An insurance claim settlement method and system based on OCR and large language model

Abstract

The application relates to the technical field of artificial intelligence and financial insurance technology, and discloses an insurance claim settlement method and system based on OCR and a large language model, the method comprising the following steps: performing semantic analysis on a to-be-processed claim settlement file to obtain the semantic role and the preliminary confidence of each text segment in the claim settlement file; performing secondary identification on a low-confidence text segment according to the preliminary confidence to obtain optimized text data; performing double semantic consistency verification on the key claim settlement information extracted from the optimized text data, and generating a final audit result of the claim settlement file according to the verification result and the semantic role. The application can improve the efficiency of an insurance claim settlement method based on OCR and a large language model.

Description

Technical Field

[0001] This invention relates to the field of interdisciplinary technology of artificial intelligence and financial insurance technology, and in particular to an insurance claims method and system based on OCR and large language model. Background Technology

[0002] Traditional insurance claims processing relies heavily on optical character recognition (OCR) technology for text extraction. However, existing OCR methods struggle to guarantee both character accuracy and semantic integrity when dealing with blurry stamps, handwriting, low-resolution scans, or complex table structures. These errors directly impact downstream review processes, particularly misreading crucial amounts, dates, or diagnostic information, leading to distortion of core claims data. Traditional post-processing mechanisms, employing only rule engines and maximum forward matching algorithms, lack the deep semantic understanding and logical reasoning capabilities of large language models, failing to automatically detect and correct contextual inconsistencies.

[0003] Current claims review processes largely rely on field extraction and rigid rule matching, lacking in-depth verification of the overall semantic roles of documents and cross-field logical consistency. Large language models possess semantic understanding, logical reasoning, and text generation capabilities in the insurance field, compensating for the shortcomings of traditional algorithms. However, current technologies have not deeply integrated OCR with insurance-specific fine-tuned large language models, failing to establish a standardized collaborative process. This leads to errors in determining the relationship between main and supplementary insurance items, requiring manual review of numerous claims conclusions on a case-by-case basis, limiting overall processing efficiency and accuracy, and hindering the improvement of automatic approval rates. Therefore, how to integrate OCR and large language models to improve insurance claims efficiency has become an urgent problem to be solved. Summary of the Invention

[0004] This invention provides an insurance claims method and system based on OCR and a large language model to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, this invention provides an insurance claims method based on OCR and a large language model, comprising: The insurance domain fine-tuned large language model is invoked to perform semantic parsing on the claims documents to be processed, so as to obtain the semantic role and preliminary confidence of each text segment in the claims documents; Based on the initial confidence level, targeted recognition prompts are generated using a large language model to perform secondary recognition on low-confidence text fragments, resulting in optimized text data. By performing dual semantic logic reasoning verification through a large language model, the key claims information extracted from the optimized text data is subjected to dual semantic consistency verification, and the final review result of the claims document is generated based on the verification result and the semantic role.

[0006] In a preferred embodiment, before semantic parsing of the claim document to be processed, the following steps are included: Denoising was performed on the collected claim document images; The initial text dataset is obtained by identifying all recognizable text regions in the denoised data. The image layout information set is input into a large language model as layout prior knowledge for semantic parsing.

[0007] In a preferred embodiment, the step of calling the insurance domain fine-tuned large language model to perform semantic parsing on the claims document to be processed, obtaining the semantic role and preliminary confidence level of each text segment in the claims document, includes: Based on the insurance semantic annotation prompt words of the large language model, combined with the positional adjacency relationship and font attributes in the image layout feature parameters, the contextual semantic role pre-annotation is performed on each initial text segment in the initial text dataset to obtain the candidate semantic role of the text segment. By using the logical coherence reasoning of a large language model, combined with the logical coherence between the text fragment and the candidate semantic roles of adjacent text fragments, and the preset document structure corresponding to the relative position of the text fragment in the claims document, a preliminary confidence level of the text fragment is generated.

[0008] In a preferred embodiment, the step of generating targeted recognition prompts using a large language model based on the initial confidence level to perform secondary recognition on low-confidence text fragments to obtain optimized text data includes: Text fragments with an initial confidence level lower than a preset matching threshold are marked as low-confidence text fragments, and the location area information of the low-confidence text fragments in the claims document is extracted; By generating targeted fine-grained recognition prompts containing character type constraints and word candidate ranges through a large language model, targeted fine-grained recognition instructions based on location region information are obtained. The image region defined by the location region information is subjected to secondary recognition according to the directional fine recognition instruction, and the returned secondary recognition candidate text is received.

[0009] In a preferred embodiment, after receiving the returned secondary recognition candidate text, the method further includes: The secondary recognition candidate text is merged with the non-low confidence text fragments in the initial text fragment to obtain the candidate optimized text dataset; By combining the semantic fluency assessment of the large language model with the traditional semantic coherence formula, the candidate texts are evaluated for semantic coherence, and the candidate text with the highest score is selected as the final replacement text. The corresponding low-confidence text fragment in the initial text fragment is replaced with the final replacement text to obtain optimized text data.

[0010] In a preferred embodiment, the semantic coherence score is calculated using the following formula: ; In the formula, Score the semantic coherence. The number of elements in the intersection of the vocabulary set of the candidate text for secondary recognition and the vocabulary set of the adjacent text on the left. This refers to the number of elements in the intersection of the vocabulary set of the candidate text for secondary recognition and the vocabulary set of the adjacent text on the right. The number of elements in the vocabulary set in the candidate text for secondary recognition. This represents the number of elements in the vocabulary set of the text adjacent to the left. This represents the number of elements in the vocabulary set of the adjacent text on the right.

[0011] In a preferred embodiment, the step of performing dual semantic consistency verification on the key claims information extracted from the optimized text data includes: By using a large language model, key claim information corresponding to preset claim fields is extracted from optimized text data to obtain structured claim information objects; Through logical reasoning using a large language model, we examine the numerical logical relationships and professional terminology associations between different field values ​​in a structured claims information object. When a contradiction is found in the numerical logic relationship or the terminology association relationship, the corresponding field is marked as a field to be corrected, and an internal consistency check result for the field to be corrected is generated.

[0012] In a preferred embodiment, the step of performing dual semantic consistency verification on the key claims information extracted from the optimized text data further includes: The fields to be corrected and the fields that have passed the verification in the structured claims information object are obtained and used as a set to be cross-validated. By extracting entities and events from a large language model, we determine whether the entities and events in the optimized text data support the set to be cross-validated. The cross-validation results are generated based on the support judgment results, wherein the cross-validation results include suggestions for correcting the field to be corrected or instructions to maintain the original field.

[0013] In a preferred embodiment, generating the final review result of the claim document based on the verification result and the semantic role includes: Based on the semantic role of the text fragment, locate the original semantic position of each field in the cross-validation result in the claims file and the corresponding insurance clause item; By matching insurance clause rules using a large language model, rule matching conclusions are generated for confirmed and corrected fields. By integrating the rule matching results with the semantic role information of the fields requiring manual review, the final review result of the claim document is generated. The final review result includes automatically approved claim items, automatically rejected claim items, and disputed items pending manual review.

[0014] To address the aforementioned problems, this invention also provides an insurance claims system based on OCR and a large language model, the system comprising: The semantic parsing and confidence assessment module is used to call the fine-tuned large language model in the insurance field to perform semantic parsing on the claims documents to be processed, and obtain the semantic role and preliminary confidence of each text segment in the claims documents; The low-confidence text orientation optimization module is used to generate orientation recognition prompts through a large language model based on the initial confidence level to perform secondary recognition on low-confidence text fragments and obtain optimized text data. The dual semantic verification and review generation module is used to perform dual semantic logical reasoning verification through a large language model, perform dual semantic consistency verification on key claims information extracted from optimized text data, and generate the final review result of the claims document based on the verification result and the semantic role.

[0015] Compared with the prior art, the present invention has the following beneficial effects: 1. By assigning a semantic role to each text fragment and calculating an initial confidence level, this invention can accurately locate low-quality recognition areas caused by ambiguity, similar shapes, or occlusion, and generate targeted, refined recognition instructions with character type constraints and a range of candidate words based on contextual semantic relationships. After secondary recognition, the semantic coherence score between the candidate text and adjacent fragments is further calculated to select the optimal replacement content, thereby significantly improving the completeness of extracting key information such as amounts and dates from complex documents, and providing a more reliable data foundation for claims processing.

[0016] 2. After obtaining the optimized text, this invention performs dual semantic consistency checks on the extracted key claims information. First, it examines the numerical logical relationships and professional knowledge associations between different fields. Then, it cross-validates the verification results by analyzing entities and events throughout the text, automatically identifying and correcting logical contradictions. Simultaneously, based on the semantic role of each field, it locates its semantic position in the original file and links it to the corresponding insurance clause rules, ultimately generating a structured review conclusion that clearly distinguishes between automatically approved, automatically rejected, and items requiring manual review. This directly improves the accuracy of claims decisions and the rate of automation. Attached Figure Description

[0017] Figure 1 This is a flowchart illustrating an insurance claims method based on OCR and a large language model, provided as an embodiment of the present invention. Figure 2 A functional module diagram of an insurance claims system based on OCR and a large language model, provided in an embodiment of the present invention; The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0018] It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0019] This application provides an insurance claims method based on OCR and a large language model. The executing entity of this OCR and large language model-based insurance claims method includes, but is not limited to, at least one of the following electronic devices that can be configured to execute the method provided in this application: a server, a terminal, etc. In other words, the OCR and large language model-based insurance claims method can be executed by software or hardware installed on a terminal device or a server device. The server includes, but is not limited to, a single server, a server cluster, a cloud server, or a cloud server cluster. The server can be an independent server or a cloud server providing basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, content delivery networks (CDNs), and big data and artificial intelligence platforms.

[0020] Reference Figure 1 The diagram shown illustrates a flowchart of an insurance claims method based on OCR and a large language model, according to an embodiment of the present invention. In this embodiment, the insurance claims method based on OCR and a large language model includes: The insurance domain fine-tuned large language model is invoked to perform semantic parsing on the claims documents to be processed, so as to obtain the semantic role and preliminary confidence level of each text segment in the claims documents.

[0021] In this embodiment of the invention, before semantic parsing of the claim document to be processed, the following steps are included: Denoising was performed on the collected claim document images; The initial text dataset is obtained by identifying all recognizable text regions in the denoised data. The image layout information set is input into a large language model as layout prior knowledge for semantic parsing.

[0022] The process involves calling a fine-tuned large language model in the insurance domain to perform semantic parsing on the claims file to obtain the semantic role and preliminary confidence level of each text segment in the claims file, including: Based on the insurance semantic annotation prompt words of the large language model, combined with the positional adjacency relationship and font attributes in the image layout feature parameters, the contextual semantic role pre-annotation is performed on each initial text segment in the initial text dataset to obtain the candidate semantic role of the text segment. By using the logical coherence reasoning of a large language model, combined with the logical coherence between the text fragment and the candidate semantic roles of adjacent text fragments, and the preset document structure corresponding to the relative position of the text fragment in the claims document, a preliminary confidence level of the text fragment is generated.

[0023] When denoising the collected claims document images, a nonlocal mean denoising algorithm is used to traverse each pixel in the image, select a neighborhood window centered on the pixel, search for other neighborhood windows in the image with similar textures to the neighborhood window, assign different weighted average weights to the corresponding center pixels according to the degree of similarity, and replace the original pixel values ​​with the weighted average pixel values, so as to maintain the edge clarity of the text strokes while smoothing the background noise and scanning noise.

[0024] When identifying text fragments corresponding to all recognizable text regions in the denoised data, the pixel connected component analysis algorithm is used to detect the bounding rectangle of the outline of each character in the image. Adjacent character bounding rectangles with horizontal spacing less than the average width of the character and vertical overlap of more than half the height are merged into text line rectangles. Then, adjacent text line rectangles that overlap vertically and have horizontal alignment are aggregated into text regions. Optical character recognition is performed on each text region to extract the text content string. All extracted text content strings constitute the initial text dataset.

[0025] When recording the position, font size, and adjacency of text fragments in the claim document image, the coordinates of the four vertices of the minimum bounding rectangle of each text area in the image coordinate system are saved as the position information of the text fragment. The median of the outline height of all characters in the text area is multiplied by the image resolution normalization coefficient to obtain the font size of the text fragment. For each text fragment, the horizontal overlap ratio and vertical distance between its bounding rectangle and the bounding rectangles of other text fragments are calculated. The upper and lower fragments with the horizontal overlap ratio exceeding a set threshold and the smallest vertical distance are recorded as vertically adjacent. The left and right fragments with the vertical overlap ratio exceeding a set threshold and the smallest horizontal distance are recorded as horizontally adjacent. The coordinate values, font sizes, and adjacency of all text fragments constitute the image layout information set of the claim document.

[0026] When performing contextual semantic role pre-annotation based on the positional adjacency relationship and font attribute in the image layout feature parameters, a pre-constructed semantic role category set for claims documents is loaded. This category set includes role tags such as insured's name, ID number, insurance type name, insurance amount, date of incident, diagnosis conclusion, total cost, and reimbursement ratio. Multiple text fragments belonging to the same cell or the same fill area in the same table are merged into semantic nodes based on adjacency relationships. The maximum font size inside each semantic node is extracted as the node's font attribute. Based on the node's font attribute and relative position, semantic nodes are divided into title nodes, table header nodes, field name nodes, and field value nodes. If there is a punctuation mark or a field name node ending with a colon in the horizontal adjacency relationship of a semantic node, the field value node horizontally adjacent to it is pre-annotated as the semantic role category corresponding to that field name node. If the semantic node is located inside the table, the semantic role category corresponding to the table header node is passed down the column direction to the corresponding field value node in each row for pre-annotation. After pre-annotation, each text fragment obtains a candidate semantic role.

[0027] When generating the initial confidence score of a text fragment, the preset document structure of the claims document is abstracted into a semantic role distribution template and order constraints on the page. The distribution template specifies that the header area contains the insured's information semantic role, the middle area contains the policy and accident information semantic role, and the tail area contains the expense and signature information semantic role. The order constraints specify that there is a constraint between the insured amount and the compensation ratio that the amount is greater than the compensation amount and the values ​​have a multiple relationship. For each text fragment, it is examined whether the allowed area corresponding to its candidate semantic role in the distribution template matches the position information of the current text fragment. At the same time, it is examined whether the candidate semantic role of the fragment satisfies the order constraints and reasonable matching relationship with the candidate semantic roles of the left and right adjacent fragments in the reading order from left to right and front to back. If the co-occurrence of adjacent roles is in the preset compliance role binary library, the coherence is established. The position matching result and the logical coherence result are combined for judgment. If the position matches and the adjacent roles are coherent, the confidence score is set to a high value. If only one of them is satisfied, the confidence score is set to a medium value. If neither of them is satisfied, the confidence score is set to a low value. In this way, the initial confidence score of each text fragment is generated.

[0028] The beneficial effects are as follows: By employing a nonlocal mean denoising algorithm to denoise the claims document images, and using a weighted average to replace pixel values ​​through neighborhood windows with similar textures within the image, the clarity of text stroke edges is fully preserved while eliminating background noise and scanning noise, enabling subsequent text region recognition to be performed on clean images. A pixel connectivity analysis algorithm is used to detect the bounding rectangles of character outlines and merge them layer by layer into text lines and text regions based on spacing and overlap. Optical character recognition is performed on each text region to extract the text content string, enabling the systematic collection of a structurally complete initial text dataset from the document images. Recording the position coordinates, font size, and adjacency relationships based on horizontal and vertical overlap ratios of text fragments forms an image layout information set, providing a precise spatial reference system for semantic analysis. A semantic role category set is loaded, and text fragments are merged into semantic nodes based on adjacency relationships. The maximum font value is used to distinguish the types of title nodes, table header nodes, field name nodes, and field value nodes. Field value nodes are pre-labeled with corresponding semantic role categories through punctuation and table header propagation mechanisms to obtain candidate semantic roles, automatically completing the semantic qualitative analysis of various types of information in the documents. The location matching results of the integrated regional distribution template and the logical coherence results of the compliance role binary library are used to generate an initial confidence score, which accurately distinguishes between text fragments with high confidence, medium confidence and low confidence, providing a reliable basis for subsequent targeted fine processing of only low-quality identification results.

[0029] Based on the initial confidence level, targeted recognition prompts are generated using a large language model to perform secondary recognition on low-confidence text fragments, resulting in optimized text data.

[0030] In this embodiment of the invention, the step of generating targeted recognition prompts using a large language model based on the initial confidence level to perform secondary recognition on low-confidence text fragments to obtain optimized text data includes: Text fragments with an initial confidence level lower than a preset matching threshold are marked as low-confidence text fragments, and the location area information of the low-confidence text fragments in the claims document is extracted; By generating targeted fine-grained recognition prompts containing character type constraints and word candidate ranges through a large language model, targeted fine-grained recognition instructions based on location region information are obtained. The image region defined by the location region information is subjected to secondary recognition according to the directional fine recognition instruction, and the returned secondary recognition candidate text is received.

[0031] After receiving the returned secondary recognition candidate text, the process also includes: The secondary recognition candidate text is merged with the non-low confidence text fragments in the initial text fragment to obtain the candidate optimized text dataset; By combining the semantic fluency assessment of the large language model with the traditional semantic coherence formula, the candidate texts are evaluated for semantic coherence, and the candidate text with the highest score is selected as the final replacement text. The corresponding low-confidence text fragment in the initial text fragment is replaced with the final replacement text to obtain optimized text data.

[0032] The formula for calculating the semantic coherence score is as follows: ; In the formula, Score the semantic coherence. The number of elements in the intersection of the vocabulary set of the candidate text for secondary recognition and the vocabulary set of the adjacent text on the left. This refers to the number of elements in the intersection of the vocabulary set of the candidate text for secondary recognition and the vocabulary set of the adjacent text on the right. The number of elements in the vocabulary set in the candidate text for secondary recognition. This represents the number of elements in the vocabulary set of the text adjacent to the left. This represents the number of elements in the vocabulary set of the adjacent text on the right.

[0033] When marking text fragments with an initial confidence level lower than a preset matching threshold as low-confidence text fragments, the initial confidence level value generated for each text fragment is read one by one, and this value is compared with the preset matching threshold. All text fragments with an initial confidence level less than the threshold are classified into the low-confidence set. The location coordinate data corresponding to each text fragment in the low-confidence set is extracted from the image layout information set as the location area information of the low-confidence text fragment in the claims document.

[0034] When generating a targeted and refined recognition instruction based on the semantic roles of low-confidence text fragments and adjacent text fragments, the process reads the candidate semantic roles carried by the low-confidence text fragment itself, as well as the candidate semantic roles of the left and right adjacent text fragments obtained through the adjacency index. Based on these three semantic roles, the contextual context is determined. The set of character types allowed in this context is retrieved from the pre-built semantic role-character type mapping table as a constraint condition for character type recognition. At the same time, the three semantic roles are used as search keys to retrieve a list of words belonging to this context from the claims domain terminology database as a context-based word candidate range. The location area information, character type constraints, and word candidate range are then assembled into a targeted and refined recognition instruction.

[0035] When performing secondary recognition on an image region defined by location information according to the directional fine recognition instruction, the local image block in the original denoised image pointed to by the location information is parsed from the directional fine recognition instruction. This local image block is sent to the optical character recognition engine to load character type constraints to limit the character category of the recognition output. After character segmentation, the optical character recognition engine matches each segmented character image with the standard character template within the candidate character category and outputs a character candidate sequence. Then, the character candidate sequences are combined into a string and compared with the words in the word candidate range for filtering. Only the strings that appear in the word candidate range are retained as secondary recognition candidate text and returned.

[0036] When merging the secondary recognition candidate text with the non-low-confidence text segments in the initial text segment, all entries marked as low-confidence text segments are removed from the initial text dataset. The remaining text segments are kept as a set of non-low-confidence text segments. The secondary recognition candidate text lists corresponding to each low-confidence text segment are attached one by one to the text region positions they originally belonged to in the initial text dataset. Together with the non-low-confidence text segments, they form a candidate optimized text dataset. The position of each low-confidence text segment in the candidate optimized text dataset retains the original text segment information and its corresponding multiple secondary recognition candidate texts.

[0037] When evaluating the semantic coherence of secondary recognition candidate texts corresponding to each low-confidence text segment in the candidate optimized text dataset, for each secondary recognition candidate text of each low-confidence text segment, the candidate text is segmented to obtain a candidate vocabulary set. At the same time, the content of the left adjacent text segment is obtained through the adjacency relation index and segmented to obtain a left vocabulary set, and the content of the right adjacent text segment is obtained through the right vocabulary set. The number of intersection words between the candidate vocabulary set and the left vocabulary set and the candidate vocabulary set and the right vocabulary set are counted respectively. The two intersection word counts are added together to obtain the total number of context-overlapping words. Then, the number of elements in the candidate vocabulary set, the number of elements in the left vocabulary set, and the number of elements in the right vocabulary set are added together to obtain the total vocabulary size. The result of dividing the total number of context-overlapping words by the total vocabulary size is used as the semantic coherence score of the secondary recognition candidate text.

[0038] When selecting the candidate text with the highest semantic coherence score as the final replacement text and replacing the corresponding low-confidence text fragment in the initial text fragment with the final replacement text, the semantic coherence scores of all secondary recognition candidate texts corresponding to the same low-confidence text fragment are sorted in descending order. The secondary recognition candidate text with the highest score is extracted as the final replacement text at that position. The original record of the low-confidence text fragment is located in the initial text dataset, and the text content field of the original record is updated with the string content of the final replacement text. After performing this replacement operation on all low-confidence text fragments one by one, the initial text dataset becomes the optimized text data.

[0039] The vocabulary set for the secondary recognition candidate text in the formula is obtained by performing word segmentation on the candidate text. During word segmentation, a maximum forward matching segmentation method based on a claims domain terminology database is used to scan the candidate text string from left to right. Whenever the longest term in the domain terminology database is matched, a word is segmented and recorded until the entire candidate text is segmented. All the segmented words constitute the vocabulary set of the secondary recognition candidate text. The vocabulary set of the left adjacent text is obtained by performing the same maximum forward matching segmentation operation on the content of the left adjacent text segment, and the vocabulary set of the right adjacent text is obtained by performing the same maximum forward matching segmentation operation on the content of the right adjacent text segment. The number of intersection elements between the vocabulary set of the secondary recognition candidate text and the vocabulary set of the left adjacent text is counted by traversing each word in the candidate vocabulary set and checking whether each word also exists in the left vocabulary set. Each time a word is found to exist in both sets, the intersection count is incremented by one. The accumulated value after the traversal is completed is the first addend in the numerator. The number of elements in the intersection of the word set of the candidate text in the secondary recognition and the word set of the adjacent text on the right is calculated by traversing the candidate word set and querying the word set on the right in the same way, and then summing the results to obtain the second addend in the numerator. In the denominator, the number of elements in the word set of the candidate text in the secondary recognition is the total number of words contained in the candidate word set after word segmentation; the number of elements in the word set of the adjacent text on the left is the total number of words contained in the left word set; and the number of elements in the word set of the adjacent text on the right is the total number of words contained in the right word set. These three are obtained by directly reading the number of elements in their respective sets.

[0040] The formula means that the semantic coherence score measures the semantic fluency of the candidate text after it is embedded in the current context by quantifying the degree of overlap between the candidate text and its left and right adjacent texts at the lexical level. When calculating the score, the sum of the number of words shared by the candidate text and the left adjacent text and the number of words shared by the candidate text and the right adjacent text is taken as the total number of words with contextual overlap. The sum of the total number of words in the candidate text itself, the total number of words in the left adjacent text, and the total number of words in the right adjacent text is taken as the total number of words used in the evaluation. The ratio obtained by dividing the total number of words with contextual overlap by the total number of words reflects the degree to which the wording of the candidate text is consistent with the wording of the immediate context. The higher the ratio, the more words the candidate text shares with the left and right adjacent texts and the more natural the semantic connection. The lower the ratio, the fewer words the candidate text shares with the left and right adjacent texts and the more awkward the semantic connection.

[0041] The formula trend shows that when the number of shared words between the candidate text and its left and right adjacent texts increases while the number of words in the candidate text itself remains constant, the numerator increases while only the increase in the denominator corresponding to the numerator increases, but the overall increase in the denominator is less than the increase in the numerator, resulting in a higher score. When the number of words in the candidate text itself increases, but the newly added words in the candidate text do not appear in either of the left or right adjacent texts, the denominator increases while the numerator remains unchanged, resulting in a lower score. When the number of words in either the left or right adjacent text increases, but the newly added words do not overlap with the candidate text, the denominator increases while the numerator remains unchanged, resulting in a lower score. The degree of overlap between the candidate text's word set and the word sets of its left and right adjacent texts directly determines the score. The proportion of the overlapping words covering the candidate text's own words and the proportion of the overlapping words covering the adjacent texts both affect the score; the higher these two proportions, the closer the score is to the upper limit of the value range.

[0042] The beneficial effects are as follows: By accurately marking text fragments with initial confidence levels below a preset matching threshold as low-confidence text fragments and extracting their location information, subsequent processing resources are focused solely on local areas where recognition quality is questionable. The semantic roles of low-confidence text fragments, together with those of adjacent text fragments, are used to lock in the context. Character type constraints are retrieved from a semantic role-character type mapping table, and a vocabulary candidate range is obtained from a claims domain terminology database. This generates targeted, fine-grained recognition instructions with both character-level and vocabulary-level constraints, significantly narrowing the search space for secondary recognition. The optical character recognition engine performs template matching under character type constraints, then compares and filters the combined strings with the vocabulary candidate range, retaining only legal words within the domain as secondary recognition candidate texts, filtering out meaningless recognition noise at the source. By merging secondary recognition candidate texts with non-low-confidence text fragments, a candidate optimized text dataset is constructed, fully preserving the original structure of the text region while embedding multiple candidate results. Each candidate text identified in the secondary recognition process is segmented and compared with the vocabulary sets of its left and right adjacent texts one by one. The number of overlapping words is counted, and a semantic coherence score is calculated to quantitatively select the candidate text with the most natural contextual connection. Finally, the candidate text with the highest semantic coherence score is extracted to replace the original low-confidence content, resulting in optimized text data that is semantically complete and contextually fluent.

[0043] By performing dual semantic logic reasoning verification through a large language model, the key claims information extracted from the optimized text data is subjected to dual semantic consistency verification, and the final review result of the claims document is generated based on the verification result and the semantic role.

[0044] In this embodiment of the invention, the step of performing dual semantic consistency verification on the key claims information extracted from the optimized text data includes: By using a large language model, key claim information corresponding to preset claim fields is extracted from optimized text data to obtain structured claim information objects; Through logical reasoning using a large language model, we examine the numerical logical relationships and professional terminology associations between different field values ​​in a structured claims information object. When a contradiction is found in the numerical logic relationship or the terminology association relationship, the corresponding field is marked as a field to be corrected, and an internal consistency check result for the field to be corrected is generated.

[0045] The dual semantic consistency verification of key claims information extracted from optimized text data also includes: The fields to be corrected and the fields that have passed the verification in the structured claims information object are obtained and used as a set to be cross-validated. By extracting entities and events from a large language model, we determine whether the entities and events in the optimized text data support the set to be cross-validated. The cross-validation results are generated based on the support judgment results, wherein the cross-validation results include suggestions for correcting the field to be corrected or instructions to maintain the original field.

[0046] The step of generating the final review result of the claim document based on the verification result and the semantic role includes: Based on the semantic role of the text fragment, locate the original semantic position of each field in the cross-validation result in the claims file and the corresponding insurance clause item; By matching insurance clause rules using a large language model, rule matching conclusions are generated for confirmed and corrected fields. By integrating the rule matching results with the semantic role information of the fields requiring manual review, the final review result of the claim document is generated. The final review result includes automatically approved claim items, automatically rejected claim items, and disputed items pending manual review.

[0047] When extracting key claims information corresponding to preset claims fields from optimized text data, a list of preset claims field names is loaded. This list includes field names such as insured's name, ID number, insurance type, insurance amount, date of incident, diagnosis conclusion, total cost, and reimbursement ratio. For each field name in the list, field value nodes that match the candidate semantic roles of the field name are retrieved from all text fragments in the optimized text data. The field name is used as the key, and the text content of the retrieved field value node is used as the value to store in a key-value pair structure. After all key-value pairs are combined, a structured claims information object is formed.

[0048] When verifying the numerical logical relationships and professional terminology associations between different field values ​​in a structured claims information object, a predefined set of numerical logical relationship rules and a set of professional terminology association rules are loaded. The numerical logical relationship rule set includes rules such as the total cost should equal the sum of all sub-costs, the compensation amount should equal the insured amount multiplied by the compensation ratio and not exceed the insured amount, and the date of the incident should be earlier than the current date and not earlier than the policy effective date. The professional terminology association rule set includes rules such as the disease name in the diagnosis conclusion should match the coverage of the insurance type name, and the weekday attribute of the incident date should be consistent with the time logic of the accident description. Each rule in the rule set is read one by one, and the fields involved in the rule are retrieved from the structured claims information object and compared and verified. The verification results of each rule are recorded as passed or failed.

[0049] When a contradiction in numerical logic or terminology is determined, all fields involved in the rule that fails the verification are marked as fields to be corrected. An internal consistency verification result record is generated for each field to be corrected. Each internal consistency verification result record includes the name of the marked field, the current field value, the description of the violated rule, and the contradiction type. The contradiction type is divided into numerical logic contradiction or terminology contradiction. All internal consistency verification result records are summarized as the internal consistency verification result of the field to be corrected.

[0050] When obtaining the fields to be corrected and the fields that have passed the validation in the structured claims information object as the set to be cross-validated, all fields of the structured claims information object are traversed, and the entries marked as fields to be corrected and the fields that have passed the validation are extracted. The field names and field values ​​of the two types of fields are kept in pairs and grouped into the same set, which is the set to be cross-validated. The set to be cross-validated also carries the semantic role information of the text fragments associated with each field during the generation process.

[0051] When determining whether entities and events in the optimized text data support the cross-validation set, named entity recognition is performed on all text fragments of the optimized text data to extract entities such as names of persons, names of organizations, dates and times, monetary values, names of diseases, and accident types. At the same time, event extraction is performed to identify events described in the text, such as the course of the accident, medical treatment, and the occurrence of expenses, as well as the participating entities and the order of occurrence. Each field value in the cross-validation set is matched with the extracted entity set. If a field value has an entity in the entity set that is completely identical or has a synonymous mapping, the field is marked as having entity support. Multiple related fields in the cross-validation set are combined into a validation proposition and compared with the semantics of the extracted events. If the participating entities and time order of the event are consistent with the field relationship in the proposition, the group of fields is marked as having event support. The support judgment result for each field is generated by combining the entity support result and the event support result.

[0052] When generating cross-validation results based on the support judgment results, for fields to be corrected that have obtained entity and event support and do not contradict the internal consistency check results, an indication to confirm and maintain the original field is generated. For fields to be corrected that have not obtained entity support or event support, the candidate value that best fits the context is retrieved from the entity set and event semantics. If there is a unique candidate value that is different from the current value of the field to be corrected, the candidate value is written as a correction suggestion into the cross-validation results. If the number of candidate values ​​exceeds the preset number, the field is determined to be a field that needs manual review and is marked in the cross-validation results. The cross-validation results include the maintenance indication for fields that have been confirmed to pass and the correction suggestion or confirmation and maintenance indication for fields to be corrected.

[0053] When locating the original semantic position of each field in the claim document and the corresponding insurance clause item in the cross-validation results based on the semantic role of the text fragment, the semantic role information retained by each field during the generation process is read from the cross-validation results. Based on the semantic role, the bounding rectangle coordinates and the text area position of the field value text fragment in the original claim document image are reverse-indexed from the image layout information set as the original semantic position. At the same time, the insurance clause mapping library is accessed using the semantic role as the search key. The insurance clause mapping library pre-stores the correspondence between the semantic role of each claim field and the specific clause number and clause text in the insurance contract. The corresponding clause number and clause text are retrieved as the insurance clause item associated with the field.

[0054] When generating rule matching conclusions for confirmed and revised fields by calling the insurance clause judgment rules that match the semantic role, a pre-configured insurance clause judgment rule library is loaded. Each rule in the insurance clause judgment rule library contains the applicable semantic role, comparison logic, and conclusion template. The comparison logic specifies that the field value is compared with the clause parameters such as the limit, deductible, waiting period, and coverage scope stipulated in the clause, or a inclusion judgment is made. For each confirmed and revised field, the judgment rule corresponding to its semantic role is matched one by one. The field value is substituted into the comparison logic and the clause parameters in the corresponding insurance clause item are used to perform the operation. Based on the operation result, the conclusion template is applied to generate a rule matching conclusion. The rule matching conclusion indicates whether the field meets the insurance clause stipulation and the corresponding claim decision tendency.

[0055] When integrating rule matching conclusions with semantic role information of fields requiring manual review to generate the final review result of the claims document, the final review result data structure is created, which includes three partitions: a list of automatically approved claims, a list of automatically rejected claims, and a list of disputed items awaiting manual review. The rule matching conclusions of all fields are traversed, and fields whose conclusions conform to the insurance terms and their corresponding claim amounts are included in the list of automatically approved claims. Fields whose conclusions do not conform to the insurance terms and their reasons for rejection are included in the list of automatically rejected claims. Fields marked as requiring manual review in the cross-validation results, along with their semantic role information, current field value, correction suggestions, and original semantic position, are included in the list of disputed items awaiting manual review. The three partition lists are then assembled to form the final review result of the claims document.

[0056] The beneficial effects are as follows: by loading a pre-defined list of claim field names, the system retrieves field value nodes that match candidate semantic roles from the optimized text data, forming a structured claim information object with a key-value pair structure, thus achieving automated and accurate collection of key claim information. The system loads a set of numerical logic relationship rules and a set of professional terminology association rules to perform line-by-line comparison and verification of the structured claim information object, automatically identifying inherent logical conflicts such as contradictions between the total cost and the sum of the itemized costs, deviations between the compensation amount and the insured amount and compensation ratio, and mismatches between disease names and the scope of insurance coverage. It then generates internal consistency verification results carrying field names, current field values, descriptions of violations, and contradiction types, accurately identifying fields to be corrected. The fields to be corrected are merged with the fields that have passed verification into a set for cross-validation. Named entity recognition and event extraction are performed on the optimized text data. The authenticity of field values ​​is verified using the full-text entity set and event semantics. Fields with dual support maintain their original values, while fields lacking support retrieve the candidate values ​​from the entity set that best fit the context as correction suggestions, generating cross-validation results containing maintenance instructions and correction suggestions. Based on semantic roles, the system uses a reverse index to locate the coordinates of fields in the original document from the image layout information set. Simultaneously, it retrieves the corresponding clause number and text from the insurance clause mapping library, and uses comparison logic and conclusion templates from the insurance clause judgment rule library to generate a clause compliance conclusion. Finally, it constructs three partitions: an automatically approved claim list, an automatically rejected claim list, and a list of disputed items awaiting manual review, achieving structured output and automated distribution of claim conclusions.

[0057] like Figure 2 The diagram shown is a functional block diagram of an insurance claims system based on OCR and a large language model, provided by an embodiment of the present invention.

[0058] The insurance claims system 100 based on OCR and a large language model described in this invention can be installed in an electronic device. Depending on the functions implemented, the insurance claims system 100 based on OCR and a large language model may include a semantic parsing and confidence assessment module 101, a low-confidence text targeting optimization module 102, and a dual semantic verification and review generation module 103. The modules described in this invention can also be referred to as units, which are a series of computer program segments that can be executed by the processor of an electronic device and can perform fixed functions, stored in the memory of the electronic device.

[0059] In this embodiment, the functions of each module / unit are as follows: The semantic parsing and confidence assessment module 101 is used to call the fine-tuned large language model in the insurance field to perform semantic parsing on the claims file to be processed, and to obtain the semantic role and preliminary confidence of each text segment in the claims file; The low-confidence text orientation optimization module 102 is used to generate orientation recognition prompts through a large language model based on the initial confidence level to perform secondary recognition on low-confidence text fragments and obtain optimized text data. The dual semantic verification and review generation module 103 is used to perform dual semantic logical reasoning verification through a large language model, perform dual semantic consistency verification on the key claims information extracted from the optimized text data, and generate the final review result of the claims document based on the verification result and the semantic role.

[0060] In the several embodiments provided by this invention, it should be understood that the disclosed methods and systems can be implemented in other ways. For example, the system embodiments described above are merely illustrative; for instance, the division of modules is only a logical functional division, and other division methods may be used in actual implementation.

[0061] The modules described as separate components may or may not be physically separate. The components shown as modules may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.

[0062] Furthermore, the functional modules in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or in the form of hardware plus software functional modules.

[0063] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention.

[0064] This application embodiment can acquire and process relevant data based on artificial intelligence technology. Artificial intelligence is the theory, method, technology, and application system that uses digital computers or machines controlled by digital computers to simulate, extend, and expand human intelligence, perceive the environment, acquire knowledge, and use that knowledge to obtain optimal results.

[0065] Finally, it should be noted that the above 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 preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. An insurance claims method based on OCR and a large language model, characterized in that, The method includes: The insurance domain fine-tuned large language model is invoked to perform semantic parsing on the claims documents to be processed, so as to obtain the semantic role and preliminary confidence of each text segment in the claims documents; Based on the initial confidence level, targeted recognition prompts are generated using a large language model to perform secondary recognition on low-confidence text fragments, resulting in optimized text data. By performing dual semantic logic reasoning verification through a large language model, the key claims information extracted from the optimized text data is subjected to dual semantic consistency verification, and the final review result of the claims document is generated based on the verification result and the semantic role.

2. The insurance claims method based on OCR and a large language model as described in claim 1, characterized in that, Before performing semantic parsing on the claim documents to be processed, the following is included: Denoising was performed on the collected claim document images; The initial text dataset is obtained by identifying all recognizable text regions in the denoised data. The image layout information set is input into a large language model as layout prior knowledge for semantic parsing.

3. The insurance claims method based on OCR and a large language model as described in claim 2, characterized in that, The process involves calling a fine-tuned large language model in the insurance domain to perform semantic parsing on the claims file to obtain the semantic role and preliminary confidence level of each text segment in the claims file, including: Based on the insurance semantic annotation prompt words of the large language model, combined with the positional adjacency relationship and font attributes in the image layout feature parameters, the contextual semantic role pre-annotation is performed on each initial text segment in the initial text dataset to obtain the candidate semantic role of the text segment. By using the logical coherence reasoning of the large language model, combined with the logical coherence between the text fragment and the candidate semantic roles of adjacent text fragments, and the preset document structure corresponding to the relative position of the text fragment in the claims document, a preliminary confidence level of the text fragment is generated.

4. The insurance claims method based on OCR and a large language model as described in claim 3, characterized in that, Based on the initial confidence level, a targeted recognition prompt word is generated using a large language model to perform secondary recognition on the low-confidence text fragments, resulting in optimized text data, including: Text fragments with an initial confidence level lower than a preset matching threshold are marked as low-confidence text fragments, and the location area information of the low-confidence text fragments in the claims document is extracted; By generating targeted fine-grained recognition prompts containing character type constraints and word candidate ranges through a large language model, targeted fine-grained recognition instructions based on location region information are obtained. The image region defined by the location region information is subjected to secondary recognition according to the directional fine recognition instruction, and the returned secondary recognition candidate text is received.

5. The insurance claims method based on OCR and a large language model as described in claim 4, characterized in that, After receiving the returned secondary recognition candidate text, the process also includes: The secondary recognition candidate text is merged with the non-low confidence text fragments in the initial text fragment to obtain the candidate optimized text dataset; By combining the semantic fluency assessment of the large language model with the traditional semantic coherence formula, the candidate texts are evaluated for semantic coherence, and the candidate text with the highest score is selected as the final replacement text. The corresponding low-confidence text fragment in the initial text fragment is replaced with the final replacement text to obtain optimized text data.

6. The insurance claims method based on OCR and a large language model as described in claim 5, characterized in that, The formula for calculating the semantic coherence score is as follows: ； In the formula, Score the semantic coherence. This is the number of elements in the intersection of the vocabulary set of the candidate text for secondary recognition and the vocabulary set of the adjacent text on the left. This refers to the number of elements in the intersection of the vocabulary set of the candidate text for secondary recognition and the vocabulary set of the adjacent text on the right. The number of elements in the vocabulary set in the candidate text for secondary recognition. This represents the number of elements in the vocabulary set of the text adjacent to the left. This represents the number of elements in the vocabulary set of the adjacent text on the right.

7. The insurance claims method based on OCR and a large language model as described in claim 1, characterized in that, The dual semantic consistency verification of key claims information extracted from optimized text data includes: By using a large language model, key claim information corresponding to preset claim fields is extracted from optimized text data to obtain structured claim information objects; Through logical reasoning using a large language model, we examine the numerical logical relationships and professional terminology associations between different field values ​​in a structured claims information object. When a contradiction is found in the numerical logic relationship or the terminology association relationship, the corresponding field is marked as a field to be corrected, and an internal consistency check result for the field to be corrected is generated.

8. The insurance claims method based on OCR and a large language model as described in claim 7, characterized in that, The dual semantic consistency verification of key claims information extracted from optimized text data also includes: The fields to be corrected and the fields that have passed the verification in the structured claims information object are obtained and used as a set to be cross-validated. By extracting entities and events from a large language model, we determine whether the entities and events in the optimized text data support the set to be cross-validated. The cross-validation results are generated based on the support judgment results, wherein the cross-validation results include suggestions for correcting the field to be corrected or instructions to maintain the original field.

9. The insurance claims method based on OCR and a large language model as described in claim 8, characterized in that, The step of generating the final review result of the claim document based on the verification result and the semantic role includes: Based on the semantic role of the text fragment, locate the original semantic position of each field in the cross-validation result in the claims file and the corresponding insurance clause item; By matching insurance clause rules using a large language model, rule matching conclusions are generated for confirmed and corrected fields. By integrating the rule matching results with the semantic role information of the fields requiring manual review, the final review result of the claim document is generated. The final review result includes automatically approved claim items, automatically rejected claim items, and disputed items pending manual review.

10. An insurance claims system based on OCR and a large language model, used to implement the insurance claims method based on OCR and a large language model as described in claim 1, the system comprising: The semantic parsing and confidence assessment module is used to call the fine-tuned large language model in the insurance field to perform semantic parsing on the claims documents to be processed, and obtain the semantic role and preliminary confidence of each text segment in the claims documents; The low-confidence text orientation optimization module is used to generate orientation recognition prompts through a large language model based on the initial confidence level to perform secondary recognition on low-confidence text fragments and obtain optimized text data. The dual semantic verification and review generation module is used to perform dual semantic logical reasoning verification through a large language model, perform dual semantic consistency verification on key claims information extracted from optimized text data, and generate the final review result of the claims document based on the verification result and the semantic role.

Images