Science popularization content intelligent recommendation method and system based on deep learning

Abstract

The invention provides a science popularization content intelligent recommendation method and system based on deep learning. The method comprises the following steps: acquiring science popularization image / video data, performing space-time decoupling processing to generate a multi-scale space-time feature tensor, and constructing multi-level feature representation of a local scene, a dynamic process and an overall context; utilizing a multi-head cross-modal attention network to calculate a semantic association degree matrix of the features and knowledge graph entities; a time sequence causal chain extraction module is introduced to generate a causal path constraint vector, a dynamic semantic alignment constraint is formed through matching with a knowledge graph relation path, and a semantic incidence matrix is optimized; and generating a uniform space-time semantic fusion vector by adopting an adaptive mapping decoder, and outputting a structured semantic tag set through a deep semantic classifier. According to the method, the limitation of traditional shallow recommendation is broken through, understanding of deep semantics and causal logic of scientific contents is realized, and the accuracy, semantic depth and interpretability of science popularization content recommendation are remarkably improved.

Description

Technical Field

[0001] This invention relates to the field of science popularization text generation technology, and in particular to a method and system for intelligent recommendation of science popularization content based on deep learning. Background Technology

[0002] With the increasingly diversified presentation of science popularization content and the widespread dissemination of multimodal content such as images and videos on online platforms, how to achieve accurate recommendations of scientific knowledge structures has become a core technical challenge for intelligent science popularization content service systems. Currently, the mainstream technologies of science popularization content recommendation systems mostly rely on shallow feature extraction and traditional classification methods, focusing primarily on text information analysis and tag association, making it difficult to capture the deep semantic relationships between multimodal content such as videos and images and scientific knowledge.

[0003] In the field of multimodal content understanding and recommendation, existing technologies mainly include deep visual feature analysis methods based on convolutional neural networks (CNN), three-dimensional convolutional networks (3D-CNN), and temporal convolutional networks (TCN), as well as knowledge graph embedding (such as TransE) and graph neural networks (GNN, GAT) and other knowledge structured modeling methods. However, these solutions are mostly limited to single-modal processing, or only simplify and statically process cross-modal alignment, and cannot achieve accurate integration of multi-scale, spatiotemporal dynamics and scientific knowledge systems.

[0004] Therefore, it is necessary to improve the way popular science texts are generated in order to overcome the shortcomings of existing technologies. Summary of the Invention

[0005] To overcome the problems existing in related technologies, one of the objectives of this invention is to provide a deep learning-based intelligent recommendation method for popular science content. This method can achieve multi-intent intelligent understanding and high-precision response in the automatic generation process of popular science text, thereby improving the accuracy and logic of the generated text.

[0006] A deep learning-based intelligent recommendation method for popular science content includes: Acquire image / video data from popular science content and record its metadata information; Spatiotemporal decoupling processing is performed on image / video data, and a three-dimensional convolutional neural network is used to extract local spatiotemporal semantic units to generate multi-scale spatiotemporal feature tensors; Based on multi-scale spatiotemporal feature tensors, multi-level feature representations of local scene features, dynamic process features, and overall context features are constructed respectively, and semantic labels of features at each scale are recorded. The multi-level feature representations are input into a multi-head cross-modal attention network to calculate the semantic association matrix between each feature scale and entity nodes in the scientific knowledge graph. Based on the semantic relevance matrix, a temporal causal chain extraction module is introduced to identify the causal order of scientific events or processes in image / video clips and generate causal path constraint vectors. The causal path constraint vector is matched with the knowledge graph relationship path to form dynamic semantic alignment constraints and update the semantic relevance matrix. An adaptive mapping decoder is used to perform weighted fusion of the updated semantic relevance matrix based on the task scenario type, generating a unified spatiotemporal semantic fusion vector; The spatiotemporal semantic fusion vector is input into a deep semantic classifier to identify the subject affiliation, knowledge depth, and scientific principles of image / video content, and generate a set of structured semantic tags. The structured semantic tag set is written into the knowledge tag database of popular science content, and provides semantic matching basis for the user interest matching module of the subsequent recommendation system.

[0007] In a preferred embodiment of the present invention, the step of acquiring image / video data from popular science content and recording its metadata includes: Collect raw image / video streams from science popularization content platforms, obtain complete visual data input, and record their original format, resolution, and encoding method; Based on video parsing technology, the collected image / video data is decomposed at the frame level, and each frame image and its corresponding timestamp information are extracted to construct a time-series visual dataset. The content of each frame of image is classified using scene recognition algorithms to identify the shooting scene category and generate a structured set of scene labels. Based on the content classification model, semantic analysis is performed on the image / video as a whole to identify its content category tag, and the tag is initially mapped to the popular science topic node in the knowledge graph; The source platform information of images / videos is extracted in a structured manner, and a platform identifier is generated as an auxiliary context feature for subsequent multimodal feature fusion.

[0008] In a preferred embodiment of the present invention, the step of performing spatiotemporal decoupling processing on image / video data and extracting local spatiotemporal semantic units using a three-dimensional convolutional neural network to generate multi-scale spatiotemporal feature tensors includes: The acquired image / video data is subjected to spatiotemporal decoupling processing. A three-dimensional convolutional neural network is used to perform sliding window sampling on the continuous video frame sequence to extract local spatiotemporal semantic units and generate the original spatiotemporal feature tensor. Based on the original spatiotemporal feature tensor, multi-scale convolutional kernels are used for parallel processing to extract short-term dynamic features, mid-term process features and long-term context features, respectively, in order to construct a multi-scale spatiotemporal feature representation; The multi-scale spatiotemporal feature representation is normalized, and the Batch Normalization method is used to standardize the feature channels at each scale. Based on the normalized multi-scale spatiotemporal features, a feature pyramid network is used to fuse cross-scale information and generate an enhanced multi-scale spatiotemporal feature tensor. The enhanced multi-scale spatiotemporal feature tensor is input into the temporal gating module, and a temporal convolutional network is used to perform temporal modeling on it, outputting a structured spatiotemporal semantic feature tensor.

[0009] In a preferred embodiment of the present invention, the construction of multi-level feature representations based on multi-scale spatiotemporal feature tensors, including local scene features, dynamic process features, and overall context features, and the recording of semantic labels for each scale feature, includes: The multi-scale spatiotemporal feature tensor is subjected to hierarchical decomposition to extract local scene feature tensors, dynamic process feature tensors and overall context feature tensors, forming a multi-level feature representation structure; Based on the local scene feature tensor, a spatial attention mechanism is used to perform semantic enhancement processing on local regions of the image, generating local scene semantic feature vectors, and mapping them to a predefined local semantic label space to obtain a set of local scene semantic labels. Based on the dynamic process feature tensor, a temporal convolutional network is used to model the dynamic changes in video segments, extract the semantic feature vector of the dynamic process, and generate a set of semantic labels of the dynamic process through a temporal semantic classifier. Based on the overall context feature tensor, global average pooling is used to compress the spatiotemporal features to generate an overall semantic feature vector, and a set of overall context semantic labels is generated through a fully connected classification network. The set of semantic tags for local scenes, the set of semantic tags for dynamic processes, and the set of semantic tags for the overall context are structurally integrated to generate a unified set of multi-level feature semantic tags and record their semantic hierarchy relationships.

[0010] In a preferred embodiment of the present invention, the step of inputting the multi-level feature representations into a multi-head cross-modal attention network and calculating the semantic association matrix between each feature scale and entity nodes in the scientific knowledge graph includes: The multi-level feature representation is subjected to multi-head projection processing to generate multiple sets of feature query vectors, key vectors and value vectors; Based on the multiple sets of query vectors, key vectors and value vectors, the initial semantic similarity between features at each scale and entity nodes in the scientific knowledge graph is calculated using the dot product attention mechanism, and an initial semantic association matrix is ​​constructed. The initial semantic association matrix is ​​subjected to Softmax normalization to generate a normalized multi-scale semantic association matrix; Based on the normalized multi-scale semantic association matrix, a multi-head attention fusion mechanism is used to weight and aggregate it to generate a unified cross-modal semantic association matrix. The unified cross-modal semantic association matrix is ​​input into the knowledge perception alignment module. Combining the structural information of the knowledge graph with the semantic embedding vector, a graph-based semantic path enhancement operation is performed to optimize the structural consistency of the semantic association matrix.

[0011] In a preferred embodiment of the present invention, the step of introducing a temporal causal chain extraction module based on a semantic correlation matrix to identify the causal order of scientific events or processes in image / video clips and generate a causal path constraint vector includes: The semantic association matrix is ​​preprocessed by temporal causal modeling, and a causal attention mechanism is used to model the temporal dependencies in the semantic association matrix to extract potential causal inference paths. Based on the causal attention output, a causal inference graph network is used to model the event evolution relationship between consecutive frames in an image / video clip, identify the causal dependency structure of scientific events or processes, and generate a preliminary set of causal paths. A knowledge graph semantic path matching operation is performed on the preliminary causal path set. A graph embedding matching algorithm is used to align the causal paths with the relational paths in the scientific knowledge graph to generate a structured causal matching path set. Based on the structured causal matching path set, a causal strength assessment model is introduced, and a path weight aggregation algorithm is used to calculate the confidence score of each causal path to generate a causal path constraint weight vector. The causal path constraint weight vector is fused with the semantic relevance matrix, and a causal enhanced semantic relevance matrix is ​​generated through a weighted fusion operation.

[0012] In a preferred embodiment of the present invention, the step of matching the causal path constraint vector with the knowledge graph relationship path to form dynamic semantic alignment constraints and updating the semantic relevance matrix includes: A knowledge graph relation path retrieval operation is performed on the causal path constraint vector. Based on the graph embedding similarity calculation method, the structured relation path that best matches the current causal path is searched in the scientific knowledge graph to obtain a set of candidate knowledge paths. Based on the candidate knowledge path set, a graph attention network is used to model the semantic similarity between causal paths and knowledge graph paths to generate path-level semantic matching weight vectors. The path-level semantic matching weight vector and the causal path constraint vector are fused together, and a weighted fusion algorithm is used to generate a joint path constraint vector. Based on the joint path constraint vector, a dynamic semantic alignment constraint modeling operation is performed, and a graph neural network is used to structurally encode the entity-relation path in the knowledge graph to generate a structured dynamic alignment constraint vector. The structured dynamic alignment constraint vector is input into the semantic consistency optimization module, and the constraint propagation algorithm is used to perform weighted update processing on the multi-scale semantic relevance matrix to generate a causal enhanced semantic relevance matrix.

[0013] In a preferred embodiment of the present invention, the step of using an adaptive mapping decoder to perform weighted fusion of the updated semantic relevance matrix based on the task scenario type specifically includes: The updated semantic association matrix is ​​processed for task scenario classification to generate task scenario classification vectors. Based on the task scenario classification vector, a learnable task-aware attention mechanism is used to weight the semantic relevance components at each scale to generate a task-driven multi-scale attention weight matrix. The task-driven multi-scale attention weight matrix and the semantic association matrix are multiplied element-wise to generate a weighted fused multi-scale semantic association feature tensor. Spatiotemporal semantic alignment optimization processing is performed on the multi-scale semantic association feature tensor. A graph attention network combined with the structural information of the knowledge graph is used to perform consistency correction on its semantic path, generating a structure-optimized spatiotemporal semantic alignment feature vector. The spatiotemporal semantic alignment feature vector is input into the adaptive decoder module, and a multilayer perceptron is used to perform nonlinear transformation processing on it to generate a unified spatiotemporal semantic fusion vector.

[0014] In a preferred embodiment of the present invention, the step of inputting the spatiotemporal semantic fusion vector into a deep semantic classifier to identify the subject affiliation, knowledge depth, and scientific principles of image / video content, and generating a structured semantic tag set, includes: The spatiotemporal semantic fusion vector is input into a multi-task deep semantic classifier, and the fusion vector is subjected to nonlinear transformation based on a pre-trained fully connected neural network to extract high-order semantic abstract features and generate a semantic abstract feature vector. Based on the semantic abstract feature vector, a multi-branch classification head structure is used to perform subject classification tasks, knowledge depth assessment tasks, and scientific principle recognition tasks respectively, generating subject classification results, knowledge depth levels, and scientific principle matching lists. The subject classification results are processed by knowledge graph topic node mapping, and the classification results are semantically aligned with the subject topic nodes in the knowledge graph using the graph embedding similarity calculation method to generate a set of structured subject affiliation tags. Based on the knowledge depth level and scientific principle matching list, and combined with the hierarchical structure information of the corresponding scientific principle in the knowledge graph, knowledge granularity alignment and hierarchical mapping operations are performed to generate a fine-grained semantic tag set composed of knowledge depth tags and scientific principle explanation tags. The set of structured subject affiliation tags and the set of fine-grained semantic tags are semantically fused and format-standardized, and then serialized and encapsulated using the JSON-LD semantic data structure to generate the set of structured semantic tags.

[0015] 10. A deep learning-based intelligent recommendation system for popular science content, characterized in that: it is used to implement the deep learning-based intelligent recommendation method for popular science content as described in any one of claims 1-9.

[0016] In a preferred embodiment of the present invention, after text generation, the method further includes: The text generation results are dynamically monitored and optimized, including content consistency verification and context overlap assessment. If the feature deviation exceeds the set threshold, the model parameters are automatically updated.

[0017] The second objective of this invention is to provide a natural language processing-based intelligent generation system for popular science texts, which implements the deep learning-based intelligent recommendation method for popular science content as described above.

[0018] The beneficial effects of this invention are as follows: This invention provides a deep learning-based intelligent recommendation method and system for popular science content. The method acquires popular science image / video data, performs spatiotemporal decoupling processing to generate multi-scale spatiotemporal feature tensors, and constructs multi-level feature representations of local scenes, dynamic processes, and overall context. It utilizes a multi-head cross-modal attention network to calculate the semantic association matrix between features and knowledge graph entities. A temporal causal chain extraction module is introduced to generate causal path constraint vectors, which are then matched with knowledge graph relationship paths to form dynamic semantic alignment constraints, optimizing the semantic association matrix. An adaptive mapping decoder generates a unified spatiotemporal semantic fusion vector, which is then output as a structured semantic tag set by a deep semantic classifier. Through "multi-scale spatiotemporal feature tensors" and "multi-level feature representations (local, dynamic, and global)," the model can simultaneously capture static objects, dynamic processes (such as chemical reactions and celestial motion), and the overall narrative context in popular science videos. This avoids the semantic loss caused by traditional methods that rely solely on keyframes or global features. The resulting "structured semantic tag set" includes not only subject categories but also specific scientific principles and processes, providing the recommendation system with precise semantic basis and significantly improving recommendation accuracy and content relevance. The "temporal causal chain extraction module" analyzes the sequence and dependencies of events in a video, enabling it to recommend other content with similar scientific principles or causal logic to what the user has watched. This achieves recommendations based on deep knowledge logic, rather than just shallow topic tag similarity. Furthermore, the "adaptive mapping decoder" dynamically adjusts the semantic fusion weights according to different scenarios, maintaining high-precision semantic alignment even in an environment where scientific knowledge is constantly updated, ensuring the long-term effectiveness of the recommendation system. Attached Figure Description

[0019] Figure 1 This is a flowchart of a deep learning-based intelligent recommendation method for popular science content, provided in an embodiment of the present invention. Detailed Implementation

[0020] Preferred embodiments of the invention will now be described in more detail with reference to the accompanying drawings. While preferred embodiments of the invention are shown in the drawings, it should be understood that the invention can be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that the invention will be thorough and complete, and will fully convey the scope of the invention to those skilled in the art.

[0021] Example 1 like Figure 1 As shown in the attached figure, this embodiment provides... Figure 1 As shown, this application provides a method for intelligent recommendation of popular science content based on deep learning, specifically including: S1: Acquire image / video data from popular science content and record its metadata information; S2: Perform spatiotemporal decoupling processing on the image / video data, and use a three-dimensional convolutional neural network to extract local spatiotemporal semantic units to generate multi-scale spatiotemporal feature tensors; S3: Based on the multi-scale spatiotemporal feature tensor, construct multi-level feature representations of local scene features, dynamic process features and overall context features respectively, and record the semantic labels of features at each scale. S4: Input the multi-level feature representations into a multi-head cross-modal attention network and calculate the semantic association matrix between each feature scale and entity nodes in the scientific knowledge graph; S5: Based on the semantic correlation matrix, a temporal causal chain extraction module is introduced to identify the causal order of scientific events or processes in image / video clips and generate causal path constraint vectors. S6: Match the causal path constraint vector with the knowledge graph relationship path to form dynamic semantic alignment constraints, and update the semantic association matrix; S7: An adaptive mapping decoder is used to perform weighted fusion of the updated semantic association matrix based on the task scenario type to generate a unified spatiotemporal semantic fusion vector; S8: Input the spatiotemporal semantic fusion vector into a deep semantic classifier to identify the subject affiliation, knowledge depth, and scientific principles of the image / video content, and generate a set of structured semantic tags; S9: Write the structured semantic tag set into the knowledge tag database of popular science content, and provide semantic matching basis for the user interest matching module of the subsequent recommendation system.

[0022] Specifically, step S1 of this application includes: Raw image / video streams from science popularization content platforms are collected to obtain complete visual data input, and their original format, resolution, and encoding method are recorded.

[0023] Based on video parsing technology, the acquired image / video data is decomposed at the frame level, and each frame image and its corresponding timestamp information are extracted to construct a time-series visual dataset.

[0024] The content of each frame of image is classified using scene recognition algorithms to identify the shooting scene category and generate a structured set of scene labels.

[0025] Based on the content classification model, semantic analysis is performed on the image / video as a whole to identify its content category tag, and the tag is initially mapped to the popular science topic node in the knowledge graph.

[0026] Semantic analysis is performed on the images / videos as a whole based on a content classification model to identify their content category tags. These tags are then preliminarily mapped to science popularization topic nodes in the knowledge graph. This sub-step is a crucial step in the image / video data acquisition and metadata recording process, aligning content semantic classification with the knowledge graph. The input data consists of the overall image / video data after frame-level decomposition and scene classification, and its structure is a multi-dimensional visual dataset containing timestamps, resolution, encoding format, and scene tags.

[0027] A multi-layer convolutional neural network (CNN) is used to perform content classification on images / videos. The input data is a normalized sequence of RGB image frames. The CNN model structure contains 5 convolutional layers and 3 fully connected layers, with kernel sizes of 3×3, 5×5, and 7×7, respectively. The activation function is ReLU, and the loss function is cross-entropy loss. in, For the first The true label of the class, To predict probabilities for the model, This represents the total number of categories. The model performs semantic content classification on the entire image / video, outputting a content category probability distribution vector. ,in Indicates that the image / video belongs to the first... Confidence level of the category.

[0028] Furthermore, the probability distribution vector is processed using the maximum probability selection algorithm. The process involves selecting the category corresponding to the highest probability value as the content category label for the image / video. ,Right now: This tag represents text-based science content categories, such as "astronomy," "biology," and "physics."

[0029] Furthermore, based on Knowledge Graph Embedding technology, content category tags are... This is mapped to the topic node space in the knowledge graph. The TransE model is used to embed the knowledge graph, and the entity embedding vector is... The relation embedding vector is ,satisfy: in, For the head entity, For tail entities, This is a relation type. The category label is calculated. Semantic similarity with each topic node in the knowledge graph: The topic node with the highest similarity is selected as the corresponding knowledge graph node for this content category to generate preliminary mapping results. .

[0030] Furthermore, by combining the hierarchical structure information of the knowledge graph, the preliminary mapping results are analyzed. Perform a hierarchical consistency check. If Top-level nodes in the knowledge graph (such as "Natural Sciences") utilize a subclass expansion mechanism, combining local semantic tag sets of image / video content. With dynamic process semantic tag set The matching degree of sub-class nodes is calculated using a weighted voting algorithm: in , , Finally, the subclass node with the highest matching degree is selected as the corrected mapping result. .

[0031] Through the multi-stage content classification and knowledge graph mapping process described above, a structured set of mapping relationships between content category tags and knowledge graph topic nodes is generated. This is used as an auxiliary semantic input for subsequent multimodal feature fusion and semantic correlation calculation, thereby improving the semantic alignment accuracy between image / video content and scientific knowledge nodes.

[0032] For example, suppose the input image / video is a science video explaining the formation of black holes, and its content category is identified as "astronomy" by the CNN model, with the probability distribution as follows: The corresponding category is "Astronomy". Using a knowledge graph embedding model, this category is mapped to the node "Astrophysics" in the knowledge graph, and its embedding vector is... Further, the hierarchical consistency is checked by combining the local semantic labels "galaxy" and "gravitational collapse" with the dynamic process label "stellar evolution". Finally, the mapping is corrected to the "black hole formation mechanism" node, forming the final mapping result for use by subsequent modules.

[0033] The source platform information of images / videos is extracted in a structured manner to identify their publishing platform attributes, and a platform identifier is generated as an auxiliary context feature for subsequent multimodal feature fusion.

[0034] Step S2 specifically includes: The acquired image / video data is subjected to spatiotemporal decoupling processing. A three-dimensional convolutional neural network is used to perform sliding window sampling on the continuous video frame sequence to extract local spatiotemporal semantic units and generate the original spatiotemporal feature tensor.

[0035] This step plays a crucial role in the multi-scale spatiotemporal feature extraction process, performing spatiotemporal decoupling and local semantic unit modeling. Its function is to decouple the acquired image / video data in terms of spatiotemporal dimensions. It employs a 3D convolutional neural network (3D-CNN) to perform sliding window sampling on continuous video frame sequences, extracting local spatiotemporal semantic units and generating the original spatiotemporal feature tensor. The input data for this step is the time-series visual dataset processed in step S1. Each frame Indicates the timestamp The following RGB image frames, , For image resolution, For the number of channels, This represents the total number of consecutive video frames. The input data also includes structured metadata information, such as timestamp sequences. Scene tag collection Content category tags With source platform identifier .

[0036] A three-dimensional convolutional neural network is used to model the spatiotemporal decoupling of video frame sequences. The core of this approach lies in capturing local spatiotemporal semantic units through a sliding window mechanism. Specifically, a sliding window function is employed. For video frame sequences Perform local sampling, where the window length is... Sliding step size Then the first The frame sequence within each window is , contains Frames, total number of windows .

[0037] Furthermore, through 3D convolution operations For the frame sequence within each window Local spatiotemporal feature extraction is performed. The 3D convolution kernel size is... The number of channels is 64, the stride is 1, and the ReLU activation function is used. The 3D convolution operation is defined as follows: in, For the first Layer convolution kernel weights, For bias terms, Represents a three-dimensional convolution operation. For the first Layer input features, For the first Layer output features.

[0038] Furthermore, the 3D convolutional output features of all windows are concatenated into a complete spatiotemporal feature tensor. Each of them Feature representations of local spatiotemporal semantic units. A temporal attention mechanism is employed to enhance the temporal continuity of local semantic units. spatiotemporal feature tensor Weighted fusion is performed to generate the original spatiotemporal feature tensor. The calculation process is as follows: in, This is a learnable attention query vector. For the first Attention weights for each window, This is the original spatiotemporal feature tensor of the final output.

[0039] By combining the aforementioned 3D convolution and temporal attention mechanisms, the original video frame sequence is transformed into a raw spatiotemporal feature tensor with local spatiotemporal semantic information, providing high-quality input data for subsequent multi-scale feature extraction and cross-modal semantic alignment. The output data of this step is the raw spatiotemporal feature tensor. Its technical effect lies in realizing spatiotemporal decoupling modeling of video content, enhancing the expressive power of local dynamic semantic units, and improving the accuracy and robustness of subsequent multi-scale feature extraction.

[0040] For example, suppose the input video is a popular science video introducing quantum mechanics experiments, containing 64 frames, a resolution of 224×224, a frame rate of 30fps, a content category of "physics," and a platform identifier of "YouTube." The above-described combination of 3D convolution and temporal attention mechanism is used for processing: Sliding window settings: Window length Sliding step size A total of 15 windows were generated; 3D convolution processing: Each window uses a 3×3×3 convolution kernel with 64 channels, ReLU activation, and the output feature tensor dimension is 224×224×64; Time-weighted attention: Learning attention weights By fusing the outputs of 15 windows, the original spatiotemporal feature tensor is generated. The dimensions are 224×224×64×15; (1) Output results: The original spatiotemporal feature tensor is used as the input of step S2.2 for parallel processing of multi-scale convolution kernels.

[0041] This processing effectively extracts local dynamic semantic units such as "double-slit interference experiment" and "quantum superposition state" from the video, providing structured semantic input for subsequent construction of multi-scale spatiotemporal feature representation and cross-modal knowledge graph alignment, and improving the alignment accuracy between image / video content and scientific knowledge nodes.

[0042] Based on the original spatiotemporal feature tensor, multi-scale convolutional kernels are used for parallel processing to extract short-term dynamic features, mid-term process features, and long-term context features, respectively, in order to construct a multi-scale spatiotemporal feature representation.

[0043] This step plays a crucial role in the multi-scale spatiotemporal feature extraction process, undertaking the key tasks of scale awareness and dynamic semantic modeling. Its function is to extract short-term dynamic features, mid-term process features, and long-term contextual features from the original spatiotemporal feature tensor using parallel processing with multi-scale convolutional kernels, thereby constructing a multi-scale spatiotemporal feature representation. The input data for this step is the original spatiotemporal feature tensor output from step S2.1. ,in This indicates the total number of windows. The input data also includes structured metadata information such as content category tags. Platform identifier and local semantic tag set .

[0044] A multi-scale 3D convolutional kernel parallel processing method is adopted, with convolutional kernel parameters set for short-term, medium-term, and long-term scales respectively. Short-term dynamic feature extraction uses... A convolutional kernel with a stride of 1 and 32 channels is used to capture rapidly changing features in local frame sequences; mid-time process feature extraction employs... A convolutional kernel with a stride of 2 and 64 channels is used to model the dynamic evolution of video clips; long-term contextual feature extraction employs... A convolutional kernel with a stride of 3 and 128 channels is used to extract contextual semantic information across windows. The three convolutional kernel operations are defined as follows: Short-time dynamic feature convolution operation: in For short-time convolution kernel weights, For bias terms, For channel index, For spatial coordinates, For time indexing.

[0045] Mid-time process feature convolution operation: in For the convolution kernel weights in the middle time, This is a bias term.

[0046] Long-term contextual feature convolution operation: in For long-term convolution kernel weights, This is a bias term.

[0047] Furthermore, the features at the three scales are subjected to nonlinear transformation using the ReLU activation function: Furthermore, a channel attention mechanism is employed to weightedly fuse multi-scale features, generating a unified multi-scale spatiotemporal feature tensor. The channel attention weights are calculated as follows: in This is a learnable attention query vector. This is the channel projection matrix.

[0048] Furthermore, the channel attention weights are weighted and fused with multi-scale features to generate the final multi-scale spatiotemporal feature tensor. ,in The fused feature dimensions: By combining multi-scale convolution with channel attention, the original spatiotemporal feature tensor is transformed into a multi-scale spatiotemporal feature representation with multi-scale semantic information, providing high-quality input for subsequent normalization and cross-scale information fusion. The output data of this step is a multi-scale spatiotemporal feature tensor. Its technical effect lies in realizing multi-scale semantic modeling of video content, enhancing the expressive power of local dynamic semantic units at different time scales, and improving the accuracy and robustness of subsequent cross-modal semantic alignment and knowledge graph mapping.

[0049] For example, suppose the input video is a science video introducing the formation of black holes, containing 64 frames, a resolution of 224×224, a frame rate of 30fps, a content category of "astronomy," and a platform identifier of "Bilibili." The above-mentioned combination of multi-scale convolution and channel attention mechanism is used for processing: Multi-scale convolution processing: Short-time convolution kernel With 32 channels and a stride of 1, it extracts rapidly changing features (such as "gravitational collapse"); the convolutional kernel is used in the intermediate time interval. 64 channels, stride 2, modeling dynamic processes (such as "stellar evolution"); long-duration convolution kernel With 128 channels and a step size of 3, the contextual semantics (such as "event horizon formation") are extracted.

[0050] Nonlinear activation processing: The ReLU function is used to perform nonlinear transformation on the features at three scales to enhance the feature representation ability.

[0051] Channel attention weighting: learning attention weights By fusing features from three scales, a multi-scale spatiotemporal feature tensor is generated. The dimensions are 224×224×256×15.

[0052] Output: The multi-scale spatiotemporal feature tensor is used as input to step S2.3 for Batch Normalization.

[0053] This processing effectively extracts dynamic semantic units at different time scales from the video, such as "gravitational collapse," "stellar evolution," and "event horizon formation," providing structured semantic input for subsequent construction of multi-level feature representations and cross-modal knowledge graph alignment, thereby improving the alignment accuracy between image / video content and scientific knowledge nodes.

[0054] The multi-scale spatiotemporal feature representation is normalized by using the Batch Normalization method to standardize the feature channels at each scale, thereby eliminating the distribution differences between features at different scales and improving the stability of subsequent semantic modeling.

[0055] Based on the normalized multi-scale spatiotemporal features, a feature pyramid network (FPN) is used to fuse cross-scale information and generate an enhanced multi-scale spatiotemporal feature tensor to improve the expressive power of local semantic units.

[0056] The enhanced multi-scale spatiotemporal feature tensor is input into the temporal gating module, and a temporal convolutional network (TCN) is used to perform temporal modeling on it to capture the dynamic dependencies between local semantic units and output a structured spatiotemporal semantic feature tensor.

[0057] Step S3 specifically includes: The multi-scale spatiotemporal feature tensor is subjected to hierarchical decomposition to extract local scene feature tensors, dynamic process feature tensors, and overall context feature tensors, forming a multi-level feature representation structure with spatial locality, temporal dynamism, and global semantic consistency.

[0058] The multi-scale spatiotemporal feature extraction process plays a crucial role in feature standardization and distribution consistency modeling. Its function is to normalize the multi-scale spatiotemporal feature representation by employing Batch Normalization to standardize the feature channels at each scale, thereby eliminating distribution differences between features at different scales and improving the stability of subsequent semantic modeling. The input data for this step is the output multi-scale spatiotemporal feature tensor. The input data contains 256 features after fusion and 15 windows. The input data also includes structured metadata such as content category tags. Platform identifier and local semantic tag set .

[0059] Batch Normalization method is used for multi-scale spatiotemporal feature tensors Each channel in the algorithm is standardized by subtracting the channel mean and dividing by the standard deviation to ensure that the characteristics of each channel follow a standard normal distribution. Specifically, for the ... Feature channels Calculate its mean at all spatial and temporal locations. With variance : in, For the first The average of each channel, This represents the variance of the channel. For spatial coordinates, For time indexing.

[0060] Furthermore, through learnable scaling parameters With translation parameters A linear transformation is applied to the standardized features to restore the network's expressive power. The standardization operation is defined as follows: in, To prevent small constants from being divided by zero, and For learnable parameters, These are the standardized feature channels.

[0061] Furthermore, the standardized features of all channels are concatenated to generate a standardized multi-scale spatiotemporal feature tensor. Its dimension is consistent with the input tensor, but the feature distribution of each channel tends to be consistent, which improves the stability of subsequent cross-scale information fusion and semantic modeling.

[0062] Furthermore, a channel correlation constraint mechanism is introduced, which is applied to the standardized feature tensor. Calculate the Pearson correlation coefficient matrix between channels Furthermore, L2 regularization was employed to suppress high correlations between channels. in, For the first With the The correlation coefficient between channels and These are the average values ​​of the two channels, respectively.

[0063] Furthermore, through L2 regularization terms Constrain channel correlation: This loss term is introduced into the training process to optimize. and The parameters reduce the correlation between channels, making the feature representation more independent and robust.

[0064] Through the combined processing of multi-stage batch normalization and channel correlation constraints, the multi-scale spatiotemporal feature tensor is transformed into a standardized feature tensor with uniform distribution characteristics, providing high-quality input for subsequent cross-scale information fusion and semantic path enhancement. The output data of this step is the standardized multi-scale spatiotemporal feature tensor. Its technical effect lies in eliminating the distribution differences between features at different scales, improving the stability and generalization ability of subsequent semantic modeling, and enhancing the accuracy and robustness of cross-modal semantic alignment.

[0065] For example, suppose the input video is a popular science video introducing quantum mechanics experiments, containing 64 frames, a resolution of 224×224, a frame rate of 30fps, a content category of "physics," and a platform identifier of "YouTube." It is processed using the above-mentioned combination of BatchNormalization and channel correlation constraints: (1) Input data: multi-scale spatiotemporal feature tensor The dimensions are 224×224×256×15; Batch Normalization: Calculate the mean for each channel. With variance and apply standardized formulas ,in and These are learnable parameters; (2) Channel correlation constraint: Calculate the correlation coefficient matrix between channels. And introduce L2 regularization term Optimize parameters to reduce redundancy and correlation between channels; (3) Output: Standardized multi-scale spatiotemporal feature tensor The dimension is 224×224×256×15, which is used as the input for step S2.4 for cross-scale information fusion of the Feature Pyramid Network (FPN).

[0066] This processing effectively reduces the distribution differences and redundant correlations between multi-scale feature channels, enabling the subsequent FPN fusion and semantic path enhancement modules to more accurately capture dynamic semantic units such as "double-slit interference experiment" and "quantum superposition state" in the video, thereby improving the alignment accuracy and stability between image / video content and scientific knowledge nodes.

[0067] S3.2: Based on the local scene feature tensor, a spatial attention mechanism is used to perform semantic enhancement processing on the local regions of the image to generate local scene semantic feature vectors, and these vectors are mapped to a predefined local semantic label space to obtain a set of local scene semantic labels.

[0068] S3.3: Based on the dynamic process feature tensor, a temporal convolutional network is used to model the dynamic changes in video segments to extract dynamic process semantic feature vectors, and a set of dynamic process semantic labels is generated through a temporal semantic classifier.

[0069] S3.4: Based on the overall context feature tensor, global average pooling is used to compress the spatiotemporal features to generate an overall semantic feature vector, and a set of overall context semantic labels is generated through a fully connected classification network.

[0070] S3.5: The set of semantic labels for local scenes, the set of semantic labels for dynamic processes, and the set of semantic labels for the overall context are structurally integrated to generate a unified set of multi-level feature semantic labels and record their semantic hierarchy relationships for use by subsequent multi-head cross-modal attention networks.

[0071] Step S4 specifically includes: Multi-head projection processing is performed on the local scene semantic feature vector, dynamic process semantic feature vector, and overall context semantic feature vector to generate multiple sets of feature query vectors, key vectors, and value vectors, which serve as inputs to a multi-head cross-modal attention network, thereby enhancing the expressive power of features in different semantic spaces.

[0072] Based on the multiple sets of query vectors, key vectors, and value vectors, a dot product attention mechanism is used to calculate the initial semantic similarity between features at each scale and entity nodes in the scientific knowledge graph, in order to construct an initial semantic association matrix and capture the potential semantic relationships between image / video clips and entities in the knowledge graph.

[0073] The initial semantic association matrix is ​​subjected to Softmax normalization to obtain the probability distribution weights of features at each scale in the knowledge graph entity space, thereby generating a normalized multi-scale semantic association matrix and improving the interpretability and consistency of semantic matching between different modalities.

[0074] Based on the normalized multi-scale semantic correlation matrix, a multi-head attention fusion mechanism is used to perform weighted aggregation processing to generate a unified cross-modal semantic correlation matrix, which integrates multi-scale semantic information and enhances the semantic expression ability of complex popular science content.

[0075] The unified cross-modal semantic correlation matrix is ​​input into the knowledge perception alignment module. Combining the structural information of the knowledge graph with the semantic embedding vector, a graph-based semantic path enhancement operation is performed to optimize the structural consistency of the semantic correlation matrix and improve the alignment accuracy between image / video content and scientific knowledge nodes.

[0076] Step S5 specifically includes: The multi-scale semantic association matrix is ​​preprocessed with temporal causal modeling. A causal attention mechanism is used to model the temporal dependencies in the semantic association matrix to extract potential causal reasoning paths, thereby providing structured input for subsequent causal chain extraction.

[0077] Based on the causal attention output, a causal inference graph network is used to model the event evolution relationship between consecutive frames in an image / video clip, in order to identify the causal dependency structure of scientific events or processes in the time dimension and generate a preliminary set of causal paths.

[0078] A knowledge graph semantic path matching operation is performed on the preliminary causal path set. A graph embedding matching algorithm is used to align the causal paths with the relational paths in the scientific knowledge graph to verify whether the causal paths conform to the known scientific logical structure and generate a structured causal matching path set.

[0079] Based on the structured causal matching path set, a causal strength assessment model is introduced, and a path weight aggregation algorithm is used to calculate the confidence score of each causal path to quantify its logical support strength in the scientific knowledge system and generate a causal path constraint weight vector.

[0080] The causal path constraint weight vector is fused with the multi-scale semantic relevance matrix, and a causal enhanced semantic relevance matrix is ​​generated through a weighted fusion operation. This matrix serves as a constraint for dynamic semantic alignment, thereby improving the accuracy and logical consistency of subsequent cross-modal alignment.

[0081] Step S6 specifically includes: A knowledge graph relation path retrieval operation is performed on the causal path constraint vector. Based on the graph embedding similarity calculation method, the structured relation path that best matches the current causal path is searched in the scientific knowledge graph to obtain a set of candidate knowledge paths for the construction of subsequent alignment constraints.

[0082] Based on the candidate knowledge path set, a graph attention network is used to model the semantic similarity between causal paths and knowledge graph paths to generate path-level semantic matching weight vectors, which are used to quantify the semantic fit between each candidate path and the current image / video segment.

[0083] The path-level semantic matching weight vector and the causal path constraint vector are fused together, and a weighted fusion algorithm is used to generate a joint path constraint vector to enhance the consistency of the expression of causal reasoning results in the semantic space of the knowledge graph.

[0084] Based on the joint path constraint vector, a dynamic semantic alignment constraint modeling operation is performed. A graph neural network is used to structurally encode the entity-relationship paths in the knowledge graph to generate a structured dynamic alignment constraint vector, which is used to guide the subsequent update of the semantic relevance matrix.

[0085] The structured dynamic alignment constraint vector is input into the semantic consistency optimization module, and the constraint propagation algorithm is used to perform weighted update processing on the multi-scale semantic relevance matrix to generate a causal enhanced semantic relevance matrix, thereby improving the dynamic semantic alignment accuracy between image / video content and scientific knowledge nodes.

[0086] Step S7 specifically includes: The updated causal enhanced semantic correlation matrix is ​​processed for task scene classification. A scene recognition model is used to identify the recommended task type of the current image / video to generate a task scene classification vector, which serves as the input for subsequent adaptive weighted fusion.

[0087] Based on the task scenario classification vector, a multi-head attention weight allocation operation is performed. A learnable task-aware attention mechanism is used to weight the semantic relevance components at each scale to generate a task-driven multi-scale attention weight matrix, which serves as the dynamic weight basis for semantic fusion.

[0088] The task-driven multi-scale attention weight matrix is ​​multiplied element-wise with the causal enhanced semantic association matrix to achieve adaptive weighted fusion of semantic features at different scales, generating a weighted fused multi-scale semantic association feature tensor, thereby enhancing the expression intensity of task-related semantic components.

[0089] The spatiotemporal semantic alignment optimization process is performed on the multi-scale semantic association feature tensor. A graph attention network (GAT) is used in conjunction with the structural information of the knowledge graph to perform consistency correction on its semantic path, so as to generate a structure-optimized spatiotemporal semantic alignment feature vector and enhance the logical consistency of cross-modal semantics.

[0090] The spatiotemporal semantic alignment feature vector is input into the adaptive decoder module, and a multilayer perceptron (MLP) is used to perform nonlinear transformation processing on it to generate a unified spatiotemporal semantic fusion vector, which is used as the input of the subsequent deep semantic classifier to achieve accurate semantic mapping between image / video content and scientific knowledge graph.

[0091] Step S8: Input the spatiotemporal semantic fusion vector into a deep semantic classifier to identify the subject matter, knowledge depth, and scientific principles of the image / video content, generating a set of structured semantic tags. Specifically, this includes: The spatiotemporal semantic fusion vector is input into a multi-task deep semantic classifier. The fusion vector is then subjected to nonlinear transformation based on a pre-trained fully connected neural network to extract high-order semantic abstract features and generate a semantic abstract feature vector.

[0092] Based on the semantic abstract feature vector, a multi-branch classification head structure is used to perform subject classification tasks, knowledge depth assessment tasks, and scientific principle recognition tasks respectively, so as to realize multi-dimensional semantic parsing of image / video content and generate subject classification results, knowledge depth levels, and scientific principle matching lists.

[0093] The subject classification results are processed by knowledge graph topic node mapping. The graph embedding similarity calculation method is used to semantically align the classification results with the subject topic nodes in the knowledge graph to generate a set of structured subject affiliation tags.

[0094] Based on the knowledge depth level and the scientific principle matching list, and combined with the hierarchical structure information of the corresponding scientific principle in the knowledge graph, knowledge granularity alignment and hierarchical mapping operations are performed to generate a fine-grained semantic tag set composed of knowledge depth tags and scientific principle explanation tags.

[0095] The set of structured subject affiliation tags and the set of fine-grained semantic tags are semantically fused and format-standardized, and then serialized and encapsulated using the JSON-LD semantic data structure to generate a set of structured semantic tags for use by the user interest matching module of the subsequent recommendation system.

[0096] Step S9 specifically includes: The structured semantic tag set is preprocessed with database mapping. The semantic tags are standardized and encoded based on the knowledge graph entity identifier to generate structured data entries that conform to the knowledge tag database schema definition, ensuring the consistency and retrieval of semantic tags and knowledge graph nodes.

[0097] Based on the standardized and coded structured data entries, the semantic tag set is written to the corresponding popular science content node in the Neo4j knowledge tag database using the graph database writing interface, so as to establish the association between image / video content and its deep semantic description and realize the persistent storage of content semantics.

[0098] An index building operation is performed on the semantic tags written to the database. The Elasticsearch full-text indexing engine is used to build an inverted index for the subject affiliation, knowledge depth and scientific principle fields of the semantic tags, so as to improve the retrieval efficiency and response speed of semantic matching queries in the subsequent recommendation system.

[0099] Based on the inverted index construction results, a content-semantic tag mapping table is generated, and the mapping table is synchronized to the user interest matching module cache of the recommendation system through a message queue mechanism to achieve real-time updates and efficient access to content semantic information.

[0100] Semantic matching adaptation processing is performed on the semantic tag mapping table cached in the user interest matching module. Based on the knowledge structure features in the user dynamic profile, a semantic similarity calculation model (such as BERTScore or knowledge graph path similarity) is used to generate a semantic matching degree vector between content and user interests, providing a semantic-level matching basis for recommendation ranking.

[0101] Example 2 This embodiment provides a natural language processing-based intelligent generation system for popular science texts, used to implement the deep learning-based intelligent recommendation method for popular science content as described above.

[0102] Unless otherwise specifically stated, the relative arrangement, numerical expressions, and values ​​of the components and steps described in these embodiments do not limit the scope of this application. It should also be understood that, for ease of description, the dimensions of the various parts shown in the drawings are not drawn to actual scale. Techniques, methods, and devices known to those skilled in the art may not be discussed in detail, but where appropriate, such techniques, methods, and devices should be considered part of the specification. In all examples shown and discussed herein, any specific values ​​should be interpreted as merely exemplary and not as limitations. Therefore, other examples of exemplary embodiments may have different values. It should be noted that similar reference numerals and letters in the following drawings denote similar items; therefore, once an item is defined in one drawing, it need not be further discussed in subsequent drawings. In the description of this application, it should be understood that the orientation or positional relationship indicated by directional terms such as "front, back, up, down, left, right", "horizontal, vertical, horizontal" and "top, bottom" is usually based on the orientation or positional relationship shown in the accompanying drawings, and is only for the convenience of describing this application and simplifying the description. Unless otherwise stated, these directional terms do not indicate or imply that the device or element referred to must have a specific orientation or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation on the scope of protection of this application; the directional terms "inner" and "outer" refer to the inner and outer contours relative to the outline of each component itself.

[0103] For ease of description, spatial relative terms such as "above," "on top of," "on the upper surface of," "above," etc., are used herein to describe the spatial positional relationship of a device or feature as shown in the figures to other devices or features. It should be understood that spatial relative terms are intended to encompass different orientations in use or operation beyond the orientation of the device as described in the figures. For example, if the device in the figures were inverted, a device described as "above" or "on top of" other devices or structures would subsequently be positioned as "below" or "under" other devices or structures. Thus, the exemplary term "above" can include both "above" and "below." The device may also be positioned in other different ways (rotated 90 degrees or in other orientations), and the spatial relative descriptions used herein will be interpreted accordingly.

[0104] Furthermore, it should be noted that the use of terms such as "first" and "second" to define components is merely for the purpose of distinguishing the corresponding components. Unless otherwise stated, these terms have no special meaning and therefore should not be construed as limiting the scope of protection of this application. The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention can have various modifications and variations. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for intelligent recommendation of popular science content based on deep learning, characterized in that, include: Acquire image / video data from popular science content and record its metadata information; Spatiotemporal decoupling processing is performed on image / video data, and a three-dimensional convolutional neural network is used to extract local spatiotemporal semantic units to generate multi-scale spatiotemporal feature tensors; Based on multi-scale spatiotemporal feature tensors, multi-level feature representations of local scene features, dynamic process features, and overall context features are constructed respectively, and semantic labels of features at each scale are recorded. The multi-level feature representations are input into a multi-head cross-modal attention network to calculate the semantic association matrix between each feature scale and entity nodes in the scientific knowledge graph. Based on the semantic relevance matrix, a temporal causal chain extraction module is introduced to identify the causal order of scientific events or processes in image / video clips and generate causal path constraint vectors. The causal path constraint vector is matched with the knowledge graph relationship path to form dynamic semantic alignment constraints, and the semantic relevance matrix is ​​updated. An adaptive mapping decoder is used to perform weighted fusion of the updated semantic relevance matrix based on the task scenario type, generating a unified spatiotemporal semantic fusion vector; The spatiotemporal semantic fusion vector is input into a deep semantic classifier to identify the subject affiliation, knowledge depth, and scientific principles of image / video content, and generate a set of structured semantic tags. The structured semantic tag set is written into the knowledge tag database of popular science content, and provides semantic matching basis for the user interest matching module of the subsequent recommendation system.

2. The intelligent recommendation method for popular science content based on deep learning according to claim 1, characterized in that: The acquisition of image / video data from popular science content and the recording of its metadata information include: Collect raw image / video streams from science popularization content platforms, obtain complete visual data input, and record their original format, resolution, and encoding method; Based on video parsing technology, the collected image / video data is decomposed at the frame level, and each frame image and its corresponding timestamp information are extracted to construct a time-series visual dataset. The content of each frame of image is classified using scene recognition algorithms to identify the shooting scene category and generate a structured set of scene labels. Based on the content classification model, semantic analysis is performed on the image / video as a whole to identify its content category tag, and the tag is initially mapped to the popular science topic node in the knowledge graph; The source platform information of images / videos is extracted in a structured manner, and a platform identifier is generated as an auxiliary context feature for subsequent multimodal feature fusion.

3. The intelligent recommendation method for popular science content based on deep learning according to claim 2, characterized in that: The process of spatiotemporal decoupling image / video data, employing a 3D convolutional neural network to extract local spatiotemporal semantic units and generate multi-scale spatiotemporal feature tensors, includes: The acquired image / video data is subjected to spatiotemporal decoupling processing. A three-dimensional convolutional neural network is used to perform sliding window sampling on the continuous video frame sequence to extract local spatiotemporal semantic units and generate the original spatiotemporal feature tensor. Based on the original spatiotemporal feature tensor, multi-scale convolutional kernels are used for parallel processing to extract short-term dynamic features, mid-term process features and long-term context features, respectively, in order to construct a multi-scale spatiotemporal feature representation; The multi-scale spatiotemporal feature representation is normalized, and the Batch Normalization method is used to standardize the feature channels at each scale. Based on the normalized multi-scale spatiotemporal features, a feature pyramid network is used to fuse cross-scale information and generate an enhanced multi-scale spatiotemporal feature tensor. The enhanced multi-scale spatiotemporal feature tensor is input into the temporal gating module, and a temporal convolutional network is used to perform temporal modeling on it, outputting a structured spatiotemporal semantic feature tensor.

4. The intelligent recommendation method for popular science content based on deep learning according to claim 1, characterized in that: The method constructs multi-level feature representations based on multi-scale spatiotemporal feature tensors, including local scene features, dynamic process features, and overall context features, and records the semantic labels of features at each scale, including: The multi-scale spatiotemporal feature tensor is subjected to hierarchical decomposition to extract local scene feature tensors, dynamic process feature tensors and overall context feature tensors, forming a multi-level feature representation structure; Based on the local scene feature tensor, a spatial attention mechanism is used to perform semantic enhancement processing on local regions of the image, generating local scene semantic feature vectors, and mapping them to a predefined local semantic label space to obtain a set of local scene semantic labels. Based on the dynamic process feature tensor, a temporal convolutional network is used to model the dynamic changes in video segments, extract the semantic feature vector of the dynamic process, and generate a set of semantic labels of the dynamic process through a temporal semantic classifier. Based on the overall context feature tensor, global average pooling is used to compress the spatiotemporal features to generate an overall semantic feature vector, and a set of overall context semantic labels is generated through a fully connected classification network. The set of semantic tags for local scenes, the set of semantic tags for dynamic processes, and the set of semantic tags for the overall context are structurally integrated to generate a unified set of multi-level feature semantic tags and record their semantic hierarchy relationships.

5. The intelligent recommendation method for popular science content based on deep learning according to claim 1, characterized in that: The step of inputting multi-level feature representations into a multi-head cross-modal attention network and calculating the semantic association matrix between each feature scale and entity nodes in the scientific knowledge graph includes: The multi-level feature representation is subjected to multi-head projection processing to generate multiple sets of feature query vectors, key vectors and value vectors; Based on the multiple sets of query vectors, key vectors and value vectors, the initial semantic similarity between features at each scale and entity nodes in the scientific knowledge graph is calculated using the dot product attention mechanism, and an initial semantic association matrix is ​​constructed. The initial semantic association matrix is ​​subjected to Softmax normalization to generate a normalized multi-scale semantic association matrix; Based on the normalized multi-scale semantic association matrix, a multi-head attention fusion mechanism is used to weight and aggregate it to generate a unified cross-modal semantic association matrix. The unified cross-modal semantic association matrix is ​​input into the knowledge perception alignment module. Combining the structural information of the knowledge graph with the semantic embedding vector, a graph-based semantic path enhancement operation is performed to optimize the structural consistency of the semantic association matrix.

6. The intelligent recommendation method for popular science content based on deep learning according to any one of claims 1-5, characterized in that: The method, based on a semantic relevance matrix, introduces a temporal causal chain extraction module to identify the causal order of scientific events or processes in image / video clips and generate causal path constraint vectors, including: The semantic association matrix is ​​preprocessed by temporal causal modeling, and a causal attention mechanism is used to model the temporal dependencies in the semantic association matrix to extract potential causal inference paths. Based on the causal attention output, a causal inference graph network is used to model the event evolution relationship between consecutive frames in an image / video clip, identify the causal dependency structure of scientific events or processes, and generate a preliminary set of causal paths. A knowledge graph semantic path matching operation is performed on the preliminary causal path set. A graph embedding matching algorithm is used to align the causal paths with the relational paths in the scientific knowledge graph to generate a structured causal matching path set. Based on the structured causal matching path set, a causal strength assessment model is introduced, and a path weight aggregation algorithm is used to calculate the confidence score of each causal path to generate a causal path constraint weight vector. The causal path constraint weight vector is fused with the semantic relevance matrix, and a causal enhanced semantic relevance matrix is ​​generated through a weighted fusion operation.

7. The intelligent recommendation method for popular science content based on deep learning according to claim 1, characterized in that: The step of matching the causal path constraint vector with the knowledge graph relationship path to form dynamic semantic alignment constraints and updating the semantic relevance matrix includes: A knowledge graph relation path retrieval operation is performed on the causal path constraint vector. Based on the graph embedding similarity calculation method, the structured relation path that best matches the current causal path is searched in the scientific knowledge graph to obtain a set of candidate knowledge paths. Based on the candidate knowledge path set, a graph attention network is used to model the semantic similarity between causal paths and knowledge graph paths to generate path-level semantic matching weight vectors. The path-level semantic matching weight vector and the causal path constraint vector are fused together, and a weighted fusion algorithm is used to generate a joint path constraint vector. Based on the joint path constraint vector, a dynamic semantic alignment constraint modeling operation is performed, and a graph neural network is used to structurally encode the entity-relation path in the knowledge graph to generate a structured dynamic alignment constraint vector. The structured dynamic alignment constraint vector is input into the semantic consistency optimization module, and the constraint propagation algorithm is used to perform weighted update processing on the multi-scale semantic relevance matrix to generate a causal enhanced semantic relevance matrix.

8. The intelligent recommendation method for popular science content based on deep learning according to any one of claims 1-5, characterized in that: The step of using an adaptive mapping decoder to weight and fuse the updated semantic relevance matrix based on task scenario type specifically includes: The updated semantic association matrix is ​​processed for task scenario classification to generate task scenario classification vectors. Based on the task scenario classification vector, a learnable task-aware attention mechanism is used to weight the semantic relevance components at each scale to generate a task-driven multi-scale attention weight matrix. The task-driven multi-scale attention weight matrix and the semantic association matrix are multiplied element-wise to generate a weighted fused multi-scale semantic association feature tensor. Spatiotemporal semantic alignment optimization processing is performed on the multi-scale semantic association feature tensor. A graph attention network combined with the structural information of the knowledge graph is used to perform consistency correction on its semantic path, generating a structure-optimized spatiotemporal semantic alignment feature vector. The spatiotemporal semantic alignment feature vector is input into the adaptive decoder module, and a multilayer perceptron is used to perform nonlinear transformation processing on it to generate a unified spatiotemporal semantic fusion vector.

9. The intelligent recommendation method for popular science content based on deep learning according to any one of claims 1-5, characterized in that: The process of inputting the spatiotemporal semantic fusion vector into a deep semantic classifier to identify the subject matter, knowledge depth, and scientific principles of image / video content, and generating a structured semantic tag set, includes: The spatiotemporal semantic fusion vector is input into a multi-task deep semantic classifier, and the fusion vector is subjected to nonlinear transformation based on a pre-trained fully connected neural network to extract high-order semantic abstract features and generate a semantic abstract feature vector. Based on the semantic abstract feature vector, a multi-branch classification head structure is used to perform subject classification tasks, knowledge depth assessment tasks, and scientific principle recognition tasks respectively, generating subject classification results, knowledge depth levels, and scientific principle matching lists. The subject classification results are processed by knowledge graph topic node mapping, and the classification results are semantically aligned with the subject topic nodes in the knowledge graph using the graph embedding similarity calculation method to generate a set of structured subject affiliation tags. Based on the knowledge depth level and scientific principle matching list, and combined with the hierarchical structure information of the corresponding scientific principle in the knowledge graph, knowledge granularity alignment and hierarchical mapping operations are performed to generate a fine-grained semantic tag set composed of knowledge depth tags and scientific principle explanation tags. The set of structured subject affiliation tags and the set of fine-grained semantic tags are semantically fused and format-standardized, and then serialized and encapsulated using the JSON-LD semantic data structure to generate the set of structured semantic tags.

10. A deep learning-based intelligent recommendation system for popular science content, characterized in that: Used to implement the deep learning-based intelligent recommendation method for popular science content as described in any one of claims 1-9.

Images