File content review method, system, computing device, and storage medium

Abstract

Provided in the present application are a file content review method, a system, a computing device, and a storage medium. The method comprises: a file content review system first acquiring a first file, the first file being a file that requires content review; then extracting a first statement from the first file on the basis of a first keyword; and, then determining a second statement on the basis of the first statement, the similarity between the second statement and the first statement satisfying a similarity threshold, such that risk information of the second statement can be used as risk information of the first statement, the risk information being used for indicating whether a corresponding statement contains risky content. By means of measuring the similarity between a statement extracted from a file requiring content review and a statement marked with risk information, the method can accurately evaluate whether the file content contains risky content.

Description

Document content review methods, systems, computing devices, and storage media

[0001] This application claims priority to Chinese Patent Application No. 202410650398.7, filed on May 23, 2024, entitled "Method, System, Computing Device and Storage Medium for Examining Document Content", the entire contents of which are incorporated herein by reference. Technical Field

[0002] This application relates to the field of artificial intelligence (AI), and more particularly to a method, system, computing device, and storage medium for document content review. Background Technology

[0003] Recognition Module

[0004] Files can be used to record various types of data. Current file content review methods mainly rely on keyword recognition technology. The specific solution is as follows: users pre-configure the keywords to be recognized. When a file needs to be reviewed, the file is scanned in its entirety, and the locations in the file where the user-configured keywords appear are marked.

[0005] It can be seen that the above keyword recognition method can only perform content recognition based on preset keywords. The keyword setting method depends on the user being able to define accurate keywords. The recognition results obtained by the above recognition method cannot indicate whether there are risk items in the file content. In other words, the above recognition method cannot effectively identify and judge the risks of the file content. Summary of the Invention

[0006] This application provides a method, system, computing device, and storage medium for reviewing document content, which can accurately assess whether the document content poses a risk.

[0007] In a first aspect, this application provides a document content review method, which is executed by a document content review system. The method includes: firstly obtaining a first document, which is a document that needs to be reviewed; then extracting a first statement from the first document based on a first keyword; then determining a second statement based on the first statement; the similarity between the second statement and the first statement meets a similarity threshold; and thus using the risk information of the second statement as the risk information of the first statement. The risk information is used to indicate whether the corresponding statement is risky content.

[0008] In this solution, the document content review system first extracts sentences containing relevant keywords from the document. Then, it calculates the similarity between the extracted sentences and other sentences marked with risk information. If the similarity between the extracted sentence and a sentence marked with risk information meets a similarity threshold, the two sentences are considered similar and are marked with the same risk information. In other words, the risk information of the previously marked sentence is used as the risk information of the currently extracted sentence. This method can effectively identify risky content in the document.

[0009] In a possible implementation, the similarity threshold can be determined as follows: The document content review system obtains multiple statements, including the second statement, which have the same risk information. Then, the multiple statements are converted into multiple feature vectors, and the multiple statements and the multiple feature vectors correspond one-to-one. Next, the similarity between each pair of the multiple statements is calculated based on the multiple feature vectors to obtain multiple similarities. Then, the similarity threshold is determined based on the multiple similarities.

[0010] In other words, a document content review system can calculate the similarity between pairs of multiple statements with the same risk information, and then determine the similarity threshold based on the calculated similarities. For example, the average of the multiple similarities can be used as the similarity threshold, or the minimum of the multiple similarities can be used as the similarity threshold, and so on.

[0011] In a possible implementation, all the aforementioned statements contain keywords belonging to the first word category. The first statement includes the first keyword, and the first keyword also belongs to the first word category. In other words, the document content review system distinguishes between different word categories. It calculates the similarity between statements that have the same risk information and contain keywords belonging to the same word category. Based on the calculated similarity, it determines the similarity threshold corresponding to that word category. Statements that do not contain keywords belonging to the same word category are not similarity calculated, thus avoiding affecting the accuracy of the similarity threshold. Each word category can calculate its corresponding similarity threshold in a similar manner. When the document content review system extracts statements containing a certain keyword from a document, it can obtain other statements with the same keyword category and risk information based on the word category to which that keyword belongs. Then, it calculates the similarity between the extracted statement and these other statements, and compares it with the similarity threshold corresponding to that word category, thereby accurately determining the risk information of the extracted statement.

[0012] In a possible implementation, the document content review system can extract the first statement in the following way: First, obtain the scanning rules, which are used to indicate the extraction of statements that meet the statement range. Then, divide the first document into multiple statements that meet the statement range, including the first statement. Next, extract the first statement based on the first keyword.

[0013] In this solution, the document content review system first divides the document to be reviewed into multiple statements that meet the statement range indicated by the scanning rules. Then, it determines whether each statement contains the corresponding keyword. Only if a statement contains the corresponding keyword will it be extracted from the document. In other words, the extracted statement must meet both the statement range requirement and the keyword requirement (possessing the corresponding keyword). This application does not specifically limit the statement range; it can be a single sentence, several consecutive sentences, a paragraph, several consecutive paragraphs, X consecutive characters, etc. The statement range can be specified in the scanning rules or set by the default settings of the document content review system.

[0014] In the possible implementations, the scanning rules and the first statement satisfy one of the following:

[0015] (1) The scanning rules include a first keyword and a first identification type. The first identification type is used to indicate the extraction of statements that contain the first keyword and meet the statement range. The first statement contains the first keyword.

[0016] (2) The scanning rules include multiple keywords and a second identification type. The second identification type is used to indicate the extraction of statements that contain these multiple keywords and meet the statement range. The first statement contains these multiple keywords.

[0017] (3) The scanning rules include a keyword set and a third identification type. The keyword set includes a first keyword. The third identification type is used to indicate that statements containing the number of keywords in the keyword set are greater than or equal to the hit number threshold and meet the statement range. The hit number threshold is less than the number of keywords in the keyword set. The first statement contains the first keyword. The number of keywords in the first statement containing the keyword set is greater than or equal to the hit number threshold.

[0018] The above (1) to (3) are actually three ways to set the scanning rules. Method (1) specifies a keyword and a first identification type in the scanning rule. In this case, the scanning rule means that the statement containing the keyword and satisfying the statement range will be extracted. That is, for any statement in the file that satisfies the statement range, as long as the statement contains the keyword specified in the scanning rule, the statement can be extracted. Method (2) specifies multiple keywords and a second identification type in the scanning rule. In this case, the scanning rule means that the statement containing multiple keywords and satisfying the statement range will be extracted. That is, for a statement in the file that satisfies the statement range, the statement will only be extracted if the statement contains multiple keywords specified in the scanning rule. Method (3) specifies a set of keywords (including N keywords, where N is a positive integer greater than 1) and a third identification type in the scanning rule. In this case, the scanning rule means that the statement containing M keywords from the N keywords and satisfying the statement range will be extracted. That is, for a statement in the file that satisfies the statement range, as long as the statement contains any M keywords from the N keywords (M is the hit count threshold), the statement can be extracted. M is a positive integer, M is less than N. M can be given in the scanning rules or it can be the system default setting (such as default M=NP, P is a positive integer).

[0019] In a possible implementation, the scanning rule also includes auxiliary judgment information, which indicates the professional field and / or regulations corresponding to the scanning rule. After the system extracts the first statement in the first document according to the scanning rule and determines the risk information of the first statement, it can provide feedback on the review result of the first document to the user. The review result includes the risk information of the first statement and the aforementioned auxiliary judgment information, thereby improving the interpretability of the review result.

[0020] In a possible implementation, the document content review system can also obtain a second file, which is the file whose content review is required. If the file format of the second file does not conform to the file type configured in the document content review system, the scanning time of the second file exceeds a first time threshold, or the waiting time of the second file exceeds a second time threshold, the system will notify the user that the second file cannot be processed temporarily.

[0021] In other words, a protection mechanism is provided here. When a file that needs content review does not conform to the file type configured by the system, the file takes too long to scan, or the file takes too long to wait for, the system can notify the user that the file cannot be processed temporarily and stop processing the file to avoid affecting the processing of other files. This helps to ensure the reliability of the system.

[0022] Secondly, this application also provides a document content review system, which includes a functional module for performing the method described in the first aspect or any possible implementation of the first aspect.

[0023] Thirdly, this application also provides a computing device including a processor and a memory. The processor is configured to execute instructions stored in the memory to cause the computing device to perform operational steps of the method as described in the first aspect or any possible implementation thereof.

[0024] Fourthly, this application also provides a computing device cluster, including at least one computing device, each computing device including a processor and a memory. The processor of the at least one computing device is used to execute instructions stored in the memory of the at least one computing device, so that the computing device cluster performs the operational steps of the method as described in the first aspect or any possible implementation thereof.

[0025] Fifthly, this application also provides a chip system including a processor and a power supply circuit, the power supply circuit being used to supply power to the processor, and the processor being used to perform operational steps of the method as described in the first aspect or any possible implementation thereof.

[0026] In a sixth aspect, this application also provides a computer-readable storage medium including computer program instructions that, when executed by a computing device, perform operational steps of the method as described in the first aspect or any possible implementation thereof.

[0027] In a seventh aspect, this application also provides a computer program product comprising instructions. When the aforementioned instructions are executed by a computing device, the computing device performs the operational steps of the method as described in the first aspect or any possible implementation thereof.

[0028] Based on the implementation methods provided in the above aspects, this application can be further combined to provide more implementation methods. Attached Figure Description

[0029] Figure 1 is a system architecture diagram provided in this application;

[0030] Figure 2 is a schematic diagram of a file parsing process provided in this application;

[0031] Figure 3 is a schematic diagram of the structure of a semantic recognition model provided in this application;

[0032] Figure 4 is a schematic diagram illustrating the stages of document content review provided in this application;

[0033] Figure 5 is a flowchart illustrating a document content review method provided in this application;

[0034] Figure 6 is a schematic diagram of the structure of a computing device provided in this application;

[0035] Figure 7 is a schematic diagram of the structure of a computing device cluster provided in this application;

[0036] Figure 8 is a schematic diagram of the two computing devices provided in this application connected via a network. Detailed Implementation

[0037] To address the problem of existing technologies failing to effectively identify and assess the risks of document content, this application provides a document content review system. This system first extracts sentences from the document based on keywords, and then uses the risk information of similar sentences as the risk information for those sentences. Specifically, the system first extracts sentences from the document requiring content review based on keywords, then calculates the similarity between the extracted sentences and other sentences marked with risk information. This risk information indicates whether the corresponding sentence is risky content. If the similarity between the extracted sentence and a sentence marked with risk information meets a similarity threshold, the two sentences are considered similar. The risk information of that previously marked sentence is then used as the risk information for the currently extracted sentence. This method effectively identifies risky content in the document.

[0038] The following is a detailed introduction to the document content review system.

[0039] Please refer to Figure 1, which is a system architecture diagram provided in this application, including a client 100 and a document content review system 200. The client 100 and the document content review system 200 are connected by communication, which can be a wired connection or a wireless connection. The number of clients 100 that establish a communication connection with the document content review system 200 can be one or more (Figure 1 is an example of one client 100), and this application does not make a specific limitation in this regard.

[0040] The client 100 is used to implement human-computer interaction and can be deployed on terminal devices or computing devices. The terminal device can be a smartphone, wearable device, laptop, tablet, in-vehicle device, or smart conferencing device, etc., and the computing device can be a server, personal computer (PC), etc., and this application does not make specific limitations in this regard.

[0041] In some specific implementations, client 100 can be an application (APP) client / mobile client running on a mobile terminal such as a smartphone or wearable device, or software or application running on a computing device (such as a PC client), or a web application / website accessed based on a web browser, or a front-end console of a cloud platform. This application does not specifically limit it in this regard.

[0042] The document content review system 200 provides document content review functionality and can be deployed on a single computing device, a cluster of multiple computing devices, or a terminal device. The computing device can be a physical server, a virtual machine, a container, or an edge computing device. A virtual machine refers to a complete computer system simulated by software, possessing full hardware system functionality and running in a completely isolated environment. When creating a virtual machine on a computing device, a portion of the physical machine's hard drive and memory capacity is used as the virtual machine's hard drive and memory capacity. Each virtual machine has an independent basic input / output system (CMOS), hard drive, and operating system, and can be operated like a physical machine. A container is a portable software unit that can combine an application and all its dependencies into a single software package. This package is not limited by the underlying host operating system, thus eliminating the need to build complex environments and simplifying the application development and deployment process. Edge computing devices refer to devices closer to the data source and end user, characterized by low latency and high bandwidth, such as intelligent routers and edge servers; this application does not specifically limit the definition. Terminal devices are described above and will not be repeated here.

[0043] Optionally, the client 100 and the document content review system 200 can be deployed on the same terminal device or computing device; or, the client 100 can be deployed on the terminal device, and the document content review system 200 can be deployed on a single computing device or a cluster of computing devices. It should be understood that the above examples are for illustrative purposes only, and the specific deployment of the client 100 and the document content review system 200 can be determined according to the actual application scenario.

[0044] Furthermore, the document content review system 200 can be divided into multiple modules according to its functions. Figure 1 exemplarily shows one way of dividing the document content review system 200, including an acquisition module 210, an extraction module 220, a determination module 230, a sending module 240, and a training module 250, as well as a semantic recognition model 260, a scanning rule base 270, and a sentence sample base 280, which will be described below.

[0045] 1. Acquisition Module 210: Used to acquire files that require content review.

[0046] Users can send one or more files (i.e., upload files individually or in batches) to the file content review system 200 through client 100, thereby obtaining the files to be reviewed by the acquisition module 210 in the file content review system 200. In addition to receiving files to be reviewed by users through client 100, the acquisition module 210 can also obtain files to be reviewed from other modules or systems (such as file systems, database systems, etc.).

[0047] This application does not limit the file type of documents that require content review. For example, file types may include plain text files (TXT), Word documents (DOCX), portable document format (PDF) documents, presentations (PowerPoint, PPT), spreadsheets (EXCEL) files, image formats (such as JPG, PNG, GIF, etc.), and so on.

[0048] The acquisition module 210 is also used to acquire scanning rules. For example, the acquisition module 210 can receive one or more scanning rules sent by the user through the client 100, or it can acquire one or more scanning rules from other modules or systems. This application does not limit this.

[0049] Scanning rules are used to indicate the keywords that need to be identified. Scanning rules can include the following three settings (1) to (3).

[0050] (1) The scanning rules include a single keyword and a first identification type, which is used to indicate the extraction of statements that contain the keyword and meet the statement range.

[0051] It should be understood that method (1) is to specify a keyword and a first identification type in the scanning rule. In this case, the scanning rule means that the statement with the keyword and that meets the statement range will be extracted. That is, for any statement in the file that meets the statement range, as long as the statement contains the keyword specified in the scanning rule, the statement can be extracted.

[0052] The statement range refers to the scope within which keywords are identified. The statement range can be a single sentence (marked with a period), multiple consecutive sentences, a paragraph (marked with a newline character), X consecutive paragraphs (X being a positive integer), Y consecutive characters (Y being a positive integer), etc. This application does not impose specific limitations on this. A specific statement range can be specified in the scanning rules, or it can be left unspecified (in which case the default statement range is used). This application does not impose any limitations on this. In other words, if a single keyword given in the scanning rules appears in a statement that satisfies the statement range, that statement can be extracted.

[0053] For example, suppose the scanning rules include the keyword "artificial intelligence," the first recognition type, and the sentence range "a paragraph." This means that the keyword "artificial intelligence" needs to be identified within a paragraph. If a document is scanned based on these scanning rules, then every paragraph in the document containing the keyword "artificial intelligence" will be extracted.

[0054] (2) The scanning rules include multiple keywords and a second identification type. The second identification type is used to indicate the extraction of statements that contain these multiple keywords and meet the statement range.

[0055] It should be understood that method (2) specifies multiple keywords and a second identification type in the scanning rule. In this case, the scanning rule means that statements that simultaneously contain these multiple keywords and meet the statement range will be extracted. That is to say, for statements in the file that meet the statement range, the statement will only be extracted if the statement contains all the keywords specified in the scanning rule. For the statement range, please refer to the introduction in method (1), which will not be repeated here.

[0056] For example, suppose the scanning rule includes the keywords "artificial intelligence" and "image" as well as a second recognition type, and the sentence range defaults to "one sentence." This means that both the keywords "artificial intelligence" and "image" need to be recognized in a single sentence. If a file is scanned based on this scanning rule, then every sentence in the file containing the keywords "artificial intelligence" and "image" will be extracted.

[0057] (3) The scanning rules include a keyword set and a third identification type. The keyword set includes N keywords, where N is a positive integer greater than 1. The third identification type is used to indicate the extraction of statements that contain more than or equal to M keywords in the keyword set and satisfy the statement range.

[0058] It should be understood that method (3) specifies a keyword set and a third identification type in the scanning rule. In this case, the scanning rule means that statements that have M keywords from the keyword set and satisfy the statement range will be extracted. That is, for a statement in the file that satisfies the statement range, as long as any M keywords from the N keywords appear in the statement, the statement can be extracted. Here, the value of M is called the hit count threshold, and M is a positive integer less than N. M can be given in the scanning rule or can be the default setting of the file content review system 200 (such as default M=NP, P is a positive integer). This application does not limit this.

[0059] For example, suppose the scanning rule includes a keyword set consisting of the keywords "artificial intelligence," "image," and "cat," as well as a third recognition type. The rule also specifies a hit threshold M of 2 and a default sentence range of "50 characters," meaning that at least two keywords from the keyword set must be identified within 50 characters. If a file is scanned based on this rule, then any sentence in the file containing at least two keywords from the keyword set and with a character count within 50 characters will be extracted.

[0060] Optionally, in addition to specifying specific keywords in the scanning rules, you can also indicate the category to which the keywords to be identified belong (in which case, specific keywords are not specified) and the identification type in the scanning rules. The identification types can be found in the first, second, and third identification types described above, and will not be repeated here.

[0061] For example, suppose the document content review system 200 categorizes words into categories such as risky actions, people / objects, and code names, and each category has one or more keywords. This application does not impose specific limitations on this. The acquisition module 210 receives a scanning rule sent by the user through the client 100. This scanning rule indicates the two word categories: risky actions and people / objects. The recognition type given in the scanning rule is the second recognition type, using the default sentence range (assuming it is one sentence). This means that it is necessary to extract sentences that contain keywords belonging to the risky actions category and keywords belonging to the people / objects category and meet the sentence range.

[0062] Suppose a file uploaded by a user to the file content review system 200 via client 100 contains the statement: "A conflict occurred during the visit, and person A assaulted person B." Based on the aforementioned scanning rules, scanning the file reveals that the statement includes the keyword "assault," which falls under the category of risky actions, and the keywords "person A" and "person B," which fall under the category of person objects. Furthermore, the statement meets the requirement of being a single sentence, therefore it can be extracted from the file.

[0063] Optionally, the scanning rules may also include auxiliary judgment information, which indicates the professional field and / or regulations corresponding to the scanning rules. When the document content review system 200 subsequently scans the user-uploaded file based on these scanning rules, it can provide the auxiliary judgment information back to the user, so that the user is aware of the professional field and / or regulations corresponding to the scanning rules adopted by the document content review system 200, thereby improving the interpretability of the document content review results output by the document content review system 200.

[0064] For example, a user can provide scanning rules on the client 100 in the form of Table 1. Each row in the table represents a scanning rule, and each scanning rule may include information such as serial number, professional field / regulation, keywords, identification type, and statement range (optional). When the acquisition module 210 receives these scanning rules sent by the user through the client 100, it can store them in the scanning rule base 270 of Figure 1.

[0065] Table 1

[0066] The scanning rule with serial number 001 specifies the field of image recognition. This rule also provides the keywords "artificial intelligence" and "image," the designated recognition type as "second recognition type," and the specified sentence range as "one sentence." This means it needs to extract sentences containing the keywords "artificial intelligence" and "image" that meet the "one sentence" requirement. The scanning rule with serial number 002 specifies the rule as "Article XX of Enterprise Information Management stipulates: It is prohibited to disclose employee login accounts and passwords." This rule provides the keywords "account," "password," "login," and "employee," the third recognition type, and specifies N=2 and the sentence range as "a paragraph." This means it needs to extract sentences containing at least two of the above four keywords that meet the "a paragraph" recognition range.

[0067] It should be noted that the scanning rules given in the above examples are for illustrative purposes only and do not constitute specific limitations. In actual application scenarios, other scanning rules can be set, and scanning rules can also include other fields.

[0068] Optionally, when a user sends a file requiring content review to the document content review system 200 via client 100, they can also specify the corresponding professional field for that file. Subsequently, the document content review system 200 can process the file using the scanning rules of the corresponding professional field. If the user does not specify the corresponding professional field for the file, the document content review system 200 can process the file using some or all of the scanning rules stored in the scanning rule base 270 of Figure 1.

[0069] 2. Extraction module 220: Used to scan the file according to the scanning rules, thereby extracting sentences containing corresponding keywords from the file.

[0070] Specifically, when the acquisition module 210 acquires a file that requires content review, it sends the file to the extraction module 220.

[0071] As shown in Figure 2, the extraction module 220 first determines the file type of the file. For example, the file can be an image file (JPG, PNG, GIF, etc.) or a file type (EXCEL, Word, PDF, or PPT, etc.). Then, it determines whether the file contains images. If the file does not contain images, the text in the file can be directly output. If the file contains images, the images need to be rotated and corrected to ensure they are in the correct position (skewed images will negatively affect text extraction). Then, the text in the images is converted into text using an Optical Character Recognition (OCR) algorithm, and the corresponding text for the file is output (including text converted from images in the file and the original text in the file). OCR is a technology that converts text in an image into editable and searchable text, which will not be discussed in detail here.

[0072] Then, the extraction module 220 scans the text corresponding to the aforementioned file based on the scanning rules, thereby extracting statements containing the corresponding keywords from the file. Specifically, the extraction module 220 first obtains the scanning rules, which are used to indicate the extraction of statements that meet the statement range and contain the corresponding keywords. The scanning rules and statement range can be referred to in the previous introduction and will not be repeated here. Next, the extraction module 220 divides the text corresponding to the aforementioned file according to the statement range indicated by the scanning rules, that is, divides the aforementioned text into multiple statements that meet the statement range, and then determines whether each divided statement contains the corresponding keyword specified in the scanning rules. If it contains the corresponding keyword, the statement is extracted. These extracted statements are the statements to be identified for risk information, and their risk information will be determined by the determination module 230 (described later).

[0073] For example, suppose the scanning rule includes the keywords "artificial intelligence" and "image" as well as a second recognition type, and the default sentence range is "one sentence." In this case, the scanning rule means extracting sentences that contain the keywords "artificial intelligence" and "image" and meet the "one sentence" sentence range. Based on this scanning rule, the extraction module 220 first divides the document to be reviewed into multiple sentences (i.e., multiple sentences that meet the "one sentence" sentence range). That is, it scans from the beginning of the document, dividing it into sentences for each period found, and a total of multiple sentences can be divided. Then, the extraction module 220 determines whether each sentence contains the keywords "artificial intelligence" and "image." If it does, the sentence is extracted from the document; otherwise, it is not extracted.

[0074] For example, suppose the scanning rule includes a keyword set consisting of the keywords "artificial intelligence," "image," and "cat," as well as a third recognition type. The scanning rule also specifies a hit threshold M of 2, and the default sentence range is "a paragraph." In this case, the scanning rule means extracting sentences that contain at least two keywords from the aforementioned keyword set and satisfy the "a paragraph" sentence range. Based on this scanning rule, the extraction module 220 first divides the document requiring content review into multiple sentences that satisfy the "a paragraph" sentence range. That is, it scans from the beginning of the document, identifying a sentence for each newline character found, resulting in multiple sentences that meet the "a paragraph" requirement. Then, the extraction module 220 determines whether each of these multiple sentences contains at least two keywords from "artificial intelligence," "image," and "cat." If so, the sentence is extracted from the document; otherwise, it is not extracted.

[0075] 3. Training module 250: Used to train semantic recognition model 260.

[0076] As shown in Figure 1, the statement sample library 280 provides multiple statement samples for training the semantic recognition model 260. These statement samples are statements that satisfy the scanning rules. Each statement sample can be a sentence, multiple consecutive sentences, a paragraph, X consecutive paragraphs, Y consecutive characters, etc., and this application does not impose specific limitations on this. Each statement sample has a corresponding risk information label, which indicates whether the corresponding statement sample is risky content. The risk information labels of the statement samples can be manually labeled. For example, first, the manual judgment is made regarding which keywords exist in the statement sample, and then the semantics of these keywords in the statement sample are used to determine whether the statement sample is risky content, and then the corresponding risk information label is labeled for the statement sample.

[0077] For example, risk information tags can be divided into two types: high-risk tags and low-risk tags. High-risk tags indicate that the corresponding statement sample is risky content (a real risk exists), while low-risk tags indicate that the corresponding statement sample is not risky content (although it hits the keywords in the scanning rules, it is not real risky content and is a false hit). Annotators can judge whether a statement sample is risky content based on experience, and then label the statement sample with the corresponding risk information tags.

[0078] For example, risk information tags can be divided into three types: high-risk tags, medium-risk tags, and low-risk tags. A high-risk tag indicates that the corresponding statement sample contains risky content (a real risk exists) and the level of risk is relatively high. A medium-risk tag indicates that the corresponding statement sample contains risky content (a real risk exists) and the level of risk is relatively low. In other words, both high-risk and medium-risk tags indicate risky content, only the degree of risk differs; whether it is high-risk or medium-risk can be determined by human experience. A low-risk tag indicates that the corresponding statement sample does not contain risky content (it belongs to a statement that has falsely matched keywords). Tagging personnel can determine whether a statement sample contains risky content based on their experience and then tag the statement sample with the appropriate risk information tag.

[0079] Optionally, users can send either scanning rules or labeled sentence samples to the document content review system 200. The document content review system 200 stores the sentence samples sent by the user in the sentence sample library 280 shown in Figure 1, and can then use the sentence samples sent by the user to train the semantic recognition model 260. Alternatively, users can only send scanning rules to the document content review system 200 without actively sending sentence samples. In this case, the document content review system 200 can use the existing sentence samples in the sentence sample library 280 shown in Figure 1 (these sentence samples can be labeled and uploaded by other personnel) to train the semantic recognition model 260.

[0080] For example, suppose a user sends scanning rule number 001 from Table 1 to the document content review system 200 via client 100. This scanning rule corresponds to the image domain and is used to instruct the extraction of sentences containing the keywords "artificial intelligence" and "image" that fall within the scope of a single sentence. The user can also obtain sample sentences containing the keywords "artificial intelligence" and "image" and, based on their experience, label these sample sentences with corresponding risk information tags. These risk information tags indicate whether the sentences are risky content. The labeled sample sentences are then sent to the document content review system 200 so that the system can train the semantic recognition model 260 based on the user-sent sample sentences.

[0081] For example, suppose a user sends scanning rule number 002 from Table 1 to the document content review system 200 via client 100. This scanning rule instructs the extraction of sentences containing at least two keywords from a keyword set that satisfy the definition of a "sentence". Specifically, the keyword set includes the keywords "account", "password", "login", and "employee". The user does not need to actively send sentence samples to the document content review system 200. In this case, the document content review system 200 can obtain sentence samples that satisfy the scanning rule from the sentence sample library 280 in Figure 1 for training the semantic recognition model 260. Sentence samples that satisfy the scanning rule can be uploaded to the sentence sample library 280 by other users; this application does not limit this.

[0082] Then, the training module 250 trains the semantic recognition model 260 based on multiple sentence samples in the sentence sample library 280. The semantic recognition model 260 here can be a language model such as a Transformer network or bi-directional encoder representations from transformers (BERT), and this application does not limit it.

[0083] The training process will be described in detail below.

[0084] First, the training module 250 acquires multiple statement samples with the same risk information, that is, the risk information of these multiple statement samples indicates that they are all risky content or none of them are risky content. Then, these multiple statement samples are input into the semantic recognition model 260.

[0085] The semantic recognition model 260 converts these multiple sentence samples into multiple feature vectors, each of which corresponds one-to-one with a sentence sample. The conversion of sentence samples into multiple feature vectors can be achieved using methods commonly used in Natural Language Processing (NLP), such as word embeddings and Term Frequency-Inverse Document Frequency (TF-IDF), and this application does not impose any limitations on this approach.

[0086] Next, the semantic recognition model 260 calculates the pairwise similarity between these multiple feature vectors. That is, the similarity between two sentence samples is obtained by calculating the similarity between the feature vectors corresponding to the two sentence samples. The calculation of the similarity between two feature vectors can be achieved by calculating the Euclidean distance (Euclidean distance measures the straight-line distance between two vectors; the smaller the value, the more similar the two vectors), Manhattan distance, cosine similarity (cosine similarity measures the similarity in direction between two vectors; the value ranges from -1 to 1; the closer the value is to 1, the more similar the two vectors), etc. This application does not limit the specific method used.

[0087] After calculating the pairwise similarity between the above multiple sentence samples, multiple similarities can be obtained. Then, the semantic recognition model 260 determines the similarity threshold based on these multiple similarities. For example, the arithmetic mean of these multiple similarities can be used as the similarity threshold, or the minimum value among these multiple similarities can be used as the similarity threshold, and so on.

[0088] Optionally, multiple different thresholds can be set (randomly or manually), and then the model score obtainable at each threshold can be evaluated. Finally, the threshold with the highest score is selected as the similarity threshold.

[0089] Specifically, first, a threshold is selected. For each of the multiple sentence samples labeled with risk information, the similarity between that sentence sample and all other sentence samples in the multiple sentence samples (the method for calculating the similarity between sentence samples can be referred to the previous introduction). If the similarity between that sentence sample and another sentence sample meets the threshold, the semantic recognition model 260 predicts the risk information of that sentence sample as the risk information of that sentence sample. Then, it determines whether the risk information labeled by that sentence sample is consistent with the risk information predicted by the semantic recognition model 260 for that sentence sample. If they are consistent, the judgment is considered accurate; if they are inconsistent, the judgment is considered inaccurate. In the above way, the number of sentence samples whose risk information is accurately judged under this threshold can be counted, and then the number of accurately judged sentence samples is used as the model score under this threshold. The model scores under other thresholds can also be determined in the same way. By comparing the model scores under different thresholds, the threshold with the highest score is finally used as the similarity threshold.

[0090] Optionally, multiple different word categories can be defined, such as number categories, risky action categories, person / object categories, and region categories, etc., which are not limited in this application. For each of the above multiple word categories, the training module 250 acquires multiple sentence samples with the same risk information and containing keywords belonging to that word category to train the semantic recognition model 260. The semantic recognition model 260 converts the above multiple sentence samples into multiple feature vectors, and then calculates the similarity between each pair of these multiple sentence samples based on these multiple feature vectors, and then determines the similarity threshold corresponding to the word category based on the obtained multiple similarities. In the above manner, the semantic recognition model 260 can determine the corresponding similarity threshold for each word category.

[0091] Optionally, as shown in Figure 3, the semantic recognition model 260 includes a common module and multiple grouping modules (the figure shows the first and second grouping modules as examples), with each grouping module corresponding to a different word category. The common module converts the input sentence samples into corresponding feature vectors, and then inputs the feature vectors of the sentence samples into the corresponding grouping modules based on the word categories to which the keywords in the sentence samples belong. The grouping modules determine the similarity threshold corresponding to the word category based on the feature vectors of the multiple sentence samples (containing keywords belonging to the word category corresponding to that grouping module) input from the common module (refer to the previous description).

[0092] For example, training module 250 acquires multiple sentence samples containing keywords belonging to the first word category and inputs them into the common module of semantic recognition model 260. The common module converts these multiple sentence samples into corresponding feature vectors. Since these multiple sentence samples all contain keywords belonging to the first word category, the common module inputs the feature vectors of these multiple sentence samples into the first grouping module (corresponding to the first word category). Then, the first grouping module calculates the pairwise similarity between the feature vectors of these multiple sentence samples, obtaining multiple similarities. Based on these multiple similarities, a similarity threshold corresponding to the first word category is determined. For example, the arithmetic mean of these multiple similarities can be used as the similarity threshold corresponding to the first word category, or the minimum value among these multiple similarities can be used as the similarity threshold corresponding to the first word category, and so on. The similarity thresholds corresponding to other word categories besides the first word category can also be determined in the same way, which will not be elaborated here.

[0093] 4. Determining module 230: Used to determine the risk information of the statements to be identified. The statements to be identified are the statements extracted by the extraction module 220 from the documents that require content review.

[0094] Specifically, after the training module 250 trains the semantic recognition model 260, the determination module 230 can use the semantic recognition model 260 to determine the risk information of the statement to be identified. Assuming the extraction module 220 extracts the first statement from the document requiring content review, the determination module 230 can input the first statement into the semantic recognition model 260. The semantic recognition model 260 first converts the first statement into a feature vector, and then calculates the similarity between this feature vector and the feature vectors corresponding to statement samples previously learned by the semantic recognition model 260. If the similarity between the feature vector of the first statement and the feature vector of the second statement (any statement sample previously learned by the semantic recognition model 260) meets the aforementioned similarity threshold, then the semantic recognition model 260 determines that the second statement is a similar statement to the first statement, and thus determines the risk information marked by the second statement as the risk information of the first statement. In other words, if the risk information marked by the second statement indicates that the second statement is risky content, then the semantic recognition model 260 determines that the first statement is also risky content; if the risk information marked by the second statement indicates that the second statement is not risky content, then the semantic recognition model 260 determines that the first statement is also not risky information.

[0095] Optionally, as described above, the semantic recognition model 260 can determine the corresponding similarity thresholds for multiple word categories during the training phase. After the semantic recognition model 260 is trained, assuming the extraction module 220 extracts the first sentence from the document requiring content review, and the first sentence contains the first keyword belonging to the first word category, the determination module 230 can input the first sentence into the semantic recognition model 260 to determine the risk information of the first sentence. The semantic recognition model 260 first converts the first sentence into a feature vector, then obtains the similarity threshold corresponding to the word category to which the first keyword belongs, and calculates the similarity between the feature vector of the first sentence and the feature vectors corresponding to the sentence samples containing the same word category that the semantic recognition model 260 has previously learned. If the similarity between the feature vector of the first sentence and the feature vector of the second sentence (any sentence sample containing the same word category that the semantic recognition model 260 has previously learned) meets the similarity threshold corresponding to the word category, then the semantic recognition model 260 determines that the second sentence is a similar sentence to the first sentence, and thus determines the risk information marked by the second sentence as the risk information of the first sentence.

[0096] Optionally, the determining module 230 can also calculate the metrics of the semantic recognition model 260, and then the sending module 240 sends the metrics to the user. The first keyword can be a keyword in the scanning rules sent by the user through the client 100, or it can be other keywords; this application does not limit this.

[0097] The above metrics may include at least one of recall and precision:

[0098] (1)

[0099] (2)

[0100] Wherein, C is the number of statements among the multiple statements containing the first keyword that are identified as risky content by the semantic recognition model 260, K is the number of statements among the statements identified as risky content by the semantic recognition model 260 that are not actually risky content, and S is the number of statements among the multiple statements among the statements identified as not risky content by the semantic recognition model 260 that are actually risky content.

[0101] For example, assuming the primary keyword is "beating," 100 statements containing this keyword are input into the semantic recognition model 260. The semantic recognition model 260 outputs risk information for each of these 100 statements, indicating whether the corresponding statement is risky content. As shown in Table 2, 80 statements were identified as risky content by the semantic recognition model 260, while 20 statements were identified as non-risky content.

[0102] Table 2

[0103] Of the 80 statements identified as risky content by the model, manual review revealed that 70 statements were correctly identified as risky, while the remaining 10 statements were incorrectly identified as such. These 10 statements were actually non-risky content but were mistakenly identified as risky by the semantic recognition model 260. Of the 20 statements identified as non-risky content by the model, manual review revealed that 18 statements were correctly identified as risky, while the remaining 2 statements were incorrectly identified as such. These 2 statements were actually risky content but were mistakenly identified as non-risky by the semantic recognition model 260.

[0104] Based on the sensitivity calculation formula (1), the corresponding keyword can be determined. Based on the hit rate calculation formula (2), the first keyword can be determined. The calculation process for sensitivity and hit rate for other keywords is similar and will not be described in detail here. Then, the sending module 240 can send at least one of the sensitivity and hit rate for the first keyword to the client 100, so that the user can view the sensitivity and / or hit rate of the semantic recognition model 260 when recognizing the first keyword on the client 100. These performance indicators can help the user set appropriate scanning rules.

[0105] Optionally, if the sensitivity corresponding to the first keyword is lower than the sensitivity threshold and / or the hit rate corresponding to the first keyword is lower than the hit rate threshold, the training module 250 can acquire more sentence samples containing the first keyword to train the semantic recognition model 260, thereby improving the sensitivity and / or hit rate corresponding to the first keyword. The sensitivity threshold and hit rate threshold can be set by the user according to actual usage needs, and this application does not limit them.

[0106] Optionally, if a file contains statements that are identified as risky content by the semantic recognition model 260, the file can be identified as a risky file (or its risk level is determined to be high); if a file does not contain statements that are identified as risky content by the semantic recognition model 260, the file can be identified as a safe file (or its risk level is determined to be low).

[0107] 5. Sending module 240: Used to send the file review results to client 100.

[0108] The document review results include statements containing corresponding keywords extracted from the document by the extraction module 220 according to the scanning rules, as well as risk information output by the semantic recognition model 260 for the statements extracted from the document. For details on how to extract statements containing keywords from the document according to the scanning rules and determine the risk information of the statements, please refer to the previous introduction; it will not be repeated here.

[0109] Optionally, the sending module 240 can send the review results of the document to the client 100 where the reviewer (a type of user) is located. The reviewer provides review comments on the review results, which are used to indicate whether the risk information in the review results is accurate, and then feeds the review comments back to the document content review system 200.

[0110] For example, suppose the extraction module 220 extracts a statement containing the keyword "beating" from a file based on scanning rules. The determination module 230 inputs this statement into the semantic recognition model 260. The model outputs risk information indicating that the statement is not risky content. The sending module 240 includes this statement and its risk information in the review result of the file, and then sends the review result to the client 100 logged in by the reviewer. The reviewer views the risk information of the statement in the review result on the client 100. Based on experience, the reviewer determines that the statement is actually risky content. Therefore, the reviewer sends a review opinion through the client 100, indicating that the risk information of the statement given in the review result is inaccurate.

[0111] Optionally, new statement samples can be constructed based on the risk information of the statements output by the semantic recognition model 260 and the review opinions given by the reviewers on the risk information. The new statement samples are then added to the statement sample library 280 for training the semantic recognition model 260, thereby improving the accuracy of the semantic recognition model 260 in identifying risky content.

[0112] For example, if the semantic recognition model 260 outputs a risk information indicating that a statement containing the keyword "artificial intelligence" from a file, and the reviewer's comments indicate that the risk information is accurate, then this statement can be considered a new statement sample, and its risk information label indicates that it is risky content. Conversely, if the semantic recognition model 260 outputs a risk information indicating that a statement containing the keyword "password" from a file, but the reviewer's comments indicate that the risk information is inaccurate, then this statement can be considered another statement sample, and its risk information label indicates that it is risky content.

[0113] This application does not specify the number of reviewers or the review process.

[0114] For example, Figure 4 provides an exemplary schematic diagram of the stages of document content review. The first stage is the system review of files uploaded by the requester (a type of user). First, the document content review system 200 receives one or more files uploaded by the requester that require content review. Then, the document content review system 200 automatically begins to review the files (refer to the descriptions of the extraction module 220 and the determination module 230) and records the review results (to be manually reviewed). The review results include statements containing keywords extracted from the file and risk information about those statements.

[0115] The second stage is the review stage. Here, it's assumed that two reviewers are assigned: a Level 1 reviewer and a Level 2 reviewer. They are responsible for reviewing the document content review results output by the document content review system 200. First, the Level 1 reviewer views the document review results. They can view the review details page for individual documents or the summary page for all documents. On the review details page for individual documents, they can view the keyword hit rate and the risk information of the sentences containing those keywords. The Level 1 reviewer can verify the accuracy of the risk information provided by the document content review system 200 regarding the keywords, then provide review comments and submit them. On the summary page for all documents, the Level 1 reviewer can manually confirm the accuracy of the risk level assessment made by the document content review system 200, and also confirm whether the system has reviewed all documents uploaded by the requester.

[0116] After the first-level reviewer completes their review, the second-level reviewer can view the document review results. They can view the review details page for individual documents, the summary page for all documents, and the first-level reviewer's comments. For details on the review details page for individual documents and the summary page for all documents, please refer to the previous explanation; it will not be repeated here. Optionally, if the first-level and second-level reviewers' comments differ, the second-level reviewer's comments will prevail, or the second stage of Figure 4 will be restarted.

[0117] The third stage is document processing and result return. Based on the reviewers' comments, it is determined whether the document passes review. If the reviewers indicate that the document does not contain any risky content, it is deemed to have passed review; otherwise, it is deemed to have failed review. Then, feedback is sent to the requester regarding the document's approval status.

[0118] The fourth stage is operational analysis, which improves the accuracy of the document content review system 200 in identifying risky content. Specifically, by recording the risk information of the sentences output by the semantic recognition model 260 in the document content review system 200, as well as the reviewers' comments on the risk information output by the model, the performance indicators of the semantic recognition model 260 can be calculated, including sensitivity and / or hit rate for various keywords (see previous introduction). If the semantic recognition model 260's sensitivity and / or hit rate for a certain keyword is substandard, the semantic recognition model 260 can be further trained using sentence samples containing that keyword.

[0119] Optionally, the document content review system 200 can be configured with a circuit breaker protection mechanism. This mechanism is used to send the sending module 240 to the requester in the event of system crash, unsupported file type, scan timeout, or waiting timeout, thereby ensuring the stable operation of the entire document content review system 200 and improving system reliability.

[0120] For example, suppose a requester sends a file to a file content review system 200 through client 100. Then, the extraction module 220 in the file content review system 200 determines the file type and finds that the file type is not currently supported by the file content review system 200. Therefore, the sending module 240 can report to client 100 that the file cannot be processed temporarily.

[0121] For example, suppose a user sends a large number of files to the file content review system 200 via client 100. The extraction module 220 in the file content review system 200 then scans these files sequentially based on the scanning rules in the scanning rule base 270 shown in Figure 1. If the scanning time for a file exceeds a first time threshold (possibly due to system overload, excessive file size, etc.), scanning of that file can be stopped, and the sending module 240 will report to client 100 that the file cannot be processed temporarily, without affecting the processing of other files. If the waiting time for a file exceeds a second time threshold (before it has been scanned), the sending module 240 can report to client 100 that the file cannot be processed temporarily. This application does not specifically limit the first and second time thresholds; they can be reasonably set according to actual usage requirements.

[0122] It should be noted that the semantic recognition model 260 and / or the statement sample library 280 in Figure 1 can be set inside the document content review system 200 or outside the document content review system 200, and this application does not limit this.

[0123] It should also be noted that the document content review system 200 in Figure 1 is only functionally and exemplaryly divided into an acquisition module 210, an extraction module 220, a determination module 230, a sending module 240, and a training module 250. In reality, the document content review system 200 in Figure 1 can also contain more or fewer modules. For example, one of the above modules can be split into multiple functional modules, or two or more of the above modules can be merged into one functional module. Other functional modules can also be added to the document content review system 200 in Figure 1. This application does not specifically limit this. The acquisition module 210, extraction module 220, determination module 230, sending module 240, and training module 250 can all be implemented in software or in hardware. For example, the implementation method of the extraction module 220 will be described below. Similarly, the implementation methods of the other modules mentioned above can refer to the implementation method of the extraction module 220.

[0124] As an example of a software functional unit, module 220 may include code running on a computing instance. The computing instance may include at least one of a physical host (computing device), a virtual machine, or a container. Further, the aforementioned computing instance may be one or more. For example, module 220 may include code running on multiple hosts / virtual machines / containers. It should be noted that the multiple hosts / virtual machines / containers used to run the code may be distributed within the same region or in different regions. Further, the multiple hosts / virtual machines / containers used to run the code may be distributed within the same availability zone (AZ) or in different AZs, each AZ including one or more geographically proximate data centers. Typically, a region may include multiple AZs. Similarly, the multiple hosts / virtual machines / containers used to run the code may be distributed within the same virtual private cloud (VPC) or in multiple VPCs. Typically, a VPC is set up within a region. Communication between two VPCs within the same region, as well as between VPCs in different regions, requires a communication gateway to be set up within each VPC. The interconnection between VPCs is achieved through the communication gateway.

[0125] As an example of a hardware functional unit, the extraction module 220 may include at least one computing device, such as a server. Alternatively, the extraction module 220 may also be a device implemented using an application-specific integrated circuit (ASIC) or a programmable logic device (PLD). The PLD may be implemented using a complex programmable logical device (CPLD), a field-programmable gate array (FPGA), generic array logic (GAL), or any combination thereof.

[0126] The extraction module 220 includes multiple computing devices that can be distributed in the same region or in different regions. Similarly, the extraction module 220 can be distributed in the same Availability Zone (AZ) or in different AZs. Likewise, the extraction module 220 can be distributed in the same Virtual Private Cloud (VPC) or in multiple VPCs. These multiple computing devices can be any combination of computing devices such as servers, ASICs, PLDs, CPLDs, FPGAs, and GALs.

[0127] Based on the preceding description of Figure 1, the following describes an embodiment of the document content review method provided in this application.

[0128] Please refer to Figure 5, which is a flowchart of a document content review method provided in this application, including steps S501 to S504.

[0129] S501, The document content review system 200 obtains the first document, which is the document that needs to be reviewed.

[0130] Optionally, the first file may be sent by the user to the file content review system 200 through the client 100, or it may be obtained by the file content review system 200 from other modules or systems (such as file systems, database systems, etc.). This application does not limit this.

[0131] Optionally, the first file can be a plain text file (TXT), a Word document (DOCX), a portable document format (PDF) document, a presentation (PowerPoint, PPT), a spreadsheet (EXCEL) file, or an image format (such as JPG, PNG, GIF, etc.), etc. This application does not limit the file type of the first file.

[0132] S502, The document content review system 200 extracts the first sentence from the first document based on the first keyword.

[0133] Optionally, the document content review system 200 obtains scanning rules. These rules instruct the extraction of statements that meet the specified statement range. The first document is then divided into multiple statements that meet this range (including the first statement), and the first statement is extracted based on the first keyword. The aforementioned statement range can be a single sentence (marked with a period), multiple consecutive sentences, a paragraph (marked with a newline character), X consecutive paragraphs (X being a positive integer), Y consecutive characters (Y being a positive integer), etc. This application does not impose specific limitations on this. A specific statement range can be specified in the scanning rules, or it can be left unspecified (in which case the document content review system 200 uses the default statement range). This application does not impose any limitations on this either.

[0134] Optionally, the scanning rules and the first statement must satisfy one of the following:

[0135] (1) The scanning rules include a first keyword and a first identification type. The first identification type is used to indicate the extraction of statements that contain the first keyword and meet the statement range. The first statement contains the first keyword.

[0136] (2) The scanning rules include multiple keywords and a second identification type. The second identification type is used to indicate the extraction of statements that contain these multiple keywords and meet the statement range. The first statement contains these multiple keywords.

[0137] (3) The scanning rules include a keyword set and a third identification type. The keyword set includes a first keyword. The third identification type is used to indicate that statements containing the number of keywords in the keyword set are greater than or equal to the hit number threshold and meet the statement range. The hit number threshold is less than the number of keywords in the keyword set. The first statement contains the first keyword. The number of keywords in the first statement containing the keyword set is greater than or equal to the hit number threshold.

[0138] The above (1) to (3) are actually three ways to set the scanning rules. Method (1) specifies a keyword and a first identification type in the scanning rule. In this case, the scanning rule means that the statement containing the keyword and satisfying the statement range will be extracted. That is, for any statement in the file that satisfies the statement range, as long as the statement contains the keyword specified in the scanning rule, the statement can be extracted. Method (2) specifies multiple keywords and a second identification type in the scanning rule. In this case, the scanning rule means that the statement containing multiple keywords and satisfying the statement range will be extracted. That is, for a statement in the file that satisfies the statement range, the statement will only be extracted if the statement contains multiple keywords specified in the scanning rule. Method (3) specifies a keyword set (including N keywords, where N is a positive integer greater than 1) and a third identification type in the scanning rule. In this case, the scanning rule means that statements containing M keywords from these N keywords (M is the hit threshold, where M is a positive integer and M is less than N) and satisfying the statement range will be extracted. That is, for a statement in the file that satisfies the statement range, as long as any M keywords from these N keywords appear in the statement, the statement can be extracted. M can be given in the scanning rule or can be the default setting of the file content review system 200 (such as default M=NP, where P is a positive integer). This application does not limit this.

[0139] Optionally, the scanning rules may also include auxiliary judgment information, which indicates the professional field and / or regulations corresponding to the scanning rules. For details on auxiliary judgment information, please refer to the previous section; it will not be repeated here.

[0140] S503, The document content review system 200 determines the second statement based on the first statement. The similarity between the second statement and the first statement meets the similarity threshold. The risk information of the second statement is used as the risk information of the first statement. The risk information is used to indicate whether the corresponding statement is risky content.

[0141] It should be understood that the first statement is a statement extracted from a document requiring content review, and the risk information of the first statement needs to be determined. The document content review system 200 stores multiple statement samples (including the second statement) that have been marked with risk information. The risk information is used to indicate whether the corresponding statement sample is risky content.

[0142] The document content review system 200 can determine the risk information of a first statement by calculating the similarity between the first statement and these statement samples. If the similarity between the first statement and the second statement meets a similarity threshold, the second statement is considered a similar statement to the first statement, and the risk information of the second statement is then used as the risk information of the first statement, which is equivalent to labeling the first statement and the second statement with the same risk information label. In other words, if the risk information labeled on the second statement indicates that the second statement is risky content, the semantic recognition model 260 determines that the first statement is also risky content; if the risk information labeled on the second statement indicates that the second statement is not risky content, the semantic recognition model 260 determines that the first statement is also not risky information.

[0143] Optionally, before step S503, the document content review system 200 may determine the aforementioned similarity threshold as follows: First, obtain multiple statements (including the second statement mentioned above), which have the same risk information. Then, convert these multiple statements into multiple feature vectors, with each statement corresponding to one of the feature vectors. For details on how to convert statements into corresponding feature vectors, please refer to the previous description; it will not be repeated here. Next, the document content review system 200 calculates the similarity between these multiple statements based on these feature vectors, obtaining multiple similarities. Then, it determines the similarity threshold based on these multiple similarities, such as using the average of the multiple similarities as the similarity threshold, or using the minimum value among the multiple similarities as the similarity threshold, etc.

[0144] Optionally, all of the aforementioned statements contain keywords belonging to the first word category. The first statement includes the first keyword, and the first keyword also belongs to the first word category. In other words, the document content review system 200 distinguishes between different word categories. It calculates the similarity between statements that have the same risk information and contain keywords belonging to the same word category. Then, based on the calculated similarity, it determines the similarity threshold corresponding to that word category. Each word category can calculate its corresponding similarity threshold in a similar manner. When the document content review system 200 extracts the first statement containing the first keyword from the document, it obtains other statements with the same keyword category and risk information based on the word category to which the first keyword belongs. Then, it calculates the similarity between the first statement and these other statements, and compares the calculated similarity with the similarity threshold corresponding to that word category, thereby accurately determining the risk information of the first statement.

[0145] Regarding how the document content review system 200 determines the risk information of statements, please refer to the semantic recognition model 260 introduced above, which will not be repeated here.

[0146] S504. The document content review system 200 sends the review result of the first document to the user. The review result of the first document includes the first statement and the risk information of the first statement.

[0147] It should be understood that, in addition to the risk information in the first statement, the examination results of the first document may also include risk information in other statements within the first document. The methods for extracting risk information from other statements and determining risk information are similar to those for the first statement, and will not be elaborated upon here. This application does not limit the presentation of risk information.

[0148] For example, as described above, the first statement includes a first keyword. If the risk information received by client 100 from the document content review system 200 indicates that the first statement is risky content, then client 100 can mark the first keyword in the first statement of the first document in red when displaying the first document, thereby indicating that the first statement containing the first keyword is risky content. If the risk information of the first statement indicates that the first statement is not risky content, then client 100 can mark the first keyword in the first statement of the first document in green when displaying the first document, thereby indicating that the first statement containing the first keyword is not risky content. Other statements in the first document can also be presented in a similar manner to indicate whether the statement is risky content.

[0149] It should be noted that, in addition to using different colors to distinguish between risky and non-risky content, different fonts, symbols or other methods can also be used to distinguish between risky and non-risky content, and this application does not make any specific restrictions on this.

[0150] Optionally, the document review result of the first document also includes auxiliary judgment information, which is derived from the scanning rules in step S501. This auxiliary judgment information is used to indicate the professional field and / or regulations corresponding to the scanning rules. Combining the auxiliary judgment information can help users better understand the risk information in the first statement.

[0151] Optionally, the document content review system 200 can also obtain a second file, which also requires content review. If the file format of the second file does not conform to the file type configured in the document content review system 200, the scanning time of the second file exceeds the first time threshold, or the waiting time of the second file exceeds the second time threshold, the system will notify the user that the second file cannot be processed temporarily. This is similar to the circuit breaker protection mechanism described earlier, and will not be elaborated upon here.

[0152] In summary, in the document content review method provided in this application, the document content review system 200 first extracts sentences from the document based on keywords, and then calculates the similarity between the extracted sentences and other sentences marked with risk information. If the similarity between the extracted sentence and a sentence marked with risk information meets the similarity threshold, the two sentences are considered similar sentences, and the risk information of the sentence marked with risk information is then used as the risk information of the currently extracted sentence. This method can effectively identify risky content in the document.

[0153] Referring to Figure 6, this application also provides a computing device 600, including a bus 602, a processor 604, a memory 606, and a communication interface 608. The processor 604, the memory 606, and the communication interface 608 communicate with each other via the bus 602. The computing device 600 can be a server, laptop, tablet, desktop computer, edge device, smartphone, smart screen, etc., and this application does not specifically limit it, nor does it limit the number of processors and memories in the computing device 600.

[0154] Bus 602 can be a Peripheral Component Interconnect (PCI) bus or an Extended Industry Standard Architecture (EISA) bus, etc. Buses can be categorized as address buses, data buses, control buses, etc. For ease of illustration, only one line is used in Figure 6, but this does not imply that there is only one bus or one type of bus. Bus 602 can include pathways for transmitting information between various components of computing device 600 (e.g., memory 606, processor 604, communication interface 608).

[0155] Processor 604 may include any one or more processors such as a central processing unit (CPU), a graphics processing unit (GPU), a microprocessor (MP), or a digital signal processor (DSP).

[0156] Memory 606 may include volatile memory, such as random access memory (RAM). Processor 604 may also include non-volatile memory, such as read-only memory (ROM), flash memory, hard disk drive (HDD), or solid state drive (SSD).

[0157] The memory 606 stores executable program code. The processor 604 executes the executable program code to implement the functions of the acquisition module 210, extraction module 220, determination module 230, transmission module 240 and training module 250 in FIG1, respectively, thereby implementing the method steps in FIG5 of this application.

[0158] The communication interface 608 uses transceiver modules, such as, but not limited to, network interface cards and transceivers, to enable communication between the computing device 600 and other devices or communication networks.

[0159] It should be understood that the computing device 600 according to this application can correspond to the document content review system 200 shown in FIG1 of this application, and each module of the computing device 600 is used to implement the operation steps in the document content review method of FIG5. In addition, the computing device 600 can correspond to the document content review system 200 shown in FIG1 and is used to implement the functions of each module in the document content review system 200. For the sake of brevity, it will not be described in detail here.

[0160] As one possible implementation, the computing device 600 may also include a chip system, which includes a processor and a power supply circuit. The power supply circuit supplies power to the processor, which executes the operation steps in the document content review method of Figure 5. For simplicity, details are omitted here. The processor can be implemented using a CPU, or it can be implemented using computing devices or AI chips such as GPUs, DPUs, NPUs, XPUs, SoCs, offloading cards, or accelerator cards.

[0161] As one possible implementation, the computing device 600 may include multiple types of processors 604, i.e., the computing device 600 is a heterogeneous device. For example, the computing device 600 may include a CPU and a GPU, and the operation steps in the method of FIG5 may be executed by at least one of the processors 604. For the sake of brevity, further details are omitted here.

[0162] As shown in Figure 7, this application also provides a computing device cluster, which includes at least one computing device 600. The memory 606 of one or more computing devices 600 in the computing device cluster may store the same instructions for implementing the method of Figure 5.

[0163] In some possible implementations, the memory 606 of one or more computing devices 600 in the computing device cluster may also store partial instructions for implementing the method of FIG5. In other words, a combination of one or more computing devices 600 can jointly execute the instructions for implementing the method of FIG5.

[0164] It should be noted that the memory 606 in different computing devices 600 within the computing device cluster can store different instructions, which are used to execute some functions of the file content review system 200 in Figure 1. That is, the instructions stored in the memory 606 of different computing devices 600 can implement the functions of one or more modules among the acquisition module 210, extraction module 220, determination module 230, sending module 240, and training module 250.

[0165] In some possible implementations, one or more computing devices in a computing device cluster can be connected via a network. This network can be a wide area network (WAN) or a local area network (LAN), etc. Figure 8 illustrates one possible implementation. As shown in Figure 8, two computing devices 600A and 600B are connected via a network. Specifically, they are connected to the network through communication interfaces in each computing device. In this type of possible implementation, the memory 606 in computing device 600A stores instructions for executing the functions of the acquisition module 210 and the transmission module 240. Simultaneously, the memory 606 in computing device 600B stores instructions for executing the functions of the extraction module 220, the determination module 230, and the training module 250.

[0166] It should be understood that the function of computing device 600A shown in Figure 8 can also be performed by multiple computing devices 600. Similarly, the function of computing device 600B can also be performed by multiple computing devices 600.

[0167] This application also provides another computing device cluster. The connection relationship between the computing devices in this computing device cluster can be similar to the connection method of the computing device cluster shown in FIG8. The difference is that the memory 606 in one or more computing devices 600 in this computing device cluster can store the same instructions for implementing the method of FIG5 above.

[0168] In some possible implementations, the memory 606 of one or more computing devices 600 in the computing device cluster may also store partial instructions for implementing the method of FIG5. In other words, a combination of one or more computing devices 600 can jointly execute the instructions for implementing the method of FIG5.

[0169] This application also provides a computer-readable storage medium. The computer-readable storage medium can be any available medium that a computing device can store, or a data storage device such as a data center containing one or more available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium (e.g., solid-state drive). The computer-readable storage medium includes instructions that instruct the computing device to perform the method steps in FIG5.

[0170] This application also provides a computer program product containing instructions. The computer program product may be a software or program product containing instructions, capable of running on a computing device or stored on any usable medium. When the computer program product is run on at least one computing device, it causes the at least one computing device to perform the operational steps in the document content review method of FIG5.

[0171] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the protection scope of the technical solutions of the embodiments of this application.

Claims

1. A method for reviewing document content, characterized in that, The method is executed by a document content review system, and the method includes: Obtain the first file, which is the file that requires content review; Extract the first statement from the first file based on the first keyword; The second statement is determined based on the first statement, and the similarity between the second statement and the first statement meets the similarity threshold. The risk information of the second statement is used as the risk information of the first statement, and the risk information is used to indicate whether the corresponding statement is risky content.

2. The method according to claim 1, characterized in that, Before determining the second statement based on the first statement, the method further includes: Obtain multiple statements, including the second statement, wherein the multiple statements have the same risk information; The multiple statements are converted into multiple feature vectors, and the multiple statements and the multiple feature vectors correspond one-to-one; The similarity between each pair of the multiple statements is calculated based on the multiple feature vectors to obtain multiple similarity values; The similarity threshold is determined based on the multiple similarities.

3. The method according to claim 2, characterized in that, The plurality of statements all contain keywords belonging to the first word category, the first statement includes the first keyword, and the first keyword belongs to the first word category.

4. The method according to any one of claims 1 to 3, characterized in that, The step of extracting the first statement from the first file based on the first keyword includes: Obtain scanning rules, which are used to indicate the extraction of statements that meet the statement range; The first file is divided into multiple statements that satisfy the statement range, and the multiple statements that satisfy the statement range include the first statement; Extract the first statement based on the first keyword.

5. The method according to claim 4, characterized in that, The scanning rule and the first statement satisfy one of the following: The scanning rules include the first keyword and the first identification type. The first identification type is used to indicate the extraction of statements that contain the first keyword and meet the statement range. The first statement contains the first keyword. The scanning rules include the plurality of keywords and the second identification type. The plurality of keywords include the first keyword. The second identification type is used to indicate the extraction of statements that contain the plurality of keywords and satisfy the statement range. The first statement contains the plurality of keywords. The scanning rules include a keyword set and a third identification type. The keyword set includes the first keyword. The third identification type is used to indicate the extraction of statements that contain a number of keywords from the keyword set that are greater than or equal to a hit count threshold and satisfy the statement range. The hit count threshold is less than the number of keywords in the keyword set. The first statement contains a number of keywords from the keyword set that are greater than or equal to the hit count threshold.

6. A document content review system, characterized in that, include: The acquisition module is used to acquire the first file, which is the file that needs content review. The extraction module is used to extract a first statement from the first file based on a first keyword; The determining module is used to determine a second statement based on the first statement, wherein the similarity between the second statement and the first statement meets a similarity threshold; The determining module is further configured to use the risk information of the second statement as the risk information of the first statement, wherein the risk information is used to indicate whether the corresponding statement is risky content.

7. The system according to claim 6, characterized in that, The acquisition module is further configured to: acquire multiple statements including the second statement, wherein the multiple statements have the same risk information; The determining module is also used for: The multiple statements are converted into multiple feature vectors, and the multiple statements and the multiple feature vectors correspond one-to-one; The similarity between each pair of the multiple statements is calculated based on the multiple feature vectors to obtain multiple similarity values; The similarity threshold is determined based on the multiple similarities.

8. The system according to claim 7, characterized in that, The plurality of statements all contain keywords belonging to the first word category, the first statement includes the first keyword, and the first keyword belongs to the first word category.

9. The system according to any one of claims 6 to 8, characterized in that, The extraction module is specifically used for: Obtain scanning rules, which are used to indicate the extraction of statements that meet the statement range; The first file is divided into multiple statements that satisfy the statement range, and the multiple statements that satisfy the statement range include the first statement; Extract the first statement based on the first keyword.

10. The system according to claim 9, characterized in that, The scanning rule and the first statement satisfy one of the following: The scanning rules include the first keyword and the first identification type. The first identification type is used to indicate the extraction of statements that contain the first keyword and meet the statement range. The first statement contains the first keyword. The scanning rules include the plurality of keywords and the second identification type. The plurality of keywords include the first keyword. The second identification type is used to indicate the extraction of statements that contain the plurality of keywords and satisfy the statement range. The first statement contains the plurality of keywords. The scanning rules include a keyword set and a third identification type. The keyword set includes the first keyword. The third identification type is used to indicate the extraction of statements that contain a number of keywords from the keyword set that are greater than or equal to a hit count threshold and satisfy the statement range. The hit count threshold is less than the number of keywords in the keyword set. The first statement contains a number of keywords from the keyword set that are greater than or equal to the hit count threshold.

11. A computing device, characterized in that, The computing device includes a processor and memory; The processor is configured to execute instructions stored in the memory to cause the computing device to perform the method as described in any one of claims 1 to 5.

12. A computing device cluster, characterized in that, It includes at least one computing device, each computing device including a processor and memory; The processor of the at least one computing device is configured to execute instructions stored in the memory of the at least one computing device to cause the cluster of computing devices to perform the method as described in any one of claims 1 to 5.

13. A chip system, characterized in that, The chip system includes a processor and a power supply circuit, the power supply circuit being used to supply power to the processor, the processor being used to perform the method as described in any one of claims 1 to 5.

14. A computer program product containing instructions, characterized in that, When the instruction is executed by the computing device cluster, the computing device cluster performs the method as described in any one of claims 1 to 5.

15. A computer-readable storage medium, characterized in that, Includes computer program instructions, which, when executed by a cluster of computing devices, perform the method as described in any one of claims 1 to 5.

Images