Text auditing method and device, storage medium and program product

Abstract

This application relates to a text review method, apparatus, storage medium, and program product. The method includes: retrieving preliminary review historical cases similar to the target text to be reviewed from a historical review case database; assembling the target text and preliminary review historical cases into preliminary review prompts, and inputting the preliminary review prompts into a preliminary review model to obtain preliminary review results; extracting key target information from the preliminary review results, and retrieving secondary review historical cases similar to the key target information from the historical review case database; for each target historical case in the union of target cases consisting of preliminary and secondary review historical cases, classifying it as a positive sample if its historical review results meet the consistency condition with the preliminary review results, otherwise classifying it as a negative sample; assembling the target text, preliminary review results, positive samples, and negative samples into secondary review prompts, and inputting the secondary review prompts into a secondary review model to obtain secondary review results. This method can improve the accuracy of text review.

Description

Technical Field

[0001] This application relates to the field of computer technology, and in particular to a text review method, apparatus, storage medium, and program product. Background Technology

[0002] Currently, the most commonly used text moderation methods in the industry are "text moderation methods based on direct prompt word engineering of large models" and "text moderation methods based on single-path vector retrieval enhanced generation (RAG)". The direct prompt word method directly concatenates a lengthy moderation rule manual and a few examples into the prompt, requiring the large model to directly output the moderation result. The traditional RAG method stores historical violating texts in a vector database; when new text is input, it retrieves the original texts of the k most similar texts, concatenates them, and then feeds them to the large model for judgment. However, traditional prompt word schemes rely on manually summarized rules. When faced with new variants of violations such as slang, code words, and puns, static rules cannot cover them, leading to missed judgments. Traditional RAG only retrieves similar "text fragments" or "rule entries," lacking the logical chain of "why this judgment is made" provided by human moderators. Large models are prone to improvising (or "hallucinations") in complex contexts. Furthermore, traditional techniques are all single-step large model inference outputs, lacking a confirmation mechanism. In summary, the accuracy of traditional text moderation methods is low. Summary of the Invention

[0003] Therefore, it is necessary to provide a text review method, device, storage medium, and program product that can improve the accuracy of text review in order to address the above-mentioned technical problems.

[0004] Firstly, this application provides a text review method, the method comprising:

[0005] Obtain the target text to be reviewed;

[0006] Retrieve at least one preliminary review case that is similar to the target text from a pre-built historical review case library;

[0007] The target text and the historical cases of the preliminary review are assembled into preliminary review prompts, and the preliminary review prompts are input into the preliminary review model to obtain the preliminary review results;

[0008] Extract the target key information from the preliminary review results, and retrieve at least one review history case that is similar to the target key information from the historical review case database;

[0009] For each target historical case in the target case set consisting of the preliminary review historical cases and the re-review historical cases, if the historical review result of the target historical case meets the conclusion consistency condition with the preliminary review result, then the target historical case is classified as a positive sample supporting the preliminary review result; otherwise, it is classified as a negative sample opposing the preliminary review result.

[0010] The target text, the initial review result, the positive sample, and the negative sample are assembled into a review prompt word, and the review prompt word is input into the review model to obtain the review result.

[0011] In one embodiment, the method further includes:

[0012] Extract unstructured first historical review data from the review system logs; the first historical review data includes the original text, the first violation conclusion of the original text, and the original category label;

[0013] Based on the first historical audit data and a pre-built security audit rule base, extract prompt words are assembled and input into the large extraction model to obtain structured second historical audit data; the second historical audit data includes the original text, the first violation conclusion, the analysis category label of the original text, the first violation point, and the first audit reasoning information;

[0014] Extract key historical information from the second historical review data; the key historical information includes at least one of the first violation conclusion, the analysis category label, the first violation point, and the first review reasoning information.

[0015] Based on the second historical audit data and the aforementioned key historical information, a historical audit case library is constructed.

[0016] In one embodiment, the historical key information includes the first violation point and the first audit reasoning information;

[0017] The construction of a historical audit case library based on the second historical audit data and the historical key information includes:

[0018] The feature vectors of the original text in the first review reasoning information and the second historical review data are extracted respectively to obtain historical feature vectors, and a vector index is constructed based on the historical feature vectors;

[0019] The original text and the first violation point are segmented into words to obtain historical word segments, and an inverted index is constructed based on the historical word segments.

[0020] The second historical audit data, the vector index, and the inverted index are stored together to construct a historical audit case library.

[0021] In one embodiment, retrieving at least one preliminary review historical case similar to the target text from a pre-built historical review case library includes:

[0022] Extract the text feature vector of the target text;

[0023] Based on the text feature vector and the vector index, a vector index retrieval is performed in a pre-built historical review case library to select at least one second historical review data similar to the target text as a first preliminary review recall case.

[0024] The target text is segmented to obtain the target text word segmentation;

[0025] Based on the target text segmentation and the inverted index, an inverted index retrieval is performed in the historical review case library to select at least one second historical review data similar to the target text as a second preliminary review recall case.

[0026] From the set of preliminary review cases consisting of the first preliminary review recall case and the second preliminary review recall case, select at least one preliminary review historical case that is similar to the target text.

[0027] In one embodiment, selecting at least one historical preliminary review case similar to the target text from the preliminary review case set consisting of the first preliminary review recall case and the second preliminary review recall case includes:

[0028] The target text and the set of preliminary cases consisting of the first preliminary review recall case and the second preliminary review recall case are input into the re-ranking model to obtain the first case sequence; each recall case in the first case sequence is arranged in descending order of its similarity to the target text.

[0029] From the first case sequence, select at least one preliminary review historical case that is most similar to the target text.

[0030] In one embodiment, the preliminary review result includes a second violation conclusion, a preliminary review category label, a second violation point, and second review reasoning information for the target text; the target key information includes the second violation point and the second review reasoning information.

[0031] The step of retrieving at least one historical review case similar to the target key information from the historical review case database includes:

[0032] Extract the feature vector of the violation point of the second violation point, and extract the inference feature vector of the second review inference information;

[0033] The second violation point is processed by word segmentation to obtain the violation point word segmentation;

[0034] Based on the inference feature vector and the vector index, a vector index retrieval is performed in the historical review case library to select at least one second historical review data that is similar to the second review inference information as a first review recall case.

[0035] Based on the violation point feature vector and the vector index, a vector index retrieval is performed in the historical review case library to select at least one second historical review data similar to the second violation point as a second review recall case.

[0036] Based on the violation point word segmentation and the inverted index, an inverted index retrieval is performed in the historical review case library to select at least one second historical review data similar to the second violation point as a third review recall case;

[0037] From the set of review cases consisting of the first review recall case, the second review recall case, and the third review recall case, at least one historical review case that is similar to the target key information is selected.

[0038] In one embodiment, selecting at least one historical review case similar to the target key information from the review case set consisting of the first review recall case, the second review recall case, and the third review recall case includes:

[0039] The target text and the union of the review cases consisting of the first review recall case, the second review recall case, and the third review recall case are input into the re-ranking model to obtain the second case sequence; each recall case in the second case sequence is arranged in descending order of its similarity to the target text.

[0040] From the second case sequence, select at least one review history case that is most similar to the target text.

[0041] In one embodiment, the historical review result includes a first violation conclusion, a historical category label, a first violation point, and first review reasoning information of the original text corresponding to the historical review result; the preliminary review result includes a second violation conclusion, a preliminary review category label, a second violation point, and second review reasoning information of the target text; the method further includes:

[0042] For each target historical case, if the first violation conclusion in the historical review result of the target historical case is consistent with the second violation conclusion in the preliminary review result, then the historical review result and the preliminary review result are determined to meet the conclusion consistency condition.

[0043] Secondly, this application provides a text review device, the device comprising:

[0044] The acquisition module is used to acquire the target text to be reviewed;

[0045] The retrieval module is used to retrieve at least one preliminary review historical case that is similar to the target text from a pre-built historical review case library;

[0046] The preliminary review module is used to assemble the target text and the preliminary review history cases into preliminary review prompts, and input the preliminary review prompts into the preliminary review model to obtain the preliminary review results.

[0047] The retrieval module is also used to extract target key information from the preliminary review results and retrieve at least one review history case similar to the target key information from the historical review case database;

[0048] The segmentation module is used to classify each target historical case in the target case set consisting of the preliminary review historical cases and the re-review historical cases as a positive sample supporting the preliminary review result if the historical review result of the target historical case meets the conclusion consistency condition with the preliminary review result; otherwise, it is classified as a negative sample opposing the preliminary review result.

[0049] The review module is used to assemble the target text, the initial review result, the positive sample, and the negative sample into review prompt words, and input the review prompt words into the review model to obtain the review result.

[0050] Thirdly, this application provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps in the various method embodiments of this application.

[0051] Fourthly, this application provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps in the various method embodiments of this application.

[0052] Fifthly, this application provides a computer program product, including a computer program that, when executed by a processor, implements the steps in the various method embodiments of this application.

[0053] The aforementioned text review method, apparatus, storage medium, and program products acquire the target text to be reviewed; retrieve at least one preliminary review historical case similar to the target text from a pre-built historical review case library; assemble the target text and the preliminary review historical case into preliminary review prompts, and input the preliminary review prompts into a preliminary review model to obtain preliminary review results; extract the target key information from the preliminary review results, and retrieve at least one secondary review historical case similar to the target key information from the historical review case library; for each target historical case in the target case union set consisting of preliminary review historical cases and secondary review historical cases, if the historical review results of the target historical case meet the conclusion consistency condition with the preliminary review results, then the target historical case is classified as a positive sample supporting the preliminary review results; otherwise, it is classified as a negative sample opposing the preliminary review results; assemble the target text, preliminary review results, positive samples, and negative samples into secondary review prompts, and input the secondary review prompts into a secondary review model to obtain secondary review results. Compared with traditional text review methods, this application improves the accuracy of text review through a two-stage retrieval and review mechanism. First, historical cases are introduced as references during the initial review stage to compensate for the shortcomings of static rules in covering new variations such as slang and homophones. Second, by extracting key information from the initial review and retrieving positive and negative samples, review prompts containing "judgment logic" and "opposite cases" are constructed to guide the large model to perform secondary verification based on a reasoning chain similar to that of a human reviewer, effectively suppressing model illusions. This application's solution forms a closed loop of "initial review case reference - positive and negative example logic verification," significantly reducing missed judgments and misjudgments, overcoming the drawbacks of traditional single-step reasoning lacking a confirmation mechanism, and improving the accuracy of text review. Attached Figure Description

[0054] To more clearly illustrate the technical solutions in the embodiments of this application or related technologies, the drawings used in the description of the embodiments of this application or related technologies will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0055] Figure 1 This is a diagram illustrating the application environment of a text review method in one embodiment.

[0056] Figure 2 This is a flowchart illustrating a text review method in one embodiment;

[0057] Figure 3 This is a flowchart illustrating the construction process of a historical review case library in one embodiment;

[0058] Figure 4 This is a flowchart illustrating the retrieval process of historical cases during the initial review in one embodiment.

[0059] Figure 5 This is a flowchart illustrating the retrieval of historical review cases and the division of positive and negative samples during the review process in one embodiment.

[0060] Figure 6 This is a flowchart illustrating the text review method in another embodiment;

[0061] Figure 7 This is a structural block diagram of a text review device in one embodiment;

[0062] Figure 8 This is an internal structural diagram of a computer device in one embodiment;

[0063] Figure 9 This is a diagram of the internal structure of a computer device in another embodiment. Detailed Implementation

[0064] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0065] The text review method provided in this application can be applied to, for example... Figure 1 In the application environment shown, terminal 102 communicates with server 104 via a network. The data storage system can be set up separately and can store the data that server 104 needs to process. The data storage system can be integrated into server 104 or placed in the cloud or on other servers. Terminal 102 can be, but is not limited to, various desktop computers, laptops, smartphones, tablets, and in-vehicle terminals. Server 104 can be a standalone physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server providing cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, cloud security, host security and other network security services, CDN, and basic cloud computing services such as big data and artificial intelligence platforms. Terminal 102 and server 104 can be directly or indirectly connected via wired or wireless communication, which is not limited herein.

[0066] Server 104 can obtain the target text to be reviewed from terminal 102; server 104 can retrieve at least one preliminary review historical case similar to the target text from a pre-built historical review case library; server 104 can assemble the target text and the preliminary review historical case into preliminary review prompts and input the preliminary review prompts into the preliminary review model to obtain the preliminary review result; server 104 can extract the target key information from the preliminary review result and retrieve at least one re-review historical case similar to the target key information from the historical review case library; for each re-review historical case, if the historical review result of the re-review historical case meets the conclusion consistency condition with the preliminary review result, server 104 can classify the re-review historical case as a positive sample supporting the preliminary review result, otherwise it can classify it as a negative sample opposing the preliminary review result; server 104 can assemble the target text, preliminary review result, positive sample, and negative sample into re-review prompts and input the re-review prompts into the re-review model to obtain the re-review result. Server 104 can feed back the re-review result to terminal 102.

[0067] It is understood that this embodiment does not impose any limitations on this. Figure 1 The application scenarios shown are for illustrative purposes only and are not limited to these.

[0068] In one embodiment, such as Figure 2 As shown, a text review method is provided. This method can be applied to computer devices, which can be terminals or servers. That is, the method can be executed independently by the terminal or the server, or it can be implemented through interaction between the terminal and the server. This embodiment illustrates the method applied to a computer device as an example, including the following steps:

[0069] Step 202: Obtain the target text to be reviewed.

[0070] Step 204: Retrieve at least one preliminary review historical case that is similar to the target text from the pre-built historical review case library.

[0071] The historical review case library is a knowledge base that can be searched online, formed by structuring closed review data and associating it with query indexes. Preliminary review historical cases are historical review cases in the historical review case library that are similar to the target text.

[0072] In one embodiment, each historical review case in the historical review case library is stored in conjunction with its corresponding query index. A computer device can retrieve at least one preliminary review historical case similar to the target text from this historical review case library based on the target text and the query index.

[0073] Step 206: Assemble the target text and the historical cases of the preliminary review into preliminary review prompts, and input the preliminary review prompts into the preliminary review model to obtain the preliminary review results.

[0074] Step 208: Extract the target key information from the preliminary review results, and retrieve at least one historical review case that is similar to the target key information from the historical review case database.

[0075] The target key information in the preliminary review results refers to the core conclusion elements output by the preliminary review model that characterize its review judgment basis. The review history cases are historical review cases in the historical review case library that are similar to the target key information.

[0076] In one embodiment, each historical audit case in the historical audit case database is stored in conjunction with its corresponding query index. A computer device can retrieve at least one historical audit case similar to the target text from this historical audit case database based on the target key information and the query index.

[0077] Step 210: For each target historical case in the target case set consisting of preliminary review historical cases and re-review historical cases, if the historical review results of the target historical case meet the conclusion consistency condition with the preliminary review results, then the target historical case is classified as a positive sample supporting the preliminary review results; otherwise, it is classified as a negative sample opposing the preliminary review results.

[0078] The target case union is the union of preliminary review historical cases and secondary review historical cases. Each target case union contains at least two target historical cases. A positive sample is a target historical case in the target case union whose historical review results and preliminary review results meet the condition of consistency. A negative sample is a target historical case in the target case union whose historical review results and preliminary review results do not meet the condition of consistency.

[0079] Specifically, the computer equipment can merge and deduplicate historical cases from the initial review and the subsequent review to form a union of target cases. For each target historical case in the union of target cases, if the historical review result of the target historical case meets the consistency condition with the initial review result, then the target historical case is classified as a positive sample supporting the initial review result; if the historical review result of the target historical case does not meet the consistency condition with the initial review result, then the target historical case is classified as a negative sample opposing the initial review result.

[0080] Step 212: Assemble the target text, initial review results, positive samples, and negative samples into review prompt words, and input the review prompt words into the review model to obtain the review results.

[0081] It is understandable that the review result may either remain the same as the initial review result, or it may be a revised version of the initial review result.

[0082] In the aforementioned text review method, the target text to be reviewed is obtained; at least one preliminary review historical case similar to the target text is retrieved from a pre-built historical review case library; the target text and the preliminary review historical case are assembled into preliminary review prompts, and the preliminary review prompts are input into a preliminary review model to obtain preliminary review results; key target information is extracted from the preliminary review results, and at least one secondary review historical case similar to the key target information is retrieved from the historical review case library; for each target historical case in the union of target cases consisting of preliminary review historical cases and secondary review historical cases, if the historical review results of the target historical case meet the conclusion consistency condition with the preliminary review results, the target historical case is classified as a positive sample supporting the preliminary review results; otherwise, it is classified as a negative sample opposing the preliminary review results; the target text, preliminary review results, positive samples, and negative samples are assembled into secondary review prompts, and the secondary review prompts are input into a secondary review model to obtain secondary review results. Compared with traditional text review methods, this application improves the accuracy of text review through a two-stage retrieval and review mechanism. First, historical cases are introduced as references during the initial review stage to compensate for the shortcomings of static rules in covering new variations such as slang and homophones. Second, by extracting key information from the initial review and retrieving positive and negative samples, review prompts containing "judgment logic" and "opposite cases" are constructed to guide the large model to perform secondary verification based on a reasoning chain similar to that of a human reviewer, effectively suppressing model illusions. This application's solution forms a closed loop of "initial review case reference - positive and negative example logic verification," significantly reducing missed judgments and misjudgments, overcoming the drawbacks of traditional single-step reasoning lacking a confirmation mechanism, and improving the accuracy of text review.

[0083] In one embodiment, the method further includes: extracting unstructured first historical audit data from audit system logs; the first historical audit data includes original text, a first violation conclusion of the original text, and an original category label; assembling extracted prompt words based on the first historical audit data and a pre-built security audit rule base, and inputting the extracted prompt words into an extraction model to obtain structured second historical audit data; the second historical audit data includes original text, a first violation conclusion, an analysis category label of the original text, a first violation point, and first audit reasoning information; extracting historical key information from the second historical audit data; the historical key information includes at least one of the first violation conclusion, analysis category label, first violation point, and first audit reasoning information; and constructing a historical audit case library based on the second historical audit data and the historical key information.

[0084] The violation conclusions include both "normal" and "violation." Category labels indicate the category to which the text belongs. Original category labels are the category labels originally annotated in the first historical review data; for example, the original category labels could be the machine review category labels from the initial machine review. Analytical category labels are the category labels extracted after analysis by the large-scale model. It's understandable that the original category labels may not be accurate enough; more accurate analytical category labels can be obtained after large-scale model analysis. Violation points refer to specific content fragments or feature descriptions in the text that violate the review rules. Review reasoning information refers to the logical chain and judgment reasons on which the large-scale model arrives at its review conclusion.

[0085] In the above embodiments, a high-quality historical review case library is constructed by extracting structured information and annotating key information from unstructured historical review logs. Compared with the original logs, the case library integrates deeper logic such as violation points and reasoning information, providing more accurate and interpretable knowledge support for subsequent retrieval, thereby improving the accuracy and traceability of automated review.

[0086] In one embodiment, the historical key information includes a first violation point and a first review reasoning information; based on the second historical review data and the historical key information, a historical review case library is constructed, including: extracting feature vectors from the original text in the first review reasoning information and the second historical review data respectively to obtain historical feature vectors, and constructing a vector index based on the historical feature vectors; performing word segmentation on the original text and the first violation point respectively to obtain historical word segmentation, and constructing an inverted index based on the historical word segmentation; and storing the second historical review data, the vector index, and the inverted index together to construct the historical review case library.

[0087] Among them, historical word segmentation is the process of segmenting the original text and the first violation point to obtain characters or words.

[0088] In one embodiment, such as Figure 3As shown, the computer device can extract unstructured first historical audit data from the audit system logs; the first historical audit data includes the original text, the first violation conclusion of the original text, and the original category label; based on the first historical audit data and a pre-built security audit rule base, extract prompt words are assembled and input into the extraction model to obtain structured second historical audit data; the second historical audit data includes the original text, the first violation conclusion, the analysis category label of the original text, the first violation point, and the first audit reasoning information; historical key information is extracted from the second historical audit data; the historical key information includes the first violation point and the first audit reasoning information; feature vectors of the original text in the first audit reasoning information and the second historical audit data are extracted respectively through the text feature model to obtain historical feature vectors, and a vector index is constructed based on the historical feature vectors; the original text and the first violation point are segmented separately to obtain historical word segments, and an inverted index is constructed based on the historical word segments using BM25 (Best Matching 25, a probabilistic retrieval algorithm used to calculate the relevance score between a document and the query text); the second historical audit data, the vector index, and the inverted index are stored together to construct a historical audit case library.

[0089] In the above embodiments, by simultaneously constructing a vector index (extracting reasoning information and text features) and an inverted index (segmenting the original text and violation points), multimodal and efficient retrieval of the case library is achieved. The vector index supports semantic similarity matching, while the inverted index ensures accurate keyword matching. The combination of the two can quickly locate historical cases with similar reasoning logic or identical violation characteristics, improving the comprehensiveness and relevance of the review and recall.

[0090] In one embodiment, retrieving at least one preliminary review historical case similar to the target text from a pre-built historical review case library includes: extracting the text feature vector of the target text; performing a vector index retrieval in the pre-built historical review case library based on the text feature vector and vector index to select at least one second historical review data similar to the target text as a first preliminary review recall case; performing word segmentation on the target text to obtain target text word segments; performing an inverted index retrieval in the historical review case library based on the target text word segments and inverted index to select at least one second historical review data similar to the target text as a second preliminary review recall case; and selecting at least one preliminary review historical case similar to the target text from the preliminary review case union set consisting of the first preliminary review recall case and the second preliminary review recall case.

[0091] Among them, target text segmentation refers to the characters or words obtained by segmenting the target text.

[0092] In one embodiment, vector index retrieval is based on the cosine similarity between the text feature vector of the target text and the historical feature vectors in the historical review case library.

[0093] To better understand inverted index retrieval, let's illustrate with an example, using the text 'You are a great smart person' as the target text to be reviewed. First, we simulate a small text corpus (4 documents in total): Doc 1: "You are a great person" (length: 5 words); Doc 2: "Great smart person" (length: 4 words); Doc 3: "You are not a great smart person" (length: 6 words); Doc 4: "The weather is really nice today" (length: 6 words). Assuming the system performs simple segmentation by character / word, as shown in Table 1, the inverted index records which documents each word appears in and the number of times it appears (tf).

[0094] Table 1

[0095] Terms Number of documents containing this word ( ) Inverted index (DocID: number of occurrences) ) you 2 (Doc1: 1), (Doc3: 1) yes 2 (Doc1: 1), (Doc3: 1) indivual 2 (Doc1: 1), (Doc3: 1) big 3 (Doc1: 1), (Doc2: 2), (Doc3: 1) good people 1 (Doc1: 1) clever 2 (Doc2: 2), (Doc3: 1) no 1 (Doc3: 1) today 1 (Doc4: 1) weather 1 (Doc4: 1) very nice 1 (Doc4: 1)

[0096] Understandable, total number of documents = 4; Average length of all documents = (5 + 4 + 6 + 6) / 4 = 5.25; During the online query phase, the target text = "You are very smart".

[0097] Step 1: Query Preprocessing (Word Segmentation): The system segments the input target text into a word array. = [“you”, “is”, “one”, “big”, “smart”];

[0098] Step 2: Inverted Index Retrieval and Candidate Set Generation: The system searches the inverted index for these 5 words to see which documents match at least one of them: "you" -> matches Doc1, Doc3; "is" -> matches Doc1, Doc3; "a" -> matches Doc1, Doc3; "big" -> matches Doc1, Doc2, Doc3; "smart" -> matches Doc2, Doc3; Candidate set generated: Doc1, Doc2, Doc3;

[0099] Step 3: Calculate the IDF (Inverse Document Frequency) for each query term: The IDF formula used by BM25 is:

[0100] ;

[0101] in, That is, the number of documents in Table 1. .

[0102] Step 4: Calculate the BM25 score for the candidate documents. The scoring formula is as follows:

[0103] ;

[0104] in, b is a pre-set empirical value; D is the document length. The number of times a word appears in a document. The score for each candidate document in the candidate set can be calculated sequentially using this scoring formula.

[0105] Step 5: Sorting and Output: The system sorts the candidate documents in the candidate set in descending order according to their scores.

[0106] In the above embodiments, a dual-path recall strategy using vector indexing and inverted indexing is employed: the vector indexing captures semantically similar cases, while the inverted indexing accurately matches keyword-matched cases. The two are then combined to select the most relevant historical cases. Compared to single-path retrieval, this approach balances semantic generalization and precise matching, effectively avoiding missed or false recalls, and providing a more comprehensive and high-quality case reference for the initial review model, thereby improving the accuracy and robustness of the review conclusions.

[0107] In one embodiment, selecting at least one historical preliminary review case similar to the target text from the union of preliminary review cases consisting of the first preliminary review recall case and the second preliminary review recall case includes: inputting the target text and the union of preliminary review cases consisting of the first preliminary review recall case and the second preliminary review recall case into a re-ranking model to obtain a first case sequence; arranging each recall case in the first case sequence in descending order of its similarity to the target text; and selecting at least one historical preliminary review case most similar to the target text from the first case sequence.

[0108] In one embodiment, such as Figure 4As shown, the computer device can extract the text feature vector of the target text through a text feature model; based on the text feature vector and vector index, a vector index retrieval (i.e., semantic similarity case recall) is performed in a pre-built historical review case library to select the k most similar second historical review data to the target text as the first initial review recall cases; the target text is segmented to obtain target text segmentation; based on the target text segmentation and inverted index, an inverted index retrieval (i.e., statistical similarity case recall based on the BM25 algorithm) is performed in the historical review case library to select the m most similar second historical review data to the target text as the second initial review recall cases; the target text and the initial review cases consisting of the first and second initial review recall cases are input into the re-ranking model to obtain the first case sequence; each recall case in the first case sequence is arranged in descending order of its similarity to the target text; from the first case sequence, the n most similar initial review historical cases to the target text are selected.

[0109] In the above embodiments, a re-ranking model is introduced to finely rank the union of cases retrieved through dual-path recall, generating a high-quality case sequence based on the true similarity to the target text. Compared to direct merging or simple deduplication, this method can filter out low-relevance cases and prioritize retaining historical evidence that is semantically and logically closest, providing more accurate reference input for the initial review model, thereby improving the accuracy and reliability of the review conclusions.

[0110] In one embodiment, the initial review result includes the second violation conclusion of the target text, the initial review category label, the second violation point, and the second review reasoning information; the target key information includes the second violation point and the second review reasoning information; from the historical review case library, at least one review history case similar to the target key information is retrieved, including: extracting the violation point feature vector of the second violation point, and extracting the reasoning feature vector of the second review reasoning information; performing word segmentation on the second violation point to obtain violation point word segmentation; and performing vector index retrieval in the historical review case library based on the reasoning feature vector and vector index to select at least one second historical review case similar to the second review reasoning information from the historical review case library. The core data serves as the first review recall case; based on the violation point feature vector and vector index, a vector index retrieval is performed in the historical review case library to select at least one second historical review data similar to the second violation point as the second review recall case; based on violation point word segmentation and inverted index, an inverted index retrieval is performed in the historical review case library to select at least one second historical review data similar to the second violation point as the third review recall case; from the review cases consisting of the first review recall case, the second review recall case, and the third review recall case, at least one review historical case similar to the target key information is selected.

[0111] Among them, violation point segmentation is the character or word obtained by segmenting the second violation point.

[0112] In the above embodiments, vector retrieval is performed on the reasoning information and the violation points respectively, and an inverted index retrieval is performed on the violation points. After the three are integrated, historical cases most similar to the initial review logic are selected to ensure that positive and negative samples are highly matched at the level of judgment basis, providing logically consistent review basis for the large model, effectively suppressing illusions, and thus further improving the accuracy of text review.

[0113] In one embodiment, selecting at least one historical review case similar to the target key information from the union of review cases consisting of the first review recall case, the second review recall case, and the third review recall case includes: inputting the target text and the union of review cases consisting of the first review recall case, the second review recall case, and the third review recall case into a re-ranking model to obtain a second case sequence; arranging each recall case in the second case sequence in descending order of its similarity to the target text; and selecting at least one historical review case most similar to the target text from the second case sequence.

[0114] In one embodiment, such as Figure 5As shown, the computer device can extract the violation point feature vector of the second violation point and the inference feature vector of the second review inference information through a text feature model; perform word segmentation on the second violation point to obtain violation point word segments; based on the inference feature vector and vector index, perform vector index retrieval (i.e., semantic similarity case recall) in the historical review case library to select A second historical review data that are most similar to the second review inference information as the first review recall cases; based on the violation point feature vector and vector index, perform vector index retrieval in the historical review case library to select C second historical review data that are most similar to the second violation point. As the second review recall case; based on violation point word segmentation and inverted index, an inverted index retrieval is performed in the historical review case library (i.e., statistical similarity case recall based on the BM25 algorithm) to select B second historical review data most similar to the second violation point as the third review recall case; the target text and the review case set consisting of the first review recall case, the second review recall case and the third review recall case are input into the re-ranking model to obtain the second case sequence; each recall case in the second case sequence is arranged in descending order of its similarity to the target text; from the second case sequence, p review historical cases most similar to the target text are selected. For each target historical case in the union of target cases consisting of n initial review historical cases and p secondary review historical cases, if the historical review results of the target historical case meet the consistency condition with the initial review results, then the target historical case is divided into positive samples that support the initial review results (e.g., the three most similar positive samples to the target text can be selected to participate in the subsequent review process); otherwise, it is divided into negative samples that oppose the initial review results (e.g., the three most similar negative samples to the target text can be selected to participate in the subsequent review process).

[0115] In the above embodiments, a re-ranking model is introduced again during the review recall stage to perform fine-grained ranking of the union of cases obtained from multi-path retrieval. The most relevant historical review cases are selected based on their true similarity to the target text, ensuring that the positive and negative samples input into the large review model are of higher quality and more logically consistent, thereby improving the accuracy and consistency of the review conclusions and effectively reducing the risk of misjudgment.

[0116] In one embodiment, the historical review result includes the first violation conclusion, historical category label, first violation point, and first review reasoning information of the original text corresponding to the historical review result; the preliminary review result includes the second violation conclusion, preliminary review category label, second violation point, and second review reasoning information of the target text; the method further includes: for each target historical case, if the first violation conclusion in the historical review result of the target historical case is consistent with the second violation conclusion in the preliminary review result, then it is determined that the historical review result and the preliminary review result meet the conclusion consistency condition.

[0117] For example, if the first violation conclusion in the historical review of a target historical case is normal, and the second violation conclusion in the preliminary review is also normal, then the historical review result and the preliminary review result are deemed to meet the condition of consistency of conclusion. If the first violation conclusion in the historical review of a target historical case is normal, and the second violation conclusion in the preliminary review is also a violation, then the historical review result and the preliminary review result are deemed not to meet the condition of consistency of conclusion.

[0118] In the above embodiments, by comparing whether the violation conclusions of historical cases are consistent with the violation conclusions output by the initial review, the supporting or opposing samples are automatically divided. This can efficiently filter out logically consistent or conflicting cases, providing structured positive and negative examples for the review model and enhancing the credibility and accuracy of the review reasoning.

[0119] In one embodiment, such as Figure 6 As shown, the computer equipment can acquire the target text to be reviewed; retrieve at least one preliminary review historical case similar to the target text from a pre-built historical review case library; assemble the target text and the preliminary review historical case into preliminary review prompts, and input the preliminary review prompts into the preliminary review model to obtain the preliminary review result; extract the target key information from the preliminary review result, and retrieve at least one secondary review historical case similar to the target key information from the historical review case library; for each target historical case in the union of target cases consisting of preliminary review historical cases and secondary review historical cases, if the historical review result of the target historical case meets the conclusion consistency condition with the preliminary review result, then the target historical case is classified as a positive sample supporting the preliminary review result; otherwise, it is classified as a negative sample opposing the preliminary review result; assemble the target text, preliminary review result, positive sample, and negative sample into secondary review prompts, and input the secondary review prompts into the secondary review model to obtain the secondary review result. It can be understood that the secondary review result may specifically be the same as the preliminary review result, or it may be a revised version of the preliminary review result.

[0120] It should be understood that although the steps in the flowcharts of the above embodiments are shown sequentially, these steps are not necessarily executed in that order. Unless otherwise expressly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some of the steps in the above embodiments may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least a portion of the sub-steps or stages of other steps.

[0121] Based on the same inventive concept, this application also provides a text review apparatus for implementing the text review method described above. The solution provided by this apparatus is similar to the implementation scheme described in the above method; therefore, the specific limitations in one or more text review apparatus embodiments provided below can be found in the limitations of the text review method described above, and will not be repeated here.

[0122] In one embodiment, such as Figure 7 As shown, a text review device 700 is provided, which specifically includes:

[0123] Module 702 is used to obtain the target text to be reviewed;

[0124] The retrieval module 704 is used to retrieve at least one preliminary review historical case that is similar to the target text from a pre-built historical review case library;

[0125] The preliminary review module 706 is used to assemble the target text and historical preliminary review cases into preliminary review prompts, and input the preliminary review prompts into the preliminary review model to obtain the preliminary review results.

[0126] The retrieval module 704 is also used to extract the target key information from the preliminary review results and retrieve at least one review history case that is similar to the target key information from the historical review case database.

[0127] The partitioning module 708 is used to classify each target historical case in the target case set consisting of the preliminary review historical cases and the re-review historical cases as positive samples supporting the preliminary review results if the historical review results of the target historical cases meet the consistency condition with the preliminary review results; otherwise, it is classified as negative samples opposing the preliminary review results.

[0128] The review module 710 is used to assemble the target text, the initial review results, positive samples and negative samples into review prompt words, and input the review prompt words into the review model to obtain the review results.

[0129] In one embodiment, the apparatus further includes:

[0130] A construction module is used to extract unstructured first historical audit data from the audit system logs. The first historical audit data includes the original text, the first violation conclusion of the original text, and the original category label. Based on the first historical audit data and a pre-built security audit rule base, extraction prompt words are assembled and input into the extraction model to obtain structured second historical audit data. The second historical audit data includes the original text, the first violation conclusion, the analysis category label of the original text, the first violation point, and the first audit reasoning information. Historical key information is extracted from the second historical audit data. The historical key information includes at least one of the first violation conclusion, analysis category label, first violation point, and first audit reasoning information. Based on the second historical audit data and the historical key information, a historical audit case library is constructed.

[0131] In one embodiment, the historical key information includes a first violation point and a first review reasoning information; the construction module is also used to extract the feature vectors of the original text in the first review reasoning information and the second historical review data respectively, to obtain historical feature vectors, and to construct a vector index based on the historical feature vectors; to perform word segmentation on the original text and the first violation point respectively, to obtain historical word segmentation, and to construct an inverted index based on the historical word segmentation; and to associate and store the second historical review data, the vector index and the inverted index to construct a historical review case library.

[0132] In one embodiment, the retrieval module 704 is further configured to extract the text feature vector of the target text; perform vector index retrieval in a pre-built historical review case library based on the text feature vector and vector index, so as to select at least one second historical review data similar to the target text as a first preliminary review recall case from the historical review case library; perform word segmentation on the target text to obtain target text word segments; perform inverted index retrieval in the historical review case library based on the target text word segments and inverted index, so as to select at least one second historical review data similar to the target text as a second preliminary review recall case from the historical review case library; and select at least one preliminary review historical case similar to the target text from the preliminary review case union set consisting of the first preliminary review recall case and the second preliminary review recall case.

[0133] In one embodiment, the retrieval module 704 is further configured to input the target text and the union of the preliminary review cases consisting of the first preliminary review recall cases and the second preliminary review recall cases into the re-ranking model to obtain a first case sequence; each recall case in the first case sequence is arranged in descending order of its similarity to the target text; from the first case sequence, at least one preliminary review historical case that is most similar to the target text is selected.

[0134] In one embodiment, the initial review result includes the second violation conclusion of the target text, the initial review category label, the second violation point, and the second review reasoning information; the target key information includes the second violation point and the second review reasoning information; the retrieval module 704 is further used to extract the violation point feature vector of the second violation point and the reasoning feature vector of the second review reasoning information; perform word segmentation on the second violation point to obtain violation point word segmentation; and perform vector index retrieval in the historical review case library based on the reasoning feature vector and vector index to select at least one second historical review data similar to the second review reasoning information as the first re-review recall case. Based on the feature vector and vector index of the violation point, a vector index retrieval is performed in the historical review case library to select at least one second historical review data similar to the second violation point as the second review recall case; based on the violation point word segmentation and inverted index, an inverted index retrieval is performed in the historical review case library to select at least one second historical review data similar to the second violation point as the third review recall case; from the review cases set consisting of the first review recall case, the second review recall case, and the third review recall case, at least one review historical case similar to the target key information is selected.

[0135] In one embodiment, the retrieval module 704 is further configured to input the target text and the union of the review cases consisting of the first review recall case, the second review recall case and the third review recall case into the re-ranking model to obtain a second case sequence; each recall case in the second case sequence is arranged in descending order of its similarity to the target text; from the second case sequence, at least one review history case that is most similar to the target text is selected.

[0136] In one embodiment, the historical review result includes the first violation conclusion, historical category label, first violation point, and first review reasoning information of the original text corresponding to the historical review result; the preliminary review result includes the second violation conclusion, preliminary review category label, second violation point, and second review reasoning information of the target text; the segmentation module 708 is also used to determine that the historical review result and the preliminary review result meet the conclusion consistency condition if the first violation conclusion in the historical review result of each target historical case is consistent with the second violation conclusion in the preliminary review result.

[0137] The aforementioned text review device acquires the target text to be reviewed; retrieves at least one preliminary review historical case similar to the target text from a pre-built historical review case library; assembles the target text and the preliminary review historical case into preliminary review prompts, and inputs the preliminary review prompts into a preliminary review model to obtain the preliminary review result; extracts the target key information from the preliminary review result, and retrieves at least one secondary review historical case similar to the target key information from the historical review case library; for each target historical case in the union of target cases consisting of preliminary review historical cases and secondary review historical cases, if the historical review result of the target historical case meets the conclusion consistency condition with the preliminary review result, the target historical case is classified as a positive sample supporting the preliminary review result; otherwise, it is classified as a negative sample opposing the preliminary review result; assembles the target text, preliminary review result, positive sample, and negative sample into secondary review prompts, and inputs the secondary review prompts into a secondary review model to obtain the secondary review result. Compared with traditional text review methods, this application improves the accuracy of text review through a two-stage retrieval and review mechanism. First, historical cases are introduced as a reference in the preliminary review stage to compensate for the shortcomings of static rules in covering new variants such as slang and homophones. Secondly, by extracting key information from the initial review and retrieving positive and negative samples, a review prompt word containing "judgment logic" and "opposite cases" is constructed. This guides the large model to conduct secondary verification based on a reasoning chain similar to that of a human reviewer, effectively suppressing model illusions. This application's solution forms a closed loop of "initial review case reference - positive and negative example logic verification," significantly reducing missed judgments and misjudgments, overcoming the drawbacks of traditional single-step reasoning lacking a confirmation mechanism, and improving the accuracy of text review.

[0138] Each module in the aforementioned text review device can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the memory of a computer device as software, so that the processor can invoke and execute the corresponding operations of each module.

[0139] In one embodiment, a computer device is provided, which may be a server, and its internal structure diagram may be as follows: Figure 8As shown, this computer device includes a processor, memory, input / output (I / O) interfaces, and a communication interface. The processor, memory, and I / O interfaces are connected via a system bus, and the communication interface is also connected to the system bus via the I / O interfaces. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and databases. The internal memory provides the environment for the operating system and computer programs stored in the non-volatile storage media to run. The I / O interfaces are used for exchanging information between the processor and external devices. The communication interface is used for communicating with external terminals via a network connection. When the computer program is executed by the processor, it implements a text review method.

[0140] In one embodiment, a computer device is provided, which may be a terminal, and its internal structure diagram may be as follows: Figure 9 As shown, the computer device includes a processor, memory, input / output interfaces, a communication interface, a display unit, and an input device. The processor, memory, and input / output interfaces are connected via a system bus, and the communication interface, display unit, and input device are also connected to the system bus via the input / output interfaces. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The input / output interfaces are used for exchanging information between the processor and external devices. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, mobile cellular networks, NFC (Near Field Communication), or other technologies. When executed by the processor, the computer program implements a text review method. The display unit of the computer device is used to form a visually visible image. It can be a display screen, a projection device, or a virtual reality imaging device. The display screen can be an LCD screen or an e-ink screen. The input device of the computer device can be a touch layer covering the display screen, or buttons, trackballs, or touchpads set on the casing of the computer device, or external keyboards, touchpads, or mice, etc.

[0141] Those skilled in the art will understand that Figure 8 and Figure 9The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0142] In one embodiment, a computer device is also provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps in the above method embodiments.

[0143] In one embodiment, a computer-readable storage medium is provided storing a computer program that, when executed by a processor, implements the steps in the above method embodiments.

[0144] In one embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the steps in the above method embodiments.

[0145] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of related data must comply with the relevant laws, regulations and standards of the relevant countries and regions.

[0146] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile memory and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM). The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, artificial intelligence (AI) processors, etc., and are not limited to these.

[0147] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this application.

[0148] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. A text review method, characterized in that, The method includes: Obtain the target text to be reviewed; Retrieve at least one preliminary review case that is similar to the target text from a pre-built historical review case library; The target text and the historical cases of the preliminary review are assembled into preliminary review prompts, and the preliminary review prompts are input into the preliminary review model to obtain the preliminary review results; Extract the target key information from the preliminary review results, and retrieve at least one review history case that is similar to the target key information from the historical review case database; For each target historical case in the target case set consisting of the preliminary review historical cases and the re-review historical cases, if the historical review result of the target historical case meets the conclusion consistency condition with the preliminary review result, then the target historical case is classified as a positive sample supporting the preliminary review result; otherwise, it is classified as a negative sample opposing the preliminary review result. The target text, the initial review result, the positive sample, and the negative sample are assembled into a review prompt word, and the review prompt word is input into the review model to obtain the review result.

2. The method according to claim 1, characterized in that, The method further includes: Extract unstructured first historical review data from the review system logs; the first historical review data includes the original text, the first violation conclusion of the original text, and the original category label; Based on the first historical audit data and a pre-built security audit rule base, extract prompt words are assembled and input into the large extraction model to obtain structured second historical audit data; the second historical audit data includes the original text, the first violation conclusion, the analysis category label of the original text, the first violation point, and the first audit reasoning information; Extract key historical information from the second historical review data; the key historical information includes at least one of the first violation conclusion, the analysis category label, the first violation point, and the first review reasoning information. Based on the second historical audit data and the aforementioned key historical information, a historical audit case library is constructed.

3. The method according to claim 2, characterized in that, The historical key information includes the first violation point and the first review reasoning information; The construction of a historical audit case library based on the second historical audit data and the historical key information includes: The feature vectors of the original text in the first review reasoning information and the second historical review data are extracted respectively to obtain historical feature vectors, and a vector index is constructed based on the historical feature vectors; The original text and the first violation point are segmented into words to obtain historical word segments, and an inverted index is constructed based on the historical word segments. The second historical audit data, the vector index, and the inverted index are stored together to construct a historical audit case library.

4. The method according to claim 3, characterized in that, The step of retrieving at least one preliminary review case similar to the target text from a pre-built historical review case library includes: Extract the text feature vector of the target text; Based on the text feature vector and the vector index, a vector index retrieval is performed in a pre-built historical review case library to select at least one second historical review data similar to the target text as a first preliminary review recall case. The target text is segmented to obtain the target text word segmentation; Based on the target text segmentation and the inverted index, an inverted index retrieval is performed in the historical review case library to select at least one second historical review data similar to the target text as a second preliminary review recall case. From the set of preliminary review cases consisting of the first preliminary review recall case and the second preliminary review recall case, select at least one preliminary review historical case that is similar to the target text.

5. The method according to claim 4, characterized in that, The step of selecting at least one historical preliminary review case similar to the target text from the preliminary review case set consisting of the first preliminary review recall case and the second preliminary review recall case includes: The target text and the set of preliminary cases consisting of the first preliminary review recall case and the second preliminary review recall case are input into the re-ranking model to obtain the first case sequence; each recall case in the first case sequence is arranged in descending order of its similarity to the target text. From the first case sequence, select at least one preliminary review historical case that is most similar to the target text.

6. The method according to claim 3, characterized in that, The preliminary review result includes the second violation conclusion, the preliminary review category label, the second violation point, and the second review reasoning information of the target text; the target key information includes the second violation point and the second review reasoning information. The step of retrieving at least one historical review case similar to the target key information from the historical review case database includes: Extract the feature vector of the violation point of the second violation point, and extract the inference feature vector of the second review inference information; The second violation point is processed by word segmentation to obtain the violation point word segmentation; Based on the inference feature vector and the vector index, a vector index retrieval is performed in the historical review case library to select at least one second historical review data that is similar to the second review inference information as a first review recall case. Based on the violation point feature vector and the vector index, a vector index retrieval is performed in the historical review case library to select at least one second historical review data similar to the second violation point as a second review recall case. Based on the violation point word segmentation and the inverted index, an inverted index retrieval is performed in the historical review case library to select at least one second historical review data similar to the second violation point as a third review recall case; From the set of review cases consisting of the first review recall case, the second review recall case, and the third review recall case, at least one historical review case that is similar to the target key information is selected.

7. The method according to claim 6, characterized in that, The step of selecting at least one historical review case similar to the target key information from the review case set consisting of the first review recall case, the second review recall case, and the third review recall case includes: The target text and the union of the review cases consisting of the first review recall case, the second review recall case, and the third review recall case are input into the re-ranking model to obtain the second case sequence; each recall case in the second case sequence is arranged in descending order of its similarity to the target text. From the second case sequence, select at least one review history case that is most similar to the target text.

8. The method according to claim 1, characterized in that, The historical review result includes the first violation conclusion, historical category label, first violation point, and first review reasoning information of the original text corresponding to the historical review result; the preliminary review result includes the second violation conclusion, preliminary review category label, second violation point, and second review reasoning information of the target text; the method further includes: For each target historical case, if the first violation conclusion in the historical review result of the target historical case is consistent with the second violation conclusion in the preliminary review result, then the historical review result and the preliminary review result are determined to meet the conclusion consistency condition.

9. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 8.

10. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 8.

Images