Policy quality inspection method and device based on multi-modal large model, equipment and medium

Abstract

This invention belongs to the field of artificial intelligence technology and is applicable to the financial and medical fields. It discloses a method, apparatus, device, and medium for policy quality inspection based on a multimodal large model. The method includes: processing an image of the policy document to be reviewed and preset underwriting standards based on a pre-trained multimodal large model and preset initial multimodal large model prompts to obtain first text content; extracting the text content from the policy document image based on an optical character recognition algorithm to obtain second text content; comparing a first word segmentation set and a second word segmentation set based on a text difference comparison algorithm to identify differing text content that exists in the second word segmentation set but not in the first word segmentation set; optimizing the preset initial multimodal large model prompts based on the differing text content, and reprocessing the policy document image and preset underwriting standards to generate the quality inspection result of the policy to be reviewed. This invention improves the accuracy of policy quality inspection.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence technology and is applicable to the financial and medical fields. In particular, it relates to a method, device, equipment and medium for policy quality inspection based on a multimodal large model. Background Technology

[0002] As the insurance industry's underwriting process undergoes continuous digital and intelligent transformation, automating policy document review through artificial intelligence technology has become a key path to improve operational efficiency and optimize customer experience. For example, in auto insurance underwriting, policyholders need to submit various types and formats of documents, such as vehicle registration certificates, driver's licenses, group photos, and handwritten authorization letters. Traditional manual review methods struggle to meet the demands for efficient and accurate processing. Similarly, in the medical insurance field, policyholders need to submit various medical documents, such as hospitalization records, medical record summaries, and examination reports. The varying formats and contents of these documents further complicate the review process.

[0003] Currently, when using visual big data models for automated review of insurance policy documents, the process begins with the model directly parsing the policy document images, extracting key information and converting it into text. This text is then compared with pre-defined underwriting rules to achieve end-to-end intelligent review. However, in large-scale applications, visual big data models commonly suffer from omissions and misidentifications. They may overlook key fields in the images or misinterpret the identified content, leading to incorrect review conclusions. Therefore, a second manual review is still necessary. This not only fails to reduce labor costs but also increases overall processing time and costs due to the error correction process, severely hindering the efficiency and reliability of automated review.

[0004] Therefore, how to effectively improve the accuracy of policy quality inspection in insurance business is a technical problem that urgently needs to be solved. Summary of the Invention

[0005] This invention provides a method, apparatus, equipment, and medium for policy quality inspection based on a multimodal large model, in order to solve the technical problem of how to effectively improve the accuracy of policy quality inspection in insurance business.

[0006] In a first aspect, the present invention provides a policy quality inspection method based on a multimodal large model, comprising: Acquire images of policy documents to be reviewed, preset initial multimodal large model prompts, and preset underwriting standards; Based on the pre-trained multimodal large model and the preset initial multimodal large model prompts, the policy data image to be reviewed and the preset underwriting standards are processed to obtain the first text content of the policy data image. Based on the optical character recognition algorithm, the text content in the policy data image is extracted to obtain the second text content of the policy data image; Based on the text segmentation algorithm, the first text content and the second text content of the policy data image are segmented into words respectively to obtain the first word segmentation set and the second word segmentation set of the policy data image. Based on the text difference comparison algorithm, the first word segmentation set and the second word segmentation set of the policy data image are compared to identify the difference text content that exists in the second word segmentation set but does not exist in the first word segmentation set; Based on the difference in text content, the preset initial multimodal large model prompt words are optimized, and the policy data image and preset underwriting standards are reprocessed based on the optimized multimodal large model prompt words to generate the quality inspection results of the policy to be reviewed.

[0007] Secondly, the present invention provides a policy quality inspection device based on a multimodal large model, comprising: The acquisition module is used to acquire images of policy documents to be reviewed, preset initial multimodal large model prompts, and preset underwriting standards; The first processing module processes the policy document image to be reviewed and the preset underwriting standards based on the pre-trained multimodal large model and the preset initial multimodal large model prompt words to obtain the first text content of the policy document image. The extraction module, based on an optical character recognition algorithm, extracts the text content from the policy data image to obtain the second text content of the policy data image. The second processing module, based on a text segmentation algorithm, performs word segmentation on the first and second text contents of the policy data image, respectively, to obtain the first and second word segmentation sets of the policy data image. The comparison module, based on a text difference comparison algorithm, compares the first word segmentation set and the second word segmentation set of the policy data image to identify the difference text content that exists in the second word segmentation set but does not exist in the first word segmentation set; The generation module optimizes the preset initial multimodal large model prompt words based on the difference text content, and reprocesses the policy data image and preset underwriting standards based on the optimized multimodal large model prompt words to generate the quality inspection results of the policy to be reviewed.

[0008] Thirdly, the present invention provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the above-mentioned policy quality inspection method based on a multimodal large model.

[0009] Fourthly, the present invention provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of the above-described policy quality inspection method based on a multimodal large model.

[0010] The aforementioned policy quality inspection method, device, equipment, and medium based on a multimodal large model, in its implementation scheme, allows for the acquisition of an image of the policy documents to be reviewed, preset initial multimodal large model prompts, and preset underwriting standards via a client. Based on a pre-trained multimodal large model and the preset initial multimodal large model prompts, the image of the policy documents to be reviewed and the preset underwriting standards are processed to obtain the first text content of the policy document image. Based on an optical character recognition algorithm, the text content in the policy document image is extracted to obtain the second text content of the policy document image. Finally, based on a text segmentation algorithm... The first and second text contents of the policy document image are segmented into words to obtain a first word set and a second word set. Based on a text difference comparison algorithm, the first and second word sets are compared to identify differing text content that exists in the second word set but not in the first. Based on this differing text content, a preset initial multimodal large model prompt is optimized, and the policy document image and preset underwriting standards are reprocessed based on the optimized multimodal large model prompts to generate the quality inspection result of the policy to be reviewed. In this invention, the policy document image is processed and its features extracted in parallel using a pre-trained multimodal large model and an optical character recognition algorithm, effectively reducing the problems of missed and false recognition caused by the limitations of the large model itself. Through the text difference comparison algorithm, differing text content that the optical character recognition algorithm can recognize but the multimodal large model cannot capture is accurately extracted, and the preset initial multimodal large model prompts are automatically optimized based on this differing content. This method can achieve autonomous iterative improvement of prompt words without manual intervention, and then reprocess the policy document images with the optimized prompt words, effectively improving the overall accuracy of policy document review. Attached Figure Description

[0011] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments of the present invention will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0012] Figure 1 This is a schematic diagram of an application environment for a policy quality inspection method based on a multimodal large model, according to an embodiment of the present invention.

[0013] Figure 2 This is a flowchart illustrating a policy quality inspection method based on a multimodal large model in one embodiment of the present invention.

[0014] Figure 3 yes Figure 2 A schematic diagram of a specific implementation method for step S20.

[0015] Figure 4 yes Figure 2 A schematic diagram of a specific implementation method for step S30.

[0016] Figure 5 yes Figure 2 A flowchart illustrating a specific implementation of step S40.

[0017] Figure 6 yes Figure 2 A schematic diagram of a specific implementation method for step S50.

[0018] Figure 7 yes Figure 2 A flowchart illustrating a specific implementation of step S60.

[0019] Figure 8 This is a schematic diagram of a policy quality inspection device based on a multimodal large model in one embodiment of the present invention.

[0020] Figure 9 This is a schematic diagram of the structure of a computer device according to an embodiment of the present invention.

[0021] Figure 10 This is another structural schematic diagram of a computer device according to one embodiment of the present invention. Detailed Implementation

[0022] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0023] The policy quality inspection method based on a multimodal large model provided in this invention can be applied to, for example... Figure 1 In the application environment, Figure 1This is a schematic diagram of an application environment for a policy quality inspection method based on a multimodal large model according to an embodiment of the present invention; wherein, the client communicates with the server through the network. The server can obtain the policy document image to be reviewed, preset initial multimodal large model prompts, and preset underwriting standards from the client. Based on the pre-trained multimodal large model and the preset initial multimodal large model prompts, the server processes the policy document image and the preset underwriting standards to obtain the first text content of the policy document image. Based on an optical character recognition algorithm, the server extracts the text content from the policy document image to obtain the second text content of the policy document image. Based on a text segmentation algorithm, the server segments the first and second text content of the policy document image to obtain a first and a second segmentation set of the policy document image. Based on a text difference comparison algorithm, the server compares the first and second segmentation sets of the policy document image to identify the differential text content that exists in the second segmentation set but not in the first segmentation set. Based on the differential text content, the server optimizes the preset initial multimodal large model prompts and reprocesses the policy document image and the preset underwriting standards based on the optimized multimodal large model prompts to generate the quality inspection result of the policy to be reviewed. This invention utilizes a pre-trained multimodal large model and an optical character recognition algorithm to perform parallel processing and feature extraction on policy document images, effectively reducing the problems of missed and false recognition caused by the limitations of the large model itself. Through a text difference comparison algorithm, it accurately extracts textual content that the optical character recognition algorithm can recognize but the multimodal large model fails to capture, and automatically optimizes the preset initial multimodal large model prompts based on these differences. This method achieves autonomous iterative improvement of prompts without manual intervention, and then reprocesses the policy document images using the optimized prompts, effectively improving the overall accuracy of policy document review. The invention will be described in detail below through specific embodiments.

[0024] Please see Figure 2 As shown, Figure 2 This is a flowchart illustrating a policy quality inspection method based on a multimodal large model provided in an embodiment of the present invention. The policy quality inspection method based on a multimodal large model specifically includes the following steps: S10: Obtain the image of the policy documents to be reviewed, the preset initial multimodal large model prompts, and the preset underwriting standards. Specifically, in this embodiment of the invention, using preset initial prompts and underwriting standards can ensure the standardization of the entire review process, ensuring that different policies follow the same standards during review, and improving the consistency of the review. Clearly defining the policy information and standards to be reviewed makes subsequent processing more targeted, helps to improve efficiency, and reduces unnecessary workload. For example, in the scenario of financial auto insurance, the image of the auto insurance documents to be reviewed can be obtained, and the preset initial multimodal large model prompts can be: Please identify all the text information in this image, and pay special attention to the vehicle identification number, engine number, and registration date.

[0025] S20: Based on a pre-trained multimodal large model and preset initial multimodal large model prompts, the policy document image to be reviewed and preset underwriting standards are processed to obtain the first text content of the policy document image. Specifically, in this embodiment of the invention, the multimodal large model can process image and text information simultaneously, making full use of the characteristics of different data sources, thereby improving the comprehensiveness and accuracy of information extraction. The process of automatically generating the first text content reduces the need for manual intervention, improves processing speed and efficiency, and reduces the possibility of human error. By combining underwriting standards, the model can better understand the context of the policy content, thereby generating text content that better meets actual needs. Specifically, such as... Figure 3 The above, Figure 3 yes Figure 2 A schematic flowchart of a specific implementation of step S20 includes the following steps S21-S24: S21: Perform multi-scale feature extraction on the policy data image to obtain the visual feature vector of the policy data image. Specifically, in this embodiment of the invention, multi-scale feature extraction can capture different levels of information in the policy data image. For example, in the financial auto insurance scenario, multi-scale feature information includes vehicle owner information, vehicle information, etc.; in the medical insurance scenario, multi-scale feature information includes the patient's basic information and relevant medical terms. By generating visual feature vectors, the model can better understand the policy content. Feature extraction transforms high-dimensional image data into a low-dimensional representation, which can improve processing efficiency and reduce computational complexity.

[0026] S22: The visual feature vector of the policy data image is cross-modal semantically aligned with the preset initial multimodal large model prompts and preset underwriting standards, and a dynamic association mapping between the policy data image region and the preset underwriting standard clauses is established through an attention mechanism. Specifically, in this embodiment of the invention, cross-modal semantic alignment combines visual features with textual information, enabling a more comprehensive understanding of the specific content of the policy data. For example, in the car insurance scenario, accident descriptions can be compared with policy terms; in the medical insurance scenario, diagnostic information can be associated with insurance coverage. Furthermore, the attention mechanism allows the model to focus on specific regions in the image related to the underwriting standard clauses, improving the accuracy of information extraction.

[0027] S23: Based on the established dynamic association mapping, identify key fields in the policy document image area that are related to the preset underwriting standard clauses, and perform logical consistency verification on the identified key fields. Specifically, in this embodiment of the invention, the established dynamic association mapping can accurately identify key fields related to the underwriting standard clauses, ensuring that the review process focuses on the most important information. For example, in the car insurance scenario, the established dynamic association mapping can identify information such as the vehicle owner's name, vehicle type, and insurance amount; in the medical insurance scenario, the established dynamic association mapping can identify information such as the patient's disease name and treatment plan. Performing logical consistency verification on the identified key fields can ensure the compliance and rationality of the information. For example, in the car insurance scenario, confirm whether the accident date is within the policy validity period; in medical insurance, verify whether the medical services are covered by the insurance.

[0028] S24: Based on the key field identification results and logical consistency verification results, generate the first text content of the policy document image. Specifically, in this embodiment of the invention, the first text content generated based on key fields and verification results can reduce manual intervention and improve processing speed, especially in scenarios requiring the review of a large number of documents. The generated text content is usually structured, easy to process and analyze subsequently, and convenient for reviewers to quickly consult.

[0029] S30: Based on the optical character recognition algorithm, extract the text content from the policy data image to obtain the second text content of the policy data image. Specifically, in this embodiment of the invention, optical character recognition technology can quickly and accurately extract text from images, providing necessary text data for subsequent analysis and comparison, converting the text in the image into an editable and analyzable digital format, facilitating subsequent processing and storage. Using the optical character recognition algorithm reduces the need for manual text input, thereby reducing the risk of errors caused by manual input. Specifically, such as... Figure 4 The above, Figure 4 yes Figure 2A schematic flowchart of a specific implementation of step S30 includes the following steps S31-S33: S31: Based on an optical character recognition algorithm, extract global layout features and local character features from the policy document image. Specifically, in this embodiment of the invention, by extracting global layout features and local character features respectively, a more comprehensive understanding of the text's structure and content can be achieved. This is particularly important when processing complex policy documents; global layout features help distinguish different types of text, while local character features ensure accurate recognition of each character.

[0030] S32: The global layout feature information and local character feature information of the extracted policy document image are fused, and the text content in the policy document image is identified based on the fused feature information. Specifically, in this embodiment of the invention, fusing global feature information and local feature information can effectively improve the ability to recognize text content. Especially in complex insurance contracts and medical records, text may be difficult to directly recognize due to complex layout. By fusing global layout feature information and local character feature information, the model can better understand the contextual relationships of the text. For example, in a car insurance policy, the logical relationship between the title and the terms can be represented by global features, while the specific terms are identified by local features. After fusing feature information, similar characters can be distinguished more accurately, thereby reducing the probability of recognition errors.

[0031] S33: Perform structured sequence recombination on the text content in the identified policy data image to generate the second text content of the policy data image. Specifically, in this embodiment of the invention, performing structured sequence recombination on the identified text content makes the information clearer and easier to process and analyze later. For example, in a medical insurance scenario, patient basic information, diagnosis results, treatment plans, etc., can be organized. The structured text content can be easily used for storage, retrieval, and analysis, providing a foundation for subsequent data mining and decision support.

[0032] S40: Based on a text segmentation algorithm, the first and second text contents of the policy data image are segmented into words respectively, resulting in a first word segmentation set and a second word segmentation set for the policy data image. Specifically, in this embodiment of the invention, through word segmentation, the text is transformed into a meaningful vocabulary set, making subsequent comparison and analysis clearer and more efficient. The segmented data is easier to operate and analyze, helping to improve the processing speed and accuracy of policy data. Specifically, such as... Figure 5 The above, Figure 5 yes Figure 2 A schematic flowchart of a specific implementation of step S40 includes the following steps S41-S43: S41: Load the professional dictionary in the insurance field into the Jieba Chinese word segmentation tool, and perform word segmentation on the first text content and the second text content through the Jieba Chinese word segmentation tool to obtain the first initial word segmentation result and the second initial word segmentation result. Specifically, in the embodiments of the present invention, loading the professional dictionary in the insurance field enables the Jieba Chinese word segmentation tool to accurately identify insurance-related terms and proper nouns, such as information like the applicant, policy number, exemption clause, etc. This is crucial for improving the accuracy of word segmentation. By using the professional dictionary, word segmentation errors and ambiguities can be reduced, especially in complex insurance documents, ensuring that the extracted vocabulary is consistent with the context. For example, in the medical insurance scenario, "hospitalization" and "hospitalization expenses" can be accurately separated. Using the word segmentation tool can quickly process a large amount of text content, reducing the need for manual intervention and making subsequent data analysis more efficient.

[0033] S42: Remove punctuation marks and common stop words from the first initial word segmentation result and the second initial word segmentation result, and retain the key entity words and corresponding position identification information in the first initial word segmentation result and the second initial word segmentation result. Specifically, in the embodiments of the present invention, removing punctuation marks and common stop words, such as stop words like "of", "is", "in", etc., can reduce the interference of meaningless information and improve the accuracy and effectiveness of subsequent analysis. For example, in the financial vehicle insurance scenario, remove punctuation and common stop words from the word segmentation result, and retain key entity words such as the applicant and insurance amount and position information. Retaining the key entity words and position identification information enables the information to be stored and queried in an orderly manner.

[0034] S43: Convert the key entity words and corresponding position identification information in the first initial word segmentation result and the second initial word segmentation result into a structured word segmentation set to obtain the first word segmentation set and the second word segmentation set of the policy data image. Specifically, in the embodiments of the present invention, converting the word segmentation result into a structured word segmentation set can make the data easier to store and process.

[0035] S50: Based on the text difference comparison algorithm, compare the first word segmentation set and the second word segmentation set of the policy data image, and identify the differential text content that exists in the second word segmentation set but does not exist in the first word segmentation set. Specifically, in the embodiments of the present invention, through the comparison algorithm, compare the word list output by the optical character recognition algorithm and the word list output by the pre-trained multimodal large model, and find out the words that exist in the word list output by the optical character recognition algorithm but do not exist in the word list output by the pre-trained multimodal large model. These found words are the key information that the pre-trained multimodal large model may miss identifying. Specifically, as Figure 6 described, Figure 6 is Figure 2 a schematic flowchart of a specific implementation manner of step S50 in the figure, which specifically includes the following steps S51 - S53: S51: Based on a sequence matching algorithm, a word-by-word comparison is performed on the first and second word segments to generate a comparison result sequence containing difference markers. Specifically, in this embodiment of the invention, word-by-word comparison can clearly identify the specific differences between two texts. The generated comparison result sequence marks the differences, enabling reviewers to quickly locate changes in the document.

[0036] S52: Extract the differentially labeled lexical units from the alignment result sequence to form an initial differential vocabulary set. Specifically, in this embodiment of the invention, extracting the differentially labeled lexical units can effectively avoid interference from redundant information and improve the targeting of the analysis. Creating an initial differential vocabulary set provides a clear target for subsequent semantic filtering and analysis, making the process more efficient.

[0037] S53: Semantic filtering is performed on the initial set of differing words to obtain the original positional order of the semantically filtered differing words in the second word segmentation set. The filtered differing words are then reorganized according to their original positional order to generate structured differing text content. Specifically, in this embodiment of the invention, semantic filtering eliminates irrelevant differences, retaining only words with practical meaning, thereby ensuring the validity of the final result. Reorganizing the differing words according to their original positional order helps maintain the logic and context of the text, making the generated differing text content easy to understand and use. The generated structured differing text content can be directly used for subsequent decision support, such as reviewing insurance clauses, proposing modifications, or conducting compliance assessments.

[0038] S60: Optimize the preset initial multimodal large model prompt words based on the differing text content, and reprocess the policy data image and preset underwriting standards based on the optimized multimodal large model prompt words to generate the quality inspection result of the policy to be reviewed. Specifically, in this embodiment of the invention, through the analysis and learning of the differing text, the key information that the pre-trained multimodal large model may have missed is dynamically added back to the preset initial multimodal large model prompt words. This can form an updated, more targeted prompt word, improving the accuracy and adaptability of subsequent processing. Reprocessing the policy data after optimizing the prompt words can generate higher quality inspection results and reduce the risk of misjudgment and omission. Specifically, such as... Figure 7 The above, Figure 7 yes Figure 2 A schematic flowchart of a specific implementation of step S60 includes the following steps S61-S63: S61: Construct prompt statements based on the differing text content, and integrate the prompt statements with the preset initial multimodal large model prompts to generate optimized multimodal large model prompts. Specifically, in this embodiment of the invention, by constructing prompt statements based on differing text content, the multimodal large model can better understand the current policy, thereby improving the accuracy and effectiveness of the model in processing relevant data. Integrating the prompt statements with the preset initial prompts ensures that the model focuses on the most important information, avoids interference from irrelevant information, and improves the efficiency and quality of subsequent processing. Generating optimized prompts makes the model more adaptable, allowing for optimization for specific insurance products or underwriting standards, and improving the reliability of the review results.

[0039] S62: The policy document image, optimized multimodal large model prompts, and preset underwriting standards are input into a pre-trained multimodal large model for reprocessing to generate the third text content of the policy to be reviewed. Specifically, in this embodiment of the invention, combining the policy document image, optimized prompts, and underwriting standards into the model enables multi-level information fusion, ensuring that the final generated text content is more comprehensive and accurate. The pre-trained multimodal large model can identify and process complex insurance terms and conditions, ensuring that the generated third text content meets underwriting standards and improving the effectiveness of the review.

[0040] S63: Based on the third text content and the preset underwriting standards, generate the quality inspection result of the policy to be reviewed. Specifically, in this embodiment of the invention, based on the comparison between the third text content and the underwriting standards, the quality inspection result can be automatically generated, reducing the need for manual review and improving review efficiency and accuracy.

[0041] As can be seen, in the above-described scheme, specifically in this invention, the parallel processing and feature extraction of policy document images using a pre-trained multimodal large model and an optical character recognition algorithm effectively reduces the problems of missed and false recognition caused by the limitations of the large model itself. Through a text difference comparison algorithm, the algorithm accurately extracts textual content that the optical character recognition algorithm can recognize but the multimodal large model fails to capture, and automatically optimizes the preset initial multimodal large model prompts based on these differences. This method achieves autonomous iterative improvement of prompts without manual intervention, and then reprocesses the policy document images using the optimized prompts, effectively improving the overall accuracy of policy document review.

[0042] It should be understood that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0043] In one embodiment, a policy quality inspection device based on a multimodal large model is provided, which corresponds one-to-one with the policy quality inspection method based on a multimodal large model in the above embodiments. For example... Figure 8 As shown, Figure 8 This is a schematic diagram of a policy quality inspection device based on a multimodal large model according to an embodiment of the present invention. The device includes an acquisition module 81, a first processing module 82, an extraction module 83, a second processing module 84, a comparison module 85, and a generation module 86. Detailed descriptions of each functional module are as follows: The acquisition module 81 is used to acquire the image of the policy information to be reviewed, the preset initial multimodal large model prompt words, and the preset underwriting standards; The first processing module 82 processes the policy document image to be reviewed and the preset underwriting standards based on the pre-trained multimodal large model and the preset initial multimodal large model prompt words to obtain the first text content of the policy document image. Extraction module 83 extracts the text content from the policy data image based on optical character recognition algorithm to obtain the second text content of the policy data image; The second processing module 84, based on a text segmentation algorithm, performs word segmentation on the first text content and the second text content of the policy data image respectively, to obtain the first word segmentation set and the second word segmentation set of the policy data image. Comparison module 85, based on a text difference comparison algorithm, compares the first word set and the second word set of the policy data image to identify the difference text content that exists in the second word set but does not exist in the first word set; The generation module 86 optimizes the preset initial multimodal large model prompt words based on the difference text content, and reprocesses the policy data image and preset underwriting standards based on the optimized multimodal large model prompt words to generate the quality inspection result of the policy to be reviewed.

[0044] In one embodiment, the first processing module 82 is specifically used for: Multi-scale feature extraction is performed on the policy data image to obtain the visual feature vector of the policy data image; The visual feature vector of the policy data image is cross-modal semantically aligned with the preset initial multimodal large model prompt words and preset underwriting standards, and a dynamic association mapping between the policy data image region and the preset underwriting standard terms is established through an attention mechanism. Based on the established dynamic association mapping, key fields related to the preset underwriting standard clauses are identified in the policy data image area, and the identified key fields are logically consistent. Based on the key field identification results and logical consistency verification results, the first text content of the policy data image is generated.

[0045] In one embodiment, the extraction module 83 is specifically used for: Based on optical character recognition algorithms, global layout features and local character features are extracted from policy document images. The global layout features and local character features of the extracted policy data image are fused together, and the text content in the policy data image is identified based on the fused feature information. The text content in the identified policy data image is restructured into a structured sequence to generate the second text content of the policy data image.

[0046] In one embodiment, the second processing module 84 is specifically used for: A professional dictionary in the insurance field is loaded into the Jieba Chinese word segmentation tool, and the first text content and the second text content are segmented by the Jieba Chinese word segmentation tool to obtain the first initial word segmentation result and the second initial word segmentation result; Remove punctuation marks and common stop words from the first and second initial word segmentation results, and retain key entity words and corresponding position identifiers from the first and second initial word segmentation results; The key entity words and corresponding location identifiers in the first and second initial word segmentation results are converted into structured word sets to obtain the first and second word sets of the policy data image.

[0047] In one embodiment, the comparison module 85 is specifically used for: The first and second word segments are compared word by word using a sequence matching algorithm to generate a comparison result sequence containing difference markers. Extract the differentially labeled lexical units from the alignment result sequence to form an initial differential lexical set; The initial set of differing words is semantically filtered to obtain the original position order of the semantically filtered differing words in the second word segmentation set. The filtered differing words are then reorganized according to the original position order to generate structured differing text content.

[0048] In one embodiment, the generation module 86 is specifically used for: Based on the difference in text content, a prompt statement is constructed, and the prompt statement is integrated with the preset initial multimodal large model prompt words to generate optimized multimodal large model prompt words; The policy data image, the optimized multimodal large model prompts, and the preset underwriting standards are input into the pre-trained multimodal large model for reprocessing to generate the third text content of the policy to be reviewed. Based on the content of the third text and the preset underwriting standards, a quality inspection result for the policy to be reviewed is generated.

[0049] In one embodiment, the policy quality inspection device based on a multimodal large model is further used for: Determine whether the quality inspection results corresponding to the first text content and the third text content are consistent; When the quality inspection results corresponding to the first text content and the third text content are inconsistent, the optimized multimodal large model prompt words are stored in the prompt word knowledge base, and the quality inspection results corresponding to the third text content are used as the quality inspection results of the policy to be reviewed. When the quality inspection result corresponding to the first text content is consistent with that corresponding to the third text content, the quality inspection result corresponding to the first text content shall be used as the quality inspection result of the policy to be reviewed.

[0050] This invention provides a policy quality inspection device based on a multimodal large model. A pre-trained multimodal large model and an optical character recognition (OCR) algorithm perform parallel processing and feature extraction on policy document images, effectively reducing the problems of missed and false recognition caused by the limitations of the large model itself. Through a text difference comparison algorithm, it accurately extracts textual content that the OCR algorithm can recognize but the multimodal large model fails to capture. Based on this difference, it automatically optimizes the preset initial multimodal large model prompts. This method requires no manual intervention and can achieve autonomous iterative improvement of the prompts. The optimized prompts are then used to reprocess the policy document images, effectively improving the overall accuracy of policy document review.

[0051] Specific limitations regarding the policy quality inspection device based on a multimodal large model can be found in the limitations of the policy quality inspection method based on a multimodal large model mentioned above, and will not be repeated here. Each module in the aforementioned policy quality inspection device based on a multimodal large model 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 in hardware form, or stored in the memory of a computer device in software form, so that the processor can call and execute the corresponding operations of each module.

[0052] In one embodiment, a computer device is provided, which may be a server, and its internal structure diagram may be as follows: Figure 9 As shown, Figure 9This is a schematic diagram of a computer device according to an embodiment of the present invention. The computer device includes a processor, memory, network interface, and database connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile and / or volatile storage media and internal memory. The non-volatile storage media stores an operating system, computer programs, and a database. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage media. The network interface is used to communicate with external clients via a network connection. When the computer program is executed by the processor, it implements the functions or steps of a server-side policy quality inspection method based on a multimodal large model.

[0053] In one embodiment, a computer device is provided, which may be a client, and its internal structure diagram may be as follows: Figure 10 As shown, Figure 10 This is another schematic diagram of a computer device according to an embodiment of the present invention. The computer device includes a processor, memory, network interface, display screen, and input device connected via a system bus. The processor provides computing and control capabilities. The memory includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores an operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The network interface is used to communicate with an external server via a network connection. When the computer program is executed by the processor, it implements the client-side functions or steps of a policy quality inspection method based on a multimodal large model.

[0054] In one embodiment, a computer device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to perform the following steps: Acquire images of policy documents to be reviewed, preset initial multimodal large model prompts, and preset underwriting standards; Based on the pre-trained multimodal large model and the preset initial multimodal large model prompts, the policy data image to be reviewed and the preset underwriting standards are processed to obtain the first text content of the policy data image. Based on the optical character recognition algorithm, the text content in the policy data image is extracted to obtain the second text content of the policy data image; Based on the text segmentation algorithm, the first text content and the second text content of the policy data image are segmented into words respectively to obtain the first word segmentation set and the second word segmentation set of the policy data image. Based on the text difference comparison algorithm, the first word segmentation set and the second word segmentation set of the policy data image are compared to identify the difference text content that exists in the second word segmentation set but does not exist in the first word segmentation set; Based on the difference in text content, the preset initial multimodal large model prompt words are optimized, and the policy data image and preset underwriting standards are reprocessed based on the optimized multimodal large model prompt words to generate the quality inspection results of the policy to be reviewed.

[0055] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, the computer program performing the following steps when executed by a processor: Acquire images of policy documents to be reviewed, preset initial multimodal large model prompts, and preset underwriting standards; Based on the pre-trained multimodal large model and the preset initial multimodal large model prompts, the policy data image to be reviewed and the preset underwriting standards are processed to obtain the first text content of the policy data image. Based on the optical character recognition algorithm, the text content in the policy data image is extracted to obtain the second text content of the policy data image; Based on the text segmentation algorithm, the first text content and the second text content of the policy data image are segmented into words respectively to obtain the first word segmentation set and the second word segmentation set of the policy data image. Based on the text difference comparison algorithm, the first word segmentation set and the second word segmentation set of the policy data image are compared to identify the difference text content that exists in the second word segmentation set but does not exist in the first word segmentation set; Based on the difference in text content, the preset initial multimodal large model prompt words are optimized, and the policy data image and preset underwriting standards are reprocessed based on the optimized multimodal large model prompt words to generate the quality inspection results of the policy to be reviewed.

[0056] It should be noted that the functions or steps that can be implemented by the computer-readable storage medium or computer device described above can be referred to the relevant descriptions on the server side and client side in the foregoing method embodiments. To avoid repetition, they will not be described one by one here.

[0057] 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. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in a variety of forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0058] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is used as an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above.

[0059] The above-described embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included within the protection scope of the present invention.

Claims

1. A policy quality inspection method based on a multimodal large model, characterized in that, include: Acquire images of policy documents to be reviewed, preset initial multimodal large model prompts, and preset underwriting standards; Based on the pre-trained multimodal large model and the preset initial multimodal large model prompts, the policy document image to be reviewed and the preset underwriting standards are processed to obtain the first text content of the policy document image. Based on the optical character recognition algorithm, the text content in the policy data image is extracted to obtain the second text content of the policy data image; Based on the text segmentation algorithm, the first text content and the second text content of the policy data image are segmented into words respectively to obtain the first word segmentation set and the second word segmentation set of the policy data image. Based on the text difference comparison algorithm, the first word segmentation set and the second word segmentation set of the policy data image are compared to identify the difference text content that exists in the second word segmentation set but does not exist in the first word segmentation set; Based on the difference in text content, the preset initial multimodal large model prompt words are optimized, and the policy data image and preset underwriting standards are reprocessed based on the optimized multimodal large model prompt words to generate the quality inspection results of the policy to be reviewed.

2. The policy quality inspection method based on a multimodal large model according to claim 1, characterized in that, The pre-trained multimodal large model and preset initial multimodal large model prompts are used to process the policy document image to be reviewed and preset underwriting standards to obtain the first text content of the policy document image, including: Multi-scale feature extraction is performed on the policy data image to obtain the visual feature vector of the policy data image; The visual feature vector of the policy data image is cross-modal semantically aligned with the preset initial multimodal large model prompt words and preset underwriting standards, and a dynamic association mapping between the policy data image region and the preset underwriting standard terms is established through an attention mechanism. Based on the established dynamic association mapping, key fields related to the preset underwriting standard clauses are identified in the policy data image area, and the identified key fields are logically consistent. Based on the key field identification results and logical consistency verification results, the first text content of the policy data image is generated.

3. The policy quality inspection method based on a multimodal large model according to claim 1, characterized in that, Based on optical character recognition algorithms, the text content in the policy data image is extracted to obtain the second text content of the policy data image, including: Based on optical character recognition algorithms, global layout features and local character features are extracted from policy document images. The global layout features and local character features of the extracted policy data image are fused together, and the text content in the policy data image is identified based on the fused feature information. The text content in the identified policy data image is restructured into a structured sequence to generate the second text content of the policy data image.

4. The policy quality inspection method based on a multimodal large model according to claim 1, characterized in that, The text segmentation algorithm described above segments the first and second text contents of the policy data image into words, respectively, to obtain a first and second segmented word set for the policy data image, including: A professional dictionary in the insurance field is loaded into the Jieba Chinese word segmentation tool, and the first text content and the second text content are segmented by the Jieba Chinese word segmentation tool to obtain the first initial word segmentation result and the second initial word segmentation result; Remove punctuation marks and common stop words from the first and second initial word segmentation results, and retain key entity words and corresponding position identifiers from the first and second initial word segmentation results; The key entity words and corresponding location identifiers in the first and second initial word segmentation results are converted into structured word sets to obtain the first and second word sets of the policy data image.

5. The policy quality inspection method based on a multimodal large model according to claim 1, characterized in that, The text difference comparison algorithm compares the first and second word segments of the policy data image to identify differing text content that exists in the second word segment set but not in the first word segment set, including: The first and second word segments are compared word by word using a sequence matching algorithm to generate a comparison result sequence containing difference markers. Extract the differentially labeled lexical units from the alignment result sequence to form an initial differential lexical set; The initial set of differing words is semantically filtered to obtain the original position order of the semantically filtered differing words in the second word segmentation set. The filtered differing words are then reorganized according to the original position order to generate structured differing text content.

6. The policy quality inspection method based on a multimodal large model according to claim 1, characterized in that, The process involves optimizing the preset initial multimodal large model prompts based on the differential text content, and then reprocessing the policy data image and preset underwriting standards based on the optimized multimodal large model prompts to generate the quality inspection results for the policy to be reviewed, including: Based on the difference in text content, a prompt statement is constructed, and the prompt statement is integrated with the preset initial multimodal large model prompt words to generate optimized multimodal large model prompt words; The policy data image, the optimized multimodal large model prompts, and the preset underwriting standards are input into the pre-trained multimodal large model for reprocessing to generate the third text content of the policy to be reviewed. Based on the content of the third text and the preset underwriting standards, a quality inspection result for the policy to be reviewed is generated.

7. The policy quality inspection method based on a multimodal large model according to claim 6, characterized in that, After optimizing the preset initial multimodal large model prompts based on the differential text content, and reprocessing the policy data image and preset underwriting standards based on the optimized multimodal large model prompts to generate the quality inspection results of the policy to be reviewed, the method further includes: Determine whether the quality inspection results corresponding to the first text content and the third text content are consistent; When the quality inspection results corresponding to the first text content and the third text content are inconsistent, the optimized multimodal large model prompt words are stored in the prompt word knowledge base, and the quality inspection results corresponding to the third text content are used as the quality inspection results of the policy to be reviewed. When the quality inspection result corresponding to the first text content is consistent with that corresponding to the third text content, the quality inspection result corresponding to the first text content shall be used as the quality inspection result of the policy to be reviewed.

8. A policy quality inspection device based on a multimodal large model, characterized in that, include: The acquisition module is used to acquire images of policy documents to be reviewed, preset initial multimodal large model prompts, and preset underwriting standards; The first processing module processes the policy document image to be reviewed and the preset underwriting standards based on the pre-trained multimodal large model and the preset initial multimodal large model prompt words to obtain the first text content of the policy document image. The extraction module, based on an optical character recognition algorithm, extracts the text content from the policy data image to obtain the second text content of the policy data image. The second processing module, based on a text segmentation algorithm, performs word segmentation on the first and second text contents of the policy data image, respectively, to obtain the first and second word segmentation sets of the policy data image. The comparison module, based on a text difference comparison algorithm, compares the first word segmentation set and the second word segmentation set of the policy data image to identify the difference text content that exists in the second word segmentation set but does not exist in the first word segmentation set; The generation module optimizes the preset initial multimodal large model prompt words based on the difference text content, and reprocesses the policy data image and preset underwriting standards based on the optimized multimodal large model prompt words to generate the quality inspection results of the policy to be reviewed.

9. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the policy quality inspection method based on a multimodal large model as claimed in any one of claims 1 to 7.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the steps of the policy quality inspection method based on a multimodal large model as described in any one of claims 1 to 7.

Images