Intelligent parsing and information extraction method for unstructured documents

Abstract

This invention discloses an intelligent parsing and information extraction method for unstructured documents, including acquiring historical and real-time document information; parsing the real-time document information, which includes text, image, audio, and video information; extracting metadata from the text, image, audio, and video information; analyzing the real-time document information using the metadata; outputting the real-time document information as structured data using the metadata; validating the output structured data; and storing the valid structured data in an enterprise database to update the enterprise database. By parsing documents containing text, images, audio, and video into a unified document, it replaces manual or single-tool processing methods, which helps improve the efficiency of document processing; and by using a preset metadata framework and iterative analysis based on historical data, it achieves accurate and in-depth extraction of key business information.

Description

Technical Field

[0001] This invention relates to the field of document parsing technology, and in particular to an intelligent parsing and information extraction method for unstructured documents. Background Technology

[0002] In modern enterprise operations, documents are the core carriers of information and knowledge transfer. With the increasing complexity of business activities, the forms of enterprise documents are becoming increasingly diversified, including not only traditional written contracts and meeting minutes, but also a large amount of multimodal information such as images, audio, and video. These unstructured documents contain enormous commercial value, such as customer relationships, transaction details, and project progress. How to efficiently and accurately extract valuable information from these massive, multimodal documents and transform it into structured knowledge that can be understood and utilized by computers to build and update enterprise databases has become a key challenge for improving enterprise intelligence and decision-making efficiency. Existing unstructured document processing solutions typically adopt a pipeline model of "OCR (Optical Character Recognition) + NLP (Natural Language Processing)". First, OCR technology is used to convert document images into plain text. Then, deep learning models (such as BiLSTM-CRF or BERT) are used to perform sequence annotation on the text to extract key entities and relationships such as names, amounts, and dates.

[0003] Currently, information processing for enterprise documents primarily relies on manual reading and input, or on automated tools designed for a single modality. For example, OCR technology is used to process text in images, or Automatic Speech Recognition (ASR) technology is used to transcribe audio content. However, these existing technologies suffer from inefficiencies due to manual processing, which is prone to information omissions or errors due to subjective factors. Furthermore, single-modality processing tools cannot effectively integrate information from different sources and modalities into unified structured data (e.g., they cannot automatically link the specifications and prices of "products" discussed in meeting recordings with those in scanned contracts and historical contracts), resulting in fragmented information and hindering the formation of a complete business view. This significantly diminishes the value of data assets. Summary of the Invention

[0004] The technical problem solved by this invention is that manual processing is inefficient and prone to information omissions or errors due to subjective factors. Meanwhile, single-modal processing tools cannot effectively integrate information from different sources and modalities into unified structured data, resulting in information fragmentation and difficulty in forming a complete business view, which greatly reduces the value of data assets.

[0005] To address the aforementioned technical problems, this invention provides the following technical solution: a method for intelligent parsing and information extraction from unstructured documents, comprising: S1: Obtain historical document information and real-time document information, and parse the real-time document information, which includes text information, image information, audio information and video information; S2: Extract metadata from text, image, audio, and video information, and use the metadata to analyze real-time document information; S3: Utilize metadata to output real-time document information as structured data; S4: Validate the output structured data and store the validated structured data into the enterprise database to update the enterprise database.

[0006] As a preferred embodiment of the intelligent parsing and information extraction method for unstructured documents described in this invention, step S1 includes: S11: Retrieve historical document information from the enterprise database, and use meeting minutes and contract texts as real-time document information; S12: Extract the first document information from the text information; S13: Use image interpretation technology to obtain secondary document information from image information; S14: Use speech recognition technology to convert audio information into third-party document information; S15: Decompose the video information into a set of images composed of individual images and audio information. Use image interpretation technology to obtain the fourth document information in the image set and use speech recognition technology to obtain the fifth document information in the audio information.

[0007] As a preferred embodiment of the intelligent parsing and information extraction method for unstructured documents described in this invention, the metadata includes subject data, object data, consideration settlement data, time data, responsibility data, default data, and change / termination data.

[0008] As a preferred embodiment of the intelligent parsing and information extraction method for unstructured documents described in this invention, the main data includes the company's full name, unified social credit code, company address, personnel name, ID card number, personnel address, and contact information. The object data includes product name, brand, model, specifications, quantity, service content, scope, standards, deliverables, software usage rights, copyright authorization scope, project location, scope, and design requirements; The consideration settlement data includes unit price, total price, currency, whether tax is included, bank transfer, bank acceptance bill, cash prepayment, progress payment, quality assurance deposit, payment terms, invoice type, tax rate, and invoice issuance time; The time data includes the validity period delivery time, service completion date, start and end date of the project period, and payment deadline; The liability data includes the necessary information provided by Party A, working conditions, timely payment, Party B's guarantee of product quality, delivery information, training information, and confidentiality clauses; The default data includes the circumstances of the default, the penalty for breach of contract, and the compensation for losses. The change cancellation data includes the change conditions and cancellation conditions.

[0009] As a preferred embodiment of the intelligent parsing and information extraction method for unstructured documents described in this invention, the analysis of document information using metadata includes: S21: Using the main data, object data, consideration settlement data, time data, responsibility data, default data, and change / termination data in the real-time document information as primary keywords, extract primary keywords from the first document information, second document information, third document information, fourth document information, and fifth document information to generate a metadata database. Based on the main data, object data, consideration settlement data, time data, responsibility data, default data, and change / termination data, classify the primary keywords in the metadata database to generate a dataset. S22: Take the main data, object data, consideration settlement data, time data, responsibility data, default data and change / termination data in the historical document information as the input keywords of the first hash function, and take the first keyword as the hash address output by the first hash function.

[0010] As a preferred embodiment of the intelligent parsing and information extraction method for unstructured documents described in this invention, the method further includes: analyzing document information using metadata. S23: Input the input keyword of the first hash function into the first hash function, output the hash address of the first hash function, filter the primary keywords in the metadata database according to the hash address of the first hash function, define the filtering result as the secondary keyword, take all the metadata in the primary dataset where the secondary keyword is located as the tertiary keyword, and expand the secondary keyword; S24: Use the tertiary keyword as the input keyword of the second hash function, use the primary keyword as the hash address output by the first hash function, input the tertiary keyword into the second hash function, output the hash address of the second hash function, filter the primary keywords in the metadata database according to the hash address of the second hash function, define the filtering result as the quaternary keyword, use all the metadata in the primary dataset where the quaternary keyword is located as the quinary keyword, and expand the quaternary keyword; S25: Use the fifth keyword as the input keyword of the third hash function, use the first keyword as the hash address output by the first hash function, input the fifth keyword into the third hash function, output the hash address of the third hash function, filter the first keywords in the metadata database according to the hash address of the third hash function, define the filtering result as the sixth keyword, use all the metadata in the first dataset where the sixth keyword is located as the seventh keyword, and expand the sixth keyword; S26: Repeat the iteration n times to obtain first-order keywords, third-order keywords, fifth-order keywords, seventh-order keywords, ..., m-order keywords, where m is an odd number. Use a clustering algorithm to classify the same product names and service content among the first-order keywords, third-order keywords, fifth-order keywords, seventh-order keywords, ..., m-order keywords, obtain the classification results, and count the classification results respectively. Count the first occurrence frequency among the first-order keywords, third-order keywords, fifth-order keywords, seventh-order keywords, ..., m-order keywords, and use the first occurrence frequency as the first judgment criterion. If the occurrence frequency is greater than a predetermined frequency threshold x, it indicates that the product name or service content is a key product or service.

[0011] As a preferred embodiment of the intelligent parsing and information extraction method for unstructured documents described in this invention, the product name and service content in the first keyword are used as the query object. The second occurrence count of the product name and service content in the first keyword in the third keyword, fifth keyword, seventh keyword, ..., m keyword is traversed sequentially. The second occurrence count is used as the second judgment criterion. If the occurrence count is greater than the predetermined count threshold y, it indicates that the product name or service content is the core product or service.

[0012] As a preferred embodiment of the intelligent parsing and information extraction method for unstructured documents described in this invention, the method further includes: analyzing document information using metadata. S27: Extract the object data and subject data that are directly related to the core product or service from the secondary keywords, sort the remaining metadata in the secondary keywords from most to least using the bubble sort algorithm, and select the top 4 metadata as high-activity keywords. S28: Use text vectorization technology to vectorize highly active keywords to obtain vectors A1, A2, A3 and A4 respectively; S29: Use the cosine function to calculate vector similarity, which is then used as the corresponding text similarity. The calculation expression is as follows: ; Where Ai and Bi represent the components of vectors A and B, respectively, and n represents the number of times the highly active keyword appears in the secondary keyword.

[0013] As a preferred embodiment of the intelligent parsing and information extraction method for unstructured documents described in this invention, step S3 includes: S31, construct database templates based on business domains from historical document information, and establish the correspondence between business domains and database templates. S32, Business areas for annotating real-time document information, including procurement or sales contracts, service contracts, lease contracts, labor contracts, cooperation agreements, and technology development contracts; S33, Based on the business domain of the real-time document information and the corresponding relationship, fill the metadata into the corresponding position of the database template and output the structured data of the real-time document information; As a preferred embodiment of the intelligent parsing and information extraction method for unstructured documents described in this invention, step S4 includes: The system compares the object data, settlement data, time data, liability data, default data, and change / termination data corresponding to the key products or services last updated in the historical document information with the corresponding object data, settlement data, time data, liability data, default data, and change / termination data in the metadata. If different data is found, the different data is filtered out and marked. The tagged data undergoes manual review. Once the manual review is passed, the data passes the verification and the structured data that passes the verification is stored in the enterprise database to update the enterprise database.

[0014] The beneficial effects of this invention are as follows: By parsing documents containing text, images, audio, and video into a unified document, it replaces manual or single-tool processing methods, thereby improving document processing efficiency; through a preset metadata framework and iterative analysis based on historical data, it achieves accurate and in-depth extraction of key business information and automatically identifies key products or services, effectively improving the accuracy of information extraction; by utilizing metadata-based analysis, it deeply mines the intrinsic value of information, then generates standardized structured data through templates and updates the enterprise database; finally, a verification mechanism ensures the accuracy and reliability of knowledge base updates, breaking down barriers between multimodal information, achieving effective data integration and accumulation, and providing high-quality data support for enterprise decision analysis and risk warning, thus fully demonstrating the value of enterprise document data assets. Attached Figure Description

[0015] Figure 1 This is a schematic diagram of the basic process of an intelligent parsing and information extraction method for unstructured documents provided in one embodiment of the present invention. Detailed Implementation

[0016] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0017] Example 1, referring to Figure 1 As an embodiment of the present invention, a document analysis method based on a large model and fusing multimodal information is provided, comprising: Step S1: Obtain historical document information and real-time document information, and parse the real-time document information, which includes text information, image information, audio information and video information; Step S2: Extract metadata from text, image, audio, and video information, and use the metadata to analyze real-time document information; Step S3: Use metadata to output real-time document information as structured data; Step S4: Validate the output structured data and store the validated structured data into the enterprise database to update the enterprise database.

[0018] In this embodiment, the process of parsing documents containing text, images, audio, and video into a unified document replaces manual or single-tool processing, thus improving document processing efficiency. Through a pre-defined metadata framework and iterative analysis based on historical data, it achieves accurate and in-depth extraction of key business information and automatically identifies key products or services, effectively improving the accuracy of information extraction. Metadata-based analysis deeply mines the intrinsic value of information, then generates standardized structured data through templates and updates the enterprise database. Finally, a verification mechanism ensures the accuracy and reliability of knowledge base updates, breaking down barriers between multimodal information and achieving effective data integration and accumulation. This provides high-quality data support for enterprise decision analysis and risk warning, facilitating the full realization of the value of enterprise document data assets.

[0019] Example 2 is another embodiment of the present invention, which differs from the first embodiment in that step S1 includes: S11: Retrieve historical document information from the enterprise database, and use meeting minutes and contract texts as real-time document information; S12: Extract the first document information from the text information; directly read the content of a text file (such as docx, .txt or pdf format) and generate the first document information; S13: Use image interpretation technology to obtain second document information from image information; use image interpretation technology (such as extracting text from scanned images through OCR technology) to extract text and key visual elements (such as seals, signatures, charts) from images to generate second document information.

[0020] S14: Use speech recognition technology to convert audio information into third-party document information; use speech recognition (ASR) technology to convert the speech content in audio files (such as .mp3 or wav) or real-time audio streams into text to generate third-party document information.

[0021] S15: Decompose the video information into a set of individual images and audio information. Use image interpretation technology to obtain the fourth document information from the image set, and use speech recognition technology to obtain the fifth document information from the audio information. When a real-time document contains video, it is first decomposed into a set of individual images arranged in chronological order and independent audio information using video decoding technology. Then, image interpretation technology is applied to the image set to obtain the fourth document information, and speech recognition technology is applied to the audio information to obtain the fifth document information.

[0022] In this embodiment, a connector is preferentially used to obtain processed document information (such as meeting minutes and contract texts) from an existing document processing system, and then sends it to the document processing system. The document processing system processes the document information (such as meeting minutes and contract texts) as real-time document information. The connector first obtains processed document information from the existing document processing system, including document address, document details, etc.; or the meeting minutes or contract texts are manually input into the document processing system, which processes them as real-time document information. By integrating mature technologies such as OCR, ASR, and video decoding, automated parsing of all information modalities in the document is achieved, replacing traditional manual reading and transcription methods, which helps to improve the efficiency of information acquisition.

[0023] In one embodiment, a user-uploaded PDF procurement contract document is received and converted into a high-resolution image sequence (e.g., the resolution per page is adjusted to 300 dpi). An OCR engine (such as PaddleOCR or Tesseract) is invoked to recognize the images. The OCR engine not only outputs the text content "Party A: XX Technology Co., Ltd.", but also outputs the bounding box coordinates corresponding to each character or word, in the format [x_min, y_min, x_max, y_max].

[0024] Metadata includes entity data, object data, consideration settlement data, time data, liability data, default data, and change / termination data.

[0025] The main data includes the company's full name, unified social credit code, company address, personnel names, ID card numbers, personnel addresses, and contact information; The object data includes product name, brand, model, specifications, quantity, service content, scope, standards, deliverables, software usage rights, copyright license scope, project location, scope, and design requirements; The settlement data includes unit price, total price, currency, whether tax is included, bank transfer, bank acceptance bill, cash prepayment, progress payment, warranty deposit, payment terms, invoice type, tax rate, and invoice date; Time data includes the validity period delivery time, service completion date, start and end date of the project, and payment deadline; The responsibility data includes the necessary information provided by Party A, working conditions, timely payment, Party B's guarantee of product quality, delivery information, training information, and confidentiality clauses; Default data includes the circumstances of the default, the penalty for breach of contract, and compensation for losses. The data for changing or canceling data includes the conditions for changing the data and the conditions for canceling the data.

[0026] In this embodiment, metadata is prioritized as the core framework for information extraction. It defines the most critical information categories in enterprise business activities. Document parsing is the process of converting raw documents (i.e., real-time document information) containing different modalities such as text, images, audio, and video into a unified, machine-readable text data stream through a series of technical means, providing standardized data input for subsequent information extraction and analysis. Real-time document information from the enterprise file system or real-time data stream, along with historical document information obtained from the enterprise database, are used as auxiliary data. After parsing, the first to fifth document information in text form are passed to step S2 for analysis.

[0027] Analyzing document information using metadata includes: S21: Using the main data, object data, consideration settlement data, time data, responsibility data, default data, and change / termination data in the real-time document information as primary keywords, extract primary keywords from the first document information, second document information, third document information, fourth document information, and fifth document information to generate a metadata database. Based on the main data, object data, consideration settlement data, time data, responsibility data, default data, and change / termination data, classify the primary keywords in the metadata database to generate a dataset. S22: Take the main data, object data, consideration settlement data, time data, responsibility data, default data and change / termination data in the historical document information as the input keywords of the first hash function, and take the first keyword as the hash address output by the first hash function.

[0028] Analyzing document information using metadata also includes: S23: Input the input keyword of the first hash function into the first hash function, output the hash address of the first hash function, filter the primary keywords in the metadata database according to the hash address of the first hash function, define the filtering result as the secondary keyword, take all the metadata in the primary dataset where the secondary keyword is located as the tertiary keyword, and expand the secondary keyword; S24: Use the tertiary keyword as the input keyword of the second hash function, use the primary keyword as the hash address output by the first hash function, input the tertiary keyword into the second hash function, output the hash address of the second hash function, filter the primary keywords in the metadata database according to the hash address of the second hash function, define the filtering result as the quaternary keyword, use all the metadata in the primary dataset where the quaternary keyword is located as the quinary keyword, and expand the quaternary keyword; S25: Use the fifth keyword as the input keyword of the third hash function, use the first keyword as the hash address output by the first hash function, input the fifth keyword into the third hash function, output the hash address of the third hash function, filter the first keywords in the metadata database according to the hash address of the third hash function, define the filtering result as the sixth keyword, use all the metadata in the first dataset where the sixth keyword is located as the seventh keyword, and expand the sixth keyword; S26: Repeat the iteration n times to obtain first-order keywords, third-order keywords, fifth-order keywords, seventh-order keywords, ..., m-order keywords, where m is an odd number. Use the clustering algorithm to classify the same product names and service content among the first-order keywords, third-order keywords, fifth-order keywords, seventh-order keywords, ..., m-order keywords, obtain the classification results, and count the first occurrence of each category in the classification results. Use the first occurrence as the first judgment criterion. If the occurrence count is greater than the predetermined threshold x, it indicates that the product name or service content is a key product or service.

[0029] Using the product name and service content in a single keyword as the query object, the second occurrence count of the product name and service content in the single keyword is sequentially traversed through the three-fold, five-fold, seven-fold, ..., m-fold keywords. The second occurrence count is used as the second judgment criterion. If the occurrence count is greater than the predetermined frequency threshold y, it indicates that the product name or service content is the core product or service.

[0030] Analyzing document information using metadata also includes: S27: Extract the object data and subject data that are directly related to the core product or service from the secondary keywords, sort the remaining metadata in the secondary keywords from most to least using the bubble sort algorithm, and select the top 4 metadata as high-activity keywords. S28: Use text vectorization technology to vectorize highly active keywords to obtain vectors A1, A2, A3 and A4 respectively; S29: Use the cosine function to calculate vector similarity, which is then used as the corresponding text similarity. The calculation expression is as follows: ; Where Ai and Bi represent the components of vectors A and B, respectively, and n represents the number of times the highly active keyword appears in the secondary keyword.

[0031] In this preferred embodiment, iterative analysis is performed by introducing historical document information and employing hashing, clustering, and vectorization algorithms to achieve a deep understanding and information mining of document information. This not only allows for the analysis of images and videos that do not contain text or audio information to extract surface information, but also identifies potential connections between key products, core services, and critical business elements of an enterprise, clarifying user preferences and needs. Furthermore, it facilitates the intuitive observation of changes in the same data in real-time and historical document information of key products, core services, and critical business elements, preventing contract oversights and understanding changes in new contracts. Its analytical depth and intelligence level are higher than traditional keyword matching or rule extraction methods.

[0032] Step S3 includes: S31, construct database templates based on business domains from historical document information, and establish the correspondence between business domains and database templates. S32 marks the business areas for real-time document information. These business areas include purchase or sales contracts, service contracts, lease contracts, labor contracts, cooperation agreements, and technology development contracts. Different database templates are set up for different business areas, meaning different contract templates are used for different business areas.

[0033] S33, based on the business domain and corresponding relationship of the real-time document information, fill the metadata into the corresponding position of the same type of data in the database template (such as filling the unit price data in the metadata into the position of the unit price data in the database template), and output the structured data of the real-time document information.

[0034] In this preferred embodiment, a templated data generation mechanism is used to ensure that information extracted from different documents can be output in a unified and standardized format, which solves the problem of data chaos caused by inconsistent document formats. This allows the generated structured data to be seamlessly integrated into the company's existing ERP, CRM and other business systems, truly realizing the integration and utilization of data assets.

[0035] Step S4 includes: The system compares the object data, settlement data, time data, liability data, default data, and change / termination data corresponding to the key products or services last updated in the historical document information with the corresponding object data, settlement data, time data, liability data, default data, and change / termination data in the metadata. If different data is found, the different data is filtered out and marked. The tagged data undergoes manual review. Once the manual review is passed, the data passes the verification and the structured data that passes the verification is stored in the enterprise database to update the enterprise database.

[0036] In this preferred embodiment, the key metadata (such as object data, settlement data, etc.) in the structured data to be imported into the database is compared item by item with the corresponding data of the business entity in the enterprise database in the last update. Different data is automatically filtered out as discrepancies and highlighted or annotated to generate a verification report. Then, the verification report marked with discrepancies is pushed to a designated manual review terminal (such as a business manager or legal personnel). The manual reviewer confirms the discrepancies. If it is confirmed to be a valid update (such as a price adjustment or term change), the system clicks "Approved". If it is confirmed to be erroneous data, it is rejected and may trigger a re-parsing process. For the data that has passed the review, it is officially written into the enterprise database, completing the update of the enterprise database.

[0037] Through incremental comparison and manual review verification mechanisms, the system ensures the accuracy of data updates to the maximum extent while maintaining automation efficiency. This effectively prevents dirty data from polluting the enterprise database due to parsing errors or problems with the document itself, providing the enterprise with a trustworthy and reliable database.

[0038] In one embodiment, the verification includes format verification, logical verification, and data verification: Format verification includes checking whether the date format conforms to YYYY-MM-DD and whether the amount is a pure number or a standard currency format, marked by highlighting or annotation; Logical verification includes checking whether the extracted "signing date" is later than the "effective date". If there is a logical conflict, it is marked as low confidence and triggers manual review, marked by highlighting or annotation; Data verification includes comparing the consideration settlement data, time data, liability data, default data, and change / termination data with the corresponding object data, consideration settlement data, time data, liability data, default data, and change / termination data in the metadata. If different data is found, the different data is filtered out and marked.

[0039] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product implemented on one or more computer-usable storage media containing computer-usable program code. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk. These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0040] 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, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for intelligent parsing and information extraction of unstructured documents, characterized in that, The application relates to a method for updating an enterprise database, which comprises the following steps: S1: obtaining historical document information and real-time document information, and analyzing the real-time document information, wherein the real-time document information comprises text information, picture information, audio information and video information; S2: extracting metadata in the text information, the picture information, the audio information and the video information, and analyzing the real-time document information by using the metadata; S3: outputting the real-time document information as structured data by using the metadata; S4: checking the output structured data, and storing the structured data passing the check in an enterprise database to update the enterprise database. 2.The large model-based document analysis method of fusing multi-modal information according to claim 1, wherein: The S1 comprises the following steps: S11: obtaining historical document information from the enterprise database, and taking conference records and contract texts as real-time document information; S12: extracting first document information in the text information; S13: obtaining second document information in the picture information by using picture reading technology; S14: converting the audio information into third document information by using voice recognition technology; S15: decomposing the video information into a picture set composed of single pictures and sound information, obtaining fourth document information in the picture set by using picture reading technology, and obtaining fifth document information in the sound information by using voice recognition technology. 3.The large model-based document analysis method of fusing multi-modal information according to claim 2, wherein: The metadata comprises subject data, object data, counterparty settlement data, time data, responsibility data, breach data and change cancellation data. 4.The large model-based document analysis method of fusing multi-modal information according to claim 3, wherein: The subject data comprises the full name of a company, a unified social credit code, a company address, a personnel name, an ID card number, a personnel address and a contact method; The object data comprises product names, brands, models, specifications, quantities, service contents, ranges, standards, delivery objects, software use rights, copyright authorization ranges, engineering sites, ranges and design requirements; The counterparty settlement data comprises unit prices, total prices, currencies, whether tax-included bank transfer, accepted bills of exchange, cash advance, progress payments, warranty deposits, payment conditions, invoice types, tax rates and invoice times; The time data comprises valid period delivery times, service completion dates, construction period starts and ends and payment periods; The responsibility data comprises that Party A provides necessary materials, working conditions, timely payment, Party B guarantees product quality, delivery information, training information and confidentiality clauses; The breach data comprises breach situations, breach penalties and compensation losses; The change cancellation data comprises change conditions and cancellation conditions. 5.The large model-based document analysis method of fusing multi-modal information according to claim 4, wherein: The analysis of the document information by using the metadata comprises the following steps: S21: taking the subject data, the object data, the counterparty settlement data, the time data, the responsibility data, the breach data and the change cancellation data in the real-time document information as primary keywords, extracting the primary keywords in the first document information, the second document information, the third document information, the fourth document information and the fifth document information, generating a metadata database, classifying the primary keywords in the metadata database according to the subject data, the object data, the counterparty settlement data, the time data, the responsibility data, the breach data and the change cancellation data, and generating a primary data set; S22: The subject data, object data, counterparty settlement data, time data, responsibility data, breach data and change release data in the historical document information are taken as input keywords of the first hash function, and the first keyword is taken as a hash address output by the first hash function. 6.The large model-based document analysis method of fusing multi-modal information according to claim 5, wherein: The analysis of the document information by using the metadata further includes: S23: The input keyword of the first hash function is input into the first hash function, and a hash address of the first hash function is output. The first keyword in the metadata database is filtered according to the hash address of the first hash function, and the filtering result is defined as a second keyword. All the metadata in the first data set where the second keyword is located is taken as a third keyword, and the second keyword is expanded; S24: The third keyword is taken as an input keyword of the second hash function, the first keyword is taken as a hash address output by the first hash function, the third keyword is input into the second hash function, and a hash address of the second hash function is output. The first keyword in the metadata database is filtered according to the hash address of the second hash function, and the filtering result is defined as a fourth keyword. All the metadata in the first data set where the fourth keyword is located is taken as a fifth keyword, and the fourth keyword is expanded; S25: The fifth keyword is taken as an input keyword of the third hash function, the first keyword is taken as a hash address output by the first hash function, the fifth keyword is input into the third hash function, and a hash address of the third hash function is output. The first keyword in the metadata database is filtered according to the hash address of the third hash function, and the filtering result is defined as a sixth keyword. All the metadata in the first data set where the sixth keyword is located is taken as a seventh keyword, and the sixth keyword is expanded; S26: The iteration is repeated for n times to obtain the first keyword, the third keyword, the fifth keyword, the seventh keyword, …, and the mth keyword. m is a single number. The clustering algorithm is used to classify the same product name and service content in the first keyword, the third keyword, the fifth keyword, the seventh keyword, …, and the mth keyword. The classification result is obtained, and the classification result is counted respectively. The first occurrence number in the first keyword, the third keyword, the fifth keyword, the seventh keyword, …, and the mth keyword is counted respectively. The first occurrence number is taken as a first judgment standard. If the occurrence number is greater than a predetermined number threshold x, it indicates that the product name or service content is a key product or service. 7.The large model-based document analysis method of fusing multi-modal information according to claim 6, wherein: The product name and service content in the first keyword are taken as query objects, and the second occurrence number of the product name and service content in the first keyword in the third keyword, the fifth keyword, the seventh keyword, …, and the mth keyword is traversed in turn. The second occurrence number is taken as a second judgment standard. If the occurrence number is greater than a predetermined number threshold y, it indicates that the product name or service content is a core product or service. 8.The large model-based document analysis method of fusing multi-modal information according to claim 7, wherein: The analysis of the document information by using the metadata further includes: S27: Extract the object data and subject data in the secondary keywords that are directly related to the core product or service, and sort the remaining metadata in the secondary keywords from many to few using a bubble algorithm, and select the top 4 metadata as high-activity keywords; S28: Vectorize the high-activity keywords using a text vectorization technique to obtain vectors A1, A2, A3, and A4, respectively; S29: Calculate the vector similarity using a cosine function as the corresponding text similarity, and the calculation expression is: ； Where Ai and Bi represent the components of vectors A and B, respectively, and n represents the number of times the high-activity keyword appears in the secondary keywords. 9.The large model-based document analysis method of fusing multi-modal information according to claim 8, wherein: The S3 includes: S31, constructing a database template according to the business field of the historical document information, and establishing a corresponding relationship between the business field and the database template, S32, labeling the business field of the real-time document information, which includes procurement or sales contracts, service contracts, rental contracts, labor contracts, cooperation agreements, and technology development contracts; S33, filling the metadata into the corresponding position of the database template according to the business field of the real-time document information and the corresponding relationship, and outputting the structured data of the real-time document information. 10.The large model-based document analysis method of fusing multi-modal information according to claim 9, wherein: The S4 includes: Comparing the object data, counterparty settlement data, time data, liability data, breach data, and change release data corresponding to the key product or service in the last update of the historical document information with the corresponding object data, counterparty settlement data, time data, liability data, breach data, and change release data in the metadata, if there are different data, then the different data is filtered out and marked; The marked data is manually audited, and after passing the manual audit, it indicates that the verification is passed, and the structured data that passes the verification is stored in the enterprise database to update the enterprise database.

Images