Dynamic fusion-based self-adaptive architecture corpus authenticity research and judgment system and method

By dynamically fusion of CNN and Transformer models, combining text complexity evaluation and meta controller, the adaptability and efficiency of content authenticity discrimination of large language models in the prior art is solved, and higher accuracy and scalability are achieved , reduces the cost of manual review and enhances the transparency of the system and user trust.

CN119025677APending Publication Date: 2024-11-26NANJING GRAY HU INTELLIGENT TECHNOLOGY CO LTD
View PDF 0 Cites 2 Cited by

Patent Information

Application Number
CN202411064789.7
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-08-05
Publication Date
2024-11-26

AI Technical Summary

Technical Problem

The prior art lacks adaptability and flexibility in the authenticity discrimination of content generated by large language models, lacks computational efficiency and scalability, and lacks interpretability and robustness, making it difficult to deal with diverse and complex text content.

Method used

Adaptive architecture based on dynamic fusion, combined with CNN and Transformer models, dynamically adjust model weights through text complexity evaluation, and optimize the architecture using meta-controllers to achieve efficient processing of texts of different lengths and complexity.

Benefits of technology

It significantly improves discrimination accuracy and adaptability, reduces computing overhead, enhances the scalability and interpretability of the system, reduces the need for manual audits, and improves processing efficiency and user trust.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119025677A_ABST
    Figure CN119025677A_ABST
Patent Text Reader

Abstract

The invention discloses an adaptive architecture corpus authenticity research and judgment system based on dynamic fusion, which relates to the technical field of computers and comprises a data preparation module, a text complexity evaluation module, a dynamic adjustment module, a CNN module, a Transform module, a fusion module, a classification head and a meta controller. The invention further provides a method for studying and judging the authenticity of the self-adaptive framework corpus based on dynamic fusion. The method comprises the following steps: performing initial training by using the annotation data generated by the data preparation module; evaluating the input text using a text complexity evaluation module; adjusting the weight of the CNN and the weight of the Transform according to the evaluation result by using a dynamic adjustment module; the input is processed through the CNN module and the Transform module respectively; using a fusion module to merge the output of the CNN module and the output of the Transform module; and obtaining a final prediction authenticity judgment result through the classification head. According to the method, the influence of the illusion problem on the reliability of the system can be greatly reduced, and meanwhile, the processing efficiency and the judgment accuracy, efficiency and adaptability are improved.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention belongs to the field of computer technology, and in particular relates to a system and method for evaluating the authenticity of adaptive architecture corpus based on dynamic fusion. Background Art

[0002] In the field of authenticity judgment of content generated by large language models, existing technologies mainly include BERT-based text classification models, traditional CNN text classification models and rule-based detection methods.

[0003] BERT (Bidirectional Encoder Representations from Transformers) is a pre-trained language model proposed by Google in 2018. It captures contextual information through a bidirectional Transformer structure and performs well in a variety of natural language processing tasks. In the task of content authenticity discrimination, BERT is often used as a feature extractor, and its output is passed through a classification layer for binary classification. The application of convolutional neural networks (CNN) in text classification was first proposed by Kim in 2014. The CNN model uses convolution kernels of different sizes to capture local text features, extracts the most significant features through pooling operations, and finally classifies through a fully connected layer. Rule-based detection methods identify possible false or untrue content based on predefined rules and patterns. For example, checking the frequency of occurrence of specific keywords, or analyzing the grammatical structure and complexity of the text.

[0004] Although these technologies have achieved certain results in content authenticity identification, they still have the following three major shortcomings:

[0005] 1. Lack of adaptability and flexibility: Existing BERT, CNN models, and rule-based methods all use fixed structures or predefined rules, which are difficult to optimize for texts of different complexities and types. This leads to unstable performance when processing diverse content, especially when facing new and complex false content. At the same time, they lack the ability to dynamically adjust processing strategies based on the characteristics of the input text, making it difficult to maintain efficient performance in various scenarios.

[0006] 2. Computational efficiency and scalability issues: Complex models such as BERT have high computational overhead when processing large amounts of text, making it difficult to achieve real-time processing. Although CNN models are relatively efficient, they also face the problem of reduced efficiency when processing long texts. These issues limit the scalability of existing technologies in large-scale applications and make it difficult to meet the growing demand for content review.

[0007] 3. Insufficient interpretability and robustness: Deep learning models such as BERT and CNN face the "black box" problem, which makes it difficult to explain the basis for their judgments, which is a significant disadvantage in application scenarios that require a high degree of transparency. At the same time, the performance of existing methods tends to drop significantly when faced with noisy data, spelling errors, or non-standard expressions. In particular, the robustness of existing methods faces severe challenges when dealing with high-quality false content generated by large language models.

[0008] These shortcomings severely limit the application effect of existing technologies in the task of authenticity judgment of content generated by large models, and a new, more intelligent and flexible solution is urgently needed. Summary of the invention

[0009] The purpose of this invention is to solve the problems in the prior art that the question-answer pairs generated by large models may have hallucinations, affect the reliability of the system, lack an automated authenticity discrimination mechanism, manual review is time-consuming and labor-intensive, and the existing discrimination model is difficult to adapt to text inputs of different lengths and complexities. An adaptive architecture based on dynamic fusion of CNN and Transformer is proposed to improve the accuracy, efficiency and adaptability of discrimination.

[0010] In order to achieve the above object, the present invention adopts the following technical solutions:

[0011] A method for judging the authenticity of adaptive architecture corpus based on dynamic fusion is designed, including the following steps:

[0012] Step 1: Use the labeled data generated by the data preparation module for initial training;

[0013] Step 2. In each initial training batch:

[0014] a. Use the text complexity evaluation module to evaluate the input text;

[0015] b. Use the dynamic adjustment module to adjust the weights of CNN and Transformer according to the evaluation results;

[0016] c. Process the input through the CNN module and the Transformer module respectively;

[0017] d. Use the fusion module to merge the outputs of the CNN module and the Transformer module;

[0018] e. Obtain the final prediction authenticity judgment result through the classification head;

[0019] f. Calculate the loss and back-propagate to update the parameters;

[0020] Step 3: Use the meta-controller to periodically evaluate the current architecture and make adjustment recommendations;

[0021] Step 4: Fine-tune the model architecture based on the meta-controller’s recommendations.

[0022] Furthermore, in step 1, the specific steps of initial data generation and labeling are as follows:

[0023] a. Use the llama3 big model to generate initial question-answer pairs based on industry or company system documents;

[0024] b. Randomly select 20% of the samples and use the Claude3.5 Sonnet model for preliminary review to give a binary classification judgment on whether the corpus contains hallucinations.

[0025] C. The expert team module conducts a secondary review based on the results of the Claude3.5 Sonnet model to obtain the final high-quality annotation results.

[0026] Furthermore, in step a of step 2, the text complexity evaluation module evaluation includes the following evaluation path:

[0027] Path 1: Evaluate the path based on the number of characters, including the following steps:

[0028] Step 11, count the number of characters in all corpus records, and obtain the minimum value a1, quartile b1, average c1, three-quarter digit d1, and maximum value e1;

[0029] Step 12: According to the range of the number of characters in the corpus, assign different complexity values ​​complex_value1 (0.25, 0.50, 0.75, 1.00);

[0030] Path 2: Evaluation path based on the number of knowledge points, including the following steps:

[0031] Step 21, counting the number of knowledge points of the text information of each corpus source, and obtaining the minimum value a2, quartile b2, average c2, three-quarter digit d2, and maximum value e2 of the number of knowledge points of all corpus;

[0032] Step 22: According to the interval in which the number of knowledge points falls, assign different complexity values ​​complex_value2 (0.25, 0.50, 0.75, 1.00);

[0033] Path 3: Comprehensive assessment path, including the following steps:

[0034] Step 31: Calculate the sum of complex_value1+complex_value2

[0035] Step 32: According to the sum value, the corpus is divided into four categories: uncomplex text (≤0.5), medium-complex text (0.5-1), medium-to-high-complex text (1-1.5), and complex text (>1.5).

[0036] Furthermore, in step b of step 2, the dynamic adjustment module dynamically adjusts the weights of CNN and Transformer including:

[0037] a. Simple text: CNN weight 0.8, Transformer weight 0.2;

[0038] b. Medium complexity text: CNN weight 0.6, Transformer weight 0.4;

[0039] c. Medium to high-complexity text: CNN weight 0.4, Transformer weight 0.6;

[0040] d. Complex text: CNN weight 0.2, Transformer weight 0.8.

[0041] In order to solve the above technical problems, the present invention also proposes a dynamic fusion-based adaptive architecture corpus authenticity assessment system, which is used for the method of dynamic fusion-based adaptive architecture corpus authenticity assessment, including:

[0042] Data preparation module, used for generating and labeling initial data;

[0043] A text complexity evaluation module, used to evaluate the complexity of the input text;

[0044] Dynamic adjustment module, used to dynamically adjust the weights of CNN and Transformer according to the text complexity evaluation results;

[0045] CNN module, adopts multi-channel CNN structure;

[0046] Transformer module, which uses the standard Transformer encoder structure;

[0047] Fusion module, used to fuse the outputs of CNN and Transformer;

[0048] Classification head, used for the final prediction results, including the fully connected layer and the Softmax layer;

[0049] The fully connected layer is used to map the fused features to a binary classification space, and the Softmax layer is used to output the final classification probability;

[0050] A meta-controller is used to learn the optimal architecture combination strategy during training.

[0051] Furthermore, the CNN module includes:

[0052] Embedding layer, which converts input text into word vectors;

[0053] Multiple parallel convolutional layers are used to extract local features using convolution kernels of different sizes;

[0054] The maximum pooling layer is used to perform maximum pooling on the convolution results of each channel;

[0055] The concatenation layer is used to concatenate all pooling results into one vector.

[0056] Furthermore, the Transformer module includes:

[0057] Position encoding layer, used to add position information to the input sequence;

[0058] Multi-head self-attention layer to capture long-range dependencies in the sequence;

[0059] A feed-forward neural network layer to process the output of self-attention;

[0060] Layer Normalization layer, applied after each sub-layer to stabilize the training process.

[0061] Furthermore, the fusion module adopts a gating mechanism, including:

[0062] Calculate the gate value g = σ (W_g*[CNN_output; Transformer_output] + b_g)

[0063] Fusion result = g*CNN_output+(1-g)*Transformer_output

[0064] Among them, σ is the sigmoid function, W_g and b_g are learnable parameters.

[0065] Furthermore, the meta-controller includes:

[0066] The policy network module is used to generate architecture adjustment strategies based on the current input and model state;

[0067] Reward function module, which defines rewards based on model performance and computational efficiency;

[0068] Reinforcement learning algorithm module, used to optimize the policy network.

[0069] Compared with the prior art, the technical problem that the present invention aims to solve is: how to effectively and automatically determine the authenticity of question-answer pairs generated by a large model, and adapt to text inputs of different lengths and complexities. The solution to this problem is crucial to improving the reliability and practicality of AI question-answering systems. Through the method proposed in the present invention, the impact of the "hallucination" problem on system reliability can be greatly reduced, while improving processing efficiency. This can not only improve the user experience, but also expand the scope of application of AI question-answering systems, especially in fields that require a high degree of accuracy, such as education, medical care or legal consultation. In addition, the adaptive architecture of the present invention provides an effective solution for processing diverse text inputs, which is of great significance for building a general, high-performance NLP system. Specifically, the beneficial effects are:

[0070] (1) The present invention can identify possible "illusions" or false information by learning a large amount of annotated data through a binary classification model. It not only considers the surface features of the text, but also analyzes deep features such as semantic coherence and factual consistency. This automated discrimination mechanism greatly reduces the need for manual review and improves efficiency and consistency. At the same time, the model's discrimination results can be used as an aid to manual review, significantly improving the accuracy and speed of the overall review.

[0071] (2) The present invention adopts an adaptive architecture that dynamically integrates CNN and Transformer, and can dynamically adjust the model structure according to the complexity of the input text. Specifically: CNN is used to capture local features and process relatively simple or short texts; Transformer is used to process long-distance dependencies and process complex or long texts. By dynamically adjusting the weights of these two models, the system adaptively selects the most appropriate processing method based on the characteristics of the input text. For example, for simple question-answer pairs, the system uses CNN; and for complex professional questions, it uses the capabilities of Transformer. This flexible structure enables the model to maintain high efficiency in various types of texts.

[0072] (3) The present invention introduces a text complexity evaluation process, which comprehensively considers the text length and the number of knowledge points. The text complexity evaluation method not only considers the length of the text, but also evaluates the number of knowledge points contained in the text. This method can more comprehensively reflect the actual complexity of the text. For example, a short text may contain multiple professional terms or complex concepts, so its complexity may be higher than a longer text with simple content. Through this evaluation, the system can more accurately decide how to process each input, thereby optimizing resource allocation and processing strategies. BRIEF DESCRIPTION OF THE DRAWINGS

[0073] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:

[0074] Figure 1 It is a flowchart of the authenticity judgment method of the adaptive architecture corpus based on dynamic fusion of the present invention;

[0075] Figure 2 It is a system block diagram of the authenticity assessment of adaptive architecture corpus based on dynamic fusion of the present invention;

[0076] Figure 3 This is a relationship diagram of the network structure of CNN and Transformer dynamic fusion according to the present invention. DETAILED DESCRIPTION

[0077] The technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention; it is obvious that the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments, and all other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making creative work are within the scope of protection of the present invention.

[0078] The structural features of the present invention are now described in detail with reference to the accompanying drawings.

[0079] See also Figure 1-Figure 3 , an adaptive architecture corpus authenticity assessment system based on dynamic fusion, including: data preparation module, text complexity assessment module, dynamic adjustment module, CNN module, Transformer module, fusion module, classification head and meta-controller. Among them, the data preparation module is responsible for the generation and annotation of initial data. The specific steps of generating and annotating initial data are as follows:

[0080] a. Use the llama3 large model to generate initial question-answer pairs based on industry or company system documents. b. Randomly select 20% of the samples and use the Claude3.5 Sonnet model for preliminary review to give a binary classification judgment on whether the corpus contains hallucinations.

[0081] c. The expert team conducted a secondary review based on the results of Claude3.5 Sonnet and obtained the final high-quality annotation results.

[0082] The text complexity evaluation module is responsible for evaluating the complexity of the input text. It is connected to the input layer and the dynamic adjustment module. It displays two submodules: character count evaluation and knowledge point count evaluation. Specifically, it includes two evaluation paths:

[0083] Path 1: Evaluate the path based on the number of characters:

[0084] 11) Count the number of characters in all corpus records and obtain the minimum value a1, quartile b1, average c1, three-quarter digit d1, and maximum value e1.

[0085] 12) According to the range of the number of characters in the corpus, different complexity values ​​complex_value1 (0.25, 0.50, 0.75, 1.00) are assigned.

[0086] Path 2: Evaluation path based on the number of knowledge points:

[0087] 21) Count the number of knowledge points of the text information of each corpus source, and obtain the minimum value a2, quartile b2, average c2, three-quarter digit d2, and maximum value e2 of the knowledge points of all corpora.

[0088] 22) According to the interval in which the number of knowledge points falls, different complexity values ​​complex_value2 (0.25, 0.50, 0.75, 1.00) are assigned.

[0089] Path 3: Comprehensive assessment path:

[0090] 31) Calculate the sum of complex_value1+complex_value2.

[0091] 32) According to the sum value, the corpus is divided into four categories: uncomplex text (≤0.5), medium-complex text (0.5-1), medium-to-high-complex text (1-1.5), and complex text (>1.5).

[0092] Dynamic adjustment module, which is used to dynamically adjust the weights of CNN and Transformer according to the text complexity evaluation results. The dynamic adjustment module is connected to the path of the CNN module and the Transformer module, and the arrows for weight adjustment are displayed. The thickness indicates the weight size. Specifically, the weights are divided into:

[0093] Simple text: CNN weight 0.8, Transformer weight 0.2;

[0094] Medium complexity text: CNN weight 0.6, Transformer weight 0.4;

[0095] Medium to high complexity text: CNN weight 0.4, Transformer weight 0.6;

[0096] Complex text: CNN weight 0.2, Transformer weight 0.8.

[0097] CNN module, which adopts a multi-channel CNN structure and includes:

[0098] Embedding layer, which converts input text into word vectors;

[0099] Multiple parallel convolutional layers are used to extract local features using convolution kernels of different sizes, including multiple convolution kernels of different sizes (such as 3x3, 4x4, 5x5);

[0100] The maximum pooling layer is used to perform maximum pooling on the convolution results of each channel and pool the output of each convolution channel;

[0101] The concatenation layer is used to concatenate all pooling results into one vector.

[0102] Transformer module, which adopts the standard Transformer encoder structure, including:

[0103] Position encoding layer, used to add position information to the input sequence;

[0104] Multi-head self-attention layer, used to capture long-range dependencies in the sequence, can display multiple attention heads;

[0105] A feed-forward neural network layer to further process the output of self-attention;

[0106] Layer Normalization layer, applied after each sub-layer to stabilize the training process.

[0107] Fusion module, which is used to fuse the outputs of CNN and Transformer, displays the gating mechanism, fuses the two outputs, and adopts the gating mechanism. Specifically:

[0108] Calculate the gate value g = σ(W_g*[CNN_output; Transformer_output]+b_g);

[0109] Fusion result = g*CNN_output+(1-g)*Transformer_output

[0110] Among them, σ is the sigmoid function, W_g and b_g are learnable parameters.

[0111] The classification head is used for the final prediction result, including a fully connected layer and a Softmax layer. The fully connected layer is used to map the fused features to a binary classification space, and the Softmax layer is used to output the final classification probability.

[0112] The meta-controller is used to learn the optimal architecture combination strategy during training. The bidirectional arrows connect all the main modules, indicating the optimization and adjustment of the overall architecture.

[0113] The meta controller includes:

[0114] The policy network module is used to generate architecture adjustment strategies based on the current input and model state;

[0115] Reward function module, which defines rewards based on model performance and computational efficiency;

[0116] Reinforcement learning algorithm module, used to optimize the policy network.

[0117] This embodiment also proposes a method for judging the authenticity of adaptive architecture corpus based on dynamic fusion, which specifically includes the following steps:

[0118] S1, use the labeled data generated by the data preparation module for initial training;

[0119] S2. In each initial training batch:

[0120] a. Use the text complexity evaluation module to evaluate the input text;

[0121] b. Use the dynamic adjustment module to adjust the weights of CNN and Transformer according to the evaluation results;

[0122] c. Process the input through the CNN module and the Transformer module respectively;

[0123] d. Use the fusion module to merge the outputs of the CNN module and the Transformer module;

[0124] e. Obtain the final prediction authenticity judgment result through the classification head;

[0125] f. Calculate the loss and back-propagate to update the parameters;

[0126] S3, use the meta-controller to periodically evaluate the current architecture and make adjustment recommendations;

[0127] S4. Fine-tune the model architecture based on the meta-controller’s suggestions.

[0128] The adaptive architecture corpus authenticity judgment system and method based on dynamic fusion of the present invention improves the accuracy of judgment on the one hand, specifically: by integrating the advantages of the two models of CNN and Transformer, the accuracy of judgment is significantly improved. CNN is good at capturing local features, such as specific word combinations or phrase patterns, which is very helpful for identifying common local errors of large models. Transformer is good at processing long-distance dependencies, can understand the context of the entire text, and helps to judge the coherence and rationality of the overall semantics. The combination of these two capabilities enables the system to comprehensively analyze the text, thereby more accurately judging the authenticity of the content. Experimental data show that compared with using CNN or Transformer alone, the judgment accuracy of this system is improved by 15-20%, which means that in practical applications, it can more reliably identify false content generated by large models. On the other hand, it enhances adaptability, specifically: the text complexity evaluation module of the present invention can analyze the characteristics of the input text, such as length and knowledge point density, and then the dynamic adjustment module automatically adjusts the weights of CNN and Transformer according to these characteristics. For example, for short and simple texts, the system will give CNN a higher weight; for long and complex texts, the weight of Transformer will be increased. This adaptive mechanism enables the system to effectively process various types of texts, whether simple questions and answers or complex policy explanations. In practical applications, this adaptability enables the system to maintain stable performance and improve adaptability by about 30% when processing diverse industry or company system documents. This means that the system can better cope with texts of different styles and complexities, greatly expanding its scope of application. Furthermore, the present invention can improve computing efficiency. Specifically, the dynamic adjustment mechanism of the present invention not only improves adaptability, but also optimizes the allocation of computing resources. For simple texts, the system will allocate more resources to CNN with fast computing speed; for complex texts, the computing weight of Transformer will be increased. This intelligent allocation mechanism avoids unnecessary computing overhead. For example, for a simple yes-or-no question, using a complex Transformer structure may waste computing resources. The present system can be adjusted as needed, thereby improving efficiency while maintaining performance. Actual tests show that compared with the fixed structure model, the average computing time of the present system is reduced by 25% while maintaining the same performance. This means that when processing a large amount of text, the processing speed can be significantly improved, saving time and computing resources. In addition, the present invention can also reduce the cost of manual review. The high accuracy and automation characteristics of the present system greatly reduce the amount of content that requires manual review. In traditional methods, large amounts of text require manual review to ensure accuracy, which is time-consuming and costly. However, this system can automatically and efficiently determine the authenticity of content generated by large models, and manual intervention is only required when the system cannot make a high-confidence judgment.Practical data shows that the system can automatically process about 80% of the content, and only 20% requires further manual review. This means that human resources can be focused on handling more complex cases that require professional judgment, greatly improving work efficiency while significantly reducing labor costs. At the same time, the system scalability can be improved. The meta-controller module in the present invention gives the system the ability to self-optimize and adapt to new environments. It can automatically adjust the system architecture according to new data and task requirements, such as changing the convolution kernel size of CNN or the number of layers of Transformer. This adaptive ability enables the system to quickly adjust and optimize its own structure when faced with new types of text or updated language models without the need for complete redesign and training. Practice shows that this mechanism reduces the system's adaptation time by about 40% when dealing with new tasks. This greatly improves the scalability and versatility of the system, enabling it to continuously adapt to changing language environments and task requirements. Enhanced interpretability. Unlike many AI models that are regarded as "black boxes", this system provides higher interpretability by analyzing the weight distribution of CNN and Transformer modules. Users can understand which features the system relies more on when making judgments: local word patterns (CNN) or overall semantic structures (Transformer). This transparency makes the system's decision-making process easier to understand and verify. In practical applications, especially in scenarios involving important decisions, this explainability greatly increases users' trust in the system. It also provides valuable insights for further improving the system, allowing researchers to better understand and optimize the behavior of the model.

[0129] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention is described in detail with reference to the aforementioned embodiments, those skilled in the art can still modify the technical solutions described in the aforementioned embodiments or replace some of the technical features therein by equivalents. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A method for judging the authenticity of adaptive architecture corpus based on dynamic fusion, characterized in that: The steps include: Step 1) Use the labeled data generated by the data preparation module for initial training; Step 2) In each initial training batch: a. Use the text complexity evaluation module to evaluate the input text; b. Use the dynamic adjustment module to adjust the weights of CNN and Transformer according to the evaluation results; c. Process the input through the CNN module and the Transformer module respectively; d. Use the fusion module to merge the outputs of the CNN module and the Transformer module; e. Obtain the final prediction authenticity judgment result through the classification head; f. Calculate the loss and back-propagate to update the parameters; Step 3) Use the meta-controller to periodically evaluate the current architecture and make adjustment recommendations; Step 4) Fine-tune the model architecture based on the meta-controller’s suggestions.

2. The method for determining authenticity of adaptive architecture corpus based on dynamic fusion according to claim 1 is characterized in that: In step 1, the specific steps of initial data generation and annotation are as follows: a) Use the llama3 big model to generate initial question-answer pairs based on industry or company system documents; b) Randomly select 20% of the samples and use the Claude3.5 Sonnet model for preliminary review to give a binary classification judgment on whether the corpus contains hallucinations. c) The expert team module conducts a secondary review based on the results of the Claude3.5 Sonnet model to obtain the final high-quality annotation results.

3. The method for determining authenticity of adaptive architecture corpus based on dynamic fusion according to claim 1 is characterized in that: In step a of step 2), the text complexity evaluation module evaluation includes the following evaluation path: Path 1: Evaluate the path based on the number of characters, including the following steps: 11) Count the number of characters in all corpus records and obtain the minimum value a1, quartile b1, average c1, three-quarter digit d1, and maximum value e1; 12) According to the range of the number of characters in the corpus, different complexity values ​​complex_value1 (0.25, 0.50, 0.75, 1.00) are assigned; Path 2: Evaluation path based on the number of knowledge points, including the following steps: 21) Count the number of knowledge points of each text information of the corpus source, and obtain the minimum value a2, quartile b2, average c2, three-quarter digit d2, and maximum value e2 of all corpus knowledge points; 22) According to the interval in which the number of knowledge points falls, different complexity values ​​complex_value2 (0.25, 0.50, 0.75, 1.00) are assigned; Path 3: Comprehensive assessment path, including the following steps: 31) Calculate the sum of complex_value1+complex_value2; 32) According to the sum value, the corpus is divided into four categories: uncomplex text (≤0.5), medium-complex text (0.5-1), medium-to-high-complex text (1-1.5), and complex text (>1.5).

4. The method for determining authenticity of adaptive architecture corpus based on dynamic fusion according to claim 1 is characterized in that: In step b of step 2), the dynamic adjustment module dynamically adjusts the weights of CNN and Transformer including: a. Simple text: CNN weight 0.8, Transformer weight 0.2; b. Medium complexity text: CNN weight 0.6, Transformer weight 0.4; c. Medium to high-complexity text: CNN weight 0.4, Transformer weight 0.6; d. Complex text: CNN weight 0.2, Transformer weight 0.

8.

5. A dynamic fusion-based adaptive architecture corpus authenticity assessment system, used in the dynamic fusion-based adaptive architecture corpus authenticity assessment method as claimed in any one of claims 1 to 4, characterized in that: include: Data preparation module, used for generating and labeling initial data; A text complexity evaluation module, used to evaluate the complexity of the input text; Dynamic adjustment module, used to dynamically adjust the weights of CNN and Transformer according to the text complexity evaluation results; CNN module, adopts multi-channel CNN structure; Transformer module, which uses the standard Transformer encoder structure; Fusion module, used to fuse the outputs of CNN and Transformer; Classification head, used for the final prediction results, including the fully connected layer and the Softmax layer; The fully connected layer is used to map the fused features to a binary classification space, and the Softmax layer is used to output the final classification probability; A meta-controller is used to learn the optimal architecture combination strategy during training.

6. The dynamic fusion-based adaptive architecture corpus authenticity assessment system according to claim 5 is characterized in that: The CNN module includes: Embedding layer, which converts input text into word vectors; Multiple parallel convolutional layers are used to extract local features using convolution kernels of different sizes; The maximum pooling layer is used to perform maximum pooling on the convolution results of each channel; The concatenation layer is used to concatenate all pooling results into one vector.

7. The dynamic fusion-based adaptive architecture corpus authenticity assessment system according to claim 5 is characterized in that: The Transformer module includes: Position encoding layer, used to add position information to the input sequence; Multi-head self-attention layer to capture long-range dependencies in the sequence; A feed-forward neural network layer to process the output of self-attention; Layer Normalization layer, applied after each sub-layer to stabilize the training process.

8. The adaptive architecture corpus authenticity assessment system based on dynamic fusion according to claim 5 is characterized in that: The fusion module adopts a gating mechanism, including: Calculate the gate value g = σ (W_g*[CNN_output; Transformer_output] + b_g) Fusion result = g*CNN_output+(1-g)*Transformer_output Among them, σ is the sigmoid function, W_g and b_g are learnable parameters.

9. The adaptive architecture corpus authenticity assessment system based on dynamic fusion according to claim 5 is characterized in that: The meta-controller comprises: The policy network module is used to generate architecture adjustment strategies based on the current input and model state; Reward function module, which defines rewards based on model performance and computational efficiency; Reinforcement learning algorithm module, used to optimize the policy network.

Citation Information

Cited By

  • Rich text structured information extraction method and device and storage medium

    CN119557366A

  • Method, device and storage medium for extracting structured information from rich text

    CN119557366B