Text auditing method and related device

Abstract

The invention provides a text auditing method and a related device, and the method comprises the steps: carrying out the primary auditing of a to-be-audited text through one auditing rule of a first auditing rule and a second auditing rule, and carrying out the primary auditing of the to-be-audited text through the other auditing rule of the first auditing rule and the second auditing rule after the to-be-audited text passes the primary auditing. Performing secondary auditing on the to-be-audited text, wherein the first auditing rule and the second auditing rule are different in auditing direction; after the to-be-audited text passes the secondary audit, the to-be-audited text is published; after the to-be-audited text is published, multiple pieces of user feedback data of the to-be-audited text are obtained; according to the multiple pieces of user feedback data, performing third-level auditing on the to-be-audited text; if the to-be-audited text passes third-level audit, continuing to publish the to-be-audited And if the to-be-audited text does not pass the third-level audit, stopping publishing the to-be-audited

Description

Technical Field

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

[0002] With the development of Internet technology, more and more users are browsing text online. For example, news texts are posted online and users can browse them online. Before a text is posted, it needs to be reviewed to filter out text containing sensitive content. Text containing sensitive content is generally considered low-quality text or text with low-quality issues, and such text needs to be taken offline. Currently, there are three text review methods:

[0003] The first method, rule- and pattern-matching-based text review, uses pre-set keywords or combinations of pre-set rules for review. Regular expressions, for example, support complex rule combinations, enabling flexible identification of specific text structures or semantic patterns. For example, matching rule combinations such as "Click below.{,20} to purchase" can be used to identify whether a text contains advertising or marketing content. Pre-set keywords can be sensitive terms, such as advertising or marketing terms, or terms associated with illegal activities. AC automata are used to identify whether a text contains these keywords. This rule- and pattern-matching-based text review method demonstrates high efficiency and accuracy when processing text with clear features, such as pre-set keywords or combinations of pre-set rules, which are used to identify whether a text contains sensitive content.

[0004] The second method is a classification and discrimination method based on text attributes: This method usually defines an audit category, such as reviewing any of the following categories: pornographic content, personal injury content, or content that endangers public safety. After defining the audit category, text samples related to the audit category are collected, and the text samples are used to train a text classification model. The trained text classification model is then used to identify whether the category of the text to be published is a defined audit category. The classification and discrimination method based on text attributes can identify content related to the audit category implicit in the text, thereby realizing the automatic identification of content related to the audit category, and has a certain degree of generalization ability.

[0005] The third method is manual review: This involves users reviewing the text to be published, for example, to check for logical confusion, content conflicts, and the presence of metaphorical and sensitive words. The advantage of manual review lies in the user's ability to understand complex semantics and context, allowing for highly flexible and accurate judgments. However, its main disadvantages are high cost and susceptibility to human fatigue and subjectivity, making it difficult to achieve large-scale, efficient review.

[0006] In summary, the current text review method is to review the text before it is published. Summary of the Invention

[0007] In view of the above problems, this application provides a text review method and related devices to achieve the purpose of reviewing the text using user feedback data after the text is published. The specific solution is as follows:

[0008] The first aspect of the present application provides a text review method, comprising:

[0009] Performing a primary review of a document to be reviewed using one of a first review rule and a second review rule, and after the document to be reviewed passes the primary review, performing a secondary review of the document to be reviewed using the other of the first review rule and the second review rule, wherein the first review rule and the second review rule have different review directions;

[0010] After the text to be reviewed passes the secondary review, the text to be reviewed is released;

[0011] After the text to be reviewed is published, obtaining multiple user feedback data of the text to be reviewed;

[0012] Conducting a three-level review of the text to be reviewed based on the multiple pieces of user feedback data;

[0013] If the text to be reviewed passes the third-level review, continue to publish the text to be reviewed; if the text to be reviewed does not pass the third-level review, stop publishing the text to be reviewed.

[0014] In a possible implementation, performing a three-level review on the text to be reviewed based on the multiple pieces of user feedback data includes:

[0015] Identifying negative review data from the plurality of user feedback data;

[0016] When the distribution of the negative comment data in the plurality of user feedback data meets the large model evaluation conditions, a three-level review is performed on the text to be reviewed using the preset large model.

[0017] In a possible implementation, identifying negative comment data from the plurality of user feedback data includes:

[0018] Generate a feature vector for each piece of user feedback data;

[0019] Inputting the feature vector of the user feedback data into a comment tendency classification model to obtain the category to which the user feedback data output by the comment tendency classification model belongs, wherein the comment tendency classification model is trained based on the feature vector of historical user feedback data and the category to which the historical user feedback data belongs, and the comment tendency classification model can identify multiple comment tendency categories, wherein the target comment tendency category among the multiple comment tendency categories is a negative comment category;

[0020] If the category to which the user feedback data belongs is the target review tendency category, it is determined that the user feedback data is negative review data.

[0021] In a possible implementation, when the distribution of the negative comment data in the plurality of user feedback data meets the large model evaluation conditions, performing a three-level review of the to-be-reviewed text using a preset large model includes:

[0022] When the number of the negative comment data is not less than a first threshold, and the proportion of the negative comment data in the plurality of user feedback data is not less than a first preset proportion but less than a second preset proportion, performing a three-level review on the to-be-reviewed text using the preset large model, and the second preset proportion is greater than the first preset proportion;

[0023] When the proportion of the negative comment data in the multiple user feedback data is not less than the second preset proportion, and the number of the negative comment data is not less than the second threshold but less than the first threshold, the text to be reviewed is subjected to a three-level review through the preset large model, and the first threshold is greater than the second threshold.

[0024] In a possible implementation, the method further includes: determining that the to-be-reviewed text has failed the third-level review if the number of the negative comment data is not less than the first threshold and the proportion of the negative comment data in the plurality of user feedback data is not less than the second preset proportion;

[0025] If the number of the negative comment data is less than the first threshold, and the proportion of the negative comment data in the plurality of user feedback data is not less than the first preset proportion but less than the second preset proportion, determining that the to-be-reviewed text has passed the third-level review;

[0026] A preset number of user feedback data in the plurality of pieces of user feedback data relate to specific event content, and it is determined that the text to be reviewed has not passed the third-level review.

[0027] In a possible implementation, the method further includes: adjusting the review direction of the first review rule through the manual interaction interface of the first-level review;

[0028] and / or,

[0029] Adjusting the review direction of the second review rule through the manual interaction interface of the secondary review;

[0030] and / or,

[0031] Review the audit results of the secondary audit and provide feedback on the review results through the manual interaction interface of the secondary audit;

[0032] and / or,

[0033] The audit results of the three-level audit are reviewed and the review results are fed back through the manual interaction interface of the three-level audit.

[0034] In one possible implementation, performing a primary review on the document to be reviewed using one of the first review rule and the second review rule, and performing a secondary review on the document to be reviewed using the other of the first review rule and the second review rule after the document to be reviewed passes the primary review, includes:

[0035] Using the first review rule to review whether the text to be reviewed contains sensitive content;

[0036] If the text to be reviewed contains sensitive content, it is determined that the text to be reviewed has not passed the first-level review, and the release of the text to be reviewed is prohibited;

[0037] If the text to be reviewed does not contain sensitive content, determine that the text to be reviewed has passed the first-level review, and use the second review rule to review the text to be reviewed from a linguistic perspective to determine whether there are any text quality issues;

[0038] If the text to be reviewed has no text quality issues, it is determined that the text to be reviewed has passed the secondary review.

[0039] In a possible implementation, the using the second review rule to review whether the to-be-reviewed text has text quality issues from a linguistic perspective includes:

[0040] Obtaining a review result of the to-be-reviewed text in the review direction of the second review rule using a preset large language model, where the review direction of the second review rule includes at least one of whether there are grammatical errors, whether the text is readable, and whether the title is ambiguous;

[0041] If the review result indicates that the text to be reviewed does not have the problem indicated by the review direction, it is determined that the text to be reviewed does not have text quality problems;

[0042] If the review result indicates that the text to be reviewed has the problem indicated by the review direction, it is determined that the text to be reviewed has a text quality problem;

[0043] The method of obtaining the audit result of the text to be audited in the audit direction of the second audit rule by using the preset large language model includes: inputting the text to be audited into the preset large language model, and obtaining the audit result of the text to be audited in the audit direction of the second audit rule output by the preset large language model, wherein the preset large language model is obtained by fine-tuning the model parameters using the fine-tuning dataset of the audit direction of the second audit rule, and the fine-tuning dataset of the audit direction of the second audit rule includes historical text to be audited, the audit result of the historical text to be audited output by the preset large language model, and the historical text manually annotated according to the audit direction of the second audit rule. Alternatively, the method of obtaining the audit result of the text to be audited in the audit direction of the second audit rule by using the preset large language model includes: inputting the text to be audited and preset prompt information into the preset large language model, and obtaining the output result of the preset large language model, wherein the preset prompt information is used to indicate that the text to be audited is to be audited from the audit direction of the second audit rule, and the preset prompt information also indicates the format of the output result of the preset large language model, and the output result records the audit result in the audit direction of the second audit rule.

[0044] A second aspect of the present application provides a text review device, comprising:

[0045] a first review module configured to perform a primary review of a document to be reviewed using one of a first review rule and a second review rule, and, after the document to be reviewed passes the primary review, perform a secondary review of the document to be reviewed using the other of the first review rule and the second review rule, wherein the first review rule and the second review rule have different review directions;

[0046] A publishing module, configured to publish the text to be reviewed after the text to be reviewed passes the secondary review;

[0047] An acquisition module, configured to acquire multiple pieces of user feedback data on the text to be reviewed after the text to be reviewed is published;

[0048] A second review module is used to perform a three-level review on the text to be reviewed based on the multiple user feedback data;

[0049] The control module is used to control the publishing module to continue publishing the text to be reviewed if the text to be reviewed passes the third-level review; if the text to be reviewed does not pass the third-level review, control the publishing module to stop publishing the text to be reviewed.

[0050] A third aspect of the present application provides an electronic device, comprising at least one processor and a memory connected to the processor, wherein:

[0051] The memory is used to store computer programs;

[0052] The processor is used to execute the computer program so that the electronic device can implement the text review method of the above-mentioned first aspect or any implementation method of the first aspect.

[0053] The fourth aspect of the present application provides a computer program product, including computer-readable instructions. When the computer-readable instructions are executed on an electronic device, the electronic device implements the text review method of the first aspect or any implementation of the first aspect.

[0054] In a fifth aspect, the present application provides a computer storage medium, which carries one or more computer programs. When the one or more computer programs are executed by an electronic device, the electronic device can implement the text review method of the above-mentioned first aspect or any implementation of the first aspect.

[0055] By means of the above technical solution, the text review method and related devices provided by the present application divide the review tasks of the text to be reviewed into three main levels: first-level review, second-level review and third-level review. The first-level review utilizes one of the first review rule and the second review rule, the second-level review utilizes the other of the first review rule and the second review rule, and the review directions of the first review rule and the second review rule are different. The third-level review utilizes multiple user feedback data of the text to be reviewed, thereby forming a hierarchical serial review processing flow, so that the various review methods in the review processing flow can be effectively coordinated, and the review results of different review methods are used to make a comprehensive judgment on the text quality of the text to be reviewed, thereby improving accuracy and precision. In the hierarchical serial review processing flow, the number of texts to be reviewed for the second-level review and the third-level review is reduced, which effectively reduces the computational cost of the second-level review and the third-level review, avoids the high-cost comprehensive analysis of all texts to be reviewed, thereby avoiding a large amount of intermediate storage and computing costs, ensuring the reasonable allocation of computing resources and storage resources to reduce the repeated consumption of computing resources and storage resources, and improving operational efficiency and scalability.

[0056] In addition, user feedback data is introduced into the third-level review to form a closed loop for the review of the text to be reviewed. User feedback data is used to evaluate whether there are hidden quality problems in the text to be reviewed, effectively filling in the review blind spots that may exist in the first-level and second-level reviews, and realizing retrospective judgment and continuous monitoring of the text after release, further improving the efficiency, accuracy and precision of text review. BRIEF DESCRIPTION OF THE DRAWINGS

[0057] The above and other features, advantages, and aspects of the various embodiments of the present disclosure will become more apparent with reference to the following detailed description in conjunction with the accompanying drawings. Throughout the drawings, the same or similar reference numerals represent the same or similar elements. It should be understood that the drawings are schematic and that the originals and elements are not necessarily drawn to scale.

[0058] Figure 1 A flowchart of a text review method provided for this application;

[0059] Figure 2 A schematic diagram of the three-level review provided for this application;

[0060] Figure 3 Another flowchart of a text review method provided for this application;

[0061] Figure 4 A schematic diagram of the text review method provided for this application;

[0062] Figure 5 A schematic diagram of the single-category classification template provided for this application;

[0063] Figure 6 A schematic diagram of the structure of the text review method provided for this application;

[0064] Figure 7 This is a schematic diagram of the structure of the electronic device provided in this application. DETAILED DESCRIPTION

[0065] The following describes the embodiments of the present application in conjunction with the accompanying drawings. The terms used in the implementation methods of the present application are only used to explain the specific embodiments of the present application and are not intended to limit the present application.

[0066] The embodiments of the present application are described below in conjunction with the accompanying drawings. Those skilled in the art will appreciate that, with the development of technology and the emergence of new scenarios, the technical solutions provided in the embodiments of the present application are also applicable to similar technical problems.

[0067] The terms "first", "second", etc. in the specification and claims of the present application and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequential order. It should be understood that the terms used in this way can be interchangeable under appropriate circumstances, and this is merely a way of distinguishing the objects of the same attributes when describing them in the embodiments of the present application. In addition, the terms "including" and "having" and any of their variations are intended to cover non-exclusive inclusions, so that the process, method, system, product or equipment comprising a series of units need not be limited to those units, but may include other units that are not clearly listed or inherent to these processes, methods, systems, products or equipment.

[0068] Current text review methods include text review methods based on rules and pattern matching, classification and discrimination methods based on text attributes, and manual review methods. These three review methods all conduct reviews before the text is published, and generally ignore the use of user feedback data after the text is published. User feedback data includes user comment data, likes, dislikes, and shares, etc. User feedback data contains implicit quality problem evaluations of text quality. Implicit quality problems are quality problems that cannot be analyzed from the text content through large models. For example, if a text has multiple negative reviews, then the text has implicit quality problems. Therefore, current text review methods cannot promptly discover implicit quality problems after the text is published, resulting in the current text review methods lacking a retrospective discrimination and continuous monitoring mechanism for the published text.

[0069] Furthermore, current text review methods often operate independently, lacking a unified architecture. This independent design hinders effective collaboration between these methods and hinders comprehensive assessment of text quality based on different review results for the same document. Furthermore, independent operation can lead to redundant consumption of computing and storage resources, reducing operational efficiency and scalability.

[0070] In response to the above technical problems, the embodiment of the present application provides a text review method and related devices, which divide the review tasks of the text to be reviewed into three main levels: first-level review, second-level review and third-level review. The first-level review uses one of the first review rule and the second review rule, the second-level review uses the other review rule of the first review rule and the second review rule, and the review directions of the first review rule and the second review rule are different. The third-level review uses multiple user feedback data of the text to be reviewed, thereby forming a hierarchical serial review processing flow, so that the various review methods in the review processing flow can be effectively coordinated, and the review results of different review methods are used to make a comprehensive judgment on the text quality of the text to be reviewed, thereby improving accuracy and precision. In the hierarchical serial review processing flow, the number of texts to be reviewed for the second-level review and the third-level review is reduced, which effectively reduces the computing cost of the second-level review and the third-level review, avoids the high-cost comprehensive analysis of all texts to be reviewed, thereby avoiding a large amount of intermediate storage and computing costs, ensuring the reasonable allocation of computing resources and storage resources to reduce the repeated consumption of computing resources and storage resources, and improving operation efficiency and scalability.

[0071] Furthermore, user feedback data is introduced into the third-level review process, forming a closed loop for the review of the text to be reviewed. This allows the user feedback data to be used to evaluate whether the text to be reviewed has any hidden quality issues, effectively filling in the potential audit blind spots of the first-level and second-level reviews, and enabling retrospective identification and continuous monitoring of the published text, further improving the efficiency, accuracy, and precision of the text review. Furthermore, through the manual interaction interface, changes at any level of the review process can allow for local adjustments to the review process rather than global changes, greatly improving the maintainability and scalability of the review process.

[0072] The following describes the text review method and related devices provided by the embodiments of the present application. First, the terms involved in the embodiments of the present application are explained:

[0073] Multimodal quality matching: This involves assessing the quality of a text across multiple quality dimensions, including manually agreed-upon quality modalities, inherent text quality modalities, and externally evaluated quality modalities, using a hierarchical, serial review process to ultimately obtain a comprehensive quality evaluation result. This comprehensive quality evaluation result is used to determine whether the text to be reviewed will be taken offline or released after undergoing the hierarchical, serial review process.

[0074] Manually agreed quality mode: This refers to the use of manually preset rules to define text that is unsuitable for publication. For example, text containing sensitive content can be considered unsuitable for publication, and text that is unsuitable for distribution can also be considered low-quality text. Sensitive content can include at least one of politically sensitive content, advertising and marketing content, and vulgar content. Such text is prohibited from publication and must be taken offline.

[0075] Intrinsic Text Quality Mode: This refers to identifying, from a linguistic perspective, whether a text has quality issues, such as grammatical errors, incoherent logic, poor readability, inconsistencies between titles and content, or ambiguity. These issues typically reflect the standardization and professionalism of the text. Texts with quality issues can be prohibited from publication or released after corrections have been made to ensure they are no longer problematic.

[0076] External evaluation quality modality refers to hidden quality issues (also known as potential quality issues) revealed by analyzing user feedback data (such as user comments, likes, dislikes, shares, and semantically positive and negative opinions) after a text is published and interacts with users. This feedback provides a posteriori verification of text quality assessments. Texts with hidden quality issues (such as negative reviews) require review to determine whether they should be banned from publication.

[0077] Responsive: In response to scenarios such as audit rule adjustments or emergencies, the system provides a human interaction interface to enable real-time dynamic configuration and assembly of audit rules to meet current audit needs and quickly respond to changes, thereby improving the flexibility and responsiveness of text audits.

[0078] See Figure 1 , which shows an optional process of a text review method provided in an embodiment of the present application, which may include the following steps:

[0079] S101. Perform a first-level review on the text to be reviewed using one of the first review rule and the second review rule. After the text to be reviewed passes the first-level review, perform a second-level review on the text to be reviewed using the other of the first review rule and the second review rule.

[0080] Among them, the review directions of the first review rule and the second review rule are different. For example, the first review rule can be an audit of the manually agreed quality mode to identify whether the text to be reviewed includes sensitive content. The second review rule can be an audit of the text's inherent quality mode to review whether the text to be reviewed has text quality problems.

[0081] The first review rule (review of manually agreed-upon quality modalities) is to identify and process low-quality text, defined according to manually pre-set rules. Low-quality text can include sensitive content. These texts often have significant characteristics (e.g., sensitive content). These texts can be efficiently identified using content matching, pattern recognition, or classification models. For example, sensitive content such as politically sensitive content, advertising and marketing content, or vulgar content can be identified using keyword matching, blacklist and whitelist filtering, and multi-classification models. For example, the keyword combination "click below" + "to buy" involves advertising and marketing, making it unsuitable for push notifications on news clients and thus considered low-quality text. Text containing vulgar content requires semantic recognition by a model, and identifying such text can be summarized as a text classification problem. Therefore, review of manually agreed-upon quality modalities can be viewed as a multi-classification recognition problem. Its goal is to categorize text according to pre-set rules and to process identified low-quality text offline, specifically prohibiting its publication. Low-quality text is considered to be pending review and has failed the first-level review process.

[0082] The second review rule (review of the inherent quality modality of the text) mainly focuses on examining whether the text has text quality problems (intrinsic quality problems of the text itself) from a linguistic perspective, such as typos, missing punctuation, redundant sentences, logical confusion, and serious inconsistencies between the title and the text content. This review process can be abstracted as a multi-classification problem, but multi-classification problems require in-depth semantic understanding and logical reasoning. Therefore, the difficulty of the second review rule is greater than that of the first review rule.

[0083] In the scenario where the first review rule and the second review rule are completed with the help of a model, the complexity of the model used by the first review rule is less than the complexity of the model used by the second review rule, but the efficiency of the model used by the first review rule is greater than the efficiency of the model used by the second review rule. Therefore, a feasible method of this embodiment may be to first use the first review rule for review and then use the second review rule for review. The process may be: use the first review rule to review whether the text to be reviewed contains sensitive content; if the text to be reviewed contains sensitive content, determine that the text to be reviewed has not passed the first-level review, and prohibit the release of the text to be reviewed; if the text to be reviewed does not contain sensitive content, determine that the text to be reviewed has passed the first-level review, and use the second review rule to review whether the text to be reviewed has text quality problems from a linguistic perspective; if the text to be reviewed does not have text quality problems, determine that the text to be reviewed has passed the second-level review.

[0084] Because the multi-classification problem of the second review rule requires in-depth semantic understanding and logical reasoning, a preset large language model (LLM) can be used in the secondary review. The preset LLM can capture deeper semantic associations and structural defects in the text, thereby realizing the review of whether the text to be reviewed has text quality problems from a linguistic perspective. Text quality problems can be recorded in the output results of the preset LLM. If the text to be reviewed has text quality problems, it is determined that the text to be reviewed has not passed the secondary review, and the release of the text to be reviewed is prohibited, that is, the text to be reviewed is taken offline.

[0085] Since the first-level review has filtered out a large amount of text containing sensitive content, the amount of text entering the second-level review is relatively reduced, making it possible to use the preset LLM with relatively high computing resource consumption in the second-level review stage, thereby significantly improving the review precision and accuracy of text quality issues while ensuring efficiency.

[0086] In one possible implementation, using the second review rule to review whether the text to be reviewed has text quality issues from a linguistic perspective includes:

[0087] A preset large language model is used to obtain the review result of the text to be reviewed in the review direction of the second review rule. The review direction of the second review rule includes at least one of whether there are grammatical errors, whether it is readable, and whether the title is ambiguous. If the review result indicates that the text to be reviewed does not have problems indicated by the review direction, it is determined that the text to be reviewed has no text quality problems. If the review result indicates that the text to be reviewed has problems indicated by the review direction, it is determined that the text to be reviewed has text quality problems.

[0088] Using the preset large language model to obtain the audit result of the text to be audited in the audit direction of the second audit rule includes: inputting the text to be audited into the preset large language model, and obtaining the audit result of the text to be audited output by the preset large language model in the audit direction of the second audit rule. The preset large language model is obtained by fine-tuning the model parameters using the fine-tuning data set of the audit direction of the second audit rule. The fine-tuning data set of the audit direction of the second audit rule includes historical text to be audited, the audit result of the historical text to be audited output by the preset large language model, and the historical text manually annotated according to the audit direction of the second audit rule. Alternatively, using the preset large language model to obtain the audit result of the text to be audited in the audit direction of the second audit rule includes: inputting the text to be audited and the preset prompt information into the preset large language model, and obtaining the output result of the preset large language model. The preset prompt information is used to indicate that the text to be audited is to be audited from the audit direction of the second audit rule, and the preset prompt information also indicates the format of the output result of the preset large language model. The output result records the audit result in the audit direction of the second audit rule. It can be seen that the preset LLM mainly divides the audit method of auditing the text to be audited from the linguistic perspective into the following two ways:

[0089] The first audit method, based on direct reasoning based on preset prompts, utilizes a preset LLM, guiding it through a designed preset prompt to generate output results in a specific format. The preset prompt aims to transform complex audit tasks into instructions that the model can understand and execute, thereby stimulating the preset LLM's contextual understanding, logical reasoning, and knowledge generation capabilities. The implementation process combines the text to be audited with the preset prompt as input to the preset LLM. The preset LLM then outputs the audit results in the audit direction based on the instructions in the preset prompt. The audit results can be in the form of key-value pairs, where the key is the audit direction and the value is the key value. The meaning of the key value is searched in a dictionary table to indicate whether the text has issues indicated by the audit direction.

[0090] From the perspective of mathematical abstraction: the preset prompt can be regarded as a method to constrain and guide the preset LLM. and an audit task , construct a Prompt The default LLM audit process can be abstracted into a function ,in The output of the preset LLM is generated. The design goal of the preset Prompt is to maximize , that is, given the audit text and audit tasks, the preset LLM generates audit results that meet the preset prompt expectations. Audit tasks are used to indicate the audit directions.

[0091] An example of a preset prompt is as follows:

[0092] {

[0093] "Task": "Determine whether the news hits the following three categories of questions and output the corresponding label JSON (each item is 0 or 1).",

[0094] "Input format": { "Title": " <title> ", "content": "< / title> <content>"},

[0095] "Output format": " <json> (\"Grammatical Errors\":0, \"Lack of Readability\":0, \"Title Ambiguity\":0)< / json> ",

[0096] "Judgment criteria": {

[0097] "Grammar Errors": "The content contains typos, missing words, or incoherent sentences."

[0098] "Not very readable": [

[0100] "1. The narrative is confusing or repetitive",

[0101] "2. The topic is forcibly expanded and lacks clear information."

[0102] "3. Content is missing or abruptly ends",

[0103] 4. Inconsistencies and loopholes

[0104] ],

[0105] "Title Ambiguity": [

[0107] "1. The text does not mention the title information at all (except metaphorically)",

[0108] 2. The title and the main text have opposite views. ]

[0110] },

[0111] "Notes": {

[0112] "Low readability": [ "The main text is a newsletter or only contains a small amount of text + pictures / videos → not a hit" ],

[0113] "Title Ambiguity": [ "Only pictures / videos → no hits", "Financial News / Product Promotion → no hits",

[0114] "Title contains metaphor / innuendo → miss" ]},

[0115] "Output requirements": "Output only one line <json>The JSON result wrapped in the tag does not contain any explanation.

[0116] }

[0117] With the help of the above preset prompt, the preset LLM can review the text to be reviewed from three review directions: whether there are grammatical errors, poor readability, and ambiguous titles. If the output result has problems indicated by any of the review directions, it is determined that the text to be reviewed has text quality problems.

[0118] The second auditing method combines LoRA (Low-Rank Adaptation) fine-tuning. This approach addresses the issues of high cost or unstable output formats in existing models. This embodiment introduces LoRA fine-tuning technology. LoRA fine-tuning injects a small amount of trainable low-rank matrices into the existing model, enabling efficient domain-adaptive fine-tuning while maintaining the original model parameters. This allows the existing model to better audit text from a linguistic perspective. This LoRA fine-tuning technique significantly reduces computational resources and storage overhead while effectively improving the model's performance and output stability in auditing text quality issues. Specifically, a fine-tuning dataset is constructed for the existing model. The fine-tuning dataset can include historical pending audit text, audit results of historical pending audit text output using a preset LLM, and historical text manually annotated according to the audit direction. Then, LoRA fine-tuning is used to fine-tune specific layers of the existing model, enabling it to learn to audit text from the audit direction and maintain a fixed output format. The fine-tuned existing model maintains high audit accuracy while effectively reducing computational costs and fixing the output format. The existing model can be a pre-set LLM with linguistic analysis, and the fine-tuned existing model is used as the pre-set LLM in the secondary review. LoRA fine-tuning technology is usually applied to the linear layers in the model, especially the following key layers:

[0119] 1. The Q / K / V / O matrix of the attention layer (the core part of the self-attention mechanism); 2. The intermediate layer of the feedforward network (FFN); 3. The projection matrix of the output layer (such as the classification head). These key layers are the specific layers involved in LoRA fine-tuning.

[0120] From a mathematical point of view: Assume that the weight matrix of the existing model is The LoRA fine-tuning technology introduces two low-rank matrices and (in ), the weight update of the existing model is expressed as During the fine-tuning process, only and trained, and This significantly reduces the number of fine-tuning parameters, thereby reducing computational and storage costs. It is usually defined as the difference between the labels output by the model and the true labels, and is minimized by optimization algorithms such as gradient descent. To update and .

[0121] In practical applications, two review methods, direct reasoning based on preset prompts and fine-tuning with LoRA (Low-Rank Adaptation), are selected for secondary review based on their advantages and applicable scenarios. For example, the advantages of direct reasoning based on preset prompts include: zero training cost: no additional training is required, and existing LLMs (such as GPT-4 and Claude) can be directly called upon, making it suitable for rapid prototyping; high flexibility: the prompt can be adjusted to accommodate different review directions; applicability to black-box models: suitable for closed-source models that cannot be fine-tuned (such as OpenAI's GPT series), and applicable scenarios include: short-term / experimental tasks: such as creative writing and code generation; dynamic adjustment of multiple tasks: such as customer service systems, where prompts can be adjusted to adapt to different user queries.

[0122] The advantages of fine-tuning applications combined with LoRA (Low-Rank Adaptation) include: efficient parameters: only a small number of low-rank matrices (such as 0.1% of the parameters) need to be fine-tuned, which greatly reduces computing and storage costs. For example, full fine-tuning of LLaMA-7B requires 60GB of video memory, while LoRA only requires 23GB of video memory; stronger task performance: compared with Prompt, the fine-tuned model performs better in specific scenarios (such as medical question-answering, legal analysis, and news text review). Applicable scenarios include: vertical field tasks: such as medical diagnosis, financial analysis, legal contract parsing, and other scenarios that require professional knowledge; production-level applications: scenarios that require stable and controllable output (such as automatic customer service and code generation tools).

[0123] One thing needs to be explained here: the text classification model can be used to identify whether there are text quality problems in the text to be reviewed, but the recognition effect of the text classification model is poor. Although manual review can identify complex text quality problems, due to its high cost, limited manual workload, and susceptibility to factors such as visual fatigue, it leads to the widespread phenomenon of missed reviews of text quality problems. For example, a large number of typos have not been effectively discovered and corrected. This embodiment uses the preset LLM to capture deeper semantic associations and structural defects in the text, which can effectively and accurately identify text quality problems, improve the accuracy and precision of the review, and reduce computing costs.

[0124] S102. After the text to be reviewed passes the secondary review, the text to be reviewed is released.

[0125] S103: After the text to be reviewed is published, multiple user feedback data of the text to be reviewed are obtained.

[0126] In this embodiment, user feedback data is feedback data given by users after reading the text to be reviewed. User feedback data may include user comment data, likes, dislikes, shares, and positive and negative opinions at the semantic level, etc. User feedback data is used to indicate whether the user makes a positive or negative evaluation of the text to be reviewed.

[0127] S104. Conduct a three-level review of the text to be reviewed based on multiple user feedback data.

[0128] S105. If the text to be reviewed passes the third-level review, continue to publish the text to be reviewed.

[0129] S106. If the text to be reviewed fails to pass the third-level review, the release of the text to be reviewed will be stopped.

[0130] Among them, the three-level review can be regarded as an audit of the external evaluation quality modality. Its purpose is to determine whether the user has made a negative evaluation of the text to be reviewed, and the distribution of the negative evaluation among all the evaluations when making a negative evaluation, so as to determine whether the text to be reviewed has passed the three-level review based on the distribution of negative evaluations among all the evaluations. For example, if the proportion of negative evaluations in all the evaluations is greater than the preset proportion, or the number of negative evaluations is greater than the preset threshold, it is determined that the text to be reviewed has not passed the three-level review. Otherwise, it is determined that the text to be reviewed has passed the three-level review. Through the three-level review, user feedback data is introduced into the review processing flow, so that the review processing flow forms a closed loop, and the retrospective judgment and continuous monitoring of the text after release are realized.

[0131] For the pending text that has passed the third-level review, the pending text can continue to be published. If the pending text has not passed the third-level review, the publishing of the pending text will be stopped, that is, the pending text will be taken offline.

[0132] In one possible implementation, the process of performing a three-level review of the text to be reviewed based on multiple user feedback data includes: identifying negative comment data in the multiple user feedback data; and performing a three-level review of the text to be reviewed through a preset big model when the distribution of the negative comment data in the multiple user feedback data meets the big model evaluation conditions.

[0133] For each piece of user feedback data, sentiment and intent analysis is performed on the user feedback data to determine the evaluation category of the user feedback data, such as whether it is a positive evaluation category or a negative evaluation category. If the evaluation category of the user feedback data is a negative evaluation category, then it is determined that the user feedback data is negative comment data.

[0134] One feasible approach is: generating a feature vector for each piece of user feedback data, such as using a preset vector generation model to generate a feature vector for each piece of user feedback data; inputting the feature vector of the user feedback data into a comment tendency classification model to obtain the category to which the user feedback data output by the comment tendency classification model belongs. The comment tendency classification model is trained based on the feature vectors of historical user feedback data and the categories to which the historical user feedback data belongs. The comment tendency classification model can identify multiple comment tendency categories, and the target comment tendency category among the multiple comment tendency categories is a negative comment category; if the category to which the user feedback data belongs is the target comment tendency category, then the user feedback data is determined to be negative comment data.

[0135] The review tendency classification model can adopt at least one of the following sentiment analysis models: dictionary-based sentiment analysis method, machine learning classifier and deep learning model. For the question of whether the user feedback data is negative review data, the question can be modeled as a multi-classification task. Assume that the text to be reviewed is , and its corresponding comment tendency category is i is the total number of comment tendency categories. The goal of the comment tendency classification model is to learn a mapping function For example, the review tendency classification model can use the Softmax regression model for multi-classification:

[0136]

[0137] in, Indicates text awaiting review Belong to category The probability of It is a text awaiting review The eigenvector of and They are categories The weight vector and bias term of , T represents the matrix transpose operation. If the category to which the user feedback data belongs is the target review tendency category, the user feedback data is determined to be negative review data.

[0138] Taking news text as an example, we categorize comments based on their quality, author's perspective, and the people or events in the news. The results are shown in Table 1. There are nine categories for news text comment tendency, of which only category C1 is considered negative. If user feedback data falls into category C1, it is determined to be negative. The distribution of negative comments across multiple pieces of user feedback data is then used to determine whether to perform a three-level review on the text under review using a pre-set large model.

[0139] Table 1 Classification results of comment tendency categories of news texts

[0140] News quality Author's opinion People or things in the news front A1 A2 A3 neutral B1 B2 B3 Negative C1 C2 C3

[0141] In some examples, when the distribution of negative review data in multiple user feedback data meets the following conditions, a three-level review is performed on the text to be reviewed using a preset large model. The conditions include but are not limited to the following conditions:

[0142] When the number of negative comment data is not less than the first threshold, and the proportion of negative comment data in multiple user feedback data is not less than the first preset proportion but less than the second preset proportion, the text to be reviewed is subject to a three-level review using the preset large model, and the second preset proportion is greater than the first preset proportion.

[0143] When the proportion of negative comment data in multiple user feedback data is not less than the second preset proportion, and the number of negative comment data is not less than the second threshold but less than the first threshold, the text to be reviewed is subject to a three-level review using the preset large model, and the first threshold is greater than the second threshold.

[0144] Correspondingly, when the number of negative comment data is not less than the first threshold and the proportion of negative comment data in multiple user feedback data is not less than the second preset proportion, it is determined that the text to be reviewed has not passed the third-level review; when the number of negative comment data is less than the first threshold and the proportion of negative comment data in multiple user feedback data is not less than the first preset proportion but less than the second preset proportion, it is determined that the text to be reviewed has passed the third-level review.

[0145] Among them, not less than can be greater than or equal to, and not greater than can be less than or equal to. The values ​​of the first threshold, the second threshold, the first preset proportion, and the second preset proportion are not limited in this embodiment.

[0146] From the above process of performing three-level review based on the distribution of negative review data, it can be seen that this embodiment proposes a "three-stage method" activation rule. This "three-stage method" activation rule divides the degree of negative evaluation of a text to be reviewed (determined by the negative evaluation data) into different levels and sets corresponding activation thresholds. For example, the "three-stage method" activation rule includes:

[0147] Weak Rule 1: If the number of negative review data is greater than the first threshold, the proportion of the number of negative review data in the multiple user feedback data can be ignored, and the first preset proportion dimension in Weak Rule 1 is 0%;

[0148] Weak rule 2: the number of negative review data is greater than the second threshold and the proportion of negative review data in the multiple user feedback data is greater than the second preset proportion;

[0149] Strong rule: the number of negative comment data is greater than the first threshold and the proportion of negative comment data in the multiple user feedback data is greater than the second preset proportion.

[0150] As shown in Table 2, the "three-paragraph method" activation rules for news text are set by setting the threshold of the number and proportion of negative comment data.

[0151] Table 2 Activation rules of the "three-paragraph method" of news text

[0152] rule condition Weak Rule 1 Negative comments data reaches 10 Weak Rule 2 The number of negative comments reaches 3 and accounts for more than 30% Strong rules The number of negative comments reaches 10 and accounts for more than 30%

[0153] Weak rule 1, weak rule 2, and strong rule are used to control whether the text to be reviewed undergoes three-level review. The process of controlling the text to be reviewed by weak rule 1, weak rule 2, and strong rule is as follows:

[0154] When the negative comment data of a text to be reviewed reaches a certain threshold, such as the number of negative comment data is greater than 10, and the proportion of negative comment data in multiple user feedback data is greater than or equal to 0 but less than 30%, the text to be reviewed will be marked as a weak rule hit, or, when the proportion of negative comment data in multiple user feedback data is greater than or equal to 30%, and the number of negative comment data is greater than or equal to 3 but less than 10, the text to be reviewed can be marked as a weak rule hit. After the text to be reviewed is marked as a weak rule hit, the preset large model is called to perform a three-level review on the text to be reviewed. Specifically, the preset large model is called to identify the evaluation category to which the multiple user feedback data of the text to be reviewed belong. If the preset large model is called to identify that the evaluation category to which the preset number of user feedback data belongs is the user negative evaluation category, it is determined that the text to be reviewed has not passed the three-level review. For relevant instructions on the preset large model, please refer to the preset LLM in the second-level review, which will not be elaborated here.

[0155] When the negative comment data of a text to be reviewed reaches a higher threshold, for example, the number of negative comment data is greater than or equal to 10, and the proportion of negative comment data in multiple user feedback data is greater than or equal to 30%, then the text to be reviewed can be marked as a strong rule hit. In theory, the text to be reviewed that hits the strong rule will be directly judged as failing the third-level review, without the need to call the preset large model for the third-level review again, thereby achieving a quick response.

[0156] When the number of negative comments on a text to be reviewed is less than a certain threshold, such as the number of negative comments is less than , and the proportion of negative comment data in multiple user feedback data is greater than or equal to 0% but less than 30%, the text to be reviewed is directly determined to have passed the third-level review, and there is no need to call the preset large model for the third-level review again, thereby achieving a quick response.

[0157] Furthermore, if a predetermined number of user feedback items from among the multiple user feedback items involve specific event content, it is determined that the pending text has not passed the third-level review, thereby enabling a quick response. For example, the specific event content may be a recent sensitive event or malicious comments, which may affect the platform's atmosphere. The value of the predetermined number is not limited in this embodiment.

[0158] The mathematical expression of the activation rule can be defined as:

[0159]

[0160] in, Indicates activation action, ActivateLLM means calling the preset large model, It means it has not passed the third level review. Indicates that it has passed the third-level review. Indicates the number or proportion of negative review data. and Respectively represent the activation thresholds for weak rule hits and strong rule hits. For details, see the above process of weak rule 1, weak rule 2, and strong rule controlling the text to be reviewed. A Boolean variable indicating that a specific event content has been detected. That is, for any text to be reviewed, the review will first be conducted on the review tendency, and then the relationship between the review tendency review result and the "three-stage method" activation rule will be used to decide whether to call the preset large model for three-level review, such as Figure 2 shown.

[0161] Through this "three-stage method" activation rule, this embodiment can intelligently select the review strategy according to the severity of the negative comments. Specifically, the "three-stage method" activation rule is used to determine under what conditions to call the preset large model, under what conditions to directly determine that the third-level review has not been passed, and under what conditions to directly determine the conditions for passing the third-level review. This not only ensures the comprehensiveness of the review, but also optimizes resource utilization efficiency.

[0162] In summary, the text audit method provided in the embodiment of the present application divides the audit tasks of the text to be audited into three main levels: first-level audit, second-level audit and third-level audit. The first-level audit utilizes one of the first audit rule and the second audit rule, the second-level audit utilizes the other of the first audit rule and the second audit rule, and the audit directions of the first audit rule and the second audit rule are different. The third-level audit utilizes multiple user feedback data of the text to be audited, thereby forming a hierarchical serial audit processing flow, so that the various audit methods in the audit processing flow can be effectively coordinated, and the audit results of different audit methods are used to make a comprehensive judgment on the text quality of the text to be audited, thereby improving accuracy and precision. In the hierarchical serial audit processing flow, the number of texts to be audited for the second-level and third-level audits is reduced, which effectively reduces the computational cost of the second-level and third-level audits, avoids the high-cost comprehensive analysis of all texts to be audited, thereby avoiding a large amount of intermediate storage and computing costs, ensuring the reasonable allocation of computing resources and storage resources to reduce the repeated consumption of computing resources and storage resources, and improving operational efficiency and scalability.

[0163] In addition, user feedback data is introduced into the third-level review to form a closed loop for the review of the text to be reviewed. User feedback data is used to evaluate whether there are hidden quality problems in the text to be reviewed, effectively filling in the review blind spots that may exist in the first-level and second-level reviews, and realizing retrospective judgment and continuous monitoring of the text after release, further improving the efficiency, accuracy and precision of text review.

[0164] See Figure 3 , which shows another optional process of a text review method provided by an embodiment of the present application, Figure 3 The text review method shown may include setting a manual interaction interface in at least one of the first-level review, the second-level review, and the third-level review, and adjusting the review direction and / or reviewing the review results through the manual interaction interface. Specifically, the following steps may be included:

[0165] S101. Perform a first-level review on the text to be reviewed using one of the first review rule and the second review rule. After the text to be reviewed passes the first-level review, perform a second-level review on the text to be reviewed using the other of the first review rule and the second review rule.

[0166] S102. After the text to be reviewed passes the secondary review, the text to be reviewed is released.

[0167] S103: After the text to be reviewed is published, multiple user feedback data of the text to be reviewed are obtained.

[0168] In this embodiment, user feedback data is feedback data given by users after reading the text to be reviewed. User feedback data may include user comment data, likes, dislikes, shares, and positive and negative opinions at the semantic level, etc. User feedback data is used to indicate whether the user makes a positive or negative evaluation of the text to be reviewed.

[0169] S104. Conduct a three-level review of the text to be reviewed based on multiple user feedback data.

[0170] S105. If the text to be reviewed passes the third-level review, continue to publish the text to be reviewed.

[0171] S106. If the text to be reviewed fails to pass the third-level review, the release of the text to be reviewed will be stopped.

[0172] S107. Adjust the audit direction of the first audit rule through the manual interaction interface of the first-level audit. When the audit standards or audit requirements change, the audit direction can be dynamically added, modified, or deleted through the manual interaction interface of the first-level audit. For example, keywords, regular expressions, classification models, etc. can be dynamically added, modified, or deleted through the manual interaction interface of the first-level audit. This ensures that the audit direction is consistent with the audit standards or audit requirements through the manual interaction interface of the first-level audit, and that the new audit direction can be quickly used for the first-level audit.

[0173] During the first-level review process, different review intensities and standards may be required during different control periods or in response to emergencies. For example, for political content, there may be two or more sets of sensitive word libraries (e.g., two sets of sensitive word libraries for "loose mode" and "strict mode"). The manual interface for the first-level review allows reviewers to switch between sensitive word libraries in real time based on actual needs, thereby dynamically adjusting the level of review strictness and sensitive word libraries. This allows for macro-control of the review direction through the manual interface, enabling flexible response to changes in the external environment and meeting the review intensity and standards at different times.

[0174] S108. Adjust the audit direction of the second audit rule through the manual interaction interface of the secondary audit. In the secondary audit, the audit direction can be adjusted through the manual interaction interface. The audit direction indicates that the secondary audit process will be conducted based on text defects such as grammatical errors, grammar errors, and typos. Then, at least one of the grammatical errors, grammar errors, and typos can be adjusted through the manual interaction interface. For example, in the secondary audit, the audit scope and audit granularity of text quality issues can be expanded or contracted through the manual interaction interface of the secondary audit. The audit scope can be the part of the text to be audited, and the audit granularity corresponds to the text curve indicated by the audit direction. For example, the audit of a new type of grammatical error can be increased, or some unimportant text defects can be temporarily ignored, thereby achieving dynamic adjustment of the audit direction. In actual application, the preset prompt can be adjusted through the manual interaction interface of the secondary audit.

[0175] S109. Review the audit results of the secondary audit and provide feedback on the review results through the manual interaction interface of the secondary audit.

[0176] In this embodiment, the secondary review results of at least one document to be reviewed are sampled through the secondary review manual interface. A technician then reviews the secondary review results. The review results are fed back through the secondary review manual interface. If the review results are inconsistent with the secondary review results, the review results can be used to correct the secondary review process, such as by adjusting the preset large language model or preset prompt. By providing feedback data to the preset LLM through the secondary review manual interface, the accuracy of the audit results can be ensured by promptly identifying and correcting any misjudgments or omissions in the model while monitoring the operational accuracy and stability of the preset LLM.

[0177] S110. Review the results of the third-level review and provide feedback through the third-level review manual interface. The third-level review manual interface is similar to the second-level review manual interface. In addition, the third-level review manual interface is also used to manually review disputed documents. Disputed documents can include documents with a large amount of user feedback data or documents that record emergencies.

[0178] As can be seen from the above technical solutions, the text review method provided in this embodiment ensures automatic review while introducing a manual interaction interface to make the text review process highly flexible, controllable, and adaptable to dynamic changes. Specifically, the manual interaction interface allows technical personnel to intervene and adjust at key nodes to achieve optimized review through human-computer collaboration. Key nodes can be review rules (such as weak rules and strong rules), classification standards (such as comment tendency categories), or configuration files (such as sensitive word libraries), allowing technical personnel to adjust key nodes in real time according to review standards or emergencies, ensuring that key nodes can keep pace with review standards or external environments, and effectively respond to the constantly changing texts to be reviewed. Therefore, through the manual interaction interface, when the review changes at any level in the review process, the review process can be locally adjusted instead of globally modified, which greatly improves the maintainability and scalability of the review process.

[0179] Figure 4 The following is a schematic diagram of a text review method provided by an embodiment of the present application. Specifically, for each text to be reviewed, it will go through the following processing stages in sequence:

[0180] Level 1 audit (audit of manually agreed quality modalities):

[0181] Purpose: This system mainly uses manually preset rules to quickly and comprehensively screen sensitive content, such as advertising and marketing, politically sensitive content, and vulgar information. These modalities usually have clear characteristics and are suitable for processing using efficient rule matching and classification models.

[0182] Method: In this stage, the text to be reviewed is reviewed using at least one of the quality rule base, quality list (blacklist and whitelist), and multi-classification quality model. The quality rule base, quality list (blacklist and whitelist), and multi-classification quality model are used to identify sensitive content. If the text to be reviewed contains sensitive content, it will be immediately taken offline. Figure 4 After reviewing the text to be reviewed using the quality rule library, quality list (black and white list) and at least one of the multi-classification quality models, it is determined whether it is a hit. If it is a hit, it means that the text to be reviewed includes sensitive content, and it will be immediately taken offline. If it is not a hit, it means that the text to be reviewed does not include sensitive content, and a secondary review can be performed on the text to be reviewed.

[0183] Features: This phase processes all pending documents and requires extremely high processing efficiency to cope with the massive daily influx of new documents. This phase of review effectively filters out most pending documents that haven't passed the first-level review, reducing the burden on subsequent reviews.

[0184] Secondary review (review of the inherent quality modality of the text):

[0185] Purpose: For qualified texts that have passed the first-level review (i.e., texts awaiting review that have passed the first-level review), further identify whether there are any text quality issues from a linguistic perspective, such as grammatical errors, illogicality, poor readability, and clickbait titles.

[0186] Methodology: This phase primarily utilizes a pre-defined LLM to conduct in-depth semantic analysis and quality assessment of the texts under review. This pre-defined LLM can capture deeper semantic connections and structural flaws within the text, identifying text quality issues that are difficult to detect with traditional models. Texts under review with text quality issues (i.e., those that are considered for review) will be removed from the review process.

[0187] Features: Because the first-level review has filtered out a large amount of text, the amount of text entering the second-level review is relatively reduced, making it possible to use the preset LLM, which consumes relatively high computing resources, at this stage. This significantly improves the review accuracy of text quality issues while ensuring efficiency.

[0188] Level 3 audit (external evaluation quality mode audit or bottom-line audit):

[0189] Purpose: To conduct a retrospective review of pending texts that have passed the first and second level reviews and have been granted the opportunity to be published, using user feedback data such as user comments, to identify hidden quality issues.

[0190] Methodology: Approved texts that have generated user feedback will enter this phase. Combined with this user feedback data, this phase conducts an in-depth analysis using sentiment analysis and pre-defined large-scale models to determine whether the app has any hidden quality issues revealed by negative user feedback. If any such issues are identified, the app will be taken offline.

[0191] Features: This is a post-verification phase. By introducing real feedback from users (i.e., user feedback data), it forms a closed loop for news quality review, effectively filling in the blind spots that may exist in the first two stages and ensuring continuous monitoring of the text to be reviewed.

[0192] Human interaction interface:

[0193] Human interaction interfaces are implemented at each stage of the aforementioned first-level, second-level, and third-level review processes. These interfaces allow for temporary adjustments to review rules, sensitive word lists, and other aspects, as well as direct manual review and intervention of review results. This design significantly enhances the flexibility, controllability, and responsiveness of the review process, achieving optimized review through human-machine collaboration.

[0194] Through the above-mentioned process system combining hierarchical serial processing with manual interaction, this application can conduct an orderly, efficient and accurate review of the text to be reviewed.

[0195] In the above embodiment, text classification technology is widely used in the multi-classification quality model of the first-level review and the bottom-line review link to identify sensitive content such as advertising and marketing, content ecology, and political supervision. Text classification technology can be used in the category identification of comment tendency in the bottom-line review link. The main ones used are at least one sentiment analysis model among the dictionary-based sentiment analysis method, machine learning classifier, and deep learning model. In order to improve scalability, reusability and maintainability, this embodiment can adopt a set of single-category classification templates, the core of which covers three modules: AC automaton, easy_pattern and text classifier, such as Figure 5 The introduction of AC automaton, easy_pattern and text classifier is as follows:

[0196] The AC Automaton (Aho-Corasick Automaton) is an efficient multi-pattern string matching algorithm that can simultaneously search for multiple preset pattern strings in a single pass through the text to be reviewed. Its core construction and matching process includes the following three key steps:

[0197] Trie tree construction: First, all keywords (patterns) to be matched are constructed into a Trie tree (prefix tree). Each node in the Trie tree represents a string prefix, and the path from the root node to any node constitutes a pattern string or its prefix. Pattern strings with the same prefix share a common parent node sequence in the tree. The trie tree construction process can be expressed as follows:

[0198]

[0199] in, Represents the constructed Trie tree, express A set of pattern strings to be matched.

[0200] Construction of Fail pointer: Fail pointer (failure pointer) is the core of AC automaton. For each node in the Trie tree , whose Fail pointer points to another node , so that from the root node to The string represented is "from the root node to The longest true suffix of the string represented by the pattern, which is also a prefix of a pattern string. The construction of the Fail pointer is usually implemented using breadth-first search (BFS) to ensure that when a match fails, it can quickly jump to the next possible matching position to avoid redundant searches from the beginning. The construction of the Fail pointer can be formalized as follows:

[0201]

[0202] in, Represents the path from the root node to the node String, express The longest true suffix of a pattern that is a prefix of a pattern.

[0203] Text traversal and matching: After constructing the Trie tree and the Fail pointer, the AC automaton only needs to perform a linear scan of the input text to detect all occurrences of the pattern string. When the current character fails to match, it jumps through the Fail pointer and continues to try to match. The matching process can be described as follows:

[0204]

[0205] in, For texts awaiting review, A set of all detected matching results.

[0206] AC automaton demonstrates excellent performance in processing large-scale keyword matching scenarios, and is particularly suitable for the rapid identification of massive sensitive words and rules in first-level review.

[0207] In practical engineering applications, in addition to direct keyword matching, regular expressions are also needed to process more complex text rules, such as identifying specific phrases in news headlines. However, traditional regular expressions are complex to write, have poor reusability, and are difficult to maintain. To this end, this embodiment uses easy_pattern to simplify the use of regular expressions in text review and improve their maintainability and scalability.

[0208] easy_pattern implements modular management of regular expressions by introducing a separation between component files and rule files. Its core concept is to define commonly used regular expression fragments as reusable tags and combine them through a reference mechanism.

[0209] Components file: This file stores a series of tags and their corresponding regular expressions. A tag can correspond to one or more regular expressions, and a tag can reference the regular expressions corresponding to other tags. For example, tags can reference each other using quotation marks (for example, "Number"). This avoids rewriting the same regular expression fragments, greatly improving reusability and maintainability. For example, a regular expression representing a number can be defined as the tag "Number", and the tag "Number" can be used to reference the regular expression representing the number in multiple rules that need to match numbers.

[0210] The following is an example of a Components file. Education, Society, and Mathematics are three tags in the Components file. The regular expression corresponding to each tag is written under that tag. For example, the regular expression for Education is: elementary school | students | university | teachers and students | campus | grade | classmates | class suspension | class resumption | college entrance examination | high school entrance examination. Under the Society tag, the tag "Number" is used to reference the regular expression corresponding to the Mathematics tag under the Society tag.

[0211] educate:

[0212] - Primary school | students | university | teachers and students | campus | grade | classmates | class suspension | class resumption | college entrance examination | high school entrance examination

[0213] number:

[0214] - (?:[0-9]+|[00123456789010 ...

[0215] Rules file: This file defines classification categories and their corresponding regular expressions. Regular expressions corresponding to classification categories can directly use the tags defined in the Components file, or you can define custom multi-line expressions. This structure makes rule definitions clearer, easier to understand, and easier to modify. The following is an example of a Rules file. The categories "Funny," "Sports," "Military," "Society," and "Other" are defined in the Rules file. Each category contains at least one tag from the Components file, indicating that the regular expression associated with that tag should be used to classify the text to be reviewed, thereby determining the category to which the text belongs.

[0216] funny:

[0217] - "funny"

[0218] physical education:

[0219] - "physical education"

[0220] military:

[0221] - Military

[0222] society:

[0223] - "society"

[0224] other:

[0225] - Workplace | Education

[0226] The design concept of easy_pattern can be summarized as follows:

[0227]

[0228] in, Indicates the classification categories, Combinatorial functions that represent rules (e.g., logical AND, logical OR), Indicates the first reference from the Components file tags, Expressed as A custom regular expression fragment is combined with the regular expressions involved in the combination function according to its logical combination method. If the combination function uses logical OR, the regular expressions are logically ORed. The text to be reviewed is then identified based on the combined regular expression to obtain its classification category. This modularization and reference mechanism significantly improves the management efficiency and responsiveness of regular expressions in complex review scenarios.

[0229] In the text classifier module, this embodiment allows for flexible selection of two mainstream models, FastText and BERT, based on varying semantic understanding depth requirements. These two models each have their own advantages in text classification tasks, and can be switched based on specific scenarios to achieve the optimal balance between performance and accuracy. For example, if the volume of text to be reviewed is large but accuracy requirements are low, FastText can be used; if the volume of text to be reviewed is small but accuracy requirements are high, BERT can be used.

[0230] FastText: FastText is an open-source, efficient text classification tool developed by Facebook. Its core concept is to represent text as the average of word vectors and then classify it using a shallow neural network. FastText's advantages lie in its fast training speed and its ability to achieve accuracy comparable to deep networks when processing large corpora, while also consuming less computational resources. Its classification model can be simplified as follows:

[0231]

[0232] in, Represents a given text Belong to category The probability of is the weight matrix of the classifier, is the number of words in the text, It is FastText achieves efficient text representation and classification by directly inputting word vectors and N-gram features into a linear classifier.

[0233] BERT (Bidirectional Encoder Representations from Transformers): BERT is a pre-trained language model based on the Transformer encoder proposed by Google in 2018. Through bidirectional contextual pre-training and fine-tuning mechanisms, it has achieved significant performance improvements in multiple natural language processing tasks. BERT's advantage lies in its strong contextual understanding ability and versatility, which can capture deep semantic information in text and complex relationships between words. When encountering difficult tasks that require more refined semantic understanding, you can switch from FastText to BERT to improve prediction accuracy. The core of BERT lies in its multi-layer Transformer encoder, which encodes the input sequence through a self-attention mechanism:

[0234]

[0235] in, are query, key, and value matrices respectively, is the dimension of the key vector. BERT achieves deep semantic representation of text by stacking multiple such attention layers. In text classification tasks, a classification layer is usually added on top of BERT's output layer for fine-tuning.

[0236] By combining the efficient pattern matching of AC automata, the flexible rule management of easy_pattern, and the advantages of FastText and BERT at different depths of semantic understanding, this embodiment builds a robust and scalable text classification system, providing a solid foundation for the review of pending texts. In practical applications, this embodiment can use at least one of AC automata, easy_pattern, and a text classification model to perform a first-level review of pending texts.

[0237] The above describes a text review method provided by an embodiment of the present application. The following describes a device for executing the above text review method.

[0238] See also Figure 6 , Figure 6 This is a structural diagram of a text review device provided in an embodiment of the present application. Figure 6 As shown, the text review device includes: a first review module 10 , a publishing module 20 , an acquisition module 30 , a second review module 40 and a control module 50 .

[0239] The first audit module 10 is used to perform a first-level audit on the text to be audited by using one of the first audit rule and the second audit rule, and after the text to be audited passes the first-level audit, perform a second-level audit on the text to be audited by using the other audit rule of the first audit rule and the second audit rule. The audit directions of the first audit rule and the second audit rule are different.

[0240] The publishing module 20 is used to publish the text to be reviewed after it passes the secondary review.

[0241] The acquisition module 30 is used to acquire multiple user feedback data of the text to be reviewed after the text to be reviewed is published.

[0242] The second review module 40 is used to perform a three-level review on the text to be reviewed based on multiple user feedback data.

[0243] The control module 50 is used to control the publishing module 20 to continue publishing the text to be reviewed if the text to be reviewed passes the third-level review; if the text to be reviewed does not pass the third-level review, control the publishing module 20 to stop publishing the text to be reviewed.

[0244] In one possible implementation, the second review module 40 is used to identify negative comment data in multiple user feedback data; when the distribution of negative comment data in multiple user feedback data meets the large model evaluation conditions, a three-level review is performed on the text to be reviewed through a preset large model.

[0245] In one possible implementation, the second review module 40 is used to generate a feature vector for each piece of user feedback data; the feature vector of the user feedback data is input into a comment tendency classification model to obtain the category to which the user feedback data output by the comment tendency classification model belongs. The comment tendency classification model is trained based on the feature vectors of historical user feedback data and the categories to which the historical user feedback data belongs. The comment tendency classification model can identify multiple comment tendency categories, and the target comment tendency category among the multiple comment tendency categories is a negative comment category; if the category to which the user feedback data belongs is the target comment tendency category, then the user feedback data is determined to be negative comment data.

[0246] In one possible implementation, the second review module 40 is used to perform a three-level review on the text to be reviewed through a preset large model when the number of negative comment data is not less than a first threshold value and the proportion of negative comment data in multiple user feedback data is not less than a first preset proportion but less than a second preset proportion, and the second preset proportion is greater than the first preset proportion; when the proportion of negative comment data in multiple user feedback data is not less than the second preset proportion and the number of negative comment data is not less than the second threshold value but less than the first threshold value, the text to be reviewed is performed a three-level review through a preset large model, and the first threshold value is greater than the second threshold value.

[0247] In one possible implementation, the second review module 40 is also used to determine that the text to be reviewed has not passed the third-level review when the number of negative comment data is not less than a first threshold and the proportion of negative comment data in multiple user feedback data is not less than a second preset proportion; determine that the text to be reviewed has passed the third-level review when the number of negative comment data is less than the first threshold and the proportion of negative comment data in multiple user feedback data is not less than the first preset proportion but less than the second preset proportion; determine that the text to be reviewed has passed the third-level review when a preset number of user feedback data in multiple user feedback data involve specific event content.

[0248] In a possible implementation, the text review device further includes an adjustment module, the adjustment module being configured to adjust the review direction of the first review rule through a manual interaction interface of the first-level review;

[0249] and / or,

[0250] Adjust the audit direction of the second audit rule through the manual interaction interface of the second audit;

[0251] and / or,

[0252] Review the results of the second-level audit and provide feedback on the review results through the manual interaction interface of the second-level audit;

[0253] and / or,

[0254] Review the audit results of the third-level audit and provide feedback on the review results through the manual interaction interface of the third-level audit.

[0255] In one possible implementation, the first review module 10 is used to use the first review rules to review whether the text to be reviewed contains sensitive content; if the text to be reviewed contains sensitive content, it is determined that the text to be reviewed has not passed the first-level review, and the publishing module 20 is prohibited from publishing the text to be reviewed; if the text to be reviewed does not contain sensitive content, it is determined that the text to be reviewed has passed the first-level review, and the second review rules are used to review whether the text to be reviewed has text quality problems from a linguistic perspective; if the text to be reviewed does not have text quality problems, it is determined that the text to be reviewed has passed the second-level review.

[0256] In one possible implementation, the first review module 10 is used to use a preset large language model to obtain the review result of the text to be reviewed in the review direction of the second review rule. The review direction of the second review rule includes at least one of whether there are grammatical errors, whether it is readable, and whether the title is ambiguous; if the review result indicates that the text to be reviewed does not have the problem indicated by the review direction, it is determined that the text to be reviewed has no text quality problem; if the review result indicates that the text to be reviewed has the problem indicated by the review direction, it is determined that the text to be reviewed has text quality problem.

[0257] Using the preset large language model to obtain the audit result of the text to be audited in the audit direction of the second audit rule includes: inputting the text to be audited into the preset large language model, and obtaining the audit result of the text to be audited output by the preset large language model in the audit direction of the second audit rule, the preset large language model is obtained by fine-tuning the model parameters using the fine-tuning data set of the audit direction of the second audit rule, the fine-tuning data set of the audit direction of the second audit rule includes historical text to be audited, the audit result of the historical text to be audited output by the preset large language model, and the historical text manually annotated according to the audit direction of the second audit rule, or, using the preset large language model to obtain the audit result of the text to be audited in the audit direction of the second audit rule includes: inputting the text to be audited and preset prompt information into the preset large language model, and obtaining the output result of the preset large language model, the preset prompt information is used to indicate that the text to be audited is to be audited from the audit direction of the second audit rule, and the preset prompt information also indicates the format of the output result of the preset large language model, and the output result records the audit result in the audit direction of the second audit rule.

[0258] An electronic device is also provided in an embodiment of the present application. Figure 7 , which shows a schematic diagram of the structure of an electronic device suitable for implementing the embodiments of the present application. The electronic device in the embodiments of the present application may include, but is not limited to, fixed terminals such as mobile phones, laptops, PDAs (personal digital assistants), PADs (tablet computers), desktop computers, etc. Figure 7 The electronic device shown is merely an example and should not limit the functions and scope of use of the embodiments of the present application.

[0259] like Figure 7 As shown, the electronic device may include a processing device (such as a central processing unit, a graphics processing unit, etc.) 701, which can perform various appropriate actions and processes according to a program stored in a read-only memory (ROM) 702 or a program loaded from a storage device 708 into a random access memory (RAM) 703. When the electronic device is powered on, the RAM 703 also stores various programs and data required for the operation of the electronic device. The processing device 701, ROM 702, and RAM 703 are connected to each other via a bus 704. An input / output (I / O) interface 705 is also connected to the bus 704. The processing device 701 can be regarded as a processor in the electronic device. The electronic device may include at least one processor. The ROM 702, RAM 703, and storage device 708 can be regarded as a memory in the electronic device. The processor is connected to the memory, and the memory is used to store computer programs; the processor is used to execute the computer program so that the electronic device can implement any of the text review methods provided in the embodiments of the present application.

[0260] Typically, the following devices may be connected to the I / O interface 705: an input device 706 including, for example, a touch screen, a touchpad, a keyboard, a mouse, a camera, a microphone, an accelerometer, a gyroscope, etc.; an output device 707 including, for example, a liquid crystal display (LCD), a speaker, a vibrator, etc.; a storage device 708 including, for example, a memory card, a hard disk, etc.; and a communication device 709. The communication device 709 may allow the electronic device to communicate with other devices wirelessly or by wire to exchange data. Figure 7 The electronic device is shown with various devices, but it should be understood that it is not required to implement or possess all of the devices shown. More or fewer devices may be implemented or possessed instead.

[0261] An embodiment of the present application also provides a computer program product including computer-readable instructions. When the computer-readable instructions are executed on an electronic device, the electronic device implements any text review method provided in the embodiment of the present application.

[0262] A computer-readable storage medium is also provided in an embodiment of the present application. The storage medium carries one or more computer programs. When the one or more computer programs are executed by an electronic device, the electronic device can implement any text review method provided in the embodiment of the present application.

[0263] It should also be noted that the device embodiments described above are merely illustrative, wherein the units described as separate components may or may not be physically separate, and the components displayed as units may or may not be physical units, that is, they may be located in one place, or they may be distributed across multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the present embodiment. In addition, in the drawings of the device embodiments provided in this application, the connection relationship between the modules indicates that there is a communication connection between them, which can be specifically implemented as one or more communication buses or signal lines.

[0264] Through the description of the above embodiments, those skilled in the art can clearly understand that the present application can be implemented by means of software plus necessary general hardware, and of course can also be implemented by special hardware including application-specific integrated circuits, special CPUs, special memories, special components, etc. In general, all functions performed by computer programs can be easily implemented with corresponding hardware, and the specific hardware structures used to implement the same function can also be diverse, such as analog circuits, digital circuits or special circuits, etc. However, for the present application, software program implementation is a better implementation method in most cases. Based on this understanding, the technical solution of the present application is essentially or the part that contributes to the prior art can be embodied in the form of a software product, which is stored in a readable storage medium, such as a computer's floppy disk, USB flash drive, mobile hard disk, ROM, RAM, magnetic disk or optical disk, etc., and includes a number of instructions to enable a computer device (which can be a personal computer, training equipment, or network equipment, etc.) to execute the methods described in each embodiment of the present application.

[0265] In the above embodiments, all or part of the embodiments may be implemented by software, hardware, firmware, or any combination thereof. When implemented by software, all or part of the embodiments may be implemented in the form of a computer program product.

[0266] The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, the process or function described in the embodiment of the present application is generated in whole or in part. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable devices. The computer instructions can be stored in a computer-readable storage medium, or transmitted from one computer-readable storage medium to another computer-readable storage medium. For example, the computer instructions can be transmitted from a website, a computer, a training device or a data center by wired (e.g., coaxial cable, optical fiber, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) mode to another website, a computer, a training device or a data center. The computer-readable storage medium can be any available medium that a computer can store or a data storage device such as a training device, a data center, etc. that includes one or more available media integrations. The available medium can be a magnetic medium, (e.g., a floppy disk, a hard disk, a tape), an optical medium (e.g., a DVD), or a semiconductor medium (e.g., a solid-state drive (SSD)).< / json> < / content>

Claims

1. A text review method, characterized in that: include: Performing a primary review of a document to be reviewed using one of a first review rule and a second review rule, and after the document to be reviewed passes the primary review, performing a secondary review of the document to be reviewed using the other of the first review rule and the second review rule, wherein the first review rule and the second review rule have different review directions; After the text to be reviewed passes the secondary review, the text to be reviewed is released; After the text to be reviewed is published, obtaining multiple user feedback data of the text to be reviewed; Conducting a three-level review of the text to be reviewed based on the multiple pieces of user feedback data; If the text to be reviewed passes the third-level review, continue to publish the text to be reviewed; if the text to be reviewed does not pass the third-level review, stop publishing the text to be reviewed.

2. The method according to claim 1, characterized in that The performing three-level review on the text to be reviewed based on the plurality of user feedback data includes: Identifying negative review data from the plurality of user feedback data; When the distribution of the negative comment data in the plurality of user feedback data meets the large model evaluation conditions, a three-level review is performed on the text to be reviewed using the preset large model.

3. The method according to claim 2, characterized in that The identifying of negative comment data in the plurality of user feedback data comprises: Generate a feature vector for each piece of user feedback data; Inputting the feature vector of the user feedback data into a comment tendency classification model to obtain the category to which the user feedback data output by the comment tendency classification model belongs, wherein the comment tendency classification model is trained based on the feature vector of historical user feedback data and the category to which the historical user feedback data belongs, and the comment tendency classification model can identify multiple comment tendency categories, wherein the target comment tendency category among the multiple comment tendency categories is a negative comment category; If the category to which the user feedback data belongs is the target review tendency category, it is determined that the user feedback data is negative review data.

4. The method according to claim 2 or 3, characterized in that When the distribution of the negative review data in the plurality of user feedback data meets the large model evaluation conditions, performing a three-level review of the to-be-reviewed text using the preset large model includes: When the number of the negative comment data is not less than a first threshold, and the proportion of the negative comment data in the plurality of user feedback data is not less than a first preset proportion but less than a second preset proportion, performing a three-level review on the to-be-reviewed text using the preset large model, and the second preset proportion is greater than the first preset proportion; When the proportion of the negative comment data in the multiple user feedback data is not less than the second preset proportion, and the number of the negative comment data is not less than the second threshold but less than the first threshold, the text to be reviewed is subjected to a three-level review through the preset large model, and the first threshold is greater than the second threshold.

5. The method according to claim 4, characterized in that The method further includes: determining that the to-be-reviewed text has failed the third-level review if the number of the negative comment data is not less than the first threshold and the proportion of the negative comment data in the plurality of user feedback data is not less than the second preset proportion; If the number of the negative comment data is less than the first threshold, and the proportion of the negative comment data in the plurality of user feedback data is not less than the first preset proportion but less than the second preset proportion, determining that the to-be-reviewed text has passed the third-level review; A preset number of user feedback data in the plurality of user feedback data relate to specific event content, and it is determined that the text to be reviewed has not passed the third-level review.

6. The method according to claim 1, characterized in that The method further includes: adjusting the review direction of the first review rule through the manual interaction interface of the first-level review; and / or, Adjusting the review direction of the second review rule through the manual interaction interface of the second review; and / or, Review the audit results of the secondary audit and provide feedback on the review results through the manual interaction interface of the secondary audit; and / or, The audit results of the three-level audit are reviewed and the review results are fed back through the manual interaction interface of the three-level audit.

7. The method according to any one of claims 1 to 3 and 6, characterized in that The method of performing a primary review on the document to be reviewed by using one of the first review rule and the second review rule, and performing a secondary review on the document to be reviewed by using the other of the first review rule and the second review rule after the document to be reviewed passes the primary review, includes: Using the first review rule to review whether the text to be reviewed contains sensitive content; If the text to be reviewed contains sensitive content, it is determined that the text to be reviewed has not passed the first-level review, and the release of the text to be reviewed is prohibited; If the text to be reviewed does not contain sensitive content, determine that the text to be reviewed has passed the first-level review, and use the second review rule to review the text to be reviewed from a linguistic perspective to determine whether there are any text quality issues; If the text to be reviewed has no text quality issues, it is determined that the text to be reviewed has passed the secondary review.

8. The method according to claim 7, characterized in that The use of the second review rule to review whether the text to be reviewed has text quality issues from a linguistic perspective includes: Obtaining a review result of the to-be-reviewed text in the review direction of the second review rule using a preset large language model, where the review direction of the second review rule includes at least one of whether there are grammatical errors, whether the text is readable, and whether the title is ambiguous; If the review result indicates that the text to be reviewed does not have the problem indicated by the review direction, it is determined that the text to be reviewed does not have text quality problems; If the review result indicates that the text to be reviewed has the problem indicated by the review direction, it is determined that the text to be reviewed has a text quality problem; The method of obtaining the audit result of the text to be audited in the audit direction of the second audit rule by using the preset large language model includes: inputting the text to be audited into the preset large language model, and obtaining the audit result of the text to be audited in the audit direction of the second audit rule output by the preset large language model, wherein the preset large language model is obtained by fine-tuning the model parameters using the fine-tuning dataset of the audit direction of the second audit rule, and the fine-tuning dataset of the audit direction of the second audit rule includes historical text to be audited, the audit result of the historical text to be audited output by the preset large language model, and the historical text manually annotated according to the audit direction of the second audit rule. Alternatively, the method of obtaining the audit result of the text to be audited in the audit direction of the second audit rule by using the preset large language model includes: inputting the text to be audited and preset prompt information into the preset large language model, and obtaining the output result of the preset large language model, wherein the preset prompt information is used to indicate that the text to be audited is to be audited from the audit direction of the second audit rule, and the preset prompt information also indicates the format of the output result of the preset large language model, and the output result records the audit result in the audit direction of the second audit rule.

9. A text review device, characterized in that: include: a first review module configured to perform a primary review of a document to be reviewed using one of a first review rule and a second review rule, and, after the document to be reviewed passes the primary review, perform a secondary review of the document to be reviewed using the other of the first review rule and the second review rule, wherein the first review rule and the second review rule have different review directions; A publishing module, configured to publish the text to be reviewed after the text to be reviewed passes the secondary review; An acquisition module, configured to acquire multiple pieces of user feedback data on the text to be reviewed after the text to be reviewed is published; A second review module is used to perform a three-level review on the text to be reviewed based on the multiple user feedback data; The control module is used to control the publishing module to continue publishing the text to be reviewed if the text to be reviewed passes the third-level review; if the text to be reviewed does not pass the third-level review, control the publishing module to stop publishing the text to be reviewed.

10. An electronic device, characterized in that: comprising at least one processor and a memory connected to the processor, wherein: The memory is used to store computer programs; The processor is configured to execute the computer program so that the electronic device can implement the text review method according to any one of claims 1 to 8.

Images