Method and terminal for improving fair competition review recognition accuracy based on multi-level feature extraction
By employing a multi-level feature extraction and weighted fusion approach, the problems of low efficiency and low accuracy in fair competition review were solved, resulting in more efficient and accurate review outcomes.
Patent Information
- Application Number
- CN202511471586.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-10-15
- Publication Date
- 2026-02-10
AI Technical Summary
Existing fair competition review methods are inefficient, time-consuming, costly, and susceptible to subjective judgment. Machine recognition is difficult to adapt to complex semantics and has low accuracy in scenarios with few samples.
A multi-level feature extraction algorithm is adopted, combined with regular expressions and syntax analysis, to construct a hybrid model and perform weighted fusion. By dynamically adjusting the weights, the identification accuracy of fair competition review is optimized.
It improves the accuracy of fair competition review identification, enhances the adaptability to complex semantics, reduces the discrepancy between the model and the rules, and optimizes the output results.
Smart Images
Figure CN121502000A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The application belongs to the field of fair competition review, and particularly relates to a method for improving the identification accuracy of fair competition review based on multi-level feature extraction and a terminal. BACKGROUND
[0002] In drafting measure documents, bidding documents and the like related to the economic activities of operators, fair competition review must be carried out before they can be published, aiming to promote market fair competition and optimize the business environment. The existing review methods are mostly based on traditional manual identification, or use data processing, cascade deep learning model-based monitoring, big data and artificial intelligence technology-based monitoring and the like.
[0003] The traditional manual identification method has low review efficiency, and depends on the review personnel to analyze the reviewed text item by item, which is time-consuming and costly, and the review work result is easily affected by subjective judgment, and the review work cannot be objectively evaluated.
[0004] However, the methods based on various machine identification technologies also have the following shortcomings:
[0005] 1. The data processing-based system is difficult to adapt to complex semantics;
[0006] 2. The cascade deep learning model-based or big data and artificial intelligence technology-based monitoring model method depends on the number of samples, while fair competition review needs to make violation judgments from 66 aspects; however, in actual operation, more than 50% of the violation contents appear in few scenes, and the corresponding samples are few, thereby leading to low accuracy of the model, and finally a large amount of manual identification judgment is still needed. SUMMARY
[0007] In view of the deficiencies of the prior art, the purpose of the present application is to provide a method for improving the identification accuracy of fair competition review based on multi-level feature extraction and a terminal, which applies a multi-level feature extraction algorithm to the fair competition review of a document, realizes multi-level feature extraction of the document, and performs weighted fusion in model training, thereby improving the feature identification accuracy of the fair competition review of the measure document, and effectively solving the problems of low manual identification efficiency, high time and cost, and results easily affected by subjective judgment in the existing fair competition review, and the problems of difficulty of machine identification to adapt to complex semantics and low accuracy of few sample scenes due to dependence on the number of samples.
[0008] To achieve the above purpose, the technical scheme adopted by the present application is as follows:
[0009] The method for improving the identification accuracy of fair competition review based on multi-level feature extraction of the present application has the following steps:
[0010] 1) The structural features of the text are extracted by combining regular expressions and syntax analysis;
[0011] 2) Build a hybrid model based on the basic feature extractor-timing processing layer and train it; identify the key restrictive provisions in the text;
[0012] 3) Build a feature review strategy for fair competition review, dynamically adjust the weight based on rule hits, convert each rule feature into a unified dimension comprehensive feature for rule feature fusion;
[0013] 4) Perform multi-modal feature splicing, calculate the score of the trained hybrid model and the rule score using a weighted fusion method, and optimize the weighted fusion effect by dynamically adjusting the weight to improve the recognition accuracy of fair competition review.
[0014] Further, the step 1) specifically comprises: using regular expression to match a certain format content in the text, using text structure feature extraction function and counting function to count the number of provisions in the text to be reviewed, calculating the paragraph density by the ratio of text line segmentation number to word segmentation number, and combining the number of provisions in the text to be reviewed and the paragraph density into a two-dimensional structure feature array; using a syntax analysis tool to split the input text into multiple words or sub-word units, and processing long text according to the maximum sequence length limit dynamic truncation strategy;
[0015] Adding a label of overall semantics at the beginning of the truncated sequence and adding a label of separating text at the end of the truncated sequence, and then converting all labels into input ID sequence and segment ID sequence;
[0016] Input the input ID sequence and segment ID sequence into the BERT pre-training language model to obtain the semantic vector of each label, and then perform average pooling on the semantic vector of each label to obtain a fixed-length vector as the hierarchical semantic representation of the input text, thereby realizing the extraction of the structure feature of the text.
[0017] Further, the step 2) specifically comprises:
[0018] Using the BERT pre-training language model as the basic feature extractor, setting the weight, outputting the embedding vector, and setting all parameters to be fine-tunable; using the BiLSTM timing processing layer to process the forward and reverse structures of the sequence, the number of hidden units, the number of layers, and the input dimension vector to extract the sequence timing feature; performing weighted summation of the weight and the output to enhance the key provision feature, obtaining a hybrid model based on the basic feature extractor-timing processing layer, and finally outputting the classification result through the full connection layer as the extracted text semantic feature;
[0019] Set the learning frequency, batch size, linear learning rate decay strategy, training period and early stopping mechanism to train the hybrid model.
[0020] Further, the step 3) specifically comprises:
[0021] The qualitative features in the basic information feature, the violation type feature, and the influence degree feature of the fair competition review are encoded, and the quantitative features are normalized, so as to realize the construction of the feature review strategy of the fair competition review as the basic weight of the target rule;
[0022] The sensitive words of the fair competition review are combined by using the domain sensitive word library to realize the quick scanning of the keyword mode, and each regular expression is matched with the corresponding field category. The regular matching algorithm is used to scan the text fragments in the text to be reviewed which match the regular expression mode, and the number of matches of each regular expression mode is counted to form a text rule feature array as the text rule feature, so as to realize efficient pattern matching;
[0023] Based on the regular expression rule hit history of each domain category, the weight of the target rule is calculated according to the hit accuracy, and the calculation formula is:
[0024] The weight of the target rule = the basic weight + the accuracy weight coefficient x the hit accuracy;
[0025] The text structure feature, the text semantic feature, and the text rule feature are spliced in a predetermined order to form a comprehensive feature vector and output, realizing rule feature fusion.
[0026] Further, the step 4) specifically includes:
[0027] Each feature corresponds to a modal feature information, the feature is converted into a column vector, each column vector feature is fitted and converted to obtain a standardized feature, each standardized feature is flattened into a one-dimensional vector, and each one-dimensional standardized feature vector after flattening is spliced to realize multi-modal feature splicing to obtain a fusion feature vector;
[0028] The fusion feature vector is linearly transformed, that is, the input dimension is set as the dimension of the fusion feature vector, and the output dimension is the number of classification categories, which is used to map the fusion feature to a category, and the initial value of the set learnable parameter is set as 1, and the obtained category probability distribution is used as the score of the trained hybrid model;
[0029] The weighted fusion method is used to calculate the score of the trained hybrid model and the weight score of the target rule to obtain the final score, and the expression is:
[0030] ;
[0031] Wherein, sigma represents the sigmoid function, which is used to map the fusion result to the interval of 0-1; The initial weight of the model score is The initial weight of the rule score is The model score is output for classification. The weighted score of the target rule in step 3);
[0032] Based on the historical decision-making effects, the model prediction accuracy and rule evaluation accuracy are evaluated periodically, and the weights of model scores and rule scores are adjusted. When the model prediction accuracy decreases, the weight of rule scores is increased and the weight of model scores is decreased. When the rule base is updated, the weight of rule scores is decreased and the weight of model scores is increased. By dynamically adjusting the weights, the weighted fusion effect is optimized to improve the accuracy of fair competition review.
[0033] This invention also provides a terminal for improving the accuracy of fair competition review identification based on multi-level feature extraction, comprising:
[0034] The extraction module is used to extract the structural features of text by combining regular expressions and syntax analysis.
[0035] The recognition module is used to build and train a hybrid model based on a basic feature extractor and a temporal processing layer; it identifies key restrictive clauses in the text.
[0036] The fusion module is used to construct a feature review strategy for fair competition review. It dynamically adjusts the weights based on the historical rule hits and converts the features of each rule into a comprehensive feature of a unified dimension for rule feature fusion.
[0037] The optimization module is used to perform multimodal feature splicing. It calculates the score of the trained hybrid model and the rule score using a weighted fusion method. By dynamically adjusting the weights, the weighted fusion effect is optimized to improve the identification accuracy of fair competition review.
[0038] The beneficial effects of this invention are:
[0039] This invention enhances the adaptability to complex semantics through multi-level feature extraction (structural + semantic + document-level); the hybrid model of basic feature extractor and temporal processing layer enhances the recognition ability of key restrictive clauses; rule feature fusion (domain sensitive word library, dynamic weight adjustment) supplements the feature information of low-sample scenarios, solving the problem of low accuracy in low-sample scenarios caused by machine recognition relying on the number of samples; multimodal feature concatenation, weighted fusion based on historical decisions, and dynamic weight adjustment improve the overall recognition accuracy, reduce the difference between the model and the rules, further optimize the output results, and thus improve the overall accuracy of fair competition review. Attached Figure Description
[0040] Figure 1 This is a schematic diagram of the method of the present invention. Detailed Implementation
[0041] To facilitate understanding by those skilled in the art, the present invention will be further described below with reference to embodiments and accompanying drawings. The content mentioned in the embodiments is not intended to limit the present invention.
[0042] Reference Figure 1 As shown, the present invention provides a method for improving the accuracy of fair competition review identification based on multi-level feature extraction, comprising the following steps:
[0043] 1) The structural features of the text are extracted by combining regular expressions and syntax analysis. Specifically, this includes: using regular expressions to match a certain format content in the text, using text structure feature extraction functions and counting functions to count the number of clauses in the text under review, calculating the ratio of the number of line breaks to the number of word breaks to obtain the paragraph density, and combining the number of clauses in the text under review and the paragraph density into a two-dimensional structural feature array; using syntax analysis tools to split the input text into multiple word or sub-word units, and processing long texts according to a dynamic truncation strategy based on the maximum sequence length limit.
[0044] Add a marker for the overall semantics at the beginning of the truncated sequence and a marker for the text separator at the end of the truncated sequence. Then convert all markers into an input ID sequence and a segment ID sequence.
[0045] Input the input ID sequence and segment ID sequence into the BERT pre-trained language model to obtain the semantic vector of each tag. Then, average pooling (taking the average of all vectors) is performed on the semantic vector of each tag to obtain a fixed-length vector, which serves as the hierarchical semantic representation of the input text, thereby realizing the extraction of the structural features of the text.
[0046] 2) Construct and train a hybrid model based on a basic feature extractor and a temporal processing layer; identify key restrictive clauses in the text; specifically including:
[0047] We employ a BERT pre-trained language model as the basic feature extractor, setting weights and outputting embedding vectors, with all parameters set to be fine-tunable. A BiLSTM temporal processing layer is used to process the forward and reverse structures of the sequence, hiding the number of units and layers, and extracting temporal features from the input dimensional vector. We then weight and sum the outputs to enhance key clause features, resulting in a hybrid model based on basic feature extraction and temporal processing. Finally, a fully connected layer outputs the classification results as the extracted text semantic features.
[0048] The hybrid model is trained by setting the learning frequency, batch size, linear learning rate decay strategy, training period, and early stopping mechanism.
[0049] 3) Construct a feature-based review strategy for fair competition review, dynamically adjust weights based on historical rule hits, and convert various rule features into comprehensive features of a unified dimension for rule feature fusion; specifically including:
[0050] The qualitative features of the basic information features, violation type features, and impact degree features of fair competition review are encoded, and the quantitative features are normalized to realize the feature review strategy for fair competition review, which serves as the basic weight of the target rules.
[0051] The system achieves rapid scanning of keyword patterns by using a domain-sensitive word library. Sensitive words for fair competition review are combined, and each regular expression is assigned to a domain category. A regular expression matching algorithm is used to scan text segments in the text to be reviewed that match the regular expression patterns, and the number of matches for each regular expression pattern is counted to form a text rule feature array, which serves as the text rule features to achieve efficient pattern matching.
[0052] Dynamically adjust weights based on the hit history of regular expression rules for each domain category: Obtain the historical hit records of the target rule, calculate the weight of the target rule based on the hit accuracy, and use the following formula:
[0053] The weight of the target rule = base weight + accuracy weight coefficient × hit accuracy;
[0054] Text structural features, text semantic features, and text rule features are concatenated in a preset order to form a comprehensive feature vector, which is then output to achieve rule feature fusion.
[0055] 4) Perform multimodal feature concatenation, and calculate the score of the trained hybrid model and the rule score using a weighted fusion method. Optimize the weighted fusion effect by dynamically adjusting the weights to improve the accuracy of fair competition review; specifically including:
[0056] Each feature corresponds to a modality of feature information. The features are converted into column vectors. Each column vector feature is fitted and transformed to obtain standardized features. Each standardized feature is flattened into a one-dimensional vector. The flattened one-dimensional standardized feature vectors are concatenated to achieve multimodal feature concatenation and obtain a fused feature vector.
[0057] The fused feature vector is linearly transformed, i.e. the input dimension is set to the dimension of the fused feature vector and the output dimension is the number of classification categories. This is used to map the fused features to categories. The initial value of the learnable parameter is set to 1. The resulting category probability distribution is used as the score of the fused model after training.
[0058] The final score is obtained by calculating the score of the trained hybrid model and the weight score of the target rule using a weighted fusion method, expressed as:
[0059] ;
[0060] Where σ represents the sigmoid function, which is used to map the fusion result to the 0-1 interval; The initial weights for the model score. The initial weights for the rule scores; The model score is output for classification. The weighted score of the target rule in step 3);
[0061] Based on historical decision-making results (obtained through regular manual inspections), the model prediction accuracy and rule evaluation accuracy are periodically evaluated, and the weights of model scores and rule scores are adjusted. When the model prediction accuracy decreases, the weight of rule scores is increased and the weight of model scores is decreased. When the rule base is updated, the weight of rule scores is decreased and the weight of model scores is increased. By dynamically adjusting the weights, the weighted fusion effect is optimized to improve the accuracy of fair competition review.
[0062] This invention also provides a terminal for improving the accuracy of fair competition review identification based on multi-level feature extraction, comprising:
[0063] The extraction module is used to extract the structural features of text by combining regular expressions and syntax analysis.
[0064] The recognition module is used to build and train a hybrid model based on a basic feature extractor and a temporal processing layer; it identifies key restrictive clauses in the text.
[0065] The fusion module is used to construct a feature review strategy for fair competition review. It dynamically adjusts the weights based on the historical rule hits and converts the features of each rule into a comprehensive feature of a unified dimension for rule feature fusion.
[0066] The optimization module is used to perform multimodal feature splicing. It calculates the score of the trained hybrid model and the rule score using a weighted fusion method. By dynamically adjusting the weights, the weighted fusion effect is optimized to improve the identification accuracy of fair competition review.
[0067] This invention has many specific applications. The above description is only a preferred embodiment of this invention. It should be noted that for those skilled in the art, several improvements can be made without departing from the principle of this invention, and these improvements should also be considered within the scope of protection of this invention.
Claims
1. A method for improving the accuracy of fair competition review identification based on multi-level feature extraction, characterized in that, The steps are as follows: 1) Extract structural features of text by combining regular expressions with syntax analysis; 2) Construct and train a hybrid model based on a basic feature extractor and a temporal processing layer; identify key restrictive clauses in the text; 3) Construct a feature-based review strategy for fair competition review, dynamically adjust weights based on historical rule hits, and convert each rule feature into a unified dimension of comprehensive features for rule feature fusion; 4) Perform multimodal feature splicing, and use a weighted fusion method to calculate the score of the trained hybrid model and the rule score. By dynamically adjusting the weights, optimize the weighted fusion effect to improve the identification accuracy of fair competition review.
2. The method for improving the accuracy of fair competition review identification based on multi-level feature extraction according to claim 1, characterized in that, Step 1) specifically includes: using regular expressions to match a certain format content in the text, using text structure feature extraction functions and counting functions to count the number of text clauses under review, calculating the ratio of the number of text line breaks to the number of word breaks to obtain the paragraph density, and combining the number of text clauses under review and the paragraph density into a two-dimensional structure feature array; using a syntax analysis tool to split the input text into multiple words or sub-word units, and processing long texts according to a dynamic truncation strategy based on the maximum sequence length limit; Add a marker for the overall semantics at the beginning of the truncated sequence and a marker for the text separator at the end of the truncated sequence. Then convert all markers into an input ID sequence and a segment ID sequence. The input ID sequence and segment ID sequence are fed into the BERT pre-trained language model to obtain the semantic vector of each tag. Then, the semantic vector of each tag is averaged and pooled to obtain a fixed-length vector, which serves as the hierarchical semantic representation of the input text, thereby extracting the structural features of the text.
3. The method for improving the accuracy of fair competition review identification based on multi-level feature extraction according to claim 1, characterized in that, Step 2) specifically includes: We employ a BERT pre-trained language model as the basic feature extractor, setting weights and outputting embedding vectors, with all parameters set to be fine-tunable. A BiLSTM temporal processing layer is used to process the forward and reverse structures of the sequence, hiding the number of units and layers, and extracting temporal features from the input dimensional vector. We then weight and sum the outputs to enhance key clause features, resulting in a hybrid model based on basic feature extraction and temporal processing. Finally, a fully connected layer outputs the classification results as the extracted text semantic features. The hybrid model is trained by setting the learning frequency, batch size, linear learning rate decay strategy, training period, and early stopping mechanism.
4. The method for improving the accuracy of fair competition review identification based on multi-level feature extraction according to claim 1, characterized in that, Step 3) specifically includes: The qualitative features of the basic information features, violation type features, and impact degree features of fair competition review are encoded, and the quantitative features are normalized to realize the feature review strategy for fair competition review, which serves as the basic weight of the target rules. The system achieves rapid scanning of keyword patterns by using a domain-sensitive word library. Sensitive words for fair competition review are combined, and each regular expression is assigned to a domain category. A regular expression matching algorithm is used to scan text segments in the text to be reviewed that match the regular expression patterns, and the number of matches for each regular expression pattern is counted to form a text rule feature array, which serves as the text rule features to achieve efficient pattern matching. Dynamically adjust weights based on the hit history of regular expression rules for each domain category: Obtain the historical hit records of the target rule, calculate the weight of the target rule based on the hit accuracy, and use the following formula: The weight of the target rule = base weight + accuracy weight coefficient × hit accuracy; Text structural features, text semantic features, and text rule features are concatenated in a preset order to form a comprehensive feature vector, which is then output to achieve rule feature fusion.
5. The method for improving the accuracy of fair competition review identification based on multi-level feature extraction according to claim 1, characterized in that, Step 4) specifically includes: Each feature corresponds to a modality of feature information. The features are converted into column vectors. Each column vector feature is fitted and transformed to obtain standardized features. Each standardized feature is flattened into a one-dimensional vector. The flattened one-dimensional standardized feature vectors are concatenated to achieve multimodal feature concatenation and obtain a fused feature vector. The fused feature vector is linearly transformed, i.e. the input dimension is set to the dimension of the fused feature vector and the output dimension is the number of classification categories. This is used to map the fused features to categories. The initial value of the learnable parameter is set to 1. The resulting category probability distribution is used as the score of the fused model after training. The final score is obtained by calculating the score of the trained hybrid model and the weight score of the target rule using a weighted fusion method, expressed as: ； Where σ represents the sigmoid function, which is used to map the fusion result to the 0-1 interval; The initial weights for the model score. The initial weights for the rule scores; The model score is output for classification. The weighted score of the target rule in step 3); Based on the historical decision-making effects, the model prediction accuracy and rule evaluation accuracy are evaluated periodically, and the weights of model scores and rule scores are adjusted. When the model prediction accuracy decreases, the weight of rule scores is increased and the weight of model scores is decreased. When the rule base is updated, the weight of rule scores is decreased and the weight of model scores is increased. By dynamically adjusting the weights, the weighted fusion effect is optimized to improve the accuracy of fair competition review.
6. A terminal for improving the accuracy of fair competition review identification based on multi-level feature extraction, characterized in that, include: The extraction module is used to extract the structural features of text by combining regular expressions and syntax analysis. The recognition module is used to build and train a hybrid model based on a basic feature extractor and a temporal processing layer; it identifies key restrictive clauses in the text. The fusion module is used to construct a feature review strategy for fair competition review. It dynamically adjusts the weights based on the historical rule hits and converts the features of each rule into a comprehensive feature of a unified dimension for rule feature fusion. The optimization module is used to perform multimodal feature splicing. It calculates the score of the trained hybrid model and the rule score using a weighted fusion method. By dynamically adjusting the weights, the weighted fusion effect is optimized to improve the identification accuracy of fair competition review.