Table data exploration type visual analysis system and method based on large language model

Abstract

The invention discloses a table data exploration type visual analysis system and method based on a large language model. The method comprises the steps that original table data are analyzed, and a data model is established; automatically excavating insights and generating three different forms of the insights; user dominant exploratory data analysis and data story generation are carried out in the interaction module; an insights engine in the calculation module searches insights with structural association through an insights graph; and the big language model in the reasoning module performs insight recommendation through semantic association recommendation. According to the method, the structural table data can be modeled, so that data analysis and insight mining are performed. Potential insights in a table are automatically mined, so that a user gets rid of repeated and dense analysis operations, and interactivity, transparency and interpretability of an exploratory data analysis process are enhanced by fusing human intuition and insights, computing power of an external actuator and intelligence of a large language model, thereby promoting efficient and intuitive data analysis.

Description

Technical Field

[0001] The present invention relates to the field of big data visualization and human-computer interaction, and in particular to a system and method for exploratory visual analysis of tabular data based on a large language model. Background Art

[0002] Tabular data is a fundamental data format used to represent information in 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 makes it challenging to understand and explore. Effectively exploring tabular data is critical to making informed decisions and identifying data insights such as trends and outliers. However, raw tabular data presents a raw and specific view of information, often lacking the clarity and context required for deeper understanding. To bridge this gap, it becomes critical to transform extracted data insights into narrative data stories. Narrative visualizations provide a powerful way to connect discrete data points and create coherent data stories to help users understand the key findings of the data. By crafting engaging data stories, these visualizations can enhance comprehension and make complex information more understandable, allowing non-specialist audiences to more effectively engage with and retain data insights.

[0003] Traditional methods of creating data stories require users to manually extract insights and combine multiple story elements through logic, which is an inefficient process and highly dependent on the user's experience and skills. Extensive research has been conducted to simplify this process. One 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 fully match the user's specific analytical intent. Another approach allows users to participate in the data story creation process by connecting the extracted insights to help users discover related information. However, this approach is still 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 the exploratory data analysis process to analyze and study valuable or "interesting" features in the data. In the above process, data scientists will use graphics and visualization tools to observe the data from various angles and gradually choose specific methods to explore the data according to their interests. For flat table data, the traditional exploratory data analysis process usually requires repeated and complex operations, such as screening, classification, and integration. The manual iterative calculation process is not only time-consuming, but also has a very high barrier to use due to the need for professional domain knowledge. This makes it difficult for ordinary people who have not received professional training and do not have relevant professional domain knowledge to use the exploratory analysis process. Even professional data scientists need to consider too many potential analysis methods when dealing with large amounts of data, and the efficiency of using exploratory analysis will be affected.

[0005] Therefore, it is a highly demanded and important task to separate users from the complex and time-consuming manual data exploration and data understanding process and automatically mine and analyze data fragments of valuable information (i.e., data insights). Existing research has proposed a system framework for automatically mining data insights, and has proposed standard specifications for the types of data insights and the evaluation criteria for their importance. On this basis, many works focusing on the research and development of automated exploration systems hope to design and implement recommendation systems with higher intelligence to replace users to the greatest extent to complete the calculations and reasoning in the exploratory data analysis process, and directly return valuable and easy-to-understand exploration results to users. Generally, existing systems use automated methods to mine data insights, and then organize the obtained data insights into the best exploration sequence based on certain rules or heuristic evaluation criteria, and finally directly output the corresponding visualization representation of the exploration sequence to the user. These existing automations can output more advanced visualization results with richer content, showing the great potential of data visualization narratives, but they all generally face key limitations:

[0006] (1) The output quality is low. The data insights that make up the results are not semantically and logically related, which greatly reduces the overall quality of the generated high-level data narrative.

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

[0008] In view of the defects existing in the prior art, the purpose of the present invention is to propose a visualization and human-computer interaction method to support exploratory data analysis and interactive generation of data stories for the widely existing structured tabular data. The emergence of large language models has brought revolutionary progress to the field of natural language processing. Under the appropriate prompt word framework, large language models have shown excellent application potential in understanding user intentions, reasoning about the next action, and thus completing the user's specified goals. To this end, the key idea of ​​this method is to make full use of the excellent semantic understanding, text generation and reasoning capabilities of the large language model, and use it to perform logical reasoning of exploration paths and text description generation of semantic relationships between data insights based on questions about data insights given by users in natural language, so as to naturally and logically integrate users' active exploration needs, build high-quality data insight sequences, and support users to fully participate in the construction process of data insight exploration sequences.

[0009] Specifically, the ReInAct framework is proposed for an exploratory visual analysis system for tabular data based on a large language model. It uses a reasoning module based on reasoning based on a large language model, an interaction module based on human intuition generated through interaction, and a computational module based on computational insights. Different formats of data insights are customized according to the characteristics of different modules: visual charts of insights are presented to the interaction module, insight graphs with connections are built for the insight computation module, and text descriptions of insights are generated for the reasoning module. Under this framework, the insight engine searches for structurally relevant data insights, the large language model recommends semantically relevant data insights, and users use natural language questions and interactive selections to guide the visual exploration process, ultimately generating a coherent data story.

[0010] The exploratory visual analysis method for tabular data based on a big language model includes the following five steps: (1) parsing the original tabular data and establishing a data model; (2) automatically mining insights and generating three different forms of insights; (3) users leading exploratory data analysis and generating data stories in the interactive module; (4) the insight engine in the computing module searches for insights with structural associations through insight graphs; and (5) the big language model in the reasoning module recommends insights through semantic association recommendations.

[0011] Specific:

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

[0013] The data model provides a basic framework for data processing and analysis, ensuring that specific subspaces can be effectively located in multi-dimensional tabular data, so as to conduct in-depth analysis and extract valuable data insights. Building a data model requires clarifying data subspaces, analysis entities, insights and their relationships from coarse-grained to fine-grained. This solution is designed to meet two core requirements: one is accuracy, that is, the ability to accurately define and calculate multiple types of data insights within each data range; the other is flexibility, that is, 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. In order to support diversified data analysis, filters are set for data dimensions to accurately 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. Through this flexible filtering mechanism, the system can traverse all existing subspaces and support dynamic subspace generation to adapt to various analysis needs.

[0015] Based on the data subspace, analysis entities are further defined. Analysis entities are fine-grained divisions of data. The system groups data by selecting certain classification attributes and performs statistical aggregation operations on numerical attributes, such as calculating the maximum, minimum, or sum. By subdividing analysis entities, users can obtain meaningful information from these aggregated results, thereby ensuring that data analysis can go deeper into a more subtle level.

[0016] Insight extraction is achieved by identifying and classifying patterns in 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 text description for each insight and quantifies its importance.

[0017] Finally, in order to build a coherent analysis process, a relationship graph network between insights is established. By establishing structural and semantic associations between insights, the system is able to reveal the deeper logic behind the data. For example, some insights may show causal relationships or thematic correlations. The system connects the relationships between analytical entities through structural attributes, while semantic attributes help users understand the contextual connections between insights. The establishment of this insight relationship network helps to build a complete data narrative, allowing users to grasp the logical context of the data as a whole.

[0018] By constructing data subspaces, analysis entities, insights and their relationship networks, the abstract model effectively separates the complex data structure from the data analysis process. Its advantage is that users can directly and flexibly interactively explore different subspaces, quickly filter and extract valuable insights. At the same time, the definitions of different insights and their relationships can be integrated into the abstract model for reuse and expansion.

[0019] (2) Three different forms of automated insight mining and insight generation;

[0020] The insight mining process first divides the original tabular data into multiple analytical entities. For tabular data with multiple dimensions, all possible combinations of each dimension and its classification values ​​are enumerated. Row data is filtered through these combinations to define different subspaces. In each subspace, aggregation operations are performed based on different classification dimensions, and rows with the same classification attributes are merged to generate a new analytical entity. The numerical attributes of the entity are calculated by the preset aggregation function. Each analytical entity can represent a specific perspective of the data, allowing for more sophisticated exploration and analysis of the data.

[0021] Then, all data insights are extracted from each analytical entity. The system traverses 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 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 forms of insights will serve as input for subsequent analysis and display of the system.

[0022] (a) Input of the interaction module: Visualization diagram

[0023] The insights of tabular data are divided into three categories: shape insights, point insights, and composite insights. Point insights are designed to highlight the importance of specific data points or the deviations of these data points from the general pattern, such as dominance, top two, outliers, and prominent negative values. Shape insights focus on the overall distribution and structure of the data, helping users understand the overall trend and distribution characteristics of the data, such as trend, skewness, kurtosis, and averageness. Composite insights are used to describe complex patterns between different subsets of the data, including time correlation, linear correlation, and dependency. In order to effectively present these insights, a visualization chart form is selected for each type of insight, and the Vega-Lite specification is used to achieve the corresponding visualization expression. Through these intuitive visual representations, users can quickly understand the meaning of data insights, thereby promoting further exploration and analysis of the data. The entire interactive module helps users grasp the information behind the data more efficiently and make corresponding decisions by inputting these insight visualizations.

[0024] (b) Input of the computation module: Insight graph

[0025] Insight graphs reflect the structural relationship between data subspaces. These subspaces show hierarchical parallelism and inclusion, which can help reveal the structural correlation between different data insights. According to the characteristics of subspace locators, insight relationships are divided into two categories: sibling relationships and parent-child relationships. A sibling relationship refers to the situation where two insight locators have the same length and normative dimensions, have different attribute values ​​in one and only one dimension, and have the same attribute values ​​in other dimensions. For example, insights located in different regions but belonging to the same brand have a sibling relationship, which can analyze different data features in the same dimension. Parent-child relationships represent the refinement or generalization of data and are hierarchical and directional. When one subspace has one more dimension than another, and the dimensions of the former contain all the dimensions of the latter, a parent-child relationship is formed between the two subspaces. Parent-child relationships can help understand how data unfolds from broad to detailed. Organizing data insights based on these relationships can construct an insight graph. This insight graph is used as the input of the calculation module to support the module to filter and associate different data insights based on structural relationships.

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

[0027] In order to fully utilize the reasoning power of large language models, description is introduced as the third form of insight representation. The purpose of description is to concisely and effectively convey the key information in data insights, especially when facing huge tabular data, which helps to reduce the high demand for system resources and reduce unnecessary computational burden. By generating precise text descriptions, the context window limitation can be effectively alleviated and the risk of generating erroneous or irrelevant content can be reduced.

[0028] Tabular data often contains a lot of information. Direct processing is time-consuming and resource-intensive, and can easily lead to inaccurate or misleading results from large language models. Generating high-quality descriptions can help filter out irrelevant or low-value information, thereby reducing the computational burden of the model and the probability of hallucinations. These descriptions not only simplify the expression of insights, but also help users quickly grasp the core patterns in the data and avoid processing the entire data table.

[0029] In order to achieve automatic generation of descriptions, a method based on predefined templates is adopted. The description of each insight is structured as a tuple containing four elements: filter conditions of the data subspace, insight type, insight score, and description content. These elements provide contextual information of the insight, the nature and importance of the insight, and a concise explanation of the data pattern. The description part 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) In the interactive module, users lead exploratory data analysis and generate data stories:

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

[0033] Specifically, the user first browses the initial insights extracted, then selects the insights of interest as a starting point and asks a question. This question sets the context 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 recommendation. Such transparency allows users to understand why the system recommends certain insights and how they are related to the user's original question and previously selected insights.

[0034] Through these recommended insights and explanations, users can better decide the next exploration direction. For example, users can choose to analyze a specific topic in depth, or adjust the question to further explore new directions. The ReInAct framework supports users' iterative exploration, which means that the results of each exploration can become the basis for the next exploration, ensuring that users can comprehensively and deeply analyze and understand the data.

[0035] (4) The insight engine in the calculation module searches for insights with structural associations through the insight graph:

[0036] The computation module takes the constructed insight graph as input and uses the insight engine to perform structural filtering through the connections between insights. The mechanism first identifies all insights related to the selected insight. During this process, the insight engine acts as an external executor to filter and organize candidate insights in the source table to generate a compressed data subset. This step is designed to remove information with poor relevance and ensure that the received structured data contains necessary content and excludes interfering information. When performing structural filtering, each relationship is grouped by labels to better organize and display related insights. This grouping process is automatically completed by the insight engine, ensuring that all relevant candidate insights are presented to the subsequent reasoning module in a clear manner. The above mechanism realizes the effective organization and screening of data insights, thereby optimizing the information processing flow and improving the system's ability to identify and analyze relevant insights.

[0037] (5) The large language model in the reasoning module performs insight recommendations through semantic association recommendations:

[0038] The large language model receives the description form of the grouped candidate insights as input, and performs two-stage reasoning in combination with the questions raised by the user to recommend the next insight to explore from a semantic perspective. Since the insight engine has already determined the candidate insight set, the decision space is effectively limited, and the reasoning mode of the large language model can shift from generation to decision. In order to prevent the context window overload problem and reduce the complexity of problem analysis, the decision process is decomposed into two stages: group selection and insight ranking, guiding the large language model to use thought chain reasoning.

[0039] In the group selection phase, only group information related to the current insights and user questions is provided to the large language model. This group information is represented by identifiers and does not include specific insights. The system selects the groups most relevant to the current insights from multiple dimensions and levels. The selection of group type can effectively guide the direction of subsequent exploration. For example, groups with father relationships can help expand the context of the current insights or provide more general background information to support decision-making.

[0040] During the insight ranking phase, the large language model receives the text descriptions of all insights in the selected group. The system determines the ranking of these insights by building semantic relationships between the current insight and each candidate insight. These semantic relationships include logic, data relevance, question relevance, and insight score. Based on these criteria, the system recommends the top five most relevant insights and provides the rationale for each recommendation, explaining why it was selected and how it was derived from the current insight.

[0041] The user continuously explores the data through multiple rounds of iterations, and finally generates a data story as output. This data story contains all meaningful exploration paths, and has both the sequential nature of step-by-step reasoning and the nonlinear structural characteristics that allow multiple insights to support an analysis result. To this end, an extended tree structure is used to capture these features, with the user-selected insights as nodes and the associated relationships as edges. The exploration process tree can intuitively display the user's discovery process.

[0042] The method described in the present invention supports modeling of structured table data, and the model supports data analysis and insight mining, and can serve as the basis for exploratory data analysis recommendations. Automatic mining of potential insights in tables frees users from repetitive and intensive analysis operations, and enhances the interactivity, transparency and interpretability of the exploratory data analysis process by integrating human intuition and insights, the computing power of external actuators, and the intelligence of large language models, 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 associations: The above step (1) 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 use of rule-based methods to extract data insights from each subspace, 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 narration, thereby achieving the inventive effect.

[0045] Possess the ability to establish correlations between data insights to guide exploration: The above step (1) supports the detection and utilization of the intrinsic properties and characteristics of insights, and step (2) supports the system to use structural correlations to build an insight relationship graph, which, as the input of step (4), supports the insight engine to perform structural filtering, remove interference information and organize related insights, effectively limiting the decision space, thereby supporting the reasoning mode of the large language model in step (5) from generation to decision, helping to further clarify semantic relationships and guide analysts to conduct more effective visual story exploration, thereby achieving an inventive effect.

[0046] Possess the ability to support different stages of data story building: the above steps (1) and (2) support the system to adapt to scenarios where users already have specific analysis goals or intend to use their data exploration expertise and experience; steps (4) and (5) support the system to recommend insights to users through structural and semantic associations to inspire initial analysis, break through analysis bottlenecks and provide key guidance; step (3) supports users to flexibly change their requirements, while promoting extensive exploration and targeted, goal-driven storytelling, thereby achieving an inventive effect.

[0047] Enhance the rationality of insight recommendations: The insight description generation in step (3) above supports the efficient communication of key pattern information in the data to the large language model. The candidate insight set organized and screened in step (4) is combined to jointly optimize the reasoning process of step (5) to ensure that the insight relevance is logical and well-founded. The final result is transparently presented in the interactive module of step (3), supporting users to view the basic principles behind the recommendations, making the system explainable and thus achieving the inventive effect. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] Figure 1 An interactive exploratory data analysis framework in a specific implementation of the present invention:

[0049] Figure 2 Provides insight into the mining process for table data in a specific embodiment of the present invention;

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

[0051] Figure 4 The types of data insights and their visualization diagrams in the specific implementation manner of the present invention;

[0052] Figure 5 Design of visual mapping for exploring path tree in specific implementation of the present invention;

[0053] Figure 6 This is a diagram of the visual analysis system interface in a specific implementation manner of the present invention. DETAILED DESCRIPTION

[0054] The present invention will be further described in detail below in conjunction with the accompanying drawings and specific implementation methods.

[0055] This embodiment proposes a ReInAct framework based on a tabular data exploratory visual analysis system with a large language model. Figure 1 An interactive exploratory data analysis framework in a specific embodiment of the present invention is shown, which consists of three modules: a reasoning module supported by a large language model, an interaction module centered on user participation, and a calculation module driven by an insight engine. The extracted data insights are processed differently in each module: the visualization of the insight is input into the interaction module for user understanding, the constructed insight relationship diagram is input into the calculation module, and the text description of the insight is input into the reasoning module. Within this framework, users can select insights and ask questions, and then the insight engine filters out structurally related candidate insights through the insight relationship diagram, and then the large language model sorts the insights and recommends potential next steps. Users can build a data story tree that records their exploration process through iterative exploration.

[0056] Figure 2 The table data insight mining flow chart in the specific implementation of the present invention is shown: First, the system builds a data model by parsing the original table data, and traversally divides the table data into all subspaces based on this model. Then, for each subspace, all analysis entities therein are further divided, and finally insight calculations are performed in all analysis entities. This process ensures that the system can traverse all analysis objects to extract meaningful data insights, and realizes the comprehensive and automated extraction of data insights.

[0057] Figure 3 The following diagram shows three forms of data insights input into different modules in a specific embodiment of the present invention: the visualization diagram of the insights is input into the interaction module so that users can intuitively understand the data pattern; the insight relationship diagram is input into the calculation module to support the calculation processing of the insight engine; the text description of the insights is input into the reasoning module to assist the large language model in performing logical reasoning. The extracted data insights are converted into different forms and work together in their respective modules to help users iteratively explore tabular data and build coherent data visualization stories.

[0058] Figure 4 The types of data insights and their visualization diagrams in a specific implementation of the present invention are shown: data insights are divided into three categories: point insights, shape insights, and composite insights. These three categories are further divided into eleven subtypes, each of which represents specific features and provides a different data perspective, which is intuitively visually represented through its own visualization diagram.

[0059] Figure 5 The visual mapping design of the exploration path tree in the specific implementation of the present invention is shown: in the visual mapping of nodes, colors are used to encode categories, and icons in nodes represent specific insight types; in the visual mapping of edges, three relationships are mapped by different patterns and line thicknesses, the first of which is a custom relationship manually created by the user, and 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 interface diagram of the visual analysis system in a specific implementation manner of the present invention is shown: the user interface consists of three panels. In the data subspace selection panel on the left, the user can determine a specific data subspace by filtering the attribute values ​​in the panel, and all insights in the subspace will be updated in real time in the candidate insights panel next to it; the main interface in the middle is the iterative exploration panel, which uses a circular force-directed layout diagram to map the nodes and edges in the exploration path tree, and the user explores the data interactively through this interface; the right side is the insight information panel, which consists of the insight details view above and the exploration history view below. The former presents detailed information about key data insights and assists users in modifying the exploration path, and the latter enables users to retrace their steps.

Claims

1. An exploratory visual analysis system for tabular data based on a large language model, characterized by: include: A reasoning module powered by a large language model, an interaction module centered on user engagement, and a computational module driven by an insight engine; Present visual charts of insights to the interactive module, build connected insight graphs for the insight calculation module, and generate text 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 selections to guide the visual exploration process, ultimately generating a coherent data story.

2. An exploratory visual analysis method for tabular data based on a large language model, characterized in that: The table data exploratory visual analysis system based on a large language model as described in claim 1 is adopted, and the method comprises the following five steps: (1) Analyze the original table data and establish a data model; (2) Three different forms of automated insight mining and insight generation; (3) User-led exploratory data analysis and data story generation in the interactive module; (4) The insight engine in the computation module searches for insights with structural associations through the insight graph; (5) The large language model in the reasoning module performs insight recommendations through semantic association recommendations.

3. The method for exploratory visual analysis of tabular data based on a large language model according to claim 1, characterized in that: The specific method of step (1) is: First, the tabular data is divided into multiple data subspaces. To support diversified data analysis, filters are set for data dimensions to accurately 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 aggregation operations on numerical attributes. By segmenting analytical entities, users can obtain meaningful information from these aggregated results. Insight extraction is achieved by identifying and classifying patterns in each analyzed entity; insights represent key findings in data analysis, and the system generates a text description for each insight and quantifies its importance. Finally, a relationship graph network between insights is established; by establishing structural and semantic associations between insights, the system can reveal the deeper logic behind the data; the system connects and 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 1, characterized in that: The specific method of step (2) is: The insight mining process first divides the original table data into multiple analytical entities. For table data with multiple dimensions, all possible combinations of each dimension and its classification values ​​are enumerated. Row data is filtered through these combinations to define different subspaces. In each subspace, aggregation operations are performed based on different classification dimensions, and rows with the same classification attributes are merged to generate a new analytical entity. The numerical attributes of the entity are calculated by the preset aggregation function. Each analytical entity can represent a specific perspective of the data, allowing for more sophisticated exploration and analysis of the data. Then, all data insights are extracted from each analytical entity. The system traverses 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 will be retained. These extracted insights will generate corresponding text descriptions based on the templates to form insight charts and link them into insight maps. These three insight forms will serve as input for subsequent analysis and display of the system. (a) Input of the interaction module: Visualization diagram The insights of tabular data are divided into three categories: shape insights, point insights, and composite insights. Point insights are designed to highlight the importance of specific data points or the deviations of these data points from the normal pattern. Shape insights focus on the overall distribution and structure of the data to help users understand the overall trend and distribution characteristics of the data. Composite insights are used to describe complex patterns between different subsets of the data, including time correlation, linear correlation, and dependency. For each type of insight, a visualization chart form is selected, and the Vega-Lite specification is used to implement the corresponding visualization expression. Through these intuitive visual representations, users can quickly understand the meaning of data insights, thereby promoting further exploration and analysis of data. (b) Input of the computation module: Insight graph According to the characteristics of the subspace locator, the insight relationship is divided into two categories: sibling relationship and parent-child relationship. The sibling relationship refers to the situation where the length and normative dimension of two insight locators are the same, the attribute value of one and only one dimension is different, and the attribute values ​​of other dimensions are the same. The parent-child relationship represents the refinement or generalization of data, which is hierarchical and directional. When one subspace has one more dimension than another subspace, and the dimension of the former contains all the dimensions of the latter, a parent-child relationship is formed between the two subspaces. The parent-child relationship helps understand how data is expanded from general to detailed. Based on these relationships, data insights are organized to construct an insight map. This insight map is used as the input of the calculation module to support the module to filter and associate different data insights according to structural relationships. (c) Input to the reasoning module: Insight description A method based on predefined templates was adopted; the description of each insight was structured as a tuple containing four elements: filter conditions of the data subspace, insight type, insight score, and description content; These elements provide contextual information about the insight, the nature and significance of the insight, and a concise explanation of the data pattern, respectively; The description section explains the trends and anomalies in the data and explains the dimensions and data operations involved in the insights.

5. The method for exploratory visual analysis of tabular data based on a large language model according to claim 1, characterized in that: The specific method of step (3) is: The user first browses the initial insights extracted, then selects the insights of interest as a starting point and asks a question; this question sets the context 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; such 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; Through these recommended insights and explanations, users decide the next direction to explore.

6. The method for exploratory visual analysis of tabular data based on a large language model according to claim 1, characterized in that: The specific method of step (4) is: The computation module takes the constructed insight graph as input and uses the insight engine to perform structural filtering through the connections between insights. The mechanism first identifies all insights related to the selected insight. During this process, the insight engine acts as an external executor to filter and organize candidate insights in the source table to generate a compressed data subset. When performing structural filtering, each relationship is grouped by a label. This grouping process is automatically completed by the insight engine to ensure that all relevant candidate insights are presented to the subsequent reasoning module in a clear manner.

7. The method for exploratory visual analysis of tabular data based on a large language model according to claim 1, characterized in that: The specific method of step (5) is: The large language model receives the description form of grouped candidate insights as input, and performs two-stage reasoning based on the questions raised by the user to recommend the next insight to explore from a semantic perspective. Since the insight engine has already determined the candidate insight set and limited the decision space, the reasoning mode of the large language model shifts from generation to decision-making. The decision-making process is decomposed into two stages: group selection and insight ranking, guiding the large language model to use thought chain reasoning. In the group selection stage, only group information related to the current insights and user questions is provided to the large language model; this group information is represented by identifiers and does not include specific insights; the system selects the group most relevant to the current insights from multiple dimensions and levels; the selection of group type can effectively guide the direction of subsequent exploration; In the insight ranking phase, the large language model receives the text descriptions of all insights in the selected group; the system determines the ranking of these insights by building semantic relationships between the current insight and each candidate insight; these semantic relationships include logic, data relevance, question relevance, and insight score; Based on these criteria, the system recommends the top five most relevant insights and provides a rationale for each recommendation, explaining why it was chosen and how it was derived from the current insights; Users continuously explore data through multiple rounds of iterations, ultimately generating data stories as output; This data story contains all meaningful exploration paths, which has both the sequential nature of step-by-step reasoning and the nonlinear structural characteristics that allow multiple insights to support an analysis result. An extended tree structure is used to capture these features, with user-selected insights as nodes and associated relationships as edges. The exploration process tree can intuitively display the user's discovery process.

Images