Text content auditing method, related device, equipment and storage medium

Abstract

The invention discloses a text content auditing method, a related device, equipment and a storage medium, and the method comprises the steps: obtaining a to-be-audited text; performing prediction based on the to-be-audited text to obtain a probability value that the to-be-audited text relates to a sub-tag under each text topic; wherein any text subject comprises a first sub-tag, a second sub-tag and a third sub-tag, the first sub-tag represents that the content related to the text subject is compliant, the second sub-tag represents that the content related to the text subject is illegal, and the third sub-tag represents that the text subject is not related. According to the scheme, the accuracy and comprehensiveness of text content auditing can be improved.

Description

Technical Field

[0001] The present application relates to the field of natural language processing technology, and in particular to a text content review method and related devices, equipment and storage media. Background Art

[0002] Text content review is widely used in many scenarios ranging from daily communication to commercial activities, such as instant messaging and self-media.

[0003] At present, the existing text content review technology still has many shortcomings. For example, the text content review technology that relies on keyword matching cannot be applied to circumvention methods such as homophone phrase replacement and special symbol word segmentation, which affects the accuracy and comprehensiveness of text content review. Therefore, how to improve the accuracy and comprehensiveness of text content review has become an urgent problem to be solved. Summary of the invention

[0004] The main technical problem solved by this application is to provide a text content review method and related devices, equipment and storage media, which can improve the accuracy and comprehensiveness of text content review.

[0005] In order to solve the above-mentioned technical problems, the first aspect of the present application provides a text content review method, including: obtaining a text to be reviewed; making predictions based on the text to be reviewed to obtain probability values ​​of sub-labels under each text topic respectively involved in the text to be reviewed; wherein, any text topic includes a first sub-label, a second sub-label and a third sub-label, and the first sub-label represents that the content related to the text topic is compliant, the second sub-label represents that the content related to the text topic is illegal, and the third sub-label represents that the text topic is not involved.

[0006] In order to solve the above-mentioned technical problems, the second aspect of the present application provides a text content review device, including: an acquisition module and a prediction module, the acquisition module is used to obtain the text to be reviewed; the prediction module is used to make predictions based on the text to be reviewed, and obtain the probability values ​​of the sub-labels under each text topic that the text to be reviewed involves; wherein any text topic includes a first sub-label, a second sub-label and a third sub-label, and the first sub-label represents that the content related to the text topic is compliant, the second sub-label represents that the content related to the text topic is illegal, and the third sub-label represents that the text topic is not involved.

[0007] In order to solve the above technical problems, the third aspect of the present application provides an electronic device, which at least includes a memory and a processor coupled to each other, the memory at least stores program instructions, and the processor is used to execute the program instructions to implement the text content review method in the above first aspect.

[0008] In order to solve the above technical problems, the fourth aspect of the present application provides a computer-readable storage medium storing program instructions that can be executed by a processor, and the program instructions are used to implement the text content review method of the first aspect.

[0009] The above scheme obtains the text to be reviewed, and then makes predictions based on the text to be reviewed to obtain the probability values ​​of the sub-labels under each text topic involved in the text to be reviewed, and any text topic includes the first sub-label, the second sub-label and the third sub-label. The first sub-label represents the compliance of the content involved in the text topic, the second sub-label represents the violation of the content involved in the text topic, and the third sub-label represents the non-involvement of the text topic. Therefore, on the one hand, it is different from the keyword matching technology. By predicting the probability values ​​of the sub-labels under each text topic, it can alleviate the adverse effects of circumvention methods such as homophone phrase replacement and special symbol word segmentation on text content review as much as possible, which helps to improve the accuracy of text content. On the other hand, because the probability values ​​of the sub-labels under each text topic are predicted, it can simultaneously perform review predictions from the two perspectives of text topic and positive and negative judgment (i.e., content compliance or content violation), which helps to improve the comprehensiveness of text content review. Therefore, it can improve the accuracy and comprehensiveness of text content review. BRIEF DESCRIPTION OF THE DRAWINGS

[0010] Figure 1 It is a flowchart of an embodiment of the text content review method of the present application; Figure 2 It is a schematic diagram of the framework of an embodiment of a text review model; Figure 3 This is a process diagram of an embodiment of the text content review method of the present application; Figure 4 It is a schematic diagram of the framework of an embodiment of the text content review device of the present application; Figure 5 It is a schematic diagram of the framework of an embodiment of the electronic device of the present application; Figure 6 It is a schematic diagram of a framework of an embodiment of a computer-readable storage medium of the present application. DETAILED DESCRIPTION

[0011] The scheme of the embodiment of the present application is described in detail below in conjunction with the drawings of the specification.

[0012] In the following description, for the purpose of explanation rather than limitation, specific details such as specific system structures, interfaces, and technologies are provided to facilitate a thorough understanding of the present application.

[0013] The terms "system" and "network" are often used interchangeably in this article. The term "and / or" in this article is only a description of the association relationship of associated objects, indicating that there can be three relationships. For example, A and / or B can mean: A exists alone, A and B exist at the same time, and B exists alone. In addition, the fragment " / " in this article generally indicates that the associated objects before and after are in an "or" relationship. In addition, "many" in this article means two or more than two.

[0014] See also Figure 1 , Figure 1 This is a flowchart of an embodiment of the text content review method of this application. Specifically, it can include the following steps: Step S11: Obtain the text to be reviewed.

[0015] In an implementation scenario, the specific source of the text to be reviewed is not limited here. Exemplarily, taking the instant messaging scenario as an example, the text to be reviewed may include but is not limited to chat text, which may be text input text or voice recognition text, which is not limited here; or, taking the self-media scenario as an example, the text to be reviewed may include but is not limited to text to be published, which may be text input text or video / audio recognition text, which is not limited here. Other scenarios can be deduced by analogy, and examples are not given one by one here. It should be noted that in application scenarios such as instant messaging and self-media, the user can be prompted that the background will conduct a compliance review of the communication content, and after obtaining authorization for this purpose, the text content will be reviewed. In addition, in actual applications, the text to be reviewed may not be limited to the above-mentioned application scenarios, such as but not limited to sources such as online novels, large model outputs, user requests, etc., which are not given examples one by one here.

[0016] In an implementation scenario, the specific content of the text to be reviewed is not limited here. Taking the instant messaging scenario as an example, the text to be reviewed may include but is not limited to: "The weather will be nice tomorrow, suitable for outdoor activities", "Shock! According to a rumor that is inconvenient to disclose, a well-known actor XXX lives a corrupt and decadent life, and often goes to XX entertainment venues. It is reported that the XX entertainment venue provides XXX, XXX, XXX and other indescribable X services", etc., or, taking the self-media scenario as an example, the text to be reviewed may include but is not limited to: "XX drugs are widely used for sedation, and its production process includes the following steps:..." It should be noted that the above examples are only a few possible examples of the text to be reviewed, and do not limit the specific content of the text to be reviewed.

[0017] Step S12: Predictions are made based on the text to be reviewed to obtain probability values ​​of sub-labels under each text topic involved in the text to be reviewed.

[0018] In the embodiment of the present disclosure, any text topic may include a first sub-tag, a second sub-tag, and a third sub-tag, and the first sub-tag indicates that the content related to the text topic is compliant, the second sub-tag indicates that the content related to the text topic is illegal, and the third sub-tag indicates that it does not involve the text topic. It should be noted that the text topic can be pre-set, such as pre-setting including but not limited to the following text topics: daily chat, sports events, social news, illegal activities, etc., and the text topic is not limited here. It should be noted that the "rules" in "compliance" and "violations" described in the embodiment of the present disclosure may include laws, regulations, rules or other normative documents, such as information release regulations of network platforms, etc., which are not limited here.

[0019] In one implementation scenario, before the text to be reviewed is reviewed, the text to be reviewed may be preprocessed. For example, meaningless punctuation marks such as “%@#$” may be removed from the text to be reviewed, redundant spaces may be eliminated, and repeated words may be removed.

[0020] In an implementation scenario, several sample texts may be acquired in advance, and the sample texts are annotated with sample tag values ​​of sub-tags under each text topic. For example, if there are N text topics, each sample text may be annotated with N*3 sub-tags. When the sample tag value of a sub-tag under a text topic is set to 1, it indicates that the sample text involves the sub-tag, and when the sample tag value of a sub-tag under a text topic is set to 0, it indicates that the sample text does not involve the sub-tag. For example, taking the four text topics of daily chatting, sports events, social news, and illegal activities as an example, the sample text "The weather is nice today, let's go to XX Park for a stroll" can be annotated with {(1,0,0), (0,0,1), (0,0,1), (0,0,1)}, where each "()" represents the sample mark value of the sub-label (the first sub-label, the second sub-label, and the third sub-label, respectively) under the four text topics of daily chatting, sports events, social news, and illegal activities. That is, the above annotation information indicates that the sample text "The weather is nice today, let's go to XX Park for a stroll" is compliant with the content related to the text topic "daily chatting", does not involve the text topic "sports events", does not involve the text topic "social news", and does not involve the text topic "illegal activities". Or, the sample text "XX drug is widely used for sedation, and its production process includes the following steps:..." can be annotated with {(0,0,1), (0,0,1), (0,0,1), (0,1,0)}, and each "()" represents the sample tag value of the sub-label under the four text themes of daily chat, sports events, social news, and illegal activities (the first sub-label, the second sub-label, and the third sub-label, respectively). That is, the above annotation information indicates that the sample text does not involve the text theme "daily chat", does not involve the text theme "sports events", does not involve the text theme "social news", and involves the text theme "illegal activities" content violation (that is, suspected of drug manufacturing, and the production process involved is obviously illegal). Other situations can be deduced by analogy, and no examples are given here one by one. On this basis, the text feature representation can be extracted for each sample text involving the same sub-tag (for example, all involving the first sub-tag "Content compliance of the text topic involved" under the text topic "Daily Chat"), and the representative feature representation of the sub-tag can be obtained accordingly, thereby obtaining the representative feature representation of each sub-tag. On this basis, the text feature representation of the text to be reviewed can be extracted, and the similarity between the text feature representation of the text to be reviewed and the representative feature representation of each sub-tag can be measured as the probability value of the text to be reviewed involving each sub-tag. For example, if the feature similarity between the text to be reviewed and the first sub-tag under the text topic "Daily Chat" is 0.9, it can be determined that the probability value of the text to be reviewed involving the first sub-tag under the text topic "Daily Chat" is 0.9.Of course, the above example is only a possible example in the actual application process, and does not limit the specific numerical values ​​of the probability values ​​of each sub-label involved in the text to be reviewed.

[0021] In another implementation scenario, different from the aforementioned implementation, as another possible example, the probability value can be predicted by a text review model combining a multi-head attention mechanism and a grouped attention mechanism. It should be noted that the multi-head attention mechanism receives three input vectors: query (Q), key (K), and value (V). These vectors are first represented by linear transformation. The multi-head attention mechanism maps the input vectors Q, K, and V to multiple (usually H, i.e., the number of heads) lower-dimensional representation spaces through different linear transformations. Each head contains a set of independent linear transformation parameters. The attention score in each head is calculated by the softmax function, and then the value vector is weighted and summed according to the weight. The outputs of each head are spliced ​​together to form an output vector of multi-head attention, which is then transformed through a linear layer. The grouped attention mechanism also receives three input vectors: query (Q), key (K), and value (V). In contrast, the grouped attention mechanism divides the query Q into G groups, each of which shares a common key K and value V. The query Q in each group is obtained through their own linear transformation. The other parts are basically the same as the multi-head attention mechanism. For the specific calculation process of the multi-head attention mechanism and the group attention mechanism, please refer to the technical details of the two respectively, which will not be repeated here. In the above method, the probability value is predicted by the text audit model that combines the multi-head attention mechanism and the group attention mechanism. Since each head in the multi-head attention mechanism processes information independently, it can fully capture the multi-dimensional features in the input sequence, and the group attention mechanism can significantly reduce the KV cache, reduce the memory usage, improve the reasoning efficiency, and greatly enhance the model's performance in processing long texts and priority resource scenarios. Therefore, by combining the multi-head attention mechanism and the group attention mechanism, the text audit model can have both the powerful semantic representation ability of the multi-head attention mechanism and the extremely high reasoning efficiency of the group attention mechanism.

[0022] In a specific implementation scenario, the text review model may include but is not limited to Transformer, etc., and the network structure of the text review model is not limited here. Taking the case where the text review model includes Transformer as an example, the text review model may include L layers of Transformer-based encoders, and among the L layers of Transformer-based encoders, some layers may include a multi-head attention mechanism, and other layers may include a group attention mechanism. Of course, when the text review model adopts other network structures, it can be deduced by analogy, and examples are not given one by one here.

[0023] In a specific implementation scenario, when combining the multi-head attention mechanism and the group attention mechanism, as a possible example, the multi-head attention mechanism and the group attention mechanism can be executed several times based on the text feature representation of the text to be reviewed to obtain the first text feature of the text to be reviewed, and then classification prediction is performed based on the first text feature of the text to be reviewed to obtain the probability values ​​of the sub-labels under each text topic involved in the text to be reviewed. Please refer to Figure 2 , Figure 2 Schematic diagram of the framework of an embodiment of a text audit model. Figure 2 As shown, still taking the text review model including L layers of Transformer-based encoder as an example, the first layer can include a multi-head attention mechanism, the second layer can include a group attention mechanism, the third layer can include a multi-head attention mechanism, the fourth layer can include a group attention mechanism, and so on, and no examples are given here one by one. Among them, T1, T2, ..., TM represent each character in the text to be reviewed, and E1, E2, ..., EM represent the sub-feature representations corresponding to each character in the text feature representation. In addition, the text review model can also include a multi-layer perceptron for realizing classification prediction, that is, the first text feature can be input into the multi-layer perceptron to obtain the probability value of the text to be reviewed involving N*3 sub-labels. Among them, the encoder including the multi-head attention mechanism can include: multi-head attention mechanism, addition and normalization, feedforward layer, addition and normalization and other network layers, and the encoder including the group attention mechanism can include: group attention mechanism, addition and normalization, feedforward layer, addition and normalization and other network layers. It should be noted that the specific principles of the encoder containing the multi-head attention mechanism and the encoder containing the group attention mechanism can refer to the relevant technical details of the Transformer, which will not be repeated here. In particular, taking L set to 24 as an example, after testing, compared with the text review model set to 24 layers of encoders that only contain the multi-head attention mechanism, the model reasoning speed can be improved by 30% when the multi-head attention mechanism and the group attention mechanism are set to be executed alternately. In the case of multi-way deployment, the improvement is more obvious. In the above method, based on the text feature representation of the text to be reviewed, the multi-head attention mechanism and the group attention mechanism are executed alternately several times to obtain the first text feature of the text to be reviewed, and the classification prediction is performed based on the first text feature of the text to be reviewed to obtain the probability values ​​of the sub-labels under each text topic involved in the text to be reviewed. By alternately executing the multi-head attention mechanism and the group attention mechanism, the reasoning efficiency of the text review model can be significantly improved.

[0024] In a specific implementation scenario, when combining the multi-head attention mechanism and the group attention mechanism, as another possible example, the multi-head attention mechanism can be executed several times and then the group attention mechanism can be executed several times based on the text feature representation of the text to be reviewed, to obtain the second text feature of the text to be reviewed, and then classification prediction is performed based on the second text feature of the text to be reviewed to obtain the probability values ​​of the sub-labels under each text topic involved in the text to be reviewed. Still taking the text review model including L layers (e.g., 24 layers) of Transformer-based encoders as an example, the first L / 2 layers (e.g., the first 12 layers) can be encoders containing a multi-head attention mechanism, and the last L / 2 layers (e.g., the last 12 layers) can be encoders containing a group attention mechanism. When the text review model adopts other network structures, it can be deduced by analogy, and examples are not given one by one here. In the above method, based on the text feature representation of the text to be reviewed, the multi-head attention mechanism is first executed several times and then the group attention mechanism is executed several times to obtain the second text feature of the text to be reviewed, and then classification prediction is performed based on the second text feature of the text to be reviewed to obtain the probability values ​​of the sub-labels under each text topic involved in the text to be reviewed. Therefore, by successively executing the multi-head attention mechanism and the group attention mechanism, compared with the single execution of the multi-head attention mechanism, it can help to improve the reasoning efficiency of the text review model.

[0025] In a specific implementation scenario, when combining the multi-head attention mechanism and the group attention mechanism, as another possible example, the group attention mechanism can be executed several times and then the multi-head attention mechanism can be executed several times based on the text feature representation of the text to be reviewed, to obtain the third text feature of the text to be reviewed, and then the classification prediction is performed based on the third text feature of the text to be reviewed to obtain the probability values ​​of the sub-labels under each text topic involved in the text to be reviewed. Still taking the text review model including L layers (e.g., 24 layers) of Transformer-based encoders as an example, the first L / 2 layers (e.g., the first 12 layers) can be encoders containing group attention mechanisms, and the last L / 2 layers (e.g., the last 12 layers) can be encoders containing multi-head attention mechanisms. When the text review model adopts other network structures, it can be deduced by analogy, and examples are not given one by one here. In the above method, based on the text feature representation of the text to be reviewed, the group attention mechanism is first executed several times and then the multi-head attention mechanism is executed several times to obtain the third text feature of the text to be reviewed, and then classification prediction is performed based on the third text feature of the text to be reviewed to obtain the probability values ​​of the sub-labels under each text topic involved in the text to be reviewed. Therefore, by successively executing the group attention mechanism and the multi-head attention mechanism, compared with the single execution of the multi-head attention mechanism, it can help to improve the reasoning efficiency of the text review model.

[0026] In one implementation scenario, as described above, the probability value can be predicted by a text review model, and the text review model can be used as follows: Figure 2The network structure shown, or other network structures, are not limited here. Before using the text audit model to classify and predict the text to be audited, the text audit model can be trained first. Specifically, the sample text can be predicted based on the text audit model to obtain the predicted probability values ​​of the sub-labels under each text topic, and the sample text is annotated with the sample tag values ​​of the sub-labels under each text topic. The specific process of obtaining the predicted probability value can refer to the aforementioned process of obtaining the probability value, which will not be repeated here. In addition, the specific meaning of the sample tag value can refer to the aforementioned related description, which will not be repeated here. On this basis, for each text topic, the first loss of the sample text involving the text topic can be measured based on the predicted probability value and the specimen tag value of the same sub-label under the text topic, and then the weight factor of each text topic and the first loss of the sample text involving each text topic are weighted to obtain the second loss, so that the network parameters of the text audit model can be adjusted based on the second loss. The above method, on the one hand, sets different weight preferences for text topics during training, which is conducive to achieving customized emphasis on different text topics. On the other hand, it simultaneously trains topic discrimination and compliance discrimination during training. Compared with training topic discrimination and compliance discrimination separately, it can integrate two aspects of data information, which is conducive to increasing the robustness of the model and speeding up the reasoning speed.

[0027] In a specific implementation scenario, for each text topic, a loss function such as cross entropy can be used to measure the loss of the predicted probability value and sample label value of the same sub-label under the text topic of the sample text, and the sub-loss of the corresponding sub-label under the text topic can be obtained. For example, for the text topic "daily chat", a loss function such as cross entropy can be used to measure the loss of the predicted probability value and sample label value of the first sub-label under the text topic "daily chat" respectively involved in the sample text, and the sub-loss of the first sub-label under the text topic "daily chat" can be obtained. By analogy, the sub-loss of the second sub-label under the text topic "daily chat" and the sub-loss of the third sub-label under the text topic "daily chat" can be obtained. Other text topics can be deduced in the same way, and examples are not given one by one here. On this basis, the sub-losses of each sub-label under the same text topic can be fused by addition, averaging, etc. to obtain the first loss of the sample text involving the text topic.

[0028] In a specific implementation scenario, different from the above implementation, for each text topic, the sub-loss of the corresponding sub-label under the text topic can be obtained based on the predicted probability value and the specimen mark value of the same sub-label under the text topic respectively involved in the sample text, and then the sub-weight and sub-loss of each sub-label under the text topic are weighted to obtain the first loss of the sample text involving the text topic, and the sub-weight of each sub-label under the text topic can be obtained based on the inverse frequency statistics of the sample texts involving each sub-label under the text topic. Exemplarily, the total number of sample texts involving various sub-labels can be counted according to the sample mark value of each sample text about the sub-label, and then the sub-weights of various sub-labels can be obtained based on the total number of various sub-labels. For example, the total number of various sub-labels can be reciprocated, and then the total number of each sub-label under the same text topic can be normalized after the reciprocal is taken to obtain the sub-weight of each sub-label under the same text topic. Of course, the above example is only a possible example of obtaining sub-weights through inverse frequency statistics, and it is not limited to using other methods to measure and obtain sub-weights. For ease of description, the first loss under any text topic can be expressed as: …… (1) In the above formula (1), represents the first loss under any text topic, Respectively represent the sub-weights of each sub-tag (i.e., the first sub-tag, the second sub-tag, and the third sub-tag) under the same text topic. For example, 1. Respectively represent the sample tag values ​​of each sub-label under the same text topic, Respectively represent the predicted probability values ​​of each sub-label under the same text topic. In other words, That is, it represents the sub-loss of the first sub-label, That is, it represents the sub-loss of the second sub-label. That is, it represents the sub-loss of the third sub-label. Of course, formula (1) is only a possible way to calculate the first loss, and it does not limit the first loss to be measured by other loss functions. In the above method, based on the predicted probability values ​​and specimen marking values ​​of the same sub-labels under the text topics respectively involved in the sample text, the sub-loss of the corresponding sub-label under the text topic involved in the sample text is obtained, and then the sub-weights and sub-losses of each sub-label under the text topic are weighted to obtain the first loss of the sample text involving the text topic, and the sub-weights of each sub-label under the text topic are obtained based on the inverse frequency statistics of the sample text involving each sub-label under the text topic. Therefore, it is possible to adaptively focus on learning each sub-label according to the distribution of different sub-labels involved in the sample text, which helps to improve the training performance of the text review model.

[0029] In a specific implementation scenario, the weight factor of the text topic can be positively correlated with the discrimination accuracy requirement for the text topic. That is to say, the higher the discrimination accuracy requirement for the text topic, the larger the weight factor of the text topic can be set. Conversely, the lower the discrimination accuracy requirement for the text topic, the smaller the weight factor of the text topic can be set. For example, in the online novel review scenario, the discrimination accuracy requirement for sensitive topics such as the text topic "illegal activities" is high, then the text topic "illegal activities" and other sensitive topics can be set to have a larger weight factor accordingly. Other application scenarios can be deduced by analogy, and examples are not given one by one here. In addition, the specific value of the weight factor is not limited. On this basis, the second loss can be obtained : …… (2) In the above formula (2), Represent the weight factors of each text topic, In the above method, the weight factor of the text topic is positively correlated with the discrimination accuracy requirement of the text topic, and different weight preferences of the text topics can be set during training, which is conducive to realizing customized emphasis on different text topics.

[0030] In an implementation scenario, please refer to Figure 3 , Figure 3 This is a process diagram of an embodiment of the text content review method of this application. Figure 3 As shown, as a possible implementation method, different from directly predicting the probability values ​​of the sub-labels under each text topic after obtaining the text to be reviewed, it is also possible to perform a preliminary inspection based on the text to be reviewed after obtaining the text to be reviewed, and obtain a first detection result that characterizes whether the text to be reviewed has illegal content. Then, in response to the first detection result characterizing that the text to be reviewed has illegal content, a prediction based on the text to be reviewed can be performed to obtain the steps of the probability values ​​of the sub-labels under each text topic that the text to be reviewed respectively involves. Exemplarily, based on methods such as adding sensitive words and model thresholds, a certain diversion ratio (such as 80:20) can be formed through preliminary inspection to reduce the pressure on subsequent classification while greatly reducing the omission rate. In the above method, the text to be reviewed is first preliminarily inspected, and then the probability values ​​of each sub-label involved are predicted when the preliminary inspection is found to involve violations. It can be preliminarily screened first and then finely screened, that is, it follows the principle of coarse screening first and fine screening later, which helps to reduce the workload of fine screening and improve the efficiency of text content review.

[0031] In a specific implementation scenario, as a possible example, the text to be reviewed can be preliminarily tested by using detection methods such as sensitive word matching and simple audit models to obtain a first test result. Exemplarily, the sensitive word library may include but is not limited to: X service, sedation and other sensitive words, and the specific content of the sensitive word library is not limited here. Specifically, the sensitive word library can be composed of a large number of common ontology words and tool conversion variant words. In addition, the simple audit model can adopt FastText, etc., and the network structure of the simple audit model is not limited here. The training corpus of the simple audit model can be composed of sensitive word collection and subsequent manual annotation data. The training data can reach tens of millions. The model can eventually output the probability values ​​corresponding to the positive and negative examples (positive examples indicate no illegal content, and negative examples indicate illegal content). In the end, the positive and negative can be decided again through the threshold. Alternatively, in order to integrate different detection methods, the text to be reviewed can also be preliminarily inspected based on multiple detection methods to obtain the second detection results of various detection methods, and the second detection result of any detection method indicates whether the text to be reviewed has illegal content after detection by the detection method, so that in response to at least one second detection result indicating the existence of illegal content, it can be determined that the first detection result indicates that the text to be reviewed has illegal content, and in response to each second detection result indicating the absence of illegal content, it can be determined that the first detection result indicates that the text to be reviewed has no illegal content. In other words, as long as there is a detection method that preliminarily detects that the text to be reviewed has illegal content, the text to be reviewed needs to be sent to the next step of fine detection, and only when all detection methods preliminarily detect that the text to be reviewed has no illegal content, the text to be reviewed is judged as a positive example (i.e., no illegal content). The above method, based on multiple detection methods, preliminarily inspects the text to be reviewed, obtains the second detection results of various detection methods, and determines the first detection result of the text to be reviewed by combining the second detection results of multiple detection methods, which can complement each other and help improve the accuracy of the preliminary detection.

[0032] In a specific implementation scenario, in response to the first detection result indicating that the text to be reviewed does not contain any illegal content, the step of performing prediction based on the text to be reviewed to obtain the probability values ​​of the sub-labels under each text topic involved in the text to be reviewed can be skipped, such as Figure 3 As shown, the text to be reviewed can be directly judged as a positive example. Of course, the text to be reviewed can also be sent to the next step of fine detection, that is, the step of still performing prediction based on the text to be reviewed to obtain the probability values ​​of the sub-labels under each text topic involved in the text to be reviewed. It can be specifically set according to application needs and is not limited here. In the above method, when the first detection result indicates that the text to be reviewed does not contain any illegal content, the subsequent steps can be directly skipped, which can improve the efficiency of text review.

[0033] In an implementation scenario, after making a prediction based on the text to be reviewed and obtaining the probability values ​​of the sub-tags under each text theme respectively involving the text to be reviewed, it is also possible to determine whether the text to be reviewed involves the target theme based on the probability values ​​of the sub-tags under each text theme respectively involving the text to be reviewed. It should be noted that the target theme can be set according to the application scenario. For example, if a publishing platform mainly focuses on the text theme "illegal activities", the text theme "illegal activities" can be used as the target theme. Other situations can be deduced by analogy, and examples are not given one by one here. On this basis, in response to the text to be reviewed involving the target theme, the text to be reviewed can be directly intercepted, or it can be determined whether to intercept the text to be reviewed based on the probability value of the text to be reviewed involving the first sub-tag or the second sub-tag under the target theme. The above method determines whether to intercept the text to be reviewed through the probability values ​​of the theme discrimination and the positive and negative discrimination, which can improve the adaptability to different usage scenarios of text review.

[0034] In a specific implementation scenario, the probability value of the third sub-tag under the target topic that the text to be reviewed involves can be obtained. On this basis, the probability value of the third sub-tag under the target topic that the text to be reviewed involves can be compared with the probability values ​​of the first sub-tag and the second sub-tag under the target topic that the text to be reviewed involves. If the former is greater than the latter, it can be considered that the text to be reviewed does not involve the target topic. On the contrary, if the former is less than the latter, it can be considered that the text to be reviewed involves the target topic. For example, taking the text to be reviewed "XX drug is widely used in sedation, and its production process includes the following steps:..." as an example, the probability values ​​of the sub-labels under each text topic can be {(0.1, 0.1, 0.8), (0.2, 0.1, 0.7), (0.1, 0.1, 0.8), (0.1, 0.8, 0.1)}. When the target topic is set to the text topic "illegal activities", the probability value of 0.1 of the third sub-label "not involving text topic" under the target topic "illegal activities" can be obtained, as well as the probability value of 0.1 of the first sub-label "content involving text topic is compliant" under the target topic "illegal activities" and the probability value of 0.8 of the second sub-label "content involving text topic is illegal" under the target topic "illegal activities". Obviously, the probability value of the third sub-label is less than the probability value of the second sub-label, so it can be determined that the text to be reviewed "XX drug is widely used in sedation, and its production process includes the following steps:..." involves the target topic "illegal activities". Of course, the above example is only a possible example in the actual application process, and does not limit other possible situations. I will not give examples one by one here. In addition, other methods can be used to determine whether the text to be reviewed involves the target topic, such as setting a probability threshold of the sub-tag, and determining whether the text to be reviewed involves the target topic by comparing the probability value of the sub-tag with the probability threshold. I will not repeat them here.

[0035] In a specific implementation scenario, when text interception is performed only through topic discrimination, as long as it is determined that the text to be reviewed involves the target topic, the text to be reviewed can be directly intercepted. Exemplarily, still taking the aforementioned text to be reviewed "XX drug is widely used in sedation, and its production process includes the following steps:..." as an example, since it involves the target topic "illegal activities", the text to be reviewed can be directly intercepted. Alternatively, when text interception is performed in combination with topic discrimination and positive and negative discrimination, on the premise that it is determined that the text to be reviewed involves the target topic, it can be further determined whether to intercept the text to be reviewed based on the probability value of either the first sub-tag or the second sub-tag under the target topic that the text to be reviewed involves. Exemplarily, if the probability value of the first sub-tag under the target topic being involved in the text to be reviewed is greater than the probability value of the second sub-tag under the target topic being involved, then it can be considered that the content of the target topic being involved in the text to be reviewed is compliant, and it can be determined not to intercept the text to be reviewed; conversely, if the probability value of the first sub-tag under the target topic being involved in the text to be reviewed is less than the probability value of the second sub-tag under the target topic being involved, then it can be considered that the content of the target topic being involved in the text to be reviewed is in violation of regulations, and it can be determined to intercept the text to be reviewed. For example, the aforementioned text to be reviewed, "XX drugs are widely used in sedation, and its production process includes the following steps:...", because its probability value of 0.8 for the second sub-tag "involving content violations of the text topic" under the target topic "illegal activities" is greater than its probability value of 0.1 for the first sub-tag "involving content compliance of the text topic" under the target topic "illegal activities", it can be considered that the content of the target topic being involved in the text to be reviewed is in violation of regulations, and it can be determined to intercept the text to be reviewed. In addition, if the probability value of the first sub-tag under the target topic involved in the text to be reviewed is equal to the probability value of the second sub-tag under the target topic, it can be considered that the text to be reviewed involves the target topic but it is impossible to determine whether the content is compliant or illegal. In order to ensure the safety and controllability of information release, it can be determined to intercept the text to be reviewed, or the text to be reviewed can be manually reviewed. Of course, the above example is only a possible example of determining whether to intercept the text to be reviewed on the premise that the text to be reviewed involves the target topic. The probability threshold of the first sub-tag or the probability threshold of the second sub-tag can also be set, and by comparing the size relationship between the probability value of the sub-tag and the probability threshold, it is determined whether to intercept the text to be reviewed. I will not go into details here.The above method, after making a prediction based on the text to be reviewed and obtaining the probability values ​​of the sub-labels under each text topic that the text to be reviewed involves, determines whether the text to be reviewed involves the target topic based on the probability values ​​of the sub-labels under each text topic that the text to be reviewed involves, and then directly intercepts the text to be reviewed in response to the text to be reviewed involving the target topic, or determines whether to intercept the text to be reviewed based on the probability value of the first sub-label or the second sub-label under the target topic that the text to be reviewed involves. This method can be adaptively applied to downstream tasks of text interception based on topic discrimination and positive and negative discrimination, which helps to improve the scope of application of text interception and provide more possibilities for downstream autonomous control.

[0036] The above scheme obtains the text to be reviewed, and then makes predictions based on the text to be reviewed to obtain the probability values ​​of the sub-labels under each text topic involved in the text to be reviewed, and any text topic includes the first sub-label, the second sub-label and the third sub-label. The first sub-label represents the compliance of the content involved in the text topic, the second sub-label represents the violation of the content involved in the text topic, and the third sub-label represents the non-involvement of the text topic. Therefore, on the one hand, it is different from the keyword matching technology. By predicting the probability values ​​of the sub-labels under each text topic, it can alleviate the adverse effects of circumvention methods such as homophone phrase replacement and special symbol word segmentation on text content review as much as possible, which helps to improve the accuracy of text content. On the other hand, because the probability values ​​of the sub-labels under each text topic are predicted, it can simultaneously perform review predictions from the two perspectives of text topic and positive and negative judgment (i.e., content compliance or content violation), which helps to improve the comprehensiveness of text content review. Therefore, it can improve the accuracy and comprehensiveness of text content review.

[0037] See also Figure 4 , Figure 4 It is a schematic diagram of the framework of an embodiment of the text content audit device of the present application. The text content audit device 40 includes: an acquisition module 41 and a prediction module 42, wherein the acquisition module 41 is used to acquire the text to be audited; the prediction module 42 is used to make predictions based on the text to be audited, and obtain the probability values ​​of the sub-tags under each text topic respectively involved in the text to be audited; wherein any text topic includes a first sub-tag, a second sub-tag and a third sub-tag, and the first sub-tag represents that the content related to the text topic is compliant, the second sub-tag represents that the content related to the text topic is illegal, and the third sub-tag represents that the text topic is not involved.

[0038] In the above scheme, the text content audit device 40 obtains the text to be audited, and then predicts based on the text to be audited, and obtains the probability values ​​of the sub-tags under each text topic respectively involved in the text to be audited, and any text topic includes the first sub-tag, the second sub-tag and the third sub-tag, the first sub-tag represents the compliance of the content related to the text topic, the second sub-tag represents the violation of the content related to the text topic, and the third sub-tag represents that the text topic is not involved. Therefore, on the one hand, it is different from the keyword matching technology. By predicting the probability values ​​of the sub-tags under each text topic, it can alleviate the adverse effects of circumvention methods such as homophone phrase replacement and special symbol word segmentation on text content audit as much as possible, which helps to improve the accuracy of text content. On the other hand, because the probability values ​​of the sub-tags under each text topic are predicted, it can simultaneously perform audit prediction from the two perspectives of text topic and positive and negative judgment (i.e., content compliance or content violation), which helps to improve the comprehensiveness of text content audit. Therefore, the accuracy and comprehensiveness of text content audit can be improved.

[0039] In some disclosed embodiments, the text content review device 40 also includes a preliminary inspection module, which is used to perform a preliminary inspection based on the text to be reviewed to obtain a first detection result characterizing whether the text to be reviewed contains any illegal content; the prediction module 42 is specifically used to respond to the first detection result characterizing that the text to be reviewed contains any illegal content, perform a prediction based on the text to be reviewed, and obtain the probability values ​​of the sub-labels under each text topic involved in the text to be reviewed.

[0040] In some disclosed embodiments, the initial inspection module includes a preliminary detection submodule, which is used to perform initial inspections on the text to be reviewed based on multiple detection methods, and obtain second detection results of various detection methods; wherein the second detection result of any detection method indicates whether the text to be reviewed contains any illegal content after detection by the detection method; the initial inspection module includes a first response submodule, which is used to determine, in response to at least one second detection result indicating the presence of illegal content, that the first detection result indicates that the text to be reviewed contains illegal content; the initial inspection module includes a second detection submodule, which is used to determine, in response to each second detection result indicating the absence of illegal content, that the first detection result indicates that the text to be reviewed contains no illegal content.

[0041] In some disclosed embodiments, the text content review device 40 also includes a skipping module for skipping the step of making predictions based on the text to be reviewed and obtaining probability values ​​of sub-labels under each text topic involved in the text to be reviewed in response to a first detection result indicating that the text to be reviewed does not contain any illegal content.

[0042] In some disclosed embodiments, the probability values ​​are predicted by a text review model that combines a multi-head attention mechanism and a group attention mechanism.

[0043] In some disclosed embodiments, the prediction module 42 includes a first execution submodule, which is used to alternately execute the multi-head attention mechanism and the group attention mechanism several times based on the text feature representation of the text to be reviewed, so as to obtain the first text feature of the text to be reviewed; the prediction module 42 includes a first classification submodule, which is used to perform classification prediction based on the first text feature of the text to be reviewed, so as to obtain the probability values ​​of the sub-labels under each text topic involved in the text to be reviewed.

[0044] In some disclosed embodiments, the prediction module 42 includes a second execution submodule, which is used to first execute a multi-head attention mechanism several times and then execute a group attention mechanism several times based on the text feature representation of the text to be reviewed, so as to obtain a second text feature of the text to be reviewed; the prediction module 42 includes a second classification submodule, which is used to perform classification prediction based on the second text feature of the text to be reviewed, so as to obtain the probability values ​​of the sub-labels under each text topic involved in the text to be reviewed.

[0045] In some disclosed embodiments, the prediction module 42 includes a third execution submodule, which is used to first execute a group attention mechanism several times and then execute a multi-head attention mechanism several times based on the text feature representation of the text to be reviewed, so as to obtain a third text feature of the text to be reviewed; the prediction module 42 includes a third classification submodule, which is used to perform classification prediction based on the third text feature of the text to be reviewed, so as to obtain the probability values ​​of the sub-labels under each text topic involved in the text to be reviewed.

[0046] In some disclosed embodiments, the probability value is predicted by a text audit model, and the text content audit device 40 includes a sample prediction module for predicting the sample text based on the text audit model to obtain the predicted probability values ​​of the sub-labels under each text topic respectively involved in the sample text; wherein the sample text is annotated with sample tag values ​​of the sub-labels under each text topic respectively; the text content audit device 40 includes a first loss module for measuring, for each text topic, the first loss of the sample text involving the text topic based on the predicted probability values ​​and sample tag values ​​of the same sub-labels under the text topic respectively involved in the sample text; the text content audit device 40 includes a second loss module for weighting based on the weight factors of each text topic and the first losses of the sample text involving each text topic respectively to obtain a second loss; the text content audit device 40 includes a parameter adjustment module for adjusting the network parameters of the text audit model based on the second loss.

[0047] In some disclosed embodiments, the first loss module includes a sub-loss measurement sub-module, which is used to obtain the sub-loss of the sample text involving the corresponding sub-label under the text topic based on the predicted probability values ​​and sample label values ​​of the same sub-label under the text topic respectively involved in the sample text; the first loss module includes a sub-loss weighting sub-module, which is used to weight the sub-weights and sub-losses of each sub-label under the text topic to obtain the first loss of the sample text involving the text topic; wherein the sub-weights of each sub-label under the text topic are obtained based on inverse frequency statistics of the sample text involving each sub-label under the text topic.

[0048] In some disclosed embodiments, the weight factor of a text topic is positively correlated with the accuracy requirement for distinguishing the text topic.

[0049] In some disclosed embodiments, the text content review device 40 includes a topic determination module for determining whether the text to be reviewed involves a target topic based on the probability values ​​of the sub-tags under each text topic that the text to be reviewed involves respectively; the text content review device 40 includes a text interception module for directly intercepting the text to be reviewed in response to the text to be reviewed involving the target topic, or determining whether to intercept the text to be reviewed based on the probability value of the text to be reviewed involving either the first sub-tag or the second sub-tag under the target topic.

[0050] See also Figure 5 , Figure 5 It is a schematic diagram of the framework of an embodiment of an electronic device of the present application. The electronic device 50 includes at least a memory 51 and a processor 52 coupled to each other, the memory 51 at least stores program instructions, and the processor 52 is used to execute the program instructions to implement the steps in any of the above-mentioned text content review method embodiments. For details, please refer to the aforementioned disclosed embodiments, which will not be repeated here. As a possible example, the electronic device 50 may include but is not limited to a server, etc., and the specific type of the electronic device 50 is not limited here.

[0051] Specifically, the processor 52 is used to control itself and the memory 51 to implement the steps in any of the above-mentioned text content audit method embodiments. The processor 52 can also be called a CPU (Central Processing Unit). The processor 52 may be an integrated circuit chip with signal processing capabilities. The processor 52 can also be a general-purpose processor, a digital signal processor (Digital Signal Processor, DSP), an application-specific integrated circuit (Application Specific Integrated Circuit, ASIC), a field-programmable gate array (Field-Programmable Gate Array, FPGA) or other programmable logic devices, discrete gates or transistor logic devices, discrete hardware components. The general-purpose processor can be a microprocessor or the processor can also be any conventional processor, etc. In addition, the processor 52 can be implemented by an integrated circuit chip.

[0052] In the above scheme, the electronic device 50 obtains the text to be reviewed, and then makes predictions based on the text to be reviewed to obtain the probability values ​​of the sub-tags under each text topic respectively involved in the text to be reviewed, and any text topic includes the first sub-tag, the second sub-tag and the third sub-tag, the first sub-tag represents the compliance of the content related to the text topic, the second sub-tag represents the violation of the content related to the text topic, and the third sub-tag represents the non-involvement of the text topic. Therefore, on the one hand, it is different from the keyword matching technology. By predicting the probability values ​​of the sub-tags under each text topic, it can alleviate the adverse effects of circumvention methods such as homophone phrase replacement and special symbol word segmentation on text content review as much as possible, which helps to improve the accuracy of text content. On the other hand, because the probability values ​​of the sub-tags under each text topic are predicted, it can simultaneously perform review predictions from the two perspectives of text topic and positive and negative judgment (i.e., content compliance or content violation), which helps to improve the comprehensiveness of text content review. Therefore, the accuracy and comprehensiveness of text content review can be improved.

[0053] See also Figure 6 , Figure 6 The figure is a schematic diagram of a framework of an embodiment of a computer-readable storage medium 60 of the present application. The computer-readable storage medium 60 stores program instructions 61 that can be executed by a processor, and the program instructions 61 are used to implement the steps in any of the above-mentioned text content review method embodiments.

[0054] In the above scheme, the computer-readable storage medium 60 obtains the text to be reviewed, and then predicts based on the text to be reviewed to obtain the probability values ​​of the sub-tags under each text topic respectively involved in the text to be reviewed, and any text topic includes the first sub-tag, the second sub-tag and the third sub-tag, the first sub-tag represents the compliance of the content related to the text topic, the second sub-tag represents the violation of the content related to the text topic, and the third sub-tag represents that the text topic is not involved. Therefore, on the one hand, it is different from the keyword matching technology. By predicting the probability values ​​of the sub-tags under each text topic, it can alleviate the adverse effects of circumvention methods such as homophone phrase replacement and special symbol word segmentation on text content review as much as possible, which helps to improve the accuracy of text content. On the other hand, because the probability values ​​of the sub-tags under each text topic are predicted, it can simultaneously perform review prediction from the two perspectives of text topic and positive and negative judgment (i.e., content compliance or content violation), which helps to improve the comprehensiveness of text content review. Therefore, the accuracy and comprehensiveness of text content review can be improved.

[0055] In some embodiments, the functions or modules included in the device provided by the embodiments of the present disclosure can be used to execute the method described in the above method embodiments. The specific implementation can refer to the description of the above method embodiments, and for the sake of brevity, it will not be repeated here.

[0056] The above description of various embodiments tends to emphasize the differences between the various embodiments. The same or similar aspects can be referenced to each other, and for the sake of brevity, they will not be repeated herein.

[0057] In the several embodiments provided in the present application, it should be understood that the disclosed methods and devices can be implemented in other ways. For example, the device implementation described above is only schematic. For example, the division of modules or units is only a logical function division. There may be other division methods in actual implementation, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, and the indirect coupling or communication connection of devices or units can be electrical, mechanical or other forms.

[0058] The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed on multiple network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the present embodiment.

[0059] In addition, each functional unit in each embodiment of the present application may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit. The above-mentioned integrated unit may be implemented in the form of hardware or in the form of software functional units.

[0060] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present application is essentially or the part that contributes to the prior art or all or part of the technical solution can be embodied in the form of a software product. The computer software product is stored in a storage medium, including several instructions to enable a computer device (which can be a personal computer, server, or network device, etc.) or a processor to execute all or part of the steps of each implementation method of the present application. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), disk or optical disk and other media that can store program code.

[0061] If the technical solution of this application involves personal information, the product using the technical solution of this application has clearly informed the personal information processing rules and obtained the individual's voluntary consent before processing the personal information. If the technical solution of this application involves sensitive personal information, the product using the technical solution of this application has obtained the individual's separate consent before processing the sensitive personal information, and at the same time meets the "explicit consent" requirement. For example, on personal information collection devices such as cameras, clear and prominent signs are set to inform that the personal information collection scope has been entered and personal information will be collected. If the individual voluntarily enters the collection scope, it is deemed that he or she agrees to the collection of his or her personal information; or on the device that processes personal information, the personal information processing rules are notified by obvious signs / information, and the individual's authorization is obtained through pop-up information or by asking the individual to upload his or her personal information; among them, the personal information processing rules may include information such as the personal information processor, the purpose of personal information processing, the processing method, and the type of personal information processed.

Claims

1. A text content review method, characterized in that: include: Get the text to be reviewed; Predictions are made based on the text to be reviewed to obtain probability values ​​of sub-labels under each text topic that the text to be reviewed involves respectively; wherein, any of the text topics includes a first sub-label, a second sub-label and a third sub-label, and the first sub-label represents that the content involving the text topic is compliant, the second sub-label represents that the content involving the text topic is not in compliance, and the third sub-label represents that the text topic is not involved.

2. The method according to claim 1, characterized in that After obtaining the text to be reviewed and before making predictions based on the text to be reviewed to obtain probability values ​​of sub-labels under each text topic respectively involved in the text to be reviewed, the method further includes: Performing a preliminary inspection based on the text to be reviewed to obtain a first inspection result indicating whether the text to be reviewed contains any illegal content; The prediction based on the text to be reviewed is performed to obtain the probability values ​​of the sub-labels under each text topic involved in the text to be reviewed, including: In response to the first detection result indicating that the text to be reviewed contains illegal content, the step of making predictions based on the text to be reviewed to obtain probability values ​​of sub-labels under each text topic respectively involved in the text to be reviewed is performed.

3. The method according to claim 2, characterized in that The initial inspection based on the text to be reviewed to obtain a first detection result indicating whether the text to be reviewed contains any illegal content includes: Based on multiple detection methods, the text to be reviewed is initially checked respectively to obtain second detection results of various detection methods; wherein the second detection result of any detection method indicates whether the text to be reviewed contains any illegal content after being detected by the detection method; In response to at least one of the second detection results indicating the presence of illegal content, determining that the first detection result indicates that the text to be reviewed contains illegal content; In response to each of the second detection results indicating that no illegal content exists, it is determined that the first detection result indicates that no illegal content exists in the text to be reviewed.

4. The method according to claim 2, characterized in that: The method further comprises: In response to the first detection result indicating that the text to be reviewed does not contain any illegal content, the step of making predictions based on the text to be reviewed to obtain probability values ​​of sub-labels under each text topic respectively involved in the text to be reviewed is skipped.

5. The method according to claim 1, characterized in that The probability value is predicted by a text review model that combines multi-head attention mechanism and group attention mechanism.

6. The method according to claim 1 or 5, characterized in that: The prediction based on the text to be reviewed is performed to obtain the probability values ​​of the sub-labels under each text topic involved in the text to be reviewed, including: Based on the text feature representation of the to-be-reviewed text, a multi-head attention mechanism and a group attention mechanism are alternately executed several times to obtain a first text feature of the to-be-reviewed text; Classification prediction is performed based on the first text feature of the text to be reviewed to obtain probability values ​​of the sub-labels under each of the text topics that the text to be reviewed involves.

7. The method according to claim 1 or 5, characterized in that: The prediction based on the text to be reviewed is performed to obtain the probability values ​​of the sub-labels under each text topic involved in the text to be reviewed, including: Based on the text feature representation of the text to be reviewed, a multi-head attention mechanism is first executed several times and then a group attention mechanism is executed several times to obtain a second text feature of the text to be reviewed; Classification prediction is performed based on the second text feature of the text to be reviewed to obtain probability values ​​of the sub-labels under each of the text topics that the text to be reviewed involves.

8. The method according to claim 1 or 5, characterized in that: The prediction based on the text to be reviewed is performed to obtain the probability values ​​of the sub-labels under each text topic involved in the text to be reviewed, including: Based on the text feature representation of the text to be reviewed, a group attention mechanism is first executed several times and then a multi-head attention mechanism is executed several times to obtain a third text feature of the text to be reviewed; Classification prediction is performed based on the third text feature of the text to be reviewed to obtain probability values ​​of the sub-labels under each of the text topics that the text to be reviewed involves.

9. The method according to claim 1, characterized in that: The probability value is predicted by a text audit model, and the training steps of the text audit model include: Predicting the sample text based on the text audit model to obtain predicted probability values ​​of the sub-labels under each of the text topics respectively related to the sample text; wherein the sample text is annotated with sample tag values ​​of the sub-labels under each of the text topics respectively; For each of the text topics, based on the predicted probability values ​​and sample label values ​​of the sample texts respectively involving the same sub-labels under the text topics, a first loss of the sample texts involving the text topics is measured; Weighting is performed based on the weight factors of the respective text topics and the first losses of the sample texts respectively involving the respective text topics to obtain a second loss; Based on the second loss, network parameters of the text audit model are adjusted.

10. The method according to claim 9, characterized in that The first loss of the sample text related to the text topic is measured based on the predicted probability value and the sample label value of the same sub-label under the text topic respectively related to the sample text, including: Based on the predicted probability values ​​and sample label values ​​of the same sub-labels under the text topic respectively related to the sample text, the sub-loss corresponding to the sub-label under the text topic related to the sample text is obtained; Based on the sub-weights and sub-losses of each of the sub-tags under the text topic, weighting is performed to obtain a first loss of the sample text related to the text topic; wherein the sub-weights of each of the sub-tags under the text topic are obtained based on inverse frequency statistics of the sample texts related to each of the sub-tags under the text topic.

11. The method according to claim 9, characterized in that The weight factor of the text topic is positively correlated with the discrimination accuracy requirement for the text topic.

12. The method according to claim 1, characterized in that After performing prediction based on the text to be reviewed and obtaining the probability values ​​of the sub-labels under each text topic respectively involved in the text to be reviewed, the method further includes: Determine whether the text to be reviewed involves a target topic based on the probability values ​​of the sub-tags under each of the text topics that the text to be reviewed involves; In response to the text to be reviewed involving the target topic, the text to be reviewed is directly intercepted, or, based on the probability value of the text to be reviewed involving any one of the first subtag and the second subtag under the target topic, it is determined whether to intercept the text to be reviewed.

13. A text content review device, characterized in that: include: The acquisition module is used to obtain the text to be reviewed; A prediction module is used to make predictions based on the text to be reviewed, and obtain the probability values ​​of the sub-labels under each text topic that the text to be reviewed involves respectively; wherein, any of the text topics includes a first sub-label, a second sub-label and a third sub-label, and the first sub-label represents that the content involving the text topic is compliant, the second sub-label represents that the content involving the text topic is not in compliance, and the third sub-label represents that the text topic is not involved.

14. An electronic device, characterized in that: It at least comprises a memory and a processor coupled to each other, wherein the memory at least stores program instructions, and the processor is used to execute the program instructions to implement the text content review method described in any one of claims 1 to 12.

15. A computer-readable storage medium, characterized in that: Program instructions that can be executed by a processor are stored, and the program instructions are used to implement the text content review method described in any one of claims 1 to 12.

Images