Insurance business-based knowledge generation method, device, equipment and storage medium

Abstract

This application provides a method, apparatus, device, and storage medium for knowledge generation based on insurance business, belonging to the field of artificial intelligence technology. The method includes: acquiring multi-source, multi-modal data from an insurance business scenario, including at least two types of data: image data, text data, voice data, and video data; de-identifying the multi-source, multi-modal data; annotating the de-identified insurance data; training a pre-defined large language model based on insurance knowledge data; uniformly encoding the pre-acquired insurance business data based on a multi-modal insurance data processing model to obtain a fused feature representation; and generating knowledge from the fused feature representation based on the multi-modal insurance data processing model to obtain insurance business knowledge such as underwriting conclusions and claims conclusions. This application can be applied to insurance scenarios in fintech, breaking down information silos, uncovering deep knowledge connections, and improving the efficiency and accuracy of insurance business knowledge generation.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology and is applied to financial technology scenarios, particularly to a knowledge generation method, device, equipment, and storage medium based on insurance business. Background Technology

[0002] In the insurance scenario within fintech, insurance companies need to process massive amounts of heterogeneous insurance data to handle underwriting, claims, and other insurance business operations. In practice, insurance data is often highly complex and fragmented, originating from multiple sources and existing in multiple modalities. This includes internal data (such as scanned images of paper policies, claims investigation reports, and internal training documents) and external dynamic data (such as new regulatory rules and industry announcements). This data is typically isolated in independent systems (such as core business systems, CRM systems, and image management systems), resulting in numerous data silos and a fragmented insurance knowledge view. Furthermore, the lack of effective semantic-level fusion and correlation mechanisms between these systems makes it impossible to build a complete and unified insurance knowledge view. Cross-departmental and cross-process business collaboration faces significant difficulties. Relevant personnel (such as insurance agents and sales staff) spend considerable time organizing and analyzing data when handling underwriting and claims, and data silos can lead to errors, such as incorrect underwriting or claims conclusions. Therefore, improving the efficiency and accuracy of insurance data processing has become a pressing issue. Summary of the Invention

[0003] The main objective of this application is to propose a knowledge generation method, apparatus, device, and storage medium based on insurance business, aiming to solve the technical problem of low efficiency in insurance data processing in the current technology and improve the generation efficiency and accuracy of insurance business knowledge.

[0004] To achieve the above objectives, a first aspect of this application proposes a knowledge generation method based on insurance business, the method comprising: Acquire multi-source, multi-modal data from insurance business scenarios; wherein the multi-source, multi-modal data includes at least two of the following: image data, text data, voice data, and video data; Multi-source, multi-modal data is anonymized to ensure data security; wherein, the anonymized data is multi-modal data. Based on the anonymized insurance data, insurance knowledge data is obtained by labeling it. The preset large language model is trained based on the insurance knowledge data to obtain a multimodal insurance data processing model; The multimodal insurance data processing model is used to uniformly encode the pre-acquired insurance business data to obtain a fused feature representation; wherein the insurance data includes at least one of the following: image data, text data, voice data, and video data. Based on the multimodal insurance data processing model, knowledge is generated from the fused feature representation to obtain insurance business knowledge; wherein, the insurance business knowledge includes at least one of the following: underwriting conclusion and claims conclusion.

[0005] To achieve the above objectives, a second aspect of this application provides a knowledge generation apparatus based on insurance business, the apparatus comprising: The insurance data acquisition module is used to acquire multi-source, multi-modal data from insurance business scenarios; wherein, the multi-source, multi-modal data includes at least two of the following: image data, text data, voice data, and video data; The data desensitization module is used to desensitize multi-source, multi-modal data, and to desensitize and protect the data; wherein, the desensitized and protected data is multi-modal data; The data annotation module is used to annotate the de-identified insurance data to obtain insurance knowledge data; The model training module is used to train a preset large language model based on the insurance knowledge data to obtain a multimodal insurance data processing model. The feature encoding module is used to perform unified encoding processing on the pre-acquired insurance business data based on the multimodal insurance data processing model to obtain a fused feature representation; wherein, the insurance data includes at least one of the following: image data, text data, voice data, and video data; The insurance knowledge generation module is used to generate knowledge from the fused feature representation based on the multimodal insurance data processing model to obtain insurance business knowledge; wherein, the insurance business knowledge includes at least one of the following: underwriting conclusion and claims conclusion.

[0006] To achieve the above objectives, a third aspect of the present application provides a computer device, the computer device including a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement the method described in the first aspect.

[0007] To achieve the above objectives, a fourth aspect of the present application provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the method described in the first aspect.

[0008] The knowledge generation method, apparatus, device, and storage medium based on insurance business proposed in this application acquire multi-source, multi-modal data from insurance business scenarios, perform anonymization processing on the multi-source, multi-modal data, annotate the anonymized insurance data obtained from the anonymization process, and train a preset large language model based on the insurance knowledge data obtained from the annotation. Then, based on the trained multi-modal insurance data processing model, the insurance business data is uniformly encoded to obtain a fused feature representation. Knowledge is generated based on the fused feature representation using the multi-modal insurance data processing model to obtain insurance business knowledge such as underwriting conclusions and claims conclusions. This application, based on a large model, can process massive amounts of multi-modal heterogeneous data, break down information silos, and mine deep knowledge connections, thereby improving the efficiency and accuracy of insurance business knowledge generation. Attached Figure Description

[0009] Figure 1 This is a schematic diagram of the application environment for the knowledge generation method based on insurance business provided in the embodiments of this application; Figure 2 This is a flowchart of the knowledge generation method based on insurance business provided in the embodiments of this application; Figure 3 yes Figure 2 The flowchart for step 202 in the document; Figure 4 This is another flowchart of the knowledge generation method based on insurance business provided in the embodiments of this application; Figure 5 yes Figure 2 The flowchart for step 203 in the document; Figure 6 yes Figure 2 The flowchart for step 204 in the document; Figure 7 yes Figure 2 The flowchart for step 205 in the document; Figure 8 This is a schematic diagram of the structure of the knowledge generation device based on insurance business provided in the embodiments of this application; Figure 9 This is a schematic diagram of the hardware structure of the electronic device provided in the embodiments of this application. Detailed Implementation

[0010] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0011] It should be noted that although functional modules are divided in the device schematic diagram and a logical order is shown in the flowchart, in some cases, the steps shown or described may be performed in a different order than the module division in the device or the order in the flowchart. The terms "first," "second," etc., in the specification, claims, and the aforementioned drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence.

[0012] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used herein is for the purpose of describing embodiments of this application only and is not intended to limit this application.

[0013] First, let's analyze some of the terms used in this application: Artificial intelligence (AI) is a new branch of computer science that studies, develops, and applies theories, methods, technologies, and systems to simulate, extend, and expand human intelligence. It aims to understand the essence of intelligence and produce intelligent machines that can react in a way similar to human intelligence. Research in this field includes robotics, speech recognition, image recognition, natural language processing, and expert systems. AI can simulate the information processes of human consciousness and thought. Furthermore, AI utilizes digital computers or machines controlled by digital computers to simulate, extend, and expand human intelligence, perceiving the environment, acquiring knowledge, and using that knowledge to achieve optimal results.

[0014] Natural Language Processing (NLP): NLP uses computers to process, understand, and utilize human language (such as Chinese and English). NLP is a branch of artificial intelligence and an interdisciplinary field of computer science and linguistics, often referred to as computational linguistics. NLP includes syntactic analysis, semantic analysis, and discourse understanding. It is commonly used in machine translation, handwritten and printed character recognition, speech recognition and text-to-speech conversion, intent recognition, information extraction and filtering, text classification and clustering, sentiment analysis, and opinion mining. It involves data mining, machine learning, knowledge acquisition, knowledge engineering, artificial intelligence research, and linguistic research related to language computation.

[0015] In the insurance industry, although insurance companies have generally established data storage and management platforms, there are significant shortcomings in transforming massive, heterogeneous data into knowledge that can directly empower business and assist decision-making. These shortcomings are mainly reflected in the following aspects: 1. Data silos and fragmented knowledge views: Internal data sources within insurance companies are highly complex and fragmented; structured data (such as actuarial rate tables, underwriting rule bases, and performance reports), unstructured internal data (such as scanned images of paper policies, claims investigation reports, and internal training documents), and external dynamic data (such as new regulatory rules and industry announcements) are usually isolated in independent systems (such as core business systems, CRM systems, and image management systems); these systems lack effective semantic-level fusion and association mechanisms, making it impossible to build a complete and unified insurance knowledge view, severely hindering cross-departmental and cross-flow... The challenges include: 1. Inefficient business collaboration; 2. Limitations of current processing technologies, such as insufficient semantic understanding. For example, current technologies rely heavily on single-point technology stacks. Traditional Optical Character Recognition (OCR) can only recognize characters in images, outputting disorganized text and failing to understand the logical structure of documents (such as titles, paragraphs, and tables) or the semantic relationships between content (such as the correspondence between "annual premium" and "insured amount"). Traditional Automatic Speech Recognition (ASR) can only convert speech to text, failing to capture customer tone, emphasis, or implied intent. 3. Difficulty in knowledge extraction and association. Even after converting multimodal data to text, extracting macro-entities in the insurance field, such as "policyholder," "insured object," and "exclusion clauses," and constructing their complex relationships remains a significant challenge, such as "a clause exempts liability for compensation under certain circumstances." Furthermore, overlapping relationships and nested entities make rule-based or traditional machine learning methods costly, have weak generalization capabilities, and are difficult to maintain. Insurance professionals still need to invest a lot of time in manual screening, comparison and integration, and cannot quickly obtain comprehensive, relevant and directly applicable structured knowledge for decision-making, which greatly reduces the empowering effect of technology on core business.

[0016] Based on this, embodiments of this application provide a method, apparatus, device, and storage medium for generating knowledge based on insurance business, aiming to improve the efficiency and accuracy of generating insurance business knowledge.

[0017] The knowledge generation method, apparatus, device, and storage medium based on insurance business provided in this application are specifically described through the following embodiments. First, the knowledge generation method based on insurance business in this application embodiment is described.

[0018] The embodiments of this application can acquire and process relevant data based on artificial intelligence technology. Artificial intelligence (AI) refers to the theories, methods, technologies, and application systems that use digital computers or machines controlled by digital computers to simulate, extend, and expand human intelligence, perceive the environment, acquire knowledge, and use that knowledge to obtain optimal results.

[0019] Foundational technologies for artificial intelligence generally include sensors, dedicated AI chips, cloud computing, distributed storage, big data processing, operating / interactive systems, and mechatronics. AI software technologies mainly encompass computer vision, robotics, biometrics, speech processing, natural language processing, and machine learning / deep learning.

[0020] The knowledge generation method based on insurance business provided in this application relates to the field of artificial intelligence technology. This knowledge generation method based on insurance business can be applied to a terminal, a server, or software running on either a terminal or a server. In some embodiments, the terminal can be a smartphone, tablet, laptop, desktop computer, etc.; the server can be configured as an independent physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server providing basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, CDN, and big data and artificial intelligence platforms; the software can be an application implementing the knowledge generation method based on insurance business, but is not limited to the above forms.

[0021] This application can be used in a wide variety of general-purpose or special-purpose computer system environments or configurations. Examples include: personal computers, server computers, handheld or portable devices, tablet devices, multiprocessor systems, microprocessor-based systems, set-top boxes, programmable consumer electronics, network PCs, minicomputers, mainframe computers, and distributed computing environments including any of the above systems or devices. This application can be described in the general context of computer-executable instructions executed by a computer, such as program modules. Generally, program modules include routines, programs, objects, components, data structures, etc., that perform specific tasks or implement specific abstract data types. This application can also be practiced in distributed computing environments where tasks are performed by remote processing devices connected via a communication network. In distributed computing environments, program modules can reside in local and remote computer storage media, including storage devices.

[0022] It should be noted that in all specific embodiments of this application, when processing data related to user identity or characteristics, such as user audio, user voice, user behavior, user historical data, and user attribute information, user permission or consent is obtained first. Furthermore, the collection, use, and processing of this data comply with relevant laws, regulations, and standards. In addition, when embodiments of this application require access to sensitive personal information of users, separate permission or consent from the user is obtained through pop-ups or redirects to confirmation pages. Only after obtaining the user's separate permission or consent is the necessary user-related data required for the proper functioning of these embodiments obtained.

[0023] The knowledge generation method based on insurance business provided in this application embodiment can be applied to, for example... Figure 1 In this application environment, a communication connection is established between the user terminal and the server. The user terminal can acquire multi-source, multi-modal data from insurance business scenarios, perform anonymization processing on the multi-source, multi-modal data, annotate the anonymized insurance data obtained through the anonymization process, and train a preset large language model based on the insurance knowledge data obtained from the annotation. Then, based on the trained multi-modal insurance data processing model, the insurance business data is uniformly encoded to obtain a fused feature representation. Knowledge generation is performed on the fused feature representation based on the multi-modal insurance data processing model to obtain insurance business knowledge such as underwriting conclusions and claims conclusions. The user terminal can send the insurance business knowledge to the server, which then performs underwriting or claims approval based on the insurance business knowledge. The user terminal can be, but is not limited to, various personal computers, laptops, smartphones, tablets, and portable wearable devices. The server can be implemented using a standalone server or a server cluster consisting of multiple servers. The following detailed description of specific embodiments further illustrates this application.

[0024] Figure 2 This is an optional flowchart of the knowledge generation method based on insurance business provided in the embodiments of this application. Figure 2 The method may include, but is not limited to, steps 201 to 206.

[0025] Step 201: Obtain multi-source, multi-modal data for insurance business scenarios; wherein, the multi-source, multi-modal data includes at least two of the following: image data, text data, voice data, and video data; Step 202: De-identify the multi-source, multi-modal data to ensure data security; wherein, the de-identified data is multi-modal data; Step 203: Annotate the de-identified insurance data to obtain insurance knowledge data; Step 204: Train the preset large language model based on insurance knowledge data to obtain a multimodal insurance data processing model; Step 205: Based on the multimodal insurance data processing model, the pre-acquired insurance business data is uniformly encoded to obtain a fused feature representation; wherein, the insurance data includes at least one of the following: image data, text data, voice data, and video data; Step 206: Generate knowledge from the fused feature representation based on the multimodal insurance data processing model to obtain insurance business knowledge; wherein, the insurance business knowledge includes at least one of the following: underwriting conclusion, claims conclusion.

[0026] Steps 201 to 206 of this application embodiment involve acquiring multi-source, multi-modal data from insurance business scenarios, de-identifying the multi-source, multi-modal data, labeling the de-identified insurance data, training a preset large language model based on the labeled insurance knowledge data, and then uniformly encoding the insurance business data based on the trained multi-modal insurance data processing model to obtain a fused feature representation. Knowledge generation is then performed on the fused feature representation based on the multi-modal insurance data processing model to obtain insurance business knowledge such as underwriting conclusions and claims conclusions. This application, based on a large model, can process massive amounts of multi-modal heterogeneous data, break down information silos, and uncover deep knowledge connections, thereby improving the efficiency and accuracy of insurance business knowledge generation.

[0027] In step 201 of some embodiments, the multi-source multimodal data can originate from the insurance company's internal business systems, such as the core business system, claims system, and underwriting system. It can also originate from external data sources, such as regulatory agency websites, industry databases, and public networks. In one application scenario, multi-source multimodal data includes at least two of the following: structured data, image data, text data, voice data, and video data. In practical applications, multi-source multimodal data can include full-modal data such as image data, text data, voice data, and video data. Image data can be, for example, insurance policies, documents, inspection photos, electronic medical records, and invoices; documents can be, for example, ID cards; text data can be, for example, claims applications, medical information, bank account information, underwriting conclusions, medical examination report summaries, and health declarations; text data can be PDF files; voice data can be, for example, customer service calls, sales recordings, and inspection voice recordings; and video data can be, for example, videos of remote inspections and face-to-face visits. Furthermore, multi-source, multi-modal data can include structured data, unstructured data, and external dynamic data. Structured data can include, for example, actuarial rate tables, underwriting rule bases, and performance reports. Unstructured internal data can include, for example, scanned images of paper insurance policies, claims investigation reports, and internal training documents. External dynamic data can include, for example, new regulatory rules and industry announcements. In practice, the aforementioned structured data, unstructured data, and external dynamic data are usually isolated in independent systems, such as core business systems, CRM systems, and image management systems, resulting in data silos. Through the embodiments of this application, data silos can be broken down, providing a unified semantic standard for data from different sources and in different forms, constructing a complete and unified insurance knowledge view, and achieving efficient business collaboration across departments and processes.

[0028] Please see Figure 3 In some embodiments, step 202 may specifically include, but is not limited to: Step 301: Preprocess the multi-source multimodal data to obtain multi-source standardized data; Step 302: Perform semantic recognition on the multi-source standardized data to obtain multi-source semantic data; Step 303: Classify the multi-source semantic data to obtain sensitive categories; Step 304: De-identify the multi-source, multi-modal data based on the sensitive categories to obtain de-identified insurance data.

[0029] In steps 301 and 302 of some embodiments, to address the blurring, tilting, and shadowing issues in inspection photos and ID photos, adaptive threshold binarization, tilt correction, and shadow removal + noise reduction processing can be employed to ensure image clarity. For example, tilt correction can be achieved by adjusting the angle by ±15°. Specifically, tools such as OpenCV (Open Source Computer Vision Library) and PIL (Python Imaging Library) can be used. For PDF file preprocessing, this includes preprocessing both editable and scanned PDF files. For editable PDF files, text can be directly extracted. For scanned PDF files, they are first converted to image data before image preprocessing. For video data, the video data is first converted to image data before image preprocessing. Enhanced OCR (Optical Character Recognition) can be used for preliminary parsing of image data and PDF files, preserving layout information, thereby improving the recognition accuracy of insurance terminology and handwritten text (such as claims notes). The parsed text is matched with the layout of the original image or PDF file, and the position coordinates of each text fragment are recorded (e.g., "Insured's Name" is in the upper left corner of page 2 of the PDF, with coordinates x=100, y=200). This preserves the layout information, facilitating the location of image areas during subsequent expert annotation. Taking a scanned PDF of a medical invoice collected by a property insurance company as an example, during preprocessing, it is first converted to a 300dpi image. OpenCV is used to remove shadows and correct tilt (e.g., tilt angle 8°). Then, enhanced OCR technology (e.g., loading an insurance and medical industry dictionary) is used for parsing, thereby recognizing text such as "hospital visited, treatment items, cost amount, and medical insurance reimbursement amount," while preserving the table layout in the invoice (e.g., the row and column correspondence of the cost details table). Preprocessing blurry photos of the investigation site and then using OCR recognition can improve accuracy, thereby recognizing information such as license plate numbers and accident location markings in the photos.

[0030] For voice data, format standardization can be performed first, followed by noise reduction and silence removal to facilitate speech-to-text transcription. In one application scenario, voice data format standardization can include converting different formats (e.g., MP3, WMA) to WAV, using mono for customer service calls and stereo for remote inspections to ensure consistency. Noise reduction can include removing environmental noise (e.g., background noise in the customer service center, wind noise at the inspection site) while preserving voice clarity. Silence removal can include removing silent segments at the beginning and end of the speech (e.g., duration <0.5 seconds) to reduce invalid transcription content and improve efficiency. In practical scenarios, an insurance industry-tuned ASR model can be used to transcribe voice data, improving the accuracy of transcribing technical terms and avoiding homophone errors. Taking a customer service call recording from a life insurance company as an example, the voice data is in MP3 format with two channels. During preprocessing, it is converted to 16kHz, 16bit WAV format, and background noise is removed using Librosa, with the first and last silent segments removed. During ASR transcription, an insurance industry fine-tuning model is loaded to identify voice content such as "I want to purchase a critical illness insurance policy, how much is the premium?" and "What are the exclusions for this critical illness insurance policy?" and retain timestamps. For example, the timestamp corresponding to the customer service's reply "the premium is 5,000 yuan per year" is 00:00:10.500-00:00:12.300, with channel 1 marked as customer service and channel 2 marked as customer. After transcription, it was found that "exclusions" was mistakenly written as "exclusion clauses". This was marked and manually corrected to facilitate the subsequent identification of sensitive information.

[0031] Step 303 in some embodiments may include, but is not limited to: Sensitive identification is performed on multi-source semantic data to obtain semantically sensitive information; Based on semantically sensitive information, multi-source semantic data is classified to obtain sensitive categories.

[0032] In some embodiments, sensitive identification of multi-source semantic data can be performed based on a pre-trained sensitive identification model. This sensitive identification model can be based on the BERT model, so as to perform sensitive classification through the BERT model, and the obtained sensitive categories are used to represent the sensitive information level. In practical applications, the sensitive information level can include high level, medium level, low level, etc. The embodiments of the present application do not limit the sensitive information level. The high-level multi-source semantic data can include, but is not limited to, personal identity information, financial information, core medical information, ID card number, mobile phone number, bank card number, claim amount, core disease diagnosis (such as "lung cancer"), etc. The medium-level multi-source semantic data can include, but is not limited to, personal basic information, non-core medical information, policy information, address, occupation, policy number (incomplete), minor diseases (such as "cold"), etc. The low-level multi-source semantic data can include, but is not limited to, business-related basic information, insurance type name, insurance status, customer service staff number, etc.

[0033] In step 304 of some embodiments, for high-level multi-source semantic data, an encryption + desensitization processing method can be adopted. For medium-level multi-source semantic data, on-demand desensitization can be used for processing. For low-level multi-source semantic data, it can be displayed and marked. After desensitizing the multi-source multi-modal data through steps 301 to 304, it is used for the on-chain evidence storage processing in step 202.

[0034] In steps 303 and 304 of the actual application scenario, for text data with a high sensitive information level, such as ID card numbers and mobile phone numbers, masking desensitization can be adopted. For bank card numbers, format-preserving encryption can be used. For core medical diagnoses (such as "lung cancer"), replacement desensitization can be used, for example, replaced with "malignant tumor diseases". For image data with a high sensitive information level, for example, the face area can be blurred, such as using Gaussian blur processing. The ID card number and bank card number areas can be masked desensitized, such as using a black rectangle to mask, making the size consistent with the field area. For text data with a medium sensitive information level, partial desensitization or generalization processing can be adopted. For example, for address information, generalization processing can be performed. For example, "XX Community, C Yang District, B Beijing City" is generalized to "C Yang District, B Beijing City". For the policy number, partial masking processing can be performed. For example, after partially masking "INS20260226001", "INS20260226***" is obtained. For image data with a medium sensitive information level, for example, non-core medical bill information (such as the department of visit) can be shallowly blurred without affecting the overall recognition. For sensitive information with a low sensitive information level, only sensitive marking is performed without desensitization, and the original content is retained, such as information like "critical illness insurance" and "insurance status: in effect".

[0035] In some embodiments, after step 304, step 202 may include, but is not limited to, generating a de-identification log. Specifically, all de-identification operations are recorded and a de-identification log is generated. The de-identification operations include de-identification time, de-identification method, de-identification fields, operator / system, etc., for subsequent blockchain evidence storage and traceability in step 202.

[0036] Please see Figure 4 In some embodiments, after step 201, the knowledge generation method based on insurance business may also include, but is not limited to, performing on-chain notarization processing on multi-source, multi-modal data, specifically including but not limited to: Step 401: Extract information from multi-source, multi-modal data to obtain core insurance data; Step 402: Generate hash values ​​based on core insurance data; Step 403: Obtain data acquisition information for multi-source, multi-modal data, wherein the data acquisition information includes at least the acquisition time and data source identifier; Step 404: Based on the hash value and data collection information, perform on-chain evidence storage processing on the core insurance data.

[0037] In step 401 of some embodiments, the core insurance data includes at least one of the following: image data, text data, voice data, and video data. By extracting information from multi-source, multi-modal data, the obtained core insurance data may include, but is not limited to: photos of the policyholder / insured's ID card (front / back), scanned copies of medical examination reports, medical invoices (outpatient / inpatient invoices, expense lists), photos / videos of the investigation site, original policy documents (electronic files), original claim application forms, original underwriting conclusions, original claim settlement conclusions, and third-party authorized data (such as hospital medical records, public security identity verification information, etc.). The above data directly affects the underwriting or claim settlement results and is the core evidence, thus serving as core insurance data. Non-critical redundant data (such as irrelevant customer service chat records, repeatedly collected invalid data, etc.) and de-identified derived data are only stored as original data; the derived data is associated with the hash value of the original data and is not stored on the blockchain, thereby avoiding invalid evidence storage. In this embodiment, only the core insurance data is stored on the blockchain, without full data storage, thereby avoiding redundancy, reducing storage costs, and improving storage efficiency.

[0038] In step 402 of some embodiments, to avoid excessive on-chain storage pressure and low transmission efficiency caused by directly uploading the original core insurance data (such as large files like images and videos) to the blockchain, a mode of offline data storage and hash value uploading is adopted. The hash value generation follows the principles of uniqueness and irreversibility. In some application scenarios, the SM3 hash algorithm is used, and a differentiated hash generation method is adopted for core insurance data of different modalities to ensure the uniqueness of hash values. Specifically, for text data such as insurance policies and claims conclusions, the plain text content is directly subjected to SM3 hash calculation to generate a 32-byte hash value, which is associated with a unique text identifier, such as the policy number or claim number. For image data or video data such as ID card photos and inspection videos, the image data or video data is first divided into blocks, for example, the block size is 1MB. SM3 hash calculation is performed on each block, and then a second SM3 hash is performed on the hash values ​​of all blocks to generate the hash value of the final file. By dividing the data into blocks, it is possible to avoid the failure to detect tampering with a single block of data. Furthermore, for preprocessed data such as OCR parsing results and ASR transcribed text, SM3 hash values ​​can be generated separately, which can be associated with the hash values ​​of the original data, thereby realizing the traceability and association between the original data and the preprocessed data.

[0039] In some embodiments, the on-chain notarization process for multi-source, multi-modal data may include, but is not limited to, hash value verification. Specifically, after generating the hash value, a dual verification is performed immediately, such as local verification + node verification. Local verification compares the data integrity before and after hash value generation, while node verification involves two or more consensus nodes simultaneously calculating the hash value. Only after the hash value matches the locally generated hash value can the on-chain notarization process begin, thus preventing hash value generation errors. Further technical details regarding hash value verification are not limited in the embodiments of this application.

[0040] In step 403 of some embodiments, data collection information is also required as auxiliary evidence storage data, such as the collection time, collection personnel, system identifier, and data source information of the core insurance data. This data collection information is used to support the integrity of the traceability chain. Then, in step 404, hash value association is performed, and the core insurance data is processed for on-chain evidence storage based on the hash value and the data collection information, thereby obtaining a compliant insurance data source. Specifically, the hash value, unique data identifier (such as data ID), collection time, data source identifier, and preprocessing operation records of each core insurance data are associated to obtain a compliant insurance data source, forming a hash value-data source information mapping table for easy subsequent traceability queries.

[0041] Taking claims investigation data from a property insurance company as an example, the claims investigation data includes one investigation photo, one customer service follow-up recording, and one claims application form. After preprocessing and anonymization, the on-chain storage on the blockchain can include: generating SHA-256 hash values ​​for the original investigation photo, recording, and claims application form; generating hash values ​​for the anonymized photo (e.g., blurred face, obscured license plate number), recording (e.g., anonymized phone number), and claims application form (e.g., obscured ID card number); the packaged storage content includes: double hash values ​​and an anonymized log: "2026-02-26". At 16:00:00, Gaussian blurring was applied to the faces in the inspection photos. The operator was the "Claims System" and the data metadata was: Data ID=LC20260226002. The modal type was: text + image + video, and the business type was claims. After signing with the claims system's private key, the data was sent to the blockchain node, triggering the consensus mechanism. After consensus was reached, the evidence was packaged into a block, written into the blockchain, and stored. After storage, a storage certificate was generated for easy traceability.

[0042] In some embodiments, the anonymized insurance data has a modality category, which is used to characterize whether the anonymized insurance data is image data, text data, voice data, or video data. See also... Figure 5 In some embodiments, step 203 may include, but is not limited to: Step 501: Randomly sample and screen the de-identified insurance data based on modal categories to obtain candidate insurance data; Step 502: Perform entity annotation on the candidate insurance data to obtain entity-annotated data; Step 503: Perform relation annotation based on entity annotation data to obtain relation annotation data; Step 504: Perform sentiment annotation on the candidate insurance data to obtain sentiment-annotated data; Step 505: Based on entity annotation data, relationship annotation data, and sentiment annotation data, knowledge is constructed to obtain insurance knowledge data.

[0043] In step 501 of some embodiments, candidate insurance data can be selected by hierarchical random sampling based on modality category. For example, 15% can be randomly selected from text data, 10% from image data, and 8% from voice data. In this case, video data is first converted into image data, and de-identified insurance data from high-frequency business scenarios, such as de-identified insurance data from critical illness insurance, car insurance claims, and medical insurance reimbursement scenarios, are selected first to ensure that the labeled insurance knowledge data is representative.

[0044] In step 502 of some embodiments, the voice data is converted into text data. For the text data, the core entities in the insurance field are labeled, and the entity categories and boundaries are clarified. The core entity categories include: policyholder information, insurance products, medical-related information, and underwriting / claims-related information. Among them, policyholder information may include, but is not limited to, name, ID number, occupation, etc.; insurance products may include, but is not limited to, insurance type name, premium, and coverage period; medical-related information may include, but is not limited to, disease name, treatment items, and drugs; and underwriting / claims-related information may include, but is not limited to, underwriting conclusion, claim amount, and cause of the accident, etc.

[0045] In step 503 of some embodiments, the relationships between entities can be labeled based on the entity labeling results of step 502 to clarify the relationship types and constraints. The relationship types include core relationships, which may include, but are not limited to: insurance relationship (e.g., policyholder-insurance type), subordinate relationship (e.g., insurance type-exclusion clause), related relationship (e.g., disease-treatment items), and causal relationship (e.g., cause of loss-claim conclusion). In practical applications, the triple format of "entity 1-relationship type-entity 2" can be used for labeling, for example, labeled as: Zhang San-insurance-lifetime critical illness insurance, or labeled as: lifetime critical illness insurance-includes-exclusion clause.

[0046] In step 504 of some embodiments, sentiment annotation is performed on information such as customer service call content and claim report descriptions in the candidate insurance data to indicate sentiment tendency and intensity. Sentiment tendency can include three categories: positive, neutral, and negative. Positive sentiment can include, for example, satisfaction with claim processing efficiency, neutral sentiment can include consultation on the insurance application process, and negative sentiment can include questioning the reasons for claim rejection. Sentiment intensity can be divided into five levels, with level one to level five representing increasing sentiment intensity, with level one being the weakest and level five being the strongest. In practical application scenarios, annotation is performed in conjunction with contextual semantics to avoid misjudgment based on a single statement.

[0047] In step 505 of some embodiments, entity annotation data, relation annotation data, and sentiment annotation data are merged and knowledge is constructed to obtain insurance knowledge data. This insurance knowledge data is then stored to facilitate model training in subsequent step 204. The insurance knowledge data obtained through step 505 is a massive, high-quality insurance corpus.

[0048] By constructing the insurance domain ontology through step 505, a standardized knowledge system framework can be formed. Specifically, a top-level design can be adopted to clarify the core levels of insurance knowledge data. For example, it can be divided into four levels: insurance entities, insurance products, business processes, and knowledge constraints. Among them, insurance entities may include, but are not limited to, policyholders, insured persons, insurers, and third-party institutions; insurance products may include, but are not limited to, types of insurance, premiums, and coverage periods; business processes may include, but are not limited to, underwriting, and claims; and knowledge constraints may include, but are not limited to, exclusion clauses, underwriting rules, and regulatory requirements. The above four levels can be further subdivided into sub-levels to form a hierarchical structure. In practical applications, the policyholder includes: a person who enters into an insurance contract with the insurer and is obligated to pay the premium as stipulated in the insurance contract, and must be associated with attributes such as name, ID number, and contact information; the type of insurance includes: the type of insurance product provided by the insurer, which is divided into subcategories such as critical illness insurance, medical insurance, car insurance, and life insurance, and must be associated with attributes such as premium, coverage period, and exclusion clauses; exclusion clauses are: clauses in the insurance contract that stipulate that the insurer is not liable for compensation or payment of insurance benefits, and must be associated with attributes such as applicable scenarios and scope of exclusion; the underwriting conclusion includes: the insurer's review result of the application, which is divided into four categories: approved, rejected, postponed, and increased premium, and must be associated with attributes such as underwriting reasons and risk factors.

[0049] Based on the relationship annotation results obtained in step 503, hierarchical relationships and association relationships are determined, and relationship constraints are clarified. Hierarchical relationships can be represented as, for example, critical illness insurance - is-a - type of insurance, or as, policyholder - is-a - insured entity. Furthermore, parent-child hierarchies can be constructed to ensure the logical consistency of the knowledge system. Association relationships can include, for example, policyholder-application-type of insurance, inclusion-type of insurance, type of insurance-inclusion-exclusion clauses, review-underwriting-application, and claims-payment relationships. Policyholder-application-type of insurance, inclusion-type of insurance, review-underwriting-application, and claims-payment relationships are all possible. Each relationship has clearly defined constraints, such as the requirement that "policyholder has full civil capacity" for the "policyholder-application relationship." Constraint relationships are used to define attribute constraints, such as requiring ID number attributes to be in 18-character format, and requiring underwriting conclusions to be associated with corresponding underwriting reasons, to ensure the standardization of insurance knowledge data.

[0050] In step 203 of some embodiments, it may further include, but is not limited to: performing image annotation on candidate insurance data, specifically including: for image data such as ID card photos, medical bills, and survey photos, annotating key regions and corresponding semantics. For example, the core regions can be annotated in the way of rectangular box annotation + label description. The core regions may include, but are not limited to: the portrait region and number region of the ID card, the amount region and diagnosis and treatment item region of the medical bill, and the damaged part region of the survey photo. The annotation labels need to be associated with text entities. For example, the region "110101********1234" in the ID card photo is annotated as "ID number", associating with the text entity "ID number".

[0051] In step 204 of some embodiments, the preset large language model can be a general large language model that supports multi-modal inputs such as vision and language. In an application scenario, Qwen-VL-Chat can be used as the base model.

[0052] Please refer to Figure 6 , in step 204 of some embodiments, it may include, but is not limited to: Step 601, preprocessing the insurance knowledge data to obtain preprocessed insurance data; Step 602, extracting key information from the preprocessed insurance data to obtain insurance key information; Step 603, performing masking processing on the insurance key information in the preprocessed insurance data to obtain masked insurance data; Step 604, training the preset large language model based on the masked insurance data to obtain a multi-modal insurance data processing model.

[0053] In step 601 of some embodiments, preprocessing the insurance knowledge data may include, but is not limited to: using processing methods such as word segmentation, stop word removal, and normalization for the text data in the insurance knowledge data. Among them, the jieba word segmentation tool can be used, and an exclusive dictionary for the insurance industry is added, such as "extra premium underwriting", "subrogation recovery", "waiting period", etc. Stop word processing includes meaningless stop words such as "de", "di", "de". Normalization processing includes: unifying the date format, amount format, and term expression. For example, "auto insurance" is uniformly written without "motor vehicle insurance"; for the image data in the insurance knowledge data, size normalization, grayscale enhancement, and denoising processing are used. Among them, size normalization processing can include uniformly adjusting to 1024×768 pixels, grayscale enhancement processing can include improving the clarity of blurred bills and photos, and denoising processing can include removing noise points and shadows in the image, so as to ensure that the quality of the insurance knowledge data meets the model training requirements.

[0054] In one application scenario, textual data within insurance knowledge data can include, for example, underwriting rules, policy terms, claims cases, medical terminology, regulatory policies, and industry standards. These can be categorized into four main types based on the scenario: critical illness insurance, medical insurance, auto insurance, and life insurance. For example, a critical illness insurance clause might state: "The coverage of this lifetime critical illness insurance includes 28 major illnesses such as malignant tumors, acute myocardial infarction, and sequelae of stroke, as well as 3 minor illnesses. Upon diagnosis of a covered illness, the insurance company will pay a lump sum of the basic sum insured. Exclusions include intentional suicide, drunk driving, and illnesses caused by congenital diseases. Diagnosis must be confirmed within 90 days of the 90-day waiting period." "No liability for illness"; underwriting rules may include: "For insured persons aged 30-40 applying for medical insurance, if the medical examination report shows blood pressure ≥140 / 90mmHg, an additional 10% premium will be required for coverage; if blood pressure ≥160 / 100mmHg, coverage will be rejected directly"; a summary of a claims case may include: "The insured suffered a right femoral fracture due to an accidental fall and received treatment at a designated tertiary hospital, incurring medical expenses of 23,000 yuan, of which 12,000 yuan was reimbursed by medical insurance, and the insurance company paid the remaining 11,000 yuan according to the terms of the policy. The claim was based on the accidental medical insurance terms, with no exclusions." In steps 602 to 604 of some embodiments, key information is extracted from the insurance preprocessing data. The obtained key insurance information may include core insurance terms, key business information, etc. In this embodiment, a Masked Language Model (MLM) is used to model the key insurance information such as core insurance terms and key business information. The masking ratio can be set to 15%, with 80% of the masked positions replaced by the "<mask>" identifier, 10% of the masked positions replaced by random insurance domain terms, such as replacing "policyholder" with "insured" after masking, and 10% of the masked positions retaining the original terms for model self-verification. This avoids the model relying too much on context guessing and improves the accuracy of learning insurance domain terms.

[0055] In some application scenarios, the masking language model can be forced to predict the correct content at the mask position. By optimizing the parameters of the masking language model through gradient descent, the masking language model can learn the contextual logic of insurance terms and the expression habits of business rules, ensuring that the masking language model can accurately identify and understand core insurance vocabulary and sentence structures.

[0056] Taking the term "mask" as an example, the mask content in "<mask> refers to the person who enters into an insurance contract with the insurer and is obligated to pay the insurance premium according to the insurance contract" is predicted as "policyholder"; taking the rule mask as an example, the mask content in "If a policyholder aged 30-40 applies for medical insurance and the medical examination report shows that the blood pressure is ≥140 / 90mmHg, <mask> 10% coverage is required" is predicted as "surcharge"; in addition, the mask content in "The insured suffered a right femoral fracture due to <mask> and received treatment at a designated tertiary hospital, and the insurance company paid compensation according to the terms and conditions" can also be predicted as "accidental fall".

[0057] In some embodiments, step 204 further includes multimodal contrastive learning, specifically including: Construct multimodal contrastive sample pairs; Construct a multimodal contrast sample set based on the multimodal contrast sample pairs; The multimodal contrast sample groups are encoded to obtain sample vector representations; Loss data is obtained by calculating the loss based on the sample vector representation. The pre-defined large language model is trained on a multimodal contrastive sample group based on the loss data.

[0058] Specifically, a multimodal contrast sample pair includes data from two different modalities, such as image data and text data. A multimodal contrast sample pair can be represented as an image-text paired corpus, including positive sample pairs and negative sample pairs. Each multimodal contrast sample group contains one anchor sample, one positive sample, and four negative samples to ensure a balanced sample distribution. The positive sample pair is an image and text in the same insurance scenario, such as "hospital invoice photo" and the corresponding invoice text description. The negative sample pair is an image and text in different insurance scenarios, such as "hospital invoice photo" and "car insurance survey text description". The negative sample pair can also be an image and text in the same scenario but with semantically mismatched elements, such as "hospital invoice photo" and "outpatient invoice text description".

[0059] Specifically, the image-text pairing corpus can include, for example, image data such as ID card photos, medical bills, inspection photos, and scanned insurance policies, paired with corresponding precise text descriptions to ensure complete semantic matching between images and text. The text description must contain the core information of the image. Taking medical bill pairing as an example: the image data can be a scanned copy of a critical illness insurance policy (first page), and the corresponding text description can be: Lifetime Critical Illness Insurance Policy, Policyholder: Li Si, Insured: Li Si, Coverage Amount: 500,000 RMB, Annual Premium: 5,800 RMB, Coverage Period: Lifetime, Policy Date: September 1, 2025, Waiting Period: 90 days, Insurance Company: XX Insurance Company. In some application scenarios, a contrastive loss function is used to calculate the loss. Specifically, the image data and text data in the multimodal contrastive sample group are respectively input into the visual encoder and text encoder of the Qwen-VL-Chat model for encoding. The generated visual vector representation and text vector representation are used as sample vector representations. The loss function minimizes the vector distance of the same positive sample pair and maximizes the vector distance between the anchor sample and the negative sample pair, so that the model learns to associate the same insurance semantics of different modalities.

[0060] Taking an insurance scenario as an example, the anchor sample is image data: a fracture X-ray (hand), the positive sample is text data: "Hand fracture X-ray shows a continuous fracture of the fifth metacarpal bone of the right hand, which meets the diagnostic criteria for accidental fracture", the negative sample 1 is text data: "Leg fracture CT scan shows a fracture of the left tibia", the negative sample 2 is text data: "Outpatient invoice, patient Zhang San, treatment item: blood routine", the negative sample 3 is the text description corresponding to "car insurance survey photo", and the negative sample 4 is text data: "Critical illness insurance policy terms summary". During the model training process, the vector distance between "fracture X-ray" and the corresponding positive sample text is forced to approach 0, and the vector distance between "fracture X-ray" and the four negative sample texts is forced to be greater than a preset threshold, where the preset threshold can be set to, for example, 0.6.

[0061] In step 204, training samples containing counterfactual conditions are constructed, such as "If the insured does not receive treatment at the designated hospital, is the medical expense within the scope of the claim liability?" This training enhances the Qwen-VL-Chat model's ability to analyze causal logic and judge complex situations in underwriting, claims, and other scenarios. By employing contrastive learning techniques, different modalities describing the same insurance concept (such as a text description of "fracture" and an X-ray film) are forced to be close in distance in the vector representation space of the Qwen-VL-Chat model, while being pushed away from irrelevant concepts, thereby ensuring the consistency of the Qwen-VL-Chat model's understanding of different forms of information.

[0062] In step 204, the basic model Qwen-VL-Chat is continuously pre-trained on a large scale using a massive, high-quality insurance corpus, masked language modeling, and multimodal contrastive learning, thereby injecting rich insurance domain knowledge into the Qwen-VL-Chat model parameters.

[0063] In step 205 of some embodiments, the insurance data has a modal category, which is used to characterize the insurance data as image data, text data, voice data, or video data. Image data can be, for example, an insurance policy, document, inspection photos, electronic medical records, invoices, etc. Documents can be, for example, an ID card. Text data can be, for example, a claim application, medical information, bank account information, underwriting conclusions, medical examination report summaries, health declarations, etc. Text data can be PDF files. Voice data can be, for example, customer service calls, sales recordings, inspection voice recordings, etc. Video data can be, for example, videos of remote inspections, face-to-face interviews, etc. For video data, it is first converted into image data. This embodiment of the application employs differentiated encoding strategies for the three modal categories of text, voice, and image to ensure that the core information of each modal category can be accurately extracted, thereby laying the foundation for feature fusion.

[0064] Please see Figure 7 In one embodiment, step 205 may include, but is not limited to: Step 701: Encode the insurance data separately according to the modal category of the insurance data to obtain single-modal coded data; Step 702: Perform quality verification on the single-modal encoded data to obtain valid encoded data; Step 703: Normalize the valid encoded data to obtain normalized encoded data; Step 704: Perform cross-modal alignment on the normalized coded data to obtain aligned coded data; Step 705: Perform cross-modal attention fusion on the aligned encoded data to obtain the fused feature representation.

[0065] In step 701 of one embodiment, the insurance data is encoded separately according to the modal category of the insurance data to obtain single-modal encoded data; specifically, for video data, the video data is first converted into image data, and the image data, text data and voice data are encoded separately, so that the obtained single-modal encoded data includes image encoded data, text encoded data and voice encoded data; In step 702 of one embodiment, by performing quality verification on the single-modal encoded data and removing abnormal features such as feature distortion and missing information, effective encoded data can be obtained. Specifically, abnormal features include, for example, missing core terms in text encoded data, unrecognized core business areas in image encoded data, and too many transcription errors in speech encoded data. The original single-modal encoded data corresponding to the abnormal features is re-encoded. If the abnormality persists after multiple encodings, manual review is triggered.

[0066] In step 703 of one embodiment, L2 normalization can be used to normalize the effective encoded data, so as to unify the scale of the three single-modal features to the [0,1] interval and eliminate the fusion bias caused by the difference in feature scale.

[0067] In step 704 of one embodiment, the normalized encoded data of the three single modalities are mapped to the same 768-dimensional feature space through a cross-modal semantic alignment algorithm to ensure semantic consistency. For example, the encoded features of "femoral fracture" in the text modality, the encoded features of "femoral fracture X-ray" in the image, and the encoded features of "femoral fracture" in the speech transcription are semantically aligned to ensure that the three are close in distance in the feature space and to avoid semantic misalignment.

[0068] In step 705 of one embodiment, an improved multi-head cross-modal attention mechanism is used to achieve cross-modal attention mechanism fusion. The number of attention heads can be 16. Deep interactive fusion is performed on the encoded features of the three single modalities, and the attention weights of each modality feature are dynamically allocated. The weight allocation strategy is dynamically adjusted based on the insurance business scenario. Furthermore, the dominant modality in the core business scenario is given higher weight. Taking the underwriting scenario as an example, the feature weight of the text modality can be set to 0.45. The features of the text modality can be, for example, policy terms or medical examination report text. The feature weight of the image modality can be set to 0.4. The features of the image modality can be, for example, scanned copies of medical examination reports or ID card photos. The feature weight of the speech modality can be, for example, set to 0.15. For example, the feature weight for an image modality could be 0.45, and the features could be medical bills or inspection photos. The feature weight for a text modality could be 0.35, and the features could be a claim application or medical record summary. The feature weight for a voice modality could be 0.2, and the features could be a claim report recording with a weight of 0.2. For example, in an insurance consultation scenario, the feature weight for a voice modality could be 0.5, and the features could be customer service call recordings. The feature weight for a text modality could be 0.35, and the features could be consultation records. The feature weight for an image modality could be 0.15.

[0069] During the fusion process, the attention mechanism automatically identifies the core business information in each modality feature, assigns higher weights to features that are highly relevant to the current business scenario, and weakens redundant information. For example, in the car insurance claims scenario, the image feature weight of the inspection photo will be automatically increased to 0.5, while the weight of casual conversation content unrelated to claims in the voice features will be reduced to 0.05.

[0070] In another embodiment, step 205 may include, but is not limited to, feature optimization. Specifically, feature optimization is performed on the fused feature representation through a fully connected layer. This fully connected layer can adopt a two-layer structure, with the hidden layer dimension set to 1024. The first fully connected layer is responsible for dimensional transformation and feature reorganization of the initially fused feature representation to extract more representative business features. The second fully connected layer is responsible for mapping the optimized fused feature representation to a 768-dimensional feature space to obtain the final fused feature representation. During feature optimization, a dropout layer can be added to suppress overfitting and ensure the generalization ability of the fused features. Batch Normalization is used to eliminate feature distribution differences and improve the stability of the fused features. The final output fused feature representation is a 768-dimensional vector, with each dimension corresponding to a core business feature of the insurance multimodal data, including business logic of text, visual evidence of images, and semantic supplementation of speech, possessing completeness, consistency, and business relevance, so as to facilitate insurance knowledge generation in step 206.

[0071] In some embodiments, step 206, based on a multimodal insurance data processing model, may include, but is not limited to: Extract the business task identifier based on the fusion feature representation; The prompt macro template is invoked based on the business task identifier; the prompt macro template includes the user role, task information, and output requirements; Information is extracted based on fused feature representations to obtain target business information; Risk identification is performed based on target business information to obtain risk identification information; Construct micro-templates for prompt words based on macro-templates for prompt words and risk identification information; Knowledge is generated based on the prompt word micro-template to obtain insurance business knowledge.

[0072] In some embodiments, a business task identifier is extracted by fusing feature representations. This business task identifier indicates the business scenario, such as a car insurance claims scenario or a critical illness insurance underwriting scenario. This facilitates subsequent invocation of relevant prompt macro templates. These prompt macro templates define the overall task framework and set system-level instructions, limiting the macro boundaries of user roles, task information, and output requirements. This provides a clear task orientation for the multimodal insurance data processing model, avoiding problems such as task deviation and role misalignment. The prompt macro template includes role definitions, task descriptions, input information descriptions, and output format requirements, where the role definition clarifies... The role positioning of the multimodal insurance data processing model, combined with insurance business scenarios, positions the model as a professional insurance underwriter and a professional insurance claims adjuster, while assigning corresponding professional capabilities and business permissions to each role. For example, "You are now a professional underwriter with more than 5 years of experience in critical illness insurance underwriting, familiar with critical illness insurance underwriting rules, medical terminology, and relevant regulatory requirements of the State Administration of Financial Supervision and Management. You can accurately identify risk factors in the insurance application data and provide reasonable underwriting conclusions based on underwriting rules." Through role definition, the multimodal insurance data processing model is guided to reason from a professional business perspective, avoiding unprofessional expressions; the task description is used to clarify... Clearly define the core objectives, scope, and requirements of the current task. In conjunction with the specific business scenario, clarify the core demands of the task. For example, "Based on the fusion feature representation of the input insurance application text, medical examination report image, and insurance consultation recording, and in conjunction with critical illness insurance underwriting rules, complete the underwriting reasoning, provide a clear underwriting conclusion and detailed underwriting reasons, and the reasoning process must comply with the underwriting business process." The input information description clarifies the data type, source, and core information of the multimodal insurance data processing model, guiding the model to focus on key target business information. For example, "The input information is a uniformly coded fusion feature representation, including the policyholder's basic information and the insurance product..." "For core business information such as information, physical examination indicators, and past medical history, please extract risk factors from the fusion feature representation and combine them with underwriting rules for reasoning." This helps the multimodal insurance data processing model quickly locate the core content in the input data and improve reasoning efficiency. The output format requirements are used to clarify the macro format of the output results and guide the multimodal insurance data processing model to output results as required. For example, "The output results must strictly follow the structured format of critical illness insurance underwriting (e.g., JSON format), including mandatory fields such as underwriting conclusion, underwriting reason, risk factors, and reasoning process. Fields must not be omitted or modified, and the format must be standardized and free of syntax errors." This provides macro guidance for subsequent operations.

[0073] In some embodiments, risk identification information may include, but is not limited to, risk type, risk level, and risk description. Risk type may include health risk, occupational risk, and financial risk; risk level may include low risk, medium risk, and high risk; risk description is used to describe in detail the specific circumstances of each risk factor, and, in conjunction with the information extraction results, to explain the source of the risk factor and its impact on underwriting.

[0074] Step 206 of this embodiment solves the technical pain points of traditional prompting engineering in insurance underwriting and claims scenarios, such as chaotic reasoning logic, inconsistent output format, and poor business adaptability, through the layered design of macro templates and micro templates, the deep application of thinking chain technology, and the innovative breakthrough of structured constraint decoding technology. It realizes structured reasoning and standardized output of multimodal data, and the generated structured conclusions (such as underwriting conclusions and claims conclusions) are accurate and standardized, which can directly support core business operations such as underwriting conclusion judgment, claims amount calculation, and business process advancement, providing core technical support for the automation and intelligent implementation of intelligent underwriting and intelligent claims business.

[0075] The knowledge generation method based on insurance business provided in this application involves acquiring multi-source, multi-modal data from insurance business scenarios, de-identifying the multi-source, multi-modal data, labeling the de-identified insurance data, and training a pre-set large language model based on the labeled insurance knowledge data. The trained multi-modal insurance data processing model then performs unified encoding processing on the insurance business data to obtain a fused feature representation. Based on the multi-modal insurance data processing model, knowledge is generated from the fused feature representation to obtain insurance business knowledge such as underwriting conclusions and claims conclusions. This application, based on a large model, can process massive amounts of multi-modal heterogeneous data, break down information silos, and uncover deep knowledge connections, thereby improving the efficiency and accuracy of insurance business knowledge generation.

[0076] This application's embodiments completely liberate insurance professionals from tedious and repetitive manual data screening, comparison, and integration, allowing them to focus on higher-value analytical decision-making. This significantly reduces labor costs and dramatically improves business processing speed (such as underwriting and claims processing time). This application's embodiments provide frontline personnel (such as underwriters, claims adjusters, and agents) with comprehensive, relevant, and directly applicable structured knowledge, assisting them in making more accurate and consistent risk assessments and business judgments, reducing human error and claims disputes.

[0077] This application's embodiments can revitalize data assets and uncover deep knowledge connections. It breaks down information silos, constructing a unified and complete enterprise-level insurance knowledge view, and achieving cross-departmental and cross-process data integration and business collaboration. The multimodal insurance data processing model in this application's embodiments possesses causal reasoning capabilities, enabling the discovery of complex risk patterns and business regularities that are difficult to detect using traditional methods (such as the correlation between specific populations and specific diseases, and the deep connection between exclusion clauses and specific claims situations), providing data-driven new insights for product innovation and precision marketing.

[0078] The embodiments of this application enable deep semantic understanding and association of multimodal data. The large-model-based approach possesses powerful generalization and continuous learning capabilities, allowing for rapid adaptation to business changes such as updated insurance terms and new regulatory rules. Through these embodiments, seamless integration with downstream business systems is easily achieved, thereby enabling insurance business knowledge to empower business operations.

[0079] Please see Figure 8 This application also provides a knowledge generation apparatus based on insurance business, which can implement the above-mentioned knowledge generation method based on insurance business. The apparatus includes: The insurance data acquisition module is used to acquire multi-source, multi-modal data from insurance business scenarios; among which, multi-source, multi-modal data includes at least two of the following: image data, text data, voice data, and video data; The data anonymization module is used to anonymize multi-source, multi-modal data, specifically for data anonymization insurance purposes; the anonymized insurance data is multi-modal data. The data annotation module is used to annotate de-identified insurance data to obtain insurance knowledge data; The model training module is used to train a pre-set large language model based on insurance knowledge data to obtain a multimodal insurance data processing model. The feature encoding module is used to uniformly encode pre-acquired insurance business data based on a multimodal insurance data processing model to obtain a fused feature representation; wherein, the insurance data includes at least one of the following: image data, text data, voice data, and video data; The insurance knowledge generation module is used to generate knowledge from the fused feature representation based on the multimodal insurance data processing model, thereby obtaining insurance business knowledge; wherein, the insurance business knowledge includes at least one of the following: underwriting conclusion, claims conclusion.

[0080] In some embodiments, the data desensitization module can specifically be used to implement: Multi-source, multi-modal data is preprocessed to obtain multi-source standardized data; Semantic recognition is performed on multi-source standardized data to obtain multi-source semantic data; Classify multi-source semantic data to obtain sensitive categories; Desensitized insurance data is obtained by desensitizing multi-source, multi-modal data based on sensitive categories.

[0081] Specifically, the data desensitization module can be used to implement steps 301 to 304 above, which will not be described in detail here.

[0082] After step 201, the knowledge generation device based on insurance business can also be used for: on-chain notarization processing of multi-source, multi-modal data, specifically including but not limited to: Information is extracted from multi-source, multi-modal data to obtain core insurance data; Hash values ​​are generated based on core insurance data; Acquire data collection information from multi-source, multi-modal data, including at least the collection time and data source identifier; Based on the hash value and data collection information, the core insurance data is processed for on-chain storage.

[0083] Specifically, the knowledge generation device based on insurance business can be used to implement steps 401 to 404 above, which will not be described in detail here.

[0084] In some embodiments, the data annotation module can specifically be used to implement: Based on modality categories, de-identified insurance data is randomly sampled and screened to obtain candidate insurance data; Entity annotation is performed on the candidate insurance data to obtain entity-annotated data; Relationship annotation is performed based on entity annotation data to obtain relationship annotation data; Sentiment labeling is performed on candidate insurance data to obtain sentiment-labeled data; Insurance knowledge data is obtained by constructing knowledge based on entity-labeled data, relationship-labeled data, and sentiment-labeled data.

[0085] Specifically, the data annotation module can be used to implement steps 501 to 505 above, which will not be described in detail here.

[0086] In some embodiments, the model training module can specifically be used to implement: Insurance knowledge data is preprocessed to obtain insurance preprocessed data; Key information is extracted from insurance preprocessing data to obtain key insurance information. The key insurance information in the insurance preprocessing data is masked to obtain insurance mask data; A multimodal insurance data processing model is obtained by training a pre-defined large language model based on insurance mask data.

[0087] Specifically, the model training module can be used to implement steps 601 to 604 above, which will not be described in detail here.

[0088] In some embodiments, the feature encoding module can specifically be used to implement: Insurance data is encoded separately according to its modality category to obtain single-modality encoded data; Quality verification is performed on the single-modal encoded data to obtain valid encoded data; The valid encoded data is normalized to obtain normalized encoded data; Cross-modal alignment is performed on the normalized encoded data to obtain aligned encoded data; Cross-modal attention fusion is performed on the aligned encoded data to obtain a fused feature representation.

[0089] Specifically, the feature encoding module can be used to implement steps 701 to 705 above, which will not be described in detail here.

[0090] The specific implementation of the knowledge generation device based on insurance business is basically the same as the specific implementation of the knowledge generation method based on insurance business described above, and will not be repeated here.

[0091] This application also provides a computer device, which can be any smart terminal including tablet computers, in-vehicle computers, etc. The computer device includes a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement: Acquire multi-source, multi-modal data from insurance business scenarios; where multi-source, multi-modal data includes at least two of the following: image data, text data, voice data, and video data; Multi-source, multi-modal data is anonymized to ensure data security; the anonymized data is multi-modal data. Insurance knowledge data is obtained by annotating de-identified insurance data. A multimodal insurance data processing model is obtained by training a pre-defined large language model based on insurance knowledge data. Based on a multimodal insurance data processing model, pre-acquired insurance business data is uniformly encoded to obtain a fused feature representation; wherein, the insurance data includes at least one of the following: image data, text data, voice data, and video data; Based on the multimodal insurance data processing model, knowledge is generated from the fused feature representation to obtain insurance business knowledge; among which, insurance business knowledge includes at least one of the following: underwriting conclusion and claims conclusion.

[0092] Please see Figure 9 , Figure 9 The hardware structure of a computer device according to another embodiment is illustrated. The computer device includes: The processor 901 can be implemented using a general-purpose CPU (Central Processing Unit), microprocessor, application-specific integrated circuit (ASIC), or one or more integrated circuits, and is used to execute relevant programs to implement the technical solutions provided in the embodiments of this application. The memory 902 can be implemented as a read-only memory (ROM), a static storage device, a dynamic storage device, or a random access memory (RAM). The memory 902 can store the operating system and other applications. When the technical solutions provided in the embodiments of this specification are implemented through software or firmware, the relevant program code is stored in the memory 902 and is called and executed by the processor 901 using the knowledge generation method based on insurance business in the embodiments of this application. The input / output interface 903 is used to implement information input and output; The communication interface 904 is used to enable communication and interaction between this device and other devices. Communication can be achieved through wired means (such as USB, Ethernet cable, etc.) or wireless means (such as mobile network, WIFI, Bluetooth, etc.). Bus 905 transmits information between various components of the device (e.g., processor 901, memory 902, input / output interface 903, and communication interface 904); The processor 901, memory 902, input / output interface 903, and communication interface 904 are connected to each other within the device via bus 905.

[0093] This application embodiment also provides a storage medium, which is a computer-readable storage medium, storing a computer program that is implemented when executed by a processor. Acquire multi-source, multi-modal data from insurance business scenarios; where multi-source, multi-modal data includes at least two of the following: image data, text data, voice data, and video data; Multi-source, multi-modal data is anonymized to ensure data security; the anonymized data is multi-modal data. Insurance knowledge data is obtained by annotating de-identified insurance data. A multimodal insurance data processing model is obtained by training a pre-defined large language model based on insurance knowledge data. Based on a multimodal insurance data processing model, pre-acquired insurance business data is uniformly encoded to obtain a fused feature representation; wherein, the insurance data includes at least one of the following: image data, text data, voice data, and video data; Based on the multimodal insurance data processing model, knowledge is generated from the fused feature representation to obtain insurance business knowledge; among which, insurance business knowledge includes at least one of the following: underwriting conclusion and claims conclusion.

[0094] Memory, as a non-transitory computer-readable storage medium, can be used to store non-transitory software programs and non-transitory computer-executable programs. Furthermore, memory may include high-speed random access memory, and may also include non-transitory memory, such as at least one disk storage device, flash memory device, or other non-transitory solid-state storage device. In some embodiments, memory may optionally include memory remotely located relative to the processor, and these remote memories can be connected to the processor via a network. Examples of such networks include, but are not limited to, the Internet, intranets, local area networks, mobile communication networks, and combinations thereof.

[0095] The knowledge generation method, apparatus, device, and storage medium based on insurance business provided in this application acquire multi-source, multi-modal data from insurance business scenarios, perform anonymization processing on the multi-source, multi-modal data, annotate the anonymized insurance data obtained from the anonymization processing, and train a preset large language model based on the insurance knowledge data obtained from the annotation. Then, based on the trained multi-modal insurance data processing model, the insurance business data is uniformly encoded to obtain a fused feature representation. Knowledge is generated based on the fused feature representation using the multi-modal insurance data processing model to obtain insurance business knowledge such as underwriting conclusions and claims conclusions. This application, based on a large model, can process massive amounts of multi-modal heterogeneous data, break down information silos, and mine deep knowledge connections, thereby improving the efficiency and accuracy of insurance business knowledge generation.

[0096] The embodiments described in this application are for the purpose of more clearly illustrating the technical solutions of the embodiments of this application, and do not constitute a limitation on the technical solutions provided by the embodiments of this application. As those skilled in the art will know, with the evolution of technology and the emergence of new application scenarios, the technical solutions provided by the embodiments of this application are also applicable to similar technical problems.

[0097] Those skilled in the art will understand that the technical solutions shown in the figures do not constitute a limitation on the embodiments of this application, and may include more or fewer steps than shown, or combine certain steps, or different steps.

[0098] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.

[0099] Those skilled in the art will understand that all or some of the steps in the methods disclosed above, as well as the functional modules / units in the systems and devices, can be implemented as software, firmware, hardware, or suitable combinations thereof.

[0100] The terms “first,” “second,” “third,” “fourth,” etc. (if present) in the specification and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this application described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms “comprising” and “having,” and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0101] It should be understood that in this application, "at least one (item)" means one or more, and "more than" means two or more. "And / or" is used to describe the relationship between related objects, indicating that three relationships can exist. For example, "A and / or B" can represent three cases: only A exists, only B exists, and both A and B exist simultaneously, where A and B can be singular or plural. The character " / " generally indicates that the preceding and following related objects are in an "or" relationship. "At least one (item) of the following" or similar expressions refer to any combination of these items, including any combination of single or plural items. For example, at least one (item) of a, b, or c can represent: a, b, c, "a and b", "a and c", "b and c", or "a and b and c", where a, b, and c can be single or multiple.

[0102] In the several embodiments provided in this application, it should be understood that the disclosed apparatus and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of the units described above is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or units may be electrical, mechanical, or other forms.

[0103] The units described above as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0104] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0105] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes multiple instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this application. The aforementioned storage medium includes various media capable of storing programs, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0106] It should be noted that any AI models, software tools, or components not belonging to this company appearing in the embodiments of this application are merely illustrative examples and do not represent actual use. All user personal information involved in the embodiments of this application has been authorized (with the knowledge and consent) by the relevant parties or has been fully authorized by all parties, and the executing entity may obtain it through various legal and compliant means. The collection, storage, use, processing, transmission, provision, and disclosure of the information, data, and signals involved all comply with relevant laws and regulations and do not violate public order and good morals.

[0107] The preferred embodiments of the present application have been described above with reference to the accompanying drawings, but this does not limit the scope of the claims of the present application. Any modifications, equivalent substitutions, and improvements made by those skilled in the art without departing from the scope and substance of the embodiments of the present application shall be within the scope of the claims of the present application.

Claims

1. A knowledge generation method based on insurance business, characterized in that, The method includes: Acquire multi-source, multi-modal data from insurance business scenarios; wherein the multi-source, multi-modal data includes at least two of the following: image data, text data, voice data, and video data; Multi-source, multi-modal data is anonymized to ensure data security; wherein, the anonymized data is multi-modal data. Based on the anonymized insurance data, insurance knowledge data is obtained by labeling it. The preset large language model is trained based on the insurance knowledge data to obtain a multimodal insurance data processing model; The multimodal insurance data processing model is used to uniformly encode the pre-acquired insurance business data to obtain a fused feature representation; wherein the insurance data includes at least one of the following: image data, text data, voice data, and video data. Based on the multimodal insurance data processing model, knowledge is generated from the fused feature representation to obtain insurance business knowledge; wherein, the insurance business knowledge includes at least one of the following: underwriting conclusion and claims conclusion.

2. The method according to claim 1, characterized in that, The process of generating knowledge from the fused feature representation based on the multimodal insurance data processing model to obtain insurance business knowledge includes: Extract the business task identifier based on the fusion feature representation; The prompt macro template is invoked based on the business task identifier; wherein, the prompt macro template includes user role, task information, and output requirements; Information is extracted based on the fused feature representation to obtain target business information; Risk identification is performed based on the target business information to obtain risk identification information; Construct a micro-template for the prompt words based on the macro template and the risk identification information; The insurance business knowledge is obtained by generating knowledge based on the prompt word micro-template.

3. The method according to claim 1, characterized in that, The insurance data has modal categories. The pre-acquired insurance business data is uniformly encoded based on the multimodal insurance data processing model to obtain a fused feature representation, including: The insurance data is individually encoded according to its modality category to obtain single-modality encoded data. The single-modal encoded data is subjected to quality verification to obtain valid encoded data; The effective encoded data is normalized to obtain normalized encoded data; The normalized encoded data is cross-modal aligned to obtain aligned encoded data; Cross-modal attention fusion is performed on the aligned encoded data to obtain the fused feature representation.

4. The method according to claim 1, characterized in that, The de-identified insurance data has modal categories. The insurance knowledge data obtained by labeling the de-identified insurance data includes: Based on the modality category, the de-identified insurance data is randomly sampled and screened to obtain candidate insurance data; The candidate insurance data is annotated with entities to obtain entity-annotated data; Relationship annotation is performed based on the entity annotation data to obtain relationship annotation data; Sentiment-labeled data is obtained by performing sentiment annotation on the candidate insurance data; The insurance knowledge data is obtained by constructing knowledge based on the entity annotation data, the relationship annotation data, and the sentiment annotation data.

5. The method according to claim 1, characterized in that, The process of training a preset large language model based on the insurance knowledge data to obtain a multimodal insurance data processing model includes: Insurance knowledge data is preprocessed to obtain insurance preprocessed data; Key information is extracted from insurance preprocessing data to obtain key insurance information. The key insurance information in the insurance preprocessing data is masked to obtain insurance mask data; A multimodal insurance data processing model is obtained by training a pre-defined large language model based on insurance mask data.

6. The method according to any one of claims 1 to 5, characterized in that, The aforementioned data anonymization processing of multi-source, multi-modal data, used to protect data anonymization, includes: The multi-source, multi-modal data is preprocessed to obtain multi-source standardized data; Semantic recognition is performed on the multi-source standardized data to obtain multi-source semantic data; The multi-source semantic data is classified to obtain sensitive categories; The multi-source, multi-modal data is de-identified based on the aforementioned sensitive categories to obtain the de-identified insurance data.

7. The method according to any one of claims 1 to 5, characterized in that, After acquiring multi-source, multi-modal data related to insurance business scenarios, the method further includes: Information is extracted from the multi-source, multi-modal data to obtain core insurance data; A hash value is generated based on the core insurance data; Acquire data acquisition information of the multi-source multimodal data, wherein the data acquisition information includes at least the acquisition time and the data source identifier; Based on the hash value and the data collection information, the core insurance data is processed for on-chain storage.

8. A knowledge generation device based on insurance business, characterized in that, The device includes: The insurance data acquisition module is used to acquire multi-source, multi-modal data from insurance business scenarios; wherein, the multi-source, multi-modal data includes at least two of the following: image data, text data, voice data, and video data; The data desensitization module is used to desensitize multi-source, multi-modal data, and to desensitize and protect the data; wherein, the desensitized and protected data is multi-modal data; The data annotation module is used to annotate the de-identified insurance data to obtain insurance knowledge data; The model training module is used to train a preset large language model based on the insurance knowledge data to obtain a multimodal insurance data processing model. The feature encoding module is used to perform unified encoding processing on the pre-acquired insurance business data based on the multimodal insurance data processing model to obtain a fused feature representation; wherein, the insurance data includes at least one of the following: image data, text data, voice data, and video data; The insurance knowledge generation module is used to generate knowledge from the fused feature representation based on the multimodal insurance data processing model to obtain insurance business knowledge; wherein, the insurance business knowledge includes at least one of the following: underwriting conclusion and claims conclusion.

9. A computer device, characterized in that, The computer device includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement the method according to 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 a processor, it implements the method of any one of claims 1 to 7.

Images