National standard bid collection evaluation method based on multilayer cross recombination attention

Abstract

The invention discloses a national standard bid acquisition evaluation method based on multilayer cross recombination attention, and belongs to the technical field of national standard bid acquisition evaluation, and the method comprises the steps: extracting the text content in a to-be-evaluated international standard file, and splitting the text content into a plurality of parts of text data; inputting each piece of extracted text data into a multi-layer cross recombination attention index classification model to obtain a prediction indication feature corresponding to the text data; the multi-layer cross recombination attention index classification model comprises a data coding and feature mapping module, a deep cross attention module, a multi-scale fusion cross attention module, an attention cross recombination module and an index indication module; and performing score calculation on the international standardized file according to the prediction indication characteristics corresponding to all text data in the same international standardized file to obtain an evaluation result in the to-be-evaluated international standardized file. According to the invention, the matching degree between the international standard and the domestic standard can be automatically and accurately evaluated by using the deep learning algorithm.

Description

Technical Field

[0001] The present invention relates to the technical field of national standard adoption and evaluation, and more specifically to a national standard adoption and evaluation method based on multi-layer cross-reorganization attention. Background Art

[0002] In the process of converting international standards into national standards, a scientific evaluation process is needed to determine whether the international standards are suitable for conversion into domestic standards.

[0003] With the development of artificial intelligence technologies such as deep learning, traditional standard adoption and evaluation methods are gradually shifting towards intelligence and automation. Using deep learning algorithms to classify the national standard adoption and evaluation indicator system can complete this evaluation work efficiently and accurately, providing strong technical support for decision makers. By building an artificial intelligence evaluation model, it is possible to process and analyze a large amount of standardized data, automatically identify and classify key indicators, and thus optimize the national standard adoption and evaluation process. This method can not only improve the accuracy and efficiency of the evaluation, but also provide a more scientific basis for the formulation and revision of domestic standards.

[0004] On the one hand, the traditional standard assessment method relies on manual analysis and judgment, which is not only time-consuming and labor-intensive, but also easily affected by subjective factors, resulting in limited accuracy and consistency of the assessment results.

[0005] On the other hand, the standard adoption assessment needs to consider not only technical compatibility, but also economic, social, environmental and other factors. However, the existing assessment system or method often finds it difficult to achieve comprehensive and in-depth analysis when processing these multi-dimensional data, resulting in the assessment results not being able to fully reflect the comprehensive impact of standard adoption.

[0006] Therefore, how to use deep learning algorithms to provide an efficient and accurate national standard adoption assessment method is an urgent problem that technical personnel in this field need to solve. Summary of the invention

[0007] In view of this, the present invention provides a national standard adoption assessment method based on multi-layer cross-reorganization attention, which is used to at least solve some of the technical problems in the background technology.

[0008] In order to achieve the above object, the present invention adopts the following technical solution:

[0009] A national standard adoption evaluation method based on multi-layer cross-reorganization attention includes the following steps:

[0010] Extract the text content in the international standardization document to be evaluated, split the text content according to the sentence dimension, and obtain several copies of text data;

[0011] Input each extracted text data into a multi-layer cross-reorganization attention index classification model to obtain the prediction indication features corresponding to the text data; wherein the multi-layer cross-reorganization attention index classification model includes a data encoding and feature mapping module, a deep cross-attention module, a multi-scale fusion cross-attention module, an attention cross-reorganization module, and an indicator indication module;

[0012] The data encoding and feature mapping module is used to receive text data, and encode and map the text data to obtain corresponding mapping features; the deep cross-attention module and the multi-scale fusion cross-attention module respectively perform feature processing on the obtained mapping features; the attention cross-recombination module is used to receive the feature processing results of the deep cross-attention module and the feature processing results of the multi-scale fusion cross-attention module, and perform cross-recombination to obtain cross-recombination features; the indicator indication module is used to receive the cross-recombination features and output the prediction indication features corresponding to the text data;

[0013] The international standardized document is scored according to the prediction indicative features corresponding to all text data in the same international standardized document to obtain an evaluation result of the international standardized document to be evaluated.

[0014] Furthermore, the data encoding and feature mapping module includes: a BERT pre-training model, a first convolution unit, a first dimension mapping unit, a second convolution unit, a second dimension mapping unit, a third convolution unit, and a matrix multiplication unit:

[0015] Among them, the BERT pre-trained model is used to extract features from the input text data and obtain word vector text features; the first convolution unit is used to perform a convolution operation on the obtained word vector text features to obtain the convolution operation result of the first convolution unit; the convolution operation result passes through the first dimension mapping unit and the second convolution unit in sequence to obtain the first dimension mapping feature; the convolution operation result passes through the second dimension mapping unit and the third convolution unit in sequence to obtain the second dimension mapping feature; the first dimension mapping feature and the second dimension mapping feature are simultaneously input into the matrix multiplication unit to obtain the mapping feature corresponding to the text data.

[0016] Further, the deep cross attention module includes a splitting unit, a first cross attention unit, a second cross attention unit, a third cross attention unit and a splicing unit connected in sequence;

[0017] The splitting unit is used to split the mapping features output by the data encoding and feature mapping module along the channel dimension to obtain dimensional features of 12 different channel dimensions. , ;

[0018] The first cross attention unit is used to receive the dimension feature , and the dimension features Any three continuous dimensional features in , Perform cross attention operation and output 10 first cross operation features , ;

[0019] The second cross attention unit is used to receive the first cross operation feature , and the first crossover feature Any three continuous dimensional features in , Perform cross attention operation and output 8 second cross operation features , ;

[0020] The third cross attention unit is used to receive the second cross operation feature , and the second crossover feature Any three continuous dimensional features in , Perform cross attention operation and output 6 third cross operation features , ;

[0021] The concatenation unit is used to perform the 6 third crossover operation features of the output Perform feature splicing and output splicing features , and used as the feature processing result of the deep cross attention module.

[0022] Further, the cross attention operation in the first cross attention unit, the second cross attention unit and the third cross attention unit is implemented by the following steps:

[0023] Any three continuous dimensional features are recorded as , , , the first dimension feature of any three consecutive dimensional features Input the first convolution unit in the cross-attention algorithm model and output the feature ; The feature Input the first dimension reshape unit and output the feature ;

[0024] The first dimension feature of any three consecutive dimension features and the second dimension feature Input the first matrix multiplication unit to perform matrix multiplication, and output the feature , the feature After activation by the Softmax activation module, the output is the feature ;

[0025] The third dimension feature of any three consecutive dimensional features and Features Input the second matrix multiplication unit to perform matrix multiplication, and output the feature ; The feature Input the second convolution unit in the cross-attention algorithm model and output the feature ; The feature Input the second dimension reshape unit and output the feature ;

[0026] Reshape the output features of the unit along the first dimension and reshape the output features of the unit in the second dimension Perform matrix addition operation and output the features ;

[0027] The first dimension feature of any three consecutive dimension features and Features Input the third matrix multiplication unit to perform matrix multiplication, and output the feature ; The feature After activation by the Softmax activation module, the output is the feature ;

[0028] The second dimension feature of any three consecutive dimension features and Features Input the fourth matrix multiplication unit to perform matrix multiplication, and output the feature ; The feature After activation by the Softmax activation module, the output is the feature ;

[0029] The third dimension feature of any three consecutive dimensional features and Features Input the fifth matrix multiplication unit to perform matrix multiplication, and output the feature ;

[0030] The characteristics and Features Input matrix multiplication unit to perform matrix multiplication, and output feature ,feature That is, any three-dimensional features Cross-operation features after cross-attention operation.

[0031] Furthermore, the multi-scale fusion cross attention module includes thirteen convolution units from 1 to 13, two single-dimensional feature attention units, and a composite feature attention unit;

[0032] The first convolution unit in the multi-scale fusion cross attention module is used to perform a first convolution operation on the mapping features output by the data encoding and feature mapping module, and output a first convolution operation result;

[0033] The first convolution operation result is sequentially subjected to convolution operation by the fifth convolution unit, the sixth convolution unit and the seventh convolution unit in the multi-scale fusion cross attention module to obtain a fourth convolution operation result;

[0034] The first convolution operation result is sequentially convolved through the second convolution unit, the third convolution unit and the fourth convolution unit in the multi-scale fusion cross attention module to obtain a tenth convolution operation result;

[0035] The first convolution operation result is sequentially subjected to convolution operation by the second convolution unit, the eighth convolution unit and the ninth convolution unit in the multi-scale fusion cross attention module to obtain a seventh convolution operation result;

[0036] The first convolution operation result is sequentially subjected to convolution operation by the second convolution unit, the third convolution unit and the tenth convolution unit in the multi-scale fusion cross attention module to obtain a ninth convolution operation result;

[0037] Inputting the obtained fourth convolution operation result, seventh convolution operation result and ninth convolution operation result into the first single-dimensional feature attention unit to obtain a first single-dimensional attention feature;

[0038] Inputting the obtained tenth convolution operation result, the seventh convolution operation result and the ninth convolution operation result into the second single-dimensional feature attention unit to obtain a second single-dimensional attention feature;

[0039] The first single-dimensional attention feature and the second single-dimensional attention feature are obtained and simultaneously input into the composite feature attention unit for feature processing, thereby obtaining a feature processing result of the composite feature attention unit;

[0040] The feature processing results of the composite feature attention unit are sequentially convolved through three convolution units to obtain the final feature processing results of the multi-scale fusion cross attention module.

[0041] Furthermore, the first single-dimensional feature attention unit is used to perform feature processing on the input fourth convolution operation result, the seventh convolution operation result and the ninth convolution operation result to obtain a single-dimensional attention feature; specifically comprising the following steps:

[0042] Constructing a single-dimensional feature attention unit, wherein the constructed single-dimensional feature attention unit includes a maximum pooling layer, an average pooling layer, a convolutional layer, and three fully connected layers;

[0043] If the fourth convolution operation result, the seventh convolution operation result and the ninth convolution operation result of the input are recorded as features ,feature and Features ; then do the following:

[0044] The characteristics The maximum pooling layer and the first fully connected layer in the single-dimensional feature attention unit are input in sequence, and the feature ;

[0045] The characteristics The average pooling layer and the second fully connected layer in the single-dimensional feature attention unit are input in sequence, and the feature ;

[0046] The characteristics The convolutional layer and the third fully connected layer in the single-dimensional feature attention unit are input in sequence, and the feature ;

[0047] The characteristics and Features After the shape transformation, the matrix is ​​input into the matrix multiplication unit for matrix multiplication, and the output is the feature ;

[0048] The characteristics and Features After the shape transformation, the matrix multiplication unit is input, and then the shape transformation is output to obtain the feature ; The characteristics That is, the single-dimensional attention feature obtained by the fourth convolution operation result, the seventh convolution operation result, and the ninth convolution operation result.

[0049] Furthermore, the composite feature attention unit performs feature processing, specifically comprising the following steps:

[0050] Construct a composite feature attention unit, which includes a maximum pooling layer, an average pooling layer, a concatenation unit, three fully connected layers, and a convolutional layer;

[0051] If the first single-dimensional attention feature and the second single-dimensional attention feature of the input are recorded as features ,feature ; then do the following:

[0052] The characteristics Input the maximum pooling layer in the composite feature attention unit and output the feature ;

[0053] The characteristics Input the average pooling layer in the composite feature attention unit and output the feature ;

[0054] The characteristics and Features Input the splicing unit in the composite feature attention unit to perform feature splicing, and output the feature ;

[0055] The features obtained by the maximum pooling layer output Input the first fully connected layer in the composite feature attention unit and output the feature ;

[0056] The characteristics and Features Features obtained after splicing The second fully connected layer and the convolutional layer in the composite feature attention unit are input in sequence, and the feature ;

[0057] The features obtained by the average pooling layer output Input the third fully connected layer in the composite feature attention unit and output the feature ;

[0058] The characteristics and Features After the shape transformation, the matrix is ​​input into the matrix multiplication unit for matrix multiplication, and the output is the feature ;

[0059] The characteristics and Features After the shape transformation, the matrix multiplication unit is input, and then the shape transformation is output to obtain the feature , the characteristics That is, the composite feature attention unit performs feature processing on the input first single-dimensional attention feature and the second single-dimensional attention feature, and obtains the feature processing result.

[0060] Furthermore, the attention cross-recombination module is used to receive the feature processing results of the deep cross-attention module and the feature processing results of the multi-scale fusion cross-attention module, and perform cross-recombination to obtain cross-recombination features, specifically including the following steps:

[0061] Constructing an attention cross-reorganization module, wherein the constructed attention cross-reorganization module includes two splitting units, six convolution units, and three splicing units;

[0062] The first splitting unit in the attention cross-reorganization module is used to cross-split the feature processing results of the deep cross-attention module along the channel direction, and output the feature and Features ;

[0063] The first convolution unit in the attention cross-reorganization module is used to perform convolution operation on the feature processing results of the deep cross-attention module and output the feature ;

[0064] The second convolution unit in the attention cross-reorganization module is used to perform convolution operation on the feature processing results of the multi-scale fusion cross-attention module and output the feature ;

[0065] The second splitting unit in the attention cross-reorganization module is used to cross-split the feature processing results of the multi-scale fusion cross-attention module along the channel direction, and output the feature and Features ;

[0066] The third convolutional unit in the attention cross-reorganization module is used to Perform convolution operation and output features ;

[0067] The fourth convolutional unit in the attention cross-reorganization module is used to Perform convolution operation and output features ;

[0068] The first concatenation unit in the attention cross-reorganization module is used to output the features of the first convolution unit. And the features output by the second convolution unit Perform stitching and output stitching features ;

[0069] The fifth convolutional unit in the attention cross-reorganization module is used to Perform convolution operation and output features ;

[0070] The sixth convolutional unit in the attention cross-reorganization module is used to Perform convolution operation and output features ;

[0071] The second concatenation unit in the attention cross-reorganization module is used to adjust the features output by the third convolution unit. And the features output by the fifth convolutional unit Perform stitching and output stitching features ;

[0072] The third concatenation unit in the attention cross-reorganization module is used to adjust the features output by the fourth convolution unit. And the features output by the sixth convolutional unit Perform stitching and output stitching features ;

[0073] The stitching features output by the second stitching unit And the splicing features output by the third splicing unit After the shape transformation, the matrix multiplication unit is input for matrix multiplication, and the output is the feature ;

[0074] The stitching features output by the first stitching unit and Features After the shape transformation, the matrix multiplication unit is input for matrix multiplication, and the feature is output after the shape transformation. , ; The characteristics That is, the cross-recombination features after the cross-recombination of the attention cross-recombination module.

[0075] Furthermore, the indicator indication module is used to receive the cross-recombination feature and output the prediction indication feature corresponding to the text data, specifically comprising the following steps:

[0076] Construct an indicator module, which includes twelve convolution units, four splicing units, and two fully connected layer units;

[0077] Among them, the first convolution unit in the indicator module is used to perform convolution operation on the cross-reorganized features after the cross-reorganization of the attention cross-reorganization module, and output the feature ;

[0078] The seventh convolution unit in the indicator module is used to perform convolution operation on the cross-reorganized features after the cross-reorganization of the attention cross-reorganization module, and output the feature ;

[0079] The first splicing unit in the indicator module is used to and Features Splice and output the features ;

[0080] The third splicing unit in the indicator module is used to and Features Splice and output the features ;

[0081] The second convolution unit in the indicator module is used to output the features of the first concatenation unit. Perform convolution operation and output the features ;

[0082] The eighth convolution unit in the indicator module is used to output the features of the third concatenation unit. Perform convolution operation and output the features ;

[0083] The second splicing unit in the indicator module is used to and Features Splice and output the features ;

[0084] The fourth splicing unit in the indicator module is used to and Features Splice and output the features ;

[0085] The features obtained After passing through four convolution units for convolution operations in sequence, it passes through a fully connected layer unit and outputs the predicted indicator features. , ;

[0086] The features obtained After passing through four convolution units for convolution operations in sequence, it passes through a fully connected layer unit and outputs the predicted indicator features. , .

[0087] Furthermore, the score of the international standardized document is calculated according to the prediction indication features corresponding to all text data in the same international standardized document, which specifically includes the following steps:

[0088] Predictive indicator features obtained based on each text data in the same international standardized document , , get the corresponding index screening vector ;

[0089] According to the index filter vector corresponding to each piece of text data obtained , and use the following formula to get the total evaluation score of the international standardization document:

[0090] ;

[0091] In the formula, is the total number of sentence texts in the international standardization document to be evaluated; Represents predictive indicator features The elements; Indicates The index filter vector of sentences; It represents the score matrix obtained by assigning values ​​to indicators according to the importance of different indicators in the evaluation of national standard adoption; express , which is the indicator classification feature vector output by the indicator module.

[0092] It can be seen from the above technical solution that, compared with the prior art, the present invention discloses a national standard adoption assessment method based on multi-layer cross-reorganization attention, which has the following beneficial effects:

[0093] The present invention can use deep learning algorithms to automatically and accurately evaluate the degree of matching between international standards and domestic standards, as well as the potential impact of international standards on domestic industries, thereby improving the efficiency and accuracy of standard adoption and evaluation.

[0094] The present invention can use deep learning algorithms to conduct comprehensive analysis and evaluation of data of different dimensions, ensuring that the evaluation results can comprehensively consider various factors and provide decision makers with more comprehensive and in-depth decision support. BRIEF DESCRIPTION OF THE DRAWINGS

[0095] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying creative work.

[0096] Figure 1 It is a flow chart of the national standard adoption assessment and evaluation method based on multi-layer cross-reorganization attention of the present invention.

[0097] Figure 2 It is a structural diagram of the multi-layer cross-reorganization attention index indication network in the present invention.

[0098] Figure 3 It is a structural diagram of the data encoding and feature mapping module in the present invention.

[0099] Figure 4 It is a structural diagram of the deep cross-attention module in the present invention.

[0100] Figure 5 This is the overall structural diagram of the cross-attention algorithm model in the present invention.

[0101] Figure 6 It is a structural diagram of the multi-scale fusion cross-attention module in the present invention.

[0102] Figure 7 It is a structural diagram of the attention cross-reorganization module in the present invention.

[0103] Figure 8 It is a structural diagram of the indicator indicating module in the present invention.

[0104] Fig. 9 It is the national standard adoption assessment and evaluation index system proposed in the present invention. DETAILED DESCRIPTION

[0105] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0106] The national standard adoption evaluation method based on multi-layer cross-reorganization attention disclosed in the present invention is as follows: Figure 1 As shown, the specific steps include:

[0107] Extract the text content in the international standardization document to be evaluated, split the text content according to the sentence dimension, and obtain several copies of text data;

[0108] Each extracted text data is input into a multi-layer cross-recombination attention index classification model to obtain the prediction indicator features corresponding to the text data;

[0109] The international standardized document is scored according to the prediction indicative features corresponding to all text data in the same international standardized document to obtain an evaluation result of the international standardized document to be evaluated.

[0110] In the present invention, the content of the international standardized document to be evaluated is first split; then, each piece of split data is predicted through a multi-layer cross-recombination attention index classification model to obtain two prediction indication features for representing different level indicators; finally, the international standardized document to be evaluated is scored according to the prediction indication features of all data to obtain the evaluation score of the international standardized document to be evaluated. When the evaluation score of the international standardized document to be evaluated reaches a certain threshold, it can be considered that the international standardized document to be evaluated meets the international standard adoption standard.

[0111] The multi-layer cross-reorganization attention index classification model in the present invention is used to process the file data after the international standardized file is split, and obtain two prediction indicator features corresponding to the text data. Figure 2As shown, the multi-layer cross-recombination attention indicator classification model includes a data encoding and feature mapping module, a deep cross-attention module, a multi-scale fusion cross-attention module, an attention cross-recombination module, and an indicator indication module.

[0112] Among them, the data encoding and feature mapping module is used to receive text data, and encode and map the text data to obtain corresponding mapping features; the deep cross-attention module and the multi-scale fusion cross-attention module respectively perform feature processing on the obtained mapping features; the attention cross-recombination module is used to receive the feature processing results of the deep cross-attention module and the feature processing results of the multi-scale fusion cross-attention module, and perform cross-recombination to obtain cross-recombined features; the indicator indication module is used to receive the cross-recombined features and output the prediction indication features corresponding to the text data.

[0113] Combine the following Figure 3-Figure 8 Each module in the multi-layer cross-reorganization attention indicator classification model is further explained.

[0114] The data encoding and feature mapping module encodes and maps the text data after the international standard file is split, such as Figure 3 As shown, the data encoding and feature mapping module includes: a BERT pre-training model, a first convolution unit, a first dimension mapping unit, a second convolution unit, a second dimension mapping unit, a third convolution unit, and a matrix multiplication unit.

[0115] The BERT pre-trained model is used to extract features from input text data and obtain word vector text features. The first convolution unit is used to perform a convolution operation on the obtained word vector text features to obtain the convolution operation result of the first convolution unit; the convolution operation result passes through the first dimension mapping unit and the second convolution unit in sequence to obtain the first dimension mapping feature; the convolution operation result passes through the second dimension mapping unit and the third convolution unit in sequence to obtain the second dimension mapping feature; the first dimension mapping feature and the second dimension mapping feature are simultaneously input into the matrix multiplication unit to obtain the mapping feature corresponding to the text data.

[0116] In the specific implementation steps, the text data needs to be preprocessed before being input into the BERT pre-training model. In order to achieve the standardization of the input data, this application unifies the split text data characters into 384 characters. If the characters are less than 384, the number of characters 0 is added after the text data to supplement it to 384 characters; Figure 3 As shown, the standardized text data is input into the BERT pre-trained model, and the output is the feature . The feature In the first convolution unit in the input data encoding and feature mapping module, the output is the feature . The feature In the first dimension mapping unit in the input data encoding and feature mapping module, the output is the feature . Input into the second convolution unit, and output the feature . Input into the second dimension mapping unit, and output the feature . Input into the third convolution unit, and output the feature . and Features Input matrix multiplication unit, the output is , .

[0117] Deep Cross Attention Module Figure 4 As shown, it includes a splitting unit, a first cross attention unit, a second cross attention unit, a third cross attention unit and a splicing unit connected in sequence; wherein the splitting unit is used to split the mapping features output by the data encoding and feature mapping module along the channel dimension to obtain dimensional features of 12 different channel dimensions, , ; The first cross attention unit is used to receive the dimension feature , and Any three continuous dimensional features in , Perform cross attention operation and output 10 first cross operation features , ; The second cross attention unit is used to receive the first cross operation feature , and the features Any three continuous dimensional features in , Perform cross attention operation and output 8 second cross operation features , ; The third cross attention unit is used to receive the second cross operation feature , and the features Any three continuous dimensional features in , Perform cross attention operation and output 6 third cross operation features , ; Finally, the concatenation unit is used to output the 6 third crossover operation features Perform feature splicing and output splicing features , , and use it as the feature processing result of the deep cross-attention module.

[0118] In the embodiment of the present invention, the first cross attention unit, the second cross attention unit, and the third cross attention unit respectively use 10, 8, and 6 cross attention algorithm models of the same process to perform operations to obtain corresponding cross attention features. The 6 third cross operation features output by the third cross attention unit By performing feature concatenation, we can obtain , .

[0119] The first cross attention unit, the second cross attention unit, and the third cross attention unit use the cross attention algorithm model to perform cross operations on dimensional features. Specifically, the overall structure of the cross attention algorithm model is as follows Figure 5 As shown, the model specifically performs the following steps:

[0120] Any three continuous dimensional features are recorded as , , , the first dimension feature of any three consecutive dimensional features (In the figure use The first convolution unit in the cross attention algorithm model is input and the feature is output. ; The feature Input the first dimension reshape unit and output the feature . The first dimension feature of any three consecutive dimension features ( ) and the second dimension feature (In the figure use The first matrix multiplication unit is used to perform matrix multiplication, and the output is the feature , the feature After activation by the Softmax activation module, the output is the feature . The third dimension feature of any three consecutive dimensional features (In the figure use to represent) and features Input the second matrix multiplication unit to perform matrix multiplication, and output the feature ; The feature Input the second convolution unit in the cross-attention algorithm model and output the feature ; The feature Input the second dimension reshape unit and output the feature . Reshape the output features of the unit along the first dimension and reshape the output features of the unit in the second dimension Perform matrix addition operation and output the features . The first dimension feature of any three consecutive dimension features ( ) and features Input the third matrix multiplication unit to perform matrix multiplication, and output the feature ; The feature After activation by the Softmax activation module, the output is the feature . The second dimension feature of any three consecutive dimension features ( ) and features Input the fourth matrix multiplication unit to perform matrix multiplication, and output the feature ; The feature After activation by the Softmax activation module, the output is the feature . The third dimension feature of any three consecutive dimensional features ( ) and features Input the fifth matrix multiplication unit to perform matrix multiplication, and output the feature . The feature and Features Input matrix multiplication unit to perform matrix multiplication, and output feature ,feature That is, any three-dimensional features ( , , ) after the cross-attention operation.

[0121] In the deep cross-attention module provided by the present invention, the network front end uses a convolution unit to capture the local features of the input data, and then the data passes through a dimension reshaping module to prepare for the matrix multiplication operation, and further abstracts and combines the features. In multiple matrix multiplication units, the features are initially fused. The Softmax activation function introduces nonlinearity after the matrix multiplication unit to enhance the feature recognition ability of the model. The cross-attention algorithm module is mainly used to process the long-distance dependencies between standard text sequence data, and gradually extracts and strengthens key features by stacking multiple modules. Finally, the splicing unit integrates the features of different paths to provide support for the final classification task.

[0122] like Figure 6 In the embodiment of the present invention, the multi-scale fusion cross attention module includes 1 to 13 convolution units, two single-dimensional feature attention units, and a composite feature attention unit; wherein the mapping feature output by the data encoding and feature mapping module Input the first convolution unit in the multi-scale fusion cross attention module and output the feature . Input the fifth convolution unit in the multi-scale fusion cross attention module and output the feature . Input the sixth convolution unit and output the feature . Input the seventh convolution unit and output the feature . Input the second convolution unit and output the feature . Input the eighth convolution unit and output the feature . Input the ninth convolution unit and output the feature . Enter the third convolution unit in the multi-scale fusion cross attention module, and output the feature . Input the tenth convolution unit and output the feature . Input the fourth convolution unit in the multi-scale fusion cross attention module and output the feature . ,feature and Features Input the first single-dimensional feature attention unit in the multi-scale fusion cross attention module, and output the feature . ,feature and Features Input the second single-dimensional feature attention unit in sequence and output the feature . and Features Input the composite feature attention unit in the multi-scale fusion cross attention module and output the feature . Input the eleventh convolution unit in the multi-scale fusion cross attention module, and output the feature . Input into the twelfth convolution unit, and output the feature . Input into the thirteenth convolution unit, and output the feature , .

[0123] The multi-scale fusion cross attention module in the present invention performs multi-scale feature extraction through a series of convolution units to capture features at different levels of the input data. The model uses single-dimensional feature attention and compound feature attention mechanisms to enhance the representation ability of key features, where single-dimensional feature attention focuses on feature enhancement on a single dimension, while compound feature attention integrates multi-dimensional information to obtain richer feature representation. Features are further extracted and refined through maximum pooling, average pooling, and convolution unit processing. These features are integrated and nonlinearly transformed through fully connected layers to enhance the model's discrimination ability and feature richness.

[0124] The following is a detailed introduction to the single-dimensional feature attention unit in the multi-scale fusion cross attention module. Figure 6 As shown in the upper right corner of , , , (can be understood in the first single dimension feature attention unit Corresponding features ,feature ,feature ; In the second single-dimensional feature attention unit Corresponding features ,feature ,feature , Figure 6 The specific situation of the first single-dimensional feature attention unit shown in the upper right corner), specifically in the single-dimensional feature attention unit Input the maximum pooling layer and output the features ,feature Input the fully connected layer 1 and output the features ;feature Input the average pooling layer and output the feature ;feature Input the fully connected layer 2 and output the features ;feature and Features After the shape transformation, the matrix multiplication unit is input and the output is the feature ;feature Input convolution unit 1, output feature ;feature Input the fully connected layer 3 in the single-dimensional feature attention module and output the feature ;feature and Features After the shape transformation, the matrix multiplication unit is input, and then the shape transformation is output to obtain the feature , which is the single-dimensional attention feature obtained by three inputs.

[0125] like Figure 6 As shown in the lower left corner of , the composite feature attention unit receives two inputs, which are recorded as ; , Corresponding to the first single dimension attention features , the second single dimension attention feature Specifically, the composite feature attention unit includes a maximum pooling layer, an average pooling layer, a concatenation unit, three fully connected layers, and a convolutional layer. Input the maximum pooling layer in the composite feature attention unit to obtain the feature . The feature Input the average pooling layer to get the features .feature and Features Input the splicing unit together to get the feature . The feature Input the fully connected layer 1 and output the features . The feature Input the fully connected layer 3 and output the features . The feature Input the fully connected layer 2 to get the features . The feature Input the convolutional layer unit in the composite feature attention unit to obtain the feature . The feature and Features After the shape transformation, the matrix multiplication unit is input to obtain the feature . The feature and Features After the shape transformation, the matrix multiplication unit is input, and then the feature is obtained after the shape transformation. , which is the first single-dimensional attention feature of the composite feature attention unit to the input , the second single dimension attention feature Perform feature processing and obtain the feature processing result features .

[0126] In the embodiment of the present invention, the attention cross-reorganization module is used to receive the feature processing results of the deep cross-attention module. Feature processing results of multi-scale fusion cross-attention module , and cross-recombination is performed to obtain the cross-recombination characteristics. The overall structure is as follows Figure 7 As shown in Figure 1, the attention cross-reorganization module includes two splitting units, six convolution units, and three splicing units. The attention cross-reorganization module performs the following operations: Input the first split module in the attention cross-reorganization module, cross-split along the channel direction, and obtain the feature and Features ; The feature Input into the first convolution unit and get the feature .

[0127] The features output by the multi-scale fusion cross-attention module Input the second split module in the attention cross-reorganization module, cross-split along the channel direction, and output the feature and Features ; The feature Input into the second convolution unit, and output the feature . Input the third convolution unit to get the feature . Input the fourth convolution unit to get the feature . and Features Input the first concatenated unit in the attention cross-reorganization module to obtain the feature . Input the fifth convolution unit in the attention cross-reorganization module to obtain the feature . Input the sixth convolution unit to get the feature . and Features Input the second splicing unit and get the feature . and Features Input the third splicing unit and get the feature . and Features After the shape transformation, the matrix multiplication unit is input to obtain the feature . and Features After the shape transformation, the matrix multiplication unit is input, and the feature is output after the shape transformation. , , which is the output result of the attention cross-reorganization module.

[0128] The attention cross-reorganization module realizes feature fusion through a multi-branch structure to obtain feature representations of different granularities and levels. The input of this module is first sent to two split modules to divide the data into different feature subsets, and then each subset is extracted through independent convolution units. The first splicing unit integrates the features from the first and second convolution units, while the second and third splicing units fuse the outputs of the third and fourth convolution units and the fifth and sixth convolution units respectively. This feature fusion strategy enables the model to capture the complexity and structure of the associated features between data from different angles. The residual connection is used to combine the original input with the features processed by the convolution and splicing units to further enhance the degree of feature fusion.

[0129] After obtaining the output result of the attention cross-recombination module, the indicator module is used to receive the cross-recombination features. Further processing is performed to obtain the prediction indication features corresponding to the text data. Specifically, the indicator indication module disclosed in the present invention is as follows Figure 8 As shown in Figure 1, it includes twelve convolution units, four concatenation units, and two fully connected layer units. This module performs the following operations to transform the features The input indicator indicates the first convolution unit in the module, and the output is the feature . At the same time, the feature Input indicator indicates the seventh convolution unit in the module, and output the feature . The feature and Features Input together into the first splicing unit to obtain the feature . The feature and Features Input into the third splicing unit and get the feature . The feature Input the second convolution unit in the indicator module and get the feature . The feature Input the eighth convolution unit in the indicator module and get the feature . The feature and Features Input the second splicing unit in the indicator module and get the feature . The feature and Features Input into the fourth splicing unit and get the feature . The feature Input the third convolution unit to get the feature . The feature Input the fourth convolution unit to get the feature . The feature Input the fifth convolution unit to get the feature . The feature Input the sixth convolution unit in the indicator module to obtain the feature . The feature Input the first fully connected unit in the indicator module and get the feature , . The feature Input the ninth convolution unit in the indicator module and get the feature . The feature Input the tenth convolution unit and get the feature . The feature Input the eleventh convolution unit to get the feature . The feature Input the twelfth convolution unit and get the feature . The feature Input the second fully connected unit and output the feature , .

[0130] In the indicator indication module of the present invention, the number of neurons in the first fully connected layer unit is 3, and the number of neurons in the second fully connected layer unit is 11, which are used to correspond to the primary indicator characteristics and secondary indicator characteristics of the international standardized documents respectively.

[0131] It should be pointed out that in the process of using the multi-layer cross-recombination attention indicator classification model to predict the indicators of the international standardized documents to be evaluated and obtaining the prediction indicator features corresponding to the text data, it is also necessary to use the data set to train the multi-layer cross-recombination attention indicator classification model.

[0132] During the model training process, a large number of international standardization document texts were collected to construct a national standard adoption evaluation data set, which was divided into a training set, a validation set, and a test set. The training set and the validation set were used to obtain a deployable multi-layer cross-recombination attention index classification model. The data in the data set was input into the data encoding and feature mapping module, and the output was the feature . The feature Input into the deep cross attention module and output the feature . The feature Input into the multi-scale fusion cross attention module, and output the feature . The feature and Features Input into the attention cross-reorganization module and output the feature . The feature Input indicator module, output feature and Features .feature Responsible for predicting and classifying indicator 1, features Responsible for predicting and classifying indicator 2. Using features and Features Calculate the loss Loss and use the optimizer to perform gradient optimization on the cross-level feature cross-fusion index classification model.

[0133] Specifically, during the training process, the loss function of the multi-layer cross-recombination attention index classification model is composed of the cosine similarity loss function and the cross entropy loss function, and the final loss Loss is obtained by adding the two.

[0134] Among them, the cosine similarity loss function is defined as follows: ;

[0135] in, is the label prediction value, is the true value of the label, is the Euclidean norm.

[0136] The cross entropy loss function is defined as follows: ;

[0137] in, is the label prediction value, is the true value of the label.

[0138] The total loss function is defined as follows:

[0139] .

[0140] During the training and validation process, the SGD optimizer is used to perform gradient optimization on the multi-layer cross-recombination attention indicator classification model.

[0141] In the present invention, the international standardized documents to be evaluated and the known international standardized documents used as training sets usually meet the following standard system, and a systematic analysis is conducted on the international standardized documents, which can be disassembled and extracted into 3 first-level indicators and 11 second-level indicators. The 3 first-level indicators include "comprehensive environment", "technical content" and "expected benefits". The 11 second-level indicators include "conformity with relevant planning and industrial policies and systems", "friendliness of culture, customs and habits", "matching degree of the products, processes or services involved with my country's industrial structure", "the evaluation index system comprehensively covers the characteristics related to the evaluation purpose", "evaluation indicators at the same level are independent of each other", "whether the technical content is coordinated with my country's standards", "expected quality improvement", "expected efficiency improvement", "expected industrial development", "social impact", and "promotion effect on the ecological environment".

[0142] Indicator matching is carried out according to the correlation between the primary and secondary indicators, and the primary and secondary indicators are mapped and linked to each other.

[0143] The first-level comprehensive environmental indicators include three second-level indicators, namely, "Compliance with relevant planning and industrial policies and systems", "Friendliness of culture, customs and habits", and "Matching degree between the products, processes or services involved and my country's industrial structure". The first-level technical content indicators include three second-level indicators, namely, "The evaluation index system comprehensively covers the characteristics related to the evaluation purpose", "Evaluation indicators at the same level are independent of each other", and "Whether the technical content is coordinated with my country's standards". The first-level expected benefit indicators include five second-level indicators, namely, "Expected quality improvement", "Expected efficiency improvement", "Expected industrial development", "Social impact", and "Promotion effect on the ecological environment". The overall evaluation index system for national standard adoption assessment is as follows: Fig. 9 shown.

[0144] In the indicator analysis process, the indicator-label mapping can map the first-level indicators in the national standard adoption assessment indicator system to data label one, and map the second-level indicators in the national standard adoption assessment indicator system to data label two. There are four label ones, marked with 0, 1, and 2 respectively. 0 is the comprehensive environment, 1 is the technical content, and 2 is the expected benefit. There are 11 label twos, marked with a, b, c, d, e, f, g, h, i, j, and k respectively. a is the compliance with relevant planning and industrial policies and systems, b is the friendliness of culture, customs and habits, c is the comprehensive coverage of the evaluation indicator system and the characteristics related to the evaluation purpose, d is the comprehensive coverage of the evaluation indicator system and the characteristics related to the evaluation purpose, e is the independence of the evaluation indicators at the same level, f is whether the technical content is coordinated with the national standards, g is the expected quality improvement, h is the expected efficiency improvement, i is the expected industrial development, j is the social impact, and k is the promotion effect on the ecological environment. The labeled data set is , where the first data in each [] of the dataset is label one, the second data is label two, and the third data is the international standardized document text.

[0145] During the training process, the national standard assessment data set can be divided into training set, validation set and test set in a ratio of 6:2:2.

[0146] After using the trained multi-layer cross-recombination attention indicator classification model to obtain the indicator prediction of the international standardized document to be evaluated, the present invention uses the national standard adoption evaluation score calculation method to calculate the score of the national standard adoption text according to the prediction indicator category. When the evaluation score of the international standardized document to be evaluated reaches a certain threshold, it can be considered that the international standardized document to be evaluated meets the international standard adoption standard.

[0147] The score calculation proposed in the present invention is described in detail below.

[0148] Scores are assigned to the secondary indicators according to the importance of different indicators to the evaluation of national standard adoption. Among them, compliance with relevant planning and industrial policies and systems is 20 points, friendliness of culture, customs and habits is 20 points, the matching degree of the products, processes or services involved with my country's industrial structure is 20 points, the comprehensive coverage of the evaluation index system and the characteristics related to the evaluation purpose is 20 points, the independence of the evaluation indicators at the same level is 20 points, whether the technical content is coordinated with my country's standards is 20 points, the expected quality improvement is 12 points, the expected efficiency improvement is 12 points, the expected industrial development is 12 points, the social impact is 12 points, and the promotion effect on the ecological environment is 12 points. Based on the above scoring, a score matrix is ​​constructed. , ={20,20,20,20,20,20,12,12,12,12}.

[0149] Based on characteristics Construct an indicator screening vector. When the feature The first element in When it is the maximum value among three elements, construct the index screening vector , When the feature The second element When it is the maximum value among three elements, construct the index screening vector , When the feature The third element When it is the maximum value among three elements, construct the index screening vector , Each sentence has the following features: In the characteristic case, there is only one index screening vector.

[0150] Calculate the score of the national standard adoption document. Split the national standard adoption document according to the sentence dimension. Assume that a national standard adoption document can be split into sentences, each of which is obtained through a multi-layer cross-recombination attention index classification model and ,pass Obtain the indicator screening vector. The total score calculation of the national standard adoption document is defined as follows:

[0151] ;

[0152] in, is the number of sentence texts in the national standard adoption document, express The elements, Indicates The index filtering vector of sentences, express That is, the secondary index classification feature vector output by the multi-layer cross-recombination attention index classification model, A represents the score matrix, The purpose is to adjust the score from 60 points to 100 points.

[0153] In addition, the embodiment of the present invention also discloses an international standard transformation evaluation system based on multi-layer cross-combination attention, including:

[0154] The national standard adoption document classification module inputs the national standard adoption document text into the trained multi-layer cross-recombination attention index classification model and outputs the index classification result. The multi-layer cross-recombination attention index classification model has been described above and will not be repeated here.

[0155] The score calculation and evaluation module inputs the classification results output by the model into the national standard assessment score calculation method to calculate the evaluation score. The indicator scoring method has been described above and will not be repeated here.

[0156] The present invention also proposes an electronic device, including a memory and a processor, and computer instructions stored in the memory and running on the processor. When the computer instructions are executed by the processor, the steps described in the national standard adoption and evaluation method based on multi-layer cross-reorganization attention are completed.

[0157] The present invention also proposes a computer-readable storage medium for storing computer instructions. When the computer instructions are executed by a processor, the steps described in the national standard adoption and evaluation method based on multi-layer cross-reorganization attention are completed.

[0158] In this specification, each embodiment is described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the embodiments can be referred to each other. For the device disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and the relevant parts can be referred to the method part.

[0159] The above description of the disclosed embodiments enables one skilled in the art to implement or use the present invention. Various modifications to these embodiments will be apparent to one skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but rather to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A national standard adoption assessment method based on multi-layer cross-reorganization attention, characterized in that: The following steps are involved: Extract the text content in the international standardization document to be evaluated, split the text content according to the sentence dimension, and obtain several copies of text data; Input each extracted text data into a multi-layer cross-reorganization attention index classification model to obtain the prediction indication features corresponding to the text data; wherein the multi-layer cross-reorganization attention index classification model includes a data encoding and feature mapping module, a deep cross-attention module, a multi-scale fusion cross-attention module, an attention cross-reorganization module, and an indicator indication module; The data encoding and feature mapping module is used to receive text data, and encode and map the text data to obtain corresponding mapping features; the deep cross-attention module and the multi-scale fusion cross-attention module respectively perform feature processing on the obtained mapping features; the attention cross-recombination module is used to receive the feature processing results of the deep cross-attention module and the feature processing results of the multi-scale fusion cross-attention module, and perform cross-recombination to obtain cross-recombination features; The indicator indication module is used to receive the cross-recombination feature and output the prediction indication feature corresponding to the text data; The international standardized document is scored according to the prediction indicative features corresponding to all text data in the same international standardized document to obtain an evaluation result of the international standardized document to be evaluated.

2. According to claim 1, a method for evaluating national standard adoption based on multi-layer cross-reorganization attention is characterized in that: The data encoding and feature mapping module includes: a BERT pre-training model, a first convolution unit, a first dimension mapping unit, a second convolution unit, a second dimension mapping unit, a third convolution unit, and a matrix multiplication unit: Among them, the BERT pre-trained model is used to extract features from the input text data and obtain word vector text features; the first convolution unit is used to perform a convolution operation on the obtained word vector text features to obtain the convolution operation result of the first convolution unit; the convolution operation result passes through the first dimension mapping unit and the second convolution unit in sequence to obtain the first dimension mapping feature; the convolution operation result passes through the second dimension mapping unit and the third convolution unit in sequence to obtain the second dimension mapping feature; the first dimension mapping feature and the second dimension mapping feature are simultaneously input into the matrix multiplication unit to obtain the mapping feature corresponding to the text data.

3. According to claim 1, a method for evaluating national standard adoption based on multi-layer cross-reorganization attention is characterized in that: The deep cross attention module includes a splitting unit, a first cross attention unit, a second cross attention unit, a third cross attention unit and a splicing unit connected in sequence; The splitting unit is used to split the mapping features output by the data encoding and feature mapping module along the channel dimension to obtain dimensional features of 12 different channel dimensions. , ; The first cross attention unit is used to receive the dimension feature , and the dimension features Any three continuous dimensional features in , Perform cross attention operation and output 10 first cross operation features , ; The second cross attention unit is used to receive the first cross operation feature , and the first crossover feature Any three continuous dimensional features in , Perform cross attention operation and output 8 second cross operation features , ; The third cross attention unit is used to receive the second cross operation feature , and the second crossover feature Any three continuous dimensional features in , Perform cross attention operation and output 6 third cross operation features , ; The concatenation unit is used to perform the 6 third crossover operation features of the output Perform feature splicing and output splicing features , and used as the feature processing result of the deep cross attention module.

4. According to claim 3, a method for evaluating national standard adoption based on multi-layer cross-reorganization attention is characterized in that: The cross attention operation in the first cross attention unit, the second cross attention unit and the third cross attention unit is implemented by the following steps: Any three continuous dimensional features are recorded as , , , the first dimension feature of any three consecutive dimensional features Input the first convolution unit in the cross-attention algorithm model and output the feature ; The characteristics Input the first dimension reshape unit and output the feature ; The first dimension feature of any three consecutive dimension features and the second dimension feature Input the first matrix multiplication unit to perform matrix multiplication, and output the feature , the feature After activation by the Softmax activation module, the output is the feature ; The third dimension feature of any three consecutive dimensional features and Features Input the second matrix multiplication unit to perform matrix multiplication, and output the feature ; The characteristics Input the second convolution unit in the cross-attention algorithm model and output the feature ; The characteristics Input the second dimension reshape unit and output the feature ; Reshape the output features of the unit along the first dimension and reshape the output features of the unit in the second dimension Perform matrix addition operation and output the features ; The first dimension feature of any three consecutive dimension features and Features Input the third matrix multiplication unit to perform matrix multiplication, and output the feature ; The characteristics After activation by the Softmax activation module, the output is the feature ; The second dimension feature of any three consecutive dimension features and Features Input the fourth matrix multiplication unit to perform matrix multiplication, and output the feature ; The feature After activation by the Softmax activation module, the output is the feature ; The third dimension feature of any three consecutive dimensional features and Features Input the fifth matrix multiplication unit to perform matrix multiplication, and output the feature ; The characteristics and Features Input matrix multiplication unit to perform matrix multiplication, and output feature ,feature That is, any three-dimensional features Cross-operation features after cross-attention operation.

5. According to claim 1, a method for evaluating national standard adoption based on multi-layer cross-reorganization attention is characterized in that: The multi-scale fusion cross attention module includes thirteen convolution units from 1 to 13, two single-dimensional feature attention units, and a composite feature attention unit; The first convolution unit in the multi-scale fusion cross attention module is used to perform a first convolution operation on the mapping features output by the data encoding and feature mapping module, and output a first convolution operation result; The first convolution operation result is sequentially subjected to convolution operation by the fifth convolution unit, the sixth convolution unit and the seventh convolution unit in the multi-scale fusion cross attention module to obtain a fourth convolution operation result; The first convolution operation result is sequentially convolved through the second convolution unit, the third convolution unit and the fourth convolution unit in the multi-scale fusion cross attention module to obtain a tenth convolution operation result; The first convolution operation result is sequentially subjected to convolution operation by the second convolution unit, the eighth convolution unit and the ninth convolution unit in the multi-scale fusion cross attention module to obtain a seventh convolution operation result; The first convolution operation result is sequentially subjected to convolution operation by the second convolution unit, the third convolution unit and the tenth convolution unit in the multi-scale fusion cross attention module to obtain a ninth convolution operation result; Inputting the obtained fourth convolution operation result, seventh convolution operation result and ninth convolution operation result into the first single-dimensional feature attention unit to obtain a first single-dimensional attention feature; Inputting the obtained tenth convolution operation result, the seventh convolution operation result and the ninth convolution operation result into the second single-dimensional feature attention unit to obtain a second single-dimensional attention feature; The first single-dimensional attention feature and the second single-dimensional attention feature are obtained and simultaneously input into the composite feature attention unit for feature processing, thereby obtaining a feature processing result of the composite feature attention unit; The feature processing results of the composite feature attention unit are sequentially convolved through three convolution units to obtain the final feature processing results of the multi-scale fusion cross attention module.

6. According to claim 5, a method for evaluating national standard adoption based on multi-layer cross-combination attention is characterized in that: The first single-dimensional feature attention unit is used to perform feature processing on the input fourth convolution operation result, the seventh convolution operation result and the ninth convolution operation result to obtain a single-dimensional attention feature; specifically comprising the following steps: Constructing a single-dimensional feature attention unit, wherein the constructed single-dimensional feature attention unit includes a maximum pooling layer, an average pooling layer, a convolutional layer, and three fully connected layers; If the fourth convolution operation result, the seventh convolution operation result and the ninth convolution operation result of the input are recorded as features ,feature and Features ; then do the following: The characteristics The maximum pooling layer and the first fully connected layer in the single-dimensional feature attention unit are input in sequence, and the feature ; The characteristics The average pooling layer and the second fully connected layer in the single-dimensional feature attention unit are input in sequence, and the feature ; The characteristics The convolutional layer and the third fully connected layer in the single-dimensional feature attention unit are input in sequence, and the feature ; The characteristics and Features After the shape transformation, the matrix is ​​input into the matrix multiplication unit for matrix multiplication, and the output is the feature ; The characteristics and Features After the shape transformation, the matrix multiplication unit is input, and then the shape transformation is output to obtain the feature ; The characteristics That is, the single-dimensional attention feature obtained by the fourth convolution operation result, the seventh convolution operation result, and the ninth convolution operation result.

7. According to claim 5, a method for evaluating national standard adoption based on multi-layer cross-reorganization attention is characterized in that: The composite feature attention unit performs feature processing, specifically including the following steps: Construct a composite feature attention unit, which includes a maximum pooling layer, an average pooling layer, a concatenation unit, three fully connected layers, and a convolutional layer; If the first single-dimensional attention feature and the second single-dimensional attention feature of the input are recorded as features ,feature ; then do the following: The characteristics Input the maximum pooling layer in the composite feature attention unit and output the feature ; The characteristics Input the average pooling layer in the composite feature attention unit and output the feature ; The characteristics and Features Input the splicing unit in the composite feature attention unit to perform feature splicing, and output the feature ; The features obtained by the maximum pooling layer output Input the first fully connected layer in the composite feature attention unit and output the feature ; The characteristics and Features Features obtained after splicing The second fully connected layer and the convolutional layer in the composite feature attention unit are input in sequence, and the feature ; The features obtained by the average pooling layer output Input the third fully connected layer in the composite feature attention unit and output the feature ; The characteristics and Features After the shape transformation, the matrix is ​​input into the matrix multiplication unit for matrix multiplication, and the output is the feature ; The characteristics and Features After the shape transformation, the matrix multiplication unit is input, and then the shape transformation is output to obtain the feature , the characteristics That is, the composite feature attention unit performs feature processing on the input first single-dimensional attention feature and the second single-dimensional attention feature, and obtains the feature processing result.

8. According to claim 1, a method for evaluating national standard adoption based on multi-layer cross-combination attention is characterized in that: The attention cross-recombination module is used to receive the feature processing results of the deep cross-attention module and the feature processing results of the multi-scale fusion cross-attention module, and perform cross-recombination to obtain cross-recombination features, specifically including the following steps: Constructing an attention cross-reorganization module, wherein the constructed attention cross-reorganization module includes two splitting units, six convolution units, and three splicing units; The first splitting unit in the attention cross-reorganization module is used to cross-split the feature processing results of the deep cross-attention module along the channel direction, and output the feature and Features ; The first convolution unit in the attention cross-reorganization module is used to perform convolution operation on the feature processing results of the deep cross-attention module and output the feature ; The second convolution unit in the attention cross-reorganization module is used to perform convolution operation on the feature processing results of the multi-scale fusion cross-attention module and output the feature ; The second splitting unit in the attention cross-reorganization module is used to cross-split the feature processing results of the multi-scale fusion cross-attention module along the channel direction, and output the feature and Features ; The third convolutional unit in the attention cross-reorganization module is used to Perform convolution operation and output features ; The fourth convolutional unit in the attention cross-reorganization module is used to Perform convolution operation and output features ; The first concatenation unit in the attention cross-reorganization module is used to output the features of the first convolution unit. And the features output by the second convolution unit Perform stitching and output stitching features ; The fifth convolutional unit in the attention cross-reorganization module is used to Perform convolution operation and output features ; The sixth convolutional unit in the attention cross-reorganization module is used to Perform convolution operation and output features ; The second concatenation unit in the attention cross-reorganization module is used to adjust the features output by the third convolution unit. And the features output by the fifth convolutional unit Perform stitching and output stitching features ; The third concatenation unit in the attention cross-reorganization module is used to adjust the features output by the fourth convolution unit. And the features output by the sixth convolutional unit Perform stitching and output stitching features ; The stitching features output by the second stitching unit And the splicing features output by the third splicing unit After the shape transformation, the matrix multiplication unit is input for matrix multiplication, and the output is the feature ; The stitching features output by the first stitching unit and Features After the shape transformation, the matrix multiplication unit is input for matrix multiplication, and the feature is output after the shape transformation. , ; The characteristics That is, the cross-recombination features after the cross-recombination of the attention cross-recombination module.

9. The national standard adoption assessment method based on multi-layer cross-reorganization attention according to claim 1 is characterized in that: The indicator indication module is used to receive the cross-recombination feature and output the prediction indication feature corresponding to the text data, and specifically includes the following steps: Construct an indicator module, which includes twelve convolution units, four splicing units, and two fully connected layer units; Among them, the first convolution unit in the indicator module is used to perform convolution operation on the cross-reorganized features after the cross-reorganization of the attention cross-reorganization module, and output the feature ; The seventh convolution unit in the indicator module is used to perform convolution operation on the cross-reorganized features after the cross-reorganization of the attention cross-reorganization module, and output the feature ; The first splicing unit in the indicator module is used to and Features Splice and output the features ; The third splicing unit in the indicator module is used to and Features Splice and output the features ; The second convolution unit in the indicator module is used to output the features of the first concatenation unit. Perform convolution operation and output the features ; The eighth convolution unit in the indicator module is used to output the features of the third concatenation unit. Perform convolution operation and output the features ; The second splicing unit in the indicator module is used to and Features Splice and output the features ; The fourth splicing unit in the indicator module is used to and Features Splice and output the features ; The features obtained After passing through four convolution units for convolution operations in sequence, it passes through a fully connected layer unit and outputs the predicted indicator features. , ; The features obtained After passing through four convolution units for convolution operations in sequence, it passes through a fully connected layer unit and outputs the predicted indicator features. , .

10. The national standard adoption assessment method based on multi-layer cross-reorganization attention according to claim 9 is characterized in that: The score of the international standardized document is calculated according to the prediction indicator features corresponding to all text data in the same international standardized document, specifically including the following steps: Predictive indicator features obtained based on each text data in the same international standardized document , , get the corresponding index screening vector ; According to the index filter vector corresponding to each piece of text data obtained , and use the following formula to get the total evaluation score of the international standardization document: ； In the formula, is the total number of sentence texts in the international standardization document to be evaluated; Represents predictive indicator features The elements; Indicates The index filter vector of sentences; It represents the score matrix obtained by assigning values ​​to indicators according to the importance of different indicators in the evaluation of national standard adoption; express , which is the indicator classification feature vector output by the indicator module.

Images