Data processing method and device, equipment and storage medium

Abstract

The invention relates to a data processing method and device, equipment and a storage medium, and the method comprises the steps: obtaining to-be-audited data which is the content that a user needs to publish; extracting a target text from the to-be-audited data, and determining a target text vector corresponding to the target text; querying a first entry related to the target text from a sensitive word library; querying a second entry related to the target text vector from a rule knowledge base; and based on the to-be-audited data, the first entry and the second entry, calling a pre-trained data auditing model to obtain a target auditing result of the to-be-audited data. According to the method, judgment that data auditing model generation does not accord with facts or deviates from actual conditions can be avoided, and the auditing accuracy and reliability are improved.

Description

Technical Field

[0001] This application relates to the field of computer technology, and in particular to a data processing method, apparatus, device, and storage medium. Background Technology

[0002] Data media platforms conduct content review in order to fulfill their social responsibility and legal obligations, prevent the spread of illegal, irregular, or harmful content on the platform, and avoid negative impacts on users, especially teenagers.

[0003] In related technologies, when reviewing user-uploaded content, the user-uploaded content is first obtained, and then this content is input into a large model (such as the GPT series). The large model judges whether the content violates regulations, and then takes corresponding processing measures based on the judgment results of the large model.

[0004] However, the judgment results of the general large model are unstable. Sometimes, it will produce a "hallucination" phenomenon, that is, generate judgments that are inconsistent with the facts or deviate from the actual situation. This will lead to misjudgments of user-uploaded content that are inconsistent with the platform's review standards. It may misjudge compliant content as illegal, or allow illegal content to pass the review by chance, affecting the accuracy and reliability of the review. Summary of the Invention

[0005] This application provides a data processing method, apparatus, device, and storage medium that can input the data to be reviewed, sensitive word data related to the data to be reviewed, and rule data into a pre-trained data review model. With the help of the model's deep learning capabilities, it can accurately capture explicit violation features in the data to be reviewed and identify implicit risks hidden in semantics and context. This avoids the data review model generating judgments that are inconsistent with the facts or deviate from the actual situation, thereby improving the accuracy and reliability of the review.

[0006] In a first aspect, this application provides a data processing method, the method comprising:

[0007] Obtain the data to be reviewed, which is the content that the user needs to publish;

[0008] Extract the target text from the data to be reviewed, and determine the target text vector corresponding to the target text;

[0009] From the sensitive word database, query the first entry related to the target text;

[0010] From the rule knowledge base, query the second entry related to the target text vector;

[0011] Based on the data to be reviewed, the first entry, and the second entry, a pre-trained data review model is invoked to obtain the target review result for the data to be reviewed.

[0012] Optionally, the step of calling a pre-trained data review model based on the data to be reviewed, the first entry, and the second entry to obtain the target review result for the data to be reviewed includes:

[0013] Based on the second entry, determine the target template to be used;

[0014] Extract the target information to be used from the second entry;

[0015] Add the data to be reviewed, the first entry, and the target information to the target template to obtain the target prompt words;

[0016] The target prompt words are input into a pre-trained data review model so that the pre-trained data review model outputs the target review result of the data to be reviewed.

[0017] Optionally, the second entry includes category information, and determining the target template to be used based on the second entry includes:

[0018] Count the number of occurrences of the same category of information in all second entries;

[0019] The category information whose occurrence frequency meets the preset conditions is identified as the target category information;

[0020] In the correspondence between preset category information and templates, the template corresponding to the target category information is determined and used as the target template.

[0021] Optionally, after obtaining the target audit result for the data to be audited, the method further includes:

[0022] Obtain the baseline audit result of the data to be audited;

[0023] When the benchmark review result and the target review result are inconsistent, a pre-trained rule optimization model is used to process the benchmark review result, the target review result, and the second item to obtain the first identifier of the rule item to be optimized and the target optimization suggestion.

[0024] The rule knowledge base is optimized and updated based on the first identifier and the target optimization suggestion.

[0025] Optionally, optimizing and updating the rule knowledge base based on the first identifier and the target optimization suggestion includes:

[0026] Determine the third entry corresponding to the first identifier from the rule knowledge base;

[0027] The third entry is optimized and updated using the aforementioned target optimization suggestions.

[0028] Optionally, querying the rule knowledge base for a second entry related to the target text vector includes:

[0029] Obtain the text vector identifying each rule entry in the rule knowledge base;

[0030] Calculate the similarity between the target text vector and the text vector of each rule entry identifier to obtain the similarity corresponding to each rule entry identifier;

[0031] Based on the similarity corresponding to each rule entry identifier, a second identifier is selected from all rule entry identifiers;

[0032] The second entry is determined from the rule knowledge base based on the second identifier.

[0033] Optionally, the step of obtaining the target review result of the data to be reviewed by calling a pre-trained data review model based on the data to be reviewed, the first entry, and the second entry includes:

[0034] Obtain the user characteristics of the user;

[0035] Based on the data to be reviewed, the first item, the second item, and the user characteristics, a pre-trained data review model is invoked to obtain the target review result of the data to be reviewed.

[0036] Secondly, this application provides a data processing apparatus, the apparatus comprising:

[0037] The acquisition unit is used to acquire data to be reviewed, which is the content that the user needs to publish.

[0038] An extraction unit is used to extract target text from the data to be reviewed, and to determine the target text vector corresponding to the target text;

[0039] The first query unit is used to query a first entry related to the target text from the sensitive word library;

[0040] The second query unit is used to query a second entry related to the target text vector from the rule knowledge base;

[0041] The calling unit is used to call a pre-trained data auditing model based on the data to be audited, the first entry, and the second entry to obtain the target auditing result of the data to be audited.

[0042] Optionally, the calling unit is used for:

[0043] Based on the second entry, determine the target template to be used;

[0044] Extract the target information to be used from the second entry;

[0045] Add the data to be reviewed, the first entry, and the target information to the target template to obtain the target prompt words;

[0046] The target prompt words are input into a pre-trained data review model so that the pre-trained data review model outputs the target review result of the data to be reviewed.

[0047] Optionally, the second entry includes category information, and the calling unit is used for:

[0048] Count the number of occurrences of the same category of information in all second entries;

[0049] The category information whose occurrence frequency meets the preset conditions is identified as the target category information;

[0050] In the correspondence between preset category information and templates, the template corresponding to the target category information is determined and used as the target template.

[0051] Optionally, after obtaining the target audit result of the data to be audited, the device further includes an optimization unit, the optimization unit being used to:

[0052] Obtain the baseline audit result of the data to be audited;

[0053] When the benchmark review result and the target review result are inconsistent, a pre-trained rule optimization model is used to process the benchmark review result, the target review result, and the second item to obtain the first identifier of the rule item to be optimized and the target optimization suggestion.

[0054] The rule knowledge base is optimized and updated based on the first identifier and the target optimization suggestion.

[0055] Optionally, the optimization unit is used for:

[0056] Determine the third entry corresponding to the first identifier from the rule knowledge base;

[0057] The third entry is optimized and updated using the aforementioned target optimization suggestions.

[0058] Optionally, the second query unit is used for:

[0059] Obtain the text vector identifying each rule entry in the rule knowledge base;

[0060] Calculate the similarity between the target text vector and the text vector of each rule entry identifier to obtain the similarity corresponding to each rule entry identifier;

[0061] Based on the similarity corresponding to each rule entry identifier, a second identifier is selected from all rule entry identifiers;

[0062] The second entry is determined from the rule knowledge base based on the second identifier.

[0063] Optionally, the calling unit is used for:

[0064] Obtain the user characteristics of the user;

[0065] Based on the data to be reviewed, the first item, the second item, and the user characteristics, a pre-trained data review model is invoked to obtain the target review result of the data to be reviewed.

[0066] Thirdly, this application provides a data processing device, comprising: at least one communication interface; at least one bus connected to the at least one communication interface; at least one processor connected to the at least one bus; and at least one memory connected to the at least one bus, wherein the processor is configured to:

[0067] Obtain the data to be reviewed, which is the content that the user needs to publish;

[0068] Extract the target text from the data to be reviewed, and determine the target text vector corresponding to the target text;

[0069] From the sensitive word database, query the first entry related to the target text;

[0070] From the rule knowledge base, query the second entry related to the target text vector;

[0071] Based on the data to be reviewed, the first entry, and the second entry, a pre-trained data review model is invoked to obtain the target review result for the data to be reviewed.

[0072] Fourthly, this application provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the above-described data processing method.

[0073] Compared with the prior art, the technical solution provided in this application embodiment has the following advantages: In this application embodiment, data to be reviewed is obtained, which is content to be published by the user; target text is extracted from the data to be reviewed, and the target text vector corresponding to the target text is determined; a first entry related to the target text is queried from a sensitive word lexicon; a second entry related to the target text vector is queried from a rule knowledge base; based on the data to be reviewed, the first entry, and the second entry, a pre-trained data review model is called to obtain the target review result of the data to be reviewed. In this application embodiment, the data to be reviewed, the sensitive word data related to the data to be reviewed, and the rule data are input into a pre-trained data review model. With the deep learning capability of the model, it can accurately capture the explicit violation features in the data to be reviewed, and identify the implicit risks hidden in semantics and context, avoiding the data review model from generating judgments that are inconsistent with the facts or deviate from the actual situation, thus improving the accuracy and reliability of the review. Attached Figure Description

[0074] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with the invention and, together with the description, serve to explain the principles of the invention.

[0075] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0076] One or more embodiments are illustrated by way of example with reference numerals in the accompanying drawings. These illustrations do not constitute a limitation on the embodiments. Elements with the same reference numerals in the drawings are denoted as similar elements. Unless otherwise stated, the figures in the drawings are not to be limited by scale.

[0077] Figure 1 A flowchart illustrating a data processing method provided in an embodiment of this application;

[0078] Figure 2 A flowchart illustrating a method for determining a target audit result provided in an embodiment of this application;

[0079] Figure 3 A flowchart illustrating a rule optimization method provided in an embodiment of this application;

[0080] Figure 4 A flowchart illustrating a second entry determination method provided in an embodiment of this application;

[0081] Figure 5This is a schematic flowchart of a data processing apparatus provided in an embodiment of this application;

[0082] Figure 6 This is a schematic diagram of a data processing device provided in an embodiment of this application. Detailed Implementation

[0083] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0084] The following disclosure provides numerous different embodiments or examples for implementing various structures of the invention. To simplify the disclosure, specific examples of components and arrangements are described below. These are merely examples and are not intended to limit the scope of the invention. Furthermore, reference numerals and / or letters may be repeated in different examples. Such repetition is for simplification and clarity and does not in itself indicate a relationship between the various embodiments and / or arrangements discussed.

[0085] Content moderation by data media platforms is a way to fulfill their social responsibility and legal obligations, prevent the spread of illegal or inappropriate content, and avoid negative impacts on users, especially teenagers. In related technologies, when reviewing user-uploaded content, the uploaded content is first obtained and then input into a large model (such as the GPT series). The model then judges whether the content violates regulations and takes appropriate action based on its judgment. However, the judgment results of general-purpose large models are unstable and sometimes exhibit "illusion" phenomena, generating judgments that are inconsistent with or deviate from reality. This can lead to misjudgments of user-uploaded content that do not conform to the platform's review standards, potentially misclassifying compliant content as illegal or allowing illegal content to pass review, thus affecting the accuracy and reliability of the review process.

[0086] To address the aforementioned issues, this application provides a data processing method. This method inputs the data to be reviewed, related sensitive word data, and rule data into a pre-trained data review model. Leveraging the model's deep learning capabilities, it can accurately capture explicit violation features in the data to be reviewed and identify implicit risks hidden in semantics and context. This prevents the data review model from generating judgments that are inconsistent with or deviate from reality, thus improving the accuracy and reliability of the review process. Figure 1 As shown, the specific steps include:

[0087] Step 101: Obtain the data to be reviewed.

[0088] The data to be reviewed is the content that users need to publish, which can be text, images, or other data.

[0089] In this step, after the user has edited the content to be published on the media data platform, they can click the publish button. When the system receives the trigger command from the publish button, it will automatically retrieve the content to be published and send it to the review system for the review system to extract the data to be reviewed.

[0090] Step 102: Extract the target text from the data to be reviewed, and determine the target text vector corresponding to the target text.

[0091] In this step, when the data to be reviewed is text, it can be directly used as the target text; or it can be simply processed to remove invalid text to obtain the target text. If the data to be reviewed is an image, it is converted into text using a pre-trained image recognition model, and this text is used as the target text.

[0092] After obtaining the target text, basic NLP processing such as word segmentation and part-of-speech tagging is performed to obtain multiple keywords. The target text and keywords are then converted into vectors and fused together to obtain the target text vector. The specific steps for converting text into vectors include: inputting the text into a pre-trained feature extraction model, which extracts features from the text, ultimately obtaining a vector corresponding to the text.

[0093] Step 103: Query the first entry related to the target text from the sensitive word database.

[0094] The sensitive word database contains multiple sensitive word entries, each of which includes information such as sensitive word entry identifier, category, risk level, standard description, positive examples, negative examples, and keywords.

[0095] In this step, the target text needs to be compared with the keywords of each word entry in the sensitive word database to select the preset keywords that are closest to the content of the target text, and the sensitive word entry to which the keyword belongs is determined as the first entry.

[0096] Step 104: Query the rule knowledge base for the second entry related to the target text vector.

[0097] The rule knowledge base contains multiple rule entries, each of which includes information such as rule entry identifier (e.g., rule ID), category information, risk level, standard description, positive examples, negative examples, and keywords.

[0098] In practical applications, to improve computational efficiency, each rule entry in the rule knowledge base can be pre-processed: its standard description and keywords can be converted into vectors, and these two types of vectors can be fused to obtain the text vector corresponding to each rule entry. After vector generation is completed, the system will associate the text vector of each rule entry with its rule entry identifier and store it in the vector library.

[0099] When this step is required, the system will call all text vectors stored in the vector library, compare them one by one with the target text vector, filter out the text vectors with high similarity to the target text vector, and extract the rule entry identifiers corresponding to these vectors. Subsequently, based on these rule entry identifiers, the corresponding rule entries are retrieved from the rule knowledge base, and these entries serve as the second entries.

[0100] Step 105: Based on the data to be reviewed, the first item, and the second item, call the pre-trained data review model to obtain the target review result for the data to be reviewed.

[0101] The pre-trained data review model is a large language model, which learns from a large amount of labeled data and aims to evaluate and judge input data to produce appropriate output results. This model may employ various machine learning or deep learning algorithms, such as neural networks and decision trees. Its training process typically involves selecting a suitable algorithm framework, learning from a large amount of sample data labeled with correct review results, and adjusting the model's parameters so that it can accurately output the corresponding review results when faced with similar data. The review results include whether the data to be reviewed violates regulations, the identifier of the rule entry used, and the reason for the violation; of course, other data may also be included, but this is not limited to this.

[0102] In this step, the data to be reviewed, the first item, and the second item are input into a pre-trained data review model. The data review model processes the data to be reviewed based on the first and second items to obtain the target review result for the data to be reviewed. Alternatively, prompt words are generated based on the data to be reviewed, the first item, and the second item. These prompt words are then input into the pre-trained data review model to obtain the target review result for the data to be reviewed.

[0103] In this embodiment, data to be reviewed is obtained, which is content to be published by a user; target text is extracted from the data to be reviewed, and a target text vector corresponding to the target text is determined; a first entry related to the target text is queried from a sensitive word lexicon; a second entry related to the target text vector is queried from a rule knowledge base; based on the data to be reviewed, the first entry, and the second entry, a pre-trained data review model is invoked to obtain the target review result of the data to be reviewed. In this embodiment, the data to be reviewed, sensitive word data related to the data to be reviewed, and rule data are input into a pre-trained data review model. With the deep learning capability of the model, it can accurately capture explicit violation features in the data to be reviewed, and identify implicit risks hidden in semantics and context, avoiding the data review model from generating judgments that are inconsistent with the facts or deviate from the actual situation, thus improving the accuracy and reliability of the review.

[0104] In this embodiment, the prompt word template to be used is first determined based on the second entry, then the target information to be used is extracted from the second entry, and then the data to be reviewed, the first entry, and the target information are filled into the prompt word template to obtain the target prompt word. Finally, the target prompt word is input into a pre-trained data review model, which outputs the target review result for the data to be reviewed. Therefore, this embodiment provides a method for determining the review result, which is a further limitation of step 104, such as... Figure 2 As shown, the specific steps are as follows:

[0105] Step 201: Determine the target template to be used based on the second entry.

[0106] Among them, the target template, which is the prompt word template to be used, is an instruction framework customized for different types of content to be reviewed. Its core function is to guide the data review model to accurately focus on key review dimensions by clarifying specific review priorities and directions.

[0107] Specifically, the prompt templates cover various types, including politically sensitive templates, abusive templates, and default templates. Politically sensitive templates include targeted review guidelines such as "Please pay special attention to analyze whether there is an intention to circumvent censorship through metaphors, insinuations, or coded language..." as well as content to be filled in with specific information. Abusive templates include core review requirements such as "Please determine whether the remarks constitute a direct attack, insult, or defamation against a specific individual or group..." and also reserve a content area to be filled in. Default templates provide general review instructions along with the content to be filled in.

[0108] In this step, since the second item includes category information, the category information in the second item can be analyzed to obtain the target category information corresponding to the data to be reviewed. Based on the target category information, the target template to be used is determined from the template library.

[0109] For example, if the target category information determined based on the second entry is political, then the template corresponding to the political category can be found in the template library and identified as the target template.

[0110] Step 202: Extract the target information to be used from the second entry.

[0111] In this step, the core data in the rule entries, such as the standard descriptions, positive examples, and negative examples, provide clear judgment criteria and reference standards for the review. Therefore, in the second entry, it is necessary to extract the standard descriptions, positive examples, and negative examples, and organize this data into easily understandable text fragments as the target information.

[0112] Step 203: Add the data to be reviewed, the first item, and the target information to the target template to obtain the target prompt words.

[0113] In this step, core data such as standard descriptions, positive examples, and negative examples are extracted from the first entry. Then, the data to be reviewed, this core data, and the target information are added to the specified positions in the target template to generate target prompts.

[0114] Step 204: Input the target prompt words into the pre-trained data review model so that the pre-trained data review model outputs the target review result of the data to be reviewed.

[0115] In this step, firstly, a preset API is invoked, providing the generated target prompts as input to a pre-trained data review model. Upon receiving the target prompts, the data review model analyzes and processes this information based on the patterns and knowledge it has learned during training. Finally, the model outputs the target review result for the data to be reviewed. This result may indicate whether the data passes review, what problems exist, etc., providing a basis for determining the compliance of the content a user needs to publish.

[0116] In this embodiment, since the second item includes category information, the frequency of occurrence of the same category information can be counted from all second items, and the category information that appears most frequently at that time can be determined as the target category information corresponding to the data to be reviewed. Then, the template corresponding to the target category information can be searched from the template library and used as the target template. Therefore, this embodiment provides a method for determining a target template. The specific steps of this method include: counting the frequency of occurrence of the same category information from all second items; determining the category information whose frequency of occurrence meets a preset condition as the target category information; and determining the template corresponding to the target category information in the preset correspondence between category information and templates, and using it as the target template.

[0117] The preset condition is to determine the category information that appears most frequently as the target category information. Alternatively, the category information that appears half of the total number of times can be determined as the target category information. Other conditions are also possible, but are not limited here.

[0118] In this step, for each second entry, category information is extracted, and the count of the corresponding category information in the counter is incremented by one. This method is used to count the occurrences of the same category information across all second entries. Then, the category information with the highest occurrence is identified and designated as the target category information. A pre-defined mapping table between category information and templates is obtained, recording the templates corresponding to different categories. After determining the target category information, the template corresponding to the target category information is found based on this mapping table.

[0119] It should be noted that if the frequency of occurrence of all categories of information does not meet the preset conditions, the default template can be used as the target template.

[0120] In this embodiment, after obtaining the target review result for the data to be reviewed, it further detects whether there is a human review result for the data. If a corresponding human review result exists, the review result output by the model is compared with it to determine whether they are consistent. Since human review results are generally highly accurate, when the model's review result and the human review result are inconsistent, it indicates a deviation in the model's output, which is often due to insufficient rule precision. Therefore, when the model's review result and the human review result are inconsistent, the system activates a rule optimization model to perform in-depth analysis of the target review result, the human review result, and the second item output by the model, thereby accurately identifying which rule items are inaccurate, leading to errors in the review result. Finally, the rule optimization model outputs the first identifier of the rule item to be optimized and corresponding target optimization suggestions. Based on these first identifiers and target optimization suggestions, targeted optimization of the rule knowledge base can be carried out to improve the accuracy of subsequent data review by the data review model. Therefore, this embodiment provides a rule optimization method, which is as follows: Figure 3 As shown, the specific steps include:

[0121] Step 301: Obtain the baseline audit results for the data to be audited.

[0122] The baseline audit result is the manual audit result of the data to be audited. In practice, the data audit model can be used to audit the data to be audited first to obtain the target audit result. Then, the data to be audited can be sent to the auditors to obtain the manual audit result of the data to be audited, which is the baseline audit result.

[0123] Step 302: When the baseline audit result and the target audit result are inconsistent, the pre-trained rule optimization model is used to process the baseline audit result, the target audit result and the second item to obtain the first identifier of the rule item to be optimized and the target optimization suggestion.

[0124] The target optimization suggestions are used to optimize the rule entries to be optimized. These suggestions include suggestion type, suggestion content, and suggestion reason. Suggestion types include adding exceptions, clarifying descriptions, and adding negative examples. Suggestion content can be something like, "Add the following to the end of the rule description: 'Note: This rule does not apply to obvious satirical or comedic scenarios,'" or other content. Suggestion reasons can be something like, "The original rule was too broad and failed to cover the common context of satire, leading to misjudgments of compliant content," or other content.

[0125] In this step, the baseline review results and the target review results are compared to check if they are consistent. If they are consistent, a prompt word template for rule optimization is obtained, and the baseline review results, the target review results, and the second item are added to the prompt word template to obtain the prompt word. This prompt word is then input into a pre-trained rule optimization model to obtain the first identifier of the rule item to be optimized and the target optimization suggestion.

[0126] The pre-trained rule optimization model and the pre-trained data review model can be the same model or different models. When they are different models, the pre-trained rule optimization model is also a large language model, which learns from a large amount of labeled data and aims to evaluate and judge the input data to produce the desired output. This model may employ various machine learning or deep learning algorithms, such as neural networks and decision trees. Its training process typically involves selecting a suitable algorithm framework, learning from a large amount of labeled sample data with identifiers of the rules to be optimized, and adjusting the model's parameters so that the model can accurately output the corresponding identifiers of the rules to be optimized when faced with similar data.

[0127] Step 303: Optimize and update the rule knowledge base based on the first identifier and the target optimization suggestions.

[0128] It should be noted that when there are a large number of target audit results that are identical to the baseline audit results, this is likely due to inaccuracies in certain entries. Therefore, the cases corresponding to these target audit results can be clustered to obtain multiple case clusters. Since the cases in each case cluster are quite similar, the rules used should also be similar. Therefore, the rule optimization method described above can be applied only to the typical cases in these case clusters to optimize their corresponding rules. After implementing the rule optimization method, the results are validated based on other cases in the case clusters.

[0129] A typical case can be the case corresponding to the central text vector or the text vector closest to the central text vector. The central text vector can be the average vector of the text vectors corresponding to all cases in the case cluster.

[0130] In this embodiment, after determining the rule entry identifier of the rule to be optimized, the rule entry corresponding to the identifier is determined from the rule knowledge base, and the rule entry is optimized and updated using the target optimization suggestion. Therefore, this embodiment provides a rule entry optimization method, the specific steps of which include: determining the third entry corresponding to the first identifier from the rule knowledge base; and optimizing and updating the third entry using the target optimization suggestion.

[0131] In this step, the core content of the third item can be supplemented based on the target optimization suggestions to obtain a new third item. The target optimization suggestions can also include optimized content, which can be used directly to update the third item, resulting in an updated third item.

[0132] Before optimizing the third item based on the target optimization suggestion, it is also necessary to generate an optimization notification based on the first identifier and the target optimization suggestion. This notification is used to inform relevant personnel to optimize and update the rule knowledge base and to check whether the target optimization suggestion is correct.

[0133] In this embodiment, the text vectors of each rule entry identifier in the rule knowledge base are first obtained. Then, the similarity between the target text vector and the text vector of each rule entry identifier is calculated to obtain the similarity value corresponding to each rule entry identifier. Based on these similarity values, the rule entry identifier corresponding to the second entry is selected from all rule entry identifiers. Finally, based on the selected rule entry identifier, the second entry is determined in the rule knowledge base. Therefore, this embodiment provides a method for determining a second entry, as follows: Figure 4 As shown, the specific steps include:

[0134] Step 401: Obtain the text vector of each rule entry identifier in the rule knowledge base.

[0135] In this step, each rule entry and its corresponding rule entry identifier can be obtained from the rule knowledge base. For each rule entry identifier, the standard description and keywords in its corresponding rule entry are extracted, and these contents are converted into text vectors to obtain the text vector of each rule entry identifier.

[0136] Step 402: Calculate the similarity between the target text vector and the text vector of each rule entry identifier to obtain the similarity corresponding to each rule entry identifier.

[0137] In this step, for each rule entry identifier, the cosine value between the corresponding text vector and the target text vector is calculated, and this value is used as the similarity between the rule entry identifier and the target text vector.

[0138] Step 403: Select the second identifier from all rule entry identifiers based on the similarity corresponding to each rule entry identifier.

[0139] In this step, all rule entry identifiers are sorted from largest to smallest based on their similarity, thus generating a rule entry identifier sequence. Then, the first preset number of rule entry identifiers in this sequence are selected and designated as the second identifier.

[0140] Step 404: Determine the second entry from the rule knowledge base based on the second identifier.

[0141] In this step, the rule entry identifier in each rule entry in the rule knowledge base is compared with the second identifier. When there is a rule entry identifier that is the same as the second identifier, the rule entry identifier in the comparison is determined as the second entry.

[0142] In this embodiment, user characteristics (historical behavior, profile) provide clues to user intent, context of behavioral logic, and a personalized baseline, helping the model penetrate the surface form of content and more accurately determine its true risk—ultimately reducing false positives and false negatives, thereby significantly improving the accuracy of the review process. After obtaining the user's characteristics, the data to be reviewed, the first entry obtained from the sensitive word database, the second entry obtained from the rule knowledge base, and the user characteristics are combined and used as input to call a pre-trained data review model, thereby obtaining the target review result for the data to be reviewed. Therefore, this embodiment provides a method for determining the review result, the specific steps of which include: obtaining the user's characteristics; and, based on the data to be reviewed, the first entry, the second entry, and the user characteristics, calling a pre-trained data review model to obtain the target review result for the data to be reviewed.

[0143] User characteristics encompass both historical user behavior and user profiles. Specifically, this refers to the past user behavior and profiles of the data publishers to be reviewed. Historical user behavior includes aspects such as the percentage of violations in user-posted content, and the type and frequency of past posts. User profiles, on the other hand, are user characterizations constructed by integrating multiple dimensions of information, including age, gender, region, interests, and occupation. They help the data review model gain a more comprehensive understanding of user characteristics.

[0144] In this step, the user characteristics of the publisher of the data to be reviewed are obtained. Then, the data to be reviewed, the core data from the first item, the target data from the second item, and the user characteristics are added to the target template to obtain the target prompt words. The target prompt words are then input into a pre-trained data review model to obtain the target review result for the data to be reviewed.

[0145] like Figure 5 As shown, this application provides a data processing apparatus, which corresponds to the method embodiment, and specifically includes:

[0146] The acquisition unit 501 is used to acquire data to be reviewed, which is the content that the user needs to publish.

[0147] Extraction unit 502 is used to extract target text from the data to be reviewed, and to determine the target text vector corresponding to the target text;

[0148] The first query unit 503 is used to query a first entry related to the target text from the sensitive word library;

[0149] The second query unit 504 is used to query a second entry related to the target text vector from the rule knowledge base;

[0150] The calling unit 505 is used to call a pre-trained data auditing model based on the data to be audited, the first entry, and the second entry to obtain the target auditing result of the data to be audited.

[0151] Optionally, the calling unit 505 is used for:

[0152] Based on the second entry, determine the target template to be used;

[0153] Extract the target information to be used from the second entry;

[0154] Add the data to be reviewed, the first entry, and the target information to the target template to obtain the target prompt words;

[0155] The target prompt words are input into a pre-trained data review model so that the pre-trained data review model outputs the target review result of the data to be reviewed.

[0156] Optionally, the second entry includes category information, and the calling unit 505 is used to:

[0157] Count the number of occurrences of the same category of information in all second entries;

[0158] The category information whose occurrence frequency meets the preset conditions is identified as the target category information;

[0159] In the correspondence between preset category information and templates, the template corresponding to the target category information is determined and used as the target template.

[0160] Optionally, after obtaining the target audit result of the data to be audited, the device further includes an optimization unit 506, which is used to:

[0161] Obtain the baseline audit result of the data to be audited;

[0162] When the benchmark review result and the target review result are inconsistent, a pre-trained rule optimization model is used to process the benchmark review result, the target review result, and the second item to obtain the first identifier of the rule item to be optimized and the target optimization suggestion.

[0163] The rule knowledge base is optimized and updated based on the first identifier and the target optimization suggestion.

[0164] Optionally, the optimization unit 506 is used for:

[0165] Determine the third entry corresponding to the first identifier from the rule knowledge base;

[0166] The third entry is optimized and updated using the aforementioned target optimization suggestions.

[0167] Optionally, the second query unit 504 is used for:

[0168] Obtain the text vector identifying each rule entry in the rule knowledge base;

[0169] Calculate the similarity between the target text vector and the text vector of each rule entry identifier to obtain the similarity corresponding to each rule entry identifier;

[0170] Based on the similarity corresponding to each rule entry identifier, a second identifier is selected from all rule entry identifiers;

[0171] The second entry is determined from the rule knowledge base based on the second identifier.

[0172] Optionally, the calling unit 505 is used for:

[0173] Obtain the user characteristics of the user;

[0174] Based on the data to be reviewed, the first item, the second item, and the user characteristics, a pre-trained data review model is invoked to obtain the target review result of the data to be reviewed.

[0175] like Figure 6As shown in the figure, this application provides a data processing device, including a processor 601, a communication interface 602, a memory 603, and a communication bus 604, wherein the processor 601, the communication interface 602, and the memory 603 communicate with each other through the communication bus 604.

[0176] Memory 603 is used to store computer programs;

[0177] In one embodiment of this application, when the processor 601 executes a program stored in the memory 603, it implements the data processing method provided in any of the foregoing method embodiments, including:

[0178] Obtain the data to be reviewed, which is the content that the user needs to publish;

[0179] Extract the target text from the data to be reviewed, and determine the target text vector corresponding to the target text;

[0180] From the sensitive word database, query the first entry related to the target text;

[0181] From the rule knowledge base, query the second entry related to the target text vector;

[0182] Based on the data to be reviewed, the first entry, and the second entry, a pre-trained data review model is invoked to obtain the target review result for the data to be reviewed.

[0183] This application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the data processing steps provided in any of the foregoing method embodiments.

[0184] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.

[0185] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented using software plus a general-purpose hardware platform, or of course, using hardware. Based on this understanding, the above technical solutions, in essence or the parts that contribute to the related technology, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0186] It should be understood that the terminology used herein is for the purpose of describing particular exemplary embodiments only and is not intended to be limiting. Unless the context clearly indicates otherwise, the singular forms “a,” “an,” and “described” as used herein may also include the plural forms. The terms “comprising,” “including,” “containing,” and “having” are inclusive and therefore indicate the presence of the stated features, steps, operations, elements, and / or components, but do not exclude the presence or addition of one or more other features, steps, operations, elements, components, and / or combinations thereof. The method steps, processes, and operations described herein are not construed as requiring them to be performed in a particular order described or illustrated unless the order of performance is explicitly indicated. It should also be understood that additional or alternative steps may be used.

[0187] The above description is merely a specific embodiment of the present invention, enabling those skilled in the art to understand or implement the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the present invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features claimed herein.

Claims

1. A data processing method, characterized in that, The method includes: Obtain the data to be reviewed, which is the content that the user needs to publish; Extract the target text from the data to be reviewed, and determine the target text vector corresponding to the target text; From the sensitive word database, query the first entry related to the target text; From the rule knowledge base, query the second entry related to the target text vector; Based on the data to be reviewed, the first entry, and the second entry, a pre-trained data review model is invoked to obtain the target review result for the data to be reviewed.

2. The method according to claim 1, characterized in that, The step of calling a pre-trained data review model based on the data to be reviewed, the first entry, and the second entry to obtain the target review result for the data to be reviewed includes: Based on the second entry, determine the target template to be used; Extract the target information to be used from the second entry; Add the data to be reviewed, the first entry, and the target information to the target template to obtain the target prompt words; The target prompt words are input into a pre-trained data review model so that the pre-trained data review model outputs the target review result of the data to be reviewed.

3. The method according to claim 2, characterized in that, The second entry includes category information, and determining the target template to be used based on the second entry includes: Count the number of occurrences of the same category of information in all second entries; The category information whose occurrence frequency meets the preset conditions is identified as the target category information; In the correspondence between preset category information and templates, the template corresponding to the target category information is determined and used as the target template.

4. The method according to claim 1, characterized in that, After obtaining the target audit result for the data to be audited, the method further includes: Obtain the baseline audit result of the data to be audited; When the benchmark review result and the target review result are inconsistent, a pre-trained rule optimization model is used to process the benchmark review result, the target review result, and the second item to obtain the first identifier of the rule item to be optimized and the target optimization suggestion. The rule knowledge base is optimized and updated based on the first identifier and the target optimization suggestion.

5. The method according to claim 4, characterized in that, The step of optimizing and updating the rule knowledge base based on the first identifier and the target optimization suggestion includes: Determine the third entry corresponding to the first identifier from the rule knowledge base; The third entry is optimized and updated using the aforementioned target optimization suggestions.

6. The method according to claim 1, characterized in that, The step of querying the second entry related to the target text vector from the rule knowledge base includes: Obtain the text vector identifying each rule entry in the rule knowledge base; Calculate the similarity between the target text vector and the text vector of each rule entry identifier to obtain the similarity corresponding to each rule entry identifier; Based on the similarity corresponding to each rule entry identifier, a second identifier is selected from all rule entry identifiers; The second entry is determined from the rule knowledge base based on the second identifier.

7. The method according to claim 1, characterized in that, The step of obtaining the target review result of the data to be reviewed by calling a pre-trained data review model based on the data to be reviewed, the first item, and the second item includes: Obtain the user characteristics of the user; Based on the data to be reviewed, the first item, the second item, and the user characteristics, a pre-trained data review model is invoked to obtain the target review result of the data to be reviewed.

8. A data processing apparatus, characterized in that, The device includes: The acquisition unit is used to acquire data to be reviewed, which is the content that the user needs to publish. An extraction unit is used to extract target text from the data to be reviewed, and to determine the target text vector corresponding to the target text; The first query unit is used to query a first entry related to the target text from the sensitive word library; The second query unit is used to query a second entry related to the target text vector from the rule knowledge base; The calling unit is used to call a pre-trained data auditing model based on the data to be audited, the first entry, and the second entry to obtain the target auditing result of the data to be audited.

9. A data processing device, characterized in that, include: At least one communication interface; At least one bus connected to the at least one communication interface; at least one processor connected to the at least one bus; At least one memory connected to the at least one bus, wherein the processor is configured to: Obtain the data to be reviewed, which is the content that the user needs to publish; Extract the target text from the data to be reviewed, and determine the target text vector corresponding to the target text; From the sensitive word database, query the first entry related to the target text; From the rule knowledge base, query the second entry related to the target text vector; Based on the data to be reviewed, the first entry, and the second entry, a pre-trained data review model is invoked to obtain the target review result for the data to be reviewed.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the data processing method according to any one of claims 1 to 7.

Images