Micro-learning learning path generation method based on cross-media search

Abstract

The application discloses a micro-learning learning path generation method based on cross-media search, and comprises the following steps: obtaining a learning target of a learner; performing feature extraction according to the learning target to obtain a target feature vector; performing iterative solution on a similarity matrix based on a cross relation matrix according to the target feature vector to obtain a marked path; wherein the cross relation matrix and the similarity matrix are obtained by using a trained cross-media search model; obtaining a predecessor learning path and a successor learning path according to the marked path to realize generation of the micro-learning learning path; and converting the learning target into the target feature vector, generating the similarity matrix and the cross relation matrix of the micro-learning resource based on the cross-media search model, and generating the learning path based on the similarity and the mutual relation of the micro-learning resource content, so that the quality of the generated learning path is improved.

Description

Micro-learning path generation method based on cross-media search Technical Field

[0001] This invention relates to a micro-learning path generation method based on cross-media search, belonging to the field of learning path generation technology. Background Technology

[0002] In related technologies, existing cross-media search methods mostly use RNN algorithms based on image and text media types, which cannot meet the needs of micro-learning that is mainly based on video and audio. On the other hand, most existing learning path generation methods mainly use recommendation technology, which recommends the next stage of learning content based on the learner's learning results at each stage, thereby generating a learning path. This method faces the problem of learner motivation and the cold start problem inherent in recommendation technology. Summary of the Invention

[0003] The purpose of this invention is to overcome the shortcomings of the prior art and provide a micro-learning learning path generation method based on cross-media search. By transforming the learning objective into a target feature vector and generating a similarity matrix and cross-relationship matrix of micro-learning resources based on a cross-media search model, a learning path is generated based on the similarity and interrelationship of the micro-learning resource content, thereby improving the quality of the generated learning path.

[0004] To achieve the above objectives, the present invention is implemented using the following technical solution:

[0005] This invention discloses a micro-learning path generation method based on cross-media search, comprising:

[0006] Obtain the learner's learning objectives; based on the learning objectives, perform feature extraction to obtain the target feature vector;

[0007] Based on the target feature vector, the similarity matrix is ​​iteratively solved using the cross-relationship matrix to obtain the marked path; wherein, the cross-relationship matrix and the similarity matrix are constructed using a trained cross-media search model;

[0008] Based on the labeled path, the predecessor learning path and the successor learning path are obtained to realize the generation of micro-learning learning path;

[0009] The training method for the cross-media search model includes:

[0010] Acquire micro-learning resources, which include media datasets of various media types;

[0011] For media data in a media dataset of any media type, feature extraction is performed to obtain media feature vectors, and then a training set is constructed.

[0012] Based on the training set, the pre-constructed cross-media search model is trained using the total loss function to obtain the trained cross-media search model; wherein, the total loss function includes classification constraint loss, center constraint loss and ranking constraint loss.

[0013] Furthermore, the media types include image types, video types, audio types, and text types;

[0014] The feature extraction steps are as follows:

[0015] For media data in image-type media datasets, feature extraction is performed based on convolutional neural networks to obtain image media feature vectors.

[0016] For media data in video-type media datasets, feature vectors are obtained by extracting a uniformly predetermined number of video frames and converting them into images.

[0017] For media data in audio-type media datasets, the spectrum is obtained through Fourier transform, and features are extracted to obtain audio media feature vectors.

[0018] For media data in text-based media datasets, feature extraction is performed by generating word vector models to obtain text media feature vectors.

[0019] Furthermore, the expression for the total loss function is as follows:

[0020]

[0021] In the formula, Represents the total loss function; Represents the classification constraint loss; Indicates the central constraint loss; This indicates the ranking constraint loss;

[0022] The classification constraint loss The expression is as follows:

[0023]

[0024] In the formula, m represents the ordinal index of the media type. This indicates that the media type is an image type. This indicates that the media type is video. This indicates that the media type is audio. This indicates that the media type is text.

[0025] This represents the number of media data in a media dataset of media type m; This represents the nth media feature vector of media type m; The label representing the nth media data item of media type m; Represents the media feature vector and tags The cross-entropy loss function;

[0026] The central constraint loss The expression is as follows:

[0027]

[0028] In the formula, This represents the number of media data points in the media dataset across all media types. This represents the nth media feature vector; The tag representing the nth media data The corresponding sub-category cluster centers; Represents the square of the Euclidean norm;

[0029] The expression for the ranking constraint loss is as follows:

[0030]

[0031] In the formula, This represents the media feature vectors in any four media datasets of different media types, where two of the media feature vectors belong to the same subclass and the other two media feature vectors are different. Indicates the first equilibrium parameter; Indicates the second equilibrium parameter;

[0032] Represents media feature vectors Media feature vectors The cosine distance between them;

[0033] Represents media feature vectors Media feature vectors The cosine distance between them;

[0034] Represents media feature vectors Media feature vectors The cosine distance between them.

[0035] Furthermore, the expression for the similarity matrix is ​​as follows:

[0036]

[0037] In the formula, Represents the similarity matrix; This represents the first media feature vector; This represents the nth media feature vector; Represents media feature vectors Media feature vectors Cosine similarity; Represents media feature vectors Media feature vectors The cosine similarity.

[0038] Furthermore, the expression for the cross-relationship matrix is ​​as follows:

[0039]

[0040] In the formula, Represents the cross-relationship matrix; Represents media feature vectors Media feature vectors Cosine similarity; Represents media feature vectors Media feature vectors The cosine similarity.

[0041] Furthermore, the target feature vector Belongs to the media feature vector set The elements in , ;

[0042] The steps for obtaining the marked path by iteratively solving the similarity matrix based on the cross-relationship matrix according to the target feature vector include:

[0043] Obtain the similarity matrix In and target feature vector Sort the elements of the target row i in descending order to obtain the similarity queue. , ,in, The maximum threshold representing the similarity between preceding and following content; The minimum threshold representing the similarity between preceding and following content; Represents the th element in the similarity matrix Line number Column elements; Represents the first element in the similarity matrix. Line number Column elements; Represents the th element in the similarity matrix Line number Column elements;

[0044] According to the cross relationship matrix Based on the preset cross-relationship judgment function For similarity matrix Solve the problem to obtain the adjacent rows of the target row i. ;

[0045] Delete the similarity matrix Given the target row i and target column i, replace the target row i with the adjacent row. Repeat the iterative loop of obtaining the similarity queue to finally obtain the marked path.

[0046] Furthermore, the expression for the preset cross-relationship judgment function is:

[0047]

[0048] In the formula, Function to determine cross-relationships; Represents the first element in the cross-relation matrix P. Line number Column elements; Represents the first element in the cross-relation matrix P. Line number Column elements;

[0049] In response to Then it is marked as ;

[0050] In response to Then it is marked as .

[0051] Furthermore, the marked path The expression is:

[0052] ;

[0053] In the formula, Indicates The sequence number of the second learning unit in the preceding learning path as the goal; Indicates The sequence number of the first learning unit in the preceding learning path that serves as the goal; This indicates the sequence number of the learning unit that uses the learning objective as the starting point for path generation; Indicates The sequence number of the first learning unit in the subsequent learning path that serves as the starting point; Indicates The sequence number of the second learning unit in the subsequent learning path that serves as the starting point;

[0054] The expression for the predecessor learning path is:

[0055]

[0056] In the formula, Indicates a predecessor learning path; Indicates The second learning unit in the forward learning path that serves as the goal; Indicates The first learning unit in the forward learning path that serves as the goal; This represents the learning unit that serves as the starting point for generating the learning path;

[0057] The expression for the follower learning path is:

[0058]

[0059] In the formula, Indicates a follower learning path; This represents the learning unit that serves as the starting point for generating the learning path; Indicates The first learning unit in the follow-up learning path that serves as the starting point; Indicates The second learning unit in the follow-up learning path that serves as the starting point.

[0060] Compared with the prior art, the beneficial effects achieved by the present invention are as follows:

[0061] The present invention provides a micro-learning path generation method based on cross-media search. First, based on a cross-media search model, it maps media data from different media types to the same representation space and enables mutual searching. Second, by introducing center constraint loss and ranking constraint loss into the cross-media search model, the model's fine-grained feature learning capability is improved. Furthermore, based on similarity and cross-relationship matrices, the order in which media data are learned can be determined. Finally, by transforming the learning objective into a target feature vector and generating similarity and cross-relationship matrices of micro-learning resources based on the cross-media search model, a learning path is generated based on the similarity and interrelationships of the micro-learning resource content, thus improving the quality of the generated learning path. Attached Figure Description

[0062] Figure 1 is a flowchart of the micro-learning learning path generation method based on cross-media search provided in an embodiment of the present invention.

[0063] Figure 2 is a schematic diagram of the cross-media search model provided in an embodiment of the present invention;

[0064] Figure 3 is a schematic diagram of the principle of generating an undirected graph using a similarity matrix provided in an embodiment of the present invention;

[0065] Figure 4 is a schematic diagram of the principle of transforming an undirected graph into a directed graph and removing redundant edges using a cross relationship matrix provided in an embodiment of the present invention. Detailed Implementation

[0066] The present invention will be further described below with reference to the accompanying drawings. The following embodiments are only used to more clearly illustrate the technical solution of the present invention, and should not be used to limit the scope of protection of the present invention.

[0067] This embodiment provides a micro-learning learning path generation method based on cross-media search, as shown in Figure 1, including:

[0068] Obtain the learner's learning objectives; based on the learning objectives, perform feature extraction to obtain the target feature vector;

[0069] Based on the target feature vector, the similarity matrix is ​​iteratively solved using the cross-relationship matrix to obtain the marked path; where the cross-relationship matrix and the similarity matrix are constructed using a trained cross-media search model;

[0070] Based on the labeled path, the predecessor learning path and the successor learning path are obtained to realize the generation of micro-learning learning path;

[0071] The training methods for cross-media search models include:

[0072] Acquire micro-learning resources, which include media datasets of various media types;

[0073] For media data in a media dataset of any media type, feature extraction is performed to obtain media feature vectors, and then a training set is constructed.

[0074] Based on the training set, the pre-built cross-media search model is trained using the total loss function to obtain the trained cross-media search model; the total loss function includes classification constraint loss, center constraint loss and ranking constraint loss.

[0075] The technical concept of this invention is as follows: First, based on a cross-media search model, media data of different media types can be mapped to the same representation space, enabling mutual searching. Second, by introducing center constraint loss and ranking constraint loss into the cross-media search model, the model's fine-grained feature learning capability is improved. Furthermore, based on similarity and cross-relationship matrices, the order in which media data are learned can be determined. Finally, by transforming the learning objective into a target feature vector, and generating similarity and cross-relationship matrices of micro-learning resources based on the cross-media search model, a learning path is generated based on the similarity and interrelationships of the micro-learning resource content, thus improving the quality of the generated learning path.

[0076] Before generating the target learning path, it is necessary to build and train a cross-media search model, as well as construct a cross-relationship matrix and a similarity matrix. The specific steps are as follows:

[0077] (1) Obtain micro-learning resources, which include media datasets of various media types.

[0078] For any media type, feature extraction is performed on the media data in the media dataset to obtain media feature vectors, and then a training set is constructed.

[0079] Media types include image, video, audio, and text. The feature extraction steps are as follows:

[0080] For media data in image-type media datasets, feature extraction is performed based on convolutional neural networks to obtain image media feature vectors.

[0081] For media data in video-type media datasets, feature vectors are obtained by extracting a uniformly predetermined number of video frames and converting them into images.

[0082] For media data in audio-type media datasets, the spectrum is obtained through Fourier transform, and features are extracted to obtain audio media feature vectors.

[0083] For media data in text-based media datasets, feature extraction is performed by generating word vector models to obtain text media feature vectors.

[0084] (2) Based on the training set, the pre-constructed cross-media search model is trained using the total loss function to obtain the trained cross-media search model; where the total loss function includes classification constraint loss, center constraint loss and ranking constraint loss.

[0085] Specifically, the expression for the total loss function is as follows:

[0086]

[0087] In the formula, Represents the total loss function; Represents the classification constraint loss; Indicates the central constraint loss; This indicates the ranking constraint loss;

[0088] Classification constraint loss The expression is as follows:

[0089]

[0090] In the formula, m represents the ordinal index of the media type. This indicates that the media type is an image type. This indicates that the media type is video. This indicates that the media type is audio. This indicates that the media type is text.

[0091] This represents the number of media data in a media dataset of media type m; This represents the nth media feature vector of media type m; The label representing the nth media data item of media type m; Represents the media feature vector and tags The cross-entropy loss function is calculated as follows:

[0092]

[0093] Central constraint loss The expression is as follows:

[0094]

[0095] In the formula, This represents the number of media data points in the media dataset across all media types. This represents the nth media feature vector; The tag representing the nth media data The corresponding sub-category cluster centers; Represents the square of the Euclidean norm;

[0096] The Euclidean norm is expressed as follows:

[0097]

[0098] The expression for the ranking constraint loss is as follows:

[0099]

[0100] In the formula, This represents the media feature vectors in any four media datasets of different media types, where two of the media feature vectors belong to the same subclass and the other two media feature vectors are different. Indicates the first equilibrium parameter; Indicates the second equilibrium parameter;

[0101] Represents media feature vectors Media feature vectors The cosine distance between them;

[0102] Represents media feature vectors Media feature vectors The cosine distance between them;

[0103] Represents media feature vectors Media feature vectors The cosine distance between them.

[0104] The expression for calculating cosine distance is as follows:

[0105]

[0106] In the formula, Represents the magnitude of a vector.

[0107] The first balance parameter in this embodiment The value is 1, the second balance parameter. The value is 0.5.

[0108] During training, media data from four media types were used as input, and the trained cross-media search model was obtained by minimizing the loss function. The structure of the cross-media search model is shown in Figure 2. The learning rate was initialized to 0.001 and increased by 0.5 every three rounds.

[0109] After training, the output of the fully connected layer is extracted as a common representation of the four media types. Cosine distance is used to compare each output value with other similarities to complete the cross-media search.

[0110] By adopting the above technical solution, micro-learning course data of four media types—video, audio, image, and text—can be mapped to the same representation space and search between them can be realized. Fine-grained features of media data can be extracted through three loss functions: classification constraint, center constraint, and ranking constraint.

[0111] (3) Construct a similarity matrix and a cross-relationship matrix based on the trained cross-media search model.

[0112] Specifically, the expression for the similarity matrix is ​​as follows:

[0113]

[0114] In the formula, Represents the similarity matrix; This represents the first media feature vector; This represents the nth media feature vector; Represents media feature vectors Media feature vectors Cosine similarity; Represents media feature vectors Media feature vectors The cosine similarity.

[0115] The formula for calculating cosine similarity is: , Represents the magnitude of a vector.

[0116] The expression for the cross-relationship matrix is ​​as follows:

[0117]

[0118] In the formula, Represents the cross-relationship matrix; Represents media feature vectors Media feature vectors Cosine similarity; Represents media feature vectors Media feature vectors The cosine similarity.

[0119] When constructing learning paths, the matrix It can be viewed as an undirected graph, as shown in Figure 3, by calculating the matrix. A matrix can be obtained The direction of each edge in the middle, that is, when When b is considered as the predecessor node of a, a directed graph can be obtained. The weight of each edge is the cosine similarity between two nodes, as shown in Figure 4. Given a target feature vector as the learning objective... Learning path generation is about creating a directed graph. The generated in the middle are respectively with The minimum spanning tree with a starting point and an ending point.

[0120] By adopting the above technical solution, a similarity and cross-relationship matrix can be established for media data mapped to the same representation space using a cross-media search model, so as to determine the order of learning for each media data.

[0121] The specific steps of the micro-learning learning path generation method based on cross-media search are as follows:

[0122] Step 1: Obtain the learner's learning objectives; based on the learning objectives, perform feature extraction to obtain the target feature vector;

[0123] Target feature vector Belongs to the media feature vector set The elements in , .

[0124] Step 2: Based on the target feature vector, iteratively solve the similarity matrix using the cross-relationship matrix to obtain the marked path.

[0125] 2.1 Obtaining the similarity matrix In and target feature vector Sort the elements of the target row i in descending order to obtain the similarity queue. , ,in, The maximum threshold representing the similarity between preceding and following content; The minimum threshold representing the similarity between preceding and following content; Represents the th element in the similarity matrix Line number Column elements; Represents the th element in the similarity matrix Line number Column elements; Represents the th element in the similarity matrix Line number Column elements;

[0126] By setting a minimum threshold for the similarity between two consecutive micro-learning resources. and maximum threshold This can increase the accuracy of calculations.

[0127] 2.2. Based on the cross-relationship matrix Based on the preset cross-relationship judgment function For similarity matrix To solve this, the expression for the preset cross-relationship judgment function is:

[0128]

[0129] In the formula, Function to determine cross-relationships; Represents the first element in the cross-relation matrix P. Line number Column elements; Represents the first element in the cross-relation matrix P. Line number Column elements;

[0130] In response to Then it is marked as ;

[0131] In response to Then it is marked as .

[0132] based on For matrix Solve the problem to obtain the adjacent rows of the target row i. ;

[0133] 2.3 Deleting the similarity matrix Given target row i and target column i, replace target row i with the adjacent row. Repeat the steps in section 2.1 to obtain the similarity queue, iterating through the loop until the marked path is finally obtained. , ;

[0134] In the formula, Indicates The sequence number of the second learning unit in the preceding learning path as the goal; Indicates The sequence number of the first learning unit in the preceding learning path that serves as the goal; This indicates the sequence number of the learning unit that uses the learning objective as the starting point for path generation; Indicates The sequence number of the first learning unit in the subsequent learning path that serves as the starting point; Indicates The sequence number of the second learning unit in the subsequent learning path that serves as the starting point.

[0135] Step 3: Based on the labeled path, obtain the predecessor learning path and the successor learning path to realize the generation of micro-learning learning path;

[0136] Specifically, the expression for the predecessor learning path is:

[0137]

[0138] In the formula, Indicates a predecessor learning path; Indicates The second learning unit in the forward learning path that serves as the goal; Indicates The first learning unit in the forward learning path that serves as the goal; This represents the learning unit that serves as the starting point for generating the learning path.

[0139] The expression for the follower learning path is:

[0140]

[0141] In the formula, Indicates a follower learning path; This represents the learning unit that serves as the starting point for generating the learning path; Indicates The first learning unit in the follow-up learning path that serves as the starting point; Indicates The second learning unit in the follow-up learning path that serves as the starting point.

[0142] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0143] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create means for implementing the functions specified in one or more blocks of the flowchart illustrations and / or one or more blocks of the block diagrams.

[0144] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means that implement the functions specified in one or more flowcharts and / or one or more block diagrams.

[0145] These computer program instructions may also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer-implemented process, such that the instructions, which execute on the computer or other programmable apparatus, provide steps for implementing the functions specified in one or more flowcharts and / or one or more block diagrams.

[0146] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A micro-learning learning path generation method based on cross-media search, characterized by: include: Obtain the learner's learning objectives; Based on the learning objective, feature extraction is performed to obtain the target feature vector; Based on the target feature vector, the similarity matrix is ​​iteratively solved using the cross-relation matrix to obtain the labeled path; wherein, the cross-relation matrix and the similarity matrix are constructed using a trained cross-media search model; based on the labeled path, the predecessor learning path and the successor learning path are obtained to realize the generation of micro-learning learning paths; the step of obtaining the labeled path by iteratively solving the similarity matrix based on the cross-relation matrix according to the target feature vector includes: obtaining the similarity matrix In and target feature vector Sort the elements of the target row i in descending order to obtain the similarity queue. ， ;in, The maximum threshold representing the similarity between preceding and following content; The minimum threshold representing the similarity between preceding and following content; Represents the th element in the similarity matrix Line number Column elements; Represents the th element in the similarity matrix Line number Column elements; Represents the th element in the similarity matrix Line number The elements of the column; based on the cross-relation matrix. Based on the preset cross-relationship judgment function For similarity matrix Solve the problem to obtain the adjacent rows of the target row i. ; Delete the similarity matrix Given the target row i and target column i, replace the target row i with the adjacent row. The process of iterating back to the similarity queue is repeated until the marked path is obtained; the marked path The expression is: In the formula, Indicated by The sequence number of the second learning unit in the preceding learning path as the goal; Indicated by The sequence number of the first learning unit in the preceding learning path that serves as the goal; This indicates the sequence number of the learning unit that uses the learning objective as the starting point for path generation; Indicated by The sequence number of the first learning unit in the subsequent learning path that serves as the starting point; Indicated by The sequence number of the second learning unit in the successor learning path, which serves as the starting point; the expression for the predecessor learning path is: In the formula, Indicates a predecessor learning path; Indicated by The second learning unit in the forward learning path that serves as the goal; Indicated by The first learning unit in the forward learning path that serves as the goal; This represents the learning unit that serves as the starting point for generating the learning path; the expression for the subsequent learning path is: In the formula, Indicates a follower learning path; This represents the learning unit that serves as the starting point for generating the learning path; Indicated by The first learning unit in the follow-up learning path that serves as the starting point; Indicated by The second learning unit in the follow-up learning path, serving as the starting point; wherein, the training method of the cross-media search model includes: acquiring micro-learning resources, the micro-learning resources including media datasets of multiple media types; extracting features from the media data in the media dataset of any media type to obtain media feature vectors, and then constructing a training set; training the pre-constructed cross-media search model based on the training set and a total loss function to obtain a trained cross-media search model; wherein, the total loss function includes classification constraint loss, center constraint loss, and ranking constraint loss.

2. The micro-learning learning path generation method based on cross-media search according to claim 1, characterized in that, The media types include image, video, audio, and text types. The feature extraction steps are as follows: For the media data in the image type media dataset, feature extraction is performed based on a convolutional neural network to obtain image media feature vectors; for the media data in the video type media dataset, a uniformly preset number of video frames are extracted and converted into images for feature extraction to obtain video media feature vectors; for the media data in the audio type media dataset, the spectrum is obtained through Fourier transform for feature extraction to obtain audio media feature vectors. For media data in text-based media datasets, feature extraction is performed by generating word vector models to obtain text media feature vectors.

3. The micro-learning learning path generation method based on cross-media search according to claim 1, characterized in that, The expression for the total loss function is as follows: In the formula, Represents the total loss function; Represents the classification constraint loss; Indicates the central constraint loss; This represents the ranking constraint loss; the classification constraint loss. The expression is as follows: In the formula, m represents the ordinal index of the media type. This indicates that the media type is an image type. This indicates that the media type is video. This indicates that the media type is audio. This indicates that the media type is text. This represents the number of media data in a media dataset of media type m; This represents the nth media feature vector of media type m; The label representing the nth media data item of media type m; Represents the media feature vector and tags The cross-entropy loss function; the center constraint loss The expression is as follows: In the formula, This represents the number of media data points in the media dataset across all media types. This represents the nth media feature vector; The tag representing the nth media data The corresponding sub-category cluster centers; Let represent the square of the Euclidean norm; the expression for the ranking constraint loss is as follows: In the formula, This represents the media feature vectors in any four media datasets of different media types, where two of the media feature vectors belong to the same subclass and the other two media feature vectors are different. Indicates the first equilibrium parameter; Indicates the second equilibrium parameter; Represents media feature vectors Media feature vectors The cosine distance between them; Represents media feature vectors Media feature vectors The cosine distance between them; Represents media feature vectors Media feature vectors The cosine distance between them.

4. The micro-learning learning path generation method based on cross-media search according to claim 1, characterized in that, The expression for the similarity matrix is ​​as follows: In the formula, Represents the similarity matrix; This represents the first media feature vector; This represents the nth media feature vector; Represents media feature vectors Media feature vectors Cosine similarity; Represents media feature vectors Media feature vectors The cosine similarity.

5. The micro-learning learning path generation method based on cross-media search according to claim 4, characterized in that, The expression for the cross-relationship matrix is ​​as follows: In the formula, Represents the cross-relationship matrix; Represents media feature vectors Media feature vectors Cosine similarity; Represents media feature vectors Media feature vectors The cosine similarity.

6. The micro-learning learning path generation method based on cross-media search according to claim 5, characterized in that, The target feature vector Belongs to the media feature vector set The elements in ， 。 7. The micro-learning learning path generation method based on cross-media search according to claim 6, characterized in that, The expression for the preset cross-relationship judgment function is: In the formula, Function to determine cross-relationships; Represents the first element in the cross-relation matrix P. Line number Column elements; Represents the first element in the cross-relation matrix P. Line number Column elements; responding to Then it is marked as ; in response Then it is marked as 。

Images