An education resource recommendation method of fusing review text and review object content

Abstract

The application discloses an education resource recommendation method fusing review text and review object content, first, a pre-training model XLNET of natural language processing is used to mine and encode the text, and shallow and deep feature vectors of users and education resources are obtained; second, interest migration of a user is learned and predicted by using LSTM and an attention mechanism, and output vectors of a user module and an education resource module are obtained; and finally, a multilayer perceptron is used to model nonlinear interaction behaviors between the user and the education resource, and a predicted score is obtained. Experimental results show that the fusion of the review text and the content data effectively solves problems such as cold start and data sparsity, improves the performance of the recommendation algorithm, more truly reflects the feature preferences of the user, and can further improve the recommendation accuracy.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of education resource recommendation, and particularly relates to an education resource recommendation method fusing review text and content of a review object. BACKGROUND

[0002] With the integration of the Internet and education, various online education platforms have emerged and developed rapidly. These platforms have accumulated a large number of users with their high-quality and massive resources, and "Internet + education" has become an important driving force for the 2.0 era of education informatization. With the rise of online education, more and more learners participate in online learning and post a large number of comments on the platform. These comments record the feedback of users on teaching quality, learning resources, etc., and contain rich user behavior characteristics. By deeply mining these comment information, we can more accurately understand the interests and preferences of users and the comprehensive characteristics of education resources, thereby alleviating the problems of data sparsity and cold start and improving the recommendation accuracy of education resources. Therefore, it has theoretical significance and application value to research the fusion of review text and content of a review object into education resource recommendation. SUMMARY

[0003] The present application solves the problems of existing education resource recommendation algorithms, such as sparse rating matrix data and insufficient text feature extraction, and provides an education resource recommendation method fusing review text and content of a review object.

[0004] To solve the above problems, the present application is realized by the following technical scheme:

[0005] An education resource recommendation method fusing review text and content of a review object, comprising the following steps:

[0006] Step 1, store all education resource data browsed by a user as a user browsing history set, store all comment data of the user on the education resource as a user comment set, store content data of the education resource as an education resource content set, and store all comment data received by the education resource as an education resource comment set;

[0007] Step 2, use an XLNET model to perform feature mining on the user browsing history set and the user comment set of the user respectively, to obtain a browsing history feature vector of the user and a comment feature vector of the user, and splice the browsing history feature vector of the user and the comment feature vector of the user to obtain a feature vector of the user use an XLNET model to perform feature mining on the education resource content set and the education resource comment set of all education resources commented by the user respectively, to obtain an education resource content feature vector and an education resource comment feature vector, and splice the education resource content feature vector and the education resource comment feature vector to obtain a feature vector of the education resource commented by the user

[0008] Step 3: First, extract the user's feature vector. Feature vectors of educational resources reviewed by users The contribution α of educational resource features to user features is obtained by performing a dot product. ui Furthermore, the contribution α of educational resource characteristics to user characteristics was analyzed. ui Introduce user attention weight ω for educational resources ui The data was then normalized, and the normalization result was compared with the feature vectors of the educational resources reviewed by the user. Dot product yields the correlation between educational resource characteristics and user characteristics. Then, the correlation between all educational resource features and user features was determined. The data is fed into a Long Short-Term Memory (LSTM) network to learn and predict user interest transfer phenomena, thus obtaining the user's final feature representation.

[0009] Step 4: Use the XLNET model to perform feature mining on the educational resource content set and the educational resource comment set, respectively, to obtain the educational resource content feature vectors. and the feature vector of educational resource reviews

[0010] Step 5: First, generate the feature vector of the educational resource content. With the feature vector of educational resource reviews The dot product is used to obtain the contribution β of educational resource reviews to the content of educational resources. v Furthermore, the contribution of educational resource reviews to the content of educational resources (β) is analyzed. v Introducing attention weight ω of educational resources v Then, normalization was performed, and the normalization result was compared with the feature vector of educational resource reviews. Dot product yields the correlation between educational resource reviews and educational resource content. Then, the correlation between all educational resource reviews and the content of the educational resources was determined. The data are fed into a Long Short-Term Memory (LSTM) network to analyze the changes in the characteristics of learning and educational resources over time, resulting in a final representation of the educational resources' features.

[0011] Step 6: Represent the user's final characteristics and the final characteristics of educational resources The data are fed into a multilayer perceptron for feature interaction, and then linearly combined through a fully connected layer to obtain a predicted rating for the educational resource. Based on the predicted rating, educational resources are recommended to the user.

[0012] In the above method, u = 1, 2, …, d, d is the total number of users; i = 1, 2, …, n, n is the number of educational resources reviewed by user u; v = 1, 2, …, e, e is the total number of educational resources.

[0013] In the above step 6, the multi-layer perception network is composed of one full connection layer and three multi-layer perception layers.

[0014] In the above steps 3 and 5, the softmax function is used for normalization processing.

[0015] Compared with the prior art, the present application proposes an educational resource recommendation method based on fusion of review text and content of review object (DTMR, Deep Text Mining based Recommendation). First, the pre-training model XLNET of natural language processing is used to mine and encode the text, and the shallow and deep feature vectors of the user and the educational resource are obtained; second, the interest transfer generated by the user is learned and predicted by using the LSTM and attention mechanism, and the output vectors of the user module and the educational resource module are obtained; finally, the multi-layer perception machine is used to model the nonlinear interaction behavior between the user and the educational resource, and the predicted score is obtained. The experimental results show that the fusion of review text and content data effectively solves the problems of cold start and data sparseness, improves the performance of the recommendation algorithm, more truly reflects the feature preferences of the user, and can further improve the recommendation accuracy. BRIEF DESCRIPTION OF DRAWINGS

[0016] Figure 1 It is a schematic diagram of an educational resource recommendation method based on fusion of review text and content of review object.

[0017] Figure 2 It is a graph showing the influence of the number of hidden factors on the recommendation performance, (a) is the influence of the number of hidden factors on MAE under the Toys_and_Games data set, (b) is the influence of the number of hidden factors on RMSE under the Toys_and_Games data set, (c) is the influence of the number of hidden factors on MAE under the Movies_and_TV data set, and (d) is the influence of the number of hidden factors on RMSE under the Movies_and_TV data set.

[0018] Figure 3 It is a graph showing the influence of the number of hidden factors on the recommendation performance, (a) is the influence of the number of hidden factors on MAE under the Toys_and_Games data set, (b) is the influence of the number of hidden factors on RMSE under the Toys_and_Games data set, (c) is the influence of the number of hidden factors on MAE under the Movies_and_TV data set, and (d) is the influence of the number of hidden factors on RMSE under the Movies_and_TV data set. DETAILED DESCRIPTION

[0019] In order to make the purpose, technical scheme and advantages of the present application clearer and more apparent, the present application will be further described in detail below with specific examples.

[0020] A method for recommending educational resources that integrates comment text with the content of the comment object, such as... Figure 1 As shown, the specific steps include the following:

[0021] Step 1: Store all educational resource data that the user has browsed as the user's browsing history set; store all user comments on educational resources as the user's comment set; store the content data of educational resources as the educational resource content set; and store all comments received by the educational resources as the educational resource comment set.

[0022] Step 2: Use the XLNET model to perform feature mining on the user's browsing history and comment sets to obtain the user's browsing history feature vector and comment feature vector. Then, concatenate the user's browsing history feature vector and comment feature vector to obtain the user's feature vector. The XLNET model is used to perform feature mining on the content set and comment set of all educational resources commented by users, obtaining feature vectors for educational resource content and comments. These feature vectors are then concatenated to obtain the feature vector of the educational resources commented by the user.

[0023] Because the XLNET model contains a Transformer structure, which can dynamically assign weights to different stages of long text sequences, it can extract features from the context and improve recommendation performance.

[0024] For a user u (u = 1, 2, ..., d), their browsing history is represented by H. uk Let (k = 1, 2, ..., p) represent the range of numbers, and let R represent the number of comments he made. ug Let (g = 1, 2, ..., q) represent the total number of users, p be the number of educational resources viewed by user u, and q be the number of comments posted by user u. First, we store user u's browsing history H... uk After inputting into the XLNET model, deep features of each browsing history are obtained, and the deep features of all browsing histories are concatenated to obtain the user's browsing history feature vector. At the same time, all comments R of user u will be... ul After inputting the data into the XLNET model, deep features of each comment are obtained, and the deep features of all comments are concatenated to obtain the user's comment feature vector. Next, the user's browsing history feature vector and user comment feature vectors The user's feature vector is obtained by concatenation.

[0025]

[0026]

[0027]

[0028] For a user u (u = 1, 2, …, d), the educational resources he has commented on are represented by I ui (i = 1, 2, …, n), the content of the educational resource I ui is represented by C i , the comments received by the educational resource I ui are represented by R ij (j = 1, 2, …, m), where d is the total number of users, n is the number of educational resources commented on by user u, and m is the number of comments received by educational resource I ui . First, the content C i of the educational resource is input into the XLNET model to obtain the educational resource content feature vector At the same time, each comment R ij of the educational resource is input into the XLNET model to obtain the deep features of each comment, and the deep features of all comments are spliced to obtain the educational resource comment feature vector Then, the educational resource content feature vector and the educational resource comment feature vector are spliced to obtain the feature vector of the educational resource

[0029]

[0030]

[0031]

[0032] Step 3, first multiply the user feature vector and the feature vector of the educational resources commented on by the user to obtain the contribution degree α ui of the educational resource feature to the user feature; then normalize the contribution degree α ui of the educational resource feature to the user feature after introducing the attention weight ω ui of the user to the educational resource, and multiply the normalized result with the feature vector of the educational resources commented on by the user to obtain the correlation degree between the educational resource feature and the user feature After that, all the correlation degrees between the educational resource features and the user features are sent to the long short-term memory network (LSTM) to learn and predict the user interest transfer phenomenon, and the final feature representation of the user is obtained

[0033] Since the user's interest will change over time, inputting the browsing history and review set with time attributes into the LSTM that can process time series data can process the user's interest changes and obtain an output vector that better meets the user's interest.

[0034]

[0035]

[0036]

[0037] wherein α ui represents the contribution of the characteristics of the educational resource i to the characteristics of the user u, represents the correlation of the characteristics of the educational resource i to the characteristics of the user u, and α ui and The greater the value of the above formula, the more relevant the educational resource i is to the user u, and the more interested the user is. Softmax represents the softmax normalization function.

[0038] Step 4, the educational resource content set of the educational resource and the educational resource review set of the educational resource are respectively subjected to feature mining using the XLNET model to obtain the educational resource content feature vector and the educational resource review feature vector

[0039] For an educational resource v (v = 1, 2,..., e), all its content is represented by C v , and the reviews it receives are represented by R vj (j = 1, 2,..., m), where e is the total number of educational resources, and m is the number of reviews received by the educational resource v. The content C v of the educational resource v is input into the XLNET model to obtain the educational resource content feature vector At the same time, each review R vj of the educational resource is input into the XLNET model to obtain the deep features of each review, and the deep features of all reviews are spliced to obtain the educational resource review feature vector

[0040]

[0041]

[0042] Step 5, the educational resource content feature vector is first combined with the educational resource review feature vector Point multiplication is performed to obtain the contribution degree β of the education resource review to the content of the education resource v ; the contribution degree β of the education resource review to the content of the education resource is calculated again v The attention weight ω of the education resource is introduced v After normalization processing, the normalized result is combined with the education resource review feature vector Point multiplication is performed to obtain the contribution degree β of the education resource review to the content of the education resource After normalization processing, the normalized result is combined with the education resource review feature vector The contribution degree β of all the education resource reviews to the content of the education resource is sent to a long short-term memory network (LSTM) to learn the feature change of the education resource over time, and the final feature representation of the education resource is obtained

[0043]

[0044]

[0045]

[0046] wherein β v represents the contribution degree of the education resource review to the feature of the education resource, represents the association degree of the education resource review to the feature of the education resource, β v and The greater the value of β l , the more relevant the review of the education resource v is to the feature of the education resource v, and the more accurate the review is in describing the feature of the education resource v. The softmax represents a softmax normalization function.

[0047] Step 6, the final feature representation of the user and the final feature representation of the education resource are sent to a multi-layer perception network (MLP) to interact with the features, and the features are linearly combined through a fully connected layer to obtain a predicted score of the user for the education resource, and the education resource is recommended to the user based on the predicted score.

[0048] In the multi-layer perception network, the first layer is a fully connected layer, which is connected to the education resource feature and the user feature to realize the first feature interaction; the subsequent layers are at least one multi-layer perception layer, in the preferred embodiment of the present application, three multi-layer perception layers are stacked to realize the subsequent multiple interactions between the user and the education resource features. The multi-layer perception interaction network is as follows:

[0049]

[0050] wherein z0, z1, …, z l are the outputs of the 0th, 1st, …, lth layers of the multi-layer perception machine, w0, w1, …, wl are the weight matrices of the 0, 1, …, l-th layers of the multi-layer perception, respectively, b0, b1, …, b l are the bias vectors of the 0, 1, …, l-th layers of the multi-layer perception, respectively, and ReLU(*) represents the ReLU activation function.

[0051] After the user features and the education resource features are interacted by the multi-layer perception network, they are finally linearly combined by a fully connected layer to obtain a predicted score of the education resource by the user:

[0052]

[0053] where w u,i is a weight parameter of the fully connected layer, z l is an output of the multi-layer perception, and FC(*) represents the full connection.

[0054] Based on the predicted score of the education resource by the user, the education resources with the top K predicted scores are recommended to the user, and thus the whole recommendation process is completed.

[0055] The present application focuses on solving the serious data sparsity problem when using the rating matrix as the input data, and the data sparsity problem is well avoided by text mining on the review data and the content data of the review objects. Since the education resource content set is used as the input data, the cold start problem is also alleviated. The user interest transfer problem is processed by using the LSTM, which can more truly reflect the change of the user's preference, and the value change of the education resource over time is also considered, and the LSTM is used to process the education resource review data containing time sequence, which avoids the possibility of recommending outdated resources.

[0056] In order to verify the effectiveness of the recommendation method (DTMR model) of the present application, we selected four commonly used data sets in the public data set of Amazon containing a large amount of text content, i.e. Kindle_Store (K_S), Toys_and_Games (T_G), Movies_and_TV (M_T) and Digital_Music (D_M) of similar projects for experiments. These data sets contain different amounts of data, and the Movies_and_TV data set is the largest data set (containing a total of 1697533 data), and the Digital_Music data set is the smallest data set (containing about 64706 data). The basic data statistical information is shown in Table 2.

[0057] Table 2 Basic information of data set

[0058]

[0059] In the above four data sets, the maximum and minimum number of users are 123960 and 5541, the maximum and minimum number of items are 61934 and 3568, and the maximum and minimum number of reviews are 1697533 and 64706, respectively. The average data volume is large, and the data sets are sufficient as verification data sets for the recommendation method.

[0060] Although it can be concluded from the basic information of the data sets that the number of users and items is massive, in fact, the number of items rated by users is only a small part of the total number of items. Therefore, we counted the number of rated items, as shown in Table 3.

[0061] Table 3 Statistics of item rating data

[0062]

[0063] From Table 3, it can be seen that the number of items with less than 20 ratings accounts for about three-quarters, the number of items with less than 15 ratings accounts for one-third, and the number of items with less than 10 ratings accounts for about half. This proves that the rating data is sparse, which has a huge impact on the recommendation method using rating data.

[0064] Researchers use data filling and other methods to alleviate the problem of sparse rating matrix data. They discard rating data and mainly replace rating data with review text and item content data to alleviate data sparsity. Therefore, we counted the number of words in the project reviews, as shown in Table 4.

[0065] Table 4 Statistics of item review data

[0066]

[0067]

[0068] Among them, R / U (Reviews / User) represents the average number of reviews per user, W / U (Word / User) represents the average number of words contained in each user's multiple reviews, R / I (Reviews / Item) represents the average number of reviews per item, and W / I (Word / Item) represents the average number of words contained in each item's multiple reviews. From the information in Table 4, it can be seen that the number of reviews containing more than 70 words in the four data sets is more than 80%, and the number of reviews containing more than 90 words is about 70%. At the same time, the four indicators R / U, W / U, R / I and W / I show that the comments made by most users and the comments received by items contain a certain number of words, so a large number of user and item features can be mined from the comments in the data sets.

[0069] To prove whether the combination of review text and review object content can improve the impact of sparse rating data, the recommendation method of the application is compared with the recommendation method that also fuses review text, verifying whether the DTMR recommendation method can further reduce the prediction error by fusing the review object content and using the natural language preprocessing model. For this purpose, we selected LFM, HFT, DTMF, DeepCoNN, NARRE and DeepCLFM as the baseline recommendation methods of DTMR for comparison experiments.

[0070] Table 5 Method Comparison

[0071]

[0072] The NMF recommendation method and the LFM recommendation method only use the rating data matrix as the model input. The HFT recommendation method and the DTMF recommendation method use the review text as the model input on the basis of the rating data matrix. The DeepCoNN recommendation method uses the review text as the input and incorporates a deep learning model. The NARRE recommendation method and the DeepCLFM recommendation method add the rating data to the model on the basis of the DeepCoNN recommendation method and introduce an attention mechanism. Due to the extremely sparse rating data in the project, the DTMR recommendation method does not use the rating matrix as the input, but adds the content text of the project to the model to use natural language processing to deeply mine the text, which has better effect than using the rating matrix. In addition, the DTMR recommendation method also uses LSTM to process the interest transfer generated by the user and introduces an attention mechanism, achieving better recommendation effect.

[0073] (1) LFM: It is a classic matrix factorization model, which is an implicit semantic model that predicts the rating of unknown items by matrix factorization of rating data.

[0074] (2) HFT: This model explores the influence of user reviews and project reviews on recommendation performance in the experimental part, proving that using the project review set can better mine features and enhance recommendation performance. Therefore, the project review is input into the model, and the recommendation result is obtained by combining topic mining and matrix factorization. This is a classic model that combines rating matrix and review set.

[0075] (3) DTMF: On the basis of the HFT model, the user review set and the project review set are input into the model for topic mining, and then the predicted rating is obtained by combining the topic distribution with the rating matrix decomposition, similar to HFT.

[0076] (4) DeepCoNN: It is the first deep collaborative neural network model that combines user reviews and item reviews with rating matrix. Compared with traditional recommendation models, it has better recommendation performance after using deep learning as a feature extraction method.

[0077] (5) NARRE: It introduces attention mechanism on the basis of DeepCoNN model, uses attention mechanism to filter the review set, selects the reviews and rating data that contribute to users and items, and therefore the recommendation performance is further enhanced.

[0078] (6) DeepCLFM: It is a deep interactive latent factor model that uses rating data and item review data at the same time, and combines the shallow features and deep features contained in the text. Since the interaction information between users and items is strengthened, the recommendation performance is further improved.

[0079] The DTMR recommendation method proposed in the application is compared with the selected six baseline recommendation methods, and the experimental results are shown in Tables 6 and 7.

[0080] Table 6 Comparison of MAE results of each model on the data set

[0081]

[0082] Table 7 Comparison of RMSE results of each model on the data set

[0083]

[0084] From the analysis of the experimental results, the following conclusions can be obtained:

[0085] (1) In the comparison experiment, the matrix decomposition model LFM only uses rating matrix data, and the MAE and RMSE on the four data sets are the largest, so it obtains the lowest recommendation effect in all comparison experiments. It is proved that since the ratings between users and items are sparse, the recommendation model using only rating data as input is limited.

[0086] (2) From the results, the recommendation effect of HFT and DTMF two models is better than LFM. This result shows that introducing review text into the recommendation model can better perfect and supplement the feature information of the item, and using these more abundant information can more accurately locate the user's preferences and characteristics to optimize the model. However, the two models of HFT and DTMF both use topic mining algorithm to extract the topic distribution of the text, and do not realize the influence of word order and context. Therefore, this kind of model still has limitations.

[0087] (3) DeepCoNN model uses CNN to extract features from reviews, which effectively handles the context of the text and can better mine the information of users and items contained in the review text, so it has better recommendation effect on all four data sets. The performance of NARRE model is better than that of DeepCoNN model, which can prove that by introducing the attention mechanism, it can find the high contribution of features in the review, so it can improve the recommendation effect of the model. But through the text processing of CNN, it cannot effectively capture the long-distance features in the review set, so their recommendation performance is slightly lower than that of the DTMR recommendation method.

[0088] (4) The performance of DeepCLFM model is better than all other baseline models in this experiment. DeepCLFM uses a pre-trained BERT model to deal with the context feature problem, and combines a bidirectional GRU component to summarize the information of each review in the review set. Therefore, compared with the NARRE model, the DeepCLFM model has better recommendation performance.

[0089] (5) The DTMR model obtains the optimal MAE and RMSE values on the four data sets. First, compared with the traditional recommendation model using the rating matrix, the DTMR recommendation method does not use the rating matrix as input, but adds the content text of the item to the model to use natural language processing to deeply mine the text, so as to reduce the limitation of data sparsity on the recommendation model. It has better effect than using rating matrix; second, the model uses the currently advanced pre-training model XLNET for natural language processing of text, which can better obtain long-distance features in review text and review object content, and improve the globality of feature acquisition; third, because the traditional recommendation model does not specially process the user interest transfer, the DTMR model uses LSTM to process the user interest transfer generated, and introduces the attention mechanism, which achieves better recommendation effect; finally, in the feature interaction stage, the multi-layer perceptron is used for nonlinear interaction calculation of user features and item features, which can enhance the representation ability of user features and item features compared with the traditional direct dot product method, and improve the recommendation performance.

[0090] The following shows the influence of selecting different number of latent factors on the recommendation performance of each recommendation algorithm on different data sets. The number of latent factors is equivalent to the length of the user and item feature vectors in traditional rating matrix recommendation models and review text-based models. Because HFT and DTMF are topic models, the number of latent factors is equivalent to the number of topics K, which is not sensitive to the number of latent factors, so only the remaining four models and the DTMR model are experimented, as shown in Figure 2 Figure 2 ​It can be seen that the DTMR recommendation method can combine the latent factors of users and items in the comment text and the content of the comment object, and perform deep mining, thus exhibiting excellent recommendation performance under various parameter settings. It can also be observed that the MAE and RMSE values ​​of the matrix factorization-based LFM model gradually increase as the number of latent factors increases. This may be due to overfitting caused by an excessive number of latent factors, while other deep learning-based models do not show obvious overfitting. Comparison on two datasets reveals that the errors of the NARRE, DeepCLFM, and DTMR models are relatively stable on the Movies_and_TV dataset. This may be due to the simpler text structure and larger sample size of Movies_and_TV.

[0091] To investigate the impact of multilayer perceptrons (MLPs) on model recommendation performance, the effect of the number of hidden layers in the MLP on model performance was verified. The number of hidden layers was set to {1, 2, 3, 4}, and the experimental results are as follows. Figure 3 As shown. (Through) Figure 3 It can be seen that, generally, increasing the number of hidden layers in a perceptron improves the model's recommendation performance on the same dataset, indicating that multilayer perceptrons can better learn complex nonlinear interactions between users and items. However, increasing the number of hidden layers too much can lead to overfitting and reduce the model's ability to adapt to new data. Choosing an appropriate number of hidden layers is crucial and will enhance the performance of the recommendation model. Based on the analysis of experimental results, we believe that setting the number of hidden layers in a multilayer perceptron to 3 is most suitable.

[0092] It should be noted that although the embodiments described above are illustrative, they are not intended to limit the invention. Therefore, the invention is not limited to the specific embodiments described above. Any other embodiments obtained by those skilled in the art under the guidance of this invention without departing from its principles are considered to be within the protection scope of this invention.

Claims

1. A method for recommending educational resources that integrates comment text with the content of the comment object, characterized in that, The steps include the following: Step 1: Store all educational resource data that the user has viewed as the user's browsing history set; store all user comments on educational resources as the user's comment set; store the content data of educational resources as the educational resource content set; store all comments received by the educational resources as the educational resource comment set; Step 2: Perform feature mining on the user's browsing history and comment sets using the XLNET model to obtain the user's browsing history feature vector and comment feature vector, respectively. Then, concatenate the user's browsing history feature vector and comment feature vector to obtain the user's feature vector. The XLNET model is used to perform feature mining on both the content set and comment set of all educational resources commented by users, resulting in feature vectors for the content and comments of educational resources. These two feature vectors are then concatenated to obtain the feature vector of the educational resources commented by the user. Step 3: First, extract the user's feature vector. Feature vectors of educational resources reviewed by users The contribution α of educational resource features to user features is obtained by performing a dot product. ui Furthermore, the contribution α of educational resource characteristics to user characteristics was analyzed. ui Introduce user attention weight ω for educational resources ui The data was then normalized, and the normalization result was compared with the feature vectors of the educational resources reviewed by the user. Dot product yields the correlation between educational resource characteristics and user characteristics. Then, the correlation between all educational resource features and user features was determined. The data is fed into a Long Short-Term Memory (LSTM) network to learn and predict user interest transfer phenomena, thus obtaining the user's final feature representation. Step 4: Use the XLNET model to perform feature mining on the educational resource content set and the educational resource comment set, respectively, to obtain the educational resource content feature vectors. and the feature vector of educational resource reviews Step 5: First, generate the feature vector of the educational resource content. With the feature vector of educational resource reviews The dot product is used to obtain the contribution β of educational resource reviews to the content of educational resources. v Furthermore, the contribution of educational resource reviews to the content of educational resources (β) is analyzed. v Introducing attention weight ω of educational resources v Then, normalization was performed, and the normalization result was compared with the feature vector of educational resource reviews. Dot product yields the correlation between educational resource reviews and educational resource content. Then, the correlation between all educational resource reviews and the content of the educational resources was determined. The data are fed into a Long Short-Term Memory (LSTM) network to analyze the changes in the characteristics of learning and educational resources over time, resulting in a final representation of the educational resources' features. Step 6: Represent the user's final characteristics and the final characteristics of educational resources The data are fed into a multilayer perceptron for feature interaction, and then linearly combined through a fully connected layer to obtain the user's predicted rating for the educational resource. Based on the predicted rating, educational resources are recommended to the user.

2. The educational resource recommendation method that integrates comment text and comment object content according to claim 1, characterized in that, u =1,2,…,d, where d is the total number of users; i =1,2,…,n, where n is the number of educational resources commented by user u; v =1,2,…,e, where e is the total number of educational resources.

3. The educational resource recommendation method that integrates comment text and comment object content according to claim 1, characterized in that, In step 6, the multilayer sensing network consists of one fully connected layer and three multilayer sensing layers.

4. The educational resource recommendation method that integrates comment text and comment object content according to claim 1, characterized in that, In steps 3 and 5, the softmax function is used for normalization.

Images