An archive data recognition processing method and system based on artificial intelligence

Abstract

The present application relates to an artificial intelligence-based archive data recognition processing method and system, which avoids the technical problem that the AI large model technical advantage cannot be fully utilized in the process of archive digital application due to the recognition deviation of the original archive. The method comprises: standardizing the original archive image to form a training sample, and balancing the sample according to the training purpose; training and optimizing the basic large model based on the Transformer architecture through the training sample to form an archive data processing model; performing content recognition and semantic analysis on batch archive images through the archive data processing model to form a knowledge graph of archive entities and construct an archive database. The automatic labeling and entity association of archive topics are completed, batch processing and automatic analysis of massive archive data are realized, a structured database of household archives and an archive information knowledge graph are constructed, thereby realizing the deep development and sharing foundation of archive resources, and meeting the urgent needs of information sharing and business automation.

Description

Technical Field

[0001] This invention relates to the field of archival processing technology, specifically to an artificial intelligence-based method and system for archival data recognition and processing. Background Technology

[0002] A vast amount of historical information has been recorded and preserved as original archival materials. With the rapid development of information technology, archives have gradually encompassed various forms such as text, images, audio, and video, highlighting the increasing importance of archival management. Traditional archival management mainly relies on manual operation and simple computer assistance. Most practices involve scanning or photographing paper archives to create images, converting them into electronic documents. This means that computers can only recognize the format of the archives but do not understand the content information they contain, making it difficult to effectively develop intelligent collection and archiving capabilities for textual data in terms of entity recognition and relationship extraction. For example, due to the long history of paper archives, recording formats, writing styles, semantic meanings, and recording traces have changed over time. Current OCR technology and natural language processing models often struggle to accurately locate and analyze key information when processing government documents, failing to fully support richer data mining and utilization. With the continuous increase in the number and variety of archives, it is gradually becoming insufficient to meet the demands of modern society for efficient and accurate archival management.

[0003] Existing large language models (LLMs, such as GPT-4, LLaMA 3, and QWen) rely on the self-attention mechanism of the Transformer architecture to capture relevance features in data, focusing on language (text) understanding and generation. They can be used for intelligent processing and analysis of archival data, improving the efficiency and accuracy of archival information recognition. However, due to the complexity of the information presentation in the original archival documents, large language models cannot directly construct structured data and information knowledge graphs. Summary of the Invention

[0004] In view of the above problems, embodiments of the present invention provide an artificial intelligence-based method and system for identifying and processing archival data, avoiding the technical problem that the advantages of AI large model technology cannot be fully utilized in the process of archival digitalization applications due to the identification deviation of the original archives.

[0005] The artificial intelligence-based archival data recognition and processing method of this invention includes: Standardize the original archival images to form training samples, and balance the samples according to the training objective. The archival data processing model is formed by training and optimizing a large-scale model based on the Transformer architecture using training samples. By using an archival data processing model to perform content recognition and semantic analysis on batch archival images, a knowledge graph of archival entities is formed, and an archival database is constructed.

[0006] In one embodiment of the present invention, the standardization processing of the original archival images to form training samples, and the sample equalization according to the training objective, include: Retrieve text and image data from archive images; Text data is processed to standardize words, forming an embedded vocabulary set; Image data is enhanced and normalized to form an enhanced image set; Archival images are cropped at different scales and regions to create archival image slices; Training samples are generated by matching words from the embedded vocabulary set and images from the augmented image set with archival image slices. Adjust the proportion of training samples in the archival image slices for coverage size, connected regions, and text relevance, according to the training objectives.

[0007] In one embodiment of the present invention, the step of training and optimizing a basic large model based on the Transformer architecture using training samples to form an archival data processing model includes: The training set is formed based on the training samples, and the archival data processing model based on the Transformer architecture is iteratively trained. A convergence threshold is set, and training is stopped through an early stopping mechanism. Establish a training error measurement system, quantify the model accuracy based on the measurement indicators, and optimize the parameters based on the accuracy. The collaborative task training error of the model is quantified based on the metric, and the collaborative training error is gradually satisfied through collaborative training.

[0008] In one embodiment of the present invention, the basic large model based on the Transformer architecture adopts the QWEN model.

[0009] In one embodiment of the present invention, the step of performing content recognition and semantic analysis on batch archival images through an archival data processing model to form a knowledge graph of archival entities and construct an archival database includes: By using an archival data processing model, feature extraction and semantic analysis are performed on batch input archival images to identify core thematic words and entity words in the archives, and thematic labeling and entity association are completed. Based on the semantic analysis results of the archive content, a knowledge graph is formed within and between archive images; The knowledge graph forms the structured storage of archival data.

[0010] In one embodiment of the present invention, it further includes: The data processing model enables automated management of archival data.

[0011] In one embodiment of the present invention, it further includes: By extracting fixed layouts from batches of original archival images using an image recognition model based on LSTM architecture, layout association features of the archives are established, and training samples are formed by combining the standardized processing of the original archives.

[0012] In one embodiment of the present invention, the image recognition model based on the LSTM architecture adopts the ConvLSTM model.

[0013] The artificial intelligence-based archival data recognition and processing system of this invention includes: The training sample construction device is used to standardize the original archival images to form training samples and perform sample balancing according to the training objective. The model training and optimization device is used to train and optimize the Transformer architecture model using training samples to form an archive data processing model; The data recognition application device is used to perform content recognition and semantic analysis of batch archival images through an archival data processing model, form a knowledge graph of archival entities, and construct an archival database.

[0014] In one embodiment of the present invention, it further includes: The archive layout processing module is used to extract fixed layouts from batches of original archive images using an image recognition model based on an LSTM architecture, establish layout association features of the archives, and combine them with the standardized processing of the original archives to form training samples.

[0015] The artificial intelligence-based archival data identification and processing method and system of this invention rely on a large language model to process massive amounts of archival data in batches, complete the automatic labeling of archival topics and entity association, realize the batch processing and automated analysis of massive amounts of archival data, construct a structured database of household registration archives and an archival information knowledge graph, thereby realizing the foundation for in-depth development and sharing of archival resources, and meeting the urgent needs of archival researchers for information sharing and business automation. Attached Figure Description

[0016] Figure 1 The diagram shown is a flowchart of an artificial intelligence-based archival data recognition and processing method according to an embodiment of the present invention.

[0017] Figure 2 The diagram shows an architecture of an artificial intelligence-based archival data recognition and processing system according to an embodiment of the present invention.

[0018] Figure 3 The diagram shown is a schematic diagram of the architecture of an electronic device according to an embodiment of the present invention.

[0019] Figure 4The diagram shown is a flowchart of another embodiment of the present invention, which is a method for identifying and processing archival data based on artificial intelligence.

[0020] Figure 5 The diagram shows the architecture of an artificial intelligence-based archival data recognition and processing system according to one embodiment and another embodiment of the present invention. Detailed Implementation

[0021] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided below.

[0022] An embodiment of the present invention provides an artificial intelligence-based method for identifying and processing archival data, as follows: Figure 1 As shown. In Figure 1 In this embodiment, the following are included: Step 200: Standardize the original archive images to form training samples, and balance the samples according to the training objective.

[0023] Standardization processing includes OCR text recognition and image enhancement of the original archival images. By appropriately cropping the images, balanced training samples are generated that highlight key information recognition, core character recognition, and text segmentation.

[0024] Step 300: Train and optimize the basic large model based on the Transformer architecture using training samples to form an archive data processing model.

[0025] Large language models such as Qwen, GPT, and BERT are based on the Transformer architecture, enabling parallel computing and self-attention mechanisms to achieve processing capabilities in natural language processing, multimodal interaction, and complex semantic understanding. Iterative training of the models using training samples allows for parameter tuning, resulting in model optimization specifically for the archival data recognition process.

[0026] Step 400: Perform content recognition and semantic analysis on batch archival images using an archival data processing model to form a knowledge graph of archival entities and construct an archival database.

[0027] Text and structure recognition are performed on the image files of the original archives. Combining text content with the archive layout, structured data mapping to the archive images is formed. Logical relationships between text entities are established through semantic recognition, creating a knowledge graph of relationships between text content. This ultimately forms a relational database of archival data resources.

[0028] The artificial intelligence-based archival data recognition and processing method of this invention enhances the feature representation of training samples for original archival entities, making the training of the large language model more targeted and avoiding the optimization of the large language model from failing to achieve its intended purpose. By appropriately fusing the recognition results of the large language model with the image, an archival data association and data knowledge graph based on the archival image are formed. The constructed related database enables in-depth development and sharing of archival resources, meeting the urgent need for information sharing among archival researchers.

[0029] like Figure 1 As shown, in one embodiment of the present invention, step 200 includes: Step 210: Obtain the text data and image data from the archive image.

[0030] The original archival entity is scanned using an image scanning device to obtain archival images. Text data from these images is then acquired using OCR technology; for example, the open-source OCR-Tesseract library is used to perform OCR recognition on the image data to extract the text content. The recognition accuracy is set to PSM (Page Segmentation Mode) 3 to obtain the text content. Image recognition technology is then used to quantify the graphics within the archival images, locating image data such as markings, tables, and seals.

[0031] Step 220: Perform word standardization on the text data to form an embedded vocabulary set.

[0032] Word standardization includes, but is not limited to, using regular expressions to remove stop words, punctuation marks, and numbers from the text; using Chinese spelling rules to convert English characters into their corresponding Chinese characters; and standardizing the titles by unifying all title levels. The standardized characters, words, and phrases are then sorted to form an embedded vocabulary set for the archival images.

[0033] Step 230: Perform enhancement and normalization processing on the image data to form an enhanced image set.

[0034] Enhanced standardization processing includes, but is not limited to, improving image sharpness and quality; for blurry handwritten text, handwriting recognition algorithms are used for identification and sharpening. The images generated after each processing step collectively constitute the enhanced image set of the archival images.

[0035] Step 240: Crop the archival image at different scales and in different regions to form archival image slices.

[0036] Archival image slicing includes cropping of areas containing text, meaningful patterns, tables, etc. It also includes cropping of related areas, such as table headers and table areas, non-adjacent but related independent areas, and several independent areas with a causal relationship.

[0037] Step 250: Match the words in the embedded vocabulary set and the images in the augmented image set with the archival image slices to form training samples.

[0038] The vocabulary and images within the image slice region are combined to form a training sample for training a large language model. A series of training samples are formed based on different image slices.

[0039] Step 260: Adjust the proportion of training samples in the archive image slices for coverage size, connected regions, and text relevance according to the training objective.

[0040] Archival image slices are formed based on information elements such as keywords in text data, core graphics or patterns in image data, and the concentration of word segmentation frequency in text data. Training samples are generated using parameters such as coverage size, connected regions, and text relevance. To address the differences in the number of training samples for each information element, upsampling or downsampling of different types of training samples is used to achieve training sample balance for different information elements.

[0041] The artificial intelligence-based archival data recognition and processing method of this invention provides balanced samples for the training of a large language model by constructing training samples for archival text features and image features. This makes the training process of related sub-tasks avoid underfitting or overfitting as much as possible, and avoids the inability to converge due to task errors caused by data quality.

[0042] like Figure 1 As shown, in one embodiment of the present invention, step 300 includes: Step 310: Form a training set based on the training samples and iteratively train the archive data processing model based on the Transformer architecture. Set a convergence threshold and stop training through an early stopping mechanism.

[0043] The training sample data is randomly shuffled and then divided into two sets. Based on the amount of data and the model type, 80%-90% of the data is selected as the training set and 20%-10% of the data is selected as the test set.

[0044] To achieve the purpose of archival data identification and extraction in this invention, the large model based on the Transformer architecture uses either the QWen model or the BERT (Bidirectional Encoder Representations from Transformers) model as the base model to construct a deep neural network that includes feature extraction and semantic recognition. After full training and fine-tuning, the parameters of the base model are optimized to adapt to the characteristics of the archival data. In one embodiment of this invention, the model parameters of the QWen model are optimized using the Adam optimization algorithm. During training, an early stopping mechanism is used, stopping training when the model converges within 10 epochs. In another embodiment of this invention, the model parameters are optimized using the SGD (Stochastic Gradient Descent) optimization algorithm. During training, an early stopping mechanism is used, stopping training when the model converges within 20 epochs.

[0045] In one embodiment of the present invention, the reinforcement learning and iterative training process includes: 1. Optimization of Key Information Localization Based on Reinforcement Learning By constructing an intelligent collaborative decision-making system and employing reinforcement learning (RL) and real-time optimization algorithms, the model can dynamically adjust its information extraction strategy when processing documents and automatically learn the best location path for documents of different formats.

[0046] The specific implementation includes: Design a reward function: give a positive reward when the model correctly identifies the key field, and give a negative reward when it misidentifies it; State space definition: includes the current text position being processed, identified information, document structure features, etc. Action space design: determines the direction of the next step (such as continuing to read, extracting information, jumping to a specific area, etc.).

[0047] 2. Iterative training improves semantic understanding capabilities Through continuous iterative training, the model's ability to understand government terminology and complex expressions is constantly optimized.

[0048] Through an active learning mechanism, the model identifies uncertain samples, which are then manually labeled and incorporated into the training set, thereby improving the identification capabilities in challenging areas.

[0049] By using adversarial training to generate diverse text variants, the model's robustness to format changes is enhanced.

[0050] Domain-adaptive training: Targeted fine-tuning is performed based on the document characteristics of different organizations and units to improve the accuracy of extracting department-specific information.

[0051] 3. Multimodal fusion and iterative optimization For scanned documents and handwritten documents, the OCR recognition results are combined with the original image features, and the optimization is carried out through multiple rounds of iteration.

[0052] The specific implementation includes: Round 1: Basic OCR Recognition and Text Extraction Round Two: Evaluation of the Reliability of Reinforcement Learning Model Results Third round: Focused analysis and correction on areas with low confidence levels. Continuous iteration: Feeding the corrected results back to the model to continuously improve recognition accuracy. By comprehensively applying the aforementioned reinforcement learning and iterative training techniques, we can continuously optimize model algorithms, enhance learning capabilities and information extraction capabilities, effectively solve the challenges of locating and analyzing key information, and provide technical support for the digital processing of archives.

[0053] Step 320: Establish a training error measurement system, quantify the model accuracy based on the measurement indicators, and optimize the parameters based on the accuracy.

[0054] To address the characteristics of the Transformer architecture, a training error metric system is used as feedback signals. Error changes are quantified based on these metrics, dynamically optimizing the learning rate, optimizer parameters, and gradient processing strategies. Weights and biases are updated through backpropagation of gradient descent, supplemented by techniques such as warm-up, decay, and gradient clipping to continuously optimize model performance. The metrics in the training error metric system include, but are not limited to, key information recognition metrics, core character recognition metrics, and text segmentation metrics. Key information recognition metrics calculate accuracy, core character recognition metrics calculate recognition accuracy, and text segmentation metrics calculate text segmentation accuracy.

[0055] In one embodiment of the present invention, the key information identification measurement system includes the following measurement indicators: Field-level F1 scores are calculated for key fields in a document (such as document number, date, and household head). Semantic consistency score assesses the degree of semantic matching between the extracted content and the standard answer; Position deviation measures the degree of offset between the identified key information and the actual position (pixel level or character level).

[0056] A specific implementation example is used, with document number recognition as the test sample: Standard document number: Chaoyang

[2023] No. 15 Model recognition result: Chaoyang

[2023] No. 15 Its iterative optimization strategy is as follows: When the semantic consistency score < 0.95, enhance the bracket normalization processing module; When the position deviation > 10px, adjust the attention mechanism weight of the document layout analysis model; For high-frequency error characters, add adversarial sample training.

[0057] In an embodiment of the present invention, the metric indicators included in the core character recognition metric system: Character Error Rate (CER): (S + D + I) / N, where S = substitution error, D = deletion error, I = insertion error, N = standard text length; Recognition rate of easily confused character groups: separately count for specific easily confused characters (such as "申 / 审", "已 / 己", "政 / 攻"); Confidence-Accuracy Curve (CAC): evaluate the correspondence between the model's prediction confidence and the actual accuracy; Specific implementation examples, using historical archive handwritten text recognition as test samples: Standard text: "申请审批" OCR recognition result: "审请审批" Model confidence: [0.85, 0.72, 0.92, 0.88, 0.91] Its iterative optimization strategy is: For the "申 / 审" confusion problem, generate 1000 groups of adversarial samples and add them to the training set; When the low-confidence error rate > 30%, adjust the model's dropout rate from 0.1 to 0.15; Construct a character-level attention mechanism and increase the weight of error-prone character regions by 2 times.

[0058] In an embodiment of the present invention, the metric indicators included in the text segmentation metric system: Boundary detection F1 value: evaluate the recognition accuracy of structural boundaries such as paragraphs / tables / headings.

[0059] Semantic coherence score: evaluate the semantic integrity of the segmented text through a pre-trained language model.

[0060] Cross-page connection accuracy: measure the connection correctness when continuous content crosses pages.

[0061] Specific implementation examples, using file table segmentation as the test scenario: Table segmentation and recognition in a certain archive file Its iterative optimization strategy: When boundary_f1 < 0.85, increase the number of convolutional kernels of the table line detection module; When semantic coherence is less than 0.8, a text post-processing rule engine is introduced for structural repair. To address column merging errors, a dedicated column splitting loss function is designed: L_split = α·column_error +β·text_overflow.

[0062] Step 330: Quantify the collaborative task training error of the model according to the metric, and gradually meet the collaborative training error metric through collaborative training.

[0063] Key information recognition, core character recognition, and text segmentation are three interconnected sub-tasks of the model, and they influence each other. In one embodiment of this invention, collaborative training adopts an architecture of a shared encoder and a task-specific decoder. The shared encoder converts the input text into context-aware semantic features, the key information recognition head outputs entity labels, the core character recognition head outputs the probability of each character being a core character, and the text segmentation head outputs the judgment of whether each position is a segmentation boundary. By monitoring the training error of a single task and the comprehensive error of collaborative training, the parameters of the shared encoder are jointly driven, while ensuring the performance balance of each sub-task and achieving positive guidance between sub-tasks. By sharing the model encoder parameters, the total number of parameters is reduced by more than 60% compared to training three independent models, effectively reducing training costs. The general semantic features learned by the shared encoder allow the model to maintain good performance even in small sample scenarios. While fully utilizing the semantic modeling capabilities of the model, collaborative training solves the problems of overfitting and poor generalization ability of single-task training, making it suitable for multi-task scenarios such as text semantic understanding.

[0064] In one embodiment of the present invention, a specific multi-task collaborative training architecture includes: (1) Feature sharing layer, including: Unified Feature Encoder: A Transformer-based multimodal encoder that processes both image and text features.

[0065] Dynamic gating mechanism: Adaptively selects feature channels based on task requirements.

[0066] (2) Task interaction layer, used to form a two-way information flow mechanism, establish forward and feedback channels between tasks, and use attention-guided information transmission.

[0067] (3) Consistency Constraint Layer Cross-task consistency verification module: Ensures that the output logic of each task is consistent.

[0068] In one embodiment of the present invention, the specific multi-stage collaborative training strategy for multi-task collaborative training is as follows: Phase 1: Pre-training and initialization. Each task independently pre-trains the basic model, establishes the initial task dependency graph, and sets the initial collaborative weights: α=β=γ=δ=0.2, λ=0.2.

[0069] Phase 2: Progressive Collaborative Training.

[0070] Phase 3: Error-driven collaborative optimization. When the error of a task is significantly higher than that of other tasks, strengthen its connection with auxiliary tasks; when the consistency between tasks is low, increase the consistency constraint weight λ.

[0071] The artificial intelligence-based archival data recognition and processing method of this invention simplifies model construction complexity and improves training efficiency through an error measurement system and collaborative training. The error measurement system constructs a diversity of error measurement dimensions, and by establishing a dynamic interaction mechanism between tasks, subsystems can promote each other and evolve together, thereby significantly reducing the overall system error.

[0072] like Figure 1 As shown, in one embodiment of the present invention, step 400 includes: Step 410: Use the archive data processing model to extract features and perform semantic analysis on the batch of input archive images, identify core subject words and entity words in the archives, and complete subject labeling and entity association.

[0073] Feature extraction includes, but is not limited to, layout analysis of document structure parsing, paragraph and heading detection, table structure recognition, nested content restoration, and metadata generation, enabling batch structured processing of archival document content. Semantic analysis includes, but is not limited to, multilingual and handwritten character processing, understanding contextual relationships, and identifying archival themes and participating entities. Participating entities include, but are not limited to, individuals, institutions, dates, projects, and technical terms.

[0074] Step 420: Based on the semantic analysis results of the archive content, form a knowledge graph within and between archive images.

[0075] Semantic analysis results establish semantic relationships between key entities in archival images. These semantic relationships can be formed within a single archive or between single archives.

[0076] Step 430: Build a structured storage of archival data based on the knowledge graph.

[0077] In one embodiment of the present invention, the structured storage adopts JSON format. The structured storage is formed based on the hierarchical relationships of objects in the knowledge graph to express the connections between information in the knowledge graph.

[0078] like Figure 1 As shown, in one embodiment of the present invention, step 400 further includes: Step 440: Automated management of archival data is achieved through an archival data processing model.

[0079] Based on the deep learning capabilities of the archival data processing model, and using the already "controlled" open review data as a foundation, the system intelligently learns the "control" rules, constructs a "control" rule base and knowledge base, and performs intelligent "control" analysis on the archival data.

[0080] Based on the intelligent classification capabilities of the archival data processing model, it can be used in multiple business scenarios such as open archival review, digital output processing, assisting in archival organization, intelligent archival retrieval, and assisting in archival compilation and research.

[0081] Based on the rule engine and machine learning technology of the archival data processing model, a data quality assessment system is formed to detect consistency problems in archival data and effectively identify data defects through in-depth analysis of multi-source data.

[0082] The artificial intelligence-based archival data identification and processing method of this invention utilizes the specialization of models to automate archival data extraction and automatic structured storage. Simultaneously, it leverages the general capabilities of the models to build automated processing capabilities for different business scenarios in the archival management process. This enhances the ability to retrieve massive amounts of data and perform automated analysis of archival data. It effectively improves the intelligent collection and archiving capabilities for archival text data, and comprehensively supports richer data mining and utilization.

[0083] An embodiment of the present invention provides an artificial intelligence-based archival data recognition and processing system, such as... Figure 2 As shown. In Figure 2 In this embodiment, the following are included: The training sample construction device 20 is used to standardize the original archive images to form training samples and perform sample balancing according to the training purpose. The model training and optimization device 30 is used to train and optimize the basic large model based on the Transformer architecture using training samples to form an archive data processing model. The data recognition application device 40 is used to perform content recognition and semantic analysis of batch archival images through an archival data processing model, form a knowledge graph of archival entities, and construct an archival database.

[0084] like Figure 2 As shown, in one embodiment of the present invention, the training sample construction device 20 includes: Synchronous receiving module 21 is used to acquire text data and image data in the archive image; The vocabulary processing module 22 is used to perform word standardization processing on text data to form an embedded vocabulary set; Image processing module 23 is used to perform enhancement and standardization processing on image data to form an enhanced image set; Image segmentation module 24 is used to crop archival images at different scales and regions to form archival image slices; Sample construction module 25 is used to match words in the embedded vocabulary set and images in the augmented image set with archival image slices to form training samples; The sample balancing module 26 is used to adjust the proportion of training samples in the archival image slices for coverage size, connected regions, and text relevance according to the training objective.

[0085] like Figure 2 As shown, in one embodiment of the present invention, the model training optimization device 30 includes: The iterative training module 31 is used to form a training set based on the training samples to iteratively train the archive data processing model based on the Transformer architecture, set a convergence threshold, and stop training through an early stopping mechanism. Error measurement module 32 is used to form a training error measurement system, quantify the model accuracy according to the measurement index, and optimize parameters based on the accuracy. The collaborative training module 33 is used to quantify the collaborative task training error of the model according to the metric, and gradually meet the collaborative training error metric through collaborative training.

[0086] like Figure 2 As shown, in one embodiment of the present invention, the data identification application device 40 includes: Data extraction module 41 is used to perform feature extraction and semantic analysis on batch input archival images through archival data processing model, identify core subject words and entity words in the archives, and complete subject labeling and entity association; The data association module 42 is used to form a knowledge graph within and between archive images based on the semantic analysis results of the archive content; Data storage module 43 is used for structured storage of archival data formed based on knowledge graphs.

[0087] like Figure 2 As shown, in one embodiment of the present invention, the data recognition application device 40 further includes: The data automation module 44 is used to form automated management of archival data through the archival data processing model.

[0088] The embodiments of this application also provide a specific implementation of an electronic device capable of implementing all steps of the artificial intelligence-based archival data recognition and processing method in the above embodiments. (See attached document for details.) Figure 3 Electronic device 600 specifically includes the following: Processor 610, memory 620, communication unit 630 and bus 640; The processor 610, memory 620, and communication unit 630 communicate with each other via bus 640; the communication unit 630 is used to realize information transmission between server-side devices and terminal devices and other related devices.

[0089] The processor 610 is used to call the computer program in the memory 620. When the processor executes the computer program, it implements all the steps in the artificial intelligence-based archival data recognition and processing method in the above embodiments.

[0090] Those skilled in the art will understand that memory can be, but is not limited to, Random Access Memory (RAM), Read Only Memory (ROM), Programmable Read-Only Memory (PROM), Erasable Programmable Read-Only Memory (EPROM), and Electrically Erasable Programmable Read-Only Memory (EEPROM). The memory stores programs, which are then executed by the processor upon receiving execution instructions. Furthermore, the software programs and modules within the memory may include an operating system, which may include various software components and / or drivers for managing system tasks (e.g., memory management, storage device control, power management), and can communicate with various hardware or software components to provide an operating environment for other software components.

[0091] A processor can be an integrated circuit chip with signal processing capabilities. The aforementioned processor can be a general-purpose processor, including a Central Processing Unit (CPU), a Network Processor (NP), etc. It can implement or execute the methods, steps, and logic block diagrams disclosed in the embodiments of this application. The general-purpose processor can be a microprocessor or any conventional processor.

[0092] This application also provides a computer-readable storage medium including a program, which, when executed by a processor, is used to perform the artificial intelligence-based archival data recognition and processing method provided in any of the foregoing method embodiments.

[0093] Those skilled in the art will understand that all or part of the steps in the above-described method embodiments can be implemented by hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium. When executed, the program performs the steps of the above-described method embodiments; and the aforementioned storage medium includes various media capable of storing program code, such as ROM, RAM, magnetic disks, or optical disks, and this application does not limit the specific type of media.

[0094] Another embodiment of the present invention is an artificial intelligence-based method for identifying and processing archival data, as follows: Figure 4 As shown. In Figure 4 In addition to the above-described embodiments of the archival data identification and processing method, this embodiment further includes: Step 100: Extract fixed layouts from batches of original archive images using an image recognition model based on an LSTM (Long Short-Term Memory) architecture, establish layout association features of the archives, and combine them with the standardization processing of the original archives to form training samples.

[0095] Historical archives are diverse in type, with significant differences in document format across different organizations and periods. This results in the inconsistent location of key information (such as document number, date, name, and address), making extraction difficult using fixed rules. Archives often contain various elements such as tables, handwriting, stamps, and annotations, with key information potentially scattered across different locations or even spanning multiple pages, making it difficult for traditional models to establish complete information connections. Many historical archives are scanned or handwritten, significantly impacting their layout and resulting in low OCR recognition accuracy. This leads to poor quality data for subsequent information extraction, further increasing the difficulty of locating key information. Simple rules are insufficient for accurately locating and parsing key information.

[0096] Those skilled in the art will understand that text is essentially one-dimensional temporal data (a sequence of words / characters) and lacks a natural two-dimensional spatial structure. However, the original archives in archival files include the archive's creation time, a temporal feature closely related to the archive's background. For example, the text content of the archives is correlated with image features such as customized forms, stamps, and formatted titles in terms of overall layout. Through spatiotemporal joint modeling using the LSTM architecture, spatial features of fixed content within a single image are captured, while the temporal consistency of batch image sequences enhances recognition accuracy. Leveraging the advantages of the LSTM architecture in handling short-to-medium-term temporal tasks and small-sample scenarios for small-scale text analysis, images of original archives, based on paper entities, are rapidly analyzed to obtain pixel-level layout correlation features.

[0097] The artificial intelligence-based archival data recognition and processing method of this invention uses the correlation between the archival establishment period and image feature space as a measurement dimension of archival layout for original archives, enhances the feature representation of archival layout in training samples, and makes the training of the large language model more targeted, avoiding the failure of the optimization purpose of the large language model to achieve the desired result.

[0098] In one embodiment of the present invention, the batch of original archival images consists of archival images of the same type from the same period and department. The LSTM architecture model adopts the ConvLSTM model, including: The input layer receives standardized image data with a time step of 1.

[0099] The spatiotemporal feature extraction layer forms a two-layer ConvLSTM unit structure with residual connections. One ConvLSTM unit uses a larger convolutional kernel, while the other uses a smaller kernel to extract spatiotemporal features from different spatial receptive fields. Simultaneously, the ConvLSTM gating mechanism captures sequence dependencies between images (such as dynamic content changes). The outputs of the two ConvLSTM units form residual connections, reducing the gradient vanishing problem and enhancing feature propagation.

[0100] The attention enhancement module is used to apply attentional weights to spatiotemporal features, highlighting key regional information in space. It captures key features in both time and space dimensions through a multi-head attention mechanism, emphasizing crucial regional information within the space.

[0101] The multi-scale feature fusion output module is used to form a convolutional decoding structure, which performs convolution operations on spatiotemporal features to form a feature map that fuses multi-scale features, and then gradually maps the high-dimensional features in the feature map to the final classification result.

[0102] The output layer is used to output layout data for the archive images based on the classification results. It includes tables, headers, titles, and fixed labels.

[0103] Another embodiment of the present invention is an artificial intelligence-based archival data recognition and processing method system, such as... Figure 5 As shown. In Figure 5 Based on the aforementioned embodiments of the archival data identification and processing system, the system further includes: The archive layout processing module 10 is used to extract fixed layouts from batches of original archive images using an image recognition model based on an LSTM (Long Short-Term Memory) architecture, establish layout association features of the archives, and combine them with the standardized processing of the original archives to form training samples.

[0104] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A method for identifying and processing archival data based on artificial intelligence, characterized in that, include: Standardize the original archival images to form training samples, and balance the samples according to the training objective. The basic large model based on the Transformer architecture is trained and optimized using training samples to form an archival data processing model; By using an archival data processing model to perform content recognition and semantic analysis on batch archival images, a knowledge graph of archival entities is formed, and an archival database is constructed.

2. The method for identifying and processing archival data based on artificial intelligence as described in claim 1, characterized in that, The standardization process of the original archival images to form training samples, and the sample equalization according to the training objective, include: Retrieve text and image data from archive images; Text data is processed to standardize words, forming an embedded vocabulary set; Image data is enhanced and normalized to form an enhanced image set; Archival images are cropped at different scales and regions to create archival image slices; Training samples are generated by matching words from the embedded vocabulary set and images from the augmented image set with archival image slices. Adjust the proportion of training samples in the archival image slices for coverage size, connected regions, and text relevance, according to the training objectives.

3. The method for identifying and processing archival data based on artificial intelligence as described in claim 1, characterized in that, The process of training and optimizing the basic large model based on the Transformer architecture using training samples to form the archival data processing model includes: The training set is formed based on the training samples, and the archival data processing model based on the Transformer architecture is iteratively trained. A convergence threshold is set, and training is stopped through an early stopping mechanism. Establish a training error measurement system, quantify the model accuracy based on the measurement indicators, and optimize the parameters based on the accuracy. The collaborative task training error of the model is quantified based on the metric, and the collaborative training error is gradually satisfied through collaborative training.

4. The method for identifying and processing archival data based on artificial intelligence as described in claim 1, characterized in that, The underlying large model based on the Transformer architecture adopts the QWEN model.

5. The method for identifying and processing archival data based on artificial intelligence as described in claim 1, characterized in that, The process of performing content recognition and semantic analysis on batch archival images using an archival data processing model to form a knowledge graph of archival entities and construct an archival database includes: By using an archival data processing model, feature extraction and semantic analysis are performed on batch input archival images to identify core thematic words and entity words in the archives, and thematic labeling and entity association are completed. Based on the semantic analysis results of the archive content, a knowledge graph is formed within and between archive images; The knowledge graph forms the structured storage of archival data.

6. The method for identifying and processing archival data based on artificial intelligence as described in claim 5, characterized in that, Also includes: The data processing model enables automated management of archival data.

7. The method for identifying and processing archival data based on artificial intelligence as described in claim 1, characterized in that, Also includes: By extracting fixed layouts from batches of original archival images using an image recognition model based on LSTM architecture, layout association features of the archives are established, and training samples are formed by combining the standardized processing of the original archives.

8. The method for identifying and processing archival data based on artificial intelligence as described in claim 7, characterized in that, The image recognition model based on the LSTM architecture uses the ConvLSTM model.

9. An artificial intelligence-based archival data recognition and processing system, characterized in that, include: The training sample construction device is used to standardize the original archival images to form training samples and perform sample balancing according to the training objective. The model training and optimization device is used to train and optimize the Transformer architecture model using training samples to form an archive data processing model; The data recognition application device is used to perform content recognition and semantic analysis of batch archival images through an archival data processing model, form a knowledge graph of archival entities, and construct an archival database.

10. The artificial intelligence-based archival data recognition and processing system as described in claim 9, characterized in that, Also includes: The archive layout processing module is used to extract fixed layouts from batches of original archive images using an image recognition model based on an LSTM architecture, establish layout association features of the archives, and combine them with the standardized processing of the original archives to form training samples.

Images