Score prediction method based on end-to-end text filtering

Abstract

The application discloses a scoring prediction method based on end-to-end text filtering, comprising: obtaining a to-be-predicted review text, determining a user and a commodity corresponding to the review text; obtaining at least one historical review text of the user and at least one historical review text of the commodity; obtaining a word group level convolution feature vector corresponding to the review text and a word group level convolution feature vector corresponding to the historical review text; obtaining a feature vector corresponding to the review text and a feature vector corresponding to the historical review text; determining the similarity between the review text and the historical review text according to the feature vectors; filtering the historical review text based on the similarity, and calculating a similarity text feature vector according to the filtered historical review text; fusing the feature vector of the review text and the similarity text feature vector to obtain a global text feature vector; and determining the predicted score of the review text according to the global text feature vector. The application can improve the scoring prediction performance and precision of the review text.

Description

Technical Field

[0001] This invention relates to the field of Internet information classification technology, and in particular to a rating prediction method based on end-to-end text filtering. Background Technology

[0002] Review text rating prediction refers to the semantic analysis of user reviews on online platforms to predict product ratings, serving as a measure of user satisfaction. Review text rating prediction has wide applications in real life. Specifically, through review text rating prediction, businesses can understand consumer feedback on products and services to improve them and increase sales. Online platforms can use review text rating prediction to uncover consumer interests and preferences regarding product attributes, providing precise personalized recommendations.

[0003] However, existing comment text rating prediction methods mainly rely on mining the semantic information of comment texts to establish a mapping from comment texts to ratings. This method is limited by the local semantics of individual texts and is easily affected by the fluctuation of text semantics, resulting in poor prediction performance and accuracy. Summary of the Invention

[0004] To address some or all of the technical problems existing in the prior art, this invention provides a rating prediction method based on end-to-end text filtering.

[0005] The technical solution of the present invention is as follows:

[0006] A rating prediction method based on end-to-end text filtering is provided, the method comprising:

[0007] Obtain the comment text to be predicted, and determine the user and product corresponding to the comment text;

[0008] Obtain at least one historical review text of the user and at least one historical review text of the product;

[0009] A pre-trained convolutional neural network is used to perform convolution processing on the comment text, the user's historical comment text, and the product's historical comment text, respectively, to obtain the word-level convolutional feature vector corresponding to the comment text and the word-level convolutional feature vector corresponding to the historical comment text;

[0010] The word-level convolutional feature vectors corresponding to the comment text and the historical comment text are processed using a pre-trained first attention mechanism model to obtain the feature vectors corresponding to the comment text and the historical comment text.

[0011] Based on the feature vector, the similarity between the comment text and the historical comment text is determined;

[0012] Based on the similarity, the historical comment texts are filtered, and a similarity text feature vector is calculated based on the filtered historical comment texts;

[0013] The feature vectors of the comment text and the similar texts are fused to obtain a global text feature vector;

[0014] The predicted rating of the comment text is determined based on the global text feature vector.

[0015] In some possible implementations, a pre-trained convolutional neural network is used to perform convolution processing on the comment text to obtain the word-level convolutional feature vectors corresponding to the comment text, further including:

[0016] Identify the constituent words of the comment text;

[0017] The word-level convolutional features of each word are obtained using a pre-trained convolutional neural network;

[0018] Based on the word-level convolutional features of each word, determine the word-level convolutional feature vector corresponding to the comment text.

[0019] In some possible implementations, the feature vector corresponding to the comment text is obtained using the following formula:

[0020]

[0021] The feature vector corresponding to the user's historical comment text is obtained using the following formula:

[0022]

[0023] The feature vector corresponding to the historical review text of the product is obtained using the following formula:

[0024]

[0025] Where W and b represent the parameters of the first attention mechanism model, and tanh represents the nonlinear transformation function. c represents the word-level convolutional feature vector corresponding to the comment text. i Let represent the convolutional feature vector of the i-th word in the comment text, n represent the number of words in the comment text, and d represent the feature vector corresponding to the comment text. c represents the word-level convolutional feature vector corresponding to the user's historical comment text. U,iLet represent the convolutional feature vector of the i-th word in the user's historical comment text, u represent the number of words in the user's historical comment text, and d represent the number of words in the text. U This represents the feature vector corresponding to the user's historical comment text. c represents the phrase-level convolutional feature vector corresponding to the historical review text of a product. P,i Let d represent the convolutional feature vector of the i-th word in the historical review text of the product, p represent the number of words in the historical review text of the product, and d represent the number of words in the historical review text of the product. P This represents the feature vector corresponding to the historical review text of a product.

[0026] In some possible implementations, the similarity between the comment text and the historical comment text is calculated using the following formula:

[0027] α k =v T tanh(W1d+W2d ′ k +b)

[0028] Where v, W1, W2, and b represent the parameters of the second attention mechanism model, and α k This indicates the difference between the comment text and the historical comment text (D). ′ k The similarity between them, d ′ k Represents historical commentary text D ′ k The corresponding eigenvector, D ′ k ∈D U ∪ P D U D represents the user's historical comment text. P This refers to the historical review text of a product.

[0029] In some possible implementations, the historical comment text is filtered based on the similarity, and a similarity text feature vector is calculated based on the filtered historical comment text, including:

[0030] Based on the similarity and the preset filtering threshold, an end-to-end text filtering function for filtering the historical comment text is constructed using nonlinear transformation;

[0031] The similarity text feature vector is calculated based on the similarity, the end-to-end text filtering function, and the feature vector corresponding to the historical comment text.

[0032] In some possible implementations, the end-to-end text filtering function is represented as:

[0033]

[0034] The similar text feature vectors are calculated using the following formula:

[0035]

[0036] Where η represents the scaling factor for noise filtering, θ represents the preset filtering threshold, e represents the natural constant, and d ′ Represents the feature vector of similar texts.

[0037] In some possible implementations, a pre-trained gated neural network is used to fuse the feature vector of the comment text and the feature vector of the similar text to obtain the global text feature vector.

[0038] In some possible implementations, the global text feature vector is obtained using the following formula:

[0039]

[0040] Among them, W g and b g This represents the parameters of the gated neural network, where softmax is the normalized nonlinear transformation function. d g This represents the global text feature vector that takes into account end-to-end text filtering.

[0041] In some possible implementations, a pre-trained feedforward neural network is used to process the global text feature vector to obtain a predicted score for the comment text.

[0042] In some possible implementations, the predicted score of the comment text is obtained using the following formula:

[0043]

[0044] Among them, W y and b y Represents the parameters of the feedforward neural network. Represents the predicted rating vector. Indicates taking the prediction vector The dimension corresponding to the largest value. This indicates the predicted score.

[0045] The main advantages of the technical solution of this invention are as follows:

[0046] The end-to-end text filtering-based rating prediction method of the present invention significantly improves the rating prediction performance and accuracy of comment texts by mining and filtering similar texts of the comment texts, and then predicting the rating of the comment texts based on the filtered similar texts, thus obtaining more accurate predicted ratings. Attached Figure Description

[0047] The accompanying drawings, which are included to provide a further understanding of embodiments of the invention and constitute a part of this invention, illustrate exemplary embodiments of the invention and, together with their description, serve to explain the invention and do not constitute an undue limitation thereof. In the drawings:

[0048] Figure 1 This is a flowchart of a rating prediction method based on end-to-end text filtering according to an embodiment of the present invention. Detailed Implementation

[0049] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below in conjunction with specific embodiments and corresponding drawings. Obviously, the described embodiments are only a part of the embodiments of this invention, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0050] The technical solutions provided by the embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0051] refer to Figure 1 An embodiment of the present invention provides a rating prediction method based on end-to-end text filtering, the method comprising the following steps S1-S8:

[0052] Step S1: Obtain the comment text to be predicted and determine the user and product corresponding to the comment text.

[0053] Specifically, the text of the review to be predicted is obtained, and the user and product corresponding to the review text are identified.

[0054] Step S2: Obtain at least one historical review text from the user and at least one historical review text from the product.

[0055] In one embodiment of the present invention, the user's historical review text represents the existing review text of all products by the user, including the product corresponding to the review text to be predicted, and the product's historical review text represents the existing review text of the product by all users, including the user corresponding to the review text to be predicted.

[0056] In one embodiment of the present invention, a certain number of historical review texts are randomly selected from existing user historical review texts and product historical review texts for subsequent processing. The number of historical review texts selected is determined based on actual circumstances; generally, the more historical review texts used in the prediction process, the higher the prediction accuracy, but the higher the processing cost and time.

[0057] Step S3: Use a pre-trained convolutional neural network to perform convolution processing on the comment text, the user's historical comment text, and the product's historical comment text, respectively, to obtain the word-level convolutional feature vectors corresponding to the comment text and the word-level convolutional feature vectors corresponding to the historical comment text.

[0058] In one embodiment of the present invention, a pre-trained convolutional neural network is used to perform convolution processing on the comment text to obtain the word-level convolutional feature vector corresponding to the comment text, further comprising:

[0059] Identify the constituent words of the comment text;

[0060] The word-level convolutional features of each word are obtained using a pre-trained convolutional neural network;

[0061] Based on the word-level convolutional features of each word, determine the word-level convolutional feature vector corresponding to the comment text.

[0062] The comment text includes the comment text to be predicted, the user's historical comment text, and the product's historical comment text.

[0063] Specifically, let D be the text of the comment to be predicted, and let D be the text of the user's historical comments. U ={D U，1 , ..., D U，s The product's historical review text is represented as D. P ={D P，1 , ..., D P，t}, D U，s D represents the s-th historical comment text of a user. P，t This represents the t-th historical review text for a product.

[0064] Furthermore, taking the comment text to be predicted as an example, suppose: a comment text to be predicted consists of n words, represented as... w i This represents the i-th word.

[0065] By embedding retrieval operations, each of the n words is mapped to one of the n vectors. Then, a pre-trained convolutional neural network is used to obtain the w of each word. i Phrase-level convolutional features.

[0066] Specifically, the word w is obtained through a convolutional neural network. i The phrase-level convolutional features are represented as follows:

[0067]

[0068] Among them, c i，f The word w iThe word-level convolutional features, where f represents the nonlinear transformation function, W f Denotes the convolution kernel, b f This represents the bias coefficient. The convolution operation selects the word w. i A total of l words are convolved on both sides to form corresponding convolutional features c. i，f .

[0069] Furthermore, assuming the convolutional neural network uses k convolutional kernels for convolution operations, then for each word, a k-dimensional convolutional feature vector is formed, specifically represented as:

[0070]

[0071] Among them, c i The word w i k-dimensional convolutional feature vectors The word w represents the word obtained by the k-th convolution kernel. i Phrase-level convolutional features.

[0072] In one embodiment of the present invention, for the comment text to be predicted, each historical comment text of the user, and each historical comment text of the product, according to the above processing, the word-level convolutional feature vector corresponding to the comment text to be predicted and the word-level convolutional feature vector corresponding to each historical comment text can be obtained, which can effectively capture the word-level information of the comment text.

[0073] The parameters of the convolutional neural network are determined through automatic learning and updating during model training.

[0074] Step S4: Use the pre-trained first attention mechanism model to process the word-level convolutional feature vectors corresponding to the comment text and the word-level convolutional feature vectors corresponding to the historical comment texts, respectively, to obtain the feature vectors corresponding to the comment texts and the feature vectors corresponding to the historical comment texts.

[0075] In one embodiment of the present invention, a pre-trained first attention mechanism model is used to process the word-level convolutional feature vector corresponding to the comment text to be predicted, thereby obtaining the feature vector corresponding to the comment text to be predicted; the pre-trained first attention mechanism model is used to process the word-level convolutional feature vector corresponding to the user's historical comment text, thereby obtaining the feature vector corresponding to the user's historical comment text; the pre-trained first attention mechanism model is used to process the word-level convolutional feature vector corresponding to the product's historical comment text, thereby obtaining the feature vector corresponding to the product's historical comment text.

[0076] Specifically, the settings are as follows: c represents the word-level convolutional feature vector corresponding to the comment text to be predicted. iLet represent the convolutional feature vector of the i-th word in the comment text to be predicted, and n represent the number of words in the comment text to be predicted. c represents the word-level convolutional feature vector corresponding to the user's historical comment text. U，i Let represent the convolutional feature vector of the i-th word in the user's historical comment text, and u represent the number of words in the user's historical comment text. c represents the phrase-level convolutional feature vector corresponding to the historical review text of a product. P，i Let represent the convolutional feature vector of the i-th word in the historical review text of the product, and D represent the number of words in the historical review text of the product.

[0077] The first attention mechanism model processes the word-level convolutional feature vectors corresponding to the comment text to be predicted, as follows:

[0078]

[0079] The word-level convolutional feature vectors corresponding to the user's historical comment text are processed using the first attention mechanism model and represented as follows:

[0080]

[0081] The word-level convolutional feature vectors corresponding to the historical review text of a product are processed using the first attention mechanism model and represented as follows:

[0082]

[0083] Where W and b represent the parameters of the first attention mechanism model, and tanh represents the nonlinear transformation function. d represents the feature vector corresponding to the comment text to be predicted. U d represents the feature vector corresponding to the user's historical comment text. p This represents the feature vector corresponding to the historical review text of the product. The parameters of the first attention mechanism model are determined through automatic learning and updating during model training.

[0084] In one embodiment of the present invention, for the comment text to be predicted, each historical comment text of a user, and each historical comment text of a product, the feature vectors corresponding to the comment text to be predicted, each historical comment text of a user, and each historical comment text of a product are obtained through the above processing.

[0085] Step S5: Determine the similarity between the comment text and historical comment texts based on the feature vector.

[0086] In one embodiment of the present invention, a pre-trained second attention mechanism model is used to process the feature vectors corresponding to the comment text to be predicted and the feature vectors corresponding to the historical comment texts to determine the similarity between the comment text to be predicted and each historical comment text.

[0087] Specifically, the similarity between the comment text and historical comment texts is calculated using the following formula:

[0088] α k =v T tanh(W1d+W2d′ k +b)

[0089] Where v, W1, W2, and b represent the parameters of the second attention mechanism model, tanh represents the nonlinear transformation function, and α k This represents the comment text D and the historical comment text D′. k The similarity between them, where d represents the feature vector corresponding to comment text D, and d′ k Represents the historical commentary text D′ k The corresponding eigenvector, D′ k ∈D U UD p The parameters of the second attention mechanism model are determined through automatic learning and updating during model training.

[0090] Considering that a selected user's historical review text may be the same as a product's historical review text, in one embodiment of the present invention, for the same historical review text, only the similarity between the review text to be predicted and one of the historical review texts is calculated.

[0091] Step S6: Based on similarity, filter the historical comment texts and calculate the similarity text feature vector based on the filtered historical comment texts.

[0092] In one embodiment of the present invention, historical comment texts are filtered based on similarity, and a similarity text feature vector is calculated based on the filtered historical comment texts, further including:

[0093] Based on similarity and a preset filtering threshold, an end-to-end text filtering function for filtering historical comment text is constructed using nonlinear transformation.

[0094] Calculate similar text feature vectors based on similarity, end-to-end text filtering functions, and feature vectors corresponding to historical comment texts.

[0095] Specifically, in one embodiment of the present invention, the constructed end-to-end text filtering function is expressed as follows:

[0096]

[0097] Where η represents the scaling factor for noise filtering, η is a model hyperparameter, a larger η indicates more thorough filtering, η→0 indicates no filtering of historical comment text, α k This represents the similarity between the comment text and historical comment text, where θ represents the preset filtering threshold and e represents a natural constant. The preset filtering threshold is set according to the specific circumstances.

[0098] In one embodiment of the present invention, by constructing an end-to-end text filtering function to filter historical comment text to remove noisy text, the model can be made end-to-end differentiable for filtering.

[0099] Furthermore, in one embodiment of the present invention, the similarity text feature vector is calculated using the following formula:

[0100]

[0101] Where d′ represents the similarity text feature vector, d U D represents the user's historical comment text. U eigenvectors, d P D represents the historical review text of the product. P eigenvectors.

[0102] Step S7: Fuse the feature vector of the comment text and the feature vector of similar texts to obtain the global text feature vector.

[0103] In one embodiment of the present invention, a pre-trained gated neural network is used to fuse the feature vector of the comment text to be predicted and the feature vector of similar texts to obtain a global text feature vector.

[0104] Specifically, the feature vectors of the comment text and the feature vectors of similar texts are fused using a gated neural network and represented as follows:

[0105]

[0106] Among them, W g and b g This represents the parameters of the gated neural network, where softmax is the normalized nonlinear transformation function. d g This represents a global text feature vector that considers end-to-end text filtering. The physical meaning of a gated neural network is to automatically determine the weights of two different input vectors, and then fuse the two input vectors through a weighted sum operation. The parameters of the gated neural network are automatically learned and updated during model training.

[0107] Step S8: Determine the predicted score for the comment text based on the global text feature vector.

[0108] In one embodiment of the present invention, a pre-trained feedforward neural network is used to process the global text feature vector to obtain the predicted score of the comment text to be predicted.

[0109] Specifically, the processing of the global text feature vector using a feedforward neural network is represented as follows:

[0110]

[0111] Among them, W y and b y This represents the parameters of the feedforward neural network, where softmax is the normalized nonlinear transformation function. Represents the predicted rating vector. Indicates taking the prediction vector The dimension corresponding to the largest value. This represents the predicted score. The parameters of the feedforward neural network are determined through automatic learning and updating during model training.

[0112] In one embodiment of the present invention, the predicted scoring vector The output is used to represent the rating categories of the comment text. The number of rating categories for the comment text is determined according to the actual situation, for example, it can be 5 or 10 categories. When the number of rating categories is 10, the vector... Given a vector with dimension 10, take the vector... The dimension with the largest median value is used as the corresponding score for prediction.

[0113] Furthermore, in one embodiment of the present invention, the parameters of the convolutional neural network, the parameters of each attention mechanism model, the parameters of the gating neural network, and the parameters of the feedforward neural network are all updated and optimized through model training.

[0114] Specifically, in one embodiment of the present invention, the convolutional neural network, various attention mechanism models, gating neural network, and feedforward neural network are trained through the following steps:

[0115] Obtain the training dataset, which includes comment texts, the users and products corresponding to the comment texts, the users' historical comment texts, the products' historical comment texts, and the actual ratings of the comment texts;

[0116] Using the comment texts in the training data, the historical comment texts of users, and the historical comment texts of products as inputs, and the actual ratings of the comment texts in the training data as outputs, a convolutional neural network, various attention mechanism models, a gated neural network, and a feedforward neural network are trained using the backpropagation algorithm based on a preset objective function.

[0117] Specifically, during model training, the convolutional neural network, each attention mechanism model, the gated neural network, and the feedforward neural network are all set with initialization parameters. Based on the set initialization parameters and the aforementioned specific data processing procedures and formulas, a corresponding predicted score can be obtained for each input training data. The objective function is calculated based on the actual score and the predicted score. Based on the objective function, the backpropagation algorithm is used to continuously update the parameters of each model until the set iteration stopping condition is reached, thus completing the model training and determining the convolutional neural network, each attention mechanism model, the gated neural network, and the feedforward neural network.

[0118] In one embodiment of the present invention, the objective function is set as follows during model training:

[0119]

[0120] Where L represents the objective function, X represents the training dataset, and r represents the actual score. This indicates the predicted score.

[0121] The objective function defined above represents the error between the predicted score and the actual score. Optimizing this objective function is equivalent to minimizing this error, which is equivalent to improving the model's prediction accuracy.

[0122] Furthermore, in one embodiment of the present invention, the parameters of the model are updated using the following formula:

[0123]

[0124] Where, Θ t+1 Let Θ represent the set of parameters for all models at the (t+1)th iteration. t Let represent the set of parameters for all models at iteration t, Δ[·] represent the optimizer, η represent the learning rate, and Θ represent the set of model parameters. The optimizer can be, for example, Adam or SGD, and the learning rate needs to be preset to control the speed of parameter updates. The models include convolutional neural networks, various attention mechanism models, gated neural networks, and feedforward neural networks.

[0125] An embodiment of the present invention provides a rating prediction method based on end-to-end text filtering. By mining and filtering similar texts of the comment text, and predicting the rating of the comment text based on the filtered similar texts, the rating prediction performance and accuracy of the comment text can be significantly improved, resulting in a more accurate predicted rating.

[0126] It should be noted that, in this document, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Additionally, the terms "front," "back," "left," "right," "upper," and "lower" in this document refer to the placement shown in the accompanying drawings.

[0127] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A rating prediction method based on end-to-end text filtering, characterized in that, include: Obtain the comment text to be predicted, and determine the user and product corresponding to the comment text; Obtain at least one historical review text of the user and at least one historical review text of the product; A pre-trained convolutional neural network is used to perform convolution processing on the comment text, the user's historical comment text, and the product's historical comment text, respectively, to obtain the word-level convolutional feature vector corresponding to the comment text and the word-level convolutional feature vector corresponding to the historical comment text; The word-level convolutional feature vectors corresponding to the comment text and the historical comment text are processed using a pre-trained first attention mechanism model to obtain the feature vectors corresponding to the comment text and the historical comment text. Based on the feature vector, the similarity between the comment text and the historical comment text is determined; Based on the similarity, the historical comment texts are filtered, and a similarity text feature vector is calculated based on the filtered historical comment texts; The feature vectors of the comment text and the similar texts are fused to obtain a global text feature vector; The predicted rating of the comment text is determined based on the global text feature vector.

2. The rating prediction method based on end-to-end text filtering according to claim 1, characterized in that, The comment text is processed by convolutional neural networks pre-trained to obtain word-level convolutional feature vectors corresponding to the comment text, further including: Identify the constituent words of the comment text; The word-level convolutional features of each word are obtained using a pre-trained convolutional neural network; Based on the word-level convolutional features of each word, determine the word-level convolutional feature vector corresponding to the comment text.

3. The rating prediction method based on end-to-end text filtering according to claim 2, characterized in that, The feature vector corresponding to the comment text is obtained using the following formula: The feature vector corresponding to the user's historical comment text is obtained using the following formula: The feature vector corresponding to the historical review text of the product is obtained using the following formula: Where W and b represent the parameters of the first attention mechanism model, and tanh represents the nonlinear transformation function. c represents the word-level convolutional feature vector corresponding to the comment text. i Let represent the convolutional feature vector of the i-th word in the comment text, n represent the number of words in the comment text, and d represent the feature vector corresponding to the comment text. c represents the word-level convolutional feature vector corresponding to the user's historical comment text. U,i Let represent the convolutional feature vector of the i-th word in the user's historical comment text, u represent the number of words in the user's historical comment text, and d represent the number of words in the text. U This represents the feature vector corresponding to the user's historical comment text. c represents the phrase-level convolutional feature vector corresponding to the historical review text of a product. P,i Let d represent the convolutional feature vector of the i-th word in the historical review text of the product, p represent the number of words in the historical review text of the product, and d represent the number of words in the historical review text of the product. P This represents the feature vector corresponding to the historical review text of a product.

4. The rating prediction method based on end-to-end text filtering according to claim 3, characterized in that, The similarity between the comment text and the historical comment text is calculated using the following formula: α k =v T tanh(W1d+W2d ′ k +b) Where v, W1, W2, and b represent the parameters of the second attention mechanism model, and α k This indicates the difference between the comment text and the historical comment text (D). ′ k The similarity between them, d ′ k Represents historical commentary text D ′ k The corresponding eigenvector, D ′ k ∈D U ∪ P D U D represents the user's historical comment text. P This refers to the historical review text of a product.

5. The rating prediction method based on end-to-end text filtering according to claim 4, characterized in that, Based on the similarity, the historical comment text is filtered, and a similarity text feature vector is calculated based on the filtered historical comment text, including: Based on the similarity and the preset filtering threshold, an end-to-end text filtering function for filtering the historical comment text is constructed using nonlinear transformation; The similarity text feature vector is calculated based on the similarity, the end-to-end text filtering function, and the feature vector corresponding to the historical comment text.

6. The rating prediction method based on end-to-end text filtering according to claim 5, characterized in that, The end-to-end text filtering function is expressed as follows: The similar text feature vectors are calculated using the following formula: Where η represents the scaling factor for noise filtering, θ represents the preset filtering threshold, e represents the natural constant, and d ′ Represents the feature vector of similar texts.

7. The rating prediction method based on end-to-end text filtering according to claim 6, characterized in that, The global text feature vector is obtained by fusing the feature vector of the comment text and the feature vector of the similar text using a pre-trained gated neural network.

8. The rating prediction method based on end-to-end text filtering according to claim 7, characterized in that, The global text feature vector is obtained using the following formula: Among them, W g and b g This represents the parameters of the gated neural network, where softmax is the normalized nonlinear transformation function. d g This represents the global text feature vector that takes into account end-to-end text filtering.

9. The rating prediction method based on end-to-end text filtering according to claim 8, characterized in that, The global text feature vector is processed using a pre-trained feedforward neural network to obtain the predicted score of the comment text.

10. The rating prediction method based on end-to-end text filtering according to claim 9, characterized in that, The predicted score for the comment text is obtained using the following formula: Among them, W y and b y Represents the parameters of the feedforward neural network. Represents the predicted rating vector. Indicates taking the prediction vector The dimension corresponding to the largest value. This indicates the predicted score.

Images