CN-121982734-A - Examination and approval information structured extraction method and system based on OCR and large model

Abstract

The invention belongs to the technical field of image information recognition, and particularly relates to an approval information structured extraction method based on OCR and a large model, comprising the following steps of S100, preprocessing an approval form image to be processed, and extracting text content and corresponding text position information through an OCR technology; S200, inputting a text extracted by OCR into a large language model, carrying out semantic understanding, field completion and structured extraction through a preset prompt word guide model, S300, outputting structured data, marking confidence and missing conditions, and compared with the traditional OCR analysis mode, the technical scheme of the invention has the capability of supplementing missing information of a structural field and the capability of mining implicit semantic extraction wind control characteristics of approval opinions.

Inventors

• Qu GuangYang
• PENG ZHIYAO
• LI MINGHUI
• WEI BIN
• SUN SHITONG

Assignees

• 重庆工程学院

Dates

Publication Date

20260505

Application Date

20251210

Claims (6)

1. 1. The structural extraction method of approval information based on OCR and a large model is characterized by comprising the following steps: S100, preprocessing an approval form image to be processed, and extracting text content and corresponding text position information through an OCR technology; s200, inputting the text extracted by OCR into a large language model, and carrying out semantic understanding, field completion and structured extraction through a preset prompt word guide model; S300, outputting the structured data, and marking confidence and missing conditions.
2. 2. The method for structured extraction of approval information based on OCR and large models as set forth in claim 1, wherein the step S100 of preprocessing the image of the approval form to be processed is specifically image processing by image enhancement, graying, binarization and tilt correction.
3. 3. The method for structured extraction of approval information based on OCR and large models as set forth in claim 1, wherein the step S200 of "structured extraction" comprises: ① Directly extracting the rule field; ② And carrying out semantic analysis on the unstructured text, and extracting implicit wind control features.
4. 4. The method for structured extraction of approval information based on OCR and large models as set forth in claim 3, wherein said implicit wind control features include job stability, revenue realism, home payout structure, spouse economic contribution, implicit liabilities and risk ratings.
5. 5. The method for structured extraction of approval information based on OCR and large models as set forth in claim 1, wherein the outputting of the structured data in S300 is specifically to organize the content of each field obtained by recognition and reasoning into a standard JSON, CSV or Excel structure, and to label the confidence of field extraction and whether it is missing.
6. 6. A system for application to the extraction method according to any one of claims 1 to 5, characterized in that it comprises the following modules: the OCR text recognition module is used for extracting text and position information in the approval image; The promt extraction instruction module is used for constructing a field to be extracted and a promt instruction and guiding the large language model to carry out semantic reasoning and extraction; And the large language model structured extraction module is used for completing the semantic attribution and structured extraction tasks of natural language understanding and free expression according to the Prompt generated by the OCR analysis text and the Prompt extraction instruction module.

Description

Examination and approval information structured extraction method and system based on OCR and large model Technical Field The invention belongs to the technical field of image information identification, and particularly relates to an approval information structured extraction method and system based on OCR and a large model. Background In financial services such as automobile financing and leasing, credit consumption and the like, off-line approval is still an important link for evaluating customer risk and verifying information authenticity, and paper or scanning forms are usually generated in the approval process, and the paper or scanning forms comprise customer basic information, financial conditions, evaluation comments and the like. Such documents have the following problems: ① Structural field missing cannot be complemented, and information integrity is limited: OCR (optical character recognition) cannot recognize or fields are empty in an approval table often because of the problems of non-uniform form formats, irregular handwriting, blurred scanning and the like. The traditional information extraction method only depends on the field which is visible in the text, and the missing items cannot be complemented, so that wind control modeling basic data is incomplete, and analysis deviation is large. ② Unstructured text is difficult to analyze in a standardized way, and intelligent analysis capability is weak: The examination and approval list is mainly a scanning part, has a complex structure, is not easy to identify, is difficult to extract key information, is expressed in natural language, has flexible expression mode and strong subjectivity, contains a large amount of subjective description and hidden risk information, is difficult to realize structural analysis by a regular extraction means, and leads to the fact that wind control characteristics cannot be extracted effectively, thereby limiting the exertion of automatic wind control analysis capability. Therefore, we propose a method and a system for structurally extracting approval information based on OCR and a large model. Disclosure of Invention The invention aims to provide an approval information structured extraction method and system based on OCR and a large model, which are used for solving the problems in the background technology. In order to achieve the technical purpose, the invention adopts the following technical scheme: an approval information structured extraction method based on OCR and a large model comprises the following steps: S100, preprocessing an approval form image to be processed, and extracting text content and corresponding text position information through an OCR technology; s200, inputting the text extracted by OCR into a large language model, and carrying out semantic understanding, field completion and structured extraction through a preset prompt word guide model; S300, outputting the structured data, and marking confidence and missing conditions. As a further limitation of the technical solution of the present invention, the "preprocessing the to-be-processed approval form image" in step S100 is specifically to perform image processing by using image enhancement, graying, binarization and tilt correction. As a further limitation of the present invention, the "structured extraction" in step S200 includes: ① Directly extracting the rule field; ② And carrying out semantic analysis on the unstructured text, and extracting implicit wind control features. As a further limitation of the solution of the present invention, the implicit wind control features include job stability, revenue realism, home payout structure, spouse economic contribution, implicit liabilities and risk ratings. As a further limitation of the technical scheme of the present invention, the outputting of the structured data in S300 is specifically to organize the content of each field obtained by recognition and reasoning into a standard JSON, CSV or Excel structure, and mark the confidence level of field extraction and whether it is missing. An extraction system based on OCR and large model approval information structuring comprises the following modules: the OCR text recognition module is used for extracting text and position information in the approval image; and the promt extraction instruction module is used for constructing a field to be extracted and a promt instruction according to a task target by a user and guiding the large language model to carry out semantic reasoning and extraction. And the large language model structured extraction module is used for completing the semantic attribution and structured extraction tasks of natural language understanding and free expression according to the Prompt generated by the OCR analysis text and the Prompt extraction instruction module. Compared with the traditional OCR analysis mode, the technical scheme of the invention has the capability of supplementing the information of the structural field