Deep learning-based teaching video key knowledge point fragment extraction method and system

Abstract

The invention provides a teaching video key knowledge point fragment extraction method and system based on deep learning. The method comprises the following steps: marking and deconstructing a teaching video to obtain a teaching scene metadatabase and a multi-source data deconstruction result corresponding to the teaching video; constructing a knowledge flow dynamic tracking and evolution graph by using the teaching scene metadatabase and the multi-source data deconstruction result; obtaining a cognitive load curve corresponding to the teaching video by utilizing the knowledge flow dynamic tracking and evolution graph; key knowledge point fragment boundary positioning is carried out on the teaching video through combination of the cognitive load curve and a map entropy sudden increase area corresponding to the knowledge flow dynamic tracking and evolution map, and a key knowledge point fragment boundary is determined; and intercepting the key knowledge point fragments according to the key knowledge point fragment boundaries to obtain a plurality of key knowledge point fragments. The system comprises modules corresponding to the steps of the method.

Description

TECHNICAL FIELD

[0001] The application provides a deep learning-based key knowledge point segment extraction method and system for teaching videos, and belongs to the technical field of video segment extraction. BACKGROUND

[0002] With the vigorous development of online education, teaching videos, as the core knowledge dissemination carrier, have shown exponential growth in quantity. However, the existing teaching videos generally have problems such as long duration and redundant content, making it difficult for learners to quickly locate the core knowledge points, resulting in low learning efficiency, and also increasing the difficulty of secondary development of teaching resources (such as micro course production and knowledge point question bank construction). Therefore, how to accurately and efficiently extract key knowledge point segments from teaching videos has become an important research topic in the field of online education.

[0003] Currently, the key knowledge point segment extraction technology for teaching videos mainly includes traditional methods and deep learning-based methods. Traditional methods mostly rely on manual annotation or simple feature matching, such as manually marking the time interval of the appearance of knowledge points, or segmenting based on a single feature such as audio volume threshold or video frame picture change rate. Such methods not only have low efficiency and strong subjectivity, but also cannot capture the semantic logic of knowledge points and the complexity of teaching scenarios, making it difficult to meet the actual demand for extraction accuracy. SUMMARY

[0004] The application provides a deep learning-based key knowledge point segment extraction method and system for teaching videos to solve the technical problems existing in the prior art, and the technical solutions adopted are as follows: The deep learning-based key knowledge point segment extraction method for teaching videos includes: By annotating and deconstructing the teaching video, the teaching scenario meta database and the multi-source data deconstruction result corresponding to the teaching video are obtained; The knowledge flow dynamic tracking and evolution graph is constructed using the teaching scenario meta database and the multi-source data deconstruction result; The cognitive load curve corresponding to the teaching video is obtained using the knowledge flow dynamic tracking and evolution graph; The key knowledge point segment boundary is located by combining the cognitive load curve with the sudden increase area of the graph entropy value corresponding to the knowledge flow dynamic tracking and evolution graph, and the key knowledge point segment boundary is determined; The key knowledge point segment is intercepted according to the key knowledge point segment boundary, and a plurality of key knowledge point segments are obtained.

[0005] Further, by annotating and deconstructing the teaching video, the teaching scenario meta database and the multi-source data deconstruction result corresponding to the teaching video are obtained, including: The teaching videos are labeled to obtain various types of tags, and a teaching scenario metadata database is constructed by combining these tags with the spatiotemporal features of the teaching videos. The various types of tags include knowledge type tags, interaction intensity tags, and cognitive load tags. A spatiotemporal attention decomposition method is used to decompose the teaching video into video frames, and the information regions corresponding to the teaching video are obtained from the video frames after decomposition; wherein, the information regions include knowledge carrier regions, teaching behavior regions, and background interference regions. Feature extraction is performed on the audio stream corresponding to the teaching video to obtain the key and emphasized sentence fragments corresponding to the teaching video; Text stream extraction is performed on the blackboard writing and PPT content of the teaching video. The text stream of the teaching video is parsed using a subject-specific large model to obtain the logical relationship tree contained in the text stream of the teaching video. The information regions, emphasized sentence fragments, and logical relationship trees are the results of multi-source data deconstruction.

[0006] Furthermore, a dynamic knowledge flow tracking and evolution graph is constructed using the aforementioned teaching scenario metadata database and multi-source data deconstruction results, including: Based on the various tags contained in the teaching scenario metadata database and the deconstruction results of multi-source data, knowledge atoms are extracted from the teaching content corresponding to the teaching videos, and the knowledge atom sequence K=[k1, k2, ..., k t ], where k1, k2, ..., k t These represent the knowledge atoms corresponding to time t; A temporal association matrix R is generated by the semantic similarity and time decay coefficient between every two knowledge atoms, wherein the temporal association matrix R contains matrix elements R0. ij This refers to the logical association strength between the i-th node and the j-th node; The knowledge atoms are used to construct a dynamic tracking and evolution graph of knowledge flow, and the dynamic tracking and evolution graph of knowledge flow is dynamically updated according to the update strategy.

[0007] Furthermore, the knowledge atoms are used to construct a dynamic tracking and evolution graph G(t) of the knowledge flow, including: The logical association strength between every two knowledge atoms is retrieved, and the logical association strength is compared with a preset first association strength threshold; wherein the preset first association strength threshold is set to 0.12-0.31. Select knowledge atoms whose logical association strength is not lower than a preset first association strength threshold as nodes; The logical association strength R between every two nodes ij As an edge; Using the node and the logical association strength R between the two nodes ij Generate a dynamic tracking and evolution map of knowledge flow.

[0008] Furthermore, the update strategy includes: The nodes contained in the knowledge flow dynamic tracking and evolution graph constructed from knowledge atoms are retrieved, wherein the nodes are the knowledge elements contained in the knowledge flow dynamic tracking and evolution graph. Retrieve the logical association strength R of every two nodes for self-checking ij ; The logical association strength R between every two nodes ij and the logical association strength R ij The comparison is made with a preset second association strength threshold, wherein the preset second association strength threshold is between 0.25 and 0.42; Extract logical association strength R ij Two nodes with a correlation strength not lower than a preset second correlation strength threshold are considered as a node pair; Retrieve the time interval influence factor corresponding to the node group; The dynamic update interval T of the current knowledge flow dynamic tracking and evolution graph is determined by using the time interval influence factor of all node groups; The current knowledge flow dynamic tracking and evolution graph is updated based on the dynamic update time interval T.

[0009] Furthermore, the cognitive load curves corresponding to the instructional videos are obtained using knowledge flow dynamic tracking and evolutionary mapping, including: The content complexity parameter G1(t) of the current knowledge flow dynamic tracking and evolution graph is obtained by using the node density, the sum of node weights and the numerical logical correlation strength. The teaching expression complexity parameter G2(t) is obtained by using the teacher's actions corresponding to the current teaching video, the volume intensity of the key sentence segments, and the logical relationship tree in the text stream; Retrieve the weight coefficients corresponding to the content complexity parameter G1(t) and the teaching expression complexity parameter G2(t), and use the content complexity parameter G1(t) and the teaching expression complexity parameter G2(t) in combination with their corresponding weight coefficients to obtain the cognitive load value L(t) through a weighted summation method; By combining the cognitive load value L(t) with the t moments contained in the corresponding duration of the instructional video, cognitive load curves for all instructional videos at all moments t are generated.

[0010] Furthermore, the content complexity parameter G1(t) of the current knowledge flow dynamic tracking and evolution graph is obtained by utilizing the node density, the sum of node weights, and the numerical logical correlation strength of the current knowledge flow dynamic tracking and evolution graph, including: Statistically calculate the node density and total node weight of the current knowledge flow dynamic tracking and evolution graph; Retrieve the logical association strengths of all nodes and edges contained in the current knowledge flow dynamic tracking and evolution graph; The average logical association strength is obtained by utilizing the logical association strength of all nodes and edges contained in the current knowledge flow dynamic tracking and evolution graph. The content complexity parameter G1(t) corresponding to the knowledge flow dynamic tracking and evolution graph at the current moment is obtained by combining the node density, the sum of node weights and the average value of logical association strength with the parameter weights corresponding to the node density, the sum of node weights and the average value of logical association strength through a weighted summation method.

[0011] Furthermore, the teaching expression complexity parameter G2(t) is obtained by utilizing the teacher's actions corresponding to the current teaching video, the volume intensity of the emphasized sentence fragments, and the logical relationship tree in the text stream, including: Retrieve the skeleton key points of the teacher's teaching process from the teaching behavior area contained in the information area of ​​the current teaching video, and use the skeleton key point sequence to obtain the motion entropy H(t); The volume intensity of the key sentences and phrases emphasized in the current teaching video is used to obtain the standard deviation of the volume intensity B(t). Retrieve the logical relation tree density M(t) from the text stream corresponding to the current teaching video; The motion entropy H(t), volume intensity standard deviation B(t), and logical relation tree density M(t) are normalized to obtain the normalized motion entropy H. g (t), Standard deviation of volume intensity B g (t) and the density of the logical relation tree M g (t); The motion entropy H after normalization is used g (t), Standard deviation of volume intensity B g (t) and the density of the logical relation tree M g (t) Obtain the teaching expression complexity parameter G2(t)=[H g (t)×max(B g (t), M g (t))] 0.5 .

[0012] Furthermore, by combining cognitive load curves with dynamic tracking of knowledge flow and the corresponding regions of sudden increases in spectral entropy in the evolutionary graph, the boundaries of key knowledge point segments in the teaching videos are located, determining the boundaries of key knowledge point segments, including: Retrieve the dynamic tracking and evolution graph of the knowledge flow, and determine the graph entropy sequence corresponding to the dynamic tracking and evolution graph of the knowledge flow through the nodes and node weights contained in the dynamic tracking and evolution graph of the knowledge flow; The difference between the spectral entropy values ​​corresponding to each two adjacent spectral entropy value sequences contained in the spectral entropy value sequence is obtained by performing difference processing on the spectral entropy values. The spectrum entropy difference is compared with a preset difference threshold, and the two time points corresponding to the two spectrum entropy value sequences with a spectrum entropy difference not lower than the preset difference threshold are taken as spectrum entropy value surge point groups, thereby obtaining multiple spectrum entropy value surge point groups. Based on the regions of sudden increases in spectral entropy values ​​formed by multiple groups of sudden increases in spectral entropy values, a segment of the cognitive load curve corresponding to the time point of the group of sudden increases in spectral entropy values ​​is extracted from the cognitive load curve. The maximum value of cognitive load appearing in each segment of the cognitive load curve is taken as the target cognitive load value; Extract the time span between the times corresponding to each two adjacent target cognitive load values, and compare the time span with a preset time span threshold, wherein the preset time span threshold ranges from 10s to 15s. If the time span between the times corresponding to two adjacent target cognitive load values ​​does not exceed the preset time span threshold, then the times corresponding to the two adjacent target cognitive load values ​​will be used as the start time point and the end time point, and the start time point and the end time point will be used as the boundary of the key knowledge point segment. If the time span between two adjacent target cognitive load values ​​exceeds a preset time span threshold, then the time corresponding to each target cognitive load value is taken as the time center, and the time points 7 seconds before and after the time center are taken as the start and end time points, and the start and end time points are taken as the boundaries of key knowledge point segments.

[0013] A deep learning-based system for extracting key knowledge points from teaching videos, comprising: The database construction and deconstruction module is used to obtain the teaching scenario metadata database and multi-source data deconstruction results corresponding to the teaching videos by annotating and deconstructing the teaching videos. The knowledge flow dynamic tracking and evolution graph construction module is used to construct a knowledge flow dynamic tracking and evolution graph using the teaching scenario metadata database and multi-source data deconstruction results; The cognitive load curve acquisition module is used to obtain the cognitive load curve corresponding to the teaching video by dynamically tracking and evolving the knowledge flow graph. The Key Knowledge Point Segment Boundary Determination Module is used to locate the boundaries of key knowledge point segments in teaching videos by combining the cognitive load curve with the dynamic tracking and evolution map of knowledge flow and the corresponding region of sudden increase in map entropy value. The key knowledge point fragment extraction module is used to extract key knowledge point fragments based on their boundaries, thereby obtaining multiple key knowledge point fragments.

[0014] Beneficial effects of this invention: The deep learning-based method and system for extracting key knowledge point segments from teaching videos proposed in this invention can effectively improve the accuracy of key knowledge point identification and effectively solve the problems of blurred segment boundaries and logical breaks in existing technologies. It avoids the omission of core knowledge or the inclusion of redundant content, thereby effectively improving the matching degree between the extraction results and the course syllabus. Through cognitive load quantification, it adapts to students' cognitive patterns, ensuring that the segment content not only conforms to the teaching focus but also matches students' cognitive acceptance abilities, while meeting the practical needs of teachers in lesson preparation, micro-lesson production, and other teaching scenarios. It effectively reduces the fluctuation in the extraction accuracy of knowledge point segments in teaching videos of different disciplines such as humanities, sciences, engineering, and experimental subjects. Attached Figure Description

[0015] Figure 1 This is a flowchart of the method described in this invention; Figure 2 This is a system block diagram of the system described in this invention. Detailed Implementation

[0016] The preferred embodiments of the present invention will be described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0017] This invention proposes a method for extracting key knowledge point segments from teaching videos based on deep learning, such as... Figure 1 As shown, the method for extracting key knowledge point segments from teaching videos includes: By annotating and deconstructing teaching videos, a teaching scenario metadata database and multi-source data deconstruction results are obtained. The teaching scenario metadata database includes various types of tags combined with the spatiotemporal characteristics of the teaching videos. These various types of tags include knowledge type tags, interaction intensity tags, and cognitive load tags. Specifically, knowledge type tags include, but are not limited to, concept definitions, formula derivations, example explanations, experimental demonstrations, and summaries; interaction intensity tags include, but are not limited to, teacher monologues, teacher-student Q&A, and group discussions; and cognitive load tags include, but are not limited to, low load (L1, such as common sense review), medium load (L2, such as concept clarification), and high load (L3, such as complex deduction). The multi-source data deconstruction results include information areas (such as blackboard math areas, knowledge content demonstration areas, and video subtitle areas), key emphasized sentence fragments (such as key knowledge point fragments and key question type fragments), and logical relationship trees (such as subject-predicate, causality, and deduction).

[0018] A dynamic knowledge flow tracking and evolution graph is constructed using the teaching scenario metadata database and the results of multi-source data deconstruction. The cognitive load curves corresponding to the teaching videos are obtained by using knowledge flow dynamic tracking and evolution graphs; By combining the cognitive load curve with the dynamic tracking and evolution map of knowledge flow, the boundary of key knowledge point segments in the teaching video is located to determine the boundary of key knowledge point segments. Based on the boundaries of key knowledge point segments, extract multiple key knowledge point segments.

[0019] The working principle of the above technical solution is as follows: This embodiment takes the simulation of the teaching and cognitive process as the core, and achieves accurate extraction of key knowledge point fragments through multi-dimensional data processing and dynamic modeling: First, the teaching video is annotated with scenes and deconstructed into multi-modal data to construct a basic dataset containing teaching behavior features and multi-source cognitive dimension data; based on this dataset, the smallest knowledge unit (knowledge atom) is identified, and a knowledge evolution graph is generated through temporal correlation modeling and dynamic update mechanism to capture the logical relationship and evolution law of knowledge points; the content features and multi-modal expression features of the knowledge evolution graph and student cognitive feedback data are integrated to quantify and generate a high temporal resolution cognitive load curve; the dense range of knowledge points is located by the region of sudden increase in graph entropy value, and the initial location, multi-dimensional expansion calibration and closed-loop verification of the key fragment boundary are completed by combining the waveform features of the cognitive load curve; finally, the key knowledge point fragment is extracted according to the calibrated boundary, and the model parameters can be dynamically optimized through user feedback to continuously improve the extraction performance.

[0020] The above technical solution effectively improves the accuracy of key knowledge point identification and solves the problems of blurred segment boundaries and logical breaks in existing technologies. It avoids the omission of core knowledge or the inclusion of redundant content, thereby effectively improving the matching degree between the extraction results and the course syllabus. Through cognitive load quantification, it adapts to students' cognitive patterns, ensuring that the segment content not only aligns with teaching priorities but also matches students' cognitive abilities. Simultaneously, it meets the practical needs of teachers in lesson preparation, micro-lesson production, and other teaching scenarios. In teaching videos across different disciplines such as arts, sciences, engineering, and experimental subjects, it effectively reduces fluctuations in the accuracy of knowledge point segment extraction, ensuring the accuracy and reliability of key knowledge point judgment and extraction.

[0021] One embodiment of the present invention involves annotating and deconstructing teaching videos to obtain a teaching scenario metadata database and multi-source data deconstruction results corresponding to the teaching videos, including: By annotating teaching videos, various types of tags are obtained corresponding to the videos. A teaching scenario metadata database is constructed using these tags combined with the spatiotemporal characteristics of the videos. These tags include knowledge type tags, interaction intensity tags, and cognitive load tags. Specifically, knowledge type tags include, but are not limited to, concept definitions, formula derivations, example explanations, experimental demonstrations, and summaries. Interaction intensity tags include, but are not limited to, teacher monologues, teacher-student Q&A, and group discussions. Cognitive load tags include, but are not limited to, low load (L1, such as common sense review), medium load (L2, such as concept clarification), and high load (L3, such as complex deduction). A spatiotemporal attention decomposition method is used to decompose teaching videos into video frames, and the information regions corresponding to the teaching videos are obtained from the decomposed video frames. The information regions include knowledge carrier regions, teaching behavior regions, and background interference regions. Specifically, the knowledge carrier region is obtained by extracting the blackboard writing and PPT content in the teaching video through a semantic segmentation model; the teaching behavior region is obtained by capturing the teacher's gestures and expressions in the teaching video through a skeleton extraction model; and the background interference region is obtained by denoising and separating the teaching video through a generative adversarial network (GAN). Feature extraction is performed on the audio stream corresponding to the teaching video to obtain the key and emphasized sentence segments in the teaching video; among them, the key and emphasized sentence segments are those whose volume intensity exceeds the preset volume intensity.

[0022] Text stream extraction is performed on the blackboard writing and PPT content of the teaching video. The text stream of the teaching video is parsed using a subject-specific large model (such as MathGPT, PhysGPT) to obtain the logical relationship tree contained in the text stream of the teaching video, such as subject-verb, causal and inference relationships. Among them, the information area, the highlighted sentence fragments, and the logical relationship tree are the results of multi-source data deconstruction.

[0023] The working principle of the above technical solution is as follows: This embodiment focuses on the basic data preprocessing stage of teaching videos. It achieves the structured transformation of data through a dual path of annotation and classification combined with multimodal deconstruction: First, the teaching videos are annotated with scenes using a multi-dimensional tag system (knowledge type, interaction intensity, cognitive load), and a teaching scene meta-database is constructed by associating spatiotemporal features, giving the data teaching attributes; Second, the video stream is decomposed using spatiotemporal attention, and three types of regions are separated, namely knowledge carrier, teaching behavior, and background interference, by combining semantic segmentation, skeleton extraction, and GAN noise reduction technology; At the same time, high-volume key sentence segments in the audio stream are extracted, and the deep logical relationship tree of the text stream (such as subject-predicate, causality, inference, etc.) is analyzed through a subject-specific large model, and finally a structured deconstruction result containing visual, audio, and text multimodal elements is formed, providing standardized and scenario-based basic data support for subsequent knowledge modeling and knowledge point extraction.

[0024] The above technical solution achieves the following results: This embodiment transforms unstructured teaching videos into structured data containing teaching attribute tags, regional visual features, key audio segments, and deep textual logic, effectively solving the problem of traditional teaching videos being difficult to accurately parse and utilize. The solution optimizes the visual and audio processing workflow, accurately extracting knowledge carrier regions, suppressing background interference, identifying key audio sentences, and simultaneously parsing deep textual logical relationships, significantly reducing the interference of invalid data on subsequent processing. Relying on the combination of multi-dimensional teaching tags and subject-specific large-scale models, the solution can adapt to the teaching characteristics of different disciplines such as arts, sciences, engineering, and experiments, breaking through the scenario limitations of general data processing solutions. Furthermore, the structured multi-source deconstruction results can provide accurate input for downstream processes such as knowledge atom recognition, knowledge evolution graph construction, and cognitive load calculation, improving the efficiency of downstream knowledge point extraction and reducing the training complexity of downstream models.

[0025] Meanwhile, by leveraging GAN networks to separate background interference, the solution maintains stable data deconstruction results even in complex teaching environments with varying lighting, environmental noise, and background clutter, outperforming traditional methods. Furthermore, the solution requires only lightweight initial annotation to achieve automated multimodal deconstruction, significantly reducing processing costs and improving efficiency, meeting the batch preprocessing needs of large-scale teaching videos. The deep logical relationship tree parsing of text streams breaks through the limitations of traditional text processing that focuses on surface-level content, providing core support for capturing the logical connections between knowledge points and improving the logical integrity of knowledge evolution modeling.

[0026] One embodiment of the present invention utilizes the teaching scenario metadata database and multi-source data deconstruction results to construct a dynamic knowledge flow tracking and evolution graph, including: Based on the various tags contained in the teaching scenario metadata database and the deconstruction results of multi-source data, knowledge atoms are extracted from the teaching content corresponding to the teaching videos, and the knowledge atom sequence K=[k1, k2, ..., k t ], where k1, k2, ..., k t These represent the knowledge atoms corresponding to time points t, where each knowledge atom refers to the smallest knowledge unit appearing in the teaching video, such as "definition of trigonometric functions" and "statement of Newton's first law".

[0027] A temporal correlation matrix R is generated by the semantic similarity and time decay coefficient between every two knowledge atoms, wherein the temporal correlation matrix R contains matrix elements R0. ij =sqrt[w×S sem (k i k j )+(1-w)×S tsem (k i k j The logical association strength between the i-th node and the j-th node is represented by R, and the matrix is ​​composed of the association strength R. ij As a matrix formed by elements, and R ij This represents the logical association strength between the i-th knowledge atom and the j-th knowledge atom, where w represents the preset semantic weight coefficient, with a default value of 0.7; S sem (k i k j ) represents the cosine similarity between the semantic feature vectors of the i-th knowledge atom and the j-th knowledge atom; k i and k j They refer to the i-th knowledge atom and the j-th knowledge atom, respectively; S tsem (k i k j ) represents the time interval influence factor between the i-th knowledge atom and the j-th knowledge atom, and S tsem (k i k j )=exp(-λ×|t si -t zj |); where λ represents the time decay coefficient (default 0.05, unit: 1 / second); t si t represents the starting time of the i-th knowledge atom; zj This represents the end time of the j-th knowledge atom.

[0028] We construct a dynamic tracking and evolution graph of knowledge flow using knowledge atoms, and dynamically update the dynamic tracking and evolution graph of knowledge flow according to the update strategy.

[0029] The working principle of the above technical solution is as follows: In this embodiment, the above technical solution achieves accurate tracking and evolution modeling of knowledge flow in teaching videos through a progressive logic of knowledge unit decomposition, association strength quantification, and dynamic graph construction. First, based on the multi-dimensional tags (such as knowledge type, cognitive load, etc.) and multi-source deconstruction results (such as text logic, audio focus, visual knowledge carrier) of the teaching scenario metadata database, the smallest knowledge unit (knowledge atom) with independent teaching significance is decomposed from the teaching content to form a sequence of knowledge atoms with timestamps. Second, by integrating semantic similarity (knowledge content association) and time decay coefficient (spatiotemporal distance influence), the logical association strength between any two knowledge atoms is quantified to generate a temporal association matrix. Finally, a dynamic evolution graph is constructed with knowledge atoms as nodes and association strength as edges, and the temporal evolution law of knowledge flow with the teaching process is captured through a real-time update strategy (such as sliding window incremental update), realizing the dynamic tracking of knowledge association and structure.

[0030] The effects of the above technical solution are as follows: This embodiment accurately achieves knowledge atom identification and clear boundary definition, ensuring the independence and integrity of the smallest knowledge unit and solving the problems of vague splitting and incomplete semantics in traditional methods; it effectively captures various logical relationships between knowledge atoms through a temporal correlation matrix, providing reliable quantitative support for knowledge logic chain analysis; the dynamic evolution map can reflect the addition of knowledge, changes in association, and structural evolution in real time, effectively avoiding the lag or static bias of knowledge association; and it adapts to the characteristics of multiple disciplines through differentiated weight settings, providing structured knowledge association data for subsequent cognitive load calculation, boundary calibration, and other steps, effectively reducing the downstream computational complexity, thereby improving the overall efficiency of the knowledge point extraction process, while enhancing the logical interpretability of extracted knowledge point fragments, facilitating teacher verification and manual intervention.

[0031] Compared to existing technologies, this scheme breaks through the limitations of traditional linear weighting, fixed threshold filtering, or fixed coefficients. By using a square root nonlinear fusion method, it amplifies the synergistic effect of semantic and temporal contributions, effectively capturing various weakly related knowledge in teaching scenarios and fully covering the entire chain of knowledge associations from basic foundation to core and extension. Leveraging the nonlinear adjustment of semantic weights and temporal factors by the square root function, combined with the exponential characteristics of time decay, the association strength calculation can dynamically adapt to the knowledge association patterns of different disciplines. It has an implicit filtering effect on the consistency of semantic and temporal features, significantly reducing erroneous associations caused by interfering data such as "semantically similar but logically unrelated" or "temporally close but semantically fragmented," thus improving the noise resistance of association calculations. Through the synergistic effect of exponential decay and square root fusion, it accurately preserves long-term strong semantic associations, avoids the false reinforcement of short-term weak semantic associations, and fully restores the teaching logic chain structure from introduction to expansion and then summarization, fully adapting to the unique patterns of knowledge associations in teaching scenarios.

[0032] One embodiment of the present invention utilizes the knowledge atoms to construct a dynamic tracking and evolution graph G(t) of knowledge flow, comprising: The logical association strength between every two knowledge atoms is retrieved, and the logical association strength is compared with a preset first association strength threshold; wherein the preset first association strength threshold is set to 0.12-0.31. Select knowledge atoms whose logical association strength is not lower than a preset first association strength threshold as nodes; The logical association strength R between every two nodes ij As an edge; Using the node and the logical association strength R between the two nodes ij Generate a dynamic tracking and evolution map of knowledge flow.

[0033] The working principle of the above technical solution is as follows: In this embodiment, the above technical solution constructs a knowledge flow dynamic tracking and evolution graph through the logical construction from association strength screening to node edge construction and then to graph generation: First, the logical association strength R between knowledge atoms is retrieved. ij The association strength is compared with a preset first association strength threshold (0.12-0.31), and knowledge atoms with an association strength not lower than this threshold are selected as graph nodes; the corresponding R values ​​between nodes are used as the graph nodes. ij The values ​​serve as the weights of the connecting edges. Through graph structure modeling, the temporal associations and logical relationships of knowledge atoms are visualized, ultimately generating a dynamic graph G(t) that reflects the flow and evolution of knowledge associations during the teaching process, providing a structured association carrier for knowledge flow tracing.

[0034] The above technical solution achieves the following results: By using a specific threshold verified in teaching scenarios, weak associations are effectively filtered out, invalid edges are prevented from interfering with the graph, and the density of graph nodes to edges is reduced. This significantly improves the clarity and parsability of the graph, solving the problems of graph redundancy and diluted core associations caused by the absence of a threshold or an unreasonable threshold in existing technologies. This threshold can fully retain effective teaching associations, especially strengthening the retention of medium-strength associations, ensuring that transitional associations in the knowledge flow are not missed and fully restoring the teaching logic chain. The simplified graph not only reduces the time complexity of subsequent dynamic updates and entropy calculations, shortening processing time, but also provides more focused association data for downstream steps such as cognitive load calculation and boundary calibration, reducing the consumption of invalid computing resources. The edge weights in the solution directly adopt the association strength value, which can intuitively reflect the hierarchical differences in association strength and is more accurate than existing binarization processing, providing reliable support for the identification of core knowledge nodes. This threshold is determined based on multi-disciplinary data training, which can adapt to the association characteristics of different disciplines, effectively improving cross-scenario stability and ensuring the consistency of knowledge graph construction quality across different disciplines. Meanwhile, the precise selection of nodes and edges, combined with a dynamic update strategy, enables the graph to clearly record the complete process of knowledge association, effectively improving the traceability of knowledge evolution trajectory and providing strong support for teaching quality analysis.

[0035] In one embodiment of the present invention, the update strategy includes: The nodes contained in the knowledge flow dynamic tracking and evolution graph constructed from knowledge atoms are retrieved, wherein the nodes are the knowledge elements contained in the knowledge flow dynamic tracking and evolution graph. Retrieve the logical association strength R of every two nodes for self-checking ij ; The logical association strength R between every two nodes ij and the logical association strength R ij The second association strength threshold is compared with a preset second association strength threshold, which is set between 0.25 and 0.42. The second association strength threshold is larger than the first association strength threshold because the first association strength threshold is used for the initial construction of the dynamic knowledge flow graph, which needs to ensure the integrity of the graph and avoid missing potential key knowledge associations, so the value is relatively small. The second association strength threshold is used to determine the dynamic update time interval of the graph, which needs to filter out irrelevant weak association interference and focus on the core node group that plays a dominant role in the evolution to ensure accurate update decisions, so the value is relatively large.

[0036] Extract logical association strength R ij Two nodes with a correlation strength not lower than a preset second correlation strength threshold are considered as a node pair; Retrieve the time interval influence factor corresponding to the node group; The dynamic update interval of the knowledge flow dynamic tracking and evolution graph is determined by using the time interval influence factor of all node groups. Where T0 represents the preset initial time interval, ranging from 20s to 30s; n represents the total number of node groups; S temk S represents the time interval influence factor corresponding to the k-th node pair; temp S represents the average value of the time interval influence factor between every two nodes in the knowledge flow dynamic tracking and evolution graph; semp This represents the average cosine similarity between every two nodes contained in all nodes of the knowledge flow dynamic tracking and evolution graph; The current knowledge flow dynamic tracking and evolution graph is updated based on the dynamic update time interval T.

[0037] The working principle of the above technical solution is as follows: This update strategy dynamically adjusts the update time interval of the knowledge flow dynamic tracking and evolution graph by quantifying the strength and density of knowledge associations. First, it extracts the knowledge atomic nodes in the graph, calculates the logical association strength between nodes, and selects node groups with an association strength not lower than the second association strength threshold (0.25-0.42). Then, it obtains the time interval influence factor of these node groups, and calculates the dynamic update time interval T by combining the formula of "difference of node group time interval influence factor", "mean cosine similarity", and "mean time interval influence factor" in combination with the preset initial time interval. Finally, it updates the knowledge flow dynamic tracking and evolution graph in real time according to T, so as to achieve precise adaptation of the update frequency and the dynamics of knowledge associations.

[0038] The above technical solution achieves the following effects: By quantifying the density and strength of knowledge associations through formulas, the update time interval T is dynamically adjusted according to the tightness of the knowledge flow associations, effectively avoiding the resource waste or graph lag problems associated with fixed update intervals. The dynamic update strategy, while ensuring real-time knowledge flow tracking, reduces the computational overhead of graph updates, adapting to batch processing scenarios of large-scale teaching videos, and reducing unnecessary computational resource consumption. The dynamic adaptation of the update time interval shortens the capture delay of knowledge evolution, enabling more timely reflection of the addition, strengthening, or decay of knowledge associations during the teaching process, effectively improving the graph's accuracy in reconstructing the temporal patterns of knowledge flow, and providing more accurate dynamic knowledge association support for subsequent key segment extraction. The parameters in the formula are calculated based on the actual data of the node groups, significantly enhancing the adaptability of the update strategy in teaching videos of different subjects and teaching styles, ensuring consistency of update effects in various scenarios without manual intervention, and reducing adaptation errors. Furthermore, this update strategy focuses on node groups with high association strength, prioritizing the capture of the association evolution of core knowledge, filtering out interference from irrelevant weak associations, ensuring that the graph can focus on the dynamic changes of the core teaching logic, and effectively improving the tracking accuracy of key teaching knowledge flows.

[0039] One embodiment of the present invention utilizes knowledge flow dynamic tracking and evolutionary graphs to obtain the cognitive load curve corresponding to the instructional video, including: The content complexity parameter G1(t) of the current knowledge flow dynamic tracking and evolution graph is obtained by using the node density, the sum of node weights and the numerical logical correlation strength. The teaching expression complexity parameter G2(t) is obtained by using the teacher's actions corresponding to the current teaching video, the volume intensity of the key sentence segments, and the logical relationship tree in the text stream; The weight coefficients corresponding to the content complexity parameter G1(t) and the teaching expression complexity parameter G2(t) are retrieved, and the cognitive load value L(t) = x1×G1(t) + x2×G2(t) is obtained by weighted summation using the content complexity parameter G1(t) and the teaching expression complexity parameter G2(t) and their corresponding weight coefficients; where x1 and x2 represent the weight coefficients corresponding to the content complexity parameter G1(t) and the teaching expression complexity parameter G2(t), respectively; and the values ​​of x1 and x2 are in the range of (0-1], preferably x1=0.6 and x2=0.4; By combining the cognitive load value L(t) with the t time points contained in the corresponding duration of the instructional video, cognitive load curves are generated for all instructional videos at time t. Specifically, based on the L(t) formula, the x-axis represents time t, and the y-axis represents the corresponding value of L(t). The cognitive load curves are automatically generated using a curve generation tool.

[0040] The working principle of the above technical solution is as follows: In this embodiment, the above technical solution constructs a cognitive load curve through a dual-dimensional complexity quantification and weighted fusion approach. The core logic is as follows: Based on the structural features (node ​​density, sum of node weights, and strength of logical associations) of the knowledge flow dynamic tracking and evolution graph, the complexity of the teaching content itself is quantified to obtain the content complexity parameter G1(t); at the same time, multimodal expression features (teacher actions, volume intensity of key sentences, and text logical relationship tree) are fused to quantify the complexity of the teaching transmission process to obtain the teaching expression complexity parameter G2(t); by using preset weight coefficients x1 and x2 to perform weighted summation of G1(t) and G2(t), the cognitive load value L(t) at each time t is calculated, and finally, the cognitive load curve for the entire time period is output in combination with the duration of the teaching video, thereby realizing a quantitative characterization of the student's real-time cognitive state.

[0041] The above technical solution achieves the following effects: Content complexity parameters are accurately quantified using structured data from the knowledge evolution graph; the teaching expression complexity parameter integrates multimodal expression features to capture the difficulty of teaching delivery; and the weighted fusion of these two methods solves the problem of one-sided cognitive load assessment in existing technologies, effectively improving the comprehensiveness and accuracy of the assessment. The solution outputs a cognitive load curve with a 1-second time granularity, enabling rapid capture of instantaneous changes in cognitive load. Its dynamic response speed is superior to traditional low-resolution curves, providing accurate temporal feature support for subsequent boundary positioning. Simultaneously, this curve can completely depict the load peak during the explanation of core knowledge points and the load waveform of the complete knowledge chain, providing core temporal basis for key segment boundary positioning and calibration, effectively reducing positioning errors. Through multimodal data fusion, the solution enhances the anti-interference ability of relevant parameters against single-modal noise, reduces the fluctuation amplitude of the cognitive load curve, and ensures stable output of effective load features even in complex teaching environments.

[0042] One embodiment of the present invention utilizes the node density, the sum of node weights, and the numerical logical correlation strength of the current knowledge flow dynamic tracking and evolution graph to obtain the content complexity parameter G1(t), including: Statistically calculate the node density and total node weight of the current knowledge flow dynamic tracking and evolution graph; Retrieve the logical association strengths of all nodes and edges contained in the current knowledge flow dynamic tracking and evolution graph; The average logical association strength is obtained by utilizing the logical association strength of all nodes and edges contained in the current knowledge flow dynamic tracking and evolution graph. The content complexity parameter G1(t) corresponding to the knowledge flow dynamic tracking and evolution graph at the current moment is obtained by combining the node density, the sum of node weights and the average value of logical association strength with the parameter weights corresponding to the node density, the sum of node weights and the average value of logical association strength through a weighted summation method.

[0043] The working principle of the above technical solution is as follows: In this embodiment, the above technical solution quantifies the structural features of the knowledge evolution graph in multiple dimensions and constructs a content complexity parameter G1(t): First, it calculates the node density (reflecting the density of knowledge units per unit time) and the sum of node weights (reflecting the overall importance of knowledge units) of the current knowledge flow dynamic tracking and evolution graph; at the same time, it retrieves the logical association strength of all node edges in the graph and calculates its average value (reflecting the overall complexity of logical associations between knowledge units); finally, it assigns corresponding parameter weights to the node density, the sum of node weights, and the average value of logical association strength, and integrates the features of the three dimensions into a single content complexity parameter G1(t) through weighted summation, thereby achieving a comprehensive quantification of the complexity of the teaching content itself.

[0044] The above technical solution achieves the following results: By integrating node density, the sum of node weights, and the average value of logical association strength across multiple dimensions, it overcomes the limitations of traditional single-indicator evaluation, expanding the dimensions for describing content complexity from one to three. This effectively improves information coverage and more comprehensively reflects the combined characteristics of knowledge quantity, importance, and correlation. Quantification errors across all dimensions are controlled at a low level. After multi-dimensional weighted fusion, the correlation between the content complexity parameter and manual annotation is significantly and effectively improved, ensuring accurate quantification of content complexity. By dynamically adjusting the parameter weights of the three dimensions, it adapts to the content characteristics of different disciplines, significantly improving cross-disciplinary stability and reducing quantification adaptation errors. This parameter provides core content feature input for the cognitive load curve. The effective improvement in its quantification accuracy helps improve the accuracy of the cognitive load curve in capturing the cognitive load peak corresponding to high-complexity content, providing a reliable basis for content complexity in subsequent key segment boundary positioning. The complementarity of multi-dimensional features enhances the parameter's resistance to noise from single features, reduces numerical fluctuations, and ensures stable output of effective complexity features even in complex knowledge scenarios. Meanwhile, the continuous value of this parameter can accurately reflect subtle changes in content complexity, clearly depict the difficulty gradient, and provide quantitative support for subsequent identification of key segments that conform to the cognitive progression pattern.

[0045] One embodiment of the present invention utilizes the teacher's actions corresponding to the current teaching video, the volume intensity of emphasized sentence fragments, and the logical relationship tree in the text stream to obtain the teaching expression complexity parameter G2(t), including: Retrieve the skeleton key points of the teacher's teaching process from the teaching behavior area contained in the information area of ​​the current teaching video, and use the skeleton key point sequence to obtain the motion entropy H(t); The volume intensity of the key sentences and phrases emphasized in the current teaching video is used to obtain the standard deviation of the volume intensity B(t). Retrieve the logical relation tree density M(t) from the text stream corresponding to the current teaching video; The motion entropy H(t), volume intensity standard deviation B(t), and logical relation tree density M(t) are normalized to obtain the normalized motion entropy H. g (t), Standard deviation of volume intensity B g (t) and the density of the logical relation tree M g (t); The motion entropy H after normalization is used g (t), Standard deviation of volume intensity B g (t) and the density of the logical relation tree M g (t) Obtain the teaching expression complexity parameter G2(t)=[H g (t)×max(B g (t), M g (t))] 0.5 .

[0046] The working principle of the above technical solution is as follows: In this embodiment, the above technical solution quantifies the complexity of teaching expression through multimodal feature fusion: extracting key points of the teacher skeleton from the teaching behavior area to calculate motion entropy (reflecting action complexity), extracting volume intensity from key emphasized sentence fragments to calculate standard deviation (reflecting the degree of emphasis in audio expression), and extracting logical relationship tree from the text stream to calculate density (reflecting the logical complexity of text expression); after normalizing the three features, the teaching expression complexity parameter G2(t) is output through a formula that nonlinearly fuses the maximum value of motion entropy and volume / logical complexity, thereby achieving a comprehensive quantification of the complexity of the teaching transmission process.

[0047] The above technical solution achieves the following results: By integrating visual, audio, and textual features, it overcomes the limitations of single-modal assessment, expanding the dimension of teaching expression complexity from one-dimensional to three-dimensional, improving information coverage, and comprehensively reflecting the integrated expressive characteristics of action richness, expression emphasis, and logical density. The quantification error of each modality-related indicator is controlled at a low level. After multimodal fusion, the correlation between the teaching expression complexity parameter and manual annotation is significantly improved, ensuring accurate quantification of teaching expression complexity. The solution significantly improves the cross-scenario stability of the parameter under different subjects and teaching styles by dynamically capturing the differentiated performance of teacher actions, audio emphasis, and text logic, reducing adaptation error and accurately adapting to diverse teaching expression scenarios. This parameter provides core expressive feature input for the cognitive load curve. The improved quantification accuracy helps improve the accuracy of the cognitive load curve in capturing cognitive load fluctuations corresponding to high expressive complexity, providing a reliable basis for expression complexity for subsequent key segment boundary positioning. The complementarity of multimodal features enhances the parameter's resistance to single-modal noise, reduces numerical fluctuation amplitude, and ensures stable output of effective expression complexity features even in complex teaching environments. At the same time, the nonlinear fusion formula of parameters can highlight the dominant expressive dimensions in action, volume, and logic, and maximize the differentiation of teaching expression styles.

[0048] One embodiment of the present invention uses a cognitive load curve combined with dynamic tracking of knowledge flow and an evolutionary graph to locate the boundaries of key knowledge point segments in teaching videos, determining the boundaries of key knowledge point segments, including: Retrieve the dynamic tracking and evolution graph of the knowledge flow, and determine the graph entropy sequence corresponding to the dynamic tracking and evolution graph of the knowledge flow through the nodes and node weights contained in the dynamic tracking and evolution graph of the knowledge flow; The difference between the spectral entropy values ​​corresponding to each two adjacent spectral entropy value sequences contained in the spectral entropy value sequence is obtained by performing difference processing on the spectral entropy values. The spectrum entropy difference is compared with a preset difference threshold, and the two time points corresponding to the two spectrum entropy value sequences with a spectrum entropy difference not lower than the preset difference threshold are taken as spectrum entropy value surge point groups, thereby obtaining multiple spectrum entropy value surge point groups. Based on the regions of sudden increases in spectral entropy values ​​formed by multiple groups of sudden increases in spectral entropy values, a segment of the cognitive load curve corresponding to the time point of the group of sudden increases in spectral entropy values ​​is extracted from the cognitive load curve. The maximum value of cognitive load appearing in each segment of the cognitive load curve is taken as the target cognitive load value; Extract the time span between the times corresponding to each two adjacent target cognitive load values, and compare the time span with a preset time span threshold, wherein the preset time span threshold ranges from 10s to 15s. If the time span between the times corresponding to two adjacent target cognitive load values ​​does not exceed the preset time span threshold, then the times corresponding to the two adjacent target cognitive load values ​​will be used as the start time point and the end time point, and the start time point and the end time point will be used as the boundary of the key knowledge point segment. If the time span between two adjacent target cognitive load values ​​exceeds a preset time span threshold, then the time corresponding to each target cognitive load value is taken as the time center, and the time points 7 seconds before and after the time center are taken as the start and end time points, and the start and end time points are taken as the boundaries of key knowledge point segments.

[0049] The working principle of the above technical solution is as follows: In this embodiment, the above technical solution achieves the boundary positioning of key knowledge point segments by combining knowledge-intensive area positioning with cognitive load peak analysis and time span adaptive calibration: First, nodes and weights are extracted from the knowledge flow dynamic tracking and evolution graph, the graph entropy value sequence is calculated and the entropy value surge point group is identified to determine the graph entropy value surge area with dense knowledge points; then, the load segment of the corresponding area is intercepted on the cognitive load curve, and the target cognitive load value (maximum cognitive load value) is extracted; finally, the boundary of the key knowledge point segment is adaptively determined according to the time span of adjacent target cognitive load values—if the time span is small, the start and end times are directly used as the boundary; if the span is large, the boundary is extended by 7 seconds before and after the target time as the center, so as to achieve accurate boundary positioning that conforms to cognitive laws.

[0050] The above technical solution achieves the following effects: By co-locating the graph entropy value and the peak cognitive load, it effectively controls boundary positioning errors, ensuring the complete coverage of the cognitive chain of key segments, solving the problems of blurred boundaries and knowledge gaps in existing technologies, and avoiding the omission of core knowledge points or the inclusion of redundant content. With the help of time-span adaptive calibration logic, the segment boundaries are highly aligned with students' cognitive patterns, improving the cognitive logical integrity of the segments and aiding learners in understanding and memorization. As dynamic quantitative indicators, graph entropy value and cognitive load can adapt to the characteristics of videos from different subjects and teaching styles, overcoming the bottleneck of poor cross-scene adaptability of traditional fixed threshold methods and ensuring the stability of positioning accuracy. No manual intervention is required throughout the process, resulting in outstanding positioning efficiency and consistency, meeting the needs of large-scale batch processing of teaching videos. Simultaneously, it provides structured basic units for subsequent key segment teaching value scoring and redundancy optimization, improving segment selection accuracy, balancing technical precision and teaching practicality, and enabling rapid model adaptation and continuous evolution through a closed-loop system incorporating user feedback. Compared to existing technologies, this solution breaks through the limitations of single-dimensional positioning by capturing implicit knowledge points through the collaboration of two methods, achieving full coverage of both explicit core and implicit key points. It adaptively calibrates the time span to accommodate the varying degrees of compactness in knowledge point explanations, avoiding excessive splitting or boundary redundancy. A dual verification method combining graph entropy value logical filtering and cognitive load saliency screening enhances anti-interference capabilities in complex teaching scenarios and ensures robust positioning across different scenarios. It accurately reconstructs the teaching logic chain, avoiding logical fragmentation between knowledge points and improving the accuracy of logical correlation recognition in subsequent teaching value scoring. Under the premise of full automation, it achieves a dual breakthrough in positioning efficiency and teaching practicality, meeting the dual needs of large-scale processing and personalized teaching applications.

[0051] This invention proposes a deep learning-based system for extracting key knowledge point segments from teaching videos, such as... Figure 2 As shown, the teaching video key knowledge point segment extraction system includes: The database construction and deconstruction module is used to obtain the teaching scenario metadata database and multi-source data deconstruction results corresponding to the teaching videos by annotating and deconstructing the teaching videos. The knowledge flow dynamic tracking and evolution graph construction module is used to construct a knowledge flow dynamic tracking and evolution graph using the teaching scenario metadata database and multi-source data deconstruction results; The cognitive load curve acquisition module is used to obtain the cognitive load curve corresponding to the teaching video by dynamically tracking and evolving the knowledge flow graph. The Key Knowledge Point Segment Boundary Determination Module is used to locate the boundaries of key knowledge point segments in teaching videos by combining the cognitive load curve with the dynamic tracking and evolution map of knowledge flow and the corresponding region of sudden increase in map entropy value. The key knowledge point fragment extraction module is used to extract key knowledge point fragments based on their boundaries, thereby obtaining multiple key knowledge point fragments.

[0052] The working principle of the above technical solution is as follows: This embodiment takes the simulation of the teaching and cognitive process as the core, and achieves accurate extraction of key knowledge point fragments through multi-dimensional data processing and dynamic modeling: First, the teaching video is annotated with scenes and deconstructed into multi-modal data to construct a basic dataset containing teaching behavior features and multi-source cognitive dimension data; based on this dataset, the smallest knowledge unit (knowledge atom) is identified, and a knowledge evolution graph is generated through temporal correlation modeling and dynamic update mechanism to capture the logical relationship and evolution law of knowledge points; the content features and multi-modal expression features of the knowledge evolution graph and student cognitive feedback data are integrated to quantify and generate a high temporal resolution cognitive load curve; the dense range of knowledge points is located by the region of sudden increase in graph entropy value, and the initial location, multi-dimensional expansion calibration and closed-loop verification of the key fragment boundary are completed by combining the waveform features of the cognitive load curve; finally, the key knowledge point fragment is extracted according to the calibrated boundary, and the model parameters can be dynamically optimized through user feedback to continuously improve the extraction performance.

[0053] The above technical solution effectively improves the accuracy of key knowledge point identification and solves the problems of blurred segment boundaries and logical breaks in existing technologies. It avoids the omission of core knowledge or the inclusion of redundant content, thereby effectively improving the matching degree between the extraction results and the course syllabus. Through cognitive load quantification, it adapts to students' cognitive patterns, ensuring that the segment content not only aligns with teaching priorities but also matches students' cognitive abilities, while meeting the practical needs of teachers in lesson preparation, micro-lesson production, and other teaching scenarios. It effectively reduces fluctuations in the extraction accuracy of knowledge point segments in teaching videos across different disciplines, including arts, sciences, engineering, and experimental subjects.

[0054] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A method for extracting key knowledge point segments from teaching videos based on deep learning, characterized in that, The method for extracting key knowledge point segments from teaching videos includes: By annotating and deconstructing the teaching videos, we can obtain the teaching scenario metadata database and multi-source data deconstruction results corresponding to the teaching videos. A dynamic knowledge flow tracking and evolution graph is constructed using the teaching scenario metadata database and the results of multi-source data deconstruction. The cognitive load curves corresponding to the teaching videos are obtained by using knowledge flow dynamic tracking and evolution graphs; By combining the cognitive load curve with the dynamic tracking and evolution map of knowledge flow, the boundary of key knowledge point segments in the teaching video is located to determine the boundary of key knowledge point segments. Based on the boundaries of key knowledge point segments, extract multiple key knowledge point segments.

2. The method for extracting key knowledge point segments from teaching videos according to claim 1, characterized in that, By annotating and deconstructing the teaching videos, we obtain the teaching scenario metadata database and multi-source data deconstruction results corresponding to the teaching videos, including: The teaching videos are labeled to obtain various types of tags, and a teaching scenario metadata database is constructed by combining these tags with the spatiotemporal features of the teaching videos. The various types of tags include knowledge type tags, interaction intensity tags, and cognitive load tags. A spatiotemporal attention decomposition method is used to decompose the teaching video into video frames, and the information regions corresponding to the teaching video are obtained from the video frames after decomposition; wherein, the information regions include knowledge carrier regions, teaching behavior regions, and background interference regions. Feature extraction is performed on the audio stream corresponding to the teaching video to obtain the key and emphasized sentence fragments corresponding to the teaching video; Text stream extraction is performed on the blackboard writing and PPT content of the teaching video. The text stream of the teaching video is parsed using a subject-specific large model to obtain the logical relationship tree contained in the text stream of the teaching video. The information regions, emphasized sentence fragments, and logical relationship trees are the results of multi-source data deconstruction.

3. The method for extracting key knowledge point segments from teaching videos according to claim 1, characterized in that, Using the aforementioned teaching scenario metadata database and multi-source data deconstruction results, a dynamic knowledge flow tracking and evolution graph is constructed, including: Based on the various tags contained in the teaching scenario metadata database and the deconstruction results of multi-source data, knowledge atoms are extracted from the teaching content corresponding to the teaching videos, and the knowledge atom sequence K=[k1, k2, ..., k t ], where k1, k2, ..., k t These represent the knowledge atoms corresponding to time t; A temporal association matrix R is generated by the semantic similarity and time decay coefficient between every two knowledge atoms, wherein the temporal association matrix R contains matrix elements R0. ij This refers to the logical association strength between the i-th node and the j-th node; The knowledge atoms are used to construct a knowledge flow dynamic tracking and evolution graph G(t), and the knowledge flow dynamic tracking and evolution graph G(t) is dynamically updated according to the update strategy.

4. The method for extracting key knowledge point segments from teaching videos according to claim 3, characterized in that, Constructing a dynamic knowledge flow tracking and evolution graph G(t) using the aforementioned knowledge atoms includes: The logical association strength between every two knowledge atoms is retrieved, and the logical association strength is compared with a preset first association strength threshold; wherein the preset first association strength threshold is set to 0.12-0.

31. Select knowledge atoms whose logical association strength is not lower than a preset first association strength threshold as nodes; The logical association strength R between every two nodes ij As an edge; Using the node and the logical association strength R between the two nodes ij Generate a dynamic tracking and evolution map of knowledge flow.

5. The method for extracting key knowledge point segments from teaching videos according to claim 3, characterized in that, The update strategy includes: The nodes contained in the knowledge flow dynamic tracking and evolution graph constructed from knowledge atoms are retrieved, wherein the nodes are the knowledge elements contained in the knowledge flow dynamic tracking and evolution graph. Retrieve the logical association strength R of every two nodes for self-checking ij ; The logical association strength R between every two nodes ij and the logical association strength R ij The comparison is made with a preset second association strength threshold, wherein the preset second association strength threshold is between 0.25 and 0.42; Extract logical association strength R ij Two nodes with a correlation strength not lower than a preset second correlation strength threshold are considered as a node pair; Retrieve the time interval influence factors corresponding to the node groups; use the time interval influence factors of all node groups to determine the current dynamic update time interval T of the knowledge flow dynamic tracking and evolution graph; The current knowledge flow dynamic tracking and evolution graph is updated based on the dynamic update time interval T.

6. The method for extracting key knowledge point segments from teaching videos according to claim 1, characterized in that, The cognitive load curves corresponding to instructional videos are obtained using knowledge flow dynamic tracking and evolutionary mapping, including: The content complexity parameter G1(t) of the current knowledge flow dynamic tracking and evolution graph is obtained by using the node density, the sum of node weights and the numerical logical correlation strength. The teaching expression complexity parameter G2(t) is obtained by using the teacher's actions corresponding to the current teaching video, the volume intensity of the key sentence segments, and the logical relationship tree in the text stream; Retrieve the weight coefficients corresponding to the content complexity parameter G1(t) and the teaching expression complexity parameter G2(t), and use the content complexity parameter G1(t) and the teaching expression complexity parameter G2(t) in combination with their corresponding weight coefficients to obtain the cognitive load value L(t) through a weighted summation method; By combining the cognitive load value L(t) with the t moments contained in the corresponding duration of the instructional video, cognitive load curves for all instructional videos at all moments t are generated.

7. The method for extracting key knowledge point segments from teaching videos according to claim 6, characterized in that, The content complexity parameter G1(t) of the current knowledge flow dynamic tracking and evolution graph is obtained by utilizing the node density, the sum of node weights, and the numerical logical correlation strength. This includes: Statistically calculate the node density and total node weight of the current knowledge flow dynamic tracking and evolution graph; Retrieve the logical association strengths of all nodes and edges contained in the current knowledge flow dynamic tracking and evolution graph; The average logical association strength is obtained by utilizing the logical association strength of all nodes and edges contained in the current knowledge flow dynamic tracking and evolution graph. The content complexity parameter G1(t) corresponding to the knowledge flow dynamic tracking and evolution graph at the current moment is obtained by combining the node density, the sum of node weights and the average value of logical association strength with the parameter weights corresponding to the node density, the sum of node weights and the average value of logical association strength through a weighted summation method.

8. The method for extracting key knowledge point segments from teaching videos according to claim 6, characterized in that, The teaching expression complexity parameter G2(t) is obtained by utilizing the teacher's actions corresponding to the current teaching video, the volume intensity of key emphasized sentence fragments, and the logical relationship tree in the text stream. This includes: Retrieve the skeleton key points of the teacher's teaching process from the teaching behavior area contained in the information area of ​​the current teaching video, and use the skeleton key point sequence to obtain the motion entropy H(t); The volume intensity of the key sentences and phrases emphasized in the current teaching video is used to obtain the standard deviation of the volume intensity B(t). Retrieve the logical relation tree density M(t) from the text stream corresponding to the current teaching video; The motion entropy H(t), volume intensity standard deviation B(t), and logical relation tree density M(t) are normalized to obtain the normalized motion entropy H. g (t), Standard deviation of volume intensity B g (t) and the density of the logical relation tree M g (t); The motion entropy H after normalization is used g (t), Standard deviation of volume intensity B g (t) and the density of the logical relation tree M g (t) Obtain the teaching expression complexity parameter G2(t).

9. The method for extracting key knowledge point segments from teaching videos according to claim 1, characterized in that, By combining cognitive load curves with dynamic tracking of knowledge flow and corresponding regions of sudden increases in entropy in the evolutionary graph, the boundaries of key knowledge point segments in teaching videos are located, determining the boundaries of key knowledge point segments, including: Retrieve the dynamic tracking and evolution graph of the knowledge flow, and determine the graph entropy sequence corresponding to the dynamic tracking and evolution graph of the knowledge flow through the nodes and node weights contained in the dynamic tracking and evolution graph of the knowledge flow; The difference between the spectral entropy values ​​corresponding to each two adjacent spectral entropy value sequences contained in the spectral entropy value sequence is obtained by performing difference processing on the spectral entropy values. The spectrum entropy difference is compared with a preset difference threshold, and the two time points corresponding to the two spectrum entropy value sequences with a spectrum entropy difference not lower than the preset difference threshold are taken as spectrum entropy value surge point groups, thereby obtaining multiple spectrum entropy value surge point groups. Based on the regions of sudden increases in spectral entropy values ​​formed by multiple groups of sudden increases in spectral entropy values, a segment of the cognitive load curve corresponding to the time point of the group of sudden increases in spectral entropy values ​​is extracted from the cognitive load curve. The maximum value of cognitive load appearing in each segment of the cognitive load curve is taken as the target cognitive load value; Extract the time span between the times corresponding to each two adjacent target cognitive load values, and compare the time span with a preset time span threshold, wherein the preset time span threshold ranges from 10s to 15s. If the time span between the times corresponding to two adjacent target cognitive load values ​​does not exceed the preset time span threshold, then the times corresponding to the two adjacent target cognitive load values ​​will be used as the start time point and the end time point, and the start time point and the end time point will be used as the boundary of the key knowledge point segment. If the time span between two adjacent target cognitive load values ​​exceeds a preset time span threshold, then the time corresponding to each target cognitive load value is taken as the time center, and the time points 7 seconds before and after the time center are taken as the start and end time points, and the start and end time points are taken as the boundaries of key knowledge point segments.

10. A system for extracting key knowledge points from teaching videos based on deep learning, characterized in that, The system for extracting key knowledge points from teaching videos includes: The database construction and deconstruction module is used to obtain the teaching scenario metadata database and multi-source data deconstruction results corresponding to the teaching videos by annotating and deconstructing the teaching videos. The knowledge flow dynamic tracking and evolution graph construction module is used to construct a knowledge flow dynamic tracking and evolution graph using the teaching scenario metadata database and multi-source data deconstruction results; The cognitive load curve acquisition module is used to obtain the cognitive load curve corresponding to the teaching video by dynamically tracking and evolving the knowledge flow graph. The Key Knowledge Point Segment Boundary Determination Module is used to locate the boundaries of key knowledge point segments in teaching videos by combining the cognitive load curve with the dynamic tracking and evolution map of knowledge flow and the corresponding region of sudden increase in map entropy value. The key knowledge point fragment extraction module is used to extract key knowledge point fragments based on their boundaries, thereby obtaining multiple key knowledge point fragments.

Images