Team Sports Interactive Tactical Analysis Method Based on Large Language Model

Abstract

This invention discloses an interactive tactical analysis method for team sports based on a large language model, comprising: multimodal alignment of a visually edited tactical board and tactical description text, and extraction of tactical knowledge from the aligned tactical image-text pair; customization of scenarios based on provided visual elements and extraction of scenario knowledge from the scenarios, and fusion of scenario knowledge and tactical knowledge to form domain knowledge; construction of a prompting engineering template for tactical analysis using the large language model; input of natural language based on provided structured tags, forming input prompts based on domain knowledge and the input natural language, combined with the prompting engineering template; tactical analysis and reasoning based on the input prompts by the large language model, and visualization of the tactical analysis results. This method allows users to fully express their tactical intentions to the large language model through visualization, including complex spatiotemporal context information, and enables iterative reasoning through thought chains, outputting easily understandable tactical analysis results.

Description

Technical Field

[0001] This invention belongs to the technical field of interactive data analysis, specifically relating to an interactive tactical analysis method for team sports based on a large language model. Background Technology

[0002] Visualization technology, as a data analysis and presentation tool, has been widely used in the analysis of various sports. Through graphics, charts, and interactive interfaces, complex patterns and trends in data can be understood and interpreted more intuitively. Many specialized visualization systems have emerged to meet the data types and analytical needs of different sports. In the field of basketball, most research focuses on designing visualization systems to support game analysis.

[0003] Despite progress in various fields, combining visualization techniques with large language models (GLAMs) for basketball tactical analysis remains an unsolved challenge. While simulation tools such as BasketballGAN and BasketballFlow simulate defensive strategies using user-drawn offensive tactics, they lack detailed visualization analysis and cannot fully explain complex game scenarios.

[0004] Large language models (MLMs) are a class of deep learning-based natural language processing models capable of understanding and generating human-like language text. Trained on large-scale text data, MLMs excel in various domain tasks, including language understanding, generation, and complex data analysis. They can combine various data types to handle complex information and tasks. However, when dealing with complex tasks, natural language alone is insufficient to accurately convey user intent; therefore, efficient visualization interfaces and interactive methods are needed to help users better understand, control, and improve the output of MLMs. For example, ChartSpark proposed a visualization system that helps users obtain high-quality data charts through interactive, optimized prompts. To ensure that MLMs can complete specific tasks based on predefined knowledge, C2Idea incorporates design principles into prompts, enabling MLMs to create color schemes that adhere to these principles. To help users better understand the output of MLMs and how they work, CommonSenseVis developed an interactive visualization system to explain and explore the common-sense reasoning capabilities of natural language models. Other research views MLMs as supporting collaborative analytics and demonstrates that they can help process complex information and facilitate decision-making. However, its application in promoting sports analytics, especially basketball tactical analysis, has not been fully explored. While large language models possess strong capabilities in interpreting spatiotemporal data, enabling experts to effectively explore and analyze tactics using them remains a challenge. Furthermore, their application in tactical design and simulation is still in the exploratory stage, and existing models have certain limitations in incorporating domain knowledge and improving the quality of simulation results.

[0005] In sports, tactics typically consist of multiple complex components that constantly change throughout the game, making tactical analysis extremely complex. To assess the effectiveness of tactics in a match, existing research has incorporated domain-specific knowledge to build models that improve tactical analysis. Traditional statistical models rely on quantitative data to evaluate tactical effectiveness, but these models often struggle to fully capture the complexity and dynamic changes of tactics. While deep learning-based artificial intelligence models have shown promise in tactical analysis, modeling the complex interaction between each player's actions and the final tactical outcome remains a challenge.

[0006] In the field of tactical visualization and visual analysis, a number of studies have developed interactive visualization systems. For example, Tac-Simur designed a visual analysis system based on second-order Markov chains to calculate tactical probabilities for the simulation analysis of table tennis tactics. OBTracker proposed an interpretable off-ball movement model to evaluate individual contributions and team patterns in basketball off-ball tactics and designed an interactive visual analysis system to support the interpretation and exploration of the results. Although these visualization methods can help users better understand the model analysis results, users still need to perform complex logical reasoning to effectively align each step of the player's actions with the final tactical effect for fine-grained analysis. Summary of the Invention

[0007] In view of the above, the purpose of this invention is to provide a team sports interactive tactical analysis method based on a large language model, which allows users to fully express their tactical intentions to the large language model through visualization, including complex spatiotemporal context information, and can perform iterative reasoning through thought chains to output easily understandable tactical analysis results.

[0008] To achieve the above-mentioned objectives, an embodiment provides a team movement interactive tactical analysis method based on a large language model, comprising the following steps:

[0009] Multimodal alignment is performed on the visually edited tactical board and tactical description text, and tactical knowledge is extracted from the aligned tactical image-text pairs;

[0010] Customize the scene based on the provided visualization elements and extract scene knowledge from the scene, and integrate scene knowledge and tactical knowledge to form domain knowledge;

[0011] A prompting engineering template for building a large language model for tactical analysis includes specifying the role of the large language model, specifying the specific domain knowledge that the large language model should consider during analysis, clarifying the task of the large language model, defining the output format for the large language model, providing multi-modal input for the large language model, and requiring the large language model to reason according to the thought chain prompts;

[0012] Natural language input is performed based on the provided structured labels. Input prompts are generated based on domain knowledge and the input natural language, combined with prompt engineering templates. The large language model performs tactical analysis and reasoning based on the input prompts and visualizes the tactical analysis results.

[0013] Preferably, the visually editable tactical board includes tactical visual elements, specifically player symbols, trajectory lines, and action arrows. Visual information is extracted in terms of time and space for multimodal alignment. Specifically, a coordinate system is used to record the mapping of visual information relationships on the tactical board, and the coordinate system is used to locate the player's position on the field. Continuous coordinate points are used to represent the trajectory, and the order of actions is described according to the order edited on the tactical board to illustrate the logic of tactical execution.

[0014] Preferably, the provided visualization elements include text, charts, and tables. The scene is customized based on these visualization elements, and the scene knowledge extracted from the scene includes physical fitness information, statistical information, and behavioral information.

[0015] Among them, physical fitness information includes the height, weight, and average speed of each player in the current scenario; statistical information is used for the interactive selection of team and player matchups in the current tactics, calculating on-court / off-court statistics for each selected player matchup, specifically including five indicators: field goal percentage, three-point field goal percentage, assists, turnovers, and personal fouls; behavioral information is reflected in the scenario, allowing users to search for similar scenarios and connect similar tactical executions in real games with the current scenario.

[0016] Preferably, it further includes: the domain knowledge is used to generate embedding vectors through an embedding model and stored in a vector library for retrieval by a large language model, wherein the embedding model includes the text-embedding-3-large model.

[0017] Preferably, the tasks of the large language model include recommendation tasks, explanation tasks, and evaluation tasks. When defining the output format for the large language model, common variables and task-specific variables are defined for the three types of players. The common variables include the player's tactical role and starting position, while the task-specific variables include the event type for the recommendation task, the cause and goal of the event for the explanation task, and the event outcome and countermeasure for the evaluation task.

[0018] Preferably, when the large language model performs reasoning according to the thought chain prompts, it performs four levels of reasoning, including context-based environment-level reasoning, team-level reasoning, player-level reasoning, and action-based behavior-level reasoning.

[0019] Preferably, it also includes: during the tactical analysis and reasoning process, the tactical analysis results for each input prompt are saved as historical nodes and visualized, and used as part of the input of the large language model in subsequent interactions, so as to ensure that the large language model uses instance information to generate coherent and relevant tactical analysis results.

[0020] Preferably, the provided structured tags include descriptive tags, requirement tags, and comment tags. The descriptive text input based on the descriptive tags is used to supplement the input and is added to the input component of the prompt. The requirement text input based on the requirement tags is integrated into the task component. The comment text input based on the comment tags serves as feedback on the tactical analysis results and is input together with the current tactical analysis results for the next tactical analysis reasoning.

[0021] Compared with the prior art, the beneficial effects of the present invention include at least the following:

[0022] Customized scenarios and tactics are built based on visualized tactical boards and elements, and domain knowledge is constructed. At the same time, based on the constructed prompt engineering template and structured label input, the tactical information to be analyzed is analyzed. Multi-level and multi-task tactical analysis and reasoning are performed through a large language model, and the reasoning and analysis results are visualized to realize the exploration, recommendation and interpretation of tactical design. Attached Figure Description

[0023] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0024] Figure 1 This is the basketball tactical analysis framework provided in the embodiments that supports complex interactions with large language models;

[0025] Figure 2 This is a flowchart of the team sports interactive tactical analysis method based on a large language model provided in the embodiment;

[0026] Figure 3 This is a flowchart of the multimodal alignment provided in the embodiment, which includes two multimodal alignment templates. We use colors to describe the visually extracted information: orange represents the player character, blue represents the action, green represents the position, and red represents the trajectory;

[0027] Figure 4 This is a knowledge integration flowchart provided in the embodiment, where A represents tactical knowledge and B represents scenario-based knowledge;

[0028] Figure 5 This is a schematic diagram of the tactical analysis prompt engineering template provided in the embodiment;

[0029] Figure 6This is the interactive interface for team sports interactive tactical analysis provided in the embodiment. A is the chat view, which provides system feedback through tag selection and open-ended questions and answers, enhancing communication between the user and the system; B is the deployment view, which provides interaction during the tactical setup process, including tactical drawing, positioning analysis, and scenario retrieval; C is the simulation view, which displays tactics recommended by the intelligent coach and provides an overview and detailed explanations and evaluations; D is the history view, which records the user's tactics and provides classic tactics as a starting point for exploration. Detailed Implementation

[0030] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and do not limit the scope of protection of this invention.

[0031] This invention proposes using large language models and visualization techniques to address the problems in the background art. Large language models have demonstrated powerful capabilities in understanding complex textual descriptions and processing spatiotemporal data. This indicates that such large language models can not only understand complex tactical textual descriptions but also provide forward-looking insights through tactical reasoning, surpassing the capabilities of traditional models. In this embodiment, basketball, a representative sport, is chosen as the application scenario. Through collaboration with basketball experts, an interactive tactical analysis method for team sports based on large language models is proposed. Users can draw on a tactics board to define tactics of interest. Through visualization, users can further specify the scenario of the tactics, such as score difference or specific players and teams. The intelligent coach based on the large language model will receive these multimodal inputs and use its basketball tactical knowledge to simulate the situation on the court. The output of the intelligent coach will be transformed into a visualization, allowing users to interact step-by-step with the large language model to understand the tactical formulation process and its possible variations. The visual analysis method significantly reduces the level of knowledge required by users because the intelligent coach can assist users in complex tactical design and analysis tasks.

[0032] However, directly utilizing large language models for interactive tactical analysis also presents numerous challenges. Two main challenges were encountered during development. The first challenge is proposing a method for fine-grained tactical analysis using large language models. Given the diversity of tactical understanding among basketball experts, customized input requirements are necessary. Furthermore, to enhance the realism of the simulated environment, the results should present coherent and easily understandable decision-making steps. However, the limited domain knowledge of large language models significantly impacts the quality of simulation results, making this task challenging. The second challenge is providing effective exploration, recommendation, and explanation functions to support tactical design. Experts aim to explore the simulated scenario as comprehensively as possible, but anticipating all possible scenarios during simulation is extremely difficult. Therefore, recommendations based on the current context are needed. Additionally, since the output of large language models is limited to natural language, users struggle to quickly extract key information from large amounts of text output. Therefore, transforming textual information into a visual form and providing explanations is particularly important. Designing an interactive visualization system based on large language models to support the exploration, recommendation, and explanation of tactical design is highly challenging.

[0033] Based on this, the embodiment provides a team sports interactive tactical analysis framework based on a large language model, such as... Figure 1 As shown, it includes user multimodal input, model intelligent reasoning, and iterative exploration processes.

[0034] In multimodal user input, contextual information provides essential support for detailed tactical analysis. To help users comprehensively convey information and intent to a large language model, hand-drawn tactical board sketches are introduced to further facilitate information integration. This creates a flexible and interactive environment for basketball tactical design. Visualization enhances user input from a data perspective. Text, charts, and tables, as basic forms of visualization, can represent various types of data, making them easier for users to understand. Users can customize tactical analysis scenarios through interaction with visualization elements. Visualization provides robust support for data by integrating broader contextual information into tactical settings.

[0035] In intelligent reasoning using models, demand-driven outputs are crucial in the tactical exploration process. This enables large language models to reason progressively to understand the complex factors provided by the user. Furthermore, task categories for the large language model are summarized based on various demands. Reasoning requires the large language model to perform a top-down analysis at four levels: first, the context level, focusing on the game background and specific scenarios; second, the team level, considering the team's overall strategy and style; next, the player level, evaluating the technical abilities and performance of individual players; and finally, the action level, focusing on the details of tactical actions and their execution. This chain-thinking, progressive reasoning allows the large language model to consider multiple factors and perspectives of complex tactics. Task decomposition helps the large language model break down complex tasks into more focused components. Tasks are decomposed into three categories: recommendation tasks, explanation tasks, and evaluation tasks. Recommendation tasks enable the model to provide various options and variations, offering tailored suggestions for tactical setups. Reasoning interpretation and explanation tasks provide insights into the large language model's decision-making process, significantly enhancing the transparency and comprehensibility of tactics. Evaluation tasks involve a comprehensive analysis of simulation results to help users assess the potential and applicability of each setup.

[0036] In an iterative exploration process, this invention's framework supports complex iterative interactions in tactical exploration. User input serves as tactical settings, guiding the model towards customized scenarios. Model output provides insights and alternatives for tactical simulation, effectively transforming the analysis process into a cyclical process between settings and simulations, representing the inferential evolution of tactics. Through this structured and flexible approach, users collaborate with a large language model to explore, simulate, and interpret tactics, driving understanding and optimization of tactics.

[0037] Based on the proposed interactive framework, an interactive visual analysis method and system, Smartboard, based on a large language model, was also developed. This method and system can support users to personalize tactical analysis scenarios and iteratively explore during tactical deployment, tactical simulation, and tactical evolution.

[0038] like Figure 2 As shown in the embodiment, the team movement interactive tactical analysis method based on a large language model provided includes the following steps:

[0039] S1 performs multimodal alignment of the visually edited tactical board and tactical description text, and extracts tactical knowledge from the aligned tactical image-text pair.

[0040] Multimodal alignment can effectively enhance the visual understanding capabilities of large language models by forming image-text pairs. To align tactical images and scene text from different modalities, tactical visual elements are extracted from visually edited tactical board sketches and combined with semantic descriptions to achieve effective multimodal alignment, such as... Figure 3As shown. The tactical board contains tactical visual elements, including player symbols, trajectory lines, and action arrows. To extract visual information from the tactical board in detail, it extracts visual information through time and space for multimodal alignment. A coordinate system is used to record the mapping of visual information relationships on the tactical board. Specifically, the coordinate system is used to locate the player's position on the field, and continuous coordinate points are used to represent the trajectory to better express spatial relationships. At the same time, the order of actions is described according to the order of the tactical board sketches, that is, the order in which the tactical board is edited, to convey the logic of tactical execution.

[0041] In specific multimodal alignment, two sets of templates are used, one corresponding to player movement trajectories and the other to organize the extracted tactical visual elements for semantic integration with the tactical description text. To enhance the overall comprehensibility of the description, the templates include player interactions, integrating semantic information to complete the alignment.

[0042] S2 customizes scenarios based on provided visualization elements and extracts scenario knowledge from the scenarios, integrating scenario knowledge and tactical knowledge to form domain knowledge.

[0043] Domain knowledge is a key factor influencing the performance of large language models. While large language models excel in some specific domain tasks, they may lack sufficient expertise in tactical analysis of specific scenarios. To address this limitation, tactical and scenario knowledge is transformed into knowledge documents, and retrieval-enhanced generation (RAG) methods are utilized to improve the responsiveness of large language models by retrieving relevant information, such as... Figure 4 As shown.

[0044] The constructed domain knowledge document consists of tactical-based knowledge (i.e., tactical knowledge) and scenario-based knowledge (i.e., scenario knowledge), organized in the form of text data. This provides comprehensive information support for the understanding and response generation of large language models in tactical scenarios.

[0045] like Figure 4 The tactical-based knowledge shown in Figure A originates from multimodal aligned tactical graphic-text pairs, including visual descriptions and coordinate systems. Visual descriptions include the visual encoding of the tactical board, such as symbols, lines, and arrows. The coordinate system includes the mapping of data relationships on the tactical board. This knowledge helps large language models understand tactical details.

[0046] like Figure 4The scenario-based knowledge shown in section B aims to allow users to communicate customized scenarios to the large language model. This type of knowledge covers various aspects of tactical scenarios, including physical, statistical, and behavioral information, to provide a comprehensive understanding of the tactical scenario. Details are as follows: Physical information includes the height and weight data of each player in the current scenario. Additionally, the average speed of each player is calculated based on the selected scenario and used as physical information to assess the player's physical performance in the tactic. Statistical information is used for the interaction selection of team and player matchups in the current tactic. Specifically, on-court / off-court statistics are calculated for each selected player matchup. On-court / off-court information measures the difference in performance between offensive and defensive players on or off the court, helping users understand the impact of matchups on tactics. Five metrics are also calculated and normalized, including field goals made, three-pointers made, assists, turnovers, and personal fouls. Behavioral information is reflected in the scenario, allowing users to retrieve similar situations, enabling them to associate similar tactical executions from real games with the current situation. Behavioral information helps users customize tactical scenarios, allowing the large language model to understand and generate similar situations.

[0047] After constructing the domain knowledge, augmentation generation is performed. Specifically, embedding models such as the text-embedding-3-large model can be used to embed the domain knowledge documents. The generated embedding vectors are stored in the FAISS vector library, allowing large language models to efficiently retrieve these vectors. This method enhances the ability of large language models to access and integrate relevant external knowledge, thereby improving the quality of their responses.

[0048] S3 is a prompting engineering template for building a large language model for tactical analysis. It includes specifying the role of the large language model, specifying the specific domain knowledge that the large language model should consider during analysis, clarifying the task of the large language model, defining the output format for the large language model, providing multi-modal input for the large language model, and requiring the large language model to reason according to the thought chain prompts.

[0049] Hint engineering has been proven to effectively improve the performance of large language models in solving domain-specific tasks. Therefore, techniques such as... Figure 5 The dynamic suggestion template shown here uses a large language model to generate content based on user needs. The template consists of six parts:

[0050] (1) Specify the role of the large language model. The role involves the context description related to the large language model. Specifically, the role played by the large language model is clearly defined, and the capabilities required for the role are specified.

[0051] (2) Specify the specific domain knowledge that the large language model should consider during analysis. This knowledge enhances the response of the large language model by allowing it to retrieve specific domain knowledge from provided external knowledge documents.

[0052] (3) Define the tasks of the large language model, including the requirements of each task. For example, the recommendation task requires recommending tactical settings, the interpretation task delves into the logic behind tactical decisions, and the evaluation task focuses on analyzing the capabilities required to execute tactics and the possible outcomes.

[0053] (4) Define the output format for the large language model. The output format describes the specific output for each task, specifically defining a set of common variables and task-specific variables for all three types of tasks. For example, the recommendation task requires additional output on action type. Both the explanation and evaluation tasks require an overview and detailed output. The overview considers the overall tactics, while the detailed output focuses on individual players. For explanation, the large language model describes each player's actions in detail, including how these actions counter the tactics and their objectives. The evaluation details explore whether the player's abilities align with the objectives of the current action and assess potential opponent reactions.

[0054] (5) Provide multimodal inputs for the large language model, as inputs are the triggering factors for communication with the large language model. Basic inputs require knowledge and text-image pairs. Considering the relevance of tasks, the interpretation task requires external inputs that recommend outputs, while the evaluation task requires inputs that recommend and interpret.

[0055] (6) The large language model is required to reason according to the chain of thought prompts. Reasoning is used to guide the logical reasoning process. Chain of thought prompts allow the large language model to think about and solve problems comprehensively. Specifically, the CoT method is used to guide the large language model to reason along a specific logical path, and the model is required to reason at four levels, including the context-based environment level, team level, player level, and action level.

[0056] S4 takes natural language input based on the provided structured labels, and forms input prompts based on domain knowledge and the input natural language, combined with prompt engineering templates. The large language model performs tactical analysis and reasoning based on the input prompts and visualizes the tactical analysis results.

[0057] To provide flexible control over the generation of large language models, structured tags are provided to embed users' natural language input. These tags are divided into three categories based on different objectives: descriptive tags, requirement tags, and comment tags.

[0058] For descriptive tags, the user's description is mainly used to supplement the input. Adding it to the input component of the prompt enriches the large language model's understanding of the current context.

[0059] For demand-related tags, user needs are integrated into the task components. This allows users to add semantic requirements to their desired responses, ensuring that the output of the large language model more closely matches the user's specific requirements.

[0060] For comment-related tags, user comments are used as feedback for tactical analysis of the response. Comments are added along with the current tactical analysis results to guide the large language model in refining its response after integrating these comments. Taking into account the previous context, the large language model refers to these comments when completing the task.

[0061] Input prompts are generated based on domain knowledge and natural language input using structured tags, combined with prompt engineering templates. The large language model performs tactical analysis and reasoning based on these prompts and visualizes the results. During the tactical analysis and reasoning process, the tactical analysis results for each input prompt are saved as historical nodes and visualized, and then used as part of the large language model's input in subsequent interactions. This ensures that the large language model uses instance information to generate coherent and relevant tactical analysis results.

[0062] The above-mentioned team movement interactive tactical analysis method based on large language models uses, for example, Figure 6 The system Smartboard shown comprises four views: A is the chat view, where users (A1-A3) input text to control the iterative exploration process with the large language model; B is the deployment view, which helps users define initial tactics of interest; B1 allows users to sketch deployment details, and B2 and B3 allow users to select matchup information and view similar historical match scenarios; C is the simulation view, displaying simulated alternatives returned by the large language model. C1 represents the visual content of each alternative, and C2 represents specific explanations and evaluation information. Users can select an alternative to iterate through tactical simulations; D is the history view, recording the simulation path D1 and the user's interaction history D2.

[0063] The embodiments verify the technical effectiveness of the invention based on case studies and user experiments using real basketball datasets, and demonstrate the applicability of the invention in analyzing team tactics.

[0064] Two experts with years of coaching experience were specifically invited to conduct case studies using the system developed in this invention. These studies were based on data from the 2015-16 NBA season. After an introduction to the interactive and visualization features of Smartboard, the experts freely explored tactics and scenarios of interest for an hour, recording their actions and comments. Feedback was collected through interviews to further improve the system. The first expert conducted an in-depth analysis of the "Horns" tactic, exploring how to create scoring opportunities using screens and cuts through tactical deployment and simulation. The expert found that creating open shots through team movement was particularly important for the Detroit Pistons and adjusted the tactics according to different defensive strategies. The expert also explored the "Flex" tactic, finding that it optimized scoring opportunities for the Detroit Pistons players and validating its effectiveness through actual game data. The second expert studied the "Delay" tactic, simulating defensive strategies against the powerful offensive team, the Houston Rockets. The expert found that special defensive adjustments were needed against Howard's high pick-and-roll and proposed principles for the delay tactic to effectively address mismatches and misplacements in defense.

[0065] Experts gave Smartboard a positive review of its overall performance, believing it to be very effective in basketball tactical analysis, providing in-depth tactical analysis and further exploration of complex tactical variations.

[0066] User research was also conducted, inviting eight experts in the field of basketball tactics to evaluate the tactical analysis content in Smartboard. Six metrics were established, including accuracy, effectiveness, diversity, usability, insight, and consistency, and ratings from all participants were collected. The rating analysis results showed that all participants agreed that the tactical analysis content generated by Smartboard performed excellently on these metrics, confirming the invention's capabilities in these areas.

[0067] The specific embodiments described above illustrate the technical solution and beneficial effects of the present invention in detail. It should be understood that the above description is only the most preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, additions, and equivalent substitutions made within the scope of the principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A team sports interactive tactical analysis method based on a large language model, characterized in that, Includes the following steps: Multimodal alignment is performed on the visually edited tactical board and tactical description text, and tactical knowledge is extracted from the aligned tactical image-text pair. The visually edited tactical board contains tactical visual elements, including player symbols, trajectory lines, and action arrows. Visual information is extracted from time and space aspects for multimodal alignment. A coordinate system is used to record the mapping of visual information relationships on the tactical board. Specifically, the coordinate system is used to locate the player's position on the field and uses continuous coordinate points to represent the trajectory. At the same time, the order of actions is described according to the order of editing on the tactical board to illustrate the logic of tactical execution. Customize the scene based on the provided visualization elements and extract scene knowledge from the scene. Integrate scene knowledge and tactical knowledge to form domain knowledge. The scene knowledge extracted from the scene includes physical fitness information, statistical information and behavioral information. A prompting engineering template for building a large language model for tactical analysis includes specifying the role of the large language model, specifying the specific domain knowledge that the large language model should consider during analysis, clarifying the task of the large language model, defining the output format for the large language model, providing multi-modal input for the large language model, and requiring the large language model to reason according to the thought chain prompts; The tasks of the large language model include recommendation, explanation, and evaluation. When defining the output format for the large language model, common variables and task-specific variables are defined for the three types of tasks. The common variables include the player's tactical role and starting position, while the task-specific variables include the event type for the recommendation task, the cause and goal of the event for the explanation task, and the event outcome and countermeasure for the evaluation task. When the large language model reasones according to the thought chain prompts, it performs four levels of reasoning, including context-based environment-level reasoning, team-level reasoning, player-level reasoning, and action-based behavior-level reasoning. The system takes natural language input based on the provided structured tags, and combines domain knowledge and the input natural language with the prompting engineering template to form input prompts. The large language model performs tactical analysis and reasoning based on the input prompts and visualizes the tactical analysis results. The provided structured tags include descriptive tags, requirement tags, and comment tags.

2. The team movement interactive tactical analysis method based on a large language model according to claim 1, characterized in that, The provided visualization elements include text, charts, and tables, and scenarios can be customized based on these visualization elements; Among them, physical fitness information includes the height, weight, and average speed of each player in the current scenario; statistical information is used for the interactive selection of team and player matchups in the current tactics, calculating on-court / off-court statistics for each selected player matchup, specifically including five indicators: field goal percentage, three-point field goal percentage, assists, turnovers, and personal fouls; behavioral information is reflected in the scenario, allowing users to search for similar scenarios and connect similar tactical executions in real games with the current scenario.

3. The interactive tactical analysis method for team sports based on a large language model according to claim 1, characterized in that, Also includes: The domain knowledge is used to generate embedding vectors through an embedding model and stored in a vector library for retrieval by a large language model. The embedding model includes the text-embedding-3-large model.

4. The interactive tactical analysis method for team sports based on a large language model according to claim 1, characterized in that, Also includes: During the tactical analysis and reasoning process, the tactical analysis results for each input prompt are saved as historical nodes and visualized, and used as part of the input of the large language model in subsequent interactions, ensuring that the large language model uses instance information to generate coherent and relevant tactical analysis results.

5. The interactive tactical analysis method for team movement based on a large language model according to claim 1, characterized in that, The descriptive text input based on the descriptive tag is used to supplement the input and is added to the prompt input component. The requirement text input based on the requirement tag is integrated into the task component. The comment text input based on the comment tag serves as feedback on the tactical analysis results and is input together with the current tactical analysis results for the next tactical analysis reasoning.

Images