illustration demonstrates the visual mapping design of the exploration path tree in a specific embodiment of the present invention: In the visual mapping of nodes, color is used to encode categories, and icons within nodes represent specific insight types; in the visual mapping of edges, three relationships are mapped through different patterns and line thicknesses. The first type is a user-created custom relationship, while the other two are structural connections recommended based on a large language model. Semantic connections are displayed by hovering the mouse over the edge.

[0060] Figure 6The diagram illustrates the interface of the visual analytics system in a specific embodiment of the present invention: The user interface consists of three panels. In the data subspace selection panel on the left, users can determine a specific data subspace by filtering attribute values ​​in the filter panel, and all insights in that subspace will be updated in real time in the adjacent candidate insight panel; the main interface in the middle is the iterative exploration panel, which uses a circular force-directed layout to map nodes and edges in the exploration path tree, allowing users to interactively explore data through this interface; the right side is the insight information panel, which consists of an upper view of detailed insight information and a lower view of exploration history. The former presents detailed information about key data insights and assists users in modifying the exploration path, while the latter allows users to rewind their steps.

Claims

1. A tabular data exploration and visual analysis system based on a large language model, characterized in that, include: The system includes an inference module powered by a large language model, an interaction module centered on user engagement, and a computation module driven by an insight engine. The system presents visualizations of insights to the interaction module, builds connected insight maps for the insight computing module, and generates textual descriptions of insights for the reasoning module. The insight engine searches for structure-related data insights, the large language model recommends semantically related data insights, and users use natural language questions and interactive choices to guide the visual exploration process, ultimately generating a coherent data story. The specific method for generating insights is as follows: The insight mining process first divides the original tabular data into multiple analytical entities. For tabular data with multiple dimensions, it first enumerates all possible combinations of each dimension and its classification values. These combinations are then used to filter row data, thereby defining different subspaces. Within each subspace, aggregation operations are performed based on different classification dimensions, merging rows with the same classification attributes to generate a new analytical entity. The numerical attributes of this entity are calculated using a pre-defined aggregation function. Each analytical entity represents a specific perspective of the data, allowing for more refined exploration and analysis. Next, all data insights are extracted within each analysis entity. The system iterates through each analysis entity and extracts relevant insights based on predefined calculation rules and thresholds. Different types of insights have corresponding calculation standards, thresholds, visualization methods, and description templates. Only insights that pass the preset thresholds will be retained. These extracted insights will generate corresponding text descriptions based on the templates, form insight charts, and link them into insight maps. These three insight formats will serve as inputs for subsequent system analysis and display. (a) Input for the interactive module: Visual charts Insights into tabular data are categorized into three main types: shape insights, point insights, and composite insights. Point insights highlight the importance of specific data points or deviations from common patterns. Shape insights focus on the overall distribution and structure of the data, helping users understand the overall trends and distribution characteristics. Composite insights describe complex patterns between different subsets of the data, including temporal correlations, linear correlations, and dependencies. A visual chart format is selected for each type of insight, and the Vega-Lite specification is used to achieve the corresponding visual representation. Through these intuitive visual representations, users can quickly understand the meaning of data insights, thereby facilitating further exploration and analysis of the data. (b) Input to the calculation module: Insight map Based on the characteristics of subspace locators, insight relationships are divided into two categories: sibling relationships and parent-child relationships. Sibling relationships refer to situations where two insight locators have the same length and specified dimensions, differ in only one dimension's attribute value, and have the same attribute values ​​in all other dimensions. Parent-child relationships represent data refinement or generalization, exhibiting hierarchy and directionality. A parent-child relationship is formed when one subspace has one more dimension than another, and the former's dimensions encompass all of the latter's dimensions. Parent-child relationships help understand how data expands from general to specific. Based on these relationships, data insights are organized to construct insight maps. These insight maps serve as input to the computation module, supporting the module in filtering and associating different data insights according to structural relationships. (c) Input to the inference module: Insight description A predefined template-based approach was adopted; the description of each insight was structured into a tuple containing four elements: filtering conditions of the data subspace, insight type, insight score, and description content; these elements respectively provide contextual information about the insight, the nature and importance of the insight, and a concise explanation of the data pattern; the description section clarifies the trends and key information of anomalies in the data, and explains the dimensions and data operations involved in the insight; The specific methods for providing semantically relevant data insights using large language models are as follows: The large language model receives the descriptions of grouped candidate insights as input and performs two-stage reasoning in conjunction with the questions raised by the user, recommending the next insight to explore from a semantic perspective. Since the insight engine has already determined the set of candidate insights, the decision space is limited, and the reasoning mode of the large language model shifts from generation to decision. The decision-making process is decomposed into two stages: group selection and insight ranking, which guides the large language model to use thought chain reasoning. During the group selection phase, only grouping information relevant to the current insight and user question is provided to the large language model; this grouping information is represented by identifiers and does not include specific insight information; the system selects the group most relevant to the current insight from multiple dimensions and levels; the selection of grouping type can effectively guide the direction of subsequent exploration; During the insight ranking phase, the large language model receives textual descriptions of all insights within the selected group. The system determines the ranking of these insights by constructing semantic relationships between the current insight and each candidate insight. These semantic relationships include logicality, data relevance, question relevance, and insight score. Based on these criteria, the system recommends the top five most relevant insights and provides the basic principles behind each recommendation, explaining the reasons for its selection and how it is derived from the current insight. Users explore data through multiple iterations, ultimately generating a data story as output. This data story contains all meaningful exploration paths, possessing both a sequential, step-by-step reasoning structure and a non-linear structure that allows multiple insights to support a single analytical result. An extended tree structure is used to capture these features, with user-selected insights as nodes and relationships as edges. This exploration process tree visually demonstrates the user's discovery process.

2. A method for exploratory visual analysis of tabular data based on a large language model, characterized in that: The method using the tabular data exploration and visual analysis system based on a large language model as described in claim 1 includes the following five steps: (1) Parse the original table data and establish a data model; (2) Automated discovery of insights and generation of three different forms of insights; (3) In the interactive module, users lead exploratory data analysis and generate data stories; (4) In the calculation module, the insight engine searches for insights with structural relationships through the insight map; (5) In the reasoning module, the large language model performs insight recommendation through semantic association recommendation.

3. The method for exploratory visual analysis of tabular data based on a large language model according to claim 2, characterized in that, The specific method for step (1) is as follows: First, the tabular data is divided into multiple data subspaces. To support diverse data analysis, filters are set for data dimensions to precisely locate the target data subset. These filters allow users to specify specific conditions for certain dimensions, thereby filtering the data to a specific range. For dimensions without specified conditions, all possible attribute values ​​are retained to ensure data integrity. Based on the data subspace, analytical entities are defined; analytical entities are fine-grained divisions of data. The system groups data by selecting categorical attributes and performs statistical summary operations on numerical attributes; by subdividing analytical entities, users can obtain meaningful information from these aggregation results. Insights are extracted by identifying and classifying patterns in each analytical entity; insights represent key findings in data analysis, and the system generates a textual description for each insight and quantifies its importance. Finally, a network of relationships between insights is established; by establishing structural and semantic connections between insights, the system can reveal the deeper logic behind the data; the system analyzes the relationships between entities through structural attributes, and helps users understand the contextual connections between insights through semantic attributes.

4. The method for exploratory visual analysis of tabular data based on a large language model according to claim 2, characterized in that, The specific method for step (2) is as follows: The insight mining process first divides the original tabular data into multiple analytical entities. For tabular data with multiple dimensions, it first enumerates all possible combinations of each dimension and its classification values. These combinations are then used to filter row data, thereby defining different subspaces. Within each subspace, aggregation operations are performed based on different classification dimensions, merging rows with the same classification attributes to generate a new analytical entity. The numerical attributes of this entity are calculated using a pre-defined aggregation function. Each analytical entity represents a specific perspective of the data, allowing for more refined exploration and analysis. Next, all data insights are extracted within each analysis entity. The system iterates through each analysis entity and extracts relevant insights based on predefined calculation rules and thresholds. Different types of insights have corresponding calculation standards, thresholds, visualization methods, and description templates. Only insights that pass the preset thresholds will be retained. These extracted insights will generate corresponding text descriptions based on the templates, form insight charts, and link them into insight maps. These three insight formats will serve as inputs for subsequent system analysis and display. (a) Input for the interactive module: Visual charts Insights into tabular data are categorized into three main types: shape insights, point insights, and composite insights. Point insights highlight the importance of specific data points or deviations from common patterns. Shape insights focus on the overall distribution and structure of the data, helping users understand the overall trends and distribution characteristics. Composite insights describe complex patterns between different subsets of the data, including temporal correlations, linear correlations, and dependencies. A visual chart format is selected for each type of insight, and the Vega-Lite specification is used to achieve the corresponding visual representation. Through these intuitive visual representations, users can quickly understand the meaning of data insights, thereby facilitating further exploration and analysis of the data. (b) Input to the calculation module: Insight map Based on the characteristics of subspace locators, insight relationships are divided into two categories: sibling relationships and parent-child relationships. Sibling relationships refer to situations where two insight locators have the same length and specified dimensions, differ in only one dimension's attribute value, and have the same attribute values ​​in all other dimensions. Parent-child relationships represent data refinement or generalization, exhibiting hierarchy and directionality. A parent-child relationship is formed when one subspace has one more dimension than another, and the former's dimensions encompass all of the latter's dimensions. Parent-child relationships help understand how data expands from general to specific. Based on these relationships, data insights are organized to construct insight maps. These insight maps serve as input to the computation module, supporting the module in filtering and associating different data insights according to structural relationships. (c) Input to the inference module: Insight description A predefined template-based approach was adopted; the description of each insight was structured into a tuple containing four elements: the filtering conditions of the data subspace, the insight type, the insight score, and the description content. These elements provide contextual information about the insight, the nature and importance of the insight, and a concise explanation of the data pattern; The description section elucidates the trends and key anomalies in the data, and explains the dimensions and data manipulations involved in the insights.

5. The method for exploratory visual analysis of tabular data based on a large language model according to claim 2, characterized in that, The specific method for step (3) is as follows: Users first browse the extracted initial insights, then select an insight of interest as a starting point and ask a question; this question sets the background and direction for subsequent analysis; the system then performs a two-stage filtering process through the insight engine and related algorithms to present recommended insights related to the user's question and provide clear reasons for the recommendations; this transparency allows users to understand why the system recommends certain insights and how they relate to the user's original question and the previously selected insights; These recommendation insights and explanations help users decide on their next steps in exploration.

6. The method for exploratory visual analysis of tabular data based on a large language model according to claim 2, characterized in that, The specific method for step (4) is as follows: The computation module takes the constructed insight map as input and uses the insight engine to perform structural filtering through the relationships between insights. This mechanism first identifies all insights related to the selected insight. In this process, the insight engine acts as an external executor to filter and organize candidate insights in the source table, thereby generating a compressed subset of data. When performing structural filtering, each relationship is grouped by a label. This grouping process is completed automatically by the insight engine, ensuring that all relevant candidate insights are presented to the subsequent inference module in a clear manner.

7. The method for exploratory visual analysis of tabular data based on a large language model according to claim 2, characterized in that, The specific method for step (5) is as follows: The large language model receives the descriptions of grouped candidate insights as input and performs two-stage reasoning in conjunction with the questions raised by the user, recommending the next insight to explore from a semantic perspective. Since the insight engine has already determined the set of candidate insights, the decision space is limited, and the reasoning mode of the large language model shifts from generation to decision. The decision-making process is decomposed into two stages: group selection and insight ranking, which guides the large language model to use thought chain reasoning. During the group selection phase, only grouping information relevant to the current insight and user question is provided to the large language model; this grouping information is represented by identifiers and does not include specific insight information; the system selects the group most relevant to the current insight from multiple dimensions and levels; the selection of grouping type can effectively guide the direction of subsequent exploration; During the insight ranking phase, the large language model receives textual descriptions of all insights within the selected group; the system determines the ranking of these insights by constructing semantic relationships between the current insight and each candidate insight; these semantic relationships include logicality, data relevance, question relevance, and insight score; Based on these criteria, the system will recommend the top five most relevant insights and provide the underlying principles for each recommendation, explaining the reasons for its selection and how it was derived from the current insights. Users explore data through multiple iterations, ultimately generating data stories as output. This data story encompasses all meaningful paths of exploration, possessing both the sequential nature of step-by-step reasoning and the non-linear structural characteristic that allows multiple insights to support a single analytical outcome. Employing an extended tree structure to capture these features, with user-selected insights as nodes and relationships as edges, this exploration process tree visually illustrates the user's discovery process.

Images