Data auditing method and device, electronic equipment and storage medium
By preprocessing and keyword matching of multimodal data, and using large language models for auditing, the problems of missed and passed in multimodal data scenarios are solved, and more efficient and accurate data audit is achieved.
Patent Information
- Application Number
- CN202510031025.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-01-07
- Publication Date
- 2025-04-11
AI Technical Summary
The traditional data review method cannot meet the needs of multimodal data scenarios, and there are problems of missed inspection and pass-through inspection, and the manual review costs and low efficiency.
By receiving multimodal data to be reviewed, data preprocessing is performed, using the regular expression matching module to match the keyword collection to be reviewed from the keyword lexicon, and generating audit prompt text that can be received by the large language model, sending it to the large language model for review, and finally receiving and processing the audit results.
It reduces labor costs, improves data review efficiency and accuracy, and can more effectively identify harmful content in multimodal data.
Smart Images
Figure CN119807398A_ABST
Abstract
Description
Technical Field
[0001] The present application relates to the field of artificial intelligence technology, and in particular to a data review method, device electronic device and storage medium. Background Art
[0002] With the booming development of the Internet, the surge in user-generated content on the Internet has brought about an urgent need for data review systems to ensure the compliance of online content and protect users from harmful information. The existing data review methods include keyword matching algorithm review and manual review. Keyword matching algorithm review is an automated data review method that uses a preset keyword list to identify and filter sensitive or inappropriate content. It is efficient when processing large amounts of data and has relatively low costs. However, keyword matching technology often misses detection due to missing keywords. For example, if the publisher of harmful content uses typos, synonyms, or encoding methods to circumvent keyword filtering, then these contents may escape the detection of the algorithm. In addition, keyword matching algorithm review relies too much on keyword matching and does not review the text from the level of semantic understanding. Therefore, it may mistakenly mark harmless content as harmful, resulting in over-detection. Manual review is a traditional data review method that relies on the judgment and experience of professional reviewers. However, when the amount of data to be reviewed is huge and the content is complex, the difficulty and cost of manual review will become very high. Reviewers need to process a large amount of information within a limited time. The review efficiency of reviewers needs to be further improved, which also increases the risk of review errors. In addition, manual review often has problems of false detection and missed detection, because reviewers may not be able to accurately identify all harmful content due to differences in subjective judgment or other human factors. Summary of the invention
[0003] The main technical problem that the implementation method of the present application solves is that the traditional data review method cannot meet the needs of multimodal data scenarios.
[0004] To solve the above technical problems, the first technical solution adopted in the implementation mode of the present application is: providing a data review method, including: receiving multimodal data to be reviewed, and performing data preprocessing on the multimodal data to be reviewed to obtain text data to be processed in a preset data format; obtaining at least one preset keyword thesaurus, and a regular expression matching module corresponding to the keyword thesaurus; using the regular expression matching module to process the text data to be processed, and matching the keyword set to be reviewed corresponding to the text data to be processed from the keyword thesaurus; obtaining the review field corresponding to the multimodal data to be reviewed, and generating a review prompt text that can be received by a preset large language model according to the review field; sending the text data to be processed, the keyword set to be reviewed and the review prompt text to the preset large language model, and receiving the data review result output by the preset large language model.
[0005] Optionally, after the step of using the regular expression matching module to process the text data to be processed and matching the keywords to be reviewed corresponding to the text data to be processed from the keyword thesaurus, the method further includes: obtaining the keyword level rules set in the keyword thesaurus; traversing the keyword set to be reviewed, and generating the keyword level corresponding to each keyword to be reviewed in the keyword set to be reviewed according to the keyword level rules; and updating the keyword set to be reviewed using the keyword level corresponding to each keyword to be reviewed.
[0006] Optionally, after the step of sending the text data to be processed, the set of keywords to be reviewed and the review prompt text to the preset large language model, it also includes: generating a second review prompt text receivable by the preset large language model, the content of the second review prompt text is to identify pre-defined sensitive words or terms, and sending the second review prompt text to the preset large language model; generating a third review prompt text receivable by the preset large language model, the content of the third review prompt text is to perform contextual semantic analysis on the text data to be processed, and sending the third review prompt text to the preset large language model; generating a fourth review prompt text receivable by the preset large language model, the content of the fourth review prompt text is to perform risk assessment on the text data to be processed, and sending the fourth review prompt text to the preset large language model.
[0007] Optionally, after the step of sending the fourth review prompt text to the preset large language model, it also includes: receiving the risk assessment result returned by the preset large language model based on the fourth review prompt text; forwarding the text data to be processed and the set of keywords to be reviewed to manual review based on the risk assessment result; or, modifying the review field based on the risk assessment result, and repeating the steps between generating a review prompt text receivable by the preset large language model based on the review field to receiving the risk assessment result returned by the preset large language model based on the fourth review prompt text.
[0008] Optionally, after the step of receiving the data review result output by the preset large language model, it also includes: expanding the multimodal data to be reviewed by a preset data expansion method to obtain an expanded training data set, wherein the expanded training data set includes a special domain data subset and a general data subset; respectively establishing special domain review prompt words and general domain review prompt words that can be received by the preset large language model; cross-combining the special domain data subset and the general data subset with the special domain review prompt words and the general domain review prompt words according to a preset ratio to obtain different cross-combinations of training data; and fine-tuning the preset large language model through the different cross-combinations of training data.
[0009] Optionally, the keyword level includes a primary label, a secondary label and a tertiary label. After the step of receiving the data review result output by the preset large language model, it also includes: parsing the data review result, extracting the review result keyword set of the data review result; traversing the review result keyword set, and generating a keyword level corresponding to the review result keyword in the review result keyword set according to the keyword level; wherein the primary label is represented as a label category, the secondary label is represented as the intent analysis result logic, and the tertiary label is represented as a label explanation.
[0010] Optionally, after the step of receiving the data review results output by the preset large language model, it also includes: randomly sending the data review results to manual verification, and receiving the verification results of the manual verification; optimizing the generation process of the review prompt text and the preset large language model according to the verification results.
[0011] In order to solve the above technical problems, the second technical solution adopted by the implementation method of the present application is: to provide a data audit device, including: a data receiving module for receiving multimodal data to be audited, and performing data preprocessing on the multimodal data to be audited to obtain text data to be processed in a preset data format. A keyword thesaurus acquisition module, used to obtain at least one preset keyword thesaurus, and a regular expression matching module corresponding to the keyword thesaurus; a keyword matching module, used to process the text data to be processed using the regular expression matching module, and match the keyword thesaurus to the keyword set corresponding to the text data to be processed; an audit prompt text module, used to obtain the audit field corresponding to the multimodal data to be audited, and generate an audit prompt text that can be received by a preset large language model according to the audit field; a data audit result module, used to send the text data to be processed, the keyword set to be audited and the audit prompt text to the preset large language model, and receive the data audit result output by the preset large language model.
[0012] To solve the above technical problems, the third technical solution adopted in the implementation mode of the present application is: to provide an electronic device, comprising: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor so that the at least one processor can execute the data audit method as described above.
[0013] In order to solve the above technical problems, the fourth technical solution adopted in the implementation mode of the present application is: providing a non-volatile computer-readable storage medium, wherein the non-volatile computer-readable storage medium stores computer-executable instructions, and when the computer-executable instructions are executed by an electronic device, the electronic device executes the data audit method as described above.
[0014] Different from the related art, the present application receives multimodal data to be reviewed and performs data preprocessing on the multimodal data to be reviewed to obtain text data to be processed in a preset data format; obtains at least one preset keyword thesaurus and a regular expression matching module corresponding to the keyword thesaurus; uses the regular expression matching module to process the text data to be processed, and matches the keyword set to be reviewed corresponding to the text data to be processed from the keyword thesaurus; obtains the review field corresponding to the multimodal data to be reviewed, and generates a review prompt text that can be received by a preset large language model according to the review field; sends the text data to be processed, the keyword set to be reviewed and the review prompt text to the preset large language model, and receives the data review result output by the large language model, thereby reducing labor costs and improving data review efficiency and review accuracy. BRIEF DESCRIPTION OF THE DRAWINGS
[0015] One or more embodiments are exemplarily described by corresponding drawings, which do not constitute limitations on the embodiments. Elements with the same reference numerals in the drawings represent similar elements, and the figures in the drawings do not constitute proportional limitations unless otherwise stated.
[0016] Figure 1 It is a schematic diagram of the operating environment of the data audit method provided in the embodiment of the present application.
[0017] Figure 2 It is a schematic diagram of the execution flow of the data audit method provided in the embodiment of the present application.
[0018] Figure 3 It is a schematic diagram of the execution flow of matching the keywords to be reviewed in the data review method provided in the embodiment of the present application.
[0019] Figure 4 It is a schematic diagram of the execution flow of fine-tuning the audit function of a large language model in the data audit method provided in an embodiment of the present application.
[0020] Figure 5 It is a schematic diagram of the system structure of the data audit device provided in an embodiment of the present application.
[0021] Figure 6 It is a schematic diagram of the hardware structure of an electronic device for executing the data audit method provided in an embodiment of the present application. DETAILED DESCRIPTION
[0022] In order to make the purpose, technical solution and advantages of the present application more clearly understood, the present application is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.
[0023] It should be noted that, if there is no conflict, the various features in the embodiments of the present application can be combined with each other, all within the scope of protection of the present application. In addition, although the functional modules are divided in the device schematic diagram and the logical order is shown in the flow chart, in some cases, the steps shown or described can be performed in a different order than the module division in the device schematic diagram or the order in the flow chart.
[0024] Unless otherwise defined, all technical and scientific terms used in this specification have the same meaning as those commonly understood by those skilled in the art to which this application belongs. The terms used in the specification of this application are only for the purpose of describing specific embodiments and are not intended to limit this application. The term "and / or" used in this specification includes any and all combinations of one or more of the related listed items.
[0025] To facilitate understanding of this embodiment, a data audit method disclosed in the embodiment of this application is first described in detail. Figure 1 , Figure 1 Schematic diagram of the operating environment of the data audit method provided in the embodiment of the present application. Figure 1 As shown, the execution subject of the data audit method provided in the embodiment of the present application is generally an electronic device with certain computing capabilities, such as a computer device. In some possible implementations, the data audit method can be implemented by a processor calling a computer-readable instruction stored in a memory. Figure 1 The computer device in the term can be a server. The server can be an independent server or a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communications, middleware services, domain name services, security services, content delivery networks (CDNs), and big data and artificial intelligence platforms. It can be understood that Figure 1 The number of computer devices in the figure is only for reference and can be expanded to any number according to actual needs.
[0026] Please continue reading Figure 2 , Figure 2 is a schematic diagram of the execution flow of the data audit method provided in the embodiment of the present application, such as Figure 2 As shown, the following steps are included:
[0027] S1. Receive multimodal data to be reviewed, and perform data preprocessing on the multimodal data to be reviewed to obtain text data to be processed in a preset data format.
[0028] Among them, the embodiment of the present application combines the multimodal data processing capabilities of a large language model, receives data reviews in various modalities, such as text data, audio data, image data, video data, etc., and uniformly converts the aforementioned multimodal data to be reviewed into text data that can be analyzed and reasoned by the large language model for processing.
[0029] S2. Obtain at least one preset keyword thesaurus and a regular expression matching module corresponding to the keyword thesaurus.
[0030] Among them, in the specific implementation process, the corresponding industry field keyword thesaurus can be obtained according to the industry field characteristics of the data to be reviewed, such as the financial field, insurance field, medical field and other keyword thesaurus.
[0031] S3. Use a regular expression matching module to process the text data to be processed, and match the set of keywords to be reviewed corresponding to the text data to be processed from the keyword thesaurus.
[0032] Please continue reading Figure 3 , Figure 3 : is a schematic diagram of the execution flow of matching the keywords to be reviewed in the data review method provided in the embodiment of the present application, such as Figure 3 As shown, the following steps S31 to S33 may be included.
[0033] S31. Obtain the keyword level rules set in the keyword vocabulary.
[0034] For example, corresponding yellow-level keywords and red-level keywords can be set according to the needs of data review. If the red-level keywords are hit in the above-mentioned data to be reviewed, then the data to be reviewed can be directly determined as unqualified content data, without the need for subsequent large language model intelligent review, so as to avoid the "technical illusion" or technical misjudgment of the large language model. Audit errors caused by technical misjudgment; secondly, if the yellow-level keywords are hit in the above-mentioned data to be reviewed, but the red-level keywords are not hit, the above-mentioned data to be reviewed will be sent to the preset large language model for intelligent review.
[0035] S32: traverse the set of keywords to be reviewed, and generate a keyword level corresponding to each keyword to be reviewed in the set of keywords to be reviewed according to the keyword level rule, for example, the yellow level or the red level in the above example.
[0036] S33: Use the keyword level corresponding to each keyword to be reviewed to update the set of keywords to be reviewed.
[0037] S4. Obtain the audit field corresponding to the multimodal data to be audited, and generate an audit prompt text that can be received by a preset large language model based on the audit field.
[0038] The audit field can be a general application field or other professional knowledge field, such as finance, insurance, and medical fields. For example, when it comes to data confidentiality audit, an audit prompt text can be generated: "You are a professional confidentiality and security auditor. Please try to follow the confidentiality and security classification standards for data audit", and the preset large language model can be shaped into a professional data audit intelligent body in the simplest and fastest way.
[0039] S5. Send the text data to be processed, the keyword set to be reviewed, and the review prompt text to the preset large language model, and receive the data review result output by the large language model.
[0040] Please continue reading Figure 4 , Figure 4 is a schematic diagram of the execution flow of fine-tuning the audit function of a large language model in the data audit method provided in the embodiment of the present application, such as Figure 4 As shown, the following steps S51 to S54 may be included.
[0041] S51. Expand the multimodal data to be reviewed by a preset data expansion method to obtain an expanded training data set, where the expanded training data set includes a dedicated domain data subset and a general data subset.
[0042] For example, operations such as synonym replacement, sentence reorganization, and noise addition can be performed on the text to generate new text samples, or a pre-trained model can be used to expand the training data by fine-tuning some layers of the model to adapt to the new data set.
[0043] S52, respectively establishing special domain audit prompt words and general domain audit prompt words that can be received by the preset large language model.
[0044] S53, cross-combining the special-domain data subset and the general-domain data subset with the special-domain audit prompt words and the general-domain audit prompt words respectively according to a preset ratio to obtain different cross-combinations of training data.
[0045] For example, 70% of the special-domain audit prompt words are combined with the special-domain data subset, 30% of the special-domain audit prompt words are combined with the general data subset, 30% of the general-domain audit prompt words are combined with the special-domain data subset, and 70% of the general-domain audit prompt words are combined with the general data subset.
[0046] S54. Fine-tune the preset large language model by cross-combining different training data.
[0047] As an optional implementation, after the step of sending the text data to be processed, the keyword set to be reviewed and the review prompt text to the preset large language model, the following steps S61 to S64 may also be included.
[0048] S61: Generate a second review prompt text that can be received by a preset large language model, the content of the second review prompt text is to identify pre-defined sensitive words or terms, and send the second review prompt text to the preset large language model. That is, firstly use the preset large language model to review sensitive words or terms on the text data to be processed.
[0049] S62: Generate a third audit prompt text that can be received by the preset large language model, the content of the third audit prompt text is to perform contextual semantic analysis on the text data to be processed, and send the third audit prompt text to the preset large language model. That is, first perform semantic audit on the text data to be processed through the preset large language model.
[0050] S62: Generate a fourth audit prompt text that can be received by a preset large language model, wherein the content of the fourth audit prompt text is to perform a risk assessment on the text data to be processed, and send the fourth audit prompt text to the preset large language model.
[0051] As an optional implementation, after the above-mentioned step of sending the fourth audit prompt text to the preset large language model, it also includes: receiving the risk assessment result returned by the preset large language model based on the fourth audit prompt text; forwarding the text data to be processed and the set of keywords to be reviewed to manual review based on the risk assessment result; or, modifying the audit field according to the risk assessment result, and repeating the steps between generating the audit prompt text that can be received by the preset large language model based on the audit field and receiving the risk assessment result returned by the preset large language model based on the fourth audit prompt text, which can effectively solve the problem of incorrect audit field identification during the data audit process.
[0052] As a preferred implementation, the data audit results output by the above-mentioned preset large language model can also be optimized. First, the data audit results are parsed to extract the audit result keyword set of the data audit results. Then, the audit result keyword set is traversed, and the keyword level corresponding to the audit result keyword in the audit result keyword set is generated according to the keyword level. Among them, the first-level label is represented as a label category, the second-level label is represented as the intent analysis result logic, and the third-level label is represented as a label explanation.
[0053] Finally, in order to further prevent the preset large language model from having "technical illusions" during the review process, the embodiment of the present application also performs manual sampling processing, randomly sends data review results to manual verification, and receives the verification results of manual verification, and optimizes the generation process of the review prompt text and the preset large language model according to the verification results.
[0054] The data audit method provided in the embodiment of the present application receives multimodal data to be audited, performs data preprocessing on the multimodal data to be audited, and obtains text data to be processed in a preset data format; obtains at least one preset keyword thesaurus and a regular expression matching module corresponding to the keyword thesaurus; uses the regular expression matching module to process the text data to be processed, and matches the keyword set to be audited corresponding to the text data to be processed from the keyword thesaurus; obtains the audit field corresponding to the multimodal data to be audited, and generates an audit prompt text that can be received by a preset large language model according to the audit field; sends the text data to be processed, the keyword set to be audited and the audit prompt text to the preset large language model, and receives the data audit result output by the large language model, thereby reducing labor costs and improving data audit efficiency and audit accuracy.
[0055] Please continue reading Figure 5 , Figure 5is a schematic diagram of the system structure of the data audit device provided in the embodiment of the present application, such as Figure 5 As shown, the data audit device 50 includes: a data receiving module 51 for audit, a keyword vocabulary acquisition module 52, a keyword matching module 53, an audit prompt text module and a data audit result module 55.
[0056] The to-be-audited data receiving module 51 is used to receive multimodal to-be-audited data and perform data preprocessing on the multimodal to-be-audited data to obtain to-be-processed text data in a preset data format.
[0057] The keyword thesaurus acquisition module 52 is used to acquire at least one preset keyword thesaurus and a regular expression matching module corresponding to the keyword thesaurus.
[0058] The keyword matching module 53 is used to process the text data to be processed using a regular expression matching module, and match the keyword set to be reviewed corresponding to the text data to be processed from the keyword word library.
[0059] The audit prompt text module 54 is used to obtain the audit field corresponding to the multimodal data to be audited, and generate an audit prompt text that can be received by a preset large language model according to the audit field.
[0060] The data review result module 55 is used to send the text data to be processed, the keyword set to be reviewed and the review prompt text to the preset large language model, and receive the data review result output by the large language model.
[0061] As an optional implementation, the keyword matching module 53 is also specifically used to obtain the keyword level rules set in the keyword vocabulary; traverse the set of keywords to be reviewed, and generate the keyword level corresponding to each keyword to be reviewed in the set of keywords to be reviewed according to the keyword level rules; and use the keyword level corresponding to each keyword to be reviewed to update the set of keywords to be reviewed.
[0062] As an optional implementation, the data review result module 55 is also specifically used to generate a second review prompt text that can be received by the preset large language model, the content of the second review prompt text is to identify pre-defined sensitive words or terms, and the second review prompt text is sent to the preset large language model; generate a third review prompt text that can be received by the preset large language model, the content of the third review prompt text is to perform contextual semantic analysis on the text data to be processed, and the third review prompt text is sent to the preset large language model; generate a fourth review prompt text that can be received by the preset large language model, the content of the fourth review prompt text is to perform risk assessment on the text data to be processed, and the fourth review prompt text is sent to the preset large language model.
[0063] As an optional implementation, the data review result module 55 is also specifically used to receive the risk assessment result returned by the preset large language model based on the fourth review prompt text; based on the risk assessment result, forward the text data to be processed and the set of keywords to be reviewed to manual review; or, modify the review field based on the risk assessment result, and repeat the steps between generating a review prompt text that can be received by the preset large language model based on the review field to receiving the risk assessment result returned by the preset large language model based on the fourth review prompt text.
[0064] As an optional implementation, the data audit device 50 also includes a model fine-tuning module, which is specifically used to expand the multimodal data to be audited through a preset data expansion method to obtain an expanded training data set, wherein the expanded training data set includes a special domain data subset and a general data subset; respectively establish special domain audit prompt words and general domain audit prompt words that can be received by the preset large language model; cross-combine the special domain data subset and the general data subset with the special domain audit prompt words and the general domain audit prompt words according to a preset ratio to obtain different training data cross-combinations; and fine-tune the preset large language model through the different training data cross-combinations.
[0065] As an optional implementation, the data review device 50 also includes an review result fine-tuning module, which is specifically used to parse the data review results and extract the review result keyword set of the data review results; traverse the review result keyword set, and generate a keyword level corresponding to the review result keywords in the review result keyword set according to the keyword level; wherein the first-level label is represented as a label category, the second-level label is represented as the intent analysis result logic, and the third-level label is represented as a label explanation.
[0066] As an optional implementation, the data review device 50 also includes a manual verification module, which is specifically used to randomly send the data review results to manual verification and receive the verification results of the manual verification; optimize the generation process of the review prompt text and the preset large language model according to the verification results.
[0067] It should be noted that the above data audit device can execute the data audit method provided in the embodiment of the present application, and has the corresponding functional modules and beneficial effects of the execution method. For technical details not fully described in the data audit device embodiment, please refer to the data audit method provided in the embodiment of the present application.
[0068] Please continue reading Figure 6 , Figure 6 is a schematic diagram of the hardware structure of an electronic device for executing a data audit method provided in an embodiment of the present application, such as Figure 6 As shown, the electronic device 600 includes:
[0069] One or more processors 610 and memory 620, Figure 6 A processor 610 is taken as an example.
[0070] The processor 610 and the memory 620 may be connected via a bus or other means. Figure 6 The example of connecting through bus is taken in the following.
[0071] The memory 620 is a non-volatile computer-readable storage medium that can be used to store non-volatile software programs, non-volatile computer executable programs and modules, such as program instructions / modules corresponding to the data audit method in the embodiment of the present application. The processor 610 executes various functional applications and data processing of the server by running the non-volatile software programs, instructions and modules stored in the memory 620, that is, implementing the data audit method in the above method embodiment.
[0072] The memory 620 may include a program storage area and a data storage area, wherein the program storage area may store an operating system and an application required by at least one function; the data storage area may store data created according to the use of the data audit device, etc. In addition, the memory 620 may include a high-speed random access memory, and may also include a non-volatile memory, such as at least one disk storage device, a flash memory device, or other non-volatile solid-state storage device. In some embodiments, the memory 620 may optionally include a memory remotely arranged relative to the processor 610, and these remote memories may be connected to the data audit device via a network. Examples of the above-mentioned network include, but are not limited to, the Internet, an intranet, a local area network, a mobile communication network, and combinations thereof.
[0073] The one or more modules are stored in the memory 620, and when executed by the one or more processors 610, perform the data audit method in any of the above method embodiments, for example, perform the above described Figure 2 Steps S1 to S5 of the method, Figure 3 Steps S31 to S33 of the method, Figure 4 Steps S51 to S54 of the method are implemented Figure 5 The functions of modules 51-55 in.
[0074] The above-mentioned product can execute the method provided in the embodiment of the present application, and has the functional modules and beneficial effects corresponding to the execution method. For technical details not fully described in this embodiment, please refer to the method provided in the embodiment of the present application.
[0075] The embodiment of the present application provides a non-volatile computer-readable storage medium, wherein the non-volatile computer-readable storage medium stores computer-executable instructions, and the computer-executable instructions are executed by one or more processors, such as Figure 6 A processor 610 in the embodiment may enable the one or more processors to execute the data audit method in any of the above method embodiments, for example, to execute the above described Figure 2 Steps S1 to S5 of the method, Figure 3 Steps S31 to S33 of the method, Figure 4 Steps S51 to S54 of the method are implemented Figure 5 The functions of modules 51-55 in.
[0076] The present application provides a computer program product, wherein the computer program product includes a computer program stored on a non-volatile computer-readable storage medium, wherein the computer program includes program instructions, and when the program instructions are executed by the electronic device, the electronic device is enabled to execute the data audit method in any of the above method embodiments, for example, to execute the above described Figure 2 Steps S1 to S5 of the method, Figure 3 Steps S31 to S33 of the method, Figure 4 Steps S51 to S54 of the method are implemented Figure 5 The functions of modules 51-55 in.
[0077] The device embodiments described above are merely illustrative, wherein the units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed on multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the solution of this embodiment.
[0078] Through the description of the above implementation methods, ordinary technicians in this field can clearly understand that each implementation method can be implemented by means of software plus a general hardware platform, and of course, it can also be implemented by hardware. Ordinary technicians in this field can understand that all or part of the processes in the above-mentioned embodiment method can be completed by instructing related hardware through a computer program, and the program can be stored in a computer-readable storage medium. When the program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, the storage medium can be a disk, an optical disk, a read-only memory (ROM) or a random access memory (RAM), etc.
[0079] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit them. Under the concept of the present application, the technical features in the above embodiments or different embodiments can also be combined, the steps can be implemented in any order, and there are many other changes in different aspects of the present application as described above, which are not provided in detail for the sake of simplicity. Although the present application has been described in detail with reference to the aforementioned embodiments, a person of ordinary skill in the art should understand that the technical solutions described in the aforementioned embodiments can still be modified, or some of the technical features can be replaced by equivalents. These modifications or replacements do not deviate the essence of the corresponding technical solutions from the scope of the technical solutions of the embodiments of the present application.
Claims
1. A data audit method, characterized in that: include: Receiving multimodal data to be reviewed, and performing data preprocessing on the multimodal data to be reviewed to obtain text data to be processed in a preset data format; Obtain at least one preset keyword thesaurus and a regular expression matching module corresponding to the keyword thesaurus; Using the regular expression matching module to process the text data to be processed, and matching the keyword set to be reviewed corresponding to the text data to be processed from the keyword vocabulary; Acquire the audit field corresponding to the multimodal data to be audited, and generate an audit prompt text that can be received by a preset large language model according to the audit field; The text data to be processed, the keyword set to be reviewed and the review prompt text are sent to the preset large language model, and the data review result output by the preset large language model is received.
2. The data audit method according to claim 1, characterized in that: After the step of using the regular expression matching module to process the text data to be processed and matching the keywords to be reviewed corresponding to the text data to be processed from the keyword thesaurus, the method further includes: Obtaining keyword level rules set in the keyword vocabulary; Traversing the set of keywords to be reviewed, and generating a keyword level corresponding to each keyword to be reviewed in the set of keywords to be reviewed according to the keyword level rule; The keyword level corresponding to each keyword to be reviewed is used to update the keyword set to be reviewed.
3. The data audit method according to claim 1, characterized in that: After the step of sending the text data to be processed, the keyword set to be reviewed and the review prompt text to the preset large language model, the method further includes: Generate a second review prompt text that can be received by the preset large language model, the content of the second review prompt text is to identify predefined sensitive words or terms, and send the second review prompt text to the preset large language model; Generate a third audit prompt text that can be received by the preset large language model, wherein the content of the third audit prompt text is to perform contextual semantic analysis on the text data to be processed, and send the third audit prompt text to the preset large language model; Generate a fourth audit prompt text that can be received by the preset large language model, the content of the fourth audit prompt text is to perform a risk assessment on the text data to be processed, and send the fourth audit prompt text to the preset large language model.
4. The data audit method according to claim 3, characterized in that: After the step of sending the fourth audit prompt text to the preset large language model, the method further includes: Receive the risk assessment result returned by the preset large language model according to the fourth review prompt text; According to the risk assessment result, forwarding the text data to be processed and the keyword set to be reviewed to manual review; Or, modify the audit field according to the risk assessment result, and repeat the steps between generating an audit prompt text that can be received by a preset large language model according to the audit field and receiving the risk assessment result returned by the preset large language model according to the fourth audit prompt text.
5. The data audit method according to claim 1, characterized in that: After the step of receiving the data review result output by the preset large language model, the method further includes: Expanding the multimodal data to be reviewed by a preset data expansion method to obtain an expanded training data set, wherein the expanded training data set includes a dedicated domain data subset and a general data subset; Establishing special-domain audit prompt words and general-domain audit prompt words respectively that can be received by the preset large language model; Cross-combining the dedicated domain data subset and the general data subset with the dedicated domain audit prompt words and the general domain audit prompt words respectively according to a preset ratio to obtain different cross-combinations of training data; The preset large language model is fine-tuned by cross-combining the different training data.
6. The data audit method according to claim 2, characterized in that: The keyword level includes a primary tag, a secondary tag, and a tertiary tag. After the step of receiving the data review result output by the preset large language model, the step further includes: Parsing the data audit result, and extracting an audit result keyword set of the data audit result; Traversing the audit result keyword set, and generating a keyword level corresponding to the audit result keyword in the audit result keyword set according to the keyword level; Among them, the first-level label is represented as a label category, the second-level label is represented as the intent analysis result logic, and the third-level label is represented as a label explanation.
7. The data audit method according to claim 1, characterized in that: After the step of receiving the data review result output by the preset large language model, the method further includes: Randomly sending the data review result to manual verification, and receiving the verification result of the manual verification; The generation process of the audit prompt text and the preset large language model are optimized according to the verification result.
8. A data audit device, characterized in that: include: A module for receiving data to be reviewed, used for receiving multimodal data to be reviewed, and performing data preprocessing on the multimodal data to be reviewed to obtain text data to be processed in a preset data format; A keyword thesaurus acquisition module, used to acquire at least one preset keyword thesaurus and a regular expression matching module corresponding to the keyword thesaurus; A keyword matching module, used to process the text data to be processed using the regular expression matching module, and match the keyword set to be reviewed corresponding to the text data to be processed from the keyword vocabulary; An audit prompt text module is used to obtain the audit field corresponding to the multimodal data to be audited, and generate an audit prompt text that can be received by a preset large language model according to the audit field; The data audit result module is used to send the text data to be processed, the keyword set to be audited and the audit prompt text to the preset large language model, and receive the data audit result output by the preset large language model.
9. An electronic device, characterized in that: include: at least one processor; as well as, a memory communicatively connected to the at least one processor; wherein, The memory stores instructions that can be executed by the at least one processor, and the instructions are executed by the at least one processor so that the at least one processor can execute the data audit method according to any one of claims 1 to 7.
10. A non-volatile computer-readable storage medium, characterized in that: The non-volatile computer-readable storage medium stores computer-executable instructions. When the computer-executable instructions are executed by an electronic device, the electronic device executes the data audit method according to any one of claims 1 to 7.
Citation Information
Cited By
Automatic auditing method, device and equipment for bank vouchers and vouchers, and storage medium
CN120543267A