Intelligent data standard element extraction method and system based on large model

Abstract

The invention provides a data standard element intelligent extraction method and system based on a large model, and the method comprises the steps: obtaining multi-modal heterogeneous data, carrying out the preprocessing and alignment, and mapping the data to a unified semantic space, so as to form an initial modal; respectively inputting the initial modality into a shared encoder and a private encoder of a large model to carry out domain adversarial transfer learning to obtain domain invariant features and domain specific features, and synthesizing the domain invariant features and the domain specific features into first features; executing cross-modal dynamic fusion on the first feature in a unified semantic space to obtain a fused feature; data standard elements and confidence coefficients thereof are obtained based on fusion feature decoding; comparing the confidence with a preset confidence threshold; when the confidence is not less than the confidence threshold, directly outputting corresponding data standard elements; when the confidence coefficient is smaller than the confidence coefficient threshold value, the dynamic rule engine is triggered to carry out post-processing, and the automatic verification layer executes consistency verification according to the domain knowledge graph to obtain element output after verification. And the intelligent extraction capability of the data standard elements based on the large language model is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of natural language processing, and particularly relates to a data standard element intelligent extraction method and system based on a large model. BACKGROUND

[0002] With the rapid development of artificial intelligence technology, especially the wide application of large language models in natural language processing, the structured processing and standardized management of data have become a key research direction. In financial statements, medical records, contract texts and other industry scenarios, standardized data usually contains a large amount of unstructured or semi-structured information, and traditional methods are difficult to efficiently parse and utilize. Therefore, how to realize intelligent extraction of data standard elements based on large models, improve data structuring capabilities, and improve the accuracy and automation level of information extraction has become the focus of current research.

[0003] In recent years, large language models have shown strong context modeling capabilities in natural language understanding tasks, providing a new technical path for data element extraction. However, existing methods based on large models still have the following problems: insufficient data quality and cross-domain transfer ability, and general pre-training models lack targeted optimization for vertical domain data standards, resulting in weak adaptability in professional scenarios. At the same time, due to the lack of dynamic updating mechanism, the model responds slowly when facing changes in data standards or user feedback, making it difficult to achieve real-time adaptation. In addition, existing methods have obvious limitations in multi-modal data processing, making it difficult to effectively integrate text, table, image and other different modal data to achieve unified standard element extraction.

[0004] Existing data standard element intelligent extraction methods rely on artificial rules or single-domain labeled data, making it difficult to adapt to the differences in diversified business scenarios and the logical inconsistency between cross-data standards, resulting in poor stability and consistency of the extraction results. In addition, the domain generalization ability of these methods is limited, and most models need to be labeled and trained for each new domain, making it difficult to effectively utilize limited samples to realize cross-domain knowledge transfer. This not only increases the cost of data labeling, but also significantly weakens its feasibility and efficiency in large-scale applications.

[0005] The existing data standard element intelligent extraction method is mostly a static model, which excessively depends on a fixed training data set, is difficult to cope with the dynamic changes of business rules and the continuous evolution of new data standards, and leads to problems such as not timely updating and insufficient adaptability in actual application. In addition, due to the diversity of enterprise data sources, the existing method lacks unified and effective structured representation capability and cannot meet the application requirements of data governance, standard alignment and knowledge graph construction. At the same time, the traditional rule engine or regular expression method highly depends on manual programming and has limited generalization ability, which is difficult to adapt to the dynamic changes in complex business environment, further restricting the application breadth and expansibility of the data standard element intelligent extraction technology. SUMMARY

[0006] The present application provides a data standard element intelligent extraction method and system based on a large model to solve the technical problems of insufficient data quality and cross-domain migration ability of the existing data standard element intelligent extraction method, lack of targeted optimization of vertical field data standards in general pre-training models, leading to weak adaptability in professional scenarios, and lack of dynamic updating mechanism, causing the model to respond slowly when facing data standard changes or user feedback, and difficulty in realizing real-time adaptation.

[0007] To solve the above technical problems, the present application provides a data standard element intelligent extraction method based on a large model, comprising the following steps: Step S1: acquiring multi-modal heterogeneous data and pre-processing and aligning, mapping to a unified semantic space to form an initial mode; Step S2: inputting the initial mode into the shared encoder and private encoder of the large model respectively for domain adversarial transfer learning, obtaining domain-invariant features and domain-specific features and synthesizing first features; Step S3: performing cross-modal dynamic fusion on the first features in the unified semantic space to obtain fusion features; Step S4: decoding the fusion features to obtain data standard elements and their confidence; Step S5: comparing the confidence with a preset confidence threshold: when the confidence is not less than the confidence threshold, directly outputting the corresponding data standard element; when the confidence is less than the confidence threshold, triggering a dynamic rule engine for post-processing, and performing consistency checking according to the domain knowledge graph by an automated verification layer to obtain the checked element output; Step S6: collecting the checked element output and user feedback to form a feedback signal for adaptive updating; Step S7: writing the updating result of S6 back to steps S1 to S5 for re-extraction.

[0008] Preferably, the shared encoder in step S2 multiplexes parameters of the pre-trained large model, and the private encoder superimposes lightweight Transformer layers on the basis of the pre-trained large model, and synthesizes outputs of the shared encoder and the private encoder into the first feature.

[0009] Preferably, the domain adversarial transfer learning in step S2 includes connecting a domain discriminator to the shared encoder via a gradient reversal layer, training the domain discriminator under cross-entropy loss and imposing an adversarial constraint on the shared encoder to strengthen the domain independence of the shared feature and achieve distribution alignment.

[0010] Preferably, the cross-modal dynamic fusion in step S3 includes weighting the first features corresponding to the text, table and image based on adaptive attention weights, inputting the weighted features into a cross-modal Transformer layer, and obtaining a fusion feature.

[0011] Preferably, the automatic verification layer in step S5 performs consistency verification on cross-element logical constraints according to the domain knowledge graph, and the verification result is part of the feedback signal in step S6.

[0012] Preferably, the adaptive update in step S6 includes: selecting samples based on the confidence from step S4, and forming a training set combined with user feedback; performing incremental learning based on knowledge distillation using the training set: using the pre-update model as the teacher model, linearly weighting the cross-entropy loss of the real label and the KL divergence between the teacher model and the student model to constrain the student model.

[0013] Preferably, step S6 further includes: reinforcement learning based on the feedback signal, using the verification result of the automatic verification layer and the user feedback as a reward signal, and optimizing the confidence threshold and the weight of the dynamic rule engine.

[0014] Preferably, the large model in step S2 adds knowledge graph triple embedding and translation loss constraint during model fine-tuning, performs uncertainty sampling on training samples based on prediction entropy, and only fine-tunes the adaptation layer under the constraint of parameter efficient fine-tuning.

[0015] Preferably, step S1 performs contrastive learning on multi-modal heterogeneous data, sets a temperature coefficient to smooth the similarity distribution, makes positive sample pairs more similar and negative sample pairs separate from each other, and the loss function is expressed as: ; In the formula, is a temperature coefficient, N is the number of batch samples, and is a positive sample pair.

[0016] The application also provides a large model-based data standard element intelligent extraction system, which is realized based on the above method and comprises a preprocessing and alignment module, a field adversarial transfer learning module, a cross-modal dynamic fusion module, a decoding and confidence generation module, a dynamic rule engine and automatic verification layer module, and a self-adaptive updating module. The preprocessing and alignment module is used for processing multi-modal heterogeneous data and mapping each mode to an initial mode in a unified semantic space. The field adversarial transfer learning module is used for extracting field-invariant features and field-specific features based on a shared encoder and a private encoder, and performing adversarial alignment through a gradient reversal layer and a field discriminator to obtain aligned first features. The cross-modal dynamic fusion module is used for adaptively weighting the first features and generating fusion features through cross-modal modeling. The decoding and confidence generation module is used for generating data standard elements and their confidence based on the fusion features. The dynamic rule engine and automatic verification layer module is used for performing parameterized template post-processing on the elements when the confidence is lower than a threshold, and performing consistency checking on cross-element logical relationships according to a field knowledge graph to obtain checked element output. The self-adaptive updating module is used for updating model parameters through incremental learning of knowledge distillation based on the checking result and user feedback, and optimizing the confidence threshold and the dynamic rule engine weight through reinforcement learning.

[0017] The application has at least the following beneficial effects: 1) Since the definition of data standard elements is complex and diverse, and the terms, semantics and context across different fields, traditional methods are difficult to quickly adapt to new business environments. The application constructs a field adversarial transfer learning mechanism, constructs a double-channel structure of a field shared feature space and a field private feature space, extracts field-invariant features through adversarial training, and integrates a field knowledge graph into a large model fine-tuning link, so that the model can dynamically learn and adjust entity relationship weights, and strengthen its understanding and adaptability to vertical field data standards. Eliminate the field difference in the shared feature space, combine knowledge graph embedding and uncertainty sampling strategy, realize the rapid adaptation of cross-field data standards, significantly improve the model's data standard element extraction ability in the vertical field, reduce the dependence on labeled data, and enhance the generalization performance in cross-business scenarios.

[0018] Meanwhile, the framework adopts an uncertainty sampling strategy, preferentially selects samples with high information gain, and combines migration and active learning to realize dynamic optimization of model parameters, thereby improving cross-domain migration efficiency and data quality. In addition, a domain self-supervised learning mechanism is introduced, which can automatically mine and align domain-specific semantic information even in the case of unsupervised or limited labeled data, reducing the dependence on manual labeling, thereby improving the generalization performance and stability of data element extraction.

[0019] 2) A multi-modal dynamic fusion and incremental learning mechanism is proposed to realize multi-modal data collaborative processing and real-time updating: With the diversification of business data forms, it is difficult to meet the actual needs by relying only on text information for standard element extraction. Existing methods usually cannot effectively integrate text, table, image, PDF and other multi-modal data, resulting in information loss or incomplete element extraction. To solve this problem, the present invention proposes a multi-modal dynamic fusion and incremental learning mechanism, which performs deep semantic fusion on text, structured data (such as tables), images and other data types through a cross-modal alignment model, ensuring that information from different modalities can be complementary and enhanced. Combined with adaptive attention mechanisms and multi-modal feature mapping techniques, the model can dynamically adjust the weight distribution of different modalities according to specific task requirements, ensuring accurate extraction of standard elements. At the same time, to address the problem of dynamic changes in data standards, an incremental learning mechanism is introduced, allowing the model to gradually update existing knowledge systems based on knowledge distillation and parameter efficient fine-tuning techniques when new data or business rules are added, enabling continuous optimization and real-time adaptation, and avoiding the high computational cost of retraining. In addition, through user feedback-driven active learning strategies, the model can continuously optimize extraction rules and enhance adaptability to new data standards, ensuring long-term stable and efficient operation, thereby improving the intelligent level of data governance, standard alignment and knowledge graph construction. Effectively integrating text, table, image and other multi-modal information, the present invention solves the information loss problem of traditional single-modal methods. The incremental learning mechanism combined with knowledge distillation technology supports the model to quickly adapt to new data standards with low computational overhead, avoiding performance degradation due to retraining, and ensuring long-term stability of the system. BRIEF DESCRIPTION OF DRAWINGS

[0020] Figure 1 The method flowchart of the embodiments of the present invention is shown in the accompanying drawings. DETAILED DESCRIPTION

[0021] The technical solutions in the embodiments of the present invention will be described clearly and completely in conjunction with the accompanying drawings in the embodiments 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 skilled in the art without creative labor are within the protection scope of the present invention.

[0022] AsFigure 1 As shown, the embodiment of the present application provides a data standard element intelligent extraction method based on a large model, which comprises the following steps: Step S1: Obtain multi-modal heterogeneous data and perform preprocessing and alignment, and map to a unified semantic space to form an initial mode.

[0023] Step S2: Input the initial mode into the shared encoder and private encoder of the large model for domain adversarial transfer learning, obtain domain-invariant features and domain-specific features, and synthesize first features.

[0024] Step S3: Perform cross-modal dynamic fusion on the first features in the unified semantic space to obtain fusion features.

[0025] Step S4: Decode the fusion features to obtain data standard elements and their confidence levels; Step S5: Compare the confidence level with the preset confidence threshold: when it is not less than the confidence threshold, directly output the corresponding data standard element; when it is less than the confidence threshold, trigger the dynamic rule engine for post-processing, and perform consistency checking according to the domain knowledge graph by the automatic verification layer to obtain the checked element output; Step S6: Collect the checked element output and user feedback to form a feedback signal for adaptive updating; Step S7: Write the update result of S6 back to steps S1 to S5 for re-extraction.

[0026] Specifically, first, the input multi-modal heterogeneous data is standardized. For text data, a pre-training model such as BERT is used for sentence division, entity recognition, and semantic segmentation, while irrelevant symbols and repeated content are removed, and the standard element format is unified. For table data, the layout analysis algorithm is used to extract the row and column structure and cell semantic labels; image data is processed by OCR technology to extract text information, and combined with the target detection model to locate the contract seal, signature column and other key areas. In the noise processing stage, a non-local mean denoising algorithm is used to optimize the image quality and improve the OCR recognition accuracy. After completing the data analysis, the multi-modal large model is used to map the text, image and table to a unified semantic space. The formula is as follows: ; ; ; Wherein, T , I and C represent the text, image and table parameters, respectively; , and are the encoder parameters of the text, image and table, respectively; , and The initial modal representation that maps text, images, and tables to a unified semantic space.

[0027] Furthermore, in one embodiment of the present invention, a contrastive learning strategy is used to maximize the similarity of positive sample pairs and minimize the similarity of negative sample pairs to optimize the cross-modal embedding representation. Specifically, a contrastive loss function is defined, using a temperature coefficient... Adjusting the smoothness of the similarity distribution to ensure semantic alignment of data from different modalities, the loss function... as follows: ; in, This is a temperature coefficient used to adjust the smoothness of the similarity distribution. N This represents the batch sample size. and These are positive sample pairs.

[0028] To overcome the limitations of traditional models in cross-domain transfer learning, one embodiment of this invention designs a domain adversarial transfer learning framework based on a dual-channel structure. This framework achieves the extraction of domain-invariant features and dynamic adaptation of vertical domain knowledge through the collaborative optimization of a domain-shared feature space and a domain-private feature space. First, a dual-encoder structure is constructed based on an existing large-scale model, where the shared encoder extracts cross-domain general semantic features through adversarial training. Its output features need to satisfy the domain independence constraint, that is, the distribution of samples from different domains in the shared space should be as aligned as possible. In this embodiment, the pre-trained weights of the existing large model Qwen2.5 are directly reused as a cross-domain general feature extractor. The private encoder adds a lightweight Transformer layer on the existing large model Qwen2.5 to capture features specific to the vertical domain. Furthermore, to enhance the domain independence of shared features, another embodiment of the present invention introduces a gradient reversal layer (GRL) and a domain discriminator, where the discriminator D receives the shared features. It predicts the domain label of a sample. Its loss function is cross-entropy loss. : ; In the formula, This represents the true label of the i-th sample in the domain category; This represents the shared features of the i-th sample.

[0029] To achieve multimodal data collaborative processing, one embodiment of the present invention designs an adaptive attention mechanism to dynamically allocate the weights of each modality, with the weight values... Dynamically generated using learnable parameters: ; Using a cross-modal Transformer layer, features from different modalities are mapped to a shared semantic space, supporting joint extraction and association analysis. The formula is: ; in, This indicates feature splicing.

[0030] To address the issue of dynamically updated data standards, an incremental learning mechanism is introduced. When business rules are adjusted, knowledge distillation technology is used, with the old model serving as the teacher model to guide the new model in learning the distributions of both old and new data, thus preventing forgetting. Its expression is: ; in, For balance coefficient, Indicates the true label, This represents the predicted output of the teacher model. The student model is the predicted output of the new model. For cross-entropy loss, The loss is KL divergence. Furthermore, through a user-feedback-driven active learning strategy, the model continuously optimizes the extraction rules to achieve real-time adaptation.

[0031] To enhance the model's understanding of vertical domain data standards, a Domain Knowledge Graph (DKG) is constructed based on industry standards, containing entity, attribute, and relation triples. During model fine-tuning, the DKG triples are embedded into a large language model, and a translation loss function is used to enhance the model's understanding of domain semantics. ; in, h, c, t These represent the head entity, relation, and tail entity of the knowledge graph, respectively. This is an entity encoder. To address the scarcity of labeled data, an uncertain sampling strategy is employed, prioritizing samples with high predicted entropy values ​​for labeling. Furthermore, transfer learning techniques are used to freeze the model's underlying parameters, fine-tuning only the top-level adapter layers to reduce the risk of overfitting. (Entropy value) The calculation formula is as follows: ; in x Indicates the input sample. C Represents the set of possible categories. The model represents the samples x Category c The predicted probability, when At that time, the samples were added to the training set, where The threshold is dynamic. In unlabeled data scenarios, the model is trained to reconstruct the masked content based on context by randomly masking standard features: ; in, M For the set of mask locations, This represents the original token that has been masked. This represents the remaining context after the mask.

[0032] Finally, this embodiment of the invention employs a hybrid decision-making architecture that combines rules and models. A confidence threshold is defined. When the model predicts confidence Output the result directly; if If the result is incorrect, the dynamic rule engine will be triggered for post-processing. The rule engine supports parameterized template configuration; for example, date extraction rules can be dynamically expanded to include delimiter types, improving the flexibility and maintainability of the rule base. The rule base is updated in real time based on user feedback to ensure synchronization with business needs. To achieve continuous system optimization, an automated verification layer and a reinforcement learning framework are designed. The verification layer checks the consistency of the extraction results through the logical constraints of the business knowledge graph. Errors reported by users will trigger an error analysis process, which will dynamically update the rule base and recalibrate the model parameters after identifying the problem type. For example, the feedback is transformed into a reinforcement learning reward signal to optimize the model confidence threshold and rule weights, forming a closed-loop optimization mechanism of "processing-feedback-iteration". in For the reward function, For learning rate, This represents the expected value of the reward function with respect to the parameters. θ gradient, This represents the parameters at the t-th iteration.

[0033] The method of this invention employs a dual-channel architecture of domain adversarial transfer learning, constructing a shared encoder to extract cross-domain general features and a private encoder to capture vertical domain-specific features. Combined with adversarial training of a gradient inversion layer (GRL) and a domain discriminator, it achieves efficient extraction of domain-invariant features. Through triple embedding of the domain knowledge graph and an uncertainty sampling strategy, it dynamically optimizes model parameters in scenarios with scarce labeled data, enhancing cross-domain transfer capabilities. A multimodal dynamic fusion and incremental learning mechanism is proposed, employing a cross-modal contrastive learning framework. This framework dynamically allocates weights for different modalities such as text, tables, and images through an adaptive attention mechanism, and utilizes a cross-modal Transformer layer to achieve semantic alignment and joint extraction of multimodal features. Simultaneously, a knowledge distillation-driven incremental learning mechanism is introduced to update the model with low computational cost when data standards change, avoiding catastrophic forgetting and ensuring real-time adaptability.

[0034] This invention also provides a data standard element intelligent extraction system based on a large model, implemented based on the above method, including a preprocessing and alignment module, a domain adversarial transfer learning module, a cross-modal dynamic fusion module, a decoding and confidence generation module, a dynamic rule engine and automated verification layer module, and an adaptive update module. The preprocessing and alignment module is used to process multimodal heterogeneous data and map each modality to a unified semantic space to form an initial modality; The domain adversarial transfer learning module is used to extract domain-invariant and domain-specific features based on a shared encoder and a private encoder, and to perform adversarial alignment with the domain discriminator through a gradient inversion layer to obtain the first aligned feature. The cross-modal dynamic fusion module is used to adaptively weight the first feature with attention and generate fused features through cross-modal modeling; The decoding and confidence generation module is used to generate standard data elements and their confidence levels based on fused features. The dynamic rule engine and automated verification layer module are used to perform parameterized template post-processing on elements when the confidence level is lower than the threshold, and to perform consistency verification on cross-element logical relationships based on the domain knowledge graph to obtain the verified element output. The adaptive update module is used to update model parameters incrementally through knowledge distillation based on verification results and user feedback, and to optimize the confidence threshold and dynamic rule engine weights through reinforcement learning.

[0035] The technical features of the above embodiments can be combined arbitrarily. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described; only preferred embodiments of the present invention are illustrated. The descriptions are relatively specific and detailed, but they should not be construed as limiting the scope of the present invention. As long as the combination of these technical features does not contradict each other, it should be considered within the scope of this specification.

[0036] It should be noted that those skilled in the art can make various modifications and improvements without departing from the inventive concept, and these all fall within the scope of protection of this invention. Therefore, the scope of protection of this invention should be determined by the appended claims.

Claims

1. A method for intelligent extraction of standard data features based on a large model, characterized in that: Includes the following steps: Step S1: Acquire multimodal heterogeneous data and perform preprocessing and alignment, mapping it to a unified semantic space to form an initial modality; Step S2: Input the initial modality into the shared encoder and private encoder of the large model respectively for domain adversarial transfer learning to obtain domain-invariant features and domain-specific features, and synthesize the first feature; Step S3: Perform cross-modal dynamic fusion on the first feature within the unified semantic space to obtain the fused feature; Step S4: Obtain the data standard elements and their confidence levels based on the fused feature decoding; Step S5: Compare the confidence level with a preset confidence threshold: if it is not less than the confidence threshold, directly output the corresponding data standard elements; if it is less than the confidence threshold, trigger the dynamic rule engine for post-processing, and the automated verification layer performs consistency verification based on the domain knowledge graph to obtain the verified element output. Step S6: Collect the verified element output and user feedback to form a feedback signal for adaptive updating; Step S7: Write back the update result of S6 to steps S1 to S5 for re-extraction.

2. The intelligent extraction method for standard data features based on a large model according to claim 1, characterized in that: In step S2, the shared encoder reuses the parameters of the pre-trained large model, the private encoder superimposes a lightweight Transformer layer on the pre-trained large model, and combines the outputs of the shared encoder and the private encoder to form the first feature.

3. The intelligent extraction method for standard data features based on a large model according to claim 2, characterized in that: The domain adversarial transfer learning in step S2 includes: connecting the output of the shared encoder to the domain discriminator via a gradient inversion layer, training the domain discriminator with cross-entropy loss, and imposing adversarial constraints on the shared encoder to enhance the domain independence of the shared features, thereby achieving distribution alignment.

4. The intelligent extraction method for standard data features based on a large model according to claim 1, characterized in that: The cross-modal dynamic fusion in step S3 includes: weighting the first features corresponding to text, tables and images based on adaptive attention weights, and inputting the weighted features into the cross-modal Transformer layer to obtain fused features.

5. The intelligent extraction method for standard data features based on a large model according to claim 1, characterized in that: In step S5, the automated verification layer performs consistency verification on cross-element logical constraints based on the domain knowledge graph, and uses the verification result as part of the feedback signal in step S6.

6. The intelligent extraction method for standard data features based on a large model according to claim 1, characterized in that: The adaptive update in step S6 includes: selecting samples based on the confidence level from step S4, and forming a training set by combining user feedback; using the training set for incremental learning based on knowledge distillation: using the model before the update as the teacher model, and linearly weighting the cross-entropy loss of the real labels and the KL divergence outputs of the teacher model and the student model to constrain the student model.

7. The intelligent extraction method for standard data features based on a large model according to claim 1, characterized in that: Step S6 further includes: performing reinforcement learning based on the feedback signal, using the verification results of the automated verification layer and user feedback to form a reward signal, and optimizing the confidence threshold and the weights of the dynamic rule engine.

8. The intelligent extraction method for standard data features based on a large model according to claim 1, characterized in that: In step S2, the large model incorporates knowledge graph triple embedding and translation loss constraints during model fine-tuning. It performs uncertainty sampling on training samples based on prediction entropy and only fine-tunes the adaptation layer under the constraint of efficient parameter fine-tuning.

9. The intelligent extraction method for standard data features based on a large model according to claim 1, characterized in that: In step S1, comparative learning is performed on multimodal heterogeneous data. A temperature coefficient is set to smooth the similarity distribution, making positive sample pairs more similar and negative sample pairs more separated. The loss function is... The expression is: ； In the formula, For temperature coefficient, N This represents the batch sample size. and These are positive sample pairs.

10. A data standard feature intelligent extraction system based on a large model, implemented based on the method described in any one of claims 1 to 9, characterized in that: It includes a preprocessing and alignment module, a domain adversarial transfer learning module, a cross-modal dynamic fusion module, a decoding and confidence generation module, a dynamic rule engine and automated verification layer module, and an adaptive update module; The preprocessing and alignment module is used to process multimodal heterogeneous data and map each modality to a unified semantic space to form an initial modality. The domain adversarial transfer learning module is used to extract domain-invariant features and domain-specific features based on a shared encoder and a private encoder, and to perform adversarial alignment with a domain discriminator through a gradient inversion layer to obtain the aligned first feature. The cross-modal dynamic fusion module is used to adaptively weight the first feature with attention and generate a fused feature through cross-modal modeling; The decoding and confidence generation module is used to generate data standard elements and their confidence levels based on fused features; The dynamic rule engine and automated verification layer module are used to perform parameterized template post-processing on the elements when the confidence level is lower than the threshold, and to perform consistency verification on the cross-element logical relationship based on the domain knowledge graph to obtain the verified element output. The adaptive update module is used to update the model parameters by incremental learning through knowledge distillation based on the verification results and user feedback, and to optimize the confidence threshold and dynamic rule engine weights through reinforcement learning.

Images