Automatic office data classification method

Abstract

The invention relates to the technical field of office data classification, and discloses an automatic office data classification method, which comprises the following steps: S1, firstly, receiving office data in various forms, S2, enabling a preprocessed document to enter a natural language processing model based on deep learning, S3, carrying out image content containing data, S4, carrying out classification on the image content containing the data in the classification process, and S5, carrying out classification on the image content containing the data. S5, classifying the data through a classification rule according to the operation habit and feedback of the user, S6, dynamically adjusting the classification rule according to the dimension information of the document, and S7, storing the document into a system and then docking with an office system. Different types of office documents can be quickly and accurately classified through preprocessing of the multi-modal data and semantic analysis of the deep learning model, and deep optimization can be performed through a self-supervised learning method regardless of text documents, image documents or table documents, so that the document classification accuracy and processing speed are remarkably improved.

Description

Technical Field

[0001] The present invention relates to the technical field of office document classification, and specifically provides an automated office document classification method. Background Art

[0002] With the rapid development of information technology, the types and formats of office documents have been increasing continuously, covering various forms such as text, images, and tables. Traditional document management and classification methods usually rely on manual operations or only classify through basic technologies such as simple keyword matching and rule engines. However, these methods have obvious limitations. Especially when facing complex and diverse office documents, it is difficult to handle the situation of complex document formats, diverse content, and large amounts of information. Traditional classification methods usually have limited understanding of image content, table data, and document structures, and cannot achieve efficient and accurate automatic classification.

[0003] In recent years, deep learning and natural language processing (NLP) technologies have been widely applied, especially achieving remarkable progress in fields such as text classification, sentiment analysis, and speech recognition. Deep learning models, such as BERT and GPT, have been proven to perform well in understanding complex texts and extracting key information in documents. However, there are still several problems in the existing technologies, especially in the multi-modal processing of documents and classification accuracy. Although many existing systems can effectively classify pure text documents, they still face great challenges in processing image, table, and documents containing mixed data. Summary of the Invention

[0004] Aiming at the deficiencies of the existing technology, the present invention provides an automated office document classification method to solve the problem that most current automated document classification systems only process single-type documents, and for documents containing multiple information forms, such as reports containing images, texts, and tables at the same time, they cannot effectively integrate different modal information, resulting in poor classification effects or slow classification speeds.

[0005] To achieve the above objectives, the present invention is realized through the following technical solutions: An automated office document classification method, including the following steps:

[0006] S1. First, receive various forms of office documents, automatically extract the text in the images through optical character recognition technology, and preprocess the multi-modal data to uniformly convert it into a machine-readable format;

[0007] S2. The preprocessed documents will enter a natural language processing model based on deep learning, and through context semantic analysis, extract the key information and themes of the documents;

[0008] S3. Then, for the content containing data images, the system performs joint analysis on the image content extracted by OCR and the structured data to identify the numerical and text content in the table;

[0009] S4. During the classification process, the system uses self-supervised learning methods to process a large amount of unlabeled data, extract potential features from it, and automatically adjust the classification model;

[0010] S5. Then, according to the user's operation habits and feedback, and through classification rules, the materials are classified. By analyzing the user's behavior, it is understood which document types are the most important to the user, and these document types are preferentially displayed or the classification strategy is adjusted;

[0011] S6. By dynamically adjusting the classification rules based on the document dimension information and automatically adjusting the document classification rules according to seasonal characteristics, the classified documents automatically enter the storage location for archiving management. Labels and metadata are automatically generated based on the document type and classification results, and the documents are stored in the cloud or the local file system;

[0012] S7. After the file is stored in the system, it is connected to the office system. Connecting to the office system for document archiving will immediately trigger relevant approval processes or notify relevant personnel. After passing through the structured office system, it can be used for cross-platform applications through the office system and perform document classification and management on different devices.

[0013] Preferably, in S1, the office materials include plain text files, scanned images, PDF files, and Excel tables. The multi-modal data includes documents, texts, tables, images, and voices. The optical character recognition technology includes image preprocessing, character segmentation, feature extraction, and classification recognition.

[0014] Preferably, in S2, the deep learning includes BERT and GPT. The key information includes entity information, key data, event information, relationship information, and trend information. The topics include core topics, sub-topics, and implicit topics.

[0015] Preferably, in S3, the data image content includes documents of tables and images. The image content extracted by OCR includes text content, symbols, special characters, and text layout information. The structured data includes table content, list data, and metadata.

[0016] Preferably, in S4, the self-supervised learning methods include generative self-supervised learning, contrastive self-supervised learning, and context-based self-supervised learning. The unlabeled data includes original text data, original image data, and document data containing tables. The potential features include semantic features, structural features, and visual features. The classification models include neural network models and decision tree models.

[0017] Preferably, in S5, the classification rules include rules for content features, rules for structural features, and rules for metadata. The user behaviors include document browsing behaviors, search behaviors, annotation and classification behaviors, and feedback behaviors. The document types include text documents, image documents, table documents, multimedia documents, and hybrid documents. The strategies for adjusting classification include priority adjustment strategies, feature weight adjustment strategies, category merging and splitting strategies, and personalized classification strategies. The operation habits include operation frequency habits, operation time habits, operation mode habits, and information acquisition habits.

[0018] Preferably, in S6, the dimension information includes content, time, and source. The seasonal features include time-related features and industry activity features. The classification results include the document category, category hierarchy relationship, and classification confidence. The tags include topic tags, keyword tags, classification tags, and time tags. The metadata includes document basic information, classification-related metadata, tag metadata, and source metadata.

[0019] Preferably, in S7, the office system includes an OA system and an ERP system. The different devices include PCs, mobile phones, and tablets. The notification methods include email, instant messaging, and SMS notifications.

[0020] The present invention provides an automated office document classification method, which has the following beneficial effects:

[0021] 1. Through the preprocessing of multi-modal data and the semantic analysis of the deep learning model, the present invention can quickly and accurately classify different types of office documents. Whether it is a text document, an image document, or a table document, it can be deeply optimized through self-supervised learning methods, thus significantly improving the accuracy and processing speed of document classification.

[0022] 2. By analyzing the operation habits and behavior feedback of users, the present invention can adjust the classification strategy according to the needs of users, preferentially display the document types commonly used by users, or adjust the document classification rules according to the preferences of users. It can dynamically adjust the classification rules according to the time characteristics of the documents and the changes in industry activities, ensuring that the classification method is continuously optimized with the changes in the office environment.

[0023] 3. Through the integration with existing OA systems and ERP systems, the present invention enables the classified documents to be automatically archived and trigger corresponding work processes. Whether on a PC, mobile phone, or tablet device, users can access, manage, and use the classified documents in real time, improving cross-platform office efficiency and achieving the effects of data synchronization and seamless collaboration.

[0024] 4. Through the automatic classification of documents, the system can not only automatically generate topic tags, keyword tags, and classification tags, but also extract the metadata of the documents, including the basic information, source information, and classification-related metadata of the documents. These tags and metadata not only facilitate the retrieval and management of documents, but also automatically create clear indexes when the documents are archived, facilitating subsequent quick access and operation. Description of the Drawings

[0025] Figure 1 It is a method flowchart of an automatic office document classification method of the present invention. Detailed Embodiments

[0026] Next, the technical solutions of the present invention will be clearly and completely described in conjunction with the drawings of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0027] Please refer to the attached Figure 1 , an embodiment of the present invention provides an automatic office document classification method, including the following steps:

[0028] S1. First, receive various forms of office documents, automatically extract the text in the images through optical character recognition technology, and preprocess the multi-modal data to uniformly convert it into a machine-readable format;

[0029] S2. The preprocessed documents will enter a natural language processing model based on deep learning, and through context semantic analysis, extract the key information and topics of the documents;

[0030] S3. Then, for the content containing data images, the system combines the image content extracted by OCR and the structured data for joint analysis to identify the numerical and text content in the tables;

[0031] S4. During the classification process, the system uses a self-supervised learning method to process a large amount of unlabeled data, extract potential features from it, and automatically adjust the classification model;

[0032] S5. Then, classify the documents according to the user's operation habits and feedback and through classification rules. By analyzing the user's behavior, understand which document types are the most important to the user, and preferentially display these document types or adjust the classification strategy;

[0033] S6. Dynamically adjust the classification rules based on the document dimension information and automatically adjust the document classification rules according to seasonal characteristics. After classification, the documents automatically enter the storage location for archival management. Generate labels and metadata automatically based on the document type and classification results, and store the documents in the cloud or the local file system;

[0034] S7. After the files are deposited into the system, they are connected to the office system. When connected to the office system for document archiving, it will immediately trigger relevant approval processes or notify relevant personnel. After passing through the structured office system, it can be applied across platforms through the office system, and document classification and management can be carried out on different devices.

[0035] In S1, office materials include plain text files, scanned images, PDF files, and Excel spreadsheets. Multimodal data includes documents, text, tables, images, and voices. Optical character recognition technology includes image preprocessing, character segmentation, feature extraction, and classification recognition.

[0036] Specifically, first, the system performs image preprocessing on the scanned image files, including image denoising, grayscale conversion, binarization, etc., to improve the image quality and ensure the clarity of the characters. The preprocessed image will enter the character segmentation stage. The OCR system analyzes the character intervals and structures in the image to segment the characters in the image so that each character can be independently recognized. Next, the system extracts features such as the edges, shapes, and sizes of the characters through feature extraction technology, and combines dictionaries and character templates for recognition to ensure the accurate recognition of each character. Finally, the OCR system performs classification recognition, classifies the characters according to the training data and model, extracts the text content in the image, and converts it into editable text;

[0037] For multimodal data, including documents, text, tables, images, and voices, the system will perform different processing according to each data form. PDF files may contain text, images, and tables. The system needs to extract and analyze these parts separately. The text part is directly obtained through text extraction technology, the image part is recognized through OCR technology, and the table part is extracted for structured data through a dedicated table parsing method. When processing materials containing voices, the voice part is converted through voice recognition technology, and the converted text will be further analyzed together with other contents. It can not only recognize traditional printed text but also handle complex contents such as handwritten text, special characters, and mathematical formulas, providing strong technical support for the comprehensiveness and adaptability of the document classification system, effectively combining to provide high-quality input data for subsequent deep learning classification algorithms, and further improving the accuracy and robustness of the entire automated classification method.

[0038] In S2, deep learning includes BERT and GPT. Key information includes entity information, key data, event information, relationship information, and trend information. Themes include core themes, sub-themes, and implicit themes.

[0039] Specifically, first, use the BERT and GPT models to perform deep learning analysis on the text. BERT is a pre-trained language model based on the Transformer architecture. It can capture the complex relationships between words in a sentence through bidirectional context understanding. In document classification tasks, BERT can accurately understand the meaning of sentences and paragraphs and extract entity information (such as names, locations, organizations, etc.) in the document. This information is crucial for the specific classification of the document. GPT, on the other hand, performs text generation and understanding through a generative model and can infer potential associated information from the document.

[0040] Through in-depth understanding of the context, BERT and GPT can accurately distinguish and extract the core themes, sub-themes, and implicit themes in the document, thereby ensuring the accuracy of classification. The system can perform in-depth semantic decomposition of the document and provide multi-dimensional information support for document classification. Finally, after being processed by the BERT and GPT models, the system will generate a classification label for the document based on the key information and themes of the document. The classification of the document no longer depends on simple keyword matching but is based on in-depth understanding and semantic reasoning, making the classification process more intelligent and accurate. This technology can also process various types of complex documents (such as documents containing images, tables, and mixed content), greatly improving the adaptability and efficiency of the automated office document classification method.

[0041] In S3, the data image content includes documents with tables and images. The image content extracted by OCR includes text content, symbols, special characters, and text layout information. Structured data includes table content, list data, and metadata.

[0042] Specifically, first, the system performs OCR processing on the image content. For images embedded in scanned images or PDF files, the system extracts the image content through OCR, converts the text in the image into text. The OCR processing flow includes multiple stages, among which: The text content is the basic target of OCR recognition. The system will identify all visible text content in the image through technologies such as character segmentation and feature extraction, and convert it into editable text. Symbols and special characters are a key link in OCR processing. For some documents that may contain special symbols (such as mathematical symbols, currency symbols, etc.) or special characters (such as tab characters, format markers, etc.), the OCR system will specifically process these symbols and accurately extract them through symbol recognition algorithms. The text layout information is also the content that the OCR system needs to process. OCR not only has to recognize the text in the document, but also retain the original text layout information, including line spacing, paragraph structure, and font format, etc. This is crucial for the subsequent structured parsing and classification of the document, especially in documents with complex layouts (such as legal documents, technical reports, etc.). The layout information can help the system better understand the overall layout and important content of the document. By combining the image content extracted by OCR and table parsing technology, it is possible to comprehensively extract the text, symbols, tables, and list data in the document. This comprehensive processing method is not only applicable to traditional text files, but also can efficiently process complex documents containing tables and images. By analyzing by combining the image content with structured data, the system can provide richer and more accurate features in document classification, thereby improving the classification accuracy.

[0043] In S4, self-supervised learning methods include generative self-supervised learning, contrastive self-supervised learning, and context-based self-supervised learning. The unlabeled data includes original text data, original image data, and document data containing tables. The latent features include semantic features, structural features, and visual features. The classification models include neural network models and decision tree models.

[0044] Specifically, in the image and generative self-supervised learning method, it will attempt to predict and fill in the missing parts based on known information, thereby learning the internal relevance of the data. Generative self-supervised learning is applicable to tasks such as text generation and image generation. In document classification, it helps to extract potential context information and document structure features. The contrastive self-supervised learning method learns by comparing the similarities and differences between different data samples. It mainly optimizes the classification model by constructing positive and negative sample pairs. For document data, positive samples can be documents of similar categories, while negative samples are documents from different categories. By comparing these samples, the model can learn to distinguish the document features of different categories, thereby improving the classification performance. In the text classification task, the system further strengthens the accuracy of the classification model by comparing the semantic differences between documents. The context-based self-supervised learning method learns by utilizing the context information in the document. The model automatically calibrates potential features and learns the deep structure of the data by understanding the relationship between the context information and different parts of the document. In document classification, the learning of context information can help the system understand the paragraphs, sentences and their connections in the document, thereby improving the classification accuracy. For documents containing multiple types of information (such as text, images, tables, etc.), this context-based learning can fully explore the deep semantic features of the data and ensure that the system can handle complex multi-modal documents. The unlabeled data includes original text data, original image data, and document data containing tables. These data are directly input into the model for training without being labeled. The system can extract valuable features from a large number of unlabeled documents and automatically adjust the classification model to adapt to new data types and tasks through self-supervised learning methods. Since this method does not rely on manually labeled data, it greatly reduces the labeling cost. At the same time, it can fully explore potential features in the case of rich data, improving the efficiency and accuracy of document classification.

[0045] S5. The classification rules include rules for content features, rules for structure features, and rules for metadata. User behaviors include document browsing behavior, search behavior, annotation and classification behavior, and feedback behavior. Document types include text documents, image documents, table documents, multimedia documents, and mixed documents. The strategies for adjusting classification include priority adjustment strategies, feature weight adjustment strategies, category merging and splitting strategies, and personalized classification strategies. Operating habits include operating frequency habits, operating time habits, operating method habits, and information acquisition habits.

[0046] Specifically, the classification rules mainly include rules for content features, rules for structural features, and rules for metadata. These rules help the system automatically determine the classification attribution of documents based on different features of the documents. The rules for content features can classify according to keywords, semantic information, etc. of the documents; the rules for structural features classify according to the format and layout (such as paragraphs, tables) of the documents; the metadata rules classify based on meta-information such as the author and creation date of the documents. User behavior plays a key role in the present invention, including document browsing behavior, search behavior, annotation and classification behavior, and feedback behavior. The system analyzes these behaviors to understand the user's preferences for different types of documents, so as to adjust the classification strategy;

[0047] Document types include different types of documents such as text documents, image documents, table documents, multimedia documents, and mixed documents. For different document types, the system adopts different classification rules and processing methods. The adjusted classification strategies include priority adjustment strategy, feature weight adjustment strategy, category merging and splitting strategy, and personalized classification strategy. The priority adjustment strategy adjusts the priority of documents according to factors such as user habits and document importance; the feature weight adjustment strategy automatically adjusts the weights of various features according to the feature types of the documents (such as text, image, table); the category merging and splitting strategy allows the system to merge or split categories according to actual situations to optimize the classification accuracy; the personalized classification strategy is dynamically adjusted based on the user's operation habits to ensure that the classification results meet the user's needs. By combining user behavior, classification rules, and personalized strategies, the present invention can achieve more flexible, intelligent, and user-friendly document classification, improving the efficiency and accuracy of the system.

[0048] In S6, the dimension information includes content, time, and source. The seasonal features include time-related features and industry activity features. The classification results include the category to which the document belongs, the category hierarchy relationship, and the classification confidence level. The tags include topic tags, keyword tags, classification tags, and time tags. The metadata includes document basic information, classification-related metadata, tag metadata, and source metadata.

[0049] Specifically, the dimension information provides a multi-angle understanding of the document, helping the system make classification decisions from different dimensions. The time-related features automatically adjust the classification rules by analyzing the time when the document is published (such as quarters, years, holidays, etc.) to ensure that the system adapts to the needs of different time periods. The industry activity features dynamically adjust the documents according to the periodic activities of the industry (such as earnings report releases, project updates, etc.) to optimize the classification results. The category to which the document belongs refers to the main category assigned to the document; the category hierarchy relationship defines the hierarchical structure between categories; the classification confidence level represents the confidence level of the system in the classification results and is used to quantify the classification accuracy;

[0050] Thematic tags refer to the core themes of documents; keyword tags contain the key terms in the documents; classification tags indicate the specific categories to which the documents belong; time tags are generated based on the time characteristics of the documents, facilitating archiving and management by time, and the basic information of the documents such as document titles, creators, and times, etc.; classification-related metadata helps describe the classification basis of the documents; tag metadata corresponds to the document tags and is used to further optimize the classification accuracy; source metadata includes information such as the source channels and authors of the documents. Through the analysis of the information and features of these dimensions, the system can classify more accurately according to the multiple attributes of the documents, ensuring the flexibility and accuracy of document classification.

[0051] In S7, the office system includes the OA system and the ERP system, different devices include PCs, mobile phones, and tablets, and the notification methods include email, instant messaging, and SMS notifications.

[0052] Specifically, the office system includes the OA system (Office Automation System) and the ERP system (Enterprise Resource Planning System). Through the integration with these office systems, the classified documents can be automatically archived into relevant modules or databases, triggering corresponding workflow processes. By deploying clients or web terminals on various devices, users can view and manage the classified documents on different devices. Employees can view and operate office documents on the PC side, and can also access them through the mobile phone or tablet side when they are out, ensuring access to and use of the latest classified documents anytime and anywhere.

[0053] The notification methods include forms such as email, instant messaging, and SMS notifications. When the document classification is completed or important changes occur, the system can inform relevant personnel in real time through these notification methods. For example, after the financial statements are classified, the system will send notifications via email or instant messaging tools (such as Slack, Teams, etc.) to remind relevant personnel to view or process them. In addition, the system can also inform relevant personnel of documents that need to be urgently processed through SMS notifications, ensuring timely follow-up. Through the integration with office systems, different devices, and multiple notification methods, the automated office document classification method of the present invention not only improves the efficiency of document management, but also ensures cross-platform operation and real-time notification, effectively supporting the daily operations and business processes of enterprises.

[0054] Although the embodiments of the present invention have been shown and described, it will be understood by those of ordinary skill in the art that various changes, modifications, substitutions, and variations can be made to these embodiments without departing from the principles and spirit of the present invention, and the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. An automated office document classification method, characterized in that: The following steps are involved: S1. First, receive office documents in various forms, automatically extract text from images through optical character recognition technology, and pre-process multimodal data and convert them into machine-readable formats; S2. The preprocessed documents will enter the natural language processing model based on deep learning, and extract the key information and topics of the documents through contextual semantic analysis; S3, then the system performs a joint analysis on the image content containing data, combined with the image content extracted by OCR and the structured data, to identify the numerical values ​​and text content in the table; S4. During the classification process, the system uses self-supervised learning methods to process a large amount of unlabeled data, extract potential features from it and automatically adjust the classification model; S5. Then classify the data according to the user's operating habits and feedback and classification rules. By analyzing the user's behavior, understand which document types are most important to the user, and give priority to displaying these document types or adjust the classification strategy; S6. By dynamically adjusting the classification rules based on the document dimension information and automatically adjusting the document classification rules based on seasonal characteristics, the classified documents are automatically stored in the storage location for archiving management, and tags and metadata are automatically generated based on the document type and classification results, and the documents are stored in the cloud or local file system; S7. After the file is stored in the system, it will be connected to the office system. After being connected to the office system for document archiving, it will immediately trigger the relevant approval process or notify relevant personnel. After the office system is structured, it can be used across platforms through the office system to classify and manage documents on different devices.

2. The method for automatic office document classification according to claim 1, characterized in that: In S1, the office materials include plain text files, scanned images, PDF files, and Excel tables, the multimodal data include documents, texts, tables, images, and voices, and the optical character recognition technology includes image preprocessing, character segmentation, feature extraction, and classification recognition.

3. The method for automatic office document classification according to claim 1, characterized in that: In S2, the deep learning includes BERT and GPT, the key information includes entity information, key data, event information, relationship information and trend information, and the topics include core topics, sub-topics and implicit topics.

4. The method for automatic office document classification according to claim 1, characterized in that: In S3, the data image content includes a document of tables and images, the image content extracted by OCR includes text content, symbols and special characters and text layout information, and the structured data includes table content, list data and metadata.

5. The method for automatic office document classification according to claim 1, characterized in that: In S4, the self-supervised learning method includes generative self-supervised learning, contrastive self-supervised learning and context-based self-supervised learning, the unlabeled data includes original text data, original image data and document data containing tables, the potential features include semantic features, structural features and visual features, and the classification model includes a neural network model and a decision tree model.

6. The method for automatic office document classification according to claim 1, characterized in that: S5. The classification rules include rules for content features, rules for structural features and rules for metadata; the user behaviors include document browsing behaviors, search behaviors, annotation and classification behaviors and feedback behaviors; the document types include text documents, image documents, table documents, multimedia documents and mixed documents; the strategies for adjusting classification include priority adjustment strategies, feature weight adjustment strategies, category merging and splitting strategies and personalized classification strategies; the operating habits include operating frequency habits, operating time habits, operating method habits and information acquisition habits.

7. The method for automatic office document classification according to claim 1, characterized in that: In S6, the dimensional information includes content, time, and source; the seasonal characteristics include time-related characteristics and industry activity characteristics; the classification results include the category to which the document belongs, the category hierarchy, and the classification confidence; the tags include subject tags, keyword tags, category tags, and time tags; and the metadata includes basic document information, classification-related metadata, tag metadata, and source metadata.

8. The method for automatic office document classification according to claim 1, characterized in that: In S7, the office system includes an OA system and an ERP system, the different devices include a PC, a mobile phone, and a tablet, and the notification method includes email, instant messaging, and SMS notification.

Images