Educational resource intelligent recommendation method and system

Abstract

The invention relates to the technical field of data analysis, in particular to an educational resource intelligent recommendation method and system. An educational resource intelligent recommendation system comprises a user demand description text data management module, an educational resource matching module and a user demand updating module. According to the method, the user demand is expanded through the target educational resource corresponding to the click operation of the user, the demand direction of the user can be analyzed, more accurate user demand description is obtained, the educational resource is recommended based on the more accurate user demand description, and the user experience is improved. The accuracy of educational resource recommendation can be effectively improved, and the use experience of a user is improved. And when the educational resources are matched, the educational resource label word vectors in the same dimension space are used, so that the matching deviation caused by expression difference can be effectively avoided.

Description

Technical Field

[0001] The present invention relates to the technical field of data analysis, and in particular to an intelligent recommendation method and system for educational resources. Background Art

[0002] At present, with the explosive growth of online educational resources, how users can quickly and accurately find educational resources that meet their own needs has become an urgent problem to be solved. Most of the existing educational resource recommendation methods rely on users' keyword searches. However, in actual applications, the user demand descriptions initially entered by users are often relatively simple and general. For example, only "primary school mathematics" or "English grammar" is entered, which is difficult to accurately express their personalized learning needs. Such simple demand descriptions will cause the system to be difficult to understand the true intentions of users, resulting in a low matching degree between the recommended educational resources and the actual needs of users, and reducing users' satisfaction with the recommendation system. Summary of the Invention

[0003] The present invention expands the user's needs through the target educational resources corresponding to the user's click operations, can analyze the user's demand direction, and then obtain a more accurate description of the user's needs. Based on the more accurate description of the user's needs, the educational resources are recommended, which can effectively improve the accuracy rate of educational resource recommendation and enhance the user experience; and when matching educational resources, the educational resource label word vectors in the same dimensional space are used, which can effectively avoid the matching deviation caused by expression differences.

[0004] The present invention provides an intelligent recommendation method for educational resources, including:

[0005] Step S1: Obtain user demand description text data, perform named entity recognition operations on the user demand description text data, and output a user demand word set. The user demand word set includes several educational resource label words derived from the user demand description text. Perform word embedding operations on the educational resource label words in the user demand word set to construct a user demand word vector set, and the user demand word vector set stores educational resource label word vectors;

[0006] Step S2: Traverse all educational resources. For each educational resource, perform the following steps: Obtain the set of characteristic words corresponding to the educational resource. The set of characteristic words includes educational resource label words from the educational resource. Perform word embedding operations on the educational resource label words in the set of characteristic words to construct a set of characteristic word vectors. The set of characteristic word vectors stores the educational resource label word vectors. Match the user demand word vector set with the set of characteristic word vectors. The matching method is as follows: Denote the similarity between any educational resource label word vector in the user demand word vector set and any educational resource label word vector in the set of characteristic word vectors as the matching value. If the matching value is greater than the matching threshold, it is regarded as a successful match. If the matching value is not greater than the matching threshold, it is regarded as a failed match. Record the number of successful matches as the matching degree corresponding to the educational resource. Until all educational resources have been traversed, output the matching degrees corresponding to all educational resources, and arrange all educational resources in descending order according to the corresponding matching degrees. Select the top N educational resources to form an educational resource recommendation list;

[0007] Step S3: Determine the educational resource viewed by the user based on the user's operation feedback and denote it as the target educational resource. Denote the set of characteristic words corresponding to the target educational resource as the target word set. Perform word embedding operations on the educational resource label words in the target word set to construct a target word vector set. Then, send the user demand word vector set and the target word vector set into the user demand extension model for processing, and output an extended demand word set. The extended demand word set includes several educational resource label words, and the number of educational resource label words in the extended demand word set is greater than the number of educational resource label words in the user demand word set. Replace the user demand word set with the extended demand word set to update the user demand word set, and return to step S2.

[0008] As a preferred aspect, the user demand expansion model includes a user demand word vector update layer, a user demand analysis layer, a user demand weight analysis layer and a user demand expansion layer, wherein the user demand word vector update layer is used to update the educational resource label word vector in the user demand word vector set based on the target word vector set to construct an updated demand word vector feature map; the user demand analysis layer includes N decoding units, and the user demand analysis layer is used to process the updated demand word vector feature map to construct an extended demand word vector feature; the user demand weight analysis layer is used to splice the user demand word vector set and the target word vector set and send them into a multi-layer perceptron for processing, and the input Output the user demand weight vector; the user demand expansion layer is used to perform a full connection operation on the extended demand word vector features, and output the extended demand word probability vector. The extended demand word probability vector includes the probability values ​​corresponding to all educational resource label words. The extended demand word probability vector is enhanced by the user demand weight vector to obtain an enhanced extended demand word probability vector. The enhanced extended demand word probability vector also includes the probability values ​​corresponding to all educational resource label words. The enhanced extended demand word probability vector is traversed, and all educational resource label words in the enhanced extended demand word probability vector whose probability values ​​are greater than the value threshold are combined into an extended demand word set for output.

[0009] As a preferred aspect, the educational resource label word vectors in the user demand word vector set are updated based on the target word vector set through the user demand word vector update layer to construct an updated demand word vector feature map, which specifically includes the following steps:

[0010] For the i-th educational resource label word vector F in the user demand word vector set i , i = 1, 2, 3, ..., I, I is the total number of educational resource label word vectors in the user demand word vector set, perform the following operations, traverse the target word vector set, and calculate the i-th educational resource label word vector F in the user demand word vector set i and the jth educational resource label word vector H in the target word vector set j The similarity between them is denoted as B ij , j = 1, 2, 3, ..., J, where J is the total number of educational resource label word vectors in the target word vector set. If the i-th educational resource label word vector F in the user demand word vector set i and the jth educational resource label word vector H in the target word vector set j The similarity between ij If the similarity is greater than the threshold, the j-th educational resource label word vector H in the target word vector set is j The i-th educational resource label word vector F stored in the user demand word vector set i In the corresponding neighbor set M(i), if the i-th educational resource label word vector F in the user demand word vector seti and the similarity B j with the j-th educational resource label word vector H in the target word vector set ij is not greater than the similarity threshold, no operation is performed; until the target word vector set is traversed, the i-th educational resource label word vector F in the user demand word vector set is output i The corresponding neighbor set, for the i-th educational resource label word vector F in the user demand word vector set through the following formula i is updated as follows:

[0011]

[0012] where U i is the i-th updated demand word vector feature vector, σ() is the non-linear activation function, E k is the k-th educational resource label word vector in the neighbor set M(i), α ik is the i-th educational resource label word vector F in the user demand word vector set i is the normalized similarity between the i-th educational resource label word vector F in the user demand word vector set and the k-th educational resource label word vector in the neighbor set M(i), and W is the linear transformation matrix;

[0013] Concatenate all the updated demand word vector feature vectors Ui from top to bottom to construct an updated demand word vector feature map.

[0014] As a preferred aspect, the updated demand word vector feature map is processed by the user demand analysis layer to construct an extended demand word vector feature, which specifically includes the following steps:

[0015] For the first decoding unit, the updated demand word vector feature map is respectively multiplied by the value weight matrix, the key weight matrix, and the query weight matrix to construct the corresponding value vector V, key vector K, and query vector Q, and the self-attention mechanism operation is performed through the following formula: R = softmax(QK T / D 0.5 )V, where R is the output of the decoding unit, T is the matrix transpose operation, and D is the dimension size of the key vector K;

[0016] In the processing of the second decoding unit to the N-th decoding unit, the output of the previous decoding unit is multiplied by the query weight matrix to construct the corresponding query vector Q, and the remaining operations are the same as those of the first decoding unit.

[0017] As a preferred aspect, the user demand extension model is trained, including the following steps:

[0018] Obtain several user requirement expansion training samples. The user requirement expansion training samples include a user requirement word vector set and a random word vector set. The user requirement word vector set is derived from the user requirement description document, and the random word vector set is derived from the educational resources during the user's search process. Label the user requirement expansion training samples with the target word vector set, and form the user requirement expansion training set with all the labeled user requirement expansion training samples. Train the user requirement expansion model with the user requirement expansion training set, calculate the user requirement expansion loss value, and determine whether the user requirement expansion loss value is within the first preset range. If the user requirement expansion loss value is within the first preset range, output the trained user requirement expansion model; otherwise, continue to train the user requirement expansion model with the user requirement expansion training set.

[0019] As a preferred aspect, the similarity calculation uses the cosine similarity algorithm.

[0020] The present invention also provides an intelligent educational resource recommendation system, including:

[0021] A user requirement description text data management module, which obtains user requirement description text data, performs named entity recognition operations on the user requirement description text data, and outputs a user requirement word set. The user requirement word set includes several educational resource label words derived from the user requirement description text. Perform word embedding operations on the educational resource label words in the user requirement word set to construct a user requirement word vector set, and the user requirement word vector set stores educational resource label word vectors.

[0022] An educational resource matching module, which is used to traverse all educational resources. For each educational resource, perform the following steps: obtain the corresponding feature word set of the educational resource. The feature word set includes educational resource label words derived from the educational resource. Perform word embedding operations on the educational resource label words in the feature word set to construct a feature word vector set, and the feature word vector set stores educational resource label word vectors. Match the user requirement word vector set with the feature word vector set. The matching method is as follows: Denote the similarity between any educational resource label word vector in the user requirement word vector set and any educational resource label word vector in the feature word vector set as the matching value. If the matching value is greater than the matching threshold, it is regarded as a successful match; if the matching value is not greater than the matching threshold, it is regarded as a failed match. Denote the number of successful matches as the matching degree corresponding to the educational resource. Until all educational resources are traversed, output the matching degrees corresponding to all educational resources, and arrange all educational resources in descending order according to the corresponding matching degrees, and select the top N educational resources to form an educational resource recommendation list.

[0023] A user requirement update module is used to determine the educational resources viewed by the user based on the user's operation feedback and record them as target educational resources, record the set of characteristic words corresponding to the target educational resources as the target word set, perform word embedding operations on the educational resource label words in the target word set to construct a target word vector set, and then send the user requirement word vector set and the target word vector set into the user requirement expansion model for processing to output an expanded requirement word set. The expanded requirement word set includes several educational resource label words, and the number of educational resource label words in the expanded requirement word set is greater than the number of educational resource label words in the user requirement word set; replace the user requirement word set with the expanded requirement word set to achieve the update of the user requirement word set.

[0024] The present invention has the following advantages:

[0025] The present invention expands the user requirements through the target educational resources corresponding to the user's click operation, can analyze the user's demand direction, and then obtain a more accurate description of the user's requirements. Based on the more accurate description of the user's requirements to recommend educational resources, it can effectively improve the accuracy of educational resource recommendation and enhance the user experience; and when matching educational resources, the educational resource label word vectors in the same dimensional space are used, which can effectively avoid the matching deviation caused by expression differences. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] Figure 1 It is a schematic structural diagram of an educational resource intelligent recommendation system adopted in an embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0027] In order to enable those skilled in the art to better understand the technical solutions in the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention.

[0028] Embodiment 1, an educational resource intelligent recommendation method, includes:

[0029] Step S1: Obtain the user demand description text data. Here, the user demand description text data refers to the description statements input by the user when searching for educational resources, such as "Learning Python", "Efficient postgraduate entrance examination mathematics", etc. Perform named entity recognition operation on the user demand description text data. Here, the named entity recognition uses existing named entity extraction models, such as the BERT model, and output the user demand word set. The user demand word set includes several educational resource label words derived from the user demand description text. These educational resource label words refer to words derived from educational resources, such as "Python", "Postgraduate entrance examination mathematics", etc. Perform word embedding operation on the educational resource label words in the user demand word set. The word embedding operation uses the Word2Vec model to construct the user demand word vector set. The user demand word vector set stores the educational resource label word vectors;

[0030] Step S2: Traverse all educational resources. There will be a large number of educational resources on the educational resource recommendation platform. Here, the educational resources include but are not limited to text, audio, and video, etc., such as "Detailed Explanation of Postgraduate Entrance Examination Mathematics", "Oral English Training", etc. These educational resources are generally uploaded by the author or obtained by the platform through purchase. For each educational resource, perform the following steps: Obtain the corresponding feature word set of the educational resource. The feature word set includes the educational resource label words derived from the educational resource. It should be noted that whenever an educational resource is uploaded to the platform, named entity recognition operation will be performed on the overview of the educational resource to obtain the corresponding feature word set, and the educational resource label word here is used as a kind of upper-level concept. Although both the user demand word set and the feature word set include educational resource label words, the educational resource label words in the user demand word set and the feature word set are not exactly the same. For example, the user demand word set includes "Python", but if the current educational resource is in English, the corresponding feature word set will not include "Python". Perform word embedding operation on the educational resource label words in the feature word set to construct the feature word vector set. The feature word vector set stores the educational resource label word vectors. Match the user demand word vector set with the feature word vector set. The matching method is as follows: Denote the similarity between any educational resource label word vector in the user demand word vector set and any educational resource label word vector in the feature word vector set as the matching value. The similarity calculation can use the cosine similarity algorithm. If the matching value is greater than the matching threshold, the matching threshold is set in advance by the developer, generally 0.7, it is regarded as a successful match. If the matching value is not greater than the matching threshold, it is regarded as a failed match; Denote the number of successful matches as the matching degree corresponding to the educational resource; Until all educational resources are traversed, output the matching degrees corresponding to all educational resources, and arrange all educational resources in descending order according to the corresponding matching degrees, and select the top N educational resources to form the educational resource recommendation list, where N is determined by the developer;

[0031] Step S3: Determine the educational resource viewed by the user according to the user's operation feedback and record it as the target educational resource. Here, the operation feedback generally refers to the user's click operation. Denote the set of characteristic words corresponding to the target educational resource as the target word set. Perform word embedding operations on the educational resource label words in the target word set to construct a target word vector set. Then, send the user demand word vector set and the target word vector set into the user demand extension model for processing, and output an extended demand word set. The extended demand word set includes several educational resource label words, and the number of educational resource label words in the extended demand word set is greater than the number of educational resource label words in the user demand word set. It should be noted that the user's initial input of the user demand description is often relatively simple, and the user clicks to select a certain educational resource, indicating that the selected educational resource conforms to the user's demand direction. Therefore, the user demand extension model is used to further characterize the user's demand direction, obtain more accurately described extended demand words, and perform recommendation operations based on the more accurately described extended demand words, which can improve the accuracy of educational resource recommendation and enhance the user experience. Replace the user demand word set with the extended demand word set to update the user demand word set, and return to step S2;

[0032] It should be noted that each time the user performs a click operation, the user demand word set will be updated, thereby generating a new educational resource recommendation list, and continuously performing the educational resource recommendation work until the user completes the search for educational resources (such as logging out of the educational resource recommendation platform);

[0033] This application expands the user demand through the target educational resource corresponding to the user's click operation, can analyze the user's demand direction, and then obtain a more accurate user demand description. Recommending educational resources based on the more accurate user demand description can effectively improve the accuracy of educational resource recommendation and enhance the user experience. Moreover, when matching educational resources, the educational resource label word vectors in the same dimensional space are used, which can effectively avoid matching deviations caused by expression differences.

[0034] The user demand extension model includes a user demand word vector update layer, a user demand analysis layer, a user demand weight analysis layer, and a user demand extension layer. The user demand word vector update layer is used to update the educational resource label word vectors in the user demand word vector set based on the target word vector set to construct an updated demand word vector feature map. During the process of updating the educational resource label word vectors in the user demand word vector set through the target word vector set, the fusion of the educational resource label word vectors corresponding to the target educational resources can be achieved, providing reference information for subsequent demand direction analysis. The user demand analysis layer includes N decoding units, where N is generally 6. The decoding units are established based on the decoder in the Transformer model. The user demand analysis layer is used to process the updated demand word vector feature map to construct an extended demand word vector feature. During the process of processing the updated demand word vector feature map, the interdependent relationships in the updated demand word vector feature map can be analyzed, and the user's demand direction can be analyzed. The user demand weight analysis layer is used to splice the user demand word vector set and the target word vector set and send them to a multi-layer perceptron for processing, outputting a user demand weight vector. The user demand weight vector can represent the weight ratio of the user demand direction on different educational resource label words. The user demand extension layer is used to perform a fully connected operation on the extended demand word vector feature, outputting an extended demand word probability vector. The extended demand word probability vector includes the probability values corresponding to all educational resource label words. Here, all educational resource label words refer to those corresponding to all educational resources on the current platform. The extended demand word probability vector is strengthened through the user demand weight vector to obtain a strengthened extended demand word probability vector. The strengthened extended demand word probability vector also includes the probability values corresponding to all educational resource label words. Traverse the strengthened extended demand word probability vector, and form an extended demand word set by the educational resource label words whose probability values in the strengthened extended demand word probability vector are greater than the value threshold for output. The value threshold is determined by the developer;

[0035] Based on the target word vector set, the user demand word vector update layer updates the educational resource label word vectors in the user demand word vector set to construct an updated demand word vector feature map, which specifically includes the following steps:

[0036] For the i-th educational resource label word vector F in the user demand word vector set i , i = 1, 2, 3,..., I, where I is the total number of educational resource label word vectors in the user demand word vector set, perform the following operations. Traverse the target word vector set and calculate the similarity between the i-th educational resource label word vector F in the user demand word vector set i and the j-th educational resource label word vector H in the target word vector set j , denoted as B ij, j = 1, 2, 3, ..., J, where J is the total number of educational resource label word vectors in the target word vector set. If the i-th educational resource label word vector F in the user demand word vector set i and the jth educational resource label word vector H in the target word vector set j The similarity between ij is greater than the similarity threshold, which is set by the developer. The j-th educational resource label word vector H in the target word vector set is j The i-th educational resource label word vector F stored in the user demand word vector set i In the corresponding neighbor set M(i), if the i-th educational resource label word vector F in the user demand word vector set i and the jth educational resource label word vector H in the target word vector set j The similarity between ij If the similarity is not greater than the threshold, no operation is performed; until the target word vector set is traversed, the i-th educational resource label word vector F in the user demand word vector set is output. i The corresponding neighbor set is calculated by the following formula for the i-th educational resource label word vector F in the user demand word vector set: i To update:

[0037]

[0038] Among them U i is the feature vector of the i-th update demand word vector, σ() is the nonlinear activation function, E k is the kth educational resource label word vector in the neighbor set M(i), α ik is the i-th educational resource label word vector F in the user demand word vector set i The normalized similarity between the kth educational resource label word vector in the neighbor set M(i) and the normalization here can be performed using the softmax function to calculate the i-th educational resource label word vector F in the user demand word vector set. i The similarity between all educational resource label word vectors in the neighbor set M(i) is normalized, and W is a linear change matrix, which is adjusted according to user needs to expand model training;

[0039] It should be noted that the educational resource label word vectors related to the i-th educational resource label word vector are searched in the target word vector set. These found educational resource label word vectors indicate the demand direction that the user may need, and the i-th educational resource label word vector in the user demand word vector set is updated according to the found educational resource label word vector, which can guide the direction of subsequent user demand expansion;

[0040] Concatenate all the updated requirement word vector feature vectors Ui from top to bottom to construct an updated requirement word vector feature map.

[0041] Process the updated requirement word vector feature map through the user requirement analysis layer to construct extended requirement word vector features, which specifically include the following steps:

[0042] For the first decoding unit, perform multiplication operations on the updated requirement word vector feature map with the value weight matrix, key weight matrix, and query weight matrix respectively to construct the corresponding value vector V, key vector K, and query vector Q, and perform the self-attention mechanism operation through the following formula: R = softmax(QK T / D 0.5 )V, where R is the output of the decoding unit, T is the matrix transpose operation, and D is the dimension size of the key vector K; it should be noted that the specific self-attention mechanism operation refers to the Transformer model;

[0043] In the processing of the second decoding unit to the Nth decoding unit, perform a multiplication operation on the output of the previous decoding unit with the query weight matrix to construct the corresponding query vector Q, and the remaining operations are the same as those of the first decoding unit.

[0044] In the user requirement extension layer, strengthen the extended requirement word probability vector through the user requirement weight vector to obtain a strengthened extended requirement word probability vector, which specifically includes the following steps: perform a Hadamard multiplication operation on the normalized user requirement weight vector and the extended requirement word probability vector to obtain the strengthened extended requirement word probability vector.

[0045] Train the user requirement extension model, including the following steps:

[0046] Obtain a number of user requirement extension training samples. The user requirement extension training samples include a user requirement word vector set and a random word vector set. The user requirement word vector set is derived from the user requirement description document, and the random word vector set is derived from the educational resources during the user search process. It should be noted that each acquisition of the user requirement extension training sample is an actual educational resource search process. Label the user requirement extension training samples through the target word vector set. Here, the target word vector set is derived from the educational resources finally selected in the actual educational resource search action. Combine all the labeled user requirement extension training samples into a user requirement extension training set, train the user requirement extension model through the user requirement extension training set, calculate the user requirement extension loss value, and determine whether the user requirement extension loss value is within the first preset range. The first preset range is set by the developer. If the user requirement extension loss value is within the first preset range, output the trained user requirement extension model; otherwise, continue to train the user requirement extension model through the user requirement extension training set.

[0047] Example 2, an intelligent educational resource recommendation system, as Figure 1 shown, includes:

[0048] A user demand description text data management module, which is used to obtain user demand description text data. Here, the user demand description text data refers to the description statements input by users when searching for educational resources, such as "learning Python", "efficient postgraduate entrance examination mathematics", etc. Perform named entity recognition operations on the user demand description text data. Here, the named entity recognition uses existing named entity extraction models, such as the BERT model, and outputs a user demand word set. The user demand word set includes several educational resource label words derived from the user demand description text. These educational resource label words refer to words derived from educational resources, such as "Python", "postgraduate entrance examination mathematics", etc. Perform word embedding operations on the educational resource label words in the user demand word set. The word embedding operation uses the Word2Vec model to construct a user demand word vector set. The user demand word vector set stores educational resource label word vectors;

[0049] An educational resource matching module is used to traverse all educational resources. There are a large number of educational resources on the educational resource recommendation platform. The educational resources here include, but are not limited to, texts, audios, videos, etc. For example, "Comprehensive Explanation of Postgraduate Entrance Examination Mathematics", "Oral English Training", etc. These educational resources are generally uploaded by the author or obtained by the platform through purchase. For each educational resource, the following steps are executed: Obtain the set of characteristic words corresponding to the educational resource. The set of characteristic words includes the educational resource label words derived from the educational resource. It should be noted that whenever an educational resource is uploaded to the platform, a named entity recognition operation is performed on the overview of the educational resource to obtain the corresponding set of characteristic words, and the educational resource label words here are used as a kind of upper-level concept. Although both the user demand word set and the set of characteristic words include educational resource label words, the educational resource label words in the user demand word set and the set of characteristic words are not exactly the same. For example, the user demand word set includes "Python", but the current educational resource is in English, and there will be no "Python" in the corresponding set of characteristic words. Perform word embedding operations on the educational resource label words in the set of characteristic words to construct a set of characteristic word vectors. The set of characteristic word vectors stores the educational resource label word vectors. Match the user demand word vector set with the set of characteristic word vectors. The matching method is as follows: Denote the similarity between any educational resource label word vector in the user demand word vector set and any educational resource label word vector in the set of characteristic word vectors as the matching value. The similarity calculation can use the cosine similarity algorithm. If the matching value is greater than the matching threshold, which is set in advance by the developer and is generally 0.7, it is regarded as a successful match. If the matching value is not greater than the matching threshold, it is regarded as a failed match; Record the number of successful matches as the matching degree corresponding to the educational resource; Until all educational resources have been traversed, output the matching degrees corresponding to all educational resources, and arrange all educational resources in descending order according to the corresponding matching degrees. Select the top N educational resources to form an educational resource recommendation list, where N is determined by the developer;

[0050] The user requirement update module is used to determine the educational resources viewed by the user based on the user's operation feedback and record them as target educational resources. Here, the operation feedback generally refers to the user's click operation. The set of characteristic words corresponding to the target educational resources is recorded as the target word set. The educational resource label words in the target word set are subjected to word embedding operations to construct a target word vector set. Then, the user requirement word vector set and the target word vector set are sent to the user requirement extension model for processing, and an extended requirement word set is output. The extended requirement word set includes several educational resource label words, and the number of educational resource label words in the extended requirement word set is greater than the number of educational resource label words in the user requirement word set. It should be noted that the user's initial input of the user requirement description is often relatively simple, and the user clicks to select a certain educational resource, indicating that the selected educational resource meets the user's requirement direction. Therefore, the user requirement extension model is used to further characterize the user's requirement direction to obtain more accurately described extended requirement words. Performing the recommendation operation based on the more accurately described extended requirement words can improve the accuracy of educational resource recommendation and enhance the user experience. Replace the user requirement word set with the extended requirement word set to update the user requirement word set.

[0051] It should be understood that those of ordinary skill in the art can make improvements or changes based on the above description, and all such improvements and changes should fall within the protection scope of the appended claims of the present invention. The parts not described in detail in this specification belong to the prior art well-known to those of ordinary skill in the art.

Claims

1. An intelligent recommendation method for educational resources, characterized in that, Including: Step S1: Obtain the user requirement description text data, perform named entity recognition operation on the user requirement description text data, and output the user requirement word set. The user requirement word set includes several educational resource label words derived from the user requirement description text. Perform word embedding operation on the educational resource label words in the user requirement word set to construct the user requirement word vector set, and the user requirement word vector set stores the educational resource label word vectors; Step S2: Traverse all educational resources. For each educational resource, perform the following steps: Obtain the corresponding feature word set of the educational resource. The feature word set includes the educational resource label words derived from the educational resource. Perform word embedding operation on the educational resource label words in the feature word set to construct the feature word vector set, and the feature word vector set stores the educational resource label word vectors. Match the user requirement word vector set with the feature word vector set. The matching method is as follows: Denote the similarity between any educational resource label word vector in the user requirement word vector set and any educational resource label word vector in the feature word vector set as the matching value. If the matching value is greater than the matching threshold, it is regarded as a successful match; if the matching value is not greater than the matching threshold, it is regarded as a failed match. Denote the number of successful matches as the matching degree corresponding to the educational resource. Until all educational resources are traversed, output the matching degrees corresponding to all educational resources, and arrange all educational resources in descending order according to the corresponding matching degrees, and select the top N educational resources to form the educational resource recommendation list; Step S3: Determine the educational resource viewed by the user according to the user's operation feedback and denote it as the target educational resource. Denote the corresponding feature word set of the target educational resource as the target word set. Perform word embedding operation on the educational resource label words in the target word set to construct the target word vector set. Then send the user requirement word vector set and the target word vector set into the user requirement extension model for processing, and output the extended requirement word set. The extended requirement word set includes several educational resource label words, and the number of educational resource label words in the extended requirement word set is greater than the number of educational resource label words in the user requirement word set. Replace the user requirement word set with the extended requirement word set to update the user requirement word set, and return to Step S2.

2. The intelligent recommendation method for educational resources according to claim 1, wherein The user requirement extension model includes a user requirement word vector update layer, a user requirement analysis layer, a user requirement weight analysis layer, and a user requirement extension layer. Among them, the user requirement word vector update layer is used to update the educational resource label word vectors in the user requirement word vector set based on the target word vector set to construct an updated requirement word vector feature map; the user requirement analysis layer includes N decoding units, and the user requirement analysis layer is used to process the updated requirement word vector feature map to construct an extended requirement word vector feature; The user requirement weight analysis layer is used to splice the user requirement word vector set and the target word vector set and then send them into a multi-layer perceptron for processing, and output the user requirement weight vector; the user requirement extension layer is used to perform a fully connected operation on the extended requirement word vector features, and output the extended requirement word probability vector. The extended requirement word probability vector includes the probability values corresponding to all educational resource label words. The extended requirement word probability vector is strengthened by the user requirement weight vector to obtain the strengthened extended requirement word probability vector, which also includes the probability values corresponding to all educational resource label words. Traverse the strengthened extended requirement word probability vector, and form an extended requirement word set by the educational resource label words in the strengthened extended requirement word probability vector whose probability values are greater than the value threshold for output.

3. The intelligent education resource recommendation method according to claim 2, characterized in that, Based on the target word vector set, the user requirement word vector update layer updates the educational resource label word vectors in the user requirement word vector set to construct an updated requirement word vector feature map, which specifically includes the following steps: For the i-th educational resource tag vector F in the user demand word vector set i , where i = 1, 2, 3, …, I, and I is the total number of educational resource tag vectors in the user demand word vector set, perform the following operations. Traverse the target word vector set and calculate the similarity between the i-th educational resource tag vector F i in the user demand word vector set and the j-th educational resource tag vector H j in the target word vector set, denoted as B ij , where j = 1, 2, 3, …, J, and J is the total number of educational resource tag vectors in the target word vector set. If the similarity B i between the i-th educational resource tag vector F j in the user demand word vector set and the j-th educational resource tag vector H ij in the target word vector set is greater than the similarity threshold, store the j-th educational resource tag vector H j in the target word vector set into the neighbor set M(i) corresponding to the i-th educational resource tag vector F i in the user demand word vector set. If the similarity B i between the i-th educational resource tag vector F j in the user demand word vector set and the j-th educational resource tag vector H ij in the target word vector set is not greater than the similarity threshold, no operation is performed; until the target word vector set is traversed completely, output the neighbor set corresponding to the i-th educational resource tag vector F i in the user demand word vector set. Update the i-th educational resource tag vector F i in the user demand word vector set through the following formula: Among them, U i is the feature vector of the i-th updated demand word vector, σ() is a non-linear activation function, and E k is the k-th educational resource label word vector in the neighbor set M(i), and α ik is the i-th educational resource label word vector F in the user demand word vector set i is the normalized similarity between the educational resource label word vector and the k-th educational resource label word vector in the neighbor set M(i), and W is a linear transformation matrix; Splice all the updated requirement word vector feature vectors Ui from top to bottom to construct an updated requirement word vector feature map.

4. An intelligent recommendation method for educational resources according to claim 3, characterized in that, The user requirement analysis layer processes the updated requirement word vector feature map to construct extended requirement word vector features, which specifically includes the following steps: For the first decoding unit, the updated required word vector feature map is respectively multiplied with the value weight matrix, the key weight matrix, and the query weight matrix to construct the corresponding value vector V, key vector K, and query vector Q, and the self-attention mechanism operation is performed through the following formula: R = softmax(QK T / D 0.5 )V, where R is the output of the decoding unit, T is the matrix transpose operation, and D is the dimension size of the key vector K; In the processing of the second decoding unit to the Nth decoding unit, the output of the previous decoding unit is multiplied by the query weight matrix to construct the corresponding query vector Q, and the remaining operations are the same as those of the first decoding unit.

5. An intelligent education resource recommendation method according to claim 4, characterized in that, Training the user requirement extension model includes the following steps: Obtain a number of user requirement extension training samples. The user requirement extension training samples include a user requirement word vector set and a random word vector set. The user requirement word vector set is derived from the user requirement description document, and the random word vector set is derived from the educational resources during the user search process. The user requirement extension training samples are labeled by the target word vector set, and all the labeled user requirement extension training samples are combined into a user requirement extension training set. The user requirement extension model is trained by the user requirement extension training set, calculate the user requirement extension loss value, and judge whether the user requirement extension loss value is within the first preset range. If the user requirement extension loss value is within the first preset range, output the trained user requirement extension model; otherwise, continue to train the user requirement extension model by the user requirement extension training set.

6. The intelligent recommendation method of educational resources according to claim 5, characterized in that The similarity calculation uses the cosine similarity algorithm.

7. An intelligent educational resource recommendation system, characterized in that, The system applies the educational resource intelligent recommendation method described in any one of claims 1-6 above, including: The user requirement description text data management module obtains the user requirement description text data, performs named entity recognition operations on the user requirement description text data, and outputs the user requirement word set. The user requirement word set includes several educational resource label words derived from the user requirement description text. Perform word embedding operations on the educational resource label words in the user requirement word set to construct a user requirement word vector set, and the user requirement word vector set stores the educational resource label word vectors; An educational resource matching module is used to traverse all educational resources. For each educational resource, the following steps are executed: Obtain the set of characteristic words corresponding to the educational resource. The set of characteristic words includes educational resource label words derived from the educational resource. Perform word embedding operations on the educational resource label words in the set of characteristic words to construct a set of characteristic word vectors. The set of characteristic word vectors stores educational resource label word vectors. Match the user demand word vector set with the set of characteristic word vectors in the following way: Denote the similarity between any educational resource label word vector in the user demand word vector set and any educational resource label word vector in the set of characteristic word vectors as the matching value. If the matching value is greater than the matching threshold, it is regarded as a successful match. If the matching value is not greater than the matching threshold, it is regarded as a failed match. Record the number of successful matches as the matching degree corresponding to the educational resource. Until all educational resources have been traversed, output the matching degrees corresponding to all educational resources, and arrange all educational resources in descending order according to the corresponding matching degrees. Select the top N educational resources to form an educational resource recommendation list; A user demand update module is used to determine the educational resource viewed by the user according to the user's operation feedback and record it as the target educational resource. Denote the set of characteristic words corresponding to the target educational resource as the target word set. Perform word embedding operations on the educational resource label words in the target word set to construct a target word vector set. Then, send the user demand word vector set and the target word vector set into the user demand expansion model for processing, and output an expanded demand word set. The expanded demand word set includes several educational resource label words, and the number of educational resource label words in the expanded demand word set is greater than the number of educational resource label words in the user demand word set. Replace the user demand word set with the expanded demand word set to achieve the update of the user demand word set.

Images