Insurance data analysis method, device, equipment and medium

Abstract

The application relates to the technical field of data analysis, and discloses an insurance data analysis method, device, equipment and medium, which comprises the following steps: obtaining insurance application verification information uploaded by an applicant, and analyzing the text into structured information in JSON format. Based on a cross-modal attention mechanism, image data and structured information are semantically aligned to identify logical consistency conflicts in time and space. Conflict fields, types, severity and application value fields are extracted, and a rule engine is used to map a confidence level. If the confidence level does not exceed a threshold value, it is determined that the verification is passed and a report is generated; otherwise, the information is sent to an artificial review client, and a report is generated according to feedback data. The scheme can be applied to a financial technology business system platform and can realize confidence analysis of insurance data. The application improves the efficiency and accuracy of insurance data analysis.

Description

Technical Field

[0001] This invention relates to the field of data analysis technology, and in particular to an insurance data analysis method, apparatus, equipment, and medium. Background Technology

[0002] Against the backdrop of rapid development in fintech, the insurance industry is gradually introducing technologies such as artificial intelligence, big data, and computer vision to improve the automation and intelligence of its business processes. However, the current agricultural insurance underwriting process still faces the following technical bottlenecks: First, underwriting information relies on manual review, making it difficult to cope with periodic concentrated policy issuance pressures. Furthermore, existing systems only support field-level comparisons, failing to achieve consistency verification of text and image semantics. Second, insurance attachments (such as ID photos, pigsty photos, disease prevention records, government subsidy documents, etc.) are mostly unstructured multimodal data. Traditional OCR tools have an accuracy rate of less than 60% for recognizing blurred images, handwritten text, and nested PDFs, leading to the omission of crucial information. Third, underwriting rules are hard-coded, requiring development intervention for updates, with response cycles as long as 1-2 weeks, making it impossible to adapt to policy differences and dynamic adjustment needs across regions and insurance types. Fourth, disaster warnings rely on single-source data such as meteorological bureau APIs, lacking a closed-loop linkage with the underwriting system. Warning information lacks targeted handling suggestions, and post-disaster loss assessment still relies primarily on manual on-site inspections, which is time-consuming and highly subjective. These problems constrain the efficiency and accuracy of agricultural insurance risk control. Summary of the Invention

[0003] This invention provides an insurance data analysis method, apparatus, computer equipment, and medium to solve the problems of low efficiency and low accuracy of existing insurance data analysis methods on the market.

[0004] Firstly, an insurance data analysis method is provided, including: Obtain insurance application verification information uploaded by the policyholder; The text content in the insurance application verification information is parsed into JSON format to obtain structured verification information; Based on the cross-modal attention mechanism, the image data in the insurance underwriting verification information is semantically aligned with the structured verification information to form aligned image data and aligned verification information; Identify temporal logical consistency conflicts between the aligned image data and the aligned verification information, and identify spatial logical consistency conflicts between the aligned image data and the aligned verification information; Extract the conflict fields, conflict types, and conflict severity from the time logic consistency conflict and the spatial logic consistency conflict, as well as the insurance value field from the structured verification information; Based on a preset rule engine, the confidence level of the insurance underwriting verification information is mapped according to the insured value field, the conflict field, the conflict type, and the conflict severity. Determine whether the confidence level is greater than a preset level threshold; If not, the insurance application verification information is deemed to have passed verification, and an insurance data analysis report is generated based on the confidence level and the structured verification information. If so, the insurance application verification information is sent to a preset manual review client, the feedback data returned by the manual review client is obtained, and an insurance data analysis report is generated based on the feedback data and the structured verification information.

[0005] Secondly, an insurance data analysis device is provided, comprising: The data acquisition module is used to acquire the insurance application verification information uploaded by the policyholder, and parse the text content in the insurance application verification information into JSON format to obtain structured verification information; The semantic alignment module is used to semantically align the image data in the insurance underwriting verification information with the structured verification information based on a cross-modal attention mechanism, forming aligned image data and aligned verification information. The conflict verification module is used to identify temporal logical consistency conflicts between the aligned image data and the aligned verification information, and to identify spatial logical consistency conflicts between the aligned image data and the aligned verification information. The rule verification module is used to extract the conflict fields, conflict types, and conflict severity in the time logic consistency conflict and the spatial logic consistency conflict, as well as the insurance value field in the structured verification information. Based on a preset rule engine, the module maps the confidence level of the insurance underwriting verification information according to the insurance value field, the conflict fields, the conflict types, and the conflict severity. The confidence level judgment module is used to determine whether the confidence level is greater than a preset level threshold. If not, it determines that the insurance application verification information has passed the verification. Based on the confidence level and the structured verification information, an insurance data analysis report is generated. If yes, the insurance application verification information is sent to a preset manual review client, and feedback data returned by the manual review client is obtained. Based on the feedback data and the structured verification information, an insurance data analysis report is generated.

[0006] Thirdly, a computer device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the above-described insurance data analysis method.

[0007] Fourthly, a computer-readable storage medium is provided, which stores a computer program that, when executed by a processor, implements the steps of the aforementioned insurance data analysis method.

[0008] The aforementioned insurance data analysis method, device, computer equipment, and storage medium can acquire various insurance verification materials submitted by policyholders, extract text information from them, and convert it into structured JSON data. This transforms unstructured raw materials into a standardized format that computers can efficiently process, laying the foundation for subsequent automated analysis. Subsequently, based on a cross-modal attention mechanism, fine-grained semantic alignment of image content and structured text is performed to obtain alignment results with mapping relationships, achieving deep semantic association between images and text, significantly improving the accuracy and recall of key information extraction. Furthermore, logical consistency conflicts between images and text are identified from both temporal and spatial dimensions, such as date discrepancies or geographical location deviations, and conflict fields, conflict types, and severity are extracted. The system extracts key numerical fields such as insured area and sum insured from structured data. It then utilizes a pre-defined dynamic rule engine to map conflicting information and numerical fields to a confidence level for the current verification information. This allows for dynamic rule configuration and phased rollout, enabling second-level adjustments to the verification strategy. Simultaneously, it quantifies multi-dimensional risk signals into intuitive confidence scores for easier subsequent business decisions. If the confidence level does not exceed a preset threshold, the verification is considered successful, and an analysis report is generated based on the confidence level and structured verification information, enabling rapid and automatic approval of low-risk policies to improve efficiency. If the confidence level exceeds the threshold, the verification information is pushed to a manual review client. After receiving feedback from the reviewer, a final analysis report is generated based on the structured verification information, improving the efficiency and accuracy of insurance data analysis. Attached Figure Description

[0009] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments of the present invention will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0010] Figure 1 This is a schematic diagram of an application environment for an insurance data analysis method according to an embodiment of the present invention; Figure 2 This is a flowchart illustrating an insurance data analysis method according to one embodiment of the present invention; Figure 3 This is a schematic diagram of an insurance data analysis device according to an embodiment of the present invention; Figure 4This is a schematic diagram of the structure of a computer device according to an embodiment of the present invention; Figure 5 This is another structural schematic diagram of a computer device according to one embodiment of the present invention. Detailed Implementation

[0011] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0012] The insurance data analysis method provided in this invention can be applied to, for example... Figure 1 In this application environment, the client communicates with the server via a network. The server can obtain the insurance verification materials submitted by the policyholder through the client, extract the text information, and convert it into structured JSON data. Subsequently, based on a cross-modal attention mechanism, the image content and structured text are semantically aligned with fine granularity to obtain an alignment result with mapping relationships. On this basis, logical consistency conflicts between images and text are identified from the time and spatial dimensions, such as date contradictions or geographical location discrepancies, and conflict fields, conflict types, and severity are extracted. At the same time, key numerical fields such as insured area and insured amount are extracted from the structured data. The system further utilizes a preset dynamic rule engine to jointly map the above conflict information and numerical fields to the confidence level of the verification information. If the confidence level does not exceed a preset threshold, the verification is deemed successful, and an analysis report is generated based on the confidence level and structured verification information. If the confidence level exceeds the threshold, the verification information is pushed to the manual review client. After the manual feedback data is returned, the final analysis report is generated in combination with the structured verification information, improving the efficiency and accuracy of insurance data analysis. The client 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 invention will now be described in detail through specific embodiments.

[0013] Please see Figure 2 As shown, Figure 2 A flowchart illustrating the insurance data analysis method provided in this embodiment of the invention includes the following steps: S1. Obtain the insurance application verification information uploaded by the policyholder.

[0014] In this embodiment of the invention, obtaining the insurance application verification information uploaded by the policyholder refers to receiving various electronic supporting documents submitted by the policyholder during the agricultural insurance application process. These documents include image files such as ID card photos, ownership certificates, public notice photos, pigsty photos, disease prevention records, and government subsidy documents, as well as any accompanying text information (such as the content of the application form). This information is typically uploaded via mobile apps, mini-programs, or web forms. The system calls a file receiving interface or message queue to obtain the raw data and temporarily stores it in a temporary storage area, providing basic input for subsequent multimodal parsing, semantic alignment, and risk identification.

[0015] In the fintech field, the insurance application verification information refers to a multimodal electronic data set submitted by the policyholder during the online application process to verify the authenticity, compliance, and risk level of the insurance application. Typical examples include images / PDF / Word files such as ID card photos, ownership certificates, public notice photos, actual photos of farms, and government subsidy documents, as well as text fields filled in the application form (such as the policyholder's name, ID number, type of insured property, area, sum insured, and premium rate). This type of data combines unstructured (images and text, handwritten, stamped) and semi-structured (form fields) characteristics.

[0016] S2. Parse the text content in the insurance application verification information into JSON format to obtain structured verification information.

[0017] In this embodiment of the invention, parsing the text content in the insurance application verification information into JSON format to obtain structured verification information means extracting key data such as the policyholder's name, ID number, insured area, insured amount, premium rate, address, and time from the application verification information through OCR text recognition or by directly reading fields from a spreadsheet, and organizing them into a unified JSON key-value pair structure according to a predefined field mapping relationship.

[0018] In this embodiment of the invention, by parsing the text content in the insurance application verification information into JSON format, structured verification information is obtained, which improves the efficiency of subsequent data processing.

[0019] S3. Based on the cross-modal attention mechanism, the image data in the insurance underwriting verification information is semantically aligned with the structured verification information to form aligned image data and aligned verification information.

[0020] In this embodiment of the invention, the step of semantically aligning the image data in the insurance underwriting verification information with the structured verification information based on a cross-modal attention mechanism to form aligned image data and aligned verification information includes: Extract the feature vectors from the image data in the insurance application verification information to obtain an image feature vector set; The text feature vector set is obtained by combining the feature vector of each verification information in the structured verification information. Calculate the cosine similarity between each image feature vector in the image feature vector set and each text feature vector in the text feature vector set; Construct attention weight matrices for the image feature vector set and the text feature vector set based on the cosine similarity; Find the image feature vector in the image feature vector set that has the highest similarity to each text feature vector in the text feature vector set, and generate a list of matching vector pairs; Based on the matching vector pair list, a mapping relationship between the image data and the structured verification information is established to obtain the mapped image data and the mapped verification information. The attention weight matrix is ​​used to filter data pairs with confidence levels less than a preset threshold in the mapped image data and the mapped verification information to obtain filtered image data and filtered verification information. The filtered image data and the filtered verification information are packaged into a preset data format to form aligned image data and aligned verification information.

[0021] In detail, the step of extracting the feature vectors of the image data in the insurance application verification information to obtain the image feature vector set involves inputting each image uploaded by the policyholder (such as ID card photos, public notice photos, pigsty photos, etc.) into a pre-trained cross-modal visual encoder (Vision Transformer). This encoder maps the original pixels into dense vectors of fixed dimensions, with each image corresponding to an image feature vector. The feature vectors of all images are then aggregated to form the image feature vector set.

[0022] In detail, the feature vector of each verification information in the structured verification information is used to obtain a text feature vector set. This is achieved by concatenating the field name and field value of each field in the structured verification information that has been parsed into JSON format into short text (such as "Insurance Time: 2025-03-10" "Insured Area: 120.5 mu"), and inputting them one by one into a text encoder (such as CLIP's Text Transformer) jointly trained with the visual encoder, and outputting the corresponding text feature vector. The text feature vectors of all fields together constitute the text feature vector set, thereby ensuring that the image and text are in the same semantic space.

[0023] In detail, the calculation of the cosine similarity between each image feature vector in the image feature vector set and each text feature vector in the text feature vector set involves calculating the dot product of each pair of image vectors and text vectors and dividing it by the product of their moduli. The resulting cosine value ranges from -1 to 1, with a larger value indicating a greater semantic relevance between the image and the text. This calculation can be performed in batches using matrix multiplication to obtain a similarity matrix with the shape of (number of images × number of text fields).

[0024] In detail, the construction of the attention weight matrix of the image feature vector set and the text feature vector set based on the cosine similarity directly uses the cosine similarity matrix as the original score of the attention weight matrix without any additional transformation; the value in the i-th row and j-th column of this matrix is ​​the attention intensity of the i-th image to the j-th text field, and subsequent softmax normalization can be performed on each row to obtain the probability distribution, but the original similarity itself can already be used as a weight.

[0025] In detail, the step of finding the image feature vector with the highest similarity to each text feature vector in the image feature vector set and generating a list of matching vector pairs involves finding the row index of the maximum value in the corresponding column of the similarity matrix for each text feature vector, thereby determining the image most relevant to the current text semantics; similarly, the most relevant text field can be found for each image, and finally a series of matching pairs of "(image ID, text field name, similarity)" are generated and stored in the list of matching vector pairs.

[0026] In detail, the step of establishing a mapping relationship between the image data and the structured verification information based on the matching vector pair list to obtain the mapped image data and the mapped verification information involves adding an "associated field" tag to the metadata of each image (e.g., associating "public notice photo" with the "insurance time" field) according to the matching vector pair list, and adding an "associated image ID" attribute to each field in the structured verification information. This mapping relationship is recorded in the form of bidirectional pointers, forming the mapped image data and the mapped verification information, which facilitates quick location of the corresponding items during subsequent consistency verification.

[0027] In detail, the step of using the attention weight matrix to filter data pairs with confidence scores less than a preset threshold in the mapped image data and the mapped verification information to obtain filtered image data and filtered verification information involves taking the cosine similarity corresponding to each matching pair from the attention weight matrix as a confidence score and setting a threshold (e.g., 0.7). If the similarity of a matching pair is lower than the threshold, the mapping relationship is invalidated, the corresponding "association field" label is removed from the mapped image data, and "weak alignment" is marked in the corresponding field of the mapped verification information. If the similarity is higher than the threshold, it is retained, thereby filtering out unreliable semantic associations.

[0028] In this embodiment of the invention, the image data in the insurance underwriting verification information is semantically aligned with the structured verification information based on a cross-modal attention mechanism, thereby improving the accuracy of subsequent temporal and spatial logical consistency checks.

[0029] S4. Identify temporal logical consistency conflicts between the aligned image data and the aligned verification information, and identify spatial logical consistency conflicts between the aligned image data and the aligned verification information.

[0030] In this embodiment of the invention, identifying temporal logical consistency conflicts between the aligned image data and the aligned verification information includes: Extract the time point data from the aligned image data to obtain an image time point list; Extract the time point data from the aligned verification information to obtain a text time point list; Obtain the mapping relationship between the aligned image data and the aligned verification information to obtain a mapping relationship list; Based on the mapping relationship list, the image time point list is matched with the time points to obtain a time matching pair list; A time conflict determination table is obtained by performing time logic verification on each time matching pair in the time matching pair list using preset time logic verification rules. A structured conflict entry is generated for each conflict determination result in the time conflict determination table to obtain the time logic consistency conflict.

[0031] In detail, the step of extracting time point data from the aligned image data to obtain an image time point list involves calling the multimodal time entity recognition module one by one from a set of image files that have already undergone semantic alignment. This module first uses an OCR engine to recognize text regions in the image, and then uses a CRF-based or deep learning-based time expression parser (such as SUTime or a self-trained time extraction model) to extract strings that conform to date, time, and relative time (such as "yesterday" or "March 2025") expressions, and standardizes them into the "YYYY-MM-DD" format, while recording the position coordinates of each time entity in the image.

[0032] In detail, the step of extracting the time point data from the aligned verification information to obtain a text time point list involves iterating through all fields in the structured JSON verification information with associated image information, filtering out fields whose field names contain keywords such as "time," "date," "period," "year," "month," and "day" (e.g., insurance application time, public notice date, document validity period, shooting time, etc.), and reading their field values. If the value is a standardized time string, it is used directly; if it is a relative description, it is converted into a specific date based on the context of the insurance application, and is also uniformly converted into the "YYYY-MM-DD" format.

[0033] In detail, obtaining the mapping relationship between the aligned image data and the aligned verification information to obtain the mapping relationship list is achieved by directly reading the "association field" tag that has been generated and attached to the metadata of the aligned image data during the semantic alignment stage, as well as the "association image ID" tag attached to each field of the aligned verification information.

[0034] In detail, the step of matching the image time point list with the text time points based on the mapping relationship list to obtain a time matching pair list involves traversing each mapping in the mapping relationship list, finding all time entities of the image from the image time point list based on the image ID, finding the corresponding text time value from the text time point list based on the associated text field name, and then forming a matching pair between each image time entity and the corresponding text time value, which is stored in the time matching pair list; if an image has multiple time entities, multiple matching pairs are generated.

[0035] In detail, the step of performing time logic verification on each time matching pair in the time matching pair list using preset time logic verification rules to obtain a time conflict determination table involves substituting the image time value and text time value in each matching pair into the agricultural insurance business rule base (e.g., the posting date in the publicized photo should be ≤ the insurance application date; the validity period of the ID card should include the insurance application date; the photo shooting time should be earlier than the underwriting time, etc.). The rule is determined by comparing the size or chronological relationship of the two time values. If the rule is met, it is marked as "consistent". If the rule is not met, the conflict type (e.g., "time inversion", "overdue", "not effective") and severity (high / medium / low) are recorded, and each determination result (matching pair identifier, whether there is a conflict, conflict type, severity, specific value) is stored in the time conflict determination table.

[0036] In detail, generating a structured conflict entry for each conflict determination result in the time conflict determination table to obtain the time logical consistency conflict involves filtering all records determined as "conflicts" from the time conflict determination table and generating a JSON object for each conflict containing the following fields: conflict ID (auto-incrementing), involved image name, involved text field name, conflict type (e.g., "publication date is later than insurance date"), conflict severity, image time value, text time value, and suggested handling (e.g., "manually verify the publicized photo").

[0037] In this embodiment of the invention, identifying spatial logical consistency conflicts between the aligned image data and the aligned verification information includes: Extract the spatial information contained in the aligned image data to obtain an image space list; Find the fields related to geographical location in the aligned verification information to obtain a list of text spaces; The image space list and the text space list are converted into coordinate format data to obtain an image coordinate list and a text coordinate list; The image coordinate list and the text coordinate list are paired according to the mapping relationship list to obtain a spatial matching pair list; Calculate the spatial distance between each pair of coordinate data in the spatial matching pair list to obtain a spatial distance list; Spatial distances in the spatial distance list that are greater than a preset distance threshold are identified as having spatial conflicts, and a determination result is obtained. The spatial logical consistency conflict is generated based on the determination result.

[0038] In detail, the process of finding the geographically related fields in the aligned verification information to obtain a text space list involves traversing all fields in the structured JSON verification information, filtering out fields whose names contain keywords such as "address," "location," "location," "GPS," "coordinates," "township," "village," "land parcel," and "farm" (e.g., "farm location," "insured land parcel address," "ownership certificate location"), and reading their field values. If the value is in coordinate format (e.g., "31.23, 121.47"), it is extracted directly; if it is a text address, the original string is recorded; and if it is a land parcel number, the number is temporarily stored for later conversion through the GIS system.

[0039] In detail, the process of converting the image space list and the text space list into coordinate format data to obtain the image coordinate list and the text coordinate list involves the following steps for each item in the image space list: if it is a GPS coordinate, it is directly converted into a decimal coordinate pair (latitude, longitude); if it is an OCR address text, a geocoding API (such as the Gaode Map API or local GIS service) is called to parse the address into coordinates. If the parsing is successful, the coordinates and matching confidence are recorded; if it fails, it is marked as "unable to locate". For each item in the text space list: if coordinates already exist, they are used directly; if it is address text, the geocoding API is also called for conversion; if it is a plot number, the coordinates of the plot boundary center point are queried from the GIS database. Each successfully converted item generates a record containing coordinates (latitude, longitude) and a source identifier, which are stored in the image coordinate list and the text coordinate list respectively.

[0040] In detail, the step of pairing the image coordinate list with the text coordinate list according to the mapping relationship list to obtain a spatial matching pair list involves reading the mapping relationship list generated during the semantic alignment stage (each mapping record contains an image ID and an associated text field name), and performing a Cartesian product pairing on all coordinate items corresponding to the image ID in the image coordinate list and all coordinate items corresponding to the associated text field name in the text coordinate list (an image may have multiple coordinates extracted, and a text field may also correspond to multiple coordinates) to generate the spatial matching pair list.

[0041] In detail, the calculation of the spatial distance between each pair of coordinate data in the spatial matching pair list to obtain the spatial distance list involves using the Haversine formula to calculate the great circle distance (in meters) between the two coordinate points (image coordinates and text coordinates) in each spatial matching pair. This formula calculates a relatively accurate distance between two points using the difference in latitude and longitude and the Earth's radius (taken as 6371 km). If a coordinate comes from geocoding and has low confidence, a warning mark can be added to the distance result. All calculated distance values ​​are stored according to the matching pair identifier to form the spatial distance list.

[0042] In detail, the determination of spatial distances in the spatial distance list that are greater than a preset distance threshold as having spatial conflicts and obtaining the determination result is based on setting reasonable distance thresholds according to the agricultural insurance business scenario (for example, the distance threshold between a pig farm photo and a text address is set to 500 meters, and the distance threshold between the village name identified from the public photo and the insured plot is set to 2 kilometers). Each distance value in the spatial distance list is compared with the corresponding business threshold one by one; if the distance is ≤ the threshold, it is determined as "spatial consistency", and if the distance is > the threshold, it is determined as "spatial conflict", and further classified into severity levels according to the degree of exceedance (mild: exceedance but within the same township; moderate: across townships but within the same county; severe: across counties).

[0043] In detail, generating the spatial logical consistency conflict based on the judgment result involves filtering all records of "spatial conflict" from the judgment result set and generating a structured JSON conflict entry for each conflict, including: conflict ID, involved image name, involved text field name, conflict type (e.g., "coordinate distance exceeds threshold"), severity (high / medium / low), specific distance value, preset threshold, and suggested handling (e.g., "manually verify pigsty location" or "re-upload photos with GPS").

[0044] In this embodiment of the invention, the step of extracting the spatial information contained in the aligned image data to obtain an image space list includes: Determine whether the aligned image data contains GPS metadata; If the aligned image data contains GPS metadata, then the GPS metadata is converted into spatial information to obtain an image spatial list; If the aligned image data does not contain GPS metadata, then determine whether the aligned image data contains camera watermark text. If the image contains camera watermark text, the camera watermark text is converted into spatial information to obtain an image spatial list. If the image does not contain camera watermark text, then scene semantic recognition is performed on the aligned image data to obtain the scene semantic recognition result; The scene semantic recognition results are converted into spatial information to obtain an image spatial list.

[0045] In detail, the determination of whether the aligned image data contains GPS metadata is achieved by using an image processing library (such as PIL, ExifRead) to read the EXIF ​​information block of each image file and check whether there is a GPSInfo tag (usually containing sub-tags such as GPSLatitude, GPSLongitude, GPSLatitudeRef, GPSLongitudeRef); if it exists and the data is complete, it is determined that GPS metadata is contained; otherwise, the next level of judgment is performed.

[0046] In detail, if the aligned image data contains GPS metadata, the GPS metadata is converted into spatial information to obtain an image spatial list. This involves extracting the latitude, longitude (degrees, minutes, seconds) values ​​and the reference markers for north / south latitude and east / west longitude from the EXIF's GPSInfo. The degrees, minutes, and seconds are then converted into decimal degrees (formula: degrees + minutes / 60 + seconds / 3600). The sign is determined based on the north / south latitude (north is positive, south is negative) and east / west longitude (east is positive, west is negative). Finally, a standard coordinate pair (latitude, longitude) is generated, along with the source marker "GPS" and a confidence level of 1.0, and stored in the image spatial list.

[0047] In detail, if the aligned image data does not contain GPS metadata, then determining whether the aligned image data contains camera watermark text involves calling an OCR engine (such as PaddleOCR) to perform text recognition on the image and performing regular expression matching on common mobile phone camera watermark keyword patterns (such as "Shot on iPhone", "MI 9", "taken at", "location:", "GPS:", etc.). If the recognized text contains content that matches the above pattern, then it is considered that camera watermark text exists; otherwise, proceed to the next level of judgment.

[0048] In detail, if the text contains camera watermark text, the camera watermark text is converted into spatial information to obtain an image spatial list. This involves extracting possible geographic coordinates (such as "31.23°N 121.47°E" or "31°14'23"N, 121°28'35"E") or address descriptions (such as "Shanghai Pudong New Area") from the watermark text using regular expressions. The coordinate formats are directly converted to obtain coordinate pairs, and the address descriptions are converted into coordinates using a geocoding API. At the same time, the resolution confidence level is recorded (based on text clarity and matching degree). Finally, the converted coordinates, along with the source marker "watermark", are stored in the image spatial list.

[0049] In detail, if the image does not contain camera watermark text, then scene semantic recognition is performed on the aligned image data to obtain scene semantic recognition results. This involves inputting the image into a pre-trained scene classification model (such as a model trained on the Places365 dataset based on ResNet-50) and outputting the three most likely scene labels of the image and their corresponding probabilities, such as "pigsty", "farmland", "notice board", "residential area", etc.

[0050] In detail, the conversion of the scene semantic recognition result into spatial information to obtain an image spatial list is based on weak alignment of the scene label with the text address in the underwriting information: for example, if it is identified as "pigsty" and the insurance information contains the text address of "location of the farm", then the address is converted into coordinates through geocoding, and a "scene semantic association" tag and a low confidence level (such as 0.5) are attached and stored in the image spatial list; if no text address can be associated, then only the scene label is recorded but no coordinates are recorded, and it is not used as the basis for spatial conflict determination.

[0051] In this embodiment of the invention, by identifying temporal logical consistency conflicts between the aligned image data and the aligned verification information, and by identifying spatial logical consistency conflicts between the aligned image data and the aligned verification information, data input is provided for subsequent rule verification.

[0052] S5. Extract the conflict fields, conflict types, and conflict severity from the time logic consistency conflict and the spatial logic consistency conflict, as well as the insurance value field from the structured verification information.

[0053] In this embodiment of the invention, the extraction of conflict fields, conflict types, and conflict severity from the temporal and spatial logical consistency conflicts, as well as the insurance value fields from the structured verification information, involves extracting the specific data item names (e.g., "publication date," "farm GPS coordinates"), the type of logical error (e.g., "time inversion," "coordinate distance exceeding threshold"), and the high, medium, or low risk level of each conflict from the temporal and spatial logical conflict lists. Simultaneously, numerical data directly related to risk quantification, such as insured area, total insured amount, premium rate, subsidy ratio, and number of livestock, are specifically selected from the already parsed standardized insurance information. This extracted information collectively constitutes the input material for the rule engine: conflict fields, their types, and severity are used for deductions or anomaly detection, while the insurance value fields provide the baseline conditions for risk assessment (e.g., "area exceeding 50 mu and time conflict exists"), thereby achieving a comprehensive evaluation of the confidence level of each insurance verification information.

[0054] S6. Based on the preset rule engine, the confidence level of the insurance application verification information is mapped according to the insurance application value field, the conflict field, the conflict type and the conflict severity.

[0055] In this embodiment of the invention, the preset rule engine maps the confidence level of the insurance underwriting verification information based on the insured value field, the conflict field, the conflict type, and the conflict severity, including: The insured value field, the conflict field, the conflict type, and the conflict severity are converted into key-value pairs to obtain the object to be verified. By using the object to be verified to traverse each preset rule in the rule engine, the matching result between the object to be verified and each preset rule is obtained; Calculate the confidence score of the object to be verified against each preset rule based on the matching results; Calculate the average of all confidence scores and map the average to a confidence level.

[0056] In detail, the step of converting the insured value field, the conflict field, the conflict type, and the conflict severity into a key-value pair format to obtain the object to be verified involves creating a dictionary structure (such as a Python dict or a JSON object) and uniformly representing all types of input data as key-value pairs of "field name: value".

[0057] In detail, the step of using the object to be verified to traverse each preset rule in the rule engine and obtain the matching result between the object to be verified and each preset rule involves retrieving the predefined rule list in the rule engine (each rule contains a conditional expression and a corresponding action) one by one, and using the key-value pairs in the object to be verified as facts to match the conditional part of the rule. The matching process can use a rule engine (such as Drools, EasyRules, or a self-developed conditional matcher) to check whether the object to be verified meets the condition of the rule (for example, if the condition is if has_time_conflict == true and time_conflict_severity == "high", then if the key exists in the object to be verified and the value is equal, then the match is successful). The matching result of each rule is recorded as "hit" or "miss", and the ID of the hit rule and its weight or deduction / addition setting are retained.

[0058] In detail, the calculation of the confidence score between the object to be verified and each preset rule based on the matching result involves calculating a score contribution value for each matched rule according to the preset score adjustment logic in the rule. For example, the rule can set basic deduction items: "30 points for high-severity time conflict", "15 points for medium-severity spatial conflict", "5 points for insured area exceeding 500 acres"; it can also have bonus rules: "5 points for no conflict and reasonable value". For each matched rule, the confidence score of the rule is obtained according to its preset score change (positive or negative) (it can be a single score or an adjustment value relative to the initial base). If a rule is not matched, its score contribution is 0. Finally, the score (or score increment) corresponding to each rule is obtained.

[0059] In detail, calculating the average of all confidence scores and mapping this average to a confidence level involves summing the scores calculated for all rules (including the increments for hits and 0 points for misses) and dividing by the total number of rules to obtain an average score between 0 and 100 (assuming an initial base of 100, the average score may be lower than 100 after deductions). A weighted average can also be used, assigning different weights based on the priority of the rules. After obtaining the average score, it is compared with a preset level threshold table: for example, an average score ≥ 80 maps to "high confidence," 60 ≤ average score < 80 maps to "medium confidence," and an average score < 60 maps to "low confidence."

[0060] S7. Determine whether the confidence level is greater than a preset level threshold.

[0061] If not, proceed to step S8, determine that the insurance application verification information has passed verification, and generate an insurance data analysis report based on the confidence level and the structured verification information.

[0062] In this embodiment of the invention, the step of generating an insurance data analysis report based on the confidence level and the structured verification information involves using the final determined confidence level (high / medium / low) and structured insurance information (such as policyholder, insured object, area, sum insured, etc.) as core content to automatically generate an analysis document for underwriters or managers. The report will detail the confidence level and its corresponding score, the main reasons for deductions, the triggered conflict items (inconsistencies in time or space), key insurance numerical fields, and include the verification conclusion (e.g., "verification passed").

[0063] If so, then execute S9, send the insurance application verification information to the preset manual review client, obtain the feedback data returned by the manual review client, and generate an insurance data analysis report based on the feedback data and the structured verification information.

[0064] In this embodiment of the invention, generating an insurance data analysis report based on the feedback data and the structured verification information includes: Extract the anomalies contained in the feedback data to obtain an anomaly list; Extract the review conclusions contained in the feedback data; The preset indicator data in the structured verification information are used to obtain the indicator dataset; The list of anomalies, the review conclusions, the indicator dataset, and the confidence level are mapped to a preset report template to generate the insurance data analysis report.

[0065] In detail, the step of extracting anomalies from the feedback data to obtain an anomaly list involves parsing all information items marked as anomalies from the feedback data returned by the manual review client. The feedback data is typically transmitted in JSON format, containing the anomaly type selected or filled in by the reviewer (e.g., "image and text time inconsistency," "excessive spatial coordinate deviation," "blurred and unrecognizable document," etc.), as well as free-text comments for each anomaly. The system iterates through the "anomaly list" field in the feedback data, extracts a summary of each anomaly (anomaly type, involved fields, severity, and comments), converts it into a uniformly structured object, and stores it in the anomaly list.

[0066] In detail, extracting the review conclusion contained in the feedback data involves reading the review conclusion field from the feedback data. This field typically contains the final judgment result of the manual review, such as "verification passed," "verification rejected," "supplementary materials required," or "passed after manual modification." Additional information such as the reviewer's ID, review time, and review comments (e.g., reason for rejection or explanation of modifications) is also extracted.

[0067] In detail, the preset indicator data in the structured verification information, resulting in an indicator dataset, is a set of key statistical indicators that need to be displayed in the report, predefined according to business needs. These indicators include, for example, total insured amount, total insured area / total number of heads, subsidy amount, premium rate, total number of anomalies, and number of high-severity anomalies. The system extracts the corresponding values ​​or descriptions from the standardized and potentially manually corrected structured verification information, one by one according to the indicator name, generating a set of key-value pairs (e.g., {"Total Insured Amount":50000, "Insured Area":120.5, "Total Number of Anomalies":2, "High-Severity Anomalies":1}), which serves as the indicator dataset.

[0068] In detail, the process of mapping the anomaly list, the review conclusion, and the confidence level to a preset report template to generate the insurance data analysis report involves using a template engine (such as Jinja2 or a placeholder-based string replacement) to load a predefined report template (HTML or Markdown format). The template contains dynamic placeholders such as {{anomaly list}}, {{review conclusion}}, {{confidence level}}, and {{indicator dataset}}. The system fills in the data extracted in the previous three steps according to the template's format requirements (e.g., the anomaly list is rendered as a table, the review conclusion as a text paragraph, and the confidence level as a colored label). It can also append automatically generated timestamps, report numbers, version information, etc. After completion, a readable report text is generated.

[0069] In the field of fintech, this solution can be applied to intelligent underwriting and risk management in agricultural insurance. Through fintech means such as multimodal large models, cross-modal semantic alignment, dynamic rule engines, and remote sensing image analysis, it can automatically parse and logically verify the unstructured verification information uploaded by policyholders, such as ID cards, ownership certificates, and on-site photos, significantly reducing the burden of manual underwriting and fraud risks. At the same time, it combines multi-source data such as meteorology and satellite remote sensing to conduct disaster early warning and automatic disaster loss calculation, forming a closed loop from pre-identification of underwriting risks and business linkage control to post-disaster feedback. This provides insurance companies with an efficient, explainable, and evolvable intelligent risk control solution, promoting the transformation of agricultural insurance from "passive claims settlement" to "proactive prevention."

[0070] As can be seen, the above-mentioned scheme proposes an insurance data analysis method aimed at solving problems such as low underwriting efficiency, weak multimodal document review capabilities, lagging rule updates, isolated disaster early warning information, and strong subjectivity in disaster loss assessment in traditional agricultural insurance. The scheme first uses a multimodal large model (such as Qwen-VL) to perform end-to-end parsing of images and text files uploaded by policyholders, including ID cards, public notice photos, ownership certificates, and pigsty photos. It then utilizes a cross-modal attention mechanism to achieve fine-grained semantic alignment between image content and structured verification information, generating an image-text semantic alignment map. Next, it performs logical consistency checks on temporal entities (such as public notice date and insurance application time) and spatial entities (such as GPS coordinates and addresses) in the images and text, identifying conflicts such as time inversion and coordinate distance discrepancies. Finally, it combines a dynamic rule engine (supporting plug-and-play JSON rules and canary release) to map conflicting fields, conflict types, severity, and insured numerical fields (area, insured amount, etc.) to confidence levels. For low-confidence insurance policies, the system pushes risk interception instructions to the core policy issuance system via RESTful API and message queues. Manual review results are fed back to the AI ​​model, forming a closed-loop evolution. Simultaneously, the solution integrates real-time meteorological data, satellite remote sensing imagery, drone aerial photography, and GIS layers. It uses U-Net+++ Transformer to identify disaster types such as floods and droughts and simulate their spread paths, generating early warning information with disposal suggestions. After a disaster, a dual-branch contrastive learning network is used to calculate the disaster loss index, automatically calculating the amount of loss and displaying it on a map visualization platform. Finally, disposal tasks are pushed through an app, farmers upload before-and-after comparison images, AI analyzes the effectiveness of the disposal, and all data is fed back to optimize the model, achieving a closed-loop process of "early warning → disposal → feedback," significantly improving agricultural insurance underwriting efficiency and risk control capabilities.

[0071] It should be understood that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0072] In one embodiment, an insurance data analysis device is provided, which corresponds one-to-one with the insurance data analysis method described in the above embodiments. For example... Figure 3 As shown, the insurance data analysis device includes a data acquisition module 101, a semantic alignment module 102, a conflict verification module 103, a rule verification module 104, and a confidence level judgment module 105. Detailed descriptions of each functional module are as follows: Data acquisition module 101 is used to acquire insurance application verification information uploaded by the policyholder, and parse the text content in the insurance application verification information into JSON format to obtain structured verification information; Semantic alignment module 102 is used to semantically align the image data in the insurance underwriting verification information with the structured verification information based on a cross-modal attention mechanism, so as to form aligned image data and aligned verification information; The conflict verification module 103 is used to identify temporal logical consistency conflicts between the aligned image data and the aligned verification information, and to identify spatial logical consistency conflicts between the aligned image data and the aligned verification information. The rule verification module 104 is used to extract the conflict fields, conflict types, and conflict severity in the time logic consistency conflict and the spatial logic consistency conflict, as well as the insurance value field in the structured verification information. Based on a preset rule engine, it maps the confidence level of the insurance application verification information according to the insurance value field, the conflict fields, the conflict types, and the conflict severity. The confidence level judgment module 105 is used to determine whether the confidence level is greater than a preset level threshold. If not, it determines that the insurance application verification information has passed the verification and generates an insurance data analysis report based on the confidence level and the structured verification information. If yes, it sends the insurance application verification information to a preset manual review client, obtains the feedback data returned by the manual review client, and generates an insurance data analysis report based on the feedback data and the structured verification information.

[0073] In one embodiment, the semantic alignment module 102, when performing the semantic alignment of the image data in the insurance underwriting verification information with the structured verification information based on the cross-modal attention mechanism to form aligned image data and aligned verification information, is specifically used for: Extract the feature vectors from the image data in the insurance application verification information to obtain an image feature vector set; The text feature vector set is obtained by combining the feature vector of each verification information in the structured verification information. Calculate the cosine similarity between each image feature vector in the image feature vector set and each text feature vector in the text feature vector set; Construct attention weight matrices for the image feature vector set and the text feature vector set based on the cosine similarity; Find the image feature vector in the image feature vector set that has the highest similarity to each text feature vector in the text feature vector set, and generate a list of matching vector pairs; Based on the matching vector pair list, a mapping relationship between the image data and the structured verification information is established to obtain the mapped image data and the mapped verification information. The attention weight matrix is ​​used to filter data pairs with confidence levels less than a preset threshold in the mapped image data and the mapped verification information to obtain filtered image data and filtered verification information. The filtered image data and the filtered verification information are packaged into a preset data format to form aligned image data and aligned verification information.

[0074] In one embodiment, the conflict verification module 103, when performing the step of identifying a temporal logical consistency conflict between the aligned image data and the aligned verification information, is specifically used for: Extract the time point data from the aligned image data to obtain an image time point list; Extract the time point data from the aligned verification information to obtain a text time point list; Obtain the mapping relationship between the aligned image data and the aligned verification information to obtain a mapping relationship list; Based on the mapping relationship list, the image time point list is matched with the time points to obtain a time matching pair list; A time conflict determination table is obtained by performing time logic verification on each time matching pair in the time matching pair list using preset time logic verification rules. A structured conflict entry is generated for each conflict determination result in the time conflict determination table to obtain the time logic consistency conflict.

[0075] In one embodiment, the conflict verification module 103, when performing the identification of spatial logical consistency conflicts between the aligned image data and the aligned verification information, is specifically used for: Extract the spatial information contained in the aligned image data to obtain an image space list; Find the fields related to geographical location in the aligned verification information to obtain a list of text spaces; The image space list and the text space list are converted into coordinate format data to obtain an image coordinate list and a text coordinate list; The image coordinate list and the text coordinate list are paired according to the mapping relationship list to obtain a spatial matching pair list; Calculate the spatial distance between each pair of coordinate data in the spatial matching pair list to obtain a spatial distance list; Spatial distances in the spatial distance list that are greater than a preset distance threshold are identified as having spatial conflicts, and a determination result is obtained. The spatial logical consistency conflict is generated based on the determination result.

[0076] In one embodiment, the conflict verification module 103, when performing the extraction of spatial information contained in the aligned image data to obtain an image space list, is specifically used for: Determine whether the aligned image data contains GPS metadata; If the aligned image data contains GPS metadata, then the GPS metadata is converted into spatial information to obtain an image spatial list; If the aligned image data does not contain GPS metadata, then determine whether the aligned image data contains camera watermark text. If the image contains camera watermark text, the camera watermark text is converted into spatial information to obtain an image spatial list. If the image does not contain camera watermark text, then scene semantic recognition is performed on the aligned image data to obtain the scene semantic recognition result; The scene semantic recognition results are converted into spatial information to obtain an image spatial list.

[0077] In one embodiment, the rule verification module 104, when executing the preset rule engine to map the confidence level of the insurance underwriting verification information based on the insured value field, the conflict field, the conflict type, and the conflict severity, is specifically used for: The insured value field, the conflict field, the conflict type, and the conflict severity are converted into key-value pairs to obtain the object to be verified. By using the object to be verified to traverse each preset rule in the rule engine, the matching result between the object to be verified and each preset rule is obtained; Calculate the confidence score of the object to be verified against each preset rule based on the matching results; Calculate the average of all confidence scores and map the average to a confidence level.

[0078] In one embodiment, the confidence level judgment module 105, when performing the step of generating an insurance data analysis report based on the feedback data and the structured verification information, is specifically used for: Extract the anomalies contained in the feedback data to obtain an anomaly list; Extract the review conclusions contained in the feedback data; The preset indicator data in the structured verification information are used to obtain the indicator dataset; The list of anomalies, the review conclusions, the indicator dataset, and the confidence level are mapped to a preset report template to generate the insurance data analysis report.

[0079] This invention provides an insurance data analysis device aimed at solving problems such as low underwriting efficiency, weak multimodal document review capabilities, lagging rule updates, isolated disaster early warning information, and strong subjectivity in disaster loss assessment in traditional agricultural insurance. The solution first uses a multimodal large model (such as Qwen-VL) to perform end-to-end parsing of images and text files uploaded by the insured, including ID cards, public notice photos, ownership certificates, and pigsty photos. A cross-modal attention mechanism is used to achieve fine-grained semantic alignment between image content and structured verification information, generating an image-text semantic alignment map. Then, logical consistency checks are performed on time entities (such as public notice date and insurance application time) and spatial entities (such as GPS coordinates and addresses) in the images and text, identifying conflicts such as time inversion and coordinate distance discrepancies. Finally, a dynamic rule engine (supporting plug-and-play JSON rules and canary release) maps conflicting fields, conflict types, severity, and insured numerical fields (area, insured amount, etc.) to confidence levels. For low-confidence insurance policies, the system pushes risk interception instructions to the core policy issuance system via RESTful API and message queues. Manual review results are fed back to the AI ​​model, forming a closed-loop evolution. Simultaneously, the solution integrates real-time meteorological data, satellite remote sensing imagery, drone aerial photography, and GIS layers. It uses U-Net+++ Transformer to identify disaster types such as floods and droughts and simulate their spread paths, generating early warning information with disposal suggestions. After a disaster, a dual-branch contrastive learning network is used to calculate the disaster loss index, automatically calculating the amount of loss and displaying it on a map visualization platform. Finally, disposal tasks are pushed through an app, farmers upload before-and-after comparison images, AI analyzes the effectiveness of the disposal, and all data is fed back to optimize the model, achieving a closed-loop process of "early warning → disposal → feedback," significantly improving agricultural insurance underwriting efficiency and risk control capabilities.

[0080] Specific limitations regarding the insurance data analysis device can be found in the limitations of the insurance data analysis method described above, and will not be repeated here. Each module in the aforementioned insurance data analysis device can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device in hardware form, or stored in the memory of a computer device in software form, so that the processor can call and execute the operations corresponding to each module.

[0081] In one embodiment, a computer device is provided, which may be a server, and its internal structure diagram may be as follows: Figure 4As shown, the computer device includes a processor, memory, network interface, and database connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile and / or volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and database. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The network interface is used to communicate with external clients via a network connection. When the computer program is executed by the processor, it implements the functions or steps of an insurance data analysis method on the server side.

[0082] In one embodiment, a computer device is provided, which may be a client, and its internal structure diagram may be as follows: Figure 5 As shown, the computer device includes a processor, memory, network interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The network interface is used to communicate with an external server via a network connection. When the computer program is executed by the processor, it implements the functions or steps of a client-side insurance data analysis method.

[0083] In one embodiment, a computer device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to perform the following steps: Obtain insurance application verification information uploaded by the policyholder; The text content in the insurance application verification information is parsed into JSON format to obtain structured verification information; Based on the cross-modal attention mechanism, the image data in the insurance underwriting verification information is semantically aligned with the structured verification information to form aligned image data and aligned verification information; Identify temporal logical consistency conflicts between the aligned image data and the aligned verification information, and identify spatial logical consistency conflicts between the aligned image data and the aligned verification information; Extract the conflict fields, conflict types, and conflict severity from the time logic consistency conflict and the spatial logic consistency conflict, as well as the insurance value field from the structured verification information; Based on a preset rule engine, the confidence level of the insurance underwriting verification information is mapped according to the insured value field, the conflict field, the conflict type, and the conflict severity. Determine whether the confidence level is greater than a preset level threshold; If not, the insurance application verification information is deemed to have passed verification, and an insurance data analysis report is generated based on the confidence level and the structured verification information. If so, the insurance application verification information is sent to a preset manual review client, the feedback data returned by the manual review client is obtained, and an insurance data analysis report is generated based on the feedback data and the structured verification information.

[0084] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, the computer program performing the following steps when executed by a processor: Obtain insurance application verification information uploaded by the policyholder; The text content in the insurance application verification information is parsed into JSON format to obtain structured verification information; Based on the cross-modal attention mechanism, the image data in the insurance underwriting verification information is semantically aligned with the structured verification information to form aligned image data and aligned verification information; Identify temporal logical consistency conflicts between the aligned image data and the aligned verification information, and identify spatial logical consistency conflicts between the aligned image data and the aligned verification information; Extract the conflict fields, conflict types, and conflict severity from the time logic consistency conflict and the spatial logic consistency conflict, as well as the insurance value field from the structured verification information; Based on a preset rule engine, the confidence level of the insurance underwriting verification information is mapped according to the insured value field, the conflict field, the conflict type, and the conflict severity. Determine whether the confidence level is greater than a preset level threshold; If not, the insurance application verification information is deemed to have passed verification, and an insurance data analysis report is generated based on the confidence level and the structured verification information. If so, the insurance application verification information is sent to a preset manual review client, the feedback data returned by the manual review client is obtained, and an insurance data analysis report is generated based on the feedback data and the structured verification information.

[0085] It should be noted that the functions or steps that can be implemented by the computer-readable storage medium or computer device described above can be referred to the relevant descriptions on the server side and client side in the foregoing method embodiments. To avoid repetition, they will not be described one by one here.

[0086] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in a variety of forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0087] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is used as an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above.

[0088] It should be noted that in the technical solutions disclosed in this invention, the acquisition of user information (personal image data (e.g., facial videos or pictures, facial feature videos or pictures, etc.) and personal privacy information (e.g., name, ID number, occupation, address, etc.)) is all completed with the user's knowledge and consent, and the acquisition of the relevant user information is legal and compliant.

[0089] Finally, 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 is authorized (with knowledge and consent) by the relevant parties or fully authorized by all parties, and the executing entity can 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. The above-described embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included within the protection scope of the present invention.

Claims

1. An insurance data analysis method, characterized in that, include: Obtain insurance application verification information uploaded by the policyholder; The text content in the insurance application verification information is parsed into JSON format to obtain structured verification information; Based on the cross-modal attention mechanism, the image data in the insurance underwriting verification information is semantically aligned with the structured verification information to form aligned image data and aligned verification information; Identify temporal logical consistency conflicts between the aligned image data and the aligned verification information, and identify spatial logical consistency conflicts between the aligned image data and the aligned verification information; Extract the conflict fields, conflict types, and conflict severity from the time logic consistency conflict and the spatial logic consistency conflict, as well as the insurance value field from the structured verification information; Based on a preset rule engine, the confidence level of the insurance underwriting verification information is mapped according to the insured value field, the conflict field, the conflict type, and the conflict severity. Determine whether the confidence level is greater than a preset level threshold; If not, the insurance application verification information is deemed to have passed verification, and an insurance data analysis report is generated based on the confidence level and the structured verification information. If so, the insurance application verification information is sent to a preset manual review client, the feedback data returned by the manual review client is obtained, and an insurance data analysis report is generated based on the feedback data and the structured verification information.

2. The insurance data analysis method as described in claim 1, characterized in that, The method of semantically aligning the image data in the insurance underwriting verification information with the structured verification information based on the cross-modal attention mechanism to form aligned image data and aligned verification information includes: Extract the feature vectors from the image data in the insurance application verification information to obtain an image feature vector set; The text feature vector set is obtained by combining the feature vector of each verification information in the structured verification information. Calculate the cosine similarity between each image feature vector in the image feature vector set and each text feature vector in the text feature vector set; Construct attention weight matrices for the image feature vector set and the text feature vector set based on the cosine similarity; Find the image feature vector in the image feature vector set that has the highest similarity to each text feature vector in the text feature vector set, and generate a list of matching vector pairs; Based on the matching vector pair list, a mapping relationship between the image data and the structured verification information is established to obtain the mapped image data and the mapped verification information. The attention weight matrix is ​​used to filter data pairs with confidence levels less than a preset threshold in the mapped image data and the mapped verification information to obtain filtered image data and filtered verification information. The filtered image data and the filtered verification information are packaged into a preset data format to form aligned image data and aligned verification information.

3. The insurance data analysis method as described in claim 1, characterized in that, The identification of temporal logical consistency conflicts between the aligned image data and the aligned verification information includes: Extract the time point data from the aligned image data to obtain an image time point list; Extract the time point data from the aligned verification information to obtain a text time point list; Obtain the mapping relationship between the aligned image data and the aligned verification information to obtain a mapping relationship list; Based on the mapping relationship list, the image time point list is matched with the time points to obtain a time matching pair list; A time conflict determination table is obtained by performing time logic verification on each time matching pair in the time matching pair list using preset time logic verification rules. A structured conflict entry is generated for each conflict determination result in the time conflict determination table to obtain the time logic consistency conflict.

4. The insurance data analysis method as described in claim 1, characterized in that, The identification of spatial logical consistency conflicts between the aligned image data and the aligned verification information includes: Extract the spatial information contained in the aligned image data to obtain an image space list; Find the fields related to geographical location in the aligned verification information to obtain a list of text spaces; The image space list and the text space list are converted into coordinate format data to obtain an image coordinate list and a text coordinate list; The image coordinate list and the text coordinate list are paired according to the mapping relationship list to obtain a spatial matching pair list; Calculate the spatial distance between each pair of coordinate data in the spatial matching pair list to obtain a spatial distance list; Spatial distances in the spatial distance list that are greater than a preset distance threshold are identified as having spatial conflicts, and a determination result is obtained. The spatial logical consistency conflict is generated based on the determination result.

5. The insurance data analysis method as described in claim 4, characterized in that, The step of extracting spatial information contained in the aligned image data to obtain an image space list includes: Determine whether the aligned image data contains GPS metadata; If the aligned image data contains GPS metadata, then the GPS metadata is converted into spatial information to obtain an image spatial list; If the aligned image data does not contain GPS metadata, then determine whether the aligned image data contains camera watermark text. If the image contains camera watermark text, the camera watermark text is converted into spatial information to obtain an image spatial list. If the image does not contain camera watermark text, then scene semantic recognition is performed on the aligned image data to obtain the scene semantic recognition result; The scene semantic recognition results are converted into spatial information to obtain an image spatial list.

6. The insurance data analysis method as described in claim 1, characterized in that, The preset rule engine maps the confidence level of the insurance application verification information based on the insured value field, the conflict field, the conflict type, and the conflict severity, including: The insured value field, the conflict field, the conflict type, and the conflict severity are converted into key-value pairs to obtain the object to be verified. By using the object to be verified to traverse each preset rule in the rule engine, the matching result between the object to be verified and each preset rule is obtained; Calculate the confidence score of the object to be verified against each preset rule based on the matching results; Calculate the average of all confidence scores and map the average to a confidence level.

7. The insurance data analysis method as described in claim 1, characterized in that, The step of generating an insurance data analysis report based on the feedback data and the structured verification information includes: Extract the anomalies contained in the feedback data to obtain an anomaly list; Extract the review conclusions contained in the feedback data; The preset indicator data in the structured verification information are used to obtain the indicator dataset; The list of anomalies, the review conclusions, the indicator dataset, and the confidence level are mapped to a preset report template to generate the insurance data analysis report.

8. An insurance data analysis device, characterized in that, include: The data acquisition module is used to acquire the insurance application verification information uploaded by the policyholder, and parse the text content in the insurance application verification information into JSON format to obtain structured verification information; The semantic alignment module is used to semantically align the image data in the insurance underwriting verification information with the structured verification information based on a cross-modal attention mechanism, forming aligned image data and aligned verification information. The conflict verification module is used to identify temporal logical consistency conflicts between the aligned image data and the aligned verification information, and to identify spatial logical consistency conflicts between the aligned image data and the aligned verification information. The rule verification module is used to extract the conflict fields, conflict types, and conflict severity in the time logic consistency conflict and the spatial logic consistency conflict, as well as the insurance value field in the structured verification information. Based on a preset rule engine, the module maps the confidence level of the insurance underwriting verification information according to the insurance value field, the conflict fields, the conflict types, and the conflict severity. The confidence level judgment module is used to determine whether the confidence level is greater than a preset level threshold. If not, it determines that the insurance application verification information has passed the verification. Based on the confidence level and the structured verification information, an insurance data analysis report is generated. If yes, the insurance application verification information is sent to a preset manual review client, and feedback data returned by the manual review client is obtained. Based on the feedback data and the structured verification information, an insurance data analysis report is generated.

9. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the insurance data analysis method as described in any one of claims 1 to 7.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it implements the insurance data analysis method as described in any one of claims 1 to 7.

Images