Localization-based intelligent document question and answer processing platform and system
Through the localized intelligent document question and answer processing platform, the problems of PDF document analysis and sensitive data security are solved, efficient document processing and security model training are realized, and processing speed and office efficiency are improved.
Patent Information
- Application Number
- CN202510609258.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-05-13
- Publication Date
- 2025-08-22
AI Technical Summary
The prior art is difficult to efficiently parse and utilize text and layout information in PDF documents, and there is a lack of security training tools for sensitive documents, resulting in high risk of data leakage, difficult model training and a lot of cost.
Provides an intelligent document question and answer processing platform based on localization, including document annotation module, PDF processing module, model training module, data storage unit and reasoning application module. It uses PyMuPDF and Huggingface's transformers library to realize localized PDF document processing and model training to ensure data security and privacy.
It realizes efficient parsing of text and layout information in PDF documents, significantly improving processing speed, reducing data leakage risks, simplifying operation processes, and improving office efficiency.
Smart Images
Figure CN120523910A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the technical field of natural language processing and document intelligent processing, and specifically provides a localized intelligent document question-and-answer processing platform and system. Background Art
[0002] Amidst the booming business landscape, natural language processing (NLP) models for PDF documents have attracted significant attention but are also facing numerous challenges.
[0003] First, the PDF format itself is highly complex. It is by no means a simple pile of text. The interwoven text and layout information are intricate, and elements such as fonts, font sizes, paragraph layout, and embedded charts are nested with each other. Traditional parsing methods often lose sight of one thing while focusing on another, and are unable to accurately disassemble, integrate, and effectively utilize this information, making it difficult for the model to proceed at the initial stage of data ingestion.
[0004] Secondly, the urgent need for privacy protection has firmly constrained the pace of progress of enterprises. In business activities, various sensitive documents are frequently circulated, from confidential contracts to undisclosed business plans, all of which involve core interests. However, the current market lacks security annotation and training tools specifically for such sensitive information. If you are not careful, the risk of data leakage will follow you everywhere.
[0005] Furthermore, there are obstacles in model training. Even if companies have valuable customized business data and hope to use it to polish a layout perception model that meets their own needs, they find that the existing technical framework is difficult to adapt to these unique data. Parameter tuning is difficult and convergence is slow during training, which consumes a lot of manpower, material resources and time costs, but it is difficult to achieve ideal results. Summary of the Invention
[0006] The purpose of the present invention is to provide a localization-based intelligent document question-answering processing platform and system to solve the problems raised in the above-mentioned background technology.
[0007] To achieve the above objectives, the present invention provides the following technical solutions: a localized intelligent document question-answering processing platform, which includes a document annotation module, a PDF processing module, a model training module, a data storage unit, an inference application module, and a model management module;
[0008] Document annotation module: realizes the functions of uploading and rendering PDF files, inputting questions, highlighting answer texts, and automatically saving layout information;
[0009] PDF processing module: involves uploading, rendering and related processing of PDF files;
[0010] Model training module: provides training data selection and management, local model training, and weight storage and version control functions;
[0011] Data storage unit: uses a local SQL database to store annotation data and model weights, and provides document session management functions;
[0012] Reasoning application module: includes model selection interface, batch document processing and result visualization functions;
[0013] Model management module: covers model training, weight preservation, version control, and model selection during inference.
[0014] Preferably, the document annotation module includes:
[0015] (1) PDF file upload and rendering: users can upload PDF files, and the system uses PDF.js to render the document on the front end;
[0016] (2) Question input interface, where users input questions related to the document and the answer text is highlighted: users highlight text in the PDF as the answer, and the system automatically saves the text and layout information;
[0017] (3) Layout information is automatically saved. PyMuPDF is used to extract and save layout information on the backend to support the training of layout-aware models.
[0018] Preferably, the PDF processing module refers to accurately rendering the PDF files uploaded by users, restoring the original appearance of the documents, creating an intuitive and convenient interactive interface for users, and making the reading and operation experience smooth and easy. At the same time, the back-end uses PyMuPDF to operate efficiently and deeply mine the text content and complex layout information in the PDF, which not only realizes the accurate extraction of text, but also lays a solid foundation for subsequent answer annotation, model training and other processes, ensuring close coordination of all links in the system.
[0019] Preferably, the model training module includes:
[0020] (1) Training data selection and management: users can select which annotations to use as training data and manage a large number of PDF documents;
[0021] (2) Local model training: integrating Huggingface's transformers library to support model training on local devices and ensure data privacy;
[0022] (3) Weight preservation and version control. After training is completed, the model weights are saved in the local SQL database, which supports version control and backtracking.
[0023] Preferably, the data storage unit includes:
[0024] (1) Data storage: Using a local SQL database, we create dedicated storage space to store all kinds of annotation data generated by users during the document annotation process, as well as the weight information after model training, completely and securely, ensuring data traceability and stability;
[0025] (2) Session and version management: Build a document session management system to record user operation sessions for different documents on the one hand, and to achieve fine-grained version control of documents on the other hand. It also supports querying historical records at any time, making it convenient for users to trace back the modification history of documents and compare the differences between different versions, thus ensuring efficient and orderly document management in all aspects.
[0026] Preferably, the reasoning application module includes:
[0027] (1) Model selection interface, where users can select a trained model for inference;
[0028] (2) Batch document processing: supports batch uploading of documents for prediction, and the system automatically processes and generates results;
[0029] (3) Visual display of results: the inference results are highlighted in the PDF, helping users quickly locate the answer and its context.
[0030] Preferably, the model management module refers to the core control unit of the entire system, which shoulders multiple key responsibilities. During the model training phase, it integrates advanced technologies, allowing users to filter accurate training data according to their needs, call local resources and combine Huggingface's transformers library to start the training process, and ensure the professionalism and pertinence of the model; after the training is completed, the weight preservation mechanism is immediately activated, and the results are securely stored in the local SQL database. At the same time, version control is strictly implemented to record each iterative optimization of the model; and in the inference application, the module also provides a model selection interface to help users select the best adaptation item from a large number of local models, and efficiently drive batch document processing and result output.
[0031] Preferably, based on the above platform, the system consists of a document annotation module, a PDF processing module, a model training module, a data storage unit, an inference application module, and a model management module.
[0032] The beneficial effects of the present invention are as follows:
[0033] 1. This invention introduces automated processes through PDF document processing technology, which is like installing a high-speed engine in traditional document processing methods. In the past, manual processing of a complex document may take several hours, but now with its powerful automation function, it can be completed in just tens of minutes or even minutes, and the processing speed is increased by more than seven times. Whether it is massive contract review, technical data compilation, or market research report analysis, it can respond quickly, greatly shortening the business cycle and allowing enterprises to seize the initiative in fierce competition.
[0034] 2. This invention limits all annotation, training, and reasoning steps to the local environment. This means that the risk of sensitive corporate information such as contract terms, customer information, and R&D data being leaked to external networks is greatly reduced. This is completely different from the traditional model that relies on cloud processing and is vulnerable to attacks. It builds a solid privacy defense line for enterprises, allowing them to focus on business expansion without worrying about data leaks.
[0035] 3. The present invention uses this technology to thoughtfully create a unified user interface. No matter whether it is a novice or an experienced employee, they can quickly get started. It abandons the tedious multi-system switching and complex command input, and cleverly integrates the originally scattered functional modules. From document uploading and problem marking to model training and result viewing, it simplifies everything. Employees do not need to spend a lot of time learning professional skills and can easily master it. This greatly improves the fluency of daily office work and reduces efficiency loss caused by inconvenient operation. BRIEF DESCRIPTION OF THE DRAWINGS
[0036] Figure 1 This is a diagram of the localized intelligent document question-answering processing system of the present invention. DETAILED DESCRIPTION
[0037] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.
[0038] like Figure 1 As shown, an embodiment of the present invention provides a localized intelligent document question-answering processing platform, which includes a document annotation module, a PDF processing module, a model training module, a data storage unit, an inference application module, and a model management module;
[0039] Document annotation module: realizes the functions of uploading and rendering PDF files, inputting questions, highlighting answer texts, and automatically saving layout information;
[0040] PDF processing module: involves uploading, rendering and related processing of PDF files;
[0041] Model training module: provides training data selection and management, local model training, and weight storage and version control functions;
[0042] Data storage unit: uses a local SQL database to store annotation data and model weights, and provides document session management functions;
[0043] Reasoning application module: includes model selection interface, batch document processing and result visualization functions;
[0044] Model management module: covers model training, weight preservation, version control, and model selection during inference.
[0045] The document annotation module serves as the basis for front-end interaction and supports PDF uploading and rendering, making it easy for users to enter questions and mark answers. It also automatically saves layout information to facilitate subsequent analysis; the PDF processing module focuses on basic file operations to ensure the quality of uploading and rendering; the model training module gives users the right to choose independently and can train models locally. After completion, the weights are stored in the local SQL database using the data storage unit and the version is managed; the inference application module allows users to select models on demand, batch process documents and visualize the results; the model management module coordinates all aspects from training to inference, stringing together the entire process. The modules work closely together to empower PDF document processing.
[0046] Wherein, the document annotation module includes:
[0047] (1) PDF file upload and rendering: users can upload PDF files, and the system uses PDF.js to render the document on the front end;
[0048] (2) Question input interface, where users input questions related to the document and the answer text is highlighted: users highlight text in the PDF as the answer, and the system automatically saves the text and layout information;
[0049] (3) Layout information is automatically saved. PyMuPDF is used to extract and save layout information on the backend to support the training of layout-aware models.
[0050]
[0051] The first step is PDF file uploading and rendering. After the user easily uploads the file, the system uses the powerful PDF.js to perfectly present the document on the front end, and the reading experience is smooth and intuitive. Then, the question input interface appears. The user enters questions about the document content and can also highlight text in the PDF as answers. At this time, the system not only automatically retains the text, but also uses PyMuPDF to extract layout information on the back end, laying a solid foundation for subsequent layout perception model training.
[0052] Among them, the PDF processing module refers to accurately rendering the PDF files uploaded by users, restoring the original appearance of the documents, creating an intuitive and convenient interactive interface for users, and making the reading and operation experience smooth and easy. At the same time, the back-end uses PyMuPDF to operate efficiently and deeply mine the text content and complex layout information in the PDF, which not only realizes the accurate extraction of text, but also lays a solid foundation for subsequent answer annotation, model training and other processes, ensuring close coordination of all aspects of the system.
[0053]
[0054] Wherein, the model training module includes:
[0055] (1) Training data selection and management: users can select which annotations to use as training data and manage a large number of PDF documents;
[0056] (2) Local model training: integrating Huggingface's transformers library to support model training on local devices and ensure data privacy;
[0057] (3) Weight preservation and version control. After training is completed, the model weights are saved in the local SQL database, which supports version control and backtracking.
[0058]
[0059] First, it accurately selects appropriate annotations as training materials while easily managing massive PDF documents. Second, local model training is uniquely integrated with the advanced Huggingface transformers library, eliminating the need for users to worry about privacy leaks. Third, weight preservation and version control safeguard the continuous optimization of the model. After training is completed, the model weights are properly stored in the local SQL database. With version control and backtracking functions, the model's growth trajectory can be traced at any time.
[0060] Wherein, the data storage unit includes:
[0061] (1) Data storage: Using a local SQL database, we create dedicated storage space to store all kinds of annotation data generated by users during the document annotation process, as well as the weight information after model training, completely and securely, ensuring data traceability and stability;
[0062] (2) Session and version management: Build a document session management system to record user operation sessions for different documents on the one hand, and to achieve fine-grained version control of documents on the other hand. It also supports querying historical records at any time, making it convenient for users to trace back the modification history of documents and compare the differences between different versions, thus ensuring efficient and orderly document management in all aspects.
[0063] The local SQL database plays a pivotal role. For data storage, it meticulously creates a dedicated space. Whether it's multi-dimensional annotation data derived from document annotation or key weight intelligence after model training, everything can be properly stored. Its robust nature ensures that data is traceable, stable, and reliable, laying a solid foundation for subsequent analysis and optimization. Session and version management are equally important. The system built acts like an intelligent butler, accurately recording every interaction between users and documents, meticulously controlling document version changes, and allowing historical records to be retrieved at any time, making it easy to trace document changes and identify the pros and cons of versions, driving document management to new heights.
[0064] The reasoning application module includes:
[0065] (1) Model selection interface, where users can select a trained model for inference;
[0066] (2) Batch document processing: supports batch uploading of documents for prediction, and the system automatically processes and generates results;
[0067] (3) Visual display of results: the inference results are highlighted in the PDF, helping users quickly locate the answer and its context.
[0068] First, the model selection interface is like a master key, giving users the right to independently select well-trained models so that they can be accurately matched to different tasks; then, the batch document processing function opens a high-speed channel and supports batch uploading of documents. The system runs quickly like a precision machine, automatically predicting and outputting results, greatly improving efficiency; finally, the visual display of results is the finishing touch. The inference results are presented in a highlighted form in the PDF, like a guiding light in the night sky, helping users instantly lock in the answer and its related context, making information acquisition possible in one go.
[0069] Among them, the model management module refers to the core control unit of the entire system, shouldering multiple key responsibilities. During the model training phase, it integrates advanced technologies, allowing users to filter accurate training data according to needs, call local resources and combine Huggingface's transformers library to start the training process, and ensure the professionalism and pertinence of the model; after the training is completed, the weight preservation mechanism is immediately activated, and the results are securely stored in the local SQL database. At the same time, version control is strictly implemented to record each iterative optimization of the model; and in the inference application, the module also provides a model selection interface to help users select the best adaptation item from many local models, and efficiently drive batch document processing and result output.
[0070] The model management module serves as the core of the system, controlling the overall situation. During model training, it integrates cutting-edge technologies, filters data based on user needs, and uses local resources in conjunction with the transformers library to build professional models. After training, weights are quickly stored in the local SQL library, and refined version control records iterations. During the inference phase, a model selection interface is provided to help users choose the appropriate model, promote batch document processing, and ensure efficient operation.
[0071] Among them, based on the above platform, the system consists of a document annotation module, a PDF processing module, a model training module, a data storage unit, an inference application module, and a model management module.
[0072] It should be noted that, in this document, relational terms such as first and second, etc., are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the terms "comprises," "comprising," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that includes a list of elements includes not only those elements but also other elements not explicitly listed, or elements inherent to such process, method, article, or apparatus.
[0073] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.
Claims
1. A localized intelligent document question-answering processing platform, characterized by: The platform consists of a document annotation module, a PDF processing module, a model training module, a data storage unit, an inference application module, and a model management module; Document annotation module: realizes the functions of uploading and rendering PDF files, inputting questions, highlighting answer texts, and automatically saving layout information; PDF processing module: involves uploading, rendering and related processing of PDF files; Model training module: provides training data selection and management, local model training, and weight storage and version control functions; Data storage unit: uses a local SQL database to store annotation data and model weights, and provides document session management functions; Reasoning application module: includes model selection interface, batch document processing and result visualization functions; Model management module: covers model training, weight preservation, version control, and model selection during inference.
2. The localized intelligent document question-answering processing platform according to claim 1, characterized in that: The document annotation module includes: (1) PDF file upload and rendering: users can upload PDF files, and the system uses PDF.js to render the document on the front end; (2) Question input interface: users input questions related to the document, and the answer text is highlighted: users highlight text as answers in the PDF, and the system automatically saves the text and layout information. The answer annotation accuracy is (3) Layout information is automatically saved. PyMuPDF is used to extract and save layout information in the backend to support the training of layout-aware models. The layout feature number F of PyMuPDF is train―layout =F layout ×p eff―layout Provide strong support for model training.
3. The localized intelligent document question-answering processing platform according to claim 1, characterized in that: The PDF processing module refers to accurately rendering the PDF files uploaded by users, restoring the original appearance of the documents, creating an intuitive and convenient interactive interface for users, and making the reading and operation experience smooth and easy. At the same time, the back-end uses PyMuPDF to operate efficiently and deeply mine the text content and complex layout information in the PDF, which not only realizes the accurate extraction of text, but also lays a solid foundation for the subsequent answer annotation and model training processes, ensuring close coordination of all aspects of the system.
4. The localized intelligent document question-answering processing platform according to claim 1, characterized in that: The model training module includes: (1) Training data selection and management. Users select annotations as training data and manage a large number of PDF documents. The number of samples participating in the training is N, the feature dimension of each sample is F, and the number of model parameters is P. During the model training process, taking a simple linear regression model as an example, the loss function L is calculated during the training process. The specific formula is: where y i is a true value and w j is the weight parameter, b is the bias parameter, and the weight parameter w j It is related to the number of model parameters P and is continuously updated during training to minimize the loss function L. (2) Local model training, integrating Huggingface's transformers library to support model training on local devices and ensure data privacy; (3) Weight preservation and version control. After training is completed, the model weights are saved in the local SQL database, which supports version control and backtracking.
5. The localization-based intelligent document question-answering processing platform according to claim 1, characterized in that: The data storage unit includes: (1) Data storage: Using a local SQL database, we create dedicated storage space to store all kinds of annotation data generated by users during the document annotation process, as well as the weight information after model training, completely and securely, ensuring data traceability and stability; (2) Session and version management: Build a document session management system to record user operation sessions for different documents on the one hand, and to achieve fine-grained version control of documents on the other hand. It also supports querying historical records at any time, making it convenient for users to trace back the modification history of documents and compare the differences between different versions, thus ensuring efficient and orderly document management in all aspects.
6. The localization-based intelligent document question-answering processing platform according to claim 1, characterized in that: The reasoning application module includes: (1) Model selection interface, where users select a trained model for inference; (2) Batch document processing: supports batch uploading of documents for prediction, and the system automatically processes and generates results; (3) Visual display of results: the inference results are highlighted in the PDF, helping users quickly locate the answer and its context.
7. The localization-based intelligent document question-answering processing platform according to claim 1, characterized in that: The model management module is the core control unit of the entire system, shouldering multiple key responsibilities. During the model training phase, it integrates advanced technologies, allowing users to filter precise training data based on their needs, call local resources and combine them with Huggingface's transformers library to start the training process, ensuring the professionalism and pertinence of the model. After training, the weight preservation mechanism is immediately activated to securely store the results in the local SQL database. At the same time, version control is strictly implemented to record every iterative optimization of the model. During inference applications, the module thoughtfully provides a model selection interface to help users select the best adaptation option from a variety of local models, efficiently driving batch document processing and result output.
8. The localization-based intelligent document question-answering processing system according to any one of claims 1 to 7, characterized in that: Based on the above platform, the system consists of document annotation module, PDF processing module, model training module, data storage unit, reasoning application module, and model management module. The overall system formula is S user =αE annotation +βA pdf +γE train +δS storag .