Comment monitoring method, medium, apparatus, and computing device

Abstract

The comment monitoring method, medium, device and computing device provided by the present disclosure can accurately determine whether the target comment is a feature comment such as a violation comment, thereby improving the reliability of comment monitoring.

Description

TECHNICAL FIELD

[0001] Embodiments of the present disclosure relate to the field of Internet, and more particularly, to a comment monitoring method, medium, device and computing device. BACKGROUND

[0002] This section is intended to provide a background or context to the embodiments of the disclosure recited in the claims. The description herein does not constitute admission of prior art.

[0003] Nowadays, the Internet is full of massive knowledge and fresh information, but also full of negative content such as abusive attacks and political sensitivities. For example, for social networks and communities such as microblogs, BBS, forums, groups, etc., users are allowed to directly publish content, which becomes the focus area of garbage information and needs to be specially cleaned and purified. In contrast, news bulletins and the like published by media agencies have been audited before publication, so the compliance of the theme content can be guaranteed, but the violation comments and the like in the comments still need to be screened by the Internet platform. SUMMARY

[0004] The present disclosure provides a comment monitoring method, medium, device and computing device to solve the problem that the violation object cannot be accurately determined as a comment or a title, thereby improving the reliability of comment monitoring.

[0005] In a first aspect of the embodiments of the present disclosure, a comment monitoring method is provided, comprising: obtaining a target comment and a theme text corresponding to the target comment; obtaining a threshold value vector of the theme text, the threshold value vector representing semantic features of the theme text, and the threshold value vector being used to assist classification of the target comment; performing semantic feature extraction on the target comment to obtain a signal value vector; performing feature fusion on the threshold value vector and the signal value vector to obtain a fusion vector corresponding to the target comment; and determining whether the target comment is a feature comment according to the fusion vector.

[0006] In an embodiment of the present disclosure, obtaining the threshold value vector of the theme text comprises: in response to the threshold value vector being stored, reading the stored threshold value vector.

[0007] In another embodiment of the present disclosure, obtaining the threshold value vector of the theme text comprises: in response to the threshold value vector of the theme text not being stored, extracting semantic features of the theme text to obtain a first semantic feature matrix; extracting semantic features of the first semantic feature matrix to obtain a second semantic feature matrix; performing dimension compression on the second semantic feature matrix to obtain a first semantic feature vector; and performing normalization processing on the first semantic feature vector to obtain the threshold value vector of the theme text.

[0008] In a further embodiment of the present disclosure, the semantic feature extraction of the target review is performed to obtain a signal value vector, including: extracting the semantic features of the target review to obtain a third semantic feature matrix; and performing dimension compression on the third semantic feature matrix to obtain the signal value vector of the target review.

[0009] In a further embodiment of the present disclosure, the semantic feature extraction of the target review is performed to obtain a third semantic feature matrix, including: performing text preprocessing on the target review to obtain a review vector matrix corresponding to the target review; and extracting the semantic features of the review vector matrix corresponding to the target review to obtain the third semantic feature matrix.

[0010] In a further embodiment of the present disclosure, the determination of whether the target review is a feature review according to the fusion vector includes: determining a probability value of the target review being a feature review according to the fusion vector; determining that the target review is a feature review in response to the probability value being greater than a feature probability threshold; and determining that the target review is not a feature review in response to the probability value being less than or equal to the feature probability threshold.

[0011] In a further embodiment of the present disclosure, the determination of the probability value of the target review being a feature review according to the fusion vector includes: performing dimension compression on the fusion vector to obtain a feature fusion vector; and performing normalization processing on the feature fusion vector to obtain a probability value vector, the probability value vector including the probability value of the target review being a feature review.

[0012] In a further embodiment of the present disclosure, it further includes: determining whether the target review is a feature review according to the target review and the topic text corresponding to the target review through a review classification model, the review classification model being used to determine whether a review is a feature review.

[0013] In a further embodiment of the present disclosure, the review classification model is obtained by the following method: obtaining a plurality of training data packets, at least one sample group being included in the training data packet, a sample topic text and at least one sample review corresponding to each sample topic text being included in the sample group; and training the review classification model based on the plurality of training data packets to obtain a review classification model meeting a preset requirement.

[0014] In a further embodiment of the present disclosure, the training of the review classification model based on the plurality of training data packets to obtain a review classification model meeting a preset requirement includes: performing contrastive learning training of the review classification model based on the plurality of training data packets to obtain a review classification model meeting a preset requirement.

[0015] In a further embodiment of the present disclosure, the contrastive learning training of the review classification model based on the plurality of training data packets to obtain a review classification model meeting a preset requirement includes: performing contrastive learning training of the review classification model according to each training data packet in the plurality of training data packets using the following contrastive learning target:

[0016]

[0017]

[0018] wherein, L-CONTRA is a contrastive learning target value, L-CONTRA i,j is a contrastive learning target of the jth review of the sample group identified as i, and sim() is a cosine similarity function; is an indicator function; τ is a temperature adjustment factor, l i,j represents a classification value of the jth review of the sample group identified as i, wherein the classification value of a feature review is 1, and the classification value of a non-feature review is 0; N represents the number of sample groups; K i represents the number of sample groups; r i,j represents a fusion vector corresponding to the jth review of the sample group identified as i; r i,j represents a fusion vector corresponding to the kth review of the sample group identified as i; exp() is an exponential function; and the review classification model is obtained by taking the contrastive learning target value to be less than or equal to a set value.

[0019] In a second aspect of the embodiments of the present disclosure, a computer-readable storage medium is provided, and the computer-readable storage medium stores computer program instructions. When the computer program instructions are executed, the review monitoring method of the first aspect is implemented.

[0020] In a third aspect of the embodiments of the present disclosure, a review monitoring device is provided, and the review monitoring device comprises: a first obtaining module configured to obtain a target review and a topic text corresponding to the target review; a second obtaining module configured to obtain a threshold value vector of the topic text, the threshold value vector representing semantic features of the topic text, and the threshold value vector being used to assist in classifying the target review; a third obtaining module configured to perform semantic feature extraction on the target review to obtain a signal value vector; a feature fusion module configured to perform feature fusion on the threshold value vector and the signal value vector to obtain a fusion vector corresponding to the target review; and a determining module configured to determine whether the target review is a feature review according to the fusion vector.

[0021] In an embodiment of the present disclosure, the second obtaining module is specifically configured to: in response to the threshold value vector being stored, read the stored threshold value vector.

[0022] In another embodiment of the present disclosure, the second obtaining module comprises a first convolutional layer, a second convolutional layer, a first pooling layer and a normalization layer, wherein: the first convolutional layer is configured to extract semantic features of the topic text to obtain a first semantic feature matrix in response to a threshold value vector of the topic text; the second convolutional layer is configured to extract semantic features of the first semantic feature matrix to obtain a second semantic feature matrix; the pooling layer is configured to compress the dimensions of the second semantic feature matrix to obtain a first semantic feature vector; and the normalization layer is configured to normalize the first semantic feature vector to obtain the threshold value vector of the topic text.

[0023] In yet another embodiment of the present disclosure, the third obtaining module comprises a third convolutional layer and a second pooling layer, wherein: the third convolutional layer is configured to extract semantic features of the target review to obtain a third semantic feature matrix; and the second pooling layer is configured to compress the dimensions of the third semantic feature matrix to obtain a signal value vector of the target review.

[0024] In still another embodiment of the present disclosure, the third convolutional layer is specifically configured to: perform text preprocessing on the target review to obtain a review vector matrix corresponding to the target review; and extract semantic features of the review vector matrix corresponding to the target review to obtain the third semantic feature matrix.

[0025] In still another embodiment of the present disclosure, the determining module is specifically configured to: determine a probability value of the target review being a feature review according to the fusion vector; determine that the target review is a feature review in response to the probability value being greater than a feature probability threshold value; and determine that the target review is not a feature review in response to the probability value being less than or equal to the feature probability threshold value.

[0026] In still another embodiment of the present disclosure, the determining module is specifically configured to: compress the dimensions of the fusion vector to obtain a feature fusion vector; and perform normalization processing on the feature fusion vector to obtain a probability value vector, the probability value vector comprising the probability value of the target review being a feature review.

[0027] In still another embodiment of the present disclosure, the determining module is further configured to: determine whether the target review is a feature review according to the target review and the topic text corresponding to the target review by using a review classification model, the review classification model being configured to determine whether a review is a feature review.

[0028] In still another embodiment of the present disclosure, the training module is further configured to: train the review classification model by: obtaining a plurality of training data packets, each training data packet comprising at least one sample group, each sample group comprising a sample topic text and at least one sample review corresponding to the sample topic text; and training the review classification model based on the plurality of training data packets to obtain a review classification model meeting a preset requirement.

[0029] In a further embodiment of the present disclosure, the training module is specifically configured to perform contrastive learning training on the comment classification model based on the plurality of training data packets to obtain a comment classification model satisfying a preset requirement.

[0030] In a further embodiment of the present disclosure, the training module is specifically configured to perform contrastive learning training on the comment classification model according to each training data packet in the plurality of training data packets by using a contrastive learning objective as follows:

[0031]

[0032]

[0033] wherein L-CONTRA is the contrastive learning objective value, L-CONTRA i,j is the contrastive learning objective of the jth comment with sample grouping identification i, and sim() is a cosine similarity function; is an indicator function; τ is a temperature adjustment factor, l i,j represents a classification value of the jth comment with sample grouping identification i, wherein the classification value of a feature comment is 1, and the classification value of a non-feature comment is 0; N represents the number of sample groupings; K i represents the number of sample groupings; r i,j represents a fusion vector corresponding to the jth comment with sample grouping identification i; r i,j represents a fusion vector corresponding to the kth comment with sample grouping identification i; exp() is an exponential function; and the comment classification model with a contrastive learning objective value less than or equal to a set value is obtained.

[0034] In a fourth aspect of the embodiments of the present disclosure, a computing device is provided, comprising a memory and a processor, the memory being configured to store program instructions, and the processor being configured to invoke the program instructions in the memory to execute the comment monitoring method of the first aspect.

[0035] The comment monitoring method, medium, device, and computing device disclosed herein, after acquiring the target comment and its corresponding topic text, obtains a threshold vector of the topic text, wherein the threshold vector represents the semantic features of the topic text and is used to assist in classifying the target comment; semantic features are extracted from the target comment to obtain a signal value vector; the threshold vector and the signal value vector are fused to obtain a fusion vector corresponding to the target comment; and the target comment is determined as a feature comment based on the fusion vector. This disclosure separates the topic text and target comments to obtain threshold vectors and signal vectors, and then fuses these vectors for feature extraction. This method avoids premature fusion of topic text and target comments, which could interfere with the extraction of their complete semantic features. Furthermore, the threshold vector obtained is from the topic text, which is used to assist in the classification of target comments. In other words, the topic text is used as an aid in determining whether a target comment is a feature comment, avoiding over-reliance on the content of the topic text. Instead, the content of the target comment is used as the primary factor in determining whether it is a feature comment. This allows for accurate determination of whether suspicious local semantic information is included in the comment, thus improving the reliability of comment monitoring. Attached Figure Description

[0036] The above and other objects, features, and advantages of this disclosure will become readily apparent from the following detailed description of exemplary embodiments, taken in conjunction with the accompanying drawings. Several embodiments of this disclosure are illustrated in the drawings by way of example and not limitation, in which:

[0037] Figure 1 An application scenario diagram illustrating an embodiment of the present disclosure is shown schematically;

[0038] Figure 2 A flowchart illustrating a comment monitoring method according to an embodiment of the present disclosure is shown schematically;

[0039] Figure 3 A computation time diagram according to an embodiment of the present disclosure is illustrated schematically;

[0040] Figure 4 A training data packet generation method according to an embodiment of the present disclosure is illustrated schematically;

[0041] Figure 5a A comparative learning training method according to an embodiment of the present disclosure is illustrated schematically;

[0042] Figure 5b The diagram schematically illustrates the results corresponding to each step of the comparative learning training preprocessing method according to an embodiment of the present disclosure.

[0043] Figure 6aA flowchart of a comment monitoring method according to another embodiment of the disclosure is schematically shown;

[0044] Figure 6b A result diagram of a comment monitoring method according to another embodiment of the disclosure is schematically shown;

[0045] Figure 7 A storage medium diagram according to an embodiment of the disclosure is schematically shown;

[0046] Figure 8 A comment monitoring device according to an embodiment of the disclosure is schematically shown;

[0047] Figure 9 A structural diagram of a computing device according to an embodiment of the disclosure is schematically shown.

[0048] In the drawings, identical or corresponding numbers represent identical or corresponding parts. DETAILED DESCRIPTION

[0049] The principles and spirits of the disclosure will be described below with reference to several exemplary embodiments. It should be understood that these embodiments are given only to enable those skilled in the art to better understand and implement the disclosure, and in no way limit the scope of the disclosure. On the contrary, these embodiments are provided to make the disclosure more thorough and complete, and to fully convey the scope of the disclosure to those skilled in the art.

[0050] Those skilled in the art know that the embodiments of the disclosure can be implemented as a system, device, apparatus, method or computer program product. Therefore, the disclosure can be embodied in the form of a complete hardware, complete software (including firmware, resident software, microcode, etc.), or a combination of hardware and software.

[0051] According to embodiments of the disclosure, a comment monitoring method, medium, device and computing device are proposed.

[0052] In this document, it should be understood that any number of elements in the drawings is used for illustration only and not limitation, and any naming is only for distinction and does not have any limiting meaning.

[0053] The principles and spirits of the disclosure will be described below with reference to several representative embodiments of the disclosure. SUMMARY

[0055] In the traditional comment monitoring method, for the joint classification task of title and comment, the title and comment are mainly spliced into a single text in sequence as the input of the neural network model. This scheme is simple to implement and can directly use the main structure of the single text classification model.

[0056] The inventor found that when the title information and the comment information are text-spliced without any isolation, the information on both sides is completely mixed. Further, it is not possible to determine whether the characteristic information such as the violation information is provided by the title or the comment, resulting in a large number of false judgments in actual application.

[0057] And only the comment is detected, the content of the comment itself belonging to the characteristic comment (such as the violation comment) can be intercepted. But some comments belong to normal content, only by combining the theme content can it be determined whether it is really in violation. For this situation, if the comment and the theme are text-spliced, it is judged by a neural network model whether the comment is in violation. In actual application, the neural network model will face suspicious local semantic information when making a judgment, and it is not possible to determine whether it is provided by the title or the comment, resulting in a large number of false judgments in actual application.

[0058] Therefore, the present disclosure proposes a comment detection method, medium, device and computing device, which asymmetrically processes the title information and the comment information, and ensures that the title information is only an auxiliary for determining whether the comment information is sensitive, and the determination of sensitive information is mainly based on the content of the comment information.

[0059] After introducing the basic principles of the present disclosure, the various non-limiting embodiments of the present disclosure will be specifically introduced below.

[0060] Overview of Application Scenarios

[0061] Figure 1 An application scenario example diagram according to the embodiments of the present disclosure is schematically shown. As shown in the figure, the application scenario includes a client 110 and a server 120. For example, in the application scenario, the characteristic comment can be a violation comment. Figure 1

[0062] For example, the client 110 includes a large amount of title information and comment information, wherein one title information can correspond to at least one comment information. When a user (not marked in the figure) inputs the comment information under a certain title information in the client 110, the comment information will be sent to the server 120 by the client 110. The server 120 determines whether the comment information is a violation comment. If the server 120 determines that the comment information is a violation comment, the comment information is intercepted; if the server 120 determines that the comment information is not a violation comment, the comment information is not intercepted, so that the user can see the comment information displayed under the title information through the client 110.

[0063] ​Exemplarily, the client 110 can be a PC computer, a mobile phone, or a tablet computer. The server 120 and the client 110 can be connected by wire, or can be connected through a wireless network or Bluetooth, which is not limited in the present disclosure.

[0064] In addition, the application scenario can further include a storage device (not shown in the figure) for storing title information or comment information, which can be located in the client 110 or in the server 120.

[0065] Exemplary Method

[0066] The method for comment monitoring according to the exemplary embodiments of the present disclosure will be described below in combination with the application scenario of Figure 1 , and with reference to Figure 2 . It should be noted that the above-mentioned application scenario is only shown for the purpose of facilitating the understanding of the spirit and principles of the present disclosure, and the embodiments of the present disclosure are not limited in this respect. On the contrary, the embodiments of the present disclosure can be applied to any applicable scenario.

[0067] Figure 2 A flow chart of the method for comment monitoring according to an embodiment of the present disclosure is schematically shown. As shown in Figure 2 , the method for comment monitoring includes:

[0068] S201, obtaining a target comment and a subject text corresponding to the target comment.

[0069] Exemplarily, one subject text can correspond to one target comment, or can correspond to multiple target comments.

[0070] In addition, the subject text can be title information of a news, and the target comment can be comment information under the news title, which is not limited in the present disclosure.

[0071] The subject text and the target comment can be obtained by calling from a storage device, or can be received from a client.

[0072] S202, obtaining a threshold value vector of the subject text, the threshold value vector representing semantic features of the subject text, and the threshold value vector being used for assisting the classification of the target comment.

[0073] In the processing, the subject text and the target comment are processed separately. For the subject text, its threshold value vector is extracted.

[0074] The threshold value vector reflects semantic features of the topic text. For example, taking a news title as the topic text, the semantic features reflected by the threshold value vector include the news title itself, the type of the news title, the number of words, and the like. For example, the type of the news title can be entertainment, politics, or military, and the like. The same news title can correspond to different threshold value vectors in different types.

[0075] For example, the threshold value is a decimal number between 0 and 1. 0 represents complete closing, and 1 represents complete opening. The greater the value, the greater the influence on the corresponding dimension. If the type of the news title is military, the corresponding dimension of the military in the threshold value vector is recorded as 1, and the corresponding dimensions of the political and entertainment types in the threshold value vector are recorded as 0.

[0076] The threshold value vector can be used as an auxiliary in the subsequent monitoring of the target comment in the form of a level to determine whether the target comment is a feature comment.

[0077] S203, performing semantic feature extraction on the target comment to obtain a signal value vector.

[0078] The signal value vector reflects the semantic features of the target comment. Specifically, the signal value vector can be obtained by one semantic feature extraction, or can be obtained by multiple semantic feature extractions, wherein the output of the previous semantic feature extraction is the input of the subsequent semantic feature extraction.

[0079] The signal value vector is different from the threshold value vector. Compared with the threshold value vector, the signal value vector can contain more semantic information, while the threshold value vector more embodies the category features.

[0080] It should be understood that the signal value vector obtained after multiple semantic feature extractions is less useless information and more useful information than the signal value vector obtained after one semantic feature extraction.

[0081] S204, performing feature fusion on the threshold value vector and the signal value vector to obtain a fusion vector corresponding to the target comment.

[0082] For example, when performing feature fusion on the threshold value vector and the signal value vector, the feature fusion manner can be to multiply the two vectors, for example, the threshold value vector is [a1, a2, …, an], the signal value vector is [b1, b2, …, bn], and the fusion vector obtained by multiplying the two vectors is [a1*b1, a2*b2, …, an*bn]. However, the specific implementation of feature fusion is not limited in the present disclosure, for example, a convolution function can be introduced to perform feature fusion on the threshold value vector and the signal value vector by convolution multiplication, and the like.

[0083] It can be understood that the threshold value vector and the signal value vector are asymmetrically processed, the threshold value vector of the subject text is taken as a constraint, the target review is taken as a guide, and the two are fused. The fused vector can be considered as a signal value vector assisted by the threshold value vector. The fused vector contains at least two kinds of feature information, respectively, semantic information of the target review and type information of the subject text, etc.

[0084] In S205, whether the target review is a feature review is determined according to the fused vector.

[0085] In the present disclosure, for a target review and a subject text corresponding to the target review, the target review after feature fusion can correspond to a fused vector. By further analyzing the fused vector, the target review can be monitored.

[0086] If a subject text corresponds to multiple target reviews, each of the multiple target reviews will correspond to a fused vector. Specifically, when processing the fused vector, each fused vector can be processed independently, or multiple fused vectors under the subject text can be processed uniformly.

[0087] For example, the fused vector can be processed by a fully connected layer. The fully connected layer is a linear transformation layer, which linearly converts the fused vector from one feature space to another feature space in the form of matrix vector multiplication. Through the above steps, the local information with category distinction in the convolution layer or the pooling layer can be integrated, and finally the features of the fused vector are synthesized.

[0088] Generally, news title information is published after being reviewed by relevant agencies, so there is no violation phenomenon, that is, the threshold value vector part of the fused vector does not contain feature information; while the review is published freely by the user, which may contain sensitive information and other feature information, so the signal value vector corresponding to the target review is extracted, and after feature extraction, whether the target review is a feature review (specifically, it can be a violation review, which is taken as an example in the following) can be determined by introducing a normalization function.

[0089] In the embodiments of the present disclosure, by separately processing the topic text and the target review, the threshold value vector and the signal value vector are obtained, and then the threshold value vector and the signal value vector are fused. This method can avoid the premature fusion of the topic text and the target review, and interfere with the extraction of the complete semantic features of each other. In addition, the threshold value vector of the topic text is obtained, which is used to assist the classification of the target review, that is, the topic text is used as an auxiliary judgment for whether the target review is a feature review, so as to avoid relying too much on the content of the topic text, and to determine whether the target review is a feature review mainly based on the content of the target review, so as to accurately judge whether the local semantic information is included in the review when facing suspicious local semantic information, and improve the reliability of the review monitoring.

[0090] In some embodiments, before determining the threshold value vector and the signal value vector, the topic text and the target review can be preprocessed. The text preprocessing can be performed in the following manner:

[0091] a. The continuous text sequence is divided into a plurality of subunits (denoted as words). Specifically, each Chinese character can be divided into a word, each punctuation mark can be divided into a word, and continuous numbers or English letters can be divided into a word. This operation is used for subsequent generation of a matrix based on words as units.

[0092] b. A word embedding matrix is constructed according to a dictionary. The dictionary includes at least one (number, word) key-value pair, and the number and the word correspond one by one. The word embedding matrix is composed of (for example, number, 128-dimensional vector) key-value pairs, and the number of key-value pairs is consistent with the number of words in the dictionary. It can be understood that each word is converted into a 128-dimensional vector. The 128-dimensional vector represents the characteristics of the word. It should be understood that the 128-dimensional vector is only an example, and the present disclosure does not limit the specific dimension.

[0093] c. After each word is pre-converted into a 128-dimensional vector, for each word in the target review and the topic text, the corresponding number is first queried from the dictionary, and then the corresponding 128-dimensional vector in the word embedding matrix is extracted according to the number. Finally, the target review and the topic text can be converted from a text sequence into a vector matrix.

[0094] d. Since the matrix dimensions need to be unified during convolution calculation, it is necessary to traverse all the topic texts and the corresponding target reviews involved in the calculation. For example, if the maximum number of words in the topic text is N1 and the maximum number of words in the target review is N2. For the topic text with a length less than N1 or the target review with a length less than N2. A 128-dimensional zero vector is used to pad at the end of the matrix. Finally, the vector matrix of all the topic texts has the same size, and the matrix of all the target reviews also has the same size.

[0095] After text preprocessing, the subject text and the target review can be processed in the form of a matrix or a vector.

[0096] For example, for the calculation of the threshold value vector, S202 has been described above, and the threshold value vector can be obtained by a storage device. In some embodiments, obtaining the threshold value vector of the subject text can include: in response to the threshold value vector being stored, reading the stored threshold value vector.

[0097] This mode corresponds to a subject text corresponding to multiple target reviews. It can be understood that each time a fusion vector corresponding to a target review is calculated separately, the threshold value vector of the subject text corresponding to the target review and the signal value vector of the target review itself need to be calculated first. Therefore, the threshold value vector of the subject text can be cached in the storage device. It can be used for subsequent feature fusion of other target reviews of the subject text.

[0098] In addition, in some embodiments, the threshold value vector can be obtained by Figure 3 The change in operation rate after caching the threshold value vector is shown intuitively: Figure 3 In the above, by caching the threshold value vector, the calculation time of the detection process can be reduced. If the calculation time of each subject text corresponding to a unique target review is 100%, under the caching mechanism of the present solution, two target reviews can be reduced to 80%, and 10 target reviews can be reduced to 60%.

[0099] In the embodiments of the present disclosure, a caching mechanism of intermediate calculation results is introduced. Before feature fusion, the processing of the subject text and the target review is independent and there is no intersection, and this low coupling structure facilitates caching the threshold value vector calculated from the subject text. Therefore, it can adapt to the actual situation of "one news title corresponding to multiple reviews" in the application scenario. Therefore, by caching the threshold value vector, the calculation overhead can be effectively reduced and the prediction speed can be improved.

[0100] Correspondingly, if the threshold value vector corresponding to the subject text needs to be calculated for the first time, as an example, obtaining the threshold value vector of the subject text can include: in response to the threshold value vector of the subject text not being stored, extracting semantic features of the subject text to obtain a first semantic feature matrix; extracting semantic features of the first semantic feature matrix to obtain a second semantic feature matrix; performing dimension compression on the second semantic feature matrix to obtain a first semantic feature vector; and performing normalization processing on the first semantic feature vector to obtain the threshold value vector of the subject text.

[0101] As an example, since the threshold value vector contains fewer semantic features, more convolution layers need to be introduced to extract deep semantic features, so that the semantic features represented by the threshold value vector are more abstract.

[0102] For example, the subject text is first respectively subjected to a first convolutional layer and a second convolutional layer, then subjected to a dimension reduction processing after a pooling layer, and finally subjected to normalization by a normalization layer.

[0103] Specifically, the first convolutional layer has a convolution kernel of 5x128, a step of 1, and a filter depth of 128, and uses a linear rectifier function (ReLU) as an activation function. The convolutional layer is responsible for extracting shallow semantic features. The input of the first convolutional layer is an N1x128 two-dimensional matrix, and the output is an (N1-4)x128 low-level semantic feature matrix, which is used as a first semantic feature matrix.

[0104] The second convolutional layer has a convolution kernel of 5x128, a step of 1, and a filter depth of 128, and also uses a linear rectifier function (ReLU) as an activation function. The convolutional layer is responsible for extracting deep semantic features. The input of the second convolutional layer is an (N1-4)x128 two-dimensional matrix, and the output is an (N1-8)x128 high-level semantic feature matrix. The high-level semantic feature matrix is used as a second semantic feature matrix.

[0105] The pooling layer uses a max-pooling technique and is responsible for compressing the second semantic feature matrix into a semantic feature vector. The input of the pooling layer is an (N1-8)x128 two-dimensional matrix, and the output is a 1x128 one-dimensional vector.

[0106] The normalization layer can use a sigmoid function to process each dimension of the semantic feature vector as a floating-point number between 0 and 1, and finally output a threshold value vector corresponding to the subject text.

[0107] As for how to determine the signal value vector, on the basis of the above embodiment, for an optional implementation, the semantic feature extraction of the target review to obtain the signal value vector can include: extracting the semantic features of the target review to obtain a third semantic feature matrix; and performing dimension compression on the third semantic feature matrix to obtain the signal value vector of the target review.

[0108] Further, the semantic feature extraction of the target review to obtain the third semantic feature matrix can include: performing text preprocessing on the target review to obtain a review vector matrix corresponding to the target review; and extracting the semantic features of the review vector matrix corresponding to the target review to obtain the third semantic feature matrix.

[0109] For example, the semantic feature extraction of the target review to obtain the signal value vector can be performed by one convolutional layer and one pooling layer, as follows:

[0110] Specifically, the input of the convolution layer can be an N2x128 two-dimensional matrix, and the output is an (N2-4)x128 low-level semantic feature matrix. Since the signal value vector needs to retain more semantic features of the target review, fewer convolution layers are used to extract semantic features thereof. The (N2-4)x128 low-level semantic feature matrix is the third semantic feature matrix described above.

[0111] Further, the (N2-4)x128 low-level semantic feature matrix is input into a pooling layer, and a 1x128 one-dimensional vector is output. The role of the pooling layer is to compress the matrix, and it does not have the function of semantic feature extraction, and can filter out some useless vectors. The output 1x128 one-dimensional vector is the signal value vector described above.

[0112] Optionally, according to the fusion vector, determining whether the target review is a feature review can include: determining a probability value of the target review being a feature review according to the fusion vector; in response to the probability value being greater than a feature probability threshold, determining that the target review is a feature review; and in response to the probability value being less than or equal to the feature probability threshold, determining that the target review is not a feature review.

[0113] For example, the feature probability threshold can be 0.5, or 0.6, which is not limited in the present disclosure.

[0114] In actual applications, the feature review can be a violation review, a highlight review, or the like.

[0115] In some embodiments corresponding to the above-mentioned embodiments, determining the probability value of the target review being a feature review can include: performing dimension compression on the fusion vector to obtain a feature fusion vector; and performing normalization processing on the feature fusion vector to obtain a probability value vector, the probability value vector including the probability value of the target review being a feature review.

[0116] For example, the corresponding dimensions of the threshold value vector and the signal value vector are multiplied to obtain the fusion vector. The fusion vector is input into a full connection layer, the transformation scale of the full connection layer is 128x64, and a linear rectifier function (ReLU) is used as an activation function. This layer is responsible for further screening the semantic features of the fusion vector, while completing dimension compression. The input of the full connection layer is a 1x128 one-dimensional vector, and the output is a 1x64 one-dimensional vector, which is the feature fusion vector.

[0117] The feature fusion vector is normalized by the output layer. Specifically, the classification layer first performs a scale transformation of 64x2, and then uses a softmax function to normalize the output vector. The softmax function is a normalization exponential function that can compress one real vector into another real vector. In the compressed vector, the vector value of each dimension is between 0 and 1. Moreover, the sum of all dimension vector values is 1, i.e., normalization. Specifically, the input of the classification layer is a one-dimensional vector of 1x64, and the output is a one-dimensional vector of 1x2. The two elements in the vector are both floating-point numbers between 0 and 1, and the sum is 1, representing the probability that the target review is a feature review and the probability that the target review is not a feature review, respectively.

[0118] In some embodiments, it can also include determining whether the target review is a feature review according to the target review and the topic text corresponding to the target review through a review classification model, the review classification model being used to determine whether a review is a feature review.

[0119] That is, the target review and the topic text are monitored by the review classification model. For example, the present disclosure uses an asymmetric gated neural network.

[0120] Further, in some examples, the review classification model can be trained in the following way: obtaining a plurality of training data packets, the training data packets including at least one sample group, the sample group including a sample topic text and at least one sample review corresponding to each sample topic text; training the review classification model based on the plurality of training data packets to obtain a review classification model that meets the preset requirements.

[0121] For example, the training samples can be composed in the following way:

[0122] 1. First, construct an original sample pair. Specifically, a topic text and a target review corresponding thereto are taken as a sample pair. Therefore, if there are m target reviews under a topic text, there are m sample pairs corresponding to the topic text. Similarly, if there are n topic texts in the entire database, there are m*n sample pairs in the training samples. After the sample pairs are constructed, all sample pairs under the same topic text are grouped together. Therefore, based on the above conditions, a total of n groups can be finally obtained.

[0123] 2. Within each group, the sample pairs are randomly shuffled; that is, for target comments under the same topic text, their order is randomized. Simultaneously, each group is further divided, with a sample threshold T set. If the number of samples in a group is less than or equal to T, the group is not re-divided; if the number of samples in a group is greater than T, the portion exceeding T needs to be regenerated into a new group. In one specific implementation, the T target comments from the original group are retained, and the remaining portion is re-formed into a new group. The new group shares the same topic text as the original group.

[0124] 3. After the above processing, the groups are randomly shuffled, so that the topic text corresponding to each group is randomly distributed.

[0125] 4. Construct training data packets and set a data packet threshold B. Specifically, this can be achieved by selecting multiple groups. When the sum of sample pairs contained in these selected groups is B, these groups are grouped into one training data packet. Repeat this process to construct multiple training data packets. If the remaining groups contain fewer sample pairs than the data packet threshold B, these sample pairs can be allocated to the previously constructed data packets. Thus, each data packet contains at least B sample pairs. As a historical example, the data packet threshold B can be set to 200, and the sample threshold T can be set to 10.

[0126] Therefore, this disclosure uses limited shuffling instead of completely random sampling when constructing training data packets, to build training data packets that are more suitable for the actual situation. Specific processing includes grouping by topic text type, shuffling target comments within each group, and shuffling between groups. Each generated training data packet consists of several groups, and the target comments within each group belong to the same topic text. The above steps can be achieved through... Figure 4 To show, such as Figure 4 As shown, Figure 4 A training data packet generation method according to another embodiment of the present disclosure is illustrated schematically, the method comprising:

[0127] S401. Divide the sample pairs into multiple groups according to the type of topic text.

[0128] For example, if the subject texts are of the types such as political, military, and humanities, then samples with political subject texts will be divided into multiple groups, where the subject texts of samples within these multiple groups are all political; similarly, samples with military subject texts will also be divided into multiple groups; and samples with humanities subject texts will also be divided into multiple groups.

[0129] S402. Randomly shuffle the target comments for each group.

[0130] For example, it can disrupt the target reviews originally arranged in chronological order. This method can enhance the randomness of the training data package.

[0131] S403, determine whether the target reviews in each group exceed the sample threshold T.

[0132] If yes, execute step S405; if no, execute step S404.

[0133] S404, redivide the reviews exceeding the sample threshold T in the group into a new group.

[0134] For example, the redivided group can also randomly disrupt the target reviews again.

[0135] S405, randomly disrupt each group.

[0136] After randomly disrupting the target reviews in the group, each group will be further randomly disrupted. After the groups are randomly disrupted, the groups originally arranged according to the theme text type become randomly arranged groups.

[0137] S406, move the sample pairs of a group into the training data package.

[0138] For example, when constructing the training data package, the sample pairs in one group can be moved at a time, or the sample pairs in multiple groups can be moved at a time.

[0139] S407, determine whether the sample pairs in the training data package exceed the data package threshold B.

[0140] In order to ensure the speed and accuracy of model training, it is necessary to require that the sample pairs of the training data package do not exceed a threshold.

[0141] If no, return to step S406; if yes, execute step S408.

[0142] S408, output a constructed training data package.

[0143] In the embodiments of the present disclosure, when constructing the training sample of the comment classification model, on the one hand, the target comments are grouped according to the theme text, thereby enhancing the comparison and recognition ability of the neural network model to the details of the target comments; on the other hand, the target comments in the group are randomly shuffled, and the group is also randomly shuffled, which can avoid that all the targets come from the same theme text or the target comments under the same theme text during the model training. Since the limited shuffling training data package construction method is used, when the training data distribution is unbalanced, especially when the target comment types under some theme text are too concentrated and the sample pair richness is limited, this method as an auxiliary classification learning target can effectively enhance the generalization ability of the neural network model on unknown data. The accuracy of the operation after the model training is improved.

[0144] After the training sample is constructed, it needs to be trained. In some embodiments, the comment classification model is trained based on the plurality of training data packages to obtain a comment classification model meeting the preset requirements, which can include: performing contrastive learning training on the comment classification model based on the plurality of training data packages to obtain a comment classification model meeting the preset requirements.

[0145] It should be understood that the contrastive learning training is a reinforcement learning training on the basis training. The basis training can be understood as inputting the previously generated training sample into the comment classification model, so that the comment classification model has the ability to output the probability of whether the target comment is a feature comment after being trained. The contrastive learning training is to strengthen the comparison and analysis ability of the comment classification model to the target comments under the same theme text. That is, the target comments with the same semantics can be pulled in, and the target comments with different semantics can be pushed away.

[0146] As an example, the comment classification model is trained based on the plurality of training data packages to obtain a comment classification model meeting the preset requirements, which can include:

[0147] According to each training data package in the plurality of training data packages, the comment classification model is trained based on the following contrastive learning target:

[0148]

[0149]

[0150] Wherein, L-CONTRA is the contrastive learning target value, L-CONTRA i,j is the contrastive learning target of the jth comment of the sample group i, sim() is a cosine similarity function; is an indicator function; τ is a temperature adjustment factor, l i,jrepresents the classification value of the jth review of the sample group identified as i, wherein the classification value of the feature review is 1, and the classification value of the non-feature review is 0; N represents the number of sample group identifications; K i represents the number of sample groups; r i,j represents the fusion vector corresponding to the jth review of the sample group identified as i; r i,j represents the fusion vector corresponding to the kth review of the sample group identified as i; exp() is an exponential function;

[0151] A review classification model is obtained when the contrast learning target value is less than or equal to a set value.

[0152] As an example, the input data of the contrast learning training includes three parts: a fusion vector, a sample group identification, and a classification value vector. The fusion vector is obtained by performing feature fusion on the threshold value vector and the signal value vector. The sample group identification is the label of each group in the training data packet. Different groups of texts with the same theme have different group labels, and different groups of texts with the same theme have the same group label. For example, the groups are sequentially numbered as 1, 2, 3, …, and all samples in each group have the same theme text. Finally, the sample group identifications of all sample pairs in a data packet can form a sample group identification vector, which can be in the form of [1, 1, 2, 2, 2, 2, 3, 3, 3, …]. The classification value vector is composed of multiple classification values, wherein 0 represents that the target review is not a feature review, and 1 represents that the target review is a feature review. For example, the classification value vector can be in the form of [0, 0, 1, 0, 1, 1, 0, 0, 1, …].

[0153] In the above formula, the denominator represents the sum of the cosine similarities of all reviews in the training data packet with the sample group identification i except the jth review. The numerator represents the sum of the cosine similarities of all feature reviews in the training data packet with the sample group identification i except the jth review.

[0154] The goal of the contrast learning training is to make the value of the score tend to 1, i.e., the value of the logarithmic function tends to 0, and thus the value of the entire formula tends to 0.

[0155] In the formula, since the denominator is composed of the sum of the cosine similarities of the feature reviews and the sum of the cosine similarities of the non-feature reviews. Therefore, when the score value tends to 1, it can be considered that the cosine similarity of the non-feature review can be ignored (0). Further, it can be concluded that the similarity of the feature review (violation review) is pulled in, and the similarity between the non-feature review (normal review) and the feature review is pushed away.

[0156] It should be understood that the smaller the contrast learning target value, the better the final training effect and the more accurate the calculation result. The set value can be 0.5 or 0.3. The present disclosure does not limit it.

[0157] The above method can be further manifested by Figure 5a , Figure 5a The contrast learning training method according to still another embodiment of the present disclosure is schematically shown, which comprises: Figure 5a as shown in the figure, the method comprises:

[0158] S501, obtaining a fusion vector after feature fusion.

[0159] For example, the dimension of the fusion vector obtained here is 1*128.

[0160] S502, normalizing the fusion vector.

[0161] After normalization, the modulus value of the vector is 1.

[0162] S503, transposing the normalized fusion vector.

[0163] Here is the transpose of the matrix, which is used for subsequent matrix multiplication operation.

[0164] S504, multiplying the transposed fusion vector with the fusion vector before transposition.

[0165] The matrix multiplication here can be written as: the multiplication of the 1*128 vector and the 128*1 vector, obtaining the cosine similarity value of the fusion vector.

[0166] S505, calculating the cosine similarity sum of all feature reviews in a group according to the sample grouping identifier and the classification value vector.

[0167] S506, calculating the cosine similarity sum of all target reviews in a group according to the sample grouping identifier and the classification value vector.

[0168] S507, taking the quotient of the cosine similarity sum of all feature reviews and the cosine similarity of the target review, and performing logarithmic operation on the quotient to obtain the contrast learning target value.

[0169] Among them, steps S505, S506 and S507 are the same as the aforementioned contrast learning training method, which will not be described here.

[0170] S508, through repeated iterative calculation, when the contrast learning target value tends to 0, the contrast learning training is completed.

[0171] The iterative calculation here can be understood as: in the foregoing formula, the values of i and j are constantly replaced. By calculating the contrast learning target value of each target review and making it approach 0, the review classification model is finally trained.

[0172] Figure 5b It is shown that before the contrast learning training is performed, the classification value vector and the sample grouping identifier also need to be preprocessed, that is, a contrast learning training preprocessing method can be called. After the matrix is transposed and multiplied, the joint mask is obtained. The joint mask reflects the sample grouping identifier and the classification value vector of a sample pair, which will participate in the calculation of the cosine similarity sum of all feature reviews in the group and the cosine similarity sum of all target reviews in the group.

[0173] In the embodiments of the present disclosure, an enhanced training method based on contrast learning is used. The contrast object is a certain fusion vector after feature fusion, and the contrast range is any group in the training data packet, except for the target review corresponding to the fusion vector. The contrast process uses matrix operations of the sample grouping identifier and the classification value of the target review. Finally, the judgment ability of the review classification model for each target review under the same topic text can be enhanced, and the generalization ability is improved.

[0174] Next, the review monitoring method proposed by the present disclosure is shown through complete steps, Figure 6a The review monitoring method according to another embodiment of the present disclosure is schematically shown as Figure 6a As shown in the figure, the method comprises:

[0175] S601, preprocessing the target review to obtain a review matrix.

[0176] Each "word" in the target review is taken as a 128-dimensional vector, and finally the target review is preprocessed into a review matrix. The number of rows of the review matrix is unified to the maximum number of words N2 in the target review among all target reviews.

[0177] S602, preprocessing the topic text to obtain a topic vector matrix.

[0178] Each "word" in the topic text is taken as a 128-dimensional vector, and finally the topic text is preprocessed into a topic vector matrix. The number of rows of the topic vector matrix is unified to the maximum number of words N1 in the topic text among all topic texts.

[0179] S603, extracting semantic features of the review vector matrix through a convolution layer to obtain a third semantic feature matrix.

[0180] Generally, the review vector matrix only needs to extract semantic features, thereby ensuring the integrity of the semantic features of the target review to the greatest extent.

[0181] S604, semantic features of the topic vector matrix are extracted through the first layer of convolutional layers, to obtain a first semantic feature matrix.

[0182] Similarly, semantic features of the topic text are extracted through the convolutional layers. Since the topic text is used to assist in judging the violation of the target review, deep semantic features of the topic text need to be extracted.

[0183] S605, the third semantic feature matrix is processed by the dimension reduction of the pooling layer, to obtain a signal value vector.

[0184] The main role of the pooling layer is to compress the amount of data and parameters, and reduce overfitting. The following S607 is consistent with the role of the present step. The dimension of the processed signal value vector is 1*128.

[0185] S606, semantic features of the first semantic feature matrix are extracted through the second layer of convolutional layers, to obtain a second semantic feature matrix.

[0186] S607, the second semantic feature matrix is processed by the dimension reduction of the pooling layer, and is normalized, to obtain a threshold value vector.

[0187] The dimension of the processed threshold value vector is 1*128.

[0188] S608, the threshold value vector is multiplied by the corresponding dimension of the signal value vector, to obtain a fusion vector.

[0189] For example, the dimension of the processed threshold value vector and the signal value vector is 1*128. If the threshold value vector is (a1, a2, a3…an), and the signal value vector is (b1, b2, b3…bn), the fusion vector can be expressed as (a1*b1, a2*b2, a3*b3…an*bn).

[0190] S609, the fusion vector is further extracted through the fully connected layer.

[0191] S610, the probability value of the target review being a feature review is output through the output layer.

[0192] The vector dimension of the output of the output layer is 1*2, which can be written as (c1, c2). Wherein, c1 represents the probability of the target review being a feature review; c2 represents the probability of the target review not being a feature review, and c1+c2=1.

[0193] In addition, Figure 6b The flow result diagram of the comment monitoring method according to another embodiment of the present disclosure is schematically shown. Through Figure 6b The results of the threshold value vector and the signal value vector can be reflected, and the results of the fusion vector.

[0194] The comment monitoring method provided by the embodiments of the present disclosure first separately processes the theme text and the target comment, and then fuses the processed theme text and the target comment. This method can separately judge the features of the theme text and the target comment. Meanwhile, the theme text can be used as an auxiliary to further judge whether the target comment is a feature comment, thereby improving the monitoring accuracy.

[0195] Exemplary Medium

[0196] After introducing the method of the exemplary embodiments of the present disclosure, next, with reference to Figure 7 The storage medium of the exemplary embodiments of the present disclosure is described.

[0197] With reference to Figure 7 The program product 70 for implementing the above method according to the embodiments of the present disclosure is described, which can adopt a portable compact disc read-only memory (CD-ROM) and includes program code, and can run on a terminal device, such as a personal computer. However, the program product of the present disclosure is not limited to this.

[0198] The program product can adopt any combination of one or more readable media. The readable medium can be a readable signal medium or a readable storage medium. The readable storage medium may, for example, be but is not limited to an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, device or apparatus, or any combination of the above. More specific examples (non-exhaustive list) of the readable storage medium include an electrical connection having one or more wires, a portable disc, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disc read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the above.

[0199] The readable signal medium can include a data signal propagating in a baseband or as part of a carrier wave, in which readable program code is carried. Such a propagating data signal can take various forms, including but not limited to an electromagnetic signal, an optical signal, or any suitable combination of the above. The readable signal medium can also be any readable medium other than the readable storage medium.

[0200] The program code for carrying out operations of the present disclosure can be written in any combination of one or more programming languages, including an object oriented programming language such as Java, C++ or the like, and conventional procedural programming languages, such as the "C" programming language or similar programming languages. The program code can execute entirely on the user's computing device, partly on the user's device, as a stand-alone software package, partly on the user's device and partly on a remote computing device or entirely on the remote computing device or server. In the latter scenario, the remote computing device can be connected to the user's computing device through any type of network, including a local area network (LAN) or a wide area network (WAN), or the connection can be made to an external computing device, as in the case of the Internet using an Internet Service Provider (ISP).

[0201] Exemplary Apparatus

[0202] Having introduced the medium of the exemplary embodiments of the present disclosure, next, reference is made to the accompanying drawings, in which Figure 8 The comment monitoring apparatus of the exemplary embodiments of the present disclosure is described, which is used to implement the method in any of the method embodiments, and has similar implementation principles and technical effects, which will not be described here.

[0203] Figure 8 In the comment monitoring apparatus 800, the first obtaining module 801 is configured to obtain a target comment and a subject text corresponding to the target comment.

[0204] The first obtaining module 801 is configured to obtain a target comment and a subject text corresponding to the target comment.

[0205] The second obtaining module 802 is configured to obtain a threshold value vector of the subject text, the threshold value vector representing semantic features of the subject text, and the threshold value vector being used to assist in classifying the target comment.

[0206] The third obtaining module 803 is configured to perform semantic feature extraction on the target comment to obtain a signal value vector.

[0207] The feature fusion module 804 is configured to perform feature fusion on the threshold value vector and the signal value vector to obtain a fusion vector corresponding to the target comment.

[0208] The determining module 805 is configured to determine, according to the fusion vector, whether the target comment is a feature comment.

[0209] In an embodiment of the present disclosure, the second obtaining module is specifically configured to: in response to the threshold value vector being stored, read the stored threshold value vector.

[0210] In another embodiment of the present disclosure, the second obtaining module 802 comprises a first convolutional layer, a second convolutional layer, a first pooling layer and a normalization layer (the three are not marked in the figure), wherein: the first convolutional layer is configured to extract semantic features of the topic text to obtain a first semantic feature matrix in response to a threshold value vector of the topic text; the second convolutional layer is configured to extract semantic features of the first semantic feature matrix to obtain a second semantic feature matrix; the pooling layer is configured to perform dimension compression on the second semantic feature matrix to obtain a first semantic feature vector; and the normalization layer is configured to perform normalization processing on the first semantic feature vector to obtain the threshold value vector of the topic text.

[0211] In yet another embodiment of the present disclosure, the third obtaining module 803 comprises a third convolutional layer and a second pooling layer (the two are not marked in the figure), wherein: the third convolutional layer is configured to extract semantic features of the target review to obtain a third semantic feature matrix; and the second pooling layer is configured to perform dimension compression on the third semantic feature matrix to obtain a signal value vector of the target review.

[0212] In still another embodiment of the present disclosure, the third convolutional layer is specifically configured to: perform text preprocessing on the target review to obtain a review vector matrix corresponding to the target review; and extract semantic features of the review vector matrix corresponding to the target review to obtain the third semantic feature matrix.

[0213] In still another embodiment of the present disclosure, the determining module 805 is specifically configured to: determine a probability value of the target review being a feature review according to the fusion vector; determine that the target review is a feature review in response to the probability value being greater than a feature probability threshold value; and determine that the target review is not a feature review in response to the probability value being less than or equal to the feature probability threshold value.

[0214] In still another embodiment of the present disclosure, the determining module 805 is specifically configured to: perform dimension compression on the fusion vector to obtain a feature fusion vector; and perform normalization processing on the feature fusion vector to obtain a probability value vector, the probability value vector comprising the probability value of the target review being a feature review.

[0215] In still another embodiment of the present disclosure, the determining module 805 is further configured to: determine whether the target review is a feature review according to the target review and the topic text corresponding to the target review through a review classification model, the review classification model being configured to determine whether a review is a feature review.

[0216] In a further example of the present disclosure, a training module (not shown in the figure) is further included for training the comment classification model by: obtaining a plurality of training data packets, each of which includes at least one sample group, and each of the sample groups includes a sample topic text and at least one sample comment corresponding to the sample topic text; and training the comment classification model based on the plurality of training data packets to obtain a comment classification model meeting preset requirements.

[0217] In a further example of the present disclosure, the training module is specifically configured to: train the comment classification model based on the plurality of training data packets by contrastive learning to obtain a comment classification model meeting preset requirements.

[0218] In a further example of the present disclosure, the training module is specifically configured to: train the comment classification model based on each of the plurality of training data packets by contrastive learning using the following contrastive learning objective:

[0219]

[0220]

[0221] wherein, L-CONTRA is the contrastive learning objective value, L-CONTRA i,j is the contrastive learning objective of the jth comment of the sample group with the identifier i, and sim() is a cosine similarity function; is an indicator function; τ is a temperature adjustment factor, l i,j represents the classification value of the jth comment of the sample group with the identifier i, wherein the classification value of a feature comment is 1, and the classification value of a non-feature comment is 0; N represents the number of sample group identifiers; K i represents the number of sample groups; r i,j represents the fusion vector corresponding to the jth comment of the sample group with the identifier i; r i,j represents the fusion vector corresponding to the kth comment of the sample group with the identifier i; exp() is an exponential function; and the comment classification model with a contrastive learning objective value less than or equal to a set value is obtained.

[0222] Exemplary Computing Device

[0223] After introducing the method, medium and device of the example implementation of the present disclosure, next, reference is made to Figure 9 The computing device of the example implementation of the present disclosure is described.

[0224] Figure 9 The displayed computing device 90 is merely an example and should not impose any limitation on the functions and use range of the example implementation of the present disclosure.

[0225] As Figure 9As shown, the computing device 90 is in the form of a general-purpose computing device. The components of the computing device 90 can include, but are not limited to, the at least one processing unit 901, the at least one storage unit 902, and a bus 903 that connects the various system components, including the processing unit 901 and the storage unit 902.

[0226] The bus 903 includes a data bus, a control bus, and an address bus.

[0227] The storage unit 902 can include a readable medium that can be nonvolatile memory, such as read-only memory (ROM) 9023, and / or volatile memory, such as random access memory (RAM) 9021 and / or cache memory 9022. The storage unit 902 can further include a read-write memory, such as flash memory.

[0228] The storage unit 902 can also include a program / utility 9025 having a set of program modules 9024, including but not limited to an operating system, one or more application programs, other program modules, and program data, each of which can include implementation of a network environment, alone or in combination.

[0229] The computing device 90 can also communicate with one or more external devices 904, such as a keyboard or a pointing device, through an input / output (I / O) interface 905. Further, the computing device 90 can communicate with one or more networks, such as a local area network (LAN), a wide area network (WAN), and / or a public network, such as the Internet, through a network adapter 906. As Figure 9 As shown, the network adapter 906 communicates with the other modules of the computing device 90 through the bus 903. It should be appreciated that although the network adapter 906 is shown as a separate component of the computing device 90, the network adapter 906 can be integrated with one or more of the other components of the computing device 90, such as the processing unit 901.

[0230] It should be noted that although several units / modules or sub-units / modules of the comment monitoring apparatus are mentioned in the foregoing detailed description, such a division is merely exemplary and not mandatory. Indeed, according to embodiments of the present disclosure, features and functions of two or more units / modules described above can be embodied in one unit / module. Conversely, features and functions of one unit / module described above can be further divided into a plurality of units / modules.

[0231] Furthermore, although the operations of the online review monitoring method of the present disclosure are described in a particular, sequential order, this should not be understood as a requirement or implied that the operations be performed in anything less than the order described, and / or that all illustrated operations be performed, to achieve desirable results. Additionally or alternatively, certain steps can be omitted, combined, performed simultaneously, and / or performed separately from other steps illustrated. Additionally or alternatively, certain steps can be performed in a different order.

[0232] While the spirit and principles of the present disclosure have been described with reference to several specific implementations, it is to be understood that the disclosure is not limited to the specific implementations disclosed and that the division into aspects is not meant to imply that features from one aspect cannot be combined with features from another aspect to benefit, but is merely for ease of presentation. The present disclosure is intended to cover various modifications and equivalent arrangements included within the spirit and scope of the claims appended hereto.

Claims

1. A review monitoring method, comprising: obtaining a target review and a subject text corresponding to the target review; obtaining a threshold vector of the subject text, the threshold vector representing semantic features of the subject text, the semantic features of the subject text including types of the subject text, the threshold vector being used to assist classification of the target review, each dimension in the threshold vector corresponding to a type; extracting semantic features of the target review to obtain a signal value vector; performing feature fusion on the threshold vector and the signal value vector by multiplication to obtain a fusion vector corresponding to the target review, the fusion vector including two types of feature information: semantic information of the target review and type information of the subject text; determining, according to the fusion vector, whether the target review is a feature review, the feature review being a review with sensitive information. The obtaining of the threshold vector of the subject text comprises: in response to not storing the threshold vector of the subject text, extracting semantic features of the subject text to obtain a first semantic feature matrix; extracting semantic features of the first semantic feature matrix to obtain a second semantic feature matrix; performing dimension compression on the second semantic feature matrix to obtain a first semantic feature vector; performing normalization processing on the first semantic feature vector to obtain the threshold vector of the subject text. 2.The review monitoring method of claim 1, wherein the obtaining of the threshold vector of the subject text comprises: in response to storing the threshold vector, reading the stored threshold vector. 3.The review monitoring method of claim 1, wherein the extracting of the semantic features of the target review to obtain a signal value vector comprises: extracting semantic features of the target review to obtain a third semantic feature matrix; performing dimension compression on the third semantic feature matrix to obtain the signal value vector of the target review. 4.The review monitoring method of claim 3, wherein the extracting of the semantic features of the target review to obtain a third semantic feature matrix comprises: performing text preprocessing on the target review to obtain a review vector matrix corresponding to the target review; extracting semantic features of the review vector matrix corresponding to the target review to obtain a third semantic feature matrix. 5.The review monitoring method of claim 1, wherein the determining, according to the fusion vector, whether the target review is a feature review comprises: determining, according to the fusion vector, a probability value of the target review being a feature review; in response to the probability value being greater than a feature probability threshold, determining that the target review is a feature review; in response to the probability value being less than or equal to the feature probability threshold, determining that the target review is not a feature review. 6.The review monitoring method of claim 5, wherein the determining, according to the fusion vector, a probability value of the target review being a feature review comprises: performing dimension compression on the fusion vector to obtain a feature fusion vector; performing normalization processing on the feature fusion vector to obtain a probability value vector, the probability value vector including the probability value of the target review being a feature review.

7. The comment monitoring method of claim 1, further comprising: determining, according to the target comment and the topic text corresponding to the target comment, whether the target comment is a feature comment by a comment classification model, the comment classification model being configured to determine whether a comment is a feature comment.

8. The comment monitoring method of claim 7, wherein the comment classification model is trained by: obtaining a plurality of training data packets, each of the training data packets including at least one sample group, each of the sample groups including a sample topic text and at least one sample comment corresponding to the sample topic text; training the comment classification model based on the plurality of training data packets to obtain a comment classification model satisfying a preset requirement.

9. The comment monitoring method of claim 8, wherein the training the comment classification model based on the plurality of training data packets to obtain a comment classification model satisfying a preset requirement comprises: training the comment classification model based on the plurality of training data packets by contrastive learning to obtain a comment classification model satisfying a preset requirement.

10. The comment monitoring method of claim 9, wherein the training the comment classification model based on the plurality of training data packets by contrastive learning to obtain a comment classification model satisfying a preset requirement comprises: training the comment classification model based on each of the plurality of training data packets by contrastive learning according to a contrastive learning objective as follows: wherein, L-CONTRA is a contrastive learning target value, L-CONTRA i,j is a contrastive learning target of the jth review of the sample group identified as i, and sim() is a cosine similarity function; is an indicator function; τ is a temperature adjustment factor, l i,j represents a classification value of the jth review of the sample group identified as i, wherein the classification value of a feature review is 1, and the classification value of a non-feature review is 0; N represents the number of sample group identifications; K i represents the number of sample groups; r i,j represents a fusion vector corresponding to the jth review of the sample group identified as i; r i,j represents a fusion vector corresponding to the kth review of the sample group identified as i; exp() is an exponential function; obtaining a comment classification model with a contrastive learning objective value less than or equal to a set value.

11. A comment monitoring apparatus, comprising: a first obtaining module configured to obtain a target comment and a topic text corresponding to the target comment; a second obtaining module configured to obtain a threshold value vector of the topic text, the threshold value vector representing semantic features of the topic text, the semantic features of the topic text including types of the topic text, the threshold value vector being configured to assist classification of the target comment, each dimension of the threshold value vector corresponding to a type; a third obtaining module configured to perform semantic feature extraction on the target comment to obtain a signal value vector; a feature fusion module configured to perform feature fusion on the threshold value vector and the signal value vector by multiplication to obtain a fusion vector corresponding to the target comment, the fusion vector including two types of feature information: semantic information of the target comment and type information of the topic text; a determining module configured to determine, according to the fusion vector, whether the target comment is a feature comment, the feature comment being a comment with sensitive information; the second obtaining module includes a first convolutional layer, a second convolutional layer, a first pooling layer, and a normalization layer, wherein: the first convolutional layer is configured to extract semantic features of the topic text to obtain a first semantic feature matrix in response to the threshold value vector of the topic text not being stored; the second convolutional layer is configured to extract semantic features of the first semantic feature matrix to obtain a second semantic feature matrix; the pooling layer is configured to perform dimension compression on the second semantic feature matrix to obtain a first semantic feature vector; and The normalization layer is configured to normalize the first semantic feature vector to obtain a threshold value vector of the topic text.

12. The comment monitoring apparatus of claim 11, wherein, The second obtaining module is specifically configured to: In response to storing the threshold value vector, read the stored threshold value vector.

13. The comment monitoring apparatus according to claim 11, wherein the third obtaining module comprises a third convolutional layer and a second pooling layer, and wherein: The third convolutional layer is configured to extract semantic features of the target comment to obtain a third semantic feature matrix. The second pooling layer is configured to compress the dimension of the third semantic feature matrix to obtain a signal value vector of the target comment.

14. The comment monitoring apparatus according to claim 13, wherein the third convolutional layer is specifically configured to: perform text preprocessing on the target comment to obtain a comment vector matrix corresponding to the target comment; and extract semantic features of the comment vector matrix corresponding to the target comment to obtain the third semantic feature matrix.

15. The comment monitoring apparatus according to claim 11, wherein the determining module is specifically configured to: determine a probability value of the target comment being a feature comment according to the fusion vector; in response to the probability value being greater than a feature probability threshold value, determine that the target comment is a feature comment; and in response to the probability value being less than or equal to the feature probability threshold value, determine that the target comment is not a feature comment.

16. The comment monitoring apparatus according to claim 15, wherein the determining module is specifically configured to: compress the dimension of the fusion vector to obtain a feature fusion vector; and perform normalization processing on the feature fusion vector to obtain a probability value vector, the probability value vector comprising the probability value of the target comment being a feature comment.

17. The comment monitoring apparatus according to claim 11, wherein the determining module is further configured to: determine whether the target comment is a feature comment according to the target comment and the topic text corresponding to the target comment by using a comment classification model configured to determine whether a comment is a feature comment.

18. The comment monitoring apparatus according to claim 17, further comprising a training module configured to: train the comment classification model by: obtaining a plurality of training data packets, each of the training data packets comprising at least one sample group, each of the sample groups comprising a sample topic text and at least one sample comment corresponding to the sample topic text; and training the comment classification model based on the plurality of training data packets to obtain a comment classification model satisfying a preset requirement.

19. The comment monitoring apparatus according to claim 18, wherein the training module is specifically configured to: train the comment classification model based on the plurality of training data packets by using a contrastive learning method to obtain a comment classification model satisfying a preset requirement.

20. The comment monitoring apparatus according to claim 19, wherein the training module is specifically configured to: train the comment classification model based on each of the plurality of training data packets by using a contrastive learning target as follows: obtain a comment classification model with a contrastive learning target value less than or equal to a set value. ​ ​ ​ ​ ​ wherein, L-CONTRA is a contrastive learning target value, L-CONTRA i,j sim() is a cosine similarity function, and sim() is a cosine similarity function. is an indicator function; τ is a temperature adjustment factor, l i,j represents the classification value of the jth review of the sample group identified as i, wherein the classification value of the feature review is 1, and the classification value of the non-feature review is 0; N represents the number of sample group identifications; K i represents the number of sample groups; r i,j represents the fusion vector corresponding to the jth review of the sample group identified as i; r i,j represents the fusion vector corresponding to the kth review of the sample group identified as i; exp() is an exponential function; ​ 21. A computer readable storage medium, having stored therein computer program instructions, which when executed implement the comment monitoring method according to any one of claims 1 to 10.

22. A computing device comprising: a memory and a processor, the memory is configured to store program instructions; the processor is configured to invoke the program instructions in the memory to perform the comment monitoring method according to any one of claims 1 to 10.

Images