An intelligent claim settlement auxiliary system based on artificial intelligence

Abstract

The application discloses an intelligent claim settlement auxiliary system based on artificial intelligence, comprising the following modules: a data acquisition module for acquiring claim image data and corresponding insurance clause content and distributing a unique index; a clause analysis module for structurally analyzing the insurance clause to generate a claim settlement judgment basis set; a character recognition module for recognizing character information in the image data and outputting structured claim settlement text; a semantic coding module for context semantic coding of the claim settlement text to obtain a semantic representation vector; a clause matching module for performing matching judgment of the text and the clause; a compensation determination module for outputting a compensation prediction result, an amount estimation and a reason label; and an auxiliary report generation module for generating a structured claim settlement auxiliary report. The application realizes an integrated processing flow of automatic identification of image and text information, clause semantic matching and intelligent compensation determination.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and insurance information processing technology, and in particular to an intelligent claims assistance system based on artificial intelligence. Background Technology

[0002] With the deepening application of artificial intelligence technology in the financial and insurance fields, intelligent claims assistance technology based on image recognition and natural language processing is gradually becoming a key path to improve claims efficiency and decision-making accuracy. Existing intelligent claims systems mainly rely on image text recognition technology to extract text from claims documents and perform preliminary screening and judgment by comparing the information with insurance terms using simple rule matching or keyword-based retrieval methods. However, in practical applications, existing technologies generally suffer from the following problems:

[0003] Textual information in image data is complex and diverse. Traditional OCR models have low recognition accuracy in scenarios such as medical invoices and handwritten documents, making it difficult to guarantee the integrity and accuracy of structured data. There is a lack of deep semantic understanding between text and insurance terms. Existing solutions based on key field matching tend to ignore contextual dependencies and ambiguous expressions, leading to inaccurate judgments on the applicability of terms. The claims determination process usually fails to introduce multi-task modeling strategies, and lacks systematic integration of information such as claims results, amount estimates, and explanations of causes, making it difficult to support intelligent claims decision-making. In addition, some systems have not established a clear structured output process, and the generated auxiliary reports are loosely structured and redundant in content, which is not conducive to manual verification and subsequent auditing.

[0004] Therefore, how to provide an intelligent claims assistance system based on artificial intelligence is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose an intelligent claims assistance system based on artificial intelligence. This invention combines image recognition, context-enhanced semantic coding, multi-round rule-enhanced matching networks, and a multi-task claims determination model to comprehensively improve the structured processing capability of claims data and the accuracy of clause application judgment. It achieves comprehensive prediction of whether to pay, the recommended payout amount, and the reason for payout, and generates a structured claims assistance report. It has the advantages of strong semantic understanding capability, high accuracy of clause matching, and strong interpretability of judgment results.

[0006] According to an embodiment of the present invention, an intelligent claims assistance system based on artificial intelligence includes the following modules:

[0007] The data acquisition module is used to collect the claim image data submitted by the insured and the corresponding insurance product terms and conditions, and to assign a unique index identifier to each claim application.

[0008] The terms and conditions parsing module is used to perform structured parsing of the terms and conditions of insurance products and generate a set of criteria for claims judgment.

[0009] The text recognition module is used for text recognition and outputs structured claims text information;

[0010] The semantic encoding module is used to perform context-enhanced feature encoding on structured claims text information to obtain the context semantic representation vector of each claims text content;

[0011] The clause matching module is used to perform clause matching tasks and output clause matching scores and satisfaction status results;

[0012] The compensation determination module is used to perform compensation determination and amount estimation tasks, and output the prediction result of whether to pay, the suggested compensation amount, and the compensation reason label;

[0013] The auxiliary report generation module is used to generate structured claims auxiliary reports.

[0014] According to an embodiment of the present invention, an intelligent claims assistance method based on artificial intelligence includes the following steps:

[0015] Collect the image data of the claim submitted by the insured and the corresponding insurance product terms, and assign a unique index identifier to each claim application;

[0016] The terms and conditions of insurance products are analyzed in a structured manner to generate a set of criteria for claims judgment;

[0017] The claim image data is input into the improved PaddleOCR model for text recognition to obtain structured claim text information;

[0018] The structured claims text information is semantically encoded to obtain the contextual semantic representation vector of each claims text content;

[0019] The context semantic representation vector and the set of claims judgment criteria are input into a multi-round rule-enhanced BERT matching network to perform a clause matching task and output the clause matching score and the result of the satisfaction status judgment.

[0020] The structured claims text information, clause matching score, and satisfaction judgment result are jointly input into a multi-task semantic-driven claims determination network to perform claims determination and amount estimation tasks, and output the prediction result of whether to pay, the suggested payout amount, and the claim reason label.

[0021] A structured claims assistance report is generated based on the predicted payout amount, the suggested payout amount, and the payout reason label.

[0022] Optionally, the step of performing structured parsing of the insurance product terms to generate a set of claims judgment criteria specifically includes:

[0023] The content of insurance product terms is structured by combining rule template parsing and syntactic dependency analysis. The rule template parsing extracts fields such as compensation liability, triggering conditions, compensation ratio, exclusions, and supporting documentation requirements by matching preset insurance term language patterns.

[0024] If the rule template cannot be matched, the syntactic dependency analysis method is used. Based on verb center structure recognition and grammatical dependency path calculation, the subject-verb-object structure and modifiers related to the compensation conditions are automatically extracted from the content of the insurance product terms and conditions to complete the field extraction results.

[0025] Each extracted field is assigned a field identifier and standard semantic tag according to a preset field classification rule to construct a set of clause structure fields.

[0026] Based on the set of clause structure fields, the fields are grouped according to the similarity of compensation type, the consistency of the insured subject, and the logical dependency of the clause. The semantic association matrix between the fields is constructed by calculating the cosine similarity of sentence vectors, and a set of claims judgment criteria with structural organization and semantic mapping relationship is generated.

[0027] Optionally, the step of inputting the claim image data into the improved PaddleOCR model for text recognition to obtain structured claim text information specifically includes:

[0028] An improved PaddleOCR model is constructed, which includes a ticket structure localization module, a text detection subnetwork, a text orientation classification subnetwork, and a character recognition subnetwork.

[0029] The input claim image data is preprocessed through the bill structure localization module. Specifically, this includes performing page structure segmentation, locating the main area of ​​the medical bill, cropping interfering content at the image edges, and extracting suspected information block areas as candidate areas for structured recognition.

[0030] The structured recognition candidate regions are input into the text detection subnetwork, which adopts a detection structure based on differential feature pyramids to extract multi-scale text bounding boxes and output detection results including text region coordinates, confidence scores and candidate order labels.

[0031] The detection results are input into the text orientation classification sub-network, which uses a lightweight convolutional structure to determine the orientation angle of each text candidate box and performs orientation rotation correction on the detected text image regions to make the text images be arranged in a uniform orientation.

[0032] The corrected text image region is input into the text recognition sub-network. The convolutional feature extraction structure, sequence modeling structure and character decoding structure complete the text image feature extraction, sequence learning and text recognition operations in sequence. During the character decoding process, a medical terminology dictionary error correction mechanism is introduced. The candidate character sequence in the text recognition result is matched with the preset medical terminology dictionary. When there is a significant difference between the recognition result and the dictionary entry, it is replaced with the most similar standard term.

[0033] The final output is structured claim text information, including the time of visit, the hospital visited, the diagnosis, and the fields for cost items and amount.

[0034] Optionally, the step of semantically encoding the structured claims text information to obtain a contextual semantic representation vector for each claims text content specifically includes:

[0035] The medical time, medical institution, diagnosis, medical items and amount information in the structured claim text are concatenated according to the preset field order to construct a standard input text sequence;

[0036] The input text sequence is fed into a context-enhanced feature encoder for semantic encoding. The context-enhanced feature encoder includes a character-level convolutional feature extraction layer, a bidirectional long short-term memory network modeling layer, and a semantic attention aggregation layer. The semantic attention aggregation layer performs weighted combination of temporal hidden states to generate a fused global contextual semantic representation vector.

[0037] Optionally, the step of inputting the context semantic representation vector and the set of claims judgment criteria into a multi-round rule-enhanced BERT matching network to perform a clause matching task and outputting a clause matching score and a satisfaction judgment result specifically includes:

[0038] The context semantic representation vector and the set of claims judgment criteria are input into a multi-round rule-enhanced BERT matching network to perform a clause matching task and output the clause matching score and the result of the satisfaction status judgment.

[0039] The multi-round rule-enhanced BERT matching network includes a multi-round semantic interaction modeling layer, a structural field alignment mechanism, and a rule label-guided scoring module;

[0040] The multi-round semantic interaction modeling layer is based on a dual-input attention mechanism. In each round, it constructs a cross-attention weight matrix between the context semantic representation vector and the clause field semantic vector, and iteratively updates the matching feature representation in multiple rounds to extract matching relationships at different semantic granularities.

[0041] The structure field alignment mechanism aligns the structure fields in the structured claims text information with the semantic fields in the claims judgment basis set according to the preset field label mapping rules, and constructs a set of field-level semantic alignment vectors.

[0042] The rule-label-guided scoring module inputs the rule label representations of the field-level semantic alignment vector set and the claim judgment basis set, including the claim conditions field, exclusion liability field, and supporting document requirement field, into the scoring function. Based on cosine semantic similarity and rule satisfaction, it calculates the matching score and satisfaction judgment result for each clause. The rule satisfaction is logically judged based on the claim judgment basis set corresponding to the clause field. When the structured claim field content satisfies all the required condition labels, the rule satisfaction is recorded as 1; otherwise, it is recorded as 0.

[0043] Optionally, the multi-task semantic-driven compensation determination network specifically includes:

[0044] A multi-task semantic-driven compensation determination network is constructed, which includes a compensation determination subnetwork, an amount estimation subnetwork, and a compensation reason label generation subnetwork.

[0045] The structured claims text information, clause matching scores, and satisfaction judgment results are concatenated by field dimension to form a joint semantic feature vector;

[0046] The joint semantic feature vector is input into the compensation judgment sub-network, which adopts a dual-tower semantic classification structure. Each tower consists of two layers of Transformer encoder and feedforward network. By modeling the semantics of the claims text and the clause matching information respectively, and introducing a cross-attention mechanism to establish the coupling relationship between the semantics of compensation and the judgment basis, the prediction result of whether to pay compensation is output.

[0047] The information of the amount-related fields in the joint semantic feature vector is input into the amount estimation sub-network. The amount estimation sub-network sequentially performs feature transformation on the amount field, treatment method, and compensation ratio through a fully connected layer and a ReLU activation function, and outputs the suggested compensation amount prediction value.

[0048] The joint semantic feature vector is input into the compensation reason label generation sub-network. Specifically, the compensation reason label generation sub-network inputs the joint semantic feature vector into a one-layer feedforward fully connected network to perform nonlinear feature transformation and outputs the compensation reason label after being processed by the Sigmoid activation function.

[0049] Optionally, the generation of a structured claims assistance report based on the prediction of whether compensation will be paid, the suggested compensation amount, and the compensation reason label specifically includes:

[0050] The prediction of whether compensation will be paid, the suggested compensation amount, and the compensation reason label are integrated to generate a summary of the claims determination.

[0051] Based on the clause matching score and the satisfaction judgment result, generate clause explanation information corresponding to the claim reason label;

[0052] If the claim reason label includes a claim rejection or risk-related label, generate structured risk warning information;

[0053] Based on the preset claims assistance report format template, a structured claims assistance report is constructed, which includes a unique index identifier for the claims application, insurance product terms and conditions information, a summary of the claims determination, a recommended compensation amount, and explanatory information on the terms and conditions.

[0054] The beneficial effects of this invention are:

[0055] (1) Enhance the ability of claims data to be structured and semantically understood: By inputting claims image data into an improved OCR model for structured text extraction and introducing a context-enhanced semantic encoder, it can effectively parse non-standard format text and medical terminology, generate representation vectors with contextual semantic consistency, and significantly improve the machine parsingability and semantic understanding accuracy of claims data.

[0056] (2) Enhanced ability to judge the applicability of clause matching: This invention constructs a multi-round rule-enhanced matching network, integrates the claim text and clause semantic information, and performs clause matching tasks by combining semantic similarity and rule satisfaction. This not only improves the accuracy of clause adaptation, but also has an adjustable judgment strategy, enabling fine-grained identification of the applicability of compensation under complex clause conditions.

[0057] (3) Achieve intelligent and interpretable output of claims determination: Based on the multi-task semantic-driven claims determination network, this invention simultaneously outputs claims decision, amount estimation and claims reason label, and realizes intelligent closed loop from data identification, clause judgment to decision suggestion through the automatic generation mechanism of structured auxiliary report. It has a high level of automation and good interpretability, and significantly improves the efficiency of claims process and decision quality in the insurance industry. Attached Figure Description

[0058] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0059] Figure 1 This is a schematic diagram of the structure of an artificial intelligence-based intelligent claims assistance system proposed in this invention;

[0060] Figure 2This is a flowchart of an artificial intelligence-based intelligent claims assistance method proposed in this invention;

[0061] Figure 3 This is a framework diagram of the improved PaddleOCR model in the intelligent claims assistance method based on artificial intelligence proposed in this invention. Detailed Implementation

[0062] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0063] refer to Figure 1 An intelligent claims assistance system based on artificial intelligence includes the following modules:

[0064] The data acquisition module is used to collect the claim image data submitted by the insured and the corresponding insurance product terms and conditions, and to assign a unique index identifier to each claim application.

[0065] The terms and conditions parsing module is used to perform structured parsing of the terms and conditions of insurance products and generate a set of criteria for claims judgment.

[0066] The text recognition module is used for text recognition and outputs structured claims text information;

[0067] The semantic encoding module is used to perform context-enhanced feature encoding on structured claims text information to obtain the context semantic representation vector of each claims text content;

[0068] The clause matching module is used to perform clause matching tasks and output clause matching scores and satisfaction status results;

[0069] The compensation determination module is used to perform compensation determination and amount estimation tasks, and output the prediction result of whether to pay, the suggested compensation amount, and the compensation reason label;

[0070] The auxiliary report generation module is used to generate structured claims auxiliary reports.

[0071] refer to Figure 2-3 A smart claims assistance method based on artificial intelligence includes the following steps:

[0072] Step 1: Collect the claim image data submitted by the insured and the corresponding insurance product terms and conditions, and assign a unique index identifier to each claim application;

[0073] Step 2: Perform structured analysis of the insurance product terms and conditions to generate a set of criteria for claims judgment;

[0074] Step 3: Input the claim image data into the improved PaddleOCR model for text recognition to obtain structured claim text information;

[0075] Step 4: Semantically encode the structured claims text information to obtain the contextual semantic representation vector of each claims text content;

[0076] Step 5: Input the context semantic representation vector and the set of claims judgment criteria into the multi-round rule-enhanced BERT matching network, execute the clause matching task, and output the clause matching score and the judgment result of satisfaction status;

[0077] Step Six: Input the structured claims text information, clause matching score and satisfaction judgment result into the multi-task semantic-driven claims determination network to perform claims determination and amount estimation tasks, and output the prediction result of whether to pay, the suggested payout amount and the claim reason label;

[0078] Step 7: Generate a structured claims assistance report based on the predicted payout result, suggested payout amount, and payout reason label.

[0079] In this embodiment, the step of performing structured parsing of the insurance product terms and conditions to generate a set of claims judgment criteria specifically includes:

[0080] The content of insurance product terms is structured by combining rule template parsing and syntactic dependency analysis. The rule template parsing extracts fields such as compensation liability, triggering conditions, compensation ratio, exclusions, and supporting documentation requirements by matching preset insurance term language patterns.

[0081] If the rule template cannot be matched, the syntactic dependency analysis method is used. Based on verb center structure recognition and grammatical dependency path calculation, the subject-verb-object structure and modifiers related to the compensation conditions are automatically extracted from the content of the insurance product terms and conditions to complete the field extraction results.

[0082] Each extracted field is assigned a field identifier and standard semantic tag according to a preset field classification rule to construct a set of clause structure fields.

[0083] Based on the set of clause structure fields, the fields are grouped according to the similarity of compensation type, the consistency of the insured subject, and the logical dependency of the clause. The semantic association matrix between the fields is constructed by calculating the cosine similarity of sentence vectors, and a set of claims judgment criteria with structural organization and semantic mapping relationship is generated.

[0084] This implementation method combines rule template parsing and syntactic dependency analysis to perform structured parsing of insurance product terms. This not only ensures the accurate extraction of known semantic patterns but also enables intelligent completion of complex or non-standard expressions, significantly improving the completeness and semantic accuracy of compensation-related field extraction. It effectively solves the problem of poor adaptability of traditional parsing methods to flexible terms.

[0085] In this embodiment, the step of inputting the claim image data into the improved PaddleOCR model for text recognition to obtain structured claim text information specifically includes:

[0086] An improved PaddleOCR model is constructed, which includes a ticket structure localization module, a text detection subnetwork, a text orientation classification subnetwork, and a character recognition subnetwork.

[0087] The input claim image data is preprocessed through the bill structure localization module. Specifically, this includes performing page structure segmentation, locating the main area of ​​the medical bill, cropping interfering content at the image edges, and extracting suspected information block areas as candidate areas for structured recognition.

[0088] The structured recognition candidate regions are input into the text detection subnetwork, which adopts a detection structure based on differential feature pyramids to extract multi-scale text bounding boxes and output detection results including text region coordinates, confidence scores and candidate order labels.

[0089] The detection results are input into the text orientation classification sub-network, which uses a lightweight convolutional structure to determine the orientation angle of each text candidate box and performs orientation rotation correction on the detected text image regions to make the text images be arranged in a uniform orientation.

[0090] The corrected text image region is input into the text recognition sub-network. The convolutional feature extraction structure, sequence modeling structure and character decoding structure complete the text image feature extraction, sequence learning and text recognition operations in sequence. During the character decoding process, a medical terminology dictionary error correction mechanism is introduced. The candidate character sequence in the text recognition result is matched with the preset medical terminology dictionary. When there is a significant difference between the recognition result and the dictionary entry, it is replaced with the most similar standard term.

[0091] The final output is structured claim text information, including the time of visit, the hospital visited, the diagnosis, and the fields for cost items and amount.

[0092] This implementation method constructs an improved PaddleOCR model, introducing mechanisms such as invoice structure localization, differential feature pyramid detection, and orientation classification correction to accurately extract text regions from medical invoices. By optimizing the recognition results through character decoding and medical terminology error correction, it effectively improves the accuracy of image text recognition and the effect of field structure extraction. Especially when facing real-world scenarios such as complex invoice formats and slanted or blurred text, it can still stably output high-quality structured claims text information, providing reliable data support for subsequent clause matching and compensation determination.

[0093] In this embodiment, the step of semantically encoding the structured claims text information to obtain the context semantic representation vector of each claims text content specifically includes:

[0094] The medical time, medical institution, diagnosis, medical items and amount information in the structured claim text are concatenated according to the preset field order to construct a standard input text sequence;

[0095] The input text sequence is fed into a context-enhanced feature encoder for semantic encoding. The context-enhanced feature encoder includes a character-level convolutional feature extraction layer, a bidirectional long short-term memory network modeling layer, and a semantic attention aggregation layer. The semantic attention aggregation layer performs weighted combination of temporal hidden states to generate a fused global contextual semantic representation vector.

[0096] This implementation constructs a context-enhanced feature encoder to perform multi-layer semantic modeling on standard input text sequences. Based on character-level convolutional extraction, it introduces a bidirectional long short-term memory network and a semantic attention aggregation mechanism to achieve global contextual modeling of key information such as consultation time, medical institution, and diagnosis conclusion in the claim text, effectively improving the accuracy and discriminative ability of text semantic representation.

[0097] In this embodiment, the step of inputting the context semantic representation vector and the set of claims judgment criteria into a multi-round rule-enhanced BERT matching network to perform a clause matching task and outputting a clause matching score and a satisfaction judgment result specifically includes:

[0098] The context semantic representation vector and the set of claims judgment criteria are input into a multi-round rule-enhanced BERT matching network to perform a clause matching task and output the clause matching score and the result of the satisfaction status judgment.

[0099] The multi-round rule-enhanced BERT matching network includes a multi-round semantic interaction modeling layer, a structural field alignment mechanism, and a rule label-guided scoring module;

[0100] The multi-round semantic interaction modeling layer is based on a dual-input attention mechanism. In each round, it constructs a cross-attention weight matrix between the context semantic representation vector and the clause field semantic vector, and iteratively updates the matching feature representation in multiple rounds to extract matching relationships at different semantic granularities.

[0101] The structure field alignment mechanism aligns the structure fields in the structured claims text information with the semantic fields in the claims judgment basis set according to the preset field label mapping rules, and constructs a set of field-level semantic alignment vectors.

[0102] The rule-label-guided scoring module inputs the rule label representations of the field-level semantic alignment vector set and the claim judgment basis set, including the claim conditions field, exclusion liability field, and supporting document requirement field, into the scoring function. Based on cosine semantic similarity and rule satisfaction, it calculates the matching score and satisfaction judgment result for each clause. The rule satisfaction is logically judged based on the claim judgment basis set corresponding to the clause field. When the structured claim field content satisfies all the required condition labels, the rule satisfaction is recorded as 1; otherwise, it is recorded as 0.

[0103] This implementation introduces a multi-round rule-enhanced BERT matching network to perform deep matching modeling of the context semantic representation vector and the set of claims judgment criteria. It also integrates field alignment and rule label-guided scoring mechanisms, which not only improves the semantic accuracy of the application judgment of the terms, but also enhances the ability to make structured logical judgments on the compensation conditions and exclusions, effectively supporting the terms adaptation and intelligent decision-making process in highly complex scenarios.

[0104] In this embodiment, the multi-task semantic-driven compensation determination network specifically includes:

[0105] A multi-task semantic-driven compensation determination network is constructed, which includes a compensation determination subnetwork, an amount estimation subnetwork, and a compensation reason label generation subnetwork.

[0106] The structured claims text information, clause matching scores, and satisfaction judgment results are concatenated by field dimension to form a joint semantic feature vector;

[0107] The joint semantic feature vector is input into the compensation judgment sub-network, which adopts a dual-tower semantic classification structure. Each tower consists of two layers of Transformer encoder and feedforward network. By modeling the semantics of the claims text and the clause matching information respectively, and introducing a cross-attention mechanism to establish the coupling relationship between the semantics of compensation and the judgment basis, the prediction result of whether to pay compensation is output.

[0108] The information of the amount-related fields in the joint semantic feature vector is input into the amount estimation sub-network. The amount estimation sub-network sequentially performs feature transformation on the amount field, treatment method, and compensation ratio through a fully connected layer and a ReLU activation function, and outputs the suggested compensation amount prediction value.

[0109] The joint semantic feature vector is input into the compensation reason label generation sub-network. Specifically, the compensation reason label generation sub-network inputs the joint semantic feature vector into a one-layer feedforward fully connected network to perform nonlinear feature transformation and outputs the compensation reason label after being processed by the Sigmoid activation function.

[0110] This implementation method constructs a multi-task semantic-driven claims determination network, which integrates the semantics of claims text and clause matching using a dual-tower structure, thereby improving the accuracy and robustness of claims determination. It introduces a sub-network for amount estimation and cause label generation, enabling intelligent prediction of claims amount and interpretable output of reasons. This enhances the refined processing capability and decision transparency of the intelligent claims system, effectively supporting efficient and reliable claims assistance services.

[0111] In this embodiment, the step of generating a structured claims assistance report based on the prediction result of whether compensation will be paid, the suggested compensation amount, and the compensation reason label specifically includes:

[0112] The prediction of whether compensation will be paid, the suggested compensation amount, and the compensation reason label are integrated to generate a summary of the claims determination.

[0113] Based on the clause matching score and the satisfaction judgment result, generate clause explanation information corresponding to the claim reason label;

[0114] If the claim reason label includes a claim rejection or risk-related label, generate structured risk warning information;

[0115] Based on the preset claims assistance report format template, a structured claims assistance report is constructed, which includes a unique index identifier for the claims application, insurance product terms and conditions information, a summary of the claims determination, a recommended compensation amount, and explanatory information on the terms and conditions.

[0116] This implementation method generates a structured claims assistance report, which unifies and formats the claims prediction results, estimated amounts, and claims reason tags, significantly improving the readability and review efficiency of claims decision output. By combining structured fields and matching criteria to generate a clear semantic chain, it facilitates verification and traceability, effectively assists claims reviewers in quickly understanding the model's judgment logic, and enhances the human-machine collaboration capability and business adaptability of the intelligent claims system.

[0117] Example 1:

[0118] To verify the feasibility of this invention in practice, it was applied to the actual business processes of a large commercial health insurance company's claims service center in a certain region, focusing on the high-frequency claims scenario of outpatient and emergency room expense reimbursement for deployment and pilot testing. This insurance company receives over 2,000 outpatient and emergency room claims daily, containing a large amount of mixed structured and unstructured content, including image receipts, hospital diagnoses, treatment cost lists, and corresponding policy terms. In the traditional manual review process, claims specialists need to manually verify the image information against the policy terms; the process is cumbersome, and accuracy relies heavily on experience, easily leading to missed claims, incorrect claims, or inconsistent claims opinions.

[0119] In the pilot test, an AI-based intelligent claims assistance system proposed in this invention was deployed to process claims receipt images and insurance policy information uploaded by the insured. The system automatically collects image data and policy content, assigning each with a unique index identifier. In the image processing stage, an improved PaddleOCR model is used to perform structural localization, text recognition, and terminology correction on the medical receipt images, extracting key fields such as the hospital visited, visit time, diagnosis results, and expense items and amounts. Subsequently, the extracted structured text is input into a context-enhanced feature encoder to obtain a contextual semantic vector, and combined with the claims judgment criteria set generated by the policy parsing module, a multi-round rule-enhanced BERT matching network is used to complete the policy matching task.

[0120] The matching score and judgment result are then combined and sent to the compensation determination module. The system outputs the compensation result, the suggested compensation amount, and the compensation reason label based on the dual-tower structure and semantic cross-modeling method. The auxiliary report module generates a structured claims assistance report for claims specialists to review and make decisions.

[0121] To further quantify the performance of this invention in real business processes, key indicators such as processing efficiency, compensation accuracy, and manual review burden were recorded before and after the pilot program. The statistical results are shown in Table 1.

[0122] Table 1. Comparison of the effects of the intelligent claims assistance system before and after deployment.

[0123] Indicator Categories Pre-deployment (manual processing) Post-deployment (system assistance of this invention) Increase Average number of applications processed per day 2080 3420 +64.4％ Average processing time 18.6 minutes 7.2 minutes -61.3％ Terms matching accuracy 82.5％ 95.1％ +12.6％ Recommended payout consistency rate 79.3％ 92.4％ +13.1％ Manual review ratio 100％ 36.7％ -63.3％ User complaint rate 4.8％ 1.5％ -3.3％ Report generation time not applicable 1.2 minutes Significant optimization

[0124] As shown in Table 1, the system deployment significantly improved the number of claims processed, processing time, and review efficiency. The average daily application processing volume increased by 64.4%, demonstrating the system's strong concurrent processing capabilities and stability. The average processing time per claim was reduced to 38.7% of the original, greatly improving operational efficiency. Through clause structure parsing and semantic alignment enhancement mechanisms, the clause matching accuracy improved to 95.1%, and the consistency between the suggested payout and expert judgment also increased by 13.1 percentage points, significantly enhancing the system's performance in complex text understanding and semantic reasoning.

[0125] Regarding manual review, the system can directly output whether to pay, the recommended amount, and the reason label, enabling automated initial judgment for a large number of routine cases. Only 36.7% of applications need to enter the manual review process, effectively reducing the workload of claims personnel. In particular, by introducing a structured presentation method with a reason label for payment, the system can clearly provide the basis for the clauses for denying or limiting claims, reducing user doubts and disputes. This has resulted in a decrease in the user complaint rate from 4.8% to 1.5%, significantly improving claims transparency and service satisfaction.

[0126] The system's automatic report generation capability also demonstrates significant advantages. It can automatically summarize the prediction results and matching terms, and output a complete claims support report using standard templates. This greatly reduces the time required for traditional report writing and verification, improves case closure efficiency, and shows that the system has good adaptability and practical value in actual business operations.

[0127] This embodiment effectively verifies the intelligent processing capabilities of the AI-based smart claims assistance system proposed in this invention in the entire process of image recognition, semantic matching, clause reasoning and report generation, significantly improving claims efficiency, accuracy and user experience, and has good engineering feasibility and commercial application prospects.

[0128] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. An intelligent claims assistance system based on artificial intelligence, characterized in that, Includes the following modules: The data acquisition module is used to collect the claim image data submitted by the insured and the corresponding insurance product terms and conditions, and to assign a unique index identifier to each claim application. The terms and conditions parsing module is used to perform structured parsing of the terms and conditions of insurance products and generate a set of criteria for claims judgment. The text recognition module is used for text recognition and outputs structured claims text information; The semantic encoding module is used to perform context-enhanced feature encoding on structured claims text information to obtain the context semantic representation vector of each claims text content; The clause matching module is used to perform clause matching tasks using a multi-round rule-enhanced BERT matching network, and outputs clause matching scores and satisfaction determination results, specifically including: The context semantic representation vector and the set of claims judgment criteria are input into a multi-round rule-enhanced BERT matching network to perform a clause matching task and output the clause matching score and the result of the satisfaction status judgment. The multi-round rule-enhanced BERT matching network includes a multi-round semantic interaction modeling layer, a structural field alignment mechanism, and a rule label-guided scoring module; The multi-round semantic interaction modeling layer is based on a dual-input attention mechanism. In each round, it constructs a cross-attention weight matrix between the context semantic representation vector and the clause field semantic vector, and iteratively updates the matching feature representation in multiple rounds to extract matching relationships at different semantic granularities. The structure field alignment mechanism aligns the structure fields in the structured claims text information with the semantic fields in the claims judgment basis set according to the preset field label mapping rules, and constructs a set of field-level semantic alignment vectors. The rule-label-guided scoring module inputs the rule label representations of the field-level semantic alignment vector set and the claim judgment basis set, including the claim conditions field, exclusion liability field, and supporting document requirement field, into the scoring function. Based on cosine semantic similarity and rule satisfaction, it calculates the matching score and satisfaction judgment result for each clause. The rule satisfaction is logically judged based on the claim judgment basis set corresponding to the clause field. When the structured claim field content satisfies all the mandatory condition labels, the rule satisfaction is recorded as 1; otherwise, it is recorded as 0. The compensation determination module utilizes a multi-task semantic-driven compensation determination network to perform compensation determination and amount estimation tasks, outputting a prediction result of whether compensation is payable, a suggested compensation amount, and a compensation reason label. The multi-task semantic-driven compensation determination network specifically includes: A multi-task semantic-driven compensation determination network is constructed, which includes a compensation determination subnetwork, an amount estimation subnetwork, and a compensation reason label generation subnetwork. The structured claims text information, clause matching scores, and satisfaction judgment results are concatenated by field dimension to form a joint semantic feature vector; The joint semantic feature vector is input into the compensation judgment sub-network, which adopts a dual-tower semantic classification structure. Each tower consists of two layers of Transformer encoder and feedforward network. By modeling the semantics of the claims text and the clause matching information respectively, and introducing a cross-attention mechanism to establish the coupling relationship between the semantics of compensation and the judgment basis, the prediction result of whether to pay compensation is output. The information of the amount-related fields in the joint semantic feature vector is input into the amount estimation sub-network. The amount estimation sub-network sequentially performs feature transformation on the amount field, treatment method, and compensation ratio through a fully connected layer and a ReLU activation function, and outputs the suggested compensation amount prediction value. The joint semantic feature vector is input into the compensation reason label generation subnetwork. Specifically, the compensation reason label generation subnetwork inputs the joint semantic feature vector into a one-layer feedforward fully connected network to perform nonlinear feature transformation and outputs the compensation reason label after being processed by the Sigmoid activation function. The auxiliary report generation module is used to generate structured claims auxiliary reports.

2. The intelligent claims assistance system based on artificial intelligence according to claim 1, characterized in that, The modules are connected in the following way: Collect the image data of the claim submitted by the insured and the corresponding insurance product terms, and assign a unique index identifier to each claim application; The terms and conditions of insurance products are analyzed in a structured manner to generate a set of criteria for claims judgment; The claim image data is input into the improved PaddleOCR model for text recognition to obtain structured claim text information; The structured claims text information is semantically encoded to obtain the contextual semantic representation vector of each claims text content; The context semantic representation vector and the set of claims judgment criteria are input into a multi-round rule-enhanced BERT matching network to perform a clause matching task and output the clause matching score and the result of the satisfaction status judgment. The structured claims text information, clause matching score, and satisfaction judgment result are jointly input into a multi-task semantic-driven claims determination network to perform claims determination and amount estimation tasks, and output the prediction result of whether to pay, the suggested payout amount, and the claim reason label. A structured claims assistance report is generated based on the predicted payout amount, the suggested payout amount, and the payout reason label.

3. The intelligent claims assistance system based on artificial intelligence according to claim 2, characterized in that, The structured parsing of insurance product terms and conditions to generate a set of claims judgment criteria specifically includes: The content of insurance product terms is structured by combining rule template parsing and syntactic dependency analysis. The rule template parsing extracts fields such as compensation liability, triggering conditions, compensation ratio, exclusions, and supporting documentation requirements by matching preset insurance term language patterns. If the rule template cannot be matched, the syntactic dependency analysis method is used. Based on verb center structure recognition and grammatical dependency path calculation, the subject-verb-object structure and modifiers related to the compensation conditions are automatically extracted from the content of the insurance product terms and conditions to complete the field extraction results. Each extracted field is assigned a field identifier and standard semantic tag according to a preset field classification rule to construct a set of clause structure fields. Based on the set of clause structure fields, the fields are grouped according to the similarity of compensation type, the consistency of the insured subject, and the logical dependency of the clause. The semantic association matrix between the fields is constructed by calculating the cosine similarity of sentence vectors, and a set of claims judgment criteria with structural organization and semantic mapping relationship is generated.

4. The intelligent claims assistance system based on artificial intelligence according to claim 2, characterized in that, The process of inputting claim image data into the improved PaddleOCR model for text recognition to obtain structured claim text information specifically includes: An improved PaddleOCR model is constructed, which includes a ticket structure localization module, a text detection subnetwork, a text orientation classification subnetwork, and a character recognition subnetwork. The input claim image data is preprocessed through the bill structure localization module. Specifically, this includes performing page structure segmentation, locating the main area of ​​the medical bill, cropping interfering content at the image edges, and extracting suspected information block areas as candidate areas for structured recognition. The structured recognition candidate regions are input into the text detection subnetwork, which adopts a detection structure based on differential feature pyramids to extract multi-scale text bounding boxes and output detection results including text region coordinates, confidence scores and candidate order labels. The detection results are input into the text orientation classification sub-network, which uses a lightweight convolutional structure to determine the orientation angle of each text candidate box and performs orientation rotation correction on the detected text image regions to make the text images be arranged in a uniform orientation. The corrected text image region is input into the text recognition sub-network. The convolutional feature extraction structure, sequence modeling structure and character decoding structure complete the text image feature extraction, sequence learning and text recognition operations in sequence. During the character decoding process, a medical terminology dictionary error correction mechanism is introduced. The candidate character sequence in the text recognition result is matched with the preset medical terminology dictionary. When there is a significant difference between the recognition result and the dictionary entry, it is replaced with the most similar standard term. The final output is structured claim text information, including the time of visit, the hospital visited, the diagnosis, and the fields for cost items and amount.

5. The intelligent claims assistance system based on artificial intelligence according to claim 2, characterized in that, The step of semantically encoding the structured claims text information to obtain the contextual semantic representation vector of each claims text content specifically includes: The medical time, medical institution, diagnosis, medical items and amount information in the structured claim text are concatenated according to the preset field order to construct a standard input text sequence; The input text sequence is fed into a context-enhanced feature encoder for semantic encoding. The context-enhanced feature encoder includes a character-level convolutional feature extraction layer, a bidirectional long short-term memory network modeling layer, and a semantic attention aggregation layer. The semantic attention aggregation layer performs weighted combination of temporal hidden states to generate a fused global contextual semantic representation vector.

6. The intelligent claims assistance system based on artificial intelligence according to claim 2, characterized in that, The process of generating a structured claims assistance report based on the prediction of whether compensation will be paid, the suggested compensation amount, and the claim reason label specifically includes: The prediction of whether compensation will be paid, the suggested compensation amount, and the compensation reason label are integrated to generate a summary of the claims determination. Based on the clause matching score and the satisfaction judgment result, generate clause explanation information corresponding to the claim reason label; If the claim reason label includes a claim rejection or risk-related label, generate structured risk warning information; Based on the preset claims assistance report format template, a structured claims assistance report is constructed, which includes a unique index identifier for the claims application, insurance product terms and conditions information, a summary of the claims determination, a recommended compensation amount, and explanatory information on the terms and conditions.

Images