Method, device and storage medium for pushing product review data

Abstract

The present application provides a method and related equipment for recommending product review data, which can screen out higher-quality reviews and recommend higher-quality reviews to users. The method includes: preprocessing the target review corresponding to the target product to obtain the word vector and aspect representation vector corresponding to the target review; concatenating the word vector and the aspect representation vector and inputting them into the aspect sentiment analysis model to obtain the aspect sentiment score corresponding to the target review; concatenating the aspect sentiment score and the target review to obtain the target data; determining the user behavior data corresponding to the target review, wherein the user behavior data is the behavior data of the user corresponding to the target review; inputting the target data and the user behavior data into the true and false review analysis model to obtain the reconstruction probability of the target review; and pushing the target review to the user according to the reconstruction probability of the target review.

Description

[Technical field]

[0001] The present application belongs to the field of recommendation, and in particular relates to a method, device and storage medium for pushing product review data. [Background technology]

[0002] Online shopping platforms use review systems to provide users with valuable advice on product purchases, but due to data overload, there are more "low-quality" or "useless reviews" in product reviews, as well as "professional reviewers" who receive cashback for positive reviews. Analyzing and predicting the usefulness of online reviews is of great significance in improving the quality of online review information, assisting consumers in making purchase decisions, and promoting the optimal design of online platforms.

[0003] Currently, when predicting online reviews, most of the feature selection used adopts grammatical features (counting punctuation, function words, part-of-speech tags), semantic features (evaluation tuples, sentiment polarity), genre features (tags of attributes or writing styles), and metadata features (relying on data other than text, such as the number of votes for the review, publication time, and review star rating).

[0004] The features currently used in predicting online reviews do not mine the review text content deeply enough, resulting in inaccurate analysis and prediction of online reviews, and then recommending low-quality or useless reviews to users. [Summary of the invention]

[0005] The purpose of this application is to provide a method, device and storage medium for pushing product review data, which can display high-quality reviews corresponding to products to users and give users good shopping suggestions and experience.

[0006] A first aspect of an embodiment of the present application provides a method for pushing product review data, including:

[0007] Preprocessing the target comments corresponding to the target product to obtain the word vector and aspect representation vector corresponding to the target comment, wherein the target comment is the comment to be recommended to the user;

[0008] The word vector and the aspect representation vector are connected in series and then input into an aspect sentiment analysis model to obtain an aspect sentiment score corresponding to the target comment, wherein the aspect sentiment analysis model is a model obtained by pre-training training data and used to determine the sentiment attribute of the comment;

[0009] Concatenate the aspect sentiment score and the target comment to obtain target data;

[0010] Determining user behavior data corresponding to the target comment, wherein the user behavior data is behavior data of the user corresponding to the target comment;

[0011] Inputting the target data and the user behavior data into a true and false review analysis model to obtain a reconstruction probability of the target review;

[0012] The target comment is pushed to the user according to the reconstruction probability of the target comment.

[0013] A second aspect of the embodiment of the present application provides a device for pushing product review data, including:

[0014] A vector preprocessing unit, used to preprocess the target review corresponding to the target product to obtain a word vector and an aspect representation vector corresponding to the target review, wherein the target review is a review to be recommended to the user;

[0015] An aspect sentiment analysis unit, configured to input the word vector and the aspect representation vector in series into an aspect sentiment analysis model to obtain an aspect sentiment score corresponding to the target comment, wherein the aspect sentiment analysis model is a model obtained by pre-training training data and used to determine the sentiment attribute of the comment;

[0016] A data concatenation unit, used for concatenating the aspect sentiment score and the target comment to obtain target data;

[0017] A behavior data determination unit, configured to determine user behavior data corresponding to the target comment, wherein the user behavior data is behavior data of the user corresponding to the target comment;

[0018] A true and false review analysis unit, used to input the target data and the user behavior data into a true and false review analysis model to obtain a reconstruction probability of the target review;

[0019] A pushing unit is used to push the target comment to the user according to the reconstruction probability of the target comment.

[0020] A third aspect of an embodiment of the present application provides a computer device comprising at least one connected processor, a memory, and a transceiver, wherein the memory is used to store program code, and the processor is used to call the program code in the memory to execute the steps of the method for pushing product review data described in the first aspect above.

[0021] A fourth aspect of an embodiment of the present application provides a computer storage medium, which includes instructions, which, when executed on a computer, enable the computer to execute the steps of the method for pushing product review data described in the first aspect above.

[0022] Compared with the related art, in the embodiment provided by the present application, when recommending product reviews to users, the aspect words in the reviews are extracted, and the sentiment attribute corresponding to the aspect words is predicted, and the review is analyzed for abnormality based on the sentiment attribute of the aspect words, the review text, and the behavior data corresponding to the review, and the reconstruction probability of the review is obtained, and finally the review is pushed according to the reconstruction probability. In this way, by performing sentiment analysis on the aspect words in the reviews and using anomaly detection to find out the aspects that professional reviewers pay more attention to, higher-quality reviews can be screened out, and then higher-quality reviews can be recommended to users.

Brief Description of the Drawings

[0023] Figure 1 A flowchart of a method for pushing product review data provided in an embodiment of the present application;

[0024] Figure 2 A schematic diagram of the training process of the sentiment analysis model provided in the embodiment of the present application;

[0025] Figure 3 A schematic diagram of the training process of the true and false review analysis model provided in the embodiment of the present application;

[0026] Figure 4 A virtual structural diagram of a product review data recommendation device provided in an embodiment of the present application;

[0027] Figure 5 A schematic diagram of the hardware structure of the server provided in an embodiment of the present application. [Specific implementation method]

[0028] The technical solutions in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all of the embodiments.

[0029] The following describes a method for pushing product review data from the perspective of a product review data recommendation device. The product review data recommendation device may be a server or a service unit in a server, and is not specifically limited.

[0030] Please refer to Figure 1 , Figure 1 A flowchart of a method for pushing product review data provided in an embodiment of the present application includes:

[0031] 101. Preprocess the target comments corresponding to the target product to obtain the word vector and aspect representation vector corresponding to the target comment.

[0032] In this embodiment, when it is necessary to recommend popular reviews about a certain product (target product) to the user, the target review corresponding to the target product can be preprocessed to obtain the aspect representation vector and word vector corresponding to the target review, and the target review is the review to be recommended to the user. The aspect representation vector is the vector corresponding to the aspect word in the target review. The aspect word is a special word that appears in the review, which represents that the user mentioned this aspect. Here, the following five aspects are defined: food, price, service, environment, and others.

[0033] It should be noted that a product review consists of w1, w2, ..., w n There are n words in total. Here, the comment R is defined as R = [w1, w2, ..., w i ,...w j ,...w n ], where w represents the word in the comment R. At the same time, it is also necessary to determine the aspect sequence A of the aspect words appearing in the comment R. The aspect sequence is a sequence segment with a distance K from the center of the aspect word. A is also composed of the words in the comment R. A=[w i ,w i+1 ,...,w i+m-1 ]. For example, a restaurant review is "I feel the food is good and affordable, but the fish-flavored pork shreds are too salty. It is very close to the subway entrance and is recommended for couples to come on Friday", then the aspect words in the review are: food (food), affordable (price), fish-flavored pork shreds (food), subway entrance (environment), couples (others); as described in this example, the aspect word is a word: for example, "food", if K = 3, then consider the first 3 words and the last 3 words of "food", then the aspect sequence L1 corresponding to the aspect word "food" = "I feel the food is not bad"; the aspect word is multiple words: for example, "fish-flavored pork shreds", also consider the first 3 words and the last 3 words, then the aspect word sequence L2 corresponding to the aspect word "fish-flavored pork shreds" = "the fish-flavored pork shreds are too salty". It is understandable that when determining the aspect word sequence, it is truncated when encountering punctuation marks, because according to language habits, punctuation marks are signs of the end or pause of a sentence, and the vocabulary describing the aspect has ended. For example, "subway entrance", then the sequence L3 corresponding to the aspect word "subway entrance" = "this place is very close to the subway entrance".

[0034] After obtaining the aspect sequence corresponding to the aspect word, the aspect sequence and the words in the comment can be vectorized to obtain the aspect representation vector and word vector of the comment.

[0035] 102. The word vector and the aspect representation vector are concatenated and input into the aspect sentiment analysis model to obtain the aspect sentiment score corresponding to the target comment.

[0036] In this embodiment, after obtaining the word vector and aspect representation vector corresponding to the target comment, the product review data push device can concatenate the word vector and aspect representation vector, and input the pre-trained aspect sentiment analysis model to obtain the aspect sentiment score corresponding to the target comment. The aspect sentiment score refers to whether the user's comment on a certain aspect is negative or positive. The judgment of negative or positive can be based on the sentiment dictionary. For example, "too expensive" is negative. The pre-trained aspect sentiment analysis model is to enable the aspect sentiment analysis model to realize that in terms of price, expensive is negative, but if it is expensive in other aspects, it is not necessarily negative, for example: "The hall is decorated in a grand style, with a noble atmosphere." Here it is positive. The aspect sentiment analysis model performs sentiment analysis on the aspect categories in the product review. Taking restaurant review data as an example, the review data includes entities, aspect categories involved in the entity, and the commentator's emotions on the aspect. The following five aspects can be defined: food, price, service, environment, and others. For example, "the music in the restaurant is good", where the entity is music, the aspect involved is the environment, and the emotion on the environment is positive.

[0037] 103. Concatenate the aspect sentiment scores and target comments to obtain the target data.

[0038] In this embodiment, after obtaining the aspect sentiment score corresponding to the target review through the aspect sentiment analysis model, the product review data push device can concatenate the aspect sentiment score and the review text of the target review to obtain the target data. Here, the target data is defined as X = [x1, x2, ..., x j ,...x n ].

[0039] 104. Determine the user behavior data corresponding to the target comment.

[0040] In this embodiment, the product review data push device can determine the user behavior data corresponding to the target review, that is, determine the behavior of the user corresponding to the target review, specifically including the number of historically published reviews, usefulness voting scores, brand deviation scores, review deviations, and the maximum number of reviews per day by the reviewer. Here, the user behavior data is defined as Y = [y1, y2, ..., y i ,...y n]. It can be understood that the usefulness voting score refers to: the likes of other users for this comment. Three approval methods can be set here: Useful, funny and cool. Of course, other approval methods can also be set, and there is no specific limitation; the brand deviation score refers to how much the rating of a certain brand is offset among all similar products. For example, a restaurant has a very poor review with only 2 stars, but the average score of all local restaurants is 3.6 stars, then the brand deviation score of the restaurant is -1.6; the rating deviation is related to the user. For example, if a restaurant has an overall rating of 4.2 and the user gives it a 5-star rating, then the rating deviation is +0.8.

[0041] It should be noted that, through step 103, the aspect sentiment score and the comment text of the target comment can be concatenated to obtain the target data, and through step 104, the user behavior data can be determined. However, there is no restriction on the order of execution between these two steps. Step 103 can be executed first, or step 104 can be executed first, or they can be executed at the same time, without specific limitation.

[0042] 105. Input the target data and user behavior data into the true and false review analysis model to obtain the reconstruction probability of the target review data.

[0043] In this embodiment, after obtaining the target data and the user behavior data, the product review data push device can input the target data and the user behavior data into a pre-trained true and false review analysis model to obtain the reconstruction probability of the target review data.

[0044] It can be understood that the pre-trained true and false review analysis model is a conditional variational auto-encoder (CVAE) model, which is used for anomaly detection of product reviews. The anomaly detection task is a semi-supervised framework, and only the data of normal instances are used to train CVAE. The reconstruction probability is a probability measure that considers the variability of variable distribution. Using the generative characteristics of the variational autoencoder, data reconstruction can be derived to analyze the root cause of the anomaly. CVAE is a conditional variational autoencoder, and the condition used is the commentator behavior, such as the user behavior data in step 104, that is, the commentator behavior is used as the condition of CVAE.

[0045] 106. Recommend the target comment to the user according to the reconstruction probability of the target comment.

[0046] In this embodiment, after obtaining the reconstruction probability of the target comment, the product review data recommendation device can determine whether the reconstruction probability is less than a preset threshold. If the reconstruction probability is less than the preset threshold, it means that the target comment is an abnormal comment or a low-quality comment, and the usefulness of the target comment is reduced, such as lowering the ranking of the target comment or reducing the probability of pushing the target comment; if the reconstruction probability is greater than the preset threshold, it means that the target comment is a high-quality normal comment, and the usefulness of the target comment is increased, such as increasing the ranking of the target comment or increasing the probability of pushing the target comment. For example, the comments on the homepage of a certain product are generally hot comments, which are recommended to users to view, stimulate user consumption, and facilitate users to understand the real product.

[0047] In one embodiment, the product review data push device further performs the following operations:

[0048] Determine a quality score for each of the multiple reviews corresponding to the target product;

[0049] Classify the multiple reviews corresponding to the target product according to the quality scores to obtain a high-quality data set and a low-quality data set;

[0050] Determine a reconstruction probability of each data in the high-quality data set and a reconstruction probability of each data in the low-quality data set;

[0051] Determine a high-quality abnormal data set in the high-quality data set according to a reconstruction probability of each data in the high-quality data set;

[0052] Determine low-quality normal data in the low-quality data set according to the reconstruction probability of each data in the low-quality data set;

[0053] determining a target probability of the aspect type appearing in each high-quality anomaly data in the high-quality anomaly data set;

[0054] Determine the normality score of the target product based on the low-quality normal data;

[0055] Update the behavioral data in the true and false review analysis model based on the target probability and normality score;

[0056] Input the target data and user behavior data into the true and false review analysis model, and obtain the reconstruction probability of the target review data including:

[0057] The reconstruction probability of the target comment data is determined based on the true and false comment analysis model updated based on the behavior data.

[0058] In this embodiment, the product review data recommendation device can also predict the usefulness of the product review. The prediction index of the usefulness of the review is the ratio of the usefulness vote of the product review to the total vote. If it is greater than the preset ratio threshold, the review is considered useful, otherwise it is useless. However, a newly published comment may not receive votes or the comment may be "brushed up" by an abnormal user. Therefore, it is necessary to find all abnormal review data with a reconstruction probability less than the preset threshold, and determine the quality score of each of the multiple reviews corresponding to the target product. The review data with a quality score greater than the preset threshold is used as a high-quality data set, and the review data with a quality score less than the preset threshold is used as a low-quality data set. Then, the high-quality data set and the low-quality data set are processed separately by the true and false review analysis model to obtain a high-quality normal data set, a high-quality abnormal data set, a low-quality normal data set, and a low-quality abnormal data set. Specifically, the quality score of each of the multiple reviews corresponding to the target product can be obtained by the following formula:

[0059]

[0060] Where Quality_Score(i) is the quality score of any review i among the multiple reviews corresponding to the target product, Na i is the number of aspect words in the comment i, Na max for Na i The maximum value, Nn i Nn is the number of historical comments posted by the user corresponding to each comment in the multiple comments. max Nn i The maximum value of Na i The proportionality coefficient, β is the Nn i The proportionality coefficient is α+β=1.

[0061] Afterwards, for a high-quality anomaly dataset M, M contains textual comments, commentator behaviors, and aspect-level sentiments, where M = [m1, m2, ..., m i ,...m n ],m i =[R i ,A i ,Y i ], where m i Represents the comment R posted by user i i , the comment R i The aspect-level sentiment A contained in i and user i’s behavior Y i ;

[0062] And calculate the probability of the aspect type (the aspect type includes at least one of the five aspects defined above: food, price, service, environment and others) appearing in the aspect-level sentiment A n , taking the aspect type as the proportional coefficient, we can get the aspects that abnormal users focus on: a1a2...a k , k refers to the number of aspect words, and the probability γ i That is aspect a i For example, a restaurant review "I think the food is good and affordable, but the fish-flavored pork shreds are too salty. It is very close to the subway entrance, and I recommend couples to come on weekends." The aspect words included are: food (food), affordable (price), fish-flavored pork shreds (food), subway entrance (environment), couples (others). In this review, "food" appears twice, "price" appears once, "environment" appears once, and "others" appears once. Then the probabilities of these five aspects "food, price, service, environment, and other" are 40%, 20%, 0%, 20%, and 20% respectively. The probability here is obtained by adding up the number of aspects in the review and dividing it by the total number of aspect words to get the proportional coefficient;

[0063] For the comment data in the low-quality normal comment data set, the average score of the product is calculated based on the low-quality normal comment data, that is, the comment data will include the score of the product. The average score corresponding to all low-quality normal comment data can be obtained by averaging them, and the average score is used as the normal score of the product. Then, the probability γ i The conditions in the true and false review analysis model are updated according to the normal score of the product, that is, the behavior data Y corresponding to the true and false review analysis model is updated, and the reconstruction probability of the target review data is determined based on the true and false review analysis model updated with the behavior data.

[0064] It should be noted that the ratio of the number of votes received by a comment to the total number of votes received by the comments on the product to which the comment belongs is used here, because in online shopping, high-quality comments are very likely to be liked, and often account for the majority of the total number of likes on product comments. In addition, the behavior data Y is updated to add the proportional coefficient of aspect-level importance, so that the true and false comment analysis model can consider the aspects that users of high-quality comments care about.

[0065] In summary, it can be seen that in the embodiment provided by the present application, when recommending product reviews to users, aspect words in the reviews are extracted, and the sentiment attribute corresponding to the aspect words is predicted, and the review is analyzed for abnormality based on the sentiment attribute of the aspect words, the review text, and the behavior data corresponding to the review, and the reconstruction probability of the review is obtained, and finally the review is pushed according to the reconstruction probability. In this way, by performing sentiment analysis on the aspect words in the reviews and using abnormality detection to find out the aspects that professional reviewers are more concerned about, higher quality reviews can be screened out, and then higher quality reviews can be recommended to users.

[0066] The following is an explanation of the training of the aspect sentiment analysis model. Figure 2 , Figure 2 A schematic diagram of the training process of the sentiment analysis model provided in the embodiment of the present application includes:

[0067] 201. Obtain a training data set.

[0068] In this embodiment, the product review data push device can obtain a training data set, which includes N reviews, where N is an integer greater than or equal to 1. That is, when training a sentiment analysis model for product review data, a large amount of review data can be obtained for model training.

[0069] 202. Preprocess the comments in the training data set to obtain word vectors and aspect representation vectors corresponding to the comments in the training data set.

[0070] It can be understood that the above step 101 and the determination of the word vector and the aspect representation vector corresponding to the comment are described in detail, and the details are not repeated here.

[0071] 203. Determine the output data corresponding to the word vector and the aspect representation vector through the long short-term memory network model.

[0072] In this embodiment, a long short-term memory network (LSTM) model is used to perform aspect-level sentiment analysis. The input of the LSTM model is the word vector and aspect representation vector corresponding to the comment text in the training data set. First, the word vector and the aspect representation vector are concatenated to obtain concatenated data, and then the concatenated data is input into the LSTM model for context encoding. Compared with the recurrent neural network (RNN), LSTM solves the problem of long-term dependency. The word vector and the aspect representation vector are concatenated in series so that the aspect information can be taken into account when encoding the context information. The specific formula is as follows:

[0073] H=LSTM([w i ,v a ])

[0074] Among them, w i The review text of the training data set R = [w1,w2,...,w i ,...w n ] corresponds to the word vector of the i-th word, v a The review text of the training data set R = [w1,w2,...,w i ,...w n ] corresponds to the aspect representation vector, H = [h1,h2,...,h i ,...h n ] is the output data of the hidden layer of the aspect sentiment analysis model.

[0075] 204. Concatenate the output data with the aspect representation vector to obtain concatenated data.

[0076] In this embodiment, after obtaining the output data, the product review data push device can connect the output data and the aspect representation vector in series again, that is, H and v a Connect in series again to obtain the series data.

[0077] 205. The concatenated data is iteratively trained through the attention mechanism to obtain the aspect sentiment analysis model.

[0078] In this embodiment, the product review data push device adds an attention mechanism during the model training process, and a They are connected in series again, and the weight of each node in the hidden layer is calculated through the fully connected layer and the activation function as the importance of the corresponding word in sentiment prediction. Finally, the attention weight and the hidden layer representation are vector-multiplied to obtain the context representation with the attention mechanism introduced:

[0079] att i =W2 tanh([W1(h i ,v a )])

[0080] α=softmax(att)

[0081]

[0082] Among them, W1 and W2 are model parameters, tanh is the activation function, n is the number of words in the comment, and r is the final context representation with the attention mechanism added. Then, the context representation after the attention mechanism is added is weightedly added to the original context representation, and the result is predicted through the activation function, linear transformation layer and softmax layer:

[0083] h′=ta1nh([W r hr +W x h n ])

[0084] γ=softmax(W s h′+b s )

[0085] Among them, W r , W x , W s and b s is the model parameter, h r Finally, the hidden layer of context representation after the attention mechanism is added. Therefore, iterative training can be performed according to the above steps until the preset iteration termination condition is reached to obtain the aspect sentiment analysis model.

[0086] In summary, it can be seen that in the embodiments provided in the present application, in the process of training the model, the word vector and the aspect representation vector are respectively concatenated to obtain output data, and the aspect representation vector is again concatenated with the output data, and the model is trained based on the concatenated data. At the same time, an attention mechanism is added in the model training process, so that a better model for identifying the emotional attributes of comments can be trained.

[0087] See also Figure 3 , Figure 3 The training process diagram of the true and false review analysis model provided in the embodiment of the present application includes:

[0088] 301. Get M comments.

[0089] In this embodiment, when the product review data recommendation device needs to train the true and false review analysis model, M reviews can be obtained, where M is a positive integer greater than or equal to 1. That is, when the product review data is used to train the true and false review analysis model, a large number of review texts of normal reviewers can be obtained for model training, where the review data of the normal reviewers can be obtained by filtering through a filter.

[0090] 302. Determine the aspect sentiment attribute and behavior data corresponding to each of the M comments.

[0091] In this embodiment, the method for determining the aspect sentiment attribute corresponding to each of the M comments is not specifically limited. For example, the aspect sentiment attribute of each comment can be determined by manual annotation. Of course, it can also be determined by Figure 2 The trained aspect sentiment analysis model is used to determine; in addition, the behavioral data of each comment is as mentioned above Figure 1 The step 104 has been described in detail and will not be described again here.

[0092] 303. Concatenate each comment with the aspect sentiment attribute corresponding to each comment to obtain concatenated training data.

[0093] In this embodiment, after obtaining the comment text of each comment and the aspect sentiment attribute corresponding to each comment, the product review data recommendation device can concatenate the comment text and the aspect sentiment attribute of each comment to obtain concatenated training data. Here, the concatenated training data is defined as X = [x1, x2, ..., x j ,...x n ].

[0094] 304. Perform iterative training based on the serialized training data and behavior data to obtain a true and false review analysis model.

[0095] In this embodiment, the product review data recommendation device can perform iterative training based on the serialized training data and behavior data to obtain a true and false review analysis model. The following is a specific description:

[0096] Assume that the behavior data y has no direct relationship with the hidden variable z. The hidden variable is the content in CVAE, which refers to the data that really works. It refers to the data in a comment, including the additional information and emotional information, which really determines whether the comment is true or false. Therefore,

[0097] Conditional probability: p θ (z|y)=p θ (z)

[0098] Variational lower bound:

[0099]

[0100] First, we train CVAE using the normal reviewer’s review data X1 and the corresponding reviewer behavior data Y1 to obtain θ and φ, where θ and φ are the probability distribution parameters of p and q respectively, p is the prior distribution, q is the learned distribution, and L is the loss function. The purpose of minimizing the loss function is to make the learned distribution q(z|x) as similar as possible to the true prior distribution p(z).

[0101] The composition of CVAE includes encoding and decoding stages. Based on the reviewer's comment data (the comment data includes normal comment data and abnormal comment data), the probability encoder f φ and the probability decoder g θ Isotropic normal distributions are parameterized in the latent variable space and the original input variable space, respectively.

[0102] During the model training process, for a training sample (x i ,y i ) From the trained probability encoder f φL samples are extracted from the probability encoder fφ. For each sample from the probability encoder fφ, the probability decoder gθ outputs the mean μ and variance σ parameters. Using these parameters, the probability of generating the original data from the distribution is calculated, which is the reconstruction probability (Reconstruction Probability, abbreviated as RP). The specific calculation formula is as follows:

[0103]

[0104] If the reconstruction probability is less than the preset threshold, the training sample (x i ,y i ) is abnormal data. If the reconstruction probability is not less than the preset threshold, the training sample (x i ,y i ) is normal data, and iterative training is continued until the preset iteration termination condition is reached to obtain a true and false review analysis model.

[0105] It should be noted that after obtaining the training samples, the training samples can be divided into training data and verification data according to the proportion, the model can be trained by the training data, and the trained model can be verified by the verification data.

[0106] The above describes the embodiment of the present application from the perspective of a method for pushing product review data. The following describes the embodiment of the present application from the perspective of a device for recommending product review data.

[0107] See also Figure 4 , Figure 4 This is a virtual structural diagram of a product review data recommendation device provided in an embodiment of the present application. The product review data recommendation device 400 includes:

[0108] A vector preprocessing unit 401 is used to preprocess a target comment corresponding to a target product to obtain a word vector and an aspect representation vector corresponding to the target comment, wherein the target comment is a comment to be recommended to a user;

[0109] An aspect sentiment analysis unit 402, configured to input the word vector and the aspect representation vector in series into an aspect sentiment analysis model to obtain an aspect sentiment score corresponding to the target comment, wherein the aspect sentiment analysis model is a model obtained by pre-training training data and used to determine the sentiment attribute of a comment;

[0110] A data concatenation unit 403 is used to concatenate the aspect sentiment score and the target comment to obtain target data;

[0111] A behavior data determination unit 404 is used to determine the user behavior data corresponding to the target comment, wherein the user behavior data is the behavior data of the user corresponding to the target comment;

[0112] A true and false review analysis unit 405, used to input the target data and the user behavior data into a true and false review analysis model to obtain a reconstruction probability of the target review;

[0113] The pushing unit 406 is configured to push the target comment to the user according to the reconstruction probability of the target comment.

[0114] In one possible design, the device further includes:

[0115] The model training unit 407 is used to:

[0116] Obtain a training data set, wherein the training data set includes N comments, where N is an integer greater than or equal to 1;

[0117] Preprocessing the comments in the training data set to obtain word vectors and aspect representation vectors corresponding to the comments in the training data set;

[0118] Determining output data corresponding to the word vector and the aspect representation vector through a long short-term memory network model;

[0119] Concatenating the output data with the aspect representation vector to obtain concatenated data;

[0120] The concatenated data is iteratively trained through an attention mechanism to obtain the aspect sentiment analysis model.

[0121] In a possible design, the model training unit 407 further includes:

[0122] Get M comments, where M is a positive integer greater than or equal to 1;

[0123] Determine the aspect sentiment attribute and behavior data corresponding to each of the M comments;

[0124] Concatenate each comment and the aspect sentiment attribute corresponding to each comment to obtain concatenated training data;

[0125] Iterative training is performed based on the series training data and the behavior data to obtain the true and false review analysis model.

[0126] In a possible design, the model training unit 407 is further used for:

[0127] Determine a quality score for each of the multiple reviews corresponding to the target product;

[0128] Classifying the multiple reviews corresponding to the target product according to the quality scores to obtain a high-quality data set and a low-quality data set;

[0129] Determine a reconstruction probability of each data in the high-quality data set and a reconstruction probability of each data in the low-quality data set;

[0130] Determine a high-quality abnormal data set in the high-quality data set according to the reconstruction probability of each data in the high-quality data set;

[0131] Determine low-quality normal data in the low-quality data set according to the reconstruction probability of each data in the low-quality data set;

[0132] Determining a target probability of an aspect type appearing in each high-quality anomaly data in the high-quality anomaly data set;

[0133] Determining a normality score of the target product according to the low-quality normal data;

[0134] Updating the behavior data corresponding to the true and false review analysis model based on the target probability and the normality score;

[0135] The model training unit 407 is specifically used for:

[0136] The reconstruction probability of the target comment data is determined based on the true and false comment analysis model updated based on the behavior data.

[0137] In a possible design, the model training unit 407 determines the quality score of each of the multiple reviews corresponding to the target product, including:

[0138] The quality score of each review in the multiple reviews corresponding to the target product is obtained by the following formula:

[0139]

[0140] Where Quality_Score(i) is the quality score of any review i among the multiple reviews corresponding to the target product, Na i is the number of aspect words in the comment i, Na max for Na i The maximum value, Nn i Nn is the number of historical comments posted by the user corresponding to each comment in the multiple comments. max Nn i The maximum value of , α is the proportional coefficient of Na, and β is the proportional coefficient of Nn.

[0141] In a possible design, the model training unit 407 determines the target probability of the appearance of the aspect type in each high-quality abnormal data in the high-quality abnormal data set, including:

[0142] Determine the number of aspect words of the aspect type in target high-quality anomaly data, where the target high-quality anomaly data is any one of the high-quality anomaly data sets;

[0143] Determine the number of all words in the target high-quality anomaly data;

[0144] A target probability of the aspect type is determined according to the number of aspect words and the number of all words.

[0145] In summary, it can be seen that in the embodiment provided by the present application, when recommending product reviews to users, aspect words in the reviews are extracted, and the sentiment attribute corresponding to the aspect words is predicted, and the review is analyzed for abnormality based on the sentiment attribute of the aspect words, the review text, and the behavior data corresponding to the review, and the reconstruction probability of the review is obtained, and finally the review is pushed according to the reconstruction probability. In this way, by performing sentiment analysis on the aspect words in the reviews and using anomaly detection to find out the aspects that professional reviewers pay more attention to, higher quality reviews can be screened out, and then higher quality reviews can be recommended to users.

[0146] Figure 5 This is a schematic diagram of the structure of the application server, such as Figure 5 As shown, the server 500 of this embodiment includes at least one processor 501, at least one network interface 504 or other user interface 503, a memory 505, and at least one communication bus 502. The server 500 may optionally include a user interface 503, including a display, a keyboard or a clicking device. The memory 505 may include a high-speed RAM memory, and may also include a non-volatile memory (non-volatile memory), such as at least one disk storage. The memory 505 stores execution instructions. When the server 500 is running, the processor 501 communicates with the memory 505, and the processor 501 calls the instructions stored in the memory 505 to execute the above-mentioned method of pushing product review data. The operating system 506 includes various programs for implementing various basic businesses and processing hardware-based tasks.

[0147] The server provided in the embodiment of the present application can execute the technical solution of the embodiment of the method for pushing product review data, and its implementation principle and technical effect are similar, which will not be repeated here.

[0148] The embodiment of the present application also provides a computer-readable medium, comprising computer execution instructions, which can enable a server to execute the method for pushing product review data described in the above embodiment. The implementation principle and technical effect are similar and will not be repeated here.

[0149] Those skilled in the art can understand that all or part of the steps of implementing the above-mentioned method embodiments can be completed by hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium. When the program is executed, the steps of the above-mentioned method embodiments are executed; and the aforementioned storage medium includes: ROM, RAM, disk or optical disk and other media that can store program codes.

[0150] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit it. Although the present application has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present application.

Claims

1. A method for pushing product review data, characterized in that: include: Preprocessing the target comments corresponding to the target product to obtain the word vector and aspect representation vector corresponding to the target comment, wherein the target comment is the comment to be recommended to the user; The word vector and the aspect representation vector are connected in series and then input into an aspect sentiment analysis model to obtain an aspect sentiment score corresponding to the target comment, wherein the aspect sentiment analysis model is a model obtained by pre-training training data and used to determine the sentiment attribute of the comment; Concatenate the aspect sentiment score and the target comment to obtain target data; Determining user behavior data corresponding to the target comment, wherein the user behavior data is behavior data of the user corresponding to the target comment; Inputting the target data and the user behavior data into a true and false review analysis model to obtain a reconstruction probability of the target review; Pushing the target comment to the user according to the reconstruction probability of the target comment, The method further comprises: Determine a quality score for each of the multiple reviews corresponding to the target product; Classifying the multiple reviews corresponding to the target product according to the quality scores to obtain a high-quality data set and a low-quality data set; Determine a reconstruction probability of each data in the high-quality data set and a reconstruction probability of each data in the low-quality data set; Determine a high-quality abnormal data set in the high-quality data set according to the reconstruction probability of each data in the high-quality data set; Determine low-quality normal data in the low-quality data set according to the reconstruction probability of each data in the low-quality data set; Determining a target probability of an aspect type appearing in each high-quality anomaly data in the high-quality anomaly data set; Determining a normality score of the target product according to the low-quality normal data; updating the behavior data corresponding to the true and false review analysis model based on the target probability and the normality score, Determining the quality score of each of the multiple reviews corresponding to the target product includes: The quality score of each review in the multiple reviews corresponding to the target product is obtained by the following formula: Where Quality_Score(i) is the quality score of any review i among the multiple reviews corresponding to the target product, Na i is the number of aspect words in the comment i, Na max for Na i The maximum value, Nn i Nn is the number of historical comments posted by the user corresponding to each comment in the multiple comments. max Nn i The maximum value of Na i The proportionality coefficient, β is the Nn i The proportionality factor.

2. The method according to claim 1, characterized in that The method further comprises: Obtain a training data set, wherein the training data set includes N comments, where N is an integer greater than or equal to 1; Preprocessing the comments in the training data set to obtain word vectors and aspect representation vectors corresponding to the comments in the training data set; Determining output data corresponding to the word vector and the aspect representation vector through a long short-term memory network model; Concatenating the output data with the aspect representation vector to obtain concatenated data; The concatenated data is iteratively trained through an attention mechanism to obtain the aspect sentiment analysis model.

3. The method according to claim 1, characterized in that The method further comprises: Get M comments, where M is a positive integer greater than or equal to 1; Determine the aspect sentiment attribute and behavior data corresponding to each of the M comments; Concatenate each comment and the aspect sentiment attribute corresponding to each comment to obtain concatenated training data; Iterative training is performed based on the series training data and the behavior data to obtain the true and false review analysis model.

4. The method according to any one of claims 1 to 3, characterized in that Inputting the target data and the user behavior data into the true and false review analysis model to obtain the reconstruction probability of the target review data includes: The reconstruction probability of the target comment data is determined based on the true and false comment analysis model updated based on the behavior data.

5. The method according to claim 1, characterized in that: Determining the target probability of the appearance of an aspect type in each high-quality anomaly data in the high-quality anomaly data set comprises: Determine the number of aspect words of the aspect type in target high-quality anomaly data, where the target high-quality anomaly data is any one of the high-quality anomaly data sets; Determine the number of all words in the target high-quality anomaly data; A target probability of the aspect type is determined according to the number of aspect words and the number of all words.

6. A device for pushing product review data, characterized in that: include: A vector preprocessing unit, used to preprocess the target review corresponding to the target product to obtain a word vector and an aspect representation vector corresponding to the target review, wherein the target review is a review to be recommended to the user; An aspect sentiment analysis unit, configured to input the word vector and the aspect representation vector in series into an aspect sentiment analysis model to obtain an aspect sentiment score corresponding to the target comment, wherein the aspect sentiment analysis model is a model obtained by pre-training training data and used to determine the sentiment attribute of the comment; A data concatenation unit, used for concatenating the aspect sentiment score and the target comment to obtain target data; A behavior data determination unit, configured to determine user behavior data corresponding to the target comment, wherein the user behavior data is behavior data of the user corresponding to the target comment; A true and false review analysis unit, used to input the target data and the user behavior data into a true and false review analysis model to obtain a reconstruction probability of the target review; a pushing unit, configured to push the target comment to the user according to the reconstruction probability of the target comment, Model training unit, used to: Determine a quality score for each of the multiple reviews corresponding to the target product; Classifying the multiple reviews corresponding to the target product according to the quality scores to obtain a high-quality data set and a low-quality data set; Determine a reconstruction probability of each data in the high-quality data set and a reconstruction probability of each data in the low-quality data set; Determine a high-quality abnormal data set in the high-quality data set according to the reconstruction probability of each data in the high-quality data set; Determine low-quality normal data in the low-quality data set according to the reconstruction probability of each data in the low-quality data set; Determining a target probability of an aspect type appearing in each high-quality anomaly data in the high-quality anomaly data set; Determining a normality score of the target product according to the low-quality normal data; updating the behavior data corresponding to the true and false review analysis model based on the target probability and the normality score, Determining the quality score of each of the multiple reviews corresponding to the target product includes: The quality score of each review in the multiple reviews corresponding to the target product is obtained by the following formula: Where Quality_Score(i) is the quality score of any review i among the multiple reviews corresponding to the target product, Na i is the number of aspect words in the comment i, Na max for Na i The maximum value, Nn i Nn is the number of historical comments posted by the user corresponding to each comment in the multiple comments. max Nn i The maximum value of Na i The proportionality coefficient, β is the Nn i The proportionality factor.

7. The device according to claim 6, characterized in that The model training unit is also used for: Obtain a training data set, wherein the training data set includes N comments, where N is an integer greater than or equal to 1; Preprocessing the comments in the training data set to obtain word vectors and aspect representation vectors corresponding to the comments in the training data set; Determining output data corresponding to the word vector and the aspect representation vector through a long short-term memory network model; Concatenating the output data with the aspect representation vector to obtain concatenated data; The concatenated data is iteratively trained through an attention mechanism to obtain the aspect sentiment analysis model.

8. A computer device, characterized in that: include: At least one connected processor, memory and transceiver, wherein the memory is used to store program code, and the processor is used to call the program code in the memory to execute the method for pushing product review data as described in any one of claims 1 to 5.

9. A computer storage medium, characterized in that: include: Instructions, when the instructions are executed on a computer, enable the computer to execute the method for pushing product review data as described in any one of claims 1 to 5.

Images