Speed skating knowledge question-answering method based on ability map

Abstract

The invention discloses a speed skating knowledge question and answer method based on a capability atlas, which comprises the following steps: selecting literatures with speed skating related knowledge as original corpora, training a Qwen3-14B model, and outputting the trained Qwen3-14B model; analyzing the speed skating textbook, marking three types of entities from continuous fragments by adopting a GLiNER model, and mapping capability nodes to obtain a basic capability map; constructing a student graph neural network GNN on the ability sub-graph, training the student graph neural network GNN by using multi-view ability vector representation, and outputting student ability vector representation; calculating the cosine similarity between the capability nodes, generating a similarity edge, and obtaining a fusion capability map; performing text analysis and vector coding on a question proposed by a user to form an initial hit node; and outputting the initial hit node into the trained Qwen3-14B model, and outputting a natural language answer to the user.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of knowledge Q&A, in particular to a speed skating knowledge Q&A method based on a capability graph. BACKGROUND

[0002] In the prior art, intelligent Q&A for sports knowledge is mainly based on the following technologies:

[0003] 1. General large model fine-tuning method (such as LoRA, SFT), which has the defect of lacking speed skating field data and being insufficient in professionalism, and cannot generate fine-grained training suggestions.

[0004] 2. Traditional knowledge graph (Knowledge Graph, KG). It is based on entity-relation triples, but cannot represent the hierarchical structure of training capabilities, dependency relationships and capability evolution methods.

[0005] 3. RAG (Retrieval-Augmented Generation) technology. Such as GraphRAG, which assists large model reasoning through graph structure, but still relies on "entity type" graph and is not suitable for "capability abstraction".

[0006] 4. Sports training data extraction method. Some existing technologies have disclosed using NLP for entity recognition and relationship extraction, but have not covered the core content of "capability abstraction".

[0007] The closest solution to the present application is the GraphRAG-based retrieval enhancement method. This solution usually relies on traditional entity-type knowledge graphs and is difficult to cover the hierarchical structure of training capabilities and the corresponding relationship between capabilities and training methods and technical actions.

[0008] General language models lack professional terminology and training logic support, making it difficult to provide reliable technical action explanations and training suggestions; traditional knowledge graphs are mainly based on entity relationships and cannot express the hierarchical structure of training capabilities and the mapping relationship between capabilities and training methods; speed skating textbooks have complex structures, and the cost of manually extracting capability structures is high, lacking automated processing means; existing Q&A systems fail to associate training capability structures with language model answers, resulting in answers lacking traceable basis. SUMMARY

[0009] To overcome the above deficiencies in the prior art, the present application provides a speed skating knowledge Q&A method based on a capability graph, which realizes a knowledge Q&A method combining speed skating knowledge capability structure and Q&A model.

[0010] In order to achieve the above-mentioned application purposes, the technical solutions adopted by the present application are as follows:

[0011] A speed skating knowledge question and answer method based on a capability graph is provided, which comprises the following steps:

[0012] S1: Selecting literature with speed skating related knowledge as original corpus, extracting question and answer samples, training the Qwen3-14B model, making the Qwen3-14B model adapt to the terminology, training logic and answer style of the speed skating field, and outputting the trained Qwen3-14B model;

[0013] S2: Analyzing speed skating teaching materials, dividing the text into continuous segments, using the GLiNER model to label three types of entities from the continuous segments, and outputting structured triples; and constructing a speed skating capability ontology tree, mapping the continuous segments to capability nodes using the cosine similarity between the content of the continuous segments and the ontology node description vectors, and obtaining a basic capability graph;

[0014] S3: Based on the basic capability graph, a three-type teacher model is constructed to output multi-perspective capability vector representation, and a student graph neural network GNN is constructed on the capability subgraph, the student graph neural network GNN is trained using the multi-perspective capability vector representation, and the student capability vector representation is outputted;

[0015] S4: The student capability vector representation and the multi-perspective capability vector representation are written back to each capability node to form a multi-perspective capability description field; the cosine similarity between the capability nodes is calculated to generate a similarity edge, and a fusion capability graph is obtained;

[0016] S5: The fusion capability graph is loaded into a graph database or a graph computing engine, an integrated index is built, a question raised by a user is text analyzed and vector encoded to form an initial hit node; the initial hit node is outputted to the trained Qwen3-14B model to output a natural language answer to the user.

[0017] Further, step S1 comprises:

[0018] S11: Selecting literature with speed skating related knowledge as original corpus, segmenting, extracting and organizing the original corpus based on the Easydataset tool, constructing question and answer samples in the form of [question-answer], and forming a question and answer dataset;

[0019] S12: Removing question and answer samples with unclear, repetitive or contradictory semantics in the question and answer dataset through manual or rule review to ensure the quality of the question and answer dataset;

[0020] S13: Converting the question and answer dataset into a standard format supported by the llama factory framework, configuring training parameters, loading the Qwen3-14B model using the standard format question and answer dataset for training, and making the Qwen3-14B model adapt to the terminology, training logic and answer style of the speed skating field.

[0021] S14: Output the trained Qwen3-14B model and perform spot checks and verification to ensure that the trained Qwen3-14B model is stable in terms of output format and professionalism in the field of speed skating.

[0022] S15: Use the trained Qwen3-14B model as the generator for capability graph-assisted question answering.

[0023] Further, step S2 includes:

[0024] S21: Use PyMuPDF to parse the PDF file of the speed skating teaching material, extract the main text and structural information, and divide the text into continuous segments according to a predetermined length, which will serve as input units for subsequent entity recognition and relation extraction.

[0025] S22: Use the GLiNER model to perform named entity recognition on continuous segments and label three types of entities: capability entities, technical or theoretical knowledge entities, and training method entities;

[0026] S23: Label the three types of entities with category tags and text positions respectively, and store them uniformly;

[0027] S24: Based on entity recognition, the Qwen3-32B model, which is based on the Transformers framework, automatically allocates devices for model loading, and runs in a multi-card environment, is called. The relationship between the three types of entities is extracted according to the preset Chinese Prompt template and output as a triple in strict JSON format. The triple fields include: head entity, head entity type; relationship; tail entity, tail entity type.

[0028] S25: Summarize and clean the fields of the triples to obtain structured triples representing [capability-speed skating knowledge-training methods], providing business relationships for the basic capability graph;

[0029] S26: Construct a speed skating capability ontology tree based on speed skating training practice, and classify capabilities according to speed skating knowledge categories. The speed skating capability ontology tree serves as the structural constraint and unified namespace for the capability layer.

[0030] S27: Calculate the cosine similarity between the content of consecutive segments and the description vector of the ontology node in the speed skating capability ontology tree, and set a threshold for the cosine similarity. Map consecutive segments with a cosine similarity exceeding the threshold to a unified capability node to achieve semantic alignment between capability and capability node.

[0031] S28: Using NetworkXMultiDiGraph as the graph structure container, import capability nodes from the speed skating capability ontology tree to form the capability layer; import technical or theoretical knowledge entities into the knowledge layer and training method entities into the training layer.

[0032] Construct directed multi-edges based on structured triples, record relation types and confidence levels, and add the ontology hierarchy relations of the speed skating capability ontology tree to the graph structure container in the form of is_parent_of edges;

[0033] Generate a basic capability map that includes node types, business relationships, and ontology hierarchy, and export it in a specified format.

[0034] Further, step S3 includes:

[0035] S31: Based on the basic ability map, simulated training logs and competition / assessment data are introduced to form a multi-perspective ability vector representation;

[0036] Three types of teacher models were constructed, including the KGTteacher model, the TemporalTeacher model, and the EvalTeacher model;

[0037] The KGTeacher model is a capability representation based on a fundamental capability graph:

[0038] On the capability subgraph of the basic capability graph, the graph relationship classification or link prediction task is constructed using the GATv2Conv graph convolutional layer of PyTorchGeometric. Through training, the capability vector representation KGTeacherembedding of each capability node in the graph structure container is obtained, which reflects the structural position of the capability corresponding to the capability node in the basic capability graph and the connection pattern with other capability nodes.

[0039] The TemporalTeacher model is a capability representation based on the training process:

[0040] TransformerEncoder is used to model the time-series data of the simulated training logs. The main training capability classification task is set to enable the model to learn the correspondence between different training time sequences and target capabilities. After training, capability vector representations are extracted from the classification head or related layer weights to serve as the capability vector representation TemporalTeacherembedding from the perspective of the training process.

[0041] The EvalTeacher model is a capability representation based on performance results:

[0042] For competition / assessment data, a regression model based on MLP is constructed to fit the functional relationship between input features and ability scores; the transpose of the output layer weights or the intermediate layer weights of the regression model are used as the ability vector representation EvalTeacherembedding from the perspective of performance results.

[0043] S32: Construct a student graph neural network (GNN) on the capability subgraph. The student graph neural network (GNN) includes two layers of GATv2Conv graph convolutional layers and a projection head. The GATv2Conv graph convolutional layers are used to extract the local structural features of the capability nodes, and the projection head is used to map the local structural features to a capability representation space of a unified dimension.

[0044] S33: Train the student graph neural network (GNN) using the ability vector representations output by the three types of teacher models to obtain the student ability vector representation that integrates information on [textbook structure + training process + competition performance] in speed skating.

[0045] Furthermore, the loss function for training the student graph neural network (GNN) is:

[0046] ;

[0047] in, For the total loss, These are multi-teacher distillation loss and contrastive learning loss, respectively. To compare the weighting coefficients of the learning loss, These are the weighting coefficients for the three types of teacher models. Let $\mathbf{ ... These represent two different augmented views of the same capability node, where i and j are the numbers of the two different augmented views. This is the temperature coefficient.

[0048] Further, step S4 includes:

[0049] S41: The student ability vector representation and the ability vector representations KGTeacherembedding, TemporalTeacherembedding, and EvalTeacherembedding output by the three types of teacher models are used as ability node attributes and written back to each ability node to form a multi-perspective ability description field.

[0050] S42: Calculate the cosine similarity between ability nodes using student ability vectors. Based on the cosine similarity, select the K ability nodes with the highest similarity to each ability node. Establish a new edge of type ability_similarity between each ability node and the K ability nodes, generate similarity edges, and attach the cosine similarity weight corresponding to each similarity edge. The cosine similarity weight is used to explicitly represent the implicit ability relationship formed after the fusion of multi-source information.

[0051] S43: Store capability node attributes, business relationships, ontology hierarchical relationships, and similarity edges in a unified manner, export them in a set format, and obtain a fused capability graph. The fused capability graph can be directly used as the data foundation for downstream retrieval and question answering modules.

[0052] Further, step S5 includes:

[0053] S51: Load the exported fusion capability graph into a graph database or graph computing engine, establish a comprehensive index of node type, relationship type and similarity edge, and support multi-hop retrieval by capability node, technology or theoretical knowledge node, training method node and similarity edge;

[0054] S52: Perform text parsing and vector encoding on the questions raised by users. First, perform similarity retrieval between the text corpus and the description of capability nodes to locate the set of capability nodes most relevant to the question and their associated technical or theoretical knowledge nodes and training method nodes, forming an initial hit set.

[0055] S53: Based on the initial hit nodes in the initial hit set, perform the following operations according to the GraphRAG approach:

[0056] Extend the search along the ontology hierarchy, business relationships, and similarity edges to extract local subgraphs related to the problem; summarize the capability links related to the problem from the local subgraphs.

[0057] S54: Organize the key node descriptions, relationship descriptions, and capability links in the local subgraphs into a structured text context to form a graph summary; input the user's question and the graph summary into the trained Qwen3-14B model, and the structured text context provides the capability nodes, training methods, and technical actions corresponding to the question;

[0058] S55: The Qwen3-14B model summarizes and expresses information based on the structured text context, and outputs user-oriented natural language answers.

[0059] The beneficial effects of this invention are as follows: This invention automatically extracts teaching material content, constructs a three-dimensional ability map, trains a multi-perspective ability representation model, and uses the three-dimensional ability map of "ability-speed skating knowledge-training method" to assist in domain question answering, and proposes a multi-teacher distillation ability representation method to improve the accuracy and interpretability of question answering in speed skating training scenarios.

[0060] This invention addresses the problem of insufficient domain knowledge in general models by adapting the model to the domain through a constructed question-and-answer dataset, enabling it to output answers that conform to the training logic of speed skating. It also addresses the issue that traditional knowledge graphs cannot represent the structure of training capabilities by clarifying the correspondence between capabilities, technical actions, and training methods through a three-dimensional capability graph. Furthermore, it addresses the problems of complex textbook content structures and high manual costs by automatically extracting capability structures through entity recognition and relation extraction steps. Finally, it addresses the lack of linkage between the knowledge graph and the question-and-answer model by constructing a structured context through graph retrieval and subgraph extraction, giving answers a traceable capability link. Attached Figure Description

[0061] Figure 1 This is a flowchart of a speed skating knowledge-based question-and-answer method based on capability graphs. Detailed Implementation

[0062] The specific embodiments of the present invention are described below to enable those skilled in the art to understand the present invention. However, it should be understood that the present invention is not limited to the scope of the specific embodiments. For those skilled in the art, various changes are obvious as long as they are within the spirit and scope of the present invention as defined and determined by the appended claims. All inventions utilizing the concept of the present invention are protected.

[0063] like Figure 1 As shown, a speed skating knowledge question-answering method based on ability graphs includes the following steps:

[0064] S1: Select literature with knowledge related to speed skating as the original corpus, extract question and answer samples, train the Qwen3-14B model, adapt the Qwen3-14B model to the terminology, training logic and answer style of the speed skating field, and output the trained Qwen3-14B model.

[0065] Step S1 specifically includes:

[0066] S11: Select literature with knowledge related to speed skating as the original corpus, and use the Easydataset tool to segment, extract and organize the original corpus, construct question-and-answer samples in the form of [question-answer], and form a question-and-answer dataset;

[0067] The literature related to speed skating in this embodiment includes Chinese documents such as speed skating textbooks, training guides, technical specifications, and research papers. The question-and-answer samples cover the principles and key points of technical movements; training methods and applicable scenarios; ability components and influencing factors; training plans, error correction, and precautions.

[0068] S12: Remove semantically unclear, duplicate, or contradictory question and answer samples from the question and answer dataset through manual or rule-based review to ensure the quality of the question and answer dataset;

[0069] S13: Convert the question-and-answer dataset to a standard format supported by the llamafactory framework, configure the training parameters, load the Qwen3-14B model using the standard format question-and-answer dataset, and train the Qwen3-14B model to adapt to the terminology, training logic, and answering style of the speed skating field.

[0070] The training parameters in this embodiment include batch size, learning rate, and number of training rounds.

[0071] S14: Output the trained Qwen3-14B model and perform spot checks and verification to ensure that the trained Qwen3-14B model is stable in terms of output format and professionalism in the field of speed skating.

[0072] S15: Use the trained Qwen3-14B model as the generator for capability graph-assisted question answering.

[0073] S2: Analyze the speed skating textbook, segment the text into continuous segments, use the GLiNER model to label three types of entities from the continuous segments, and output structured triples; and construct a speed skating capability ontology tree, using the cosine similarity between the content of the continuous segments and the description vectors of the ontology nodes to map the continuous segments into capability nodes, thus obtaining a basic capability map.

[0074] Step S2 specifically includes:

[0075] S21: Use PyMuPDF to parse the PDF file of the speed skating teaching material, extract the main text and structural information, and divide the text into continuous segments according to a predetermined length, which will serve as input units for subsequent entity recognition and relation extraction.

[0076] S22: Use the GLiNER model to perform named entity recognition on continuous segments and label three types of entities: capability entities, technical or theoretical knowledge entities, and training method entities;

[0077] The three types of entities in this embodiment specifically include:

[0078] Ability-related entities: such as "explosive power", "cornering control ability", "balance ability", etc.

[0079] Technical or theoretical knowledge entities, such as "outer edge gliding," "cornering technique," and "push-off angle."

[0080] Training methods (training): such as "interval training", "specific strength training", "rhythm control exercises", etc.

[0081] S23: Label the three types of entities with category tags and text positions respectively, and store them uniformly;

[0082] S24: Based on entity recognition, the Qwen3-32B model, which is based on the Transformers framework, automatically allocates devices for model loading, and runs in a multi-card environment, is called. The relationships between the three types of entities are extracted according to the preset Chinese Prompt template, and the output is a triple in strict JSON format. The triple fields include: head entity (head), head entity type (head_type); relation (relation); tail entity (tail), tail entity type (tail_type);

[0083] The relation types in the triplet fields of this embodiment include: training role, dependency, composition, application, and precedence.

[0084] S25: Summarize and clean the fields of the triples to obtain structured triples representing [capability-speed skating knowledge-training methods], providing business relationships for the basic capability graph;

[0085] S26: Construct a speed skating capability ontology tree based on speed skating training practice, and classify capabilities according to speed skating knowledge categories. The speed skating capability ontology tree serves as the structural constraint and unified namespace for the capability layer.

[0086] This embodiment categorizes speed skating abilities based on knowledge categories, including: physical fitness, skills, tactics, psychology, knowledge, and other multi-level structures. These multi-level structures form branches of the speed skating ability ontology tree.

[0087] S27: Calculate the cosine similarity between the content of consecutive segments and the description vector of the ontology node in the speed skating capability ontology tree, and set a threshold for the cosine similarity. Map consecutive segments with a cosine similarity exceeding the threshold to a unified capability node to achieve semantic alignment between capability and capability node, and avoid the fragmentation problem caused by multiple names for the same capability.

[0088] In this embodiment, the capability nodes are predefined high-level capability nodes in the speed skating capability ontology tree, which contain multiple finer-grained sub-capability nodes, forming a hierarchical capability structure.

[0089] S28: Using NetworkXMultiDiGraph as the graph structure container, import capability nodes from the speed skating capability ontology tree to form the capability layer; import technical or theoretical knowledge entities into the knowledge layer and training method entities into the training layer.

[0090] Construct directed multi-edges based on structured triples, record relation types and confidence levels, and add the ontology hierarchy relations of the speed skating capability ontology tree to the graph structure container in the form of is_parent_of edges;

[0091] A basic capability graph containing node types, business relationships, and ontology hierarchy is generated and exported in a specified format; in this embodiment, the specified format is JSON / GraphML or similar formats.

[0092] S3: Based on the basic ability graph, construct three types of teacher models to output multi-view ability vector representations, and construct student graph neural network GNN on the ability subgraph. Use the multi-view ability vector representations to train the student graph neural network GNN and output student ability vector representations.

[0093] Step S3 specifically includes:

[0094] S31: Based on the basic ability map, simulated training logs and competition / assessment data are introduced to form a multi-perspective ability vector representation;

[0095] The simulation training logs and competition / evaluation data in this embodiment are constructed using PyTorch tools;

[0096] Simulated training log: Represents changes in training load and content over time, used to describe the training process from a specific perspective;

[0097] Competition / Assessment Data: Primarily based on the mapping between contestant characteristics and ability scores or performance indicators, used to describe the perspective of results performance.

[0098] Three types of teacher models were constructed, including the KGTteacher model, the TemporalTeacher model, and the EvalTeacher model;

[0099] The KGTeacher model is a capability representation based on a fundamental capability graph:

[0100] On the capability subgraph of the basic capability graph, the graph relationship classification or link prediction task is constructed using the GATv2Conv graph convolutional layer of PyTorchGeometric. Through training, the capability vector representation KGTeacherembedding of each capability node in the graph structure container is obtained, which reflects the structural position of the capability corresponding to the capability node in the basic capability graph and the connection pattern with other capability nodes.

[0101] The TemporalTeacher model is a capability representation based on the training process:

[0102] TransformerEncoder is used to model the time-series data of the simulated training logs. The main training capability classification task is set to enable the model to learn the correspondence between different training time sequences and target capabilities. After training, capability vector representations are extracted from the classification head or related layer weights to serve as the capability vector representation TemporalTeacherembedding from the perspective of the training process.

[0103] The EvalTeacher model is a capability representation based on performance results:

[0104] For competition / assessment data, a regression model based on MLP is constructed to fit the functional relationship between input features and ability scores. The transpose of the output layer weights or the intermediate layer weights of the regression model are used as the ability vector representation EvalTeacherembedding from the perspective of performance results, reflecting the contribution of ability to the performance dimension.

[0105] S32: Construct a student graph neural network (GNN) on the capability subgraph. The student graph neural network (GNN) includes two layers of GATv2Conv graph convolutional layers and a projection head. The GATv2Conv graph convolutional layers are used to extract the local structural features of the capability nodes, and the projection head is used to map the local structural features to a capability representation space of a unified dimension.

[0106] S33: Train the student graph neural network (GNN) using the ability vector representations output by the three types of teacher models to obtain the student ability vector representation that integrates information on [textbook structure + training process + competition performance] in speed skating;

[0107] The loss function for training the student graph neural network (GNN) is:

[0108] ;

[0109] in, For the total loss, These are multi-teacher distillation loss and contrastive learning loss, respectively. To compare the weighting coefficients of the learning loss, These are the weighting coefficients for the three types of teacher models. Let $\mathbf{ ... These represent two different augmented views of the same capability node, where i and j are the numbers of the two different augmented views. This is the temperature coefficient.

[0110] S4: Write back the student ability vector representation and the multi-view ability vector representation to each ability node to form a multi-view ability description field; calculate the cosine similarity between ability nodes, generate similarity edges, and obtain the fused ability graph.

[0111] Step S4 specifically includes:

[0112] S41: The student ability vector representation and the ability vector representations KGTeacherembedding, TemporalTeacherembedding, and EvalTeacherembedding output by the three types of teacher models are used as ability node attributes and written back to each ability node to form a multi-perspective ability description field.

[0113] S42: Calculate the cosine similarity between ability nodes using student ability vectors. Based on the cosine similarity, select the K ability nodes with the highest similarity to each ability node. Establish a new edge of type ability_similarity between each ability node and the K ability nodes, generate similarity edges, and attach the cosine similarity weight corresponding to each similarity edge. The cosine similarity weight is used to explicitly represent the implicit ability relationship formed after the fusion of multi-source information.

[0114] S43: Store the capability node attributes (including multiple capability vector representations), business relationships (structural triples), ontology hierarchical relationships, and similarity edges in a unified manner, export them in a set format, and obtain the fused capability graph. The fused capability graph can be directly used as the data foundation for downstream retrieval and question answering modules.

[0115] S5: Load the fusion capability graph into the graph database or graph computing engine, build a comprehensive index, perform text parsing and vector encoding on the questions raised by users to form initial hit nodes; output the initial hit nodes into the trained Qwen3-14B model to output natural language answers to users.

[0116] Step S5 specifically includes:

[0117] S51: Load the exported fusion capability graph into a graph database or graph computing engine, establish a comprehensive index of node type, relationship type and similarity edge, and support multi-hop retrieval by capability node, technology or theoretical knowledge node, training method node and similarity edge;

[0118] S52: Perform text parsing and vector encoding on the questions raised by users. First, perform similarity retrieval between the text corpus and the description of capability nodes to locate the set of capability nodes most relevant to the question and their associated technical or theoretical knowledge nodes and training method nodes, forming an initial hit set.

[0119] S53: Based on the initial hit nodes in the initial hit set, perform the following operations according to the GraphRAG approach:

[0120] Extend the search along the ontology hierarchy, business relationships, and similarity edges for several hops, and extract local subgraphs related to the problem; summarize the capability links related to the problem from the local subgraphs; for example, "target capability - basic capability on which it depends - corresponding training method - technical actions involved";

[0121] S54: Organize the key node descriptions, relationship descriptions, and capability links in the local subgraphs into a structured text context to form a graph summary; input the user's question and the graph summary into the trained Qwen3-14B model, and the structured text context provides the capability nodes, training methods, and technical actions corresponding to the question;

[0122] S55: The Qwen3-14B model summarizes and expresses information based on the structured text context, and outputs user-oriented natural language answers.

[0123] The corresponding capability nodes, training method nodes, and their relationship paths serve as optional structured explanation information. This invention tightly integrates the "fine-tuned speed skating domain model" with the "3D capability graph," realizing a graph-assisted question-answering processing flow based on GraphRAG.

Claims

1. A speed skating knowledge question-and-answer method based on ability graphs, characterized in that, Includes the following steps: S1: Select literature with knowledge related to speed skating as the original corpus, extract question and answer samples, train the Qwen3-14B model, make the Qwen3-14B model adapt to the terminology, training logic and answer style of the speed skating field, and output the trained Qwen3-14B model. S2: Analyze the speed skating textbook, segment the text into continuous segments, use the GLiNER model to label three types of entities from the continuous segments, and output structured triples; Furthermore, a speed skating capability ontology tree is constructed, and the cosine similarity between the content of continuous segments and the description vector of ontology nodes is used to map continuous segments to capability nodes, thereby obtaining a basic capability map. S3: Based on the basic ability graph, construct three types of teacher models to output multi-view ability vector representations, and construct student graph neural network GNN on the ability subgraph. Use the multi-view ability vector representations to train the student graph neural network GNN and output student ability vector representations. S4: Write back the student ability vector representation and the multi-view ability vector representation to each ability node to form a multi-view ability description field; calculate the cosine similarity between ability nodes, generate similarity edges, and obtain the fused ability graph. S5: Load the fusion capability graph into the graph database or graph computing engine, build a comprehensive index, perform text parsing and vector encoding on the questions raised by users to form initial hit nodes; output the initial hit nodes into the trained Qwen3-14B model to output natural language answers to users.

2. The speed skating knowledge question-and-answer method based on capability graphs according to claim 1, characterized in that, Step S1 includes: S11: Select literature with knowledge related to speed skating as the original corpus, and use the Easydataset tool to segment, extract and organize the original corpus, construct question-and-answer samples in the form of [question-answer], and form a question-and-answer dataset; S12: Remove semantically unclear, duplicate, or contradictory question and answer samples from the question and answer dataset through manual or rule-based review to ensure the quality of the question and answer dataset; S13: Convert the question-and-answer dataset to a standard format supported by the llamafactory framework, configure the training parameters, load the Qwen3-14B model using the standard format question-and-answer dataset, and train the Qwen3-14B model to adapt to the terminology, training logic, and answering style of the speed skating field. S14: Output the trained Qwen3-14B model and perform spot checks and verification to ensure that the trained Qwen3-14B model is stable in terms of output format and professionalism in the field of speed skating. S15: Use the trained Qwen3-14B model as the generator for capability graph-assisted question answering.

3. The speed skating knowledge question-and-answer method based on capability maps according to claim 2, characterized in that, Step S2 includes: S21: Use PyMuPDF to parse the PDF file of the speed skating teaching material, extract the main text and structural information, and divide the text into continuous segments according to a predetermined length, which will serve as input units for subsequent entity recognition and relation extraction. S22: Use the GLiNER model to perform named entity recognition on continuous segments and label three types of entities: capability entities, technical or theoretical knowledge entities, and training method entities; S23: Label the three types of entities with category tags and text positions respectively, and store them uniformly; S24: Based on entity recognition, the Qwen3-32B model, which is based on the Transformers framework, automatically allocates devices for model loading, and runs in a multi-card environment, is called. The relationship between the three types of entities is extracted according to the preset Chinese Prompt template and output as a triple in strict JSON format. The triple fields include: head entity, head entity type; relationship; tail entity, tail entity type. S25: Summarize and clean the fields of the triples to obtain structured triples representing [capability-speed skating knowledge-training methods], providing business relationships for the basic capability graph; S26: Construct a speed skating capability ontology tree based on speed skating training practice, and classify capabilities according to speed skating knowledge categories. The speed skating capability ontology tree serves as the structural constraint and unified namespace for the capability layer. S27: Calculate the cosine similarity between the content of consecutive segments and the description vector of the ontology node in the speed skating capability ontology tree, and set a threshold for the cosine similarity. Map consecutive segments with a cosine similarity exceeding the threshold to a unified capability node to achieve semantic alignment between capability and capability node. S28: Using NetworkXMultiDiGraph as the graph structure container, import capability nodes from the speed skating capability ontology tree to form the capability layer; import technical or theoretical knowledge entities into the knowledge layer and training method entities into the training layer. Construct directed multi-edges based on structured triples, record relation types and confidence levels, and add the ontology hierarchy relations of the speed skating capability ontology tree to the graph structure container in the form of is_parent_of edges; Generate a basic capability map that includes node types, business relationships, and ontology hierarchy, and export it in a specified format.

4. The speed skating knowledge question-and-answer method based on capability maps according to claim 3, characterized in that, Step S3 includes: S31: Based on the basic ability map, simulated training logs and competition / assessment data are introduced to form a multi-perspective ability vector representation; Three types of teacher models were constructed, including the KGTeacher model, the TemporalTeacher model, and the EvalTeacher model; The KGTeacher model is based on the capability representation of the basic capability graph: On the capability subgraph of the basic capability graph, the graph relationship classification or link prediction task is constructed using the GATv2Conv graph convolutional layer of PyTorchGeometric; through training, the capability vector representation KGTeacherembedding of each capability node in the graph structure container is obtained; The TemporalTeacher model is based on the capability representation of the training process: TransformerEncoder is used to model the time-series data of the simulated training logs. The main training capability classification task is set to enable the model to learn the correspondence between different training time sequences and target capabilities. After training, capability vector representations are extracted from the classification head or related layer weights to serve as TemporalTeacherembedding, which represents the capability vectors from the perspective of the training process. The EvalTeacher model is a capability representation based on performance results: For competition / assessment data, a regression model based on MLP is constructed to fit the functional relationship between input features and ability scores; the transpose of the output layer weights or the intermediate layer weights of the regression model are used as the ability vector representation EvalTeacherembedding from the perspective of performance results. S32: Construct a student graph neural network (GNN) on the capability subgraph. The student graph neural network (GNN) includes two layers of GATv2Conv graph convolutional layers and a projection head. The GATv2Conv graph convolutional layers are used to extract the local structural features of the capability nodes, and the projection head is used to map the local structural features to a capability representation space of a unified dimension. S33: Train the student graph neural network (GNN) using the ability vector representations output by the three types of teacher models to obtain the student ability vector representation that integrates information on [textbook structure + training process + competition performance] in speed skating.

5. The speed skating knowledge question-and-answer method based on capability maps according to claim 4, characterized in that, The loss function for training the student graph neural network (GNN) is: ； in, For the total loss, These are multi-teacher distillation loss and contrastive learning loss, respectively. To compare the weighting coefficients of the learning loss, These are the weighting coefficients for the three types of teacher models. ； Let $\mathbf{ ... These are two different augmented view representations of the same capability node, where i and j are the numbers of the two different augmented views. This is the temperature coefficient.

6. The speed skating knowledge question-and-answer method based on capability maps according to claim 4, characterized in that, Step S4 includes: S41: The student ability vector representation and the ability vector representations KGTeacherembedding, TemporalTeacherembedding, and EvalTeacherembedding output by the three types of teacher models are used as ability node attributes and written back to each ability node to form a multi-perspective ability description field. S42: Calculate the cosine similarity between ability nodes using student ability vectors. Based on the cosine similarity, select the K ability nodes with the highest similarity to each ability node. Establish a new edge of type ability_similarity between each ability node and the K ability nodes, generate similarity edges, and attach the cosine similarity weight corresponding to each similarity edge. The cosine similarity weight is used to explicitly represent the implicit ability relationship formed after the fusion of multi-source information. S43: Store capability node attributes, business relationships, ontology hierarchical relationships, and similarity edges in a unified manner, export them in a set format, and obtain a fused capability graph. The fused capability graph can be directly used as the data foundation for downstream retrieval and question answering modules.

7. The speed skating knowledge question-and-answer method based on capability maps according to claim 6, characterized in that, Step S5 includes: S51: Load the exported fusion capability graph into a graph database or graph computing engine, establish a comprehensive index of node type, relationship type and similarity edge, and support multi-hop retrieval by capability node, technology or theoretical knowledge node, training method node and similarity edge; S52: Perform text parsing and vector encoding on the questions raised by users. First, perform similarity retrieval between the text corpus and the description of capability nodes to locate the set of capability nodes most relevant to the question and their associated technical or theoretical knowledge nodes and training method nodes, forming an initial hit set. S53: Based on the initial hit nodes in the initial hit set, perform the following operations according to the GraphRAG approach: Extend the search along the ontology hierarchy, business relationships, and similarity edges to extract local subgraphs related to the problem; summarize the capability links related to the problem from the local subgraphs. S54: Organize the key node descriptions, relationship descriptions, and capability links in the local subgraphs into a structured text context to form a graph summary; input the user's question and the graph summary into the trained Qwen3-14B model, and the structured text context provides the capability nodes, training methods, and technical actions corresponding to the question; S55: The Qwen3-14B model summarizes and expresses information based on the structured text context, and outputs user-oriented natural language answers.

Images