Remote underwriting risk identification method based on artificial intelligence

Abstract

The invention discloses a remote underwriting risk identification method based on artificial intelligence, and the method comprises the steps: S1, collecting multi-source information related to a geographic position from a third-party channel according to a protocol when an insurance subject submits an insurance application; s2, performing semantic understanding and standardized conversion on the information acquired in the S1; s3, constructing a product-mechanism-area mapping table according to the product type; s4, querying the mapping table in the step S3 according to the insurance policy of the insurance subject in the step S1, and determining a legal geographic range which is allowed to be underwritten by the insurance policy; and S5, carrying out space matching on the standardized actual risk occurrence place of the insurance subject in the S2 and the legal geographic range allowed to be underwritten by the insurance policy in the S4, and outputting a risk score according to a space matching result. The method has the beneficial effects that by constructing a multi-source data fusion mechanism and integrating multi-modal information, the actual risk occurrence place of the insurance applicant is determined through a weighting algorithm, and misjudgment caused by single information deviation is avoided.

Description

Technical Field

[0001] This application relates to the technical field of the cross-application of artificial intelligence and insurance technology, specifically to an artificial intelligence-based method for identifying risks in cross-regional underwriting. Background Technology

[0002] With the rapid development of internet insurance, online sales and cross-regional insurance applications are becoming increasingly common, leading to a rise in cross-regional underwriting. Cross-regional underwriting not only involves compliance issues between insurance companies' operating areas and regulatory policies, but may also lead to increased underwriting risks, frequent insurance fraud, and difficulties in claims services, among other operational risks. According to relevant regulations, some insurance products are prohibited from being sold outside specific regions. Therefore, timely and accurate identification and control of cross-regional underwriting has become a crucial aspect of insurance companies' risk control and compliance management. Currently, traditional methods for identifying cross-regional underwriting mainly rely on manual review or rule-based system matching, comparing structured fields such as the applicant's ID card address and contact address with the insurance company's geographical location. However, these methods have several limitations: firstly, they struggle to effectively handle unstructured text information (such as freely filled-in descriptions of residence) and have weak semantic understanding capabilities; secondly, rule engines lack flexibility, cannot adapt to complex and ever-changing real-world scenarios, and are susceptible to incomplete or false information submissions, resulting in low accuracy and high rates of missed or false identifications. Summary of the Invention

[0003] The purpose of this application is to provide a method for identifying risks associated with cross-regional insurance based on artificial intelligence. The specific technical solution is as follows:

[0004] An AI-based method for identifying risks associated with cross-regional insurance coverage includes: S1, collecting multi-source information related to geographical location from third-party channels according to an agreement when the insured submits an insurance application; S2, performing semantic understanding and standardization transformation on the information collected in S1; S3, constructing a product-institution-region mapping table based on product type; S4, querying the mapping table in S3 based on the insured's policy in S1 to determine the legal geographical scope of coverage allowed by the policy; S5, spatially matching the standardized location of the insured's actual risk occurrence in S2 with the legal geographical scope of coverage allowed by the policy in S4, and outputting a risk score based on the spatial matching result.

[0005] When collecting multi-source information in S1, the following steps are included: S1.1 Collecting structured fields, specifically including the ID card registration address, permanent address, work unit address, and the region of contact information; S1.2 Collecting unstructured fields, specifically including the description of residence in the insurance declaration; S1.3 Collecting device-side information, specifically including the IP address, GPS coordinates, and device fingerprint obtained with user authorization; S1.4 Collecting historical behavioral data, specifically including the location of past claims, customer service records, and historical underwriting results.

[0006] Before collecting multi-source information in S1, it is necessary to ensure that the insurance application process has been initiated, that there is a legitimate data collection authorization mechanism, and that all external interfaces are functioning properly.

[0007] The semantic understanding and standardization conversion in S2 include: S2.1, constructing an address semantic parsing model based on a BERT-BiLSTM-CRF hybrid architecture for Chinese address named entity recognition tasks and performing confidence scoring; S2.2, constructing a training dataset based on internal anonymized historical insurance data, a public address corpus, and manually labeled samples; S2.3, training the address semantic parsing model constructed in S2.1 using the training dataset constructed in S2.2; S2.4, deploying the address semantic parsing model trained in S2.3 on a local server, parsing the input address text information, and outputting standard administrative division codes or latitude and longitude coordinates with confidence scores.

[0008] The address semantic parsing model in S2.1 includes: a backbone network with BERT-base-chinese as a pre-trained language encoder; an upper layer structure with a bidirectional LSTM to capture contextual dependencies; an output layer with a conditional random field decoder to ensure label sequence consistency; and entity category definitions: province, city, district, street, village, township, or town.

[0009] The address semantic parsing model trained in S2.3 includes the following loss function: joint loss of cross-entropy and CRF path score; optimizer: AdamW, initial learning rate 2e-5, warmup ratio 10%; batch size: 32, maximum sequence length: 128; number of training epochs: 8; evaluation metrics: F1-score, precision, recall, and overall address matching accuracy.

[0010] When constructing the mapping table in S3, the following are included: Product ID → Underwriting Company → List of Provinces / Cities Allowed for Sales; Special restrictions are specified, including local household registration only and / or local social security required.

[0011] S5 includes: S5.1, analyzing the actual location of the risk occurrence for the insured entity based on the standardized information in S2; S5.2, constructing a cross-regional risk model based on a random forest classification architecture; S5.3, constructing a training dataset based on manually confirmed real cross-regional non-compliant underwriting cases and normal local or compliant cross-regional underwriting cases; S5.4, training the cross-regional risk model constructed in S5.2 using the training dataset constructed in S5.3; S5.5, deploying the cross-regional risk model trained in S5.4 on a local server, comparing the analysis results obtained in S5.1 with the mapping table in S3 to perform risk scoring, and outputting the cross-regional underwriting risk level; S5.6, generating structured early warning events based on the risk level output in S5.5 and pushing them to relevant personnel.

[0012] The beneficial effects of this application are as follows: By constructing a multi-source data fusion mechanism, it integrates multimodal information such as ID card address, permanent residence declaration, GPS location (authorized), network IP location, and historical incident locations, and uses a weighted algorithm to determine the actual location of the insured's risk occurrence. This avoids misjudgments due to biases in single pieces of information, thus improving the scientific rigor and robustness of risk assessment. Furthermore, by deeply integrating AI recognition results with business processes, it enables automatic risk warning triggering, intelligent generation of handling suggestions, and automatic interception of high-risk policies. This reduces manual intervention, significantly improves underwriting efficiency and the real-time nature of compliance reviews, and lowers the regulatory penalty risk faced by insurance companies due to illegal cross-regional underwriting. Attached Figure Description

[0013] Figure 1 This is a flowchart illustrating the application process. Specific Implementation

[0014] To make the objectives, technical solutions, and advantages of this application clearer, the application will be further described in detail below with reference to specific embodiments and accompanying drawings. It should be understood that these descriptions are merely exemplary and not intended to limit the scope of this application. Furthermore, descriptions of well-known structures and technologies are omitted in the following description to avoid unnecessarily obscuring the concepts of this application.

[0015] like Figure 1 As shown, an artificial intelligence-based method for identifying risks in cross-regional insurance coverage includes:

[0016] S1. When the insurance applicant submits an insurance application, collect multi-source information related to geographical location from third-party channels according to the agreement. Specifically, when collecting multi-source information in S1, it includes: S1.1. Collect structured fields, specifically including the ID card registration address, permanent address, work unit address, and the region to which the contact information belongs. S1.2. Collect unstructured fields, specifically including descriptions about the place of residence in the insurance application statement (such as "I have been working in Chengdu for three years", "Temporarily on business trip to Guangzhou", etc.). S1.3. Collect device-side information, specifically including the IP address, GPS coordinates, and device fingerprint obtained after user authorization. S1.4. Collect historical behavior data, specifically including the locations where previous insurance policies had claims, customer service records, and historical underwriting results. Before collecting multi-source information, it is necessary to ensure that the insurance application process has been started, there is a legal data collection authorization mechanism, and all external interfaces are normal and available.

[0017] S2. Perform semantic understanding and standardized conversion on the information collected in S1. Specifically, when performing semantic understanding and standardized conversion in S2, it includes:

[0018] S2.1. Construct an address semantic parsing model based on the BERT-BiLSTM-CRF hybrid architecture for Chinese address named entity recognition tasks and perform confidence scoring. The address semantic parsing model includes: a backbone network, BERT-base-chinese as a pre-trained language encoder; an upper structure, a bidirectional LSTM to capture context dependencies; an output layer, a conditional random field decoder to ensure label sequence consistency; entity category definitions: province, city, district, street, village, township or town.

[0019] S2.2. Construct a training dataset based on internal desensitized historical insurance application data, public address corpora, and manually annotated samples. In actual applications, the data sources are: internal historical insurance application data (desensitized), public address corpora (such as Baidu Encyclopedia POI), and manually annotated samples; the number of samples: a total of 120,000 real address texts are collected and annotated, covering standard formats, colloquial expressions, typos, abbreviations (such as "Hu" = Shanghai), and nicknames (such as "Beijing, Shanghai, Guangzhou, Shenzhen"); data augmentation strategies: synonym replacement (such as "live in" → "reside in"), typo injection (simulating user input errors), address abbreviation (such as "No. XX, Tianfu Avenue, Wuhou District, Chengdu City" → "Wuhou District, Tianfu Avenue"); the input is character-level tokenization (using the WordPiece tokenizer); add position embeddings and词性标记 (such as "province" and "city" suffixes) to assist in recognition; introduce an external knowledge graph: access the "Administrative Division Code Table of the People's Republic of China" (GB / T 2260-2023) released by the National Bureau of Statistics for candidate address verification and completion.

[0020] It should be noted that the term "词性标记" in the original text seems to be a misspelling. It might be "词性标注" in Chinese, which is translated as "pos tags" in English. But I translated it as "词性标记" according to your request to keep the original text as it is. If this is an important error, it may need to be corrected in the original text for a more accurate translation.S2.3. Use the training dataset constructed in S2.2 to train the address semantic parsing model constructed in S2.1. During the actual application process, the training process includes: Loss function: Joint loss of cross-entropy + CRF path score; Optimizer: AdamW, initial learning rate 2e-5, warmup ratio 10%; Batch size: 32, maximum sequence length: 128; Number of training epochs: 8 epochs; Evaluation metrics: F1-score, precision, recall, overall address matching accuracy; Performance: The F1 of the test set reaches 96.3%, and the recognition accuracy of special expressions such as "drifting in Chengdu" and "working in Beijing" reaches 91.5%.

[0021] S2.4. Deploy the trained address semantic parsing model in S2.3 to the local server, parse the input address text information, and output the standard administrative division code or longitude and latitude coordinates with a confidence score. During the actual application process, export the model in ONNX format and deploy it to the local GPU server; Inference latency < 150ms / item; Support batch processing and real-time streaming inference. Output result: Each address text is parsed into a standard administrative division code (e.g., "510107" represents Wuhou District, Chengdu City, Sichuan Province) or longitude and latitude coordinates (WGS-84), and is accompanied by a confidence score (0-1). Results with low confidence (<0.7) enter the manual review queue.

[0022] S3. Construct a product-institution-region mapping table according to the product type. Specifically, when constructing the mapping table, it includes: Product ID → Underwriting company → List of allowed sales provinces / cities; Special restriction instructions, limited to local household registration and / or local social security, etc.

[0023] S4. Query the mapping table in S3 according to the insurance policy of the insured entity in S1 to determine the legal geographical scope allowed for underwriting of the insurance policy. Specifically, if a regional inclusive insurance product only allows "permanent residents within Sichuan Province" to participate in the insurance, its underwriting jurisdiction is the administrative division codes of all prefecture-level cities under Sichuan Province. The system has a built-in compliance rule library, which supports dynamic loading and version management.

[0024] S5. Perform spatial matching between the actual risk occurrence location of the standardized insured entity in S2 and the legal geographical scope allowed for underwriting of the insurance policy in S4, and output a risk score according to the spatial matching result. Specifically, it includes:

[0025] S5.1. Analyze the actual risk location of the insured entity based on the standardized information in S2. In practical applications, the logic for determining the actual risk location is as follows: If multiple address sources exist (ID card, permanent residence, IP address, GPS), a weighted fusion algorithm is used: Final risk location = w1 * permanent residence + w2 * GPS + w3 * IP address + w4 * work unit address. Example of weight settings: permanent residence (0.4), GPS (0.3), IP address (0.2), other (0.1). If there is no GPS / IP data, permanent residence is used as the primary source. If all data is unstructured text, the parsing result with the highest confidence level is used.

[0026] S5.2 Construct an off-site risk model based on a random forest classification framework. Specifically, determine whether the provincial administrative division code of the final risk location is within the insured area; inconsistency triggers a preliminary "off-site suspicion" signal. This signal, along with other features, is input into the random forest classification model to assess the probability of off-site insured risk.

[0027] S5.3. A training dataset is constructed based on manually verified real cases of illegal out-of-town insurance and normal local or compliant out-of-town insurance cases. Specifically, positive samples are manually verified real cases of illegal out-of-town insurance (such as out-of-town residents without proof of residence applying for regionally exclusive products); negative samples are normal local or compliant out-of-town insurance (such as nationwide products); the dataset contains 52,000 positive and negative samples (positive samples account for approximately 18%), spanning from 2021 to 2024.

[0028] Random forests have good resistance to overfitting and are suitable for small to medium-sized structured data; they support feature importance ranking, which facilitates the interpretation of model decisions; they have high tolerance for missing values ​​and do not require complex preprocessing; they can be trained in parallel and have high deployment efficiency.

[0029] S5.4. The off-site risk model constructed in S5.2 is trained using the training dataset constructed in S5.3. Specifically,

[0030] Training parameter settings:

[0031] Number of decision trees: n_estimators=200;

[0032] Maximum depth: max_depth=12;

[0033] Minimum sample split: min_samples_split=10;

[0034] Feature sampling ratio: max_features='sqrt';

[0035] Evaluation method: 5-fold cross-validation.

[0036] Performance:

[0037] AUC: 0.928;

[0038] Precision (top 10%): 89.6%;

[0039] Recall rate (high-risk cases): 85.4%.

[0040] Top three features by importance:

[0041] address_mismatch_level;

[0042] ip_province_match;

[0043] product_is_regional_only.

[0044] S5.5 Deploy the off-site risk model trained in S5.4 on the local server. Compare the parsing results obtained in S5.1 with the mapping table in S3 to perform risk scoring and output the off-site underwriting risk level. Specifically, use Scikit-learn for training and save as a .pkl file; integrate it into a Java / Spring Boot service (via Python subprocess call or Jython bridging); single inference time < 50ms. Judgment logic: Risk score ≥ 0.85: High risk, automatic interception; 0.7 ≤ Risk score < 0.85: Medium risk, prompt for supplementary materials; < 0.7: Low risk, continue the underwriting process.

[0045] S5.6. Generate structured early warning events based on the risk levels output in S5.5 and push them to relevant personnel. Specifically,

[0046] The warning content includes:

[0047] Risk summary (e.g., "cross-province insurance + inconsistent IP + high-frequency devices");

[0048] Key evidence chain screenshots (original address text, parsing results, IP / GPS trajectory);

[0049] Model decision path (visualization of the contribution of key features).

[0050] Suggested operating instructions (see table below):

[0051] High risk (≥0.85) will automatically suspend the policy, place it in the manual review queue, and notify the anti-fraud team to intervene.

[0052] Medium risk (0.7–0.85), will receive a text message / internal message requesting supplementary proof of residence or social security payment records.

[0053] Low risk (<0.7), continue with the automatic underwriting process.

[0054] Required conditions: Message middleware (such as Kafka / RabbitMQ), ticketing system, and SMS gateway have been integrated.

[0055] In subsequent applications, a data closed-loop mechanism will be established to enable the model to self-evolve.

[0056] Feedback data source:

[0057] Underwriters confirm / correct the warning results;

[0058] Subsequent claims investigation findings (whether it was indeed a case of insurance fraud in another location);

[0059] Customer complaint results.

[0060] Feedback process:

[0061] Feedback data is extracted daily at set times, cleaned, and then added to the training set.

[0062] Perform an incremental training task (incremental learning strategy) once a month.

[0063] A / B testing was conducted on the new model before it was launched to compare the AUC improvement of the old model on the retained test set.

[0064] If the AUC increases by ≥ 0.01, a new version will be released.

[0065] Model monitoring dashboard:

[0066] Real-time monitoring: Daily average number of warnings, false alarm rate, and interception success rate;

[0067] Model stability: A PSI (Population Stability Index) < 0.1 is considered stable;

[0068] Feature drift detection: Monthly checks on the distribution changes of the top 5 features.

[0069] Requirements: MLOps platform with support for model version management and automated training pipeline.

[0070] To make this application easier to understand, further explanation is provided below with reference to specific applications.

[0071] Example 1

[0072] When using the method described in this application to determine the risk level of cross-regional underwriting, the following are included:

[0073] Step S101: Collection of multi-source insurance information

[0074] When an insured submits an insurance application, the system automatically collects the following information via an API interface:

[0075] Structured fields: ID card registration address, contact address, workplace location;

[0076] Unstructured text: Descriptions of place of residence and permanent residence in the insurance application statement;

[0077] Device information: User-authorized IP address and GPS location coordinates;

[0078] Historical data: the location of the accident and service records in the customer's past policies.

[0079] The information is uniformly collected into a data preprocessing queue to form a set of raw inputs to be analyzed.

[0080] Step S102: Geographic semantic analysis and standardization

[0081] The collected information is input into an address semantic parsing model deployed on a local server. This model, fine-tuned based on the BERT architecture, possesses the ability to recognize address elements at four levels: province, city, district, and street. It also supports semantic inference for colloquial expressions such as "working in Sichuan," "living in Chengdu," and "long-term overseas posting." The model outputs standard administrative division codes (GB / T 2260) and latitude / longitude coordinates, along with confidence levels. For low-confidence results, the system automatically marks them and transfers them to a manual verification process.

[0082] Step S103: Determination of the insured's place of jurisdiction

[0083] Based on the type of insurance product (such as health insurance), consult the product-institution-region mapping table to obtain the geographical scope that the policy is allowed to cover. For example, if a regional health insurance product is only available to "registered or permanent residents of Sichuan Province", then its legal coverage area is the set of administrative division codes of all prefecture-level cities under the jurisdiction of Sichuan Province.

[0084] Step S104: Risk Matching and Assessment in Different Locations

[0085] The actual location of the insured's risk, determined in step S102 (primarily the place of residence, followed by a weighted calculation of IP and GPS locations), is spatially compared with the underwriting location in step S103. If the two locations are in different provincial-level administrative regions, a preliminary risk signal is triggered. Furthermore, this comparison result, along with 15 features including device fingerprint, insurance frequency, and historical claim count, is input into a random forest classification model (trained using 50,000 historical samples, AUC=0.92), outputting a risk score (0-1) for out-of-town underwriting. When the score exceeds a preset threshold of 0.7, it is determined to be a high-risk out-of-town underwriting behavior.

[0086] Step S105: Generation of Risk Warning and Handling Suggestions

[0087] The system automatically generates risk warning events, pushes them to the underwriting workbench, and generates handling suggestions based on the rules engine:

[0088] A risk score of 0.7–0.85 indicates that "supplementary proof of residence is required."

[0089] Risk score ≥ 0.85: The policy is automatically suspended and enters the manual review queue;

[0090] Frequent and concentrated insurance applications from the same IP address: This triggers an anti-fraud alert and notifies the compliance department to intervene and investigate.

[0091] The warning information includes a summary of risk points, screenshots of the evidence chain, and the model's decision path, supporting retrospective review.

[0092] Step S106: Model Feedback and Continuous Optimization

[0093] Underwriting personnel confirm or correct the warning results, and their operational results are fed back to the model training platform as labeled data. An incremental training task is performed monthly to update model parameters, and the performance of the new model is verified through A / B testing. Once the performance meets the standards, it is automatically released as a production version, realizing the closed-loop evolution of the system.

[0094] Example 2

[0095] Full-process running example

[0096] (I) A customer applied for a "Sichuan-Chongqing Exclusive Critical Illness Insurance" product online and filled in the following information:

[0097] ID card address: Wuhou District, Chengdu City, Sichuan Province

[0098] Description of permanent residence: "Currently renting an apartment near Tencent Building in Nanshan District, Shenzhen."

[0099] IP address: Shenzhen, Guangdong Province

[0100] GPS coordinates: 22.533, 113.930 (Nanshan District, Shenzhen)

[0101] I have insured the same device 5 times in the past week.

[0102] (II) System Execution Flow:

[0103] Information collection: All the above fields were successfully retrieved;

[0104] Semantic analysis: The BERT-NER model identifies "Shenzhen Nanshan District" as Nanshan District, Shenzhen City, Guangdong Province (code 440305), with a confidence level of 0.98;

[0105] Coverage location determination: This product is only available to residents of the Sichuan-Chongqing region (codes 51xx / 50xx);

[0106] Location matching: The customer's actual risk location is Guangdong, which is not covered by the insurance → Preliminary risk;

[0107] Model evaluation:

[0108] Feature vectors are input into a random forest model;

[0109] The output risk score is 0.91;

[0110] Warning generated:

[0111] Automatically suspend the policy;

[0112] A warning was sent to the underwriter: "High-risk out-of-town underwriting, we recommend rejection."

[0113] The anti-fraud team was simultaneously notified to investigate duplicate insurance claims for the equipment.

[0114] Follow-up feedback: After confirmation by the underwriter, the sample was marked as a "genuine violation" and was added to the training set for the next model update.

[0115] Example 3: Privacy and Security Mechanisms

[0116] All processing involving personal geolocation information complies with the requirements of the Personal Information Protection Law. Explicit consent from users is obtained before collection, and sensitive data is anonymized before storage. Differential privacy technology is used during model training to prevent information leakage.

[0117] In summary, through the above specific embodiments, this invention achieves intelligent and accurate identification of cross-regional underwriting activities, and has good practicality, stability and scalability, making it suitable for risk control scenarios in various insurance business systems.

Claims

1. A method for identifying risks in cross-regional insurance based on artificial intelligence, characterized in that, include: S1. When the insured submits an insurance application, multi-source information related to geographical location is collected from third-party channels in accordance with the agreement; S2. Perform semantic understanding and standardization transformation on the information collected in S1; S3. Construct a product-organization-region mapping table based on product type; S4. Based on the policy of the insured entity in S1, query the mapping table in S3 to determine the legal geographical scope that the policy is allowed to cover; S5. Spatial matching is performed between the actual location of the risk occurrence of the insured entity as standardized in S2 and the legal geographical range that the policy allows to cover in S4, and a risk score is output based on the spatial matching result.

2. The method for identifying cross-regional underwriting risks based on artificial intelligence as described in claim 1, characterized in that, The collection of multi-source information in S1 includes: S1.1 Collect structured fields, specifically including ID card registration address, permanent address, work unit address, and the region of contact information; S1.2 Collect unstructured fields, specifically including the description of the place of residence in the insurance declaration; S1.3 Collect device information, specifically including IP address, GPS coordinates and device fingerprint obtained after user authorization; S1.4 Collect historical behavioral data, specifically including past policy claims locations, customer service records, and historical underwriting results.

3. The method for identifying cross-regional underwriting risks based on artificial intelligence as described in claim 2, characterized in that, Before collecting multi-source information, S1 must ensure that the insurance application process has been initiated, that there is a legitimate data collection authorization mechanism, and that all external interfaces are functioning properly.

4. The method for identifying cross-regional underwriting risks based on artificial intelligence as described in claim 1, characterized in that, The semantic understanding and standardization conversion performed in S2 include: S2.1 Construct an address semantic parsing model based on the BERT-BiLSTM-CRF hybrid architecture for Chinese address named entity recognition tasks, and perform confidence scoring; S2.

2. Construct a training dataset based on internally anonymized historical insurance data, a public address corpus, and manually labeled samples; S2.

3. The address semantic parsing model constructed in S2.1 is trained using the training dataset constructed in S2.2; S2.4 Deploy the address semantic parsing model trained in S2.3 on a local server, parse the input address text information, and output the standard administrative division code or latitude and longitude coordinates with confidence scores.

5. The method for identifying cross-regional underwriting risks based on artificial intelligence as described in claim 4, characterized in that, The address semantic parsing model in S2.1 includes: The backbone network uses BERT-base-chinese as a pre-trained language encoder. Upper-level structure: Bidirectional LSTM captures context dependencies; The output layer, a conditional random field decoder, ensures the consistency of the label sequence; Entity category definition: province, city, district, street, village, township or town.

6. The method for identifying cross-regional underwriting risks based on artificial intelligence as described in claim 4, characterized in that, The training of the address semantic parsing model in S2.3 includes: Loss function: Joint loss of cross-entropy and CRF path score; Optimizer: AdamW, initial learning rate 2e-5, warmup ratio 10%; Batch size: 32, maximum sequence length: 128; Training epochs: 8; Evaluation metrics: F1-score, precision, recall, and overall address matching accuracy.

7. The method for identifying cross-regional underwriting risks based on artificial intelligence as described in claim 1, characterized in that, The construction of the mapping table in S3 includes: Product ID → Underwriting Company → List of Provinces / Cities Permitted for Sale; Special restrictions include: only local residents and / or those with local social security contributions are eligible.

8. The method for identifying cross-regional underwriting risks based on artificial intelligence as described in claim 1, characterized in that, S5 includes: S5.1 Analyze the actual location of the risk occurrence for the insured entity based on the standardized information in S2; S5.2 Construct a cross-regional risk model based on a random forest classification architecture; S5.

3. Construct a training dataset based on real cases of illegal out-of-town insurance that have been manually verified, as well as normal local insurance or compliant out-of-town insurance cases; S5.

4. The off-site risk model constructed in S5.2 is trained using the training dataset constructed in S5.

3. S5.5 Deploy the off-site risk model trained in S5.4 on a local server, compare the parsing results obtained in S5.1 with the mapping table in S3 to perform risk scoring, and output the off-site underwriting risk level.

9. The method for identifying cross-regional underwriting risks based on artificial intelligence as described in claim 8, characterized in that, The S5 also includes: S5.

6. Generate a structured early warning event based on the risk level output in S5.5, and push it to the relevant responsible person.

Images