International standard transformation evaluation method based on cross-granularity feature cross fusion

Abstract

The invention discloses an international standard transformation evaluation method based on cross-granularity feature cross fusion, and relates to the technical field of natural language processing, and the method comprises the steps: 1, collecting a standard class international standard text, and constructing an international standard transformation evaluation index system; 2, constructing an international standard conversion evaluation data set according to the standard international standard text; 3, constructing a cross-granularity feature cross fusion index classification model according to the CNN network, the BERT network and the bidirectional long short-term memory network; 4, training a cross-granularity feature cross fusion index classification model by using an international standard conversion evaluation data set; step 5, constructing an index score assignment rule; and 6, inputting a to-be-evaluated international standard text into the trained cross-granularity feature cross fusion index classification model to obtain a text prediction classification, and calculating an evaluation result in combination with an index assignment rule. According to the method, the international standard conversion evaluation efficiency, rationality and accuracy are effectively improved by using the deep learning algorithm.

Description

Technical Field

[0001] The present invention relates to the technical field of natural language processing, and more specifically to an international standard transformation evaluation method based on cross-granularity feature cross-fusion. Background Art

[0002] In the context of globalization, with the acceleration of the process of global economic integration, trade between different countries and regions is becoming increasingly frequent. A set of commonly recognized standards is needed to ensure the interchangeability and compatibility of products. The introduction and transformation of international standards has also become an important means to promote scientific and technological progress and economic development, and an important means to ensure the quality and safety of products and services and promote international trade. In recent years, with the development of technology and the deepening of international cooperation, the trend of transforming international standards into domestic standards has become increasingly significant. Many countries and regions have improved their own technical level and competitiveness by introducing advanced international standards. However, due to differences in laws, cultures and market environments among countries, international standards often face many challenges when they are transformed into domestic standards, such as inconsistent technical requirements, differences in intellectual property protection mechanisms, and localization adaptability. Therefore, it has become an important issue that countries must face in the process of introducing international standards to reasonably evaluate and transform international standards and formulate technical standards that meet the needs of the domestic market. This will not only help reduce friction and obstacles in technology transformation, but also help promote scientific and technological innovation and industrial upgrading, thereby enhancing the country's scientific and technological strength and international competitiveness.

[0003] At present, the evaluation work of the conversion of international standards mainly relies on expert discussion and evaluation, which has problems such as long conversion cycle and high evaluation complexity. With the rapid development of technologies in the fields of artificial intelligence and machine learning in recent years, how to use machine learning algorithms to automatically evaluate foreign standards has become a hot and difficult research issue.

[0004] Therefore, how to improve the efficiency of international standard conversion evaluation is an urgent problem that technical personnel in this field need to solve. Summary of the invention

[0005] In view of this, the present invention provides an international standard conversion evaluation method based on cross-granularity feature cross-fusion, which uses a deep learning algorithm to effectively improve the efficiency, rationality and accuracy of international standard conversion evaluation.

[0006] In order to achieve the above object, the present invention adopts the following technical solution:

[0007] The international standard transformation evaluation method based on cross-granularity feature cross-fusion includes the following steps:

[0008] Step 1: Collect normative international standard texts and build an international standard conversion evaluation index system;

[0009] Step 2: Construct an international standard conversion evaluation dataset based on the normative international standard texts;

[0010] Step 3: Construct a cross-granularity feature cross-fusion index classification model based on the CNN network, BERT network and bidirectional long short-term memory network;

[0011] Step 4: Use the international standard transformation evaluation dataset to train a cross-granularity feature cross-fusion indicator classification model;

[0012] Step 5: Construct indicator scoring rules;

[0013] Step 6: Input the international standard text to be evaluated into the trained cross-granularity feature cross-fusion index classification model to obtain the text prediction classification, and calculate the evaluation result based on the index scoring rules.

[0014] Preferably, the specific implementation process of step 1 is:

[0015] Step 11: Collect the converted texts of international standards for normative purposes and analyze them, and select primary and secondary indicators;

[0016] Step 12: Map the primary and secondary indicators to build an international standard transformation evaluation indicator system.

[0017] Preferably, 4 first-level indicators and 14 second-level indicators are selected; the first-level indicators include: veto items, comprehensive environment, technical content and expected benefits; the second-level indicators include non-compliance with laws and regulations, non-compliance with basic natural conditions, non-compliance with industry needs and environment, non-adaptability of technical content, policy environment, natural conditions, human factors, industrial structure matching, advancement, adaptability, coordination, economic benefits, social benefits and ecological and environmental benefits; among them, veto items map non-compliance with laws and regulations, non-compliance with basic natural conditions, non-compliance with industry needs and environment, non-adaptability of technical content, comprehensive environment maps policy environment, natural conditions, human factors and industrial structure matching, technical content maps advancement, adaptability and coordination, and expected benefits map economic benefits, social benefits and ecological and environmental benefits.

[0018] Preferably, the specific implementation process of step 2 is:

[0019] Step 21: Map the first-level indicators in the international standard conversion evaluation indicator system to label one, map the second-level indicators to label two, and mark label one and label two;

[0020] Step 22: Decompose the normative international standard text into sentences to obtain several standard text sentences, and construct an international standard conversion evaluation dataset based on all standard text sentences and their corresponding label one and label two annotations.

[0021] Preferably, the labels 1 of the four first-level indicators are marked as 0, 1, 2, and 3, respectively, and the labels 2 of the 14 second-level indicators are marked as a, b, c, d, e, f, g, h, i, j, k, l, m, and n, respectively; the format of the international standard transformation evaluation data set is [label one marking, label two marking, standard text sentence].

[0022] Preferably, the cross-granularity feature cross-fusion index classification model constructed in step 3 includes a word embedding and feature extraction network, a cross-granularity attention module, a character feature matrix dimension mapping network A, a character feature matrix dimension mapping network B, a multi-level character information extraction module, a cross-domain information cross module and an index mapping network; the international standard transformation evaluation data set input word embedding is embedded in the feature extraction network, and the output feature ;feature Input character feature matrix dimension mapping network A, output feature ;feature Input cross-granularity attention module, output features ;feature Input into the multi-level character information extraction module, output feature ;feature Input character feature matrix dimension mapping network B, output feature ;feature and Features Input the cross-domain information cross module and output the feature and Features ;feature and Features Input indicator mapping network to get features and Features As text prediction classification.

[0023] Preferably, the word embedding and feature extraction network includes a filling module and a BERT pre-trained model; the international standard conversion evaluation data set is input into the filling module, and each data is filled with 0 to a set number of characters to obtain enhanced data; the enhanced data is input into the BERT pre-trained model, and the characters are embedded in words and preliminarily extracted to obtain features. , , Represents the dimension of the feature.

[0024] Preferably, the character feature matrix dimension mapping network A and the character feature matrix dimension mapping network B both include a first convolution module, a second convolution module, a third convolution module, a splicing module and a dimension mapping module; in the character feature matrix dimension mapping network A, the feature Input the first convolution module and output the features ,feature Input the second convolution module and the third convolution module respectively, and output features respectively and Features ,feature and Features Both input the splicing module and output features ,feature Input dimension mapping module, output ; In the character feature matrix dimension mapping network B, the feature Input the first convolution module, pass through the second convolution module, the third convolution module, the splicing module and the dimension mapping module in turn, and output the feature .

[0025] Preferably, the cross-granularity attention module includes a splitting module, a plurality of bidirectional long short-term memory modules, a splicing module, a layer normalization module, a fully connected layer 1, a fully connected layer 2, a fully connected layer 3, a plurality of shape reshaping modules and a plurality of matrix multiplication modules; Features Input to split module, output features ,feature Input into the i-th bidirectional long short-term memory module and output feature ,feature Input into the splicing module and output features ,feature Input to the layer normalization module and output features ,feature Input fully connected layer 1, fully connected layer 2 and fully connected layer 3 respectively, and output features respectively ,feature and Features ,feature Input into the first reshape module, output feature ,feature Input into the second shape reshaping module, output feature ,feature and Features At the same time, it is input into the first matrix multiplication module and outputs the features ,feature Input into the third shape reshaping module, output feature , the feature and Features At the same time, it is input into the second matrix multiplication module and the output feature , the feature Input into the fourth shape reshaping module and output the reshaped features .

[0026] Preferably, the multi-level character information extraction module includes a layer normalization module, a first partial convolution module, a second partial convolution module, a third partial convolution module, a first depth convolution module, a second depth convolution module, a GeLu activation module, a plurality of matrix addition modules, a first deconvolution module, a second deconvolution module, an average pooling module, a plurality of matrix multiplication modules, a splicing module and a convolution module; the feature In the input layer normalization module, the output features ,feature Input the first part of the convolution module and output the features ,feature Input into the first deep convolution module, output feature ,feature Input the second part of the convolution module and output the features ,feature Input into the second deep convolution module, output feature ,feature Input into GeLU activation module, output feature ,feature and Features At the same time, it is input into the first matrix addition module and outputs the feature ,feature Input the third part of the convolution module and output the features ,feature Input into the first deconvolution module, output feature ,feature Input average pooling module, output features ,feature Input into the second deconvolution module, output feature ,feature and Features At the same time, it is input into the first matrix multiplication module and outputs the features ,feature and Features At the same time, it is input into the second matrix multiplication module and the output feature ,feature and Features At the same time, input the splicing module and output the features , the feature Input convolution module, output features ,feature and Features At the same time, it is input into the second matrix addition module and the output feature .

[0027] Preferably, the cross-domain information cross module includes a first convolution module, a second convolution module, a third convolution module, a fourth convolution module, a fifth convolution module, a sixth convolution module, a seventh convolution module, an eighth convolution module, a first normalization layer, a second normalization layer, a fully connected layer 1, a fully connected layer 2, a plurality of matrix multiplication modules, a Sigmoid activation module and a plurality of matrix addition modules; Features Input into the first convolution module, output features , the feature Input into the second convolution module, output features , the feature Input into the third convolution module, output features , the feature Input into the fourth convolution module, output feature , the feature Input into the first normalization layer, output features , the feature Input fully connected layer 1, output features , the feature Input into the fifth convolution module, output feature , the feature Input into the sixth convolution module, output feature , the feature Input into the seventh convolution module, output feature , the feature Input into the eighth convolution module, output feature , the feature Input into the second normalization layer, output feature , the feature Input to the fully connected layer 2, output features , the feature and Features At the same time, it is input into the first matrix multiplication module and outputs the features , the feature Input into the Sigmoid activation module, output features , the feature and Features At the same time, it is input into the second matrix multiplication module and the output feature , the feature and Features At the same time, it is input into the third matrix multiplication module and the output feature , the feature and Features At the same time, it is input into the first matrix addition module and outputs the feature , the feature and Features At the same time, it is input into the second matrix addition module and the output feature Among them, the characteristics ,feature .

[0028] Preferably, the indicator mapping network includes a first convolution module, a second convolution module, a third convolution module, a fourth convolution module, a plurality of shape reshaping modules, a first fully connected module, a second fully connected module, a third fully connected module, a fourth fully connected module, a fifth fully connected module, a sixth fully connected module, a seventh fully connected module, a first Softmax activation module and a second Softmax activation module; the feature Input into the first convolution module, output features , the feature Input into the second convolution module, output features , the feature Input into the third convolution module, output features , the feature Input into the fourth convolution module, output feature , the feature Input into the first reshape module, output feature , the feature Input into the second shape reshaping module, output feature , the feature Input into the first fully connected module, output feature , the feature Input into the second fully connected module, output feature , the feature Input into the third fully connected module, output feature , the feature Input into the fourth fully connected module, output feature , the feature Input into the first Softmax activation module, output feature , the feature Input into the fifth fully connected module, output feature , the feature Input into the sixth fully connected module, output feature , the feature Input into the seventh fully connected module, output feature , the feature Input into the second Softmax activation module, output feature Among them, the characteristics ,feature .

[0029] Preferably, in step 4, during the training process, a cross entropy loss function is used to calculate the label one classification loss and the label two classification loss respectively, the label one classification loss and the label two classification loss are added together to calculate the model loss function, and an SGD optimizer is used for gradient optimization.

[0030] Preferably, in step 5, a score is assigned to each secondary indicator to construct an indicator scoring rule.

[0031] Preferably, the specific implementation process of step 6 is:

[0032] Step 61: Split the international standard text to be evaluated into several groups of standard text sentences, input all standard text sentences into the trained cross-granularity feature cross-fusion index classification model, and obtain a set of features for each group of standard text sentences. and Features ;

[0033] Step 62: Based on the characteristics of each set of standard text sentences Match the corresponding first-level indicators, set the mask matrix according to the matching results, and compare the mask matrix with the feature Multiply to obtain the discarded vector A;

[0034] Step 63: Convert the indicator scoring rule into a value matrix G;

[0035] Step 64: Calculate the product of the discard vector A and the score matrix G of each standard text sentence;

[0036] Step 65: Calculate the evaluation score based on the average value of the product results of all standard text sentences, and obtain the evaluation result in combination with the set evaluation threshold.

[0037] Preferably, the feature in step 62 is a one-dimensional vector consisting of 4 values, representing the predicted classification result of label one, feature is a one-dimensional vector consisting of 14 values, representing the predicted classification result of label 2; when feature When the maximum value is the first value, set the mask matrix to a full zero vector. When the maximum value is not the first value, all the label 2s corresponding to the secondary indicators mapped by the primary indicator matching the maximum value are set to 1, and all the label 2s corresponding to other secondary indicators are set to 0 to construct a mask matrix.

[0038] Preferably, the expression for calculating the evaluation score score is:

[0039]

[0040] Among them, A represents the discarded vector; G represents the score matrix; and n represents the number of standard text sentences.

[0041] Through the above technical solutions, it can be known that compared with the prior art, the present invention discloses an international standard conversion evaluation method based on cross-granularity feature cross-fusion, constructs an international standard conversion evaluation system, maps the international standard conversion evaluation system into data labels, and converts the evaluation task into an indicator classification problem, and then uses a deep learning model to train and learn the indicators, outputs the indicator classification results through the model, and then evaluates and scores the international standard text according to the indicator classification results, and judges whether the international standard text can be converted into a national standard based on the score. The deep learning model in the present invention combines the use of cross-level and cross-granularity feature extraction algorithms to process features, and uses a bidirectional long short-term memory module containing time steps to learn context features, so that the final classification features are more accurate, thereby making the evaluation results more reasonable, which can effectively improve the efficiency of the international standard text conversion evaluation process, and achieve accurate evaluation of a large number of international standard text conversions in the same time. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] 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.

[0043] Figure 1 It is a flow chart of the international standard transformation evaluation method based on cross-granularity feature cross-fusion provided by the present invention;

[0044] Figure 2 It is the international standard conversion evaluation index system provided by the present invention;

[0045] Figure 3 It is a structural diagram of the hierarchical feature cross-fusion index classification model provided by the present invention;

[0046] Figure 4 It is a structural diagram of the word embedding and feature extraction network provided by the present invention;

[0047] Figure 5 It is a structural diagram of the character feature matrix dimension mapping network A provided by the present invention;

[0048] Figure 6 is a structural diagram of the cross-granularity attention module provided by the present invention;

[0049] Figure 7 It is a structural diagram of the multi-level character information extraction module provided by the present invention;

[0050] Figure 8It is a structural diagram of the cross-domain information cross-connection module provided by the present invention;

[0051] Fig. 9 It is a structural diagram of the indicator mapping network provided by the present invention. DETAILED DESCRIPTION

[0052] 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.

[0053] The embodiment of the present invention discloses an international standard transformation evaluation method based on cross-granularity feature cross-fusion, such as Figure 1 As shown, the specific steps include the following steps.

[0054] S1: Using normative international standard texts as data sources, construct an international standard transformation evaluation index system;

[0055] S2: Collect normative international standard texts, build an international standard conversion evaluation dataset, and divide it into training set, validation set, and test set;

[0056] S3: Build a cross-granularity feature cross-fusion indicator classification model;

[0057] S4: Use the training set and the validation set to train a cross-granularity feature cross-fusion indicator classification model to obtain a deployable cross-granularity feature cross-fusion indicator classification model;

[0058] S5: Construct the indicator scoring rules based on the evaluation results of known normative international standard texts;

[0059] S6: Deploy the cross-granularity feature cross-fusion indicator classification model, use the test set and international standard text to input the cross-granularity feature cross-fusion indicator classification model, the test set is used to evaluate the model prediction classification effect, and the international standard text input model obtains text prediction classification; combine the text prediction classification and indicator scoring rules to calculate the score of the international standardized text to obtain the evaluation score, use the evaluation score and the preset evaluation threshold to perform transformation evaluation on the international standard text to obtain the evaluation result.

[0060] In a specific embodiment, the international standard transformation evaluation index system constructed by S1 is as follows: Figure 2 As shown, the specific process is:

[0061] S1-1 summarizes 4 first-level indicators and 14 second-level indicators by disassembling and analyzing the normative international standard texts that have not been converted but the evaluation results are known; the first-level indicators include "veto items", "comprehensive environment", "technical content" and "expected benefits"; the second-level indicators include "not in compliance with the provisions of national laws, regulations, policies and systems, and mandatory standards", "not in compliance with the basic natural conditions of the country", "no industrial demand, not in line with the industrial environment and industrial reality", "basic technical content is not adapted to the country", "policy environment", "natural conditions", "human factors", "industrial structure matching", "advancedness", "adaptability", "coordination", "economic benefits", "social benefits", and "ecological and environmental benefits";

[0062] S1-2 maps the primary and secondary indicators to each other according to the mutual matching between the primary and secondary indicators; in the veto items, it includes 4 secondary indicators, namely "not in compliance with the provisions of national laws, regulations, policies and systems, and mandatory standards", "not in compliance with the basic natural conditions of the country", "no industrial demand, not in line with the industrial environment and industrial reality" and "basic technical content is not adapted to the country"; in the comprehensive environment, it includes 4 secondary indicators, namely "policy environment", "natural conditions", "human factors" and "industrial structure matching degree"; in the technical content, it includes 3 secondary indicators, namely "advancedness", "adaptability" and "coordination"; in the expected benefits, it includes 3 secondary indicators, namely "economic benefits", "social benefits" and "ecological and environmental benefits".

[0063] Furthermore, in S2, based on the normative international standard texts that have not been converted but the evaluation results are known, the specific process of constructing an international standard conversion evaluation dataset and dividing it into a training set, a validation set, and a test set is as follows:

[0064] S2-1 maps the first-level indicators in the international standard conversion evaluation indicator system to label one, and maps the second-level indicators in the international standard conversion evaluation indicator system to label two;

[0065] There are 4 labels, which are marked by 0, 1, 2, and 3, respectively, where 0 is the veto item, 1 is the comprehensive environment, 2 is the technical content, and 3 is the expected benefit;

[0066] There are 14 labels in the second label, which are marked with a, b, c, d, e, f, g, h, i, j, k, l, m, and n. Among them, a means that it does not comply with the provisions of national laws, regulations, policies, systems, and mandatory standards; b means that it does not comply with the basic natural conditions of the country; c means that it has no industrial demand and does not comply with the industrial environment and industrial reality; d means that the basic technical content is not adapted to the country; e means the policy environment; f means the natural conditions; g means the human factors; h means the matching degree of the industrial structure; i means the advancement; j means the adaptability; k means the coordination; l means the economic benefits; m means the social benefits; and n means the ecological and environmental benefits.

[0067] S2-2 constructs an international standard conversion evaluation dataset by disassembling the standard text sentences of normative international standard texts. The international standard conversion evaluation dataset can be expressed as: , where the first data in [] is the label one, the second data is the label two, and the third data is the standard text sentence disassembled from the standardized international standard text;

[0068] S2-3 divides the international standard transformation evaluation dataset into training set, validation set and test set in a ratio of 6:2:2.

[0069] Furthermore, the cross-granularity feature cross-fusion index classification model built by S3 includes word embedding and feature extraction network, cross-granularity attention module, character feature matrix dimension mapping network A, character feature matrix dimension mapping network B, multi-level character information extraction module, cross-domain information cross module and index mapping network. The data transmission relationship between modules is as follows: Figure 3 shown.

[0070] Furthermore, during the model training process of S4, the first Data B i Input word embedding and feature extraction network, output features , the feature They are input into the character feature matrix dimension mapping network A and the cross-granularity attention module respectively, and the features are output respectively. and , the feature Input into the multi-level character information extraction module, output feature , the feature Input character feature matrix dimension mapping network B, output feature , the feature and Features All are input into the cross-domain information cross module, and the output features and Features , the feature and Features All are input into the index mapping network to obtain the characteristics and Features ;feature Responsible for predicting and classifying the first-level indicators, features Responsible for predicting and classifying secondary indicators; using features and Features Calculate the label one loss and label two loss in the training model process respectively, and calculate the model loss by adding the label one loss and the label two loss , and perform gradient optimization through the optimizer to obtain a trained cross-granularity feature cross-fusion indicator classification model.

[0071] Furthermore, the loss functions used in the two loss calculations are both cross entropy loss functions, which are defined as follows:

[0072]

[0073] in, is the loss value, Indicates the sample number in the training set or validation set. Indicates the number of samples in the training set or validation set. One-hot encoding representing the true value of the label, Represents the label prediction value.

[0074] Furthermore, the SGD optimizer is used to perform gradient optimization on the cross-granularity feature cross-fusion indicator classification model.

[0075] In a specific embodiment, S4 is the process of processing the data in the data set by the cross-granularity feature cross-fusion index classification model during the training process:

[0076] S4-1: Word embedding and feature extraction network Figure 4 As shown in the figure, the international standard conversion assessment data set Data B i After filling 0s to 384 characters, the characters are input into the BERT pre-trained model, and word embedding and feature extraction are performed on the characters to output the features. , , Represents the dimension of the feature;

[0077] S4-2: Character feature matrix dimension mapping network A Figure 5 As shown, it includes a first convolution module, a second convolution module, a third convolution module, a splicing module and a dimension mapping module; the feature Input the first convolution module and output features ; The feature Input the second convolution module and the third convolution module respectively, and output features respectively and Features ; The feature and Features Both are input into the splicing module, the two features are spliced ​​along the channel direction, and the output feature ,feature The number of channels is twice that of the input; Input dimension mapping module, transform features Multiply the feature dimension by a learnable dimension mapping matrix to map the feature dimension to a feature square matrix equal to the number of characters. ;

[0078] S4-3: Cross-granularity attention module Figure 6 As shown, it includes a splitting module, several bidirectional long short-term memory modules, a splicing module, a layer normalization module, a fully connected layer 1, a fully connected layer 2, a fully connected layer 3, several shape reshaping modules and several matrix multiplication modules; the feature Input splitting module, split the features according to the first dimension, and output features ; The feature Enter to In the bidirectional long short-term memory module (i.e., each feature split out by the splitting module is input into a bidirectional long short-term memory module), the output feature ; The feature Input into the splicing module and output features ; The feature In the input layer normalization module, the output is the feature ; The feature Input fully connected layer 1, fully connected layer 2, and fully connected layer 3 respectively, and output features respectively , , ; The feature In the input reshape module, the last two dimensions are mapped to one dimension and the first dimension is swapped, and the output feature ; The feature In the input reshape module, the last two dimensions are mapped into one dimension and the output features are ; The feature , All are input into the matrix multiplication module and the output features ; The feature In the input reshape module, the last two dimensions are mapped to one dimension and the first dimension is swapped, and the output feature ; The feature , All are input into the matrix multiplication module, and the output features ; The feature In the input reshape module, the second dimension is reshaped after the positions of the two dimensions are swapped. The reshaped dimension is consistent with the feature Consistent, that is, output features ;

[0079] S4-4: Multi-level character information extraction module Figure 7As shown, it includes a layer normalization module, a first partial convolution module, a second partial convolution module, a third partial convolution module, a first depth convolution module, a second depth convolution module, a GeLu activation module, several matrix addition modules, a first deconvolution module, a second deconvolution module, an average pooling module, several matrix multiplication modules, a splicing module and a convolution module; the feature In the input layer normalization module, the output features ; The feature Input the first part of the convolution module and output the features ; The feature Input into the first deep convolution module, output feature ; The feature Input the second part of the convolution module and output the features ; The feature Input into the second deep convolution module, output feature ; The feature Input into GeLU activation module, output feature ; The feature and All are input into the matrix addition module, and the output features ; The feature Input the third part of the convolution module and output the features ; The feature Input into the first deconvolution module, output feature ; The feature Input average pooling module, output features ; The feature Input into the second deconvolution module, output feature ; The feature and All are input into the matrix multiplication module, and the output features ; The feature and All are input into the matrix multiplication module, and the output features ; The feature and Both are input into the splicing module, and the two features are spliced ​​along the channel direction to output the feature ; The feature Input convolution module, output features ; The feature and All are input into the matrix addition module, and the output is the feature ;

[0080] S4-5: The architecture of the character feature matrix dimension mapping network B is consistent with that of the character feature matrix dimension mapping network A. The two networks only differ in the number of convolution kernels in the convolution module, which results in a different number of channels for the output features. Input character feature matrix dimension mapping network B, output feature ;

[0081] S4-6: Cross-domain information cross-module Figure 8 As shown, it includes a first convolution module, a second convolution module, a third convolution module, a fourth convolution module, a fifth convolution module, a sixth convolution module, a seventh convolution module, an eighth convolution module, a first normalization layer, a second normalization layer, a fully connected layer 1, a fully connected layer 2, several matrix multiplication modules, a Sigmoid activation module and several matrix addition modules; the feature Input into the first convolution module, output features ; The feature Input into the second convolution module, output features ; The feature Input into the third convolution module, output features ; The feature Input into the fourth convolution module, output feature ; The feature Input into the first normalization layer, output features ; The feature Input fully connected layer 1, output features ; The feature Input into the fifth convolution module, output feature ; The feature Input into the sixth convolution module, output feature ; The feature Input into the seventh convolution module, output feature ; The feature Input into the eighth convolution module, output feature ; The feature Input into the second normalization layer, output feature ; The feature Input to the fully connected layer 2, output features ; The feature and All are input into the matrix multiplication module, and the output features ; The feature Input into the Sigmoid activation module, output features ; The feature and All are input into the matrix multiplication module, and the output features ; The feature and All are input into the matrix multiplication module, and the output features ; The feature and All are input into the matrix addition module, and the output features ; The feature and All are input into the matrix addition module, and the output features Among them, the characteristics ,feature ;

[0082] S4-7: Indicator mapping network Fig. 9 As shown, it includes a first convolution module, a second convolution module, a third convolution module, a fourth convolution module, several shape reshaping modules, a first fully connected module, a second fully connected module, a third fully connected module, a fourth fully connected module, a fifth fully connected module, a sixth fully connected module, a seventh fully connected module, a first Softmax activation module and a second Softmax activation module; the feature Input into the first convolution module, output features ; The feature Input into the second convolution module, output feature ; The feature Input into the third convolution module, output features ; The feature Input into the fourth convolution module, output feature ; The feature Input shape reshaping module, output features ; The feature Input shape reshaping module, output features ; The feature Input into the first fully connected module, output feature ; The feature Input into the second fully connected module, output feature ; The feature Input into the third fully connected module, output feature ; The feature Input into the fourth fully connected module, output feature ; The feature Input into the first Softmax activation module, output feature ; The feature Input into the fifth fully connected module, output feature ; The feature Input into the sixth fully connected module, output feature ; The feature Input into the seventh fully connected module, output feature ; The feature Input into the second Softmax activation module, output feature Among them, the characteristics ,feature .

[0083] Furthermore, the test set is used to test and evaluate the trained cross-granularity feature cross-fusion indicator classification model, and the optimal model is selected for deployment.

[0084] In a specific embodiment, when constructing the indicator scoring rules in S5, a numerical value is used to identify the importance of the indicator's impact on the evaluation results, where 0 points are given for non-compliance with national laws, regulations, policies, systems, and mandatory standards; 0 points are given for non-compliance with the country's basic natural conditions; 0 points are given for no industry demand and non-compliance with the industry environment and industry reality; 0 points are given for basic technical content not adapting to the country; 15 points are given for policy environment; 15 points are given for natural conditions; 15 points are given for human factors; 15 points are given for industrial structure matching; 20 points are given for advancement; 20 points are given for adaptability; 20 points are given for coordination; 20 points are given for economic benefits; 20 points are given for social benefits; and 20 points are given for ecological and environmental benefits.

[0085] Furthermore, S6 inputs the international standard text to be evaluated into the trained cross-granularity feature cross-fusion index classification model to obtain text prediction classification, including feature and ; Calculate the score of the international standard text to be evaluated based on the text prediction classification; specifically include:

[0086] S6-1: Split the international standard text to be evaluated into several groups of standard text sentences, input all standard text sentences into the trained cross-granularity feature cross-fusion index classification model, and obtain a set of features for each group of standard text sentences. and Features ; Decompose the international standard document to be evaluated into sentences, and obtain n standard text sentences in total;

[0087] S6-2: Based on the characteristics of each set of standard text sentences Match the corresponding first-level indicators, set the mask matrix according to the matching results, and compare the mask matrix with the feature Multiply to obtain the discarded vector A;

[0088] feature is a one-dimensional vector consisting of 4 values, responsible for predicting the classification of label one, feature is a one-dimensional vector consisting of 14 values, responsible for predicting the classification of label 2; when feature When the maximum value of is the first value, the prediction result is "0", that is, "veto item", then set an all-zero vector , ,Will With features Multiply to get the discard vector When the feature When the maximum value of is the second value, the prediction result is "1", that is, "comprehensive environment", then set a mask matrix ,Will With features Multiply to get the discard vector When the feature When the maximum value of is the third value, the prediction result is "2", that is, "technical content", then a mask matrix is ​​set ,Will With features Multiply to get the discard vector When the feature When the maximum value of is the fourth value, the prediction result is "3", that is, "expected benefit", then a mask matrix is ​​set ,Will With features Multiply to get the discard vector ;

[0089] S6-3: Convert the indicator scoring rules into a score matrix G; Convert the above score identifiers into a score matrix ;

[0090] S6-4: Calculate the discard vector A and score matrix for each standard text sentence The product of multiplication;

[0091] S6-5: Calculate the evaluation score according to the average value of the product results of all standard text sentences, and obtain the evaluation result in combination with the set evaluation threshold;

[0092] The evaluation score is expressed as:

[0093]

[0094] in, The function is to adjust the score to a percentage.

[0095] On the other hand, in a specific embodiment, an international standard transformation evaluation system based on cross-granularity feature cross-fusion includes:

[0096] The international standard indicator classification module inputs the international standard text to be evaluated into the trained cross-granularity feature cross-fusion indicator classification model and outputs the text prediction classification;

[0097] The score calculation and evaluation module inputs the text prediction classification into the indicator scoring rule, calculates the evaluation score, and obtains the evaluation result.

[0098] On the other hand, in a specific embodiment, an electronic device includes a memory and a processor and computer instructions stored in the memory and executed on the processor, and when the computer instructions are executed by the processor, the steps of the international standard transformation evaluation method based on cross-granularity feature cross-fusion are completed.

[0099] On the other hand, in a specific embodiment, a computer-readable storage medium is used to store computer instructions, which, when executed by a processor, complete the steps of an international standard transformation evaluation method based on cross-granularity feature cross-fusion.

[0100] 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.

[0101] 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. An international standard transformation evaluation method based on cross-granularity feature cross-fusion, characterized in that: The following steps are involved: Step 1: Collect normative international standard texts and build an international standard conversion evaluation index system; Step 2: Construct an international standard conversion evaluation dataset based on the normative international standard texts; Step 3: Construct a cross-granularity feature cross-fusion index classification model based on the CNN network, BERT network and bidirectional long short-term memory network; Step 4: Use the international standard transformation evaluation dataset to train the cross-granularity feature cross-fusion index classification model; the constructed cross-granularity feature cross-fusion index classification model includes word embedding and feature extraction network, cross-granularity attention module, character feature matrix dimension mapping network A, character feature matrix dimension mapping network B, multi-level character information extraction module, cross-domain information cross module and index mapping network; the international standard transformation evaluation dataset input word embedding is embedded in the feature extraction network, and the output feature ;feature Input character feature matrix dimension mapping network A, output feature ;feature Input cross-granularity attention module, output features ;feature Input into the multi-level character information extraction module, output feature ;feature Input character feature matrix dimension mapping network B, output feature ;feature and Features Input the cross-domain information cross module and output the feature and Features ;feature and Features Input indicator mapping network to get features and Features As a text prediction classification; Step 5: Construct indicator scoring rules; Step 6: Input the international standard text to be evaluated into the trained cross-granularity feature cross-fusion index classification model to obtain the text prediction classification, and calculate the evaluation result based on the index scoring rules.

2. The international standard transformation evaluation method based on cross-granularity feature cross-fusion according to claim 1 is characterized in that: The international standard conversion evaluation index system includes primary indicators and secondary indicators with mapping relationships. The specific implementation process of step 2 is as follows: Step 21: Map the first-level indicators in the international standard conversion evaluation indicator system to label one, map the second-level indicators to label two, and mark label one and label two; Step 22: Decompose the normative international standard text into sentences to obtain several standard text sentences, and construct an international standard conversion evaluation dataset based on all standard text sentences and their corresponding label one and label two annotations.

3. The international standard transformation evaluation method based on cross-granularity feature cross-fusion according to claim 1 is characterized in that: The word embedding and feature extraction network includes a filling module and a BERT pre-trained model; the international standard conversion evaluation data set is input into the filling module, and each data is filled with 0 to the set number of characters to obtain enhanced data; the enhanced data is input into the BERT pre-trained model, and the characters are embedded and preliminarily extracted to obtain features , , Represents the dimension of the feature.

4. The international standard transformation evaluation method based on cross-granularity feature cross-fusion according to claim 1 is characterized in that: The character feature matrix dimension mapping network A and the character feature matrix dimension mapping network B both include a first convolution module, a second convolution module, a third convolution module, a splicing module and a dimension mapping module; in the character feature matrix dimension mapping network A, the feature Input the first convolution module and output the features ,feature Input the second convolution module and the third convolution module respectively, and output features respectively and Features ,feature and Features Both input the splicing module and output features ,feature Input dimension mapping module, output ; In the character feature matrix dimension mapping network B, the feature Input the first convolution module, pass through the second convolution module, the third convolution module, the splicing module and the dimension mapping module in turn, and output the feature , Represents the dimension of the feature.

5. The international standard transformation evaluation method based on cross-granularity feature cross-fusion according to claim 1 is characterized in that: The cross-granularity attention module includes a splitting module, several bidirectional long short-term memory modules, a splicing module, a layer normalization module, a fully connected layer 1, a fully connected layer 2, a fully connected layer 3, several shape reshaping modules and several matrix multiplication modules; Features Input to split module, output features ,feature Input into the i-th bidirectional long short-term memory module and output feature ,feature Input into the splicing module and output features ,feature Input to the layer normalization module and output features ,feature Input fully connected layer 1, fully connected layer 2 and fully connected layer 3 respectively, and output features respectively ,feature and Features ,feature Input into the first reshape module, output feature ,feature Input into the second shape reshaping module, output feature ,feature and Features At the same time, it is input into the first matrix multiplication module and outputs the features ,feature Input into the third shape reshaping module, output feature , the feature and Features At the same time, it is input into the second matrix multiplication module and the output feature , the feature Input into the fourth shape reshaping module and output the reshaped features , Represents the dimension of the feature.

6. The international standard transformation evaluation method based on cross-granularity feature cross-fusion according to claim 1 is characterized in that: The multi-level character information extraction module includes a layer normalization module, a first partial convolution module, a second partial convolution module, a third partial convolution module, a first depth convolution module, a second depth convolution module, a GeLu activation module, several matrix addition modules, a first deconvolution module, a second deconvolution module, an average pooling module, several matrix multiplication modules, a splicing module and a convolution module; Features In the input layer normalization module, the output features ,feature Input the first part of the convolution module and output the features ,feature Input into the first deep convolution module, output feature ,feature Input the second part of the convolution module and output the features ,feature Input into the second deep convolution module, output feature ,feature Input into GeLU activation module, output feature ,feature and Features At the same time, it is input into the first matrix addition module and outputs the feature ,feature Input the third part of the convolution module and output the features ,feature Input into the first deconvolution module, output feature ,feature Input average pooling module, output features ,feature Input into the second deconvolution module, output feature ,feature and Features At the same time, it is input into the first matrix multiplication module and outputs the features ,feature and Features At the same time, it is input into the second matrix multiplication module and the output feature ,feature and Features At the same time, input the splicing module and output the features , the feature Input convolution module, output features ,feature and Features At the same time, it is input into the second matrix addition module and the output feature , Represents the dimension of the feature.

7. The international standard transformation evaluation method based on cross-granularity feature cross-fusion according to claim 1 is characterized in that: The cross-domain information cross module includes a first convolution module, a second convolution module, a third convolution module, a fourth convolution module, a fifth convolution module, a sixth convolution module, a seventh convolution module, an eighth convolution module, a first normalization layer, a second normalization layer, a fully connected layer 1, a fully connected layer 2, several matrix multiplication modules, a Sigmoid activation module and several matrix addition modules; Features Input into the first convolution module, output features , the feature Input into the second convolution module, output feature , the feature Input into the third convolution module, output features , the feature Input into the fourth convolution module, output feature , the feature Input into the first normalization layer, output features , the feature Input fully connected layer 1, output features , the feature Input into the fifth convolution module, output feature , the feature Input into the sixth convolution module, output feature , the feature Input into the seventh convolution module, output feature , the feature Input into the eighth convolution module, output feature , the feature Input into the second normalization layer, output feature , the feature Input fully connected layer 2, output features , the feature and Features At the same time, it is input into the first matrix multiplication module and outputs the features , the feature Input into the Sigmoid activation module, output features , the feature and Features At the same time, it is input into the second matrix multiplication module and the output feature , the feature and Features At the same time, it is input into the third matrix multiplication module and the output feature , the feature and Features At the same time, it is input into the first matrix addition module and outputs the feature , the feature and Features At the same time, it is input into the second matrix addition module and the output feature ; Among them, the characteristics ,feature , Represents the dimension of the feature.

8. The international standard transformation evaluation method based on cross-granularity feature cross-fusion according to claim 1 is characterized in that: The indicator mapping network includes a first convolution module, a second convolution module, a third convolution module, a fourth convolution module, several shape reshaping modules, a first fully connected module, a second fully connected module, a third fully connected module, a fourth fully connected module, a fifth fully connected module, a sixth fully connected module, a seventh fully connected module, a first Softmax activation module and a second Softmax activation module; the feature Input into the first convolution module, output features , the feature Input into the second convolution module, output feature , the feature Input into the third convolution module, output features , the feature Input into the fourth convolution module, output feature , the feature Input into the first reshape module, output feature , the feature Input into the second shape reshaping module, output feature , the feature Input into the first fully connected module, output feature , the feature Input into the second fully connected module, output feature , the feature Input into the third fully connected module, output feature , the feature Input into the fourth fully connected module, output feature , the feature Input into the first Softmax activation module, output feature , the feature Input into the fifth fully connected module, output feature , the feature Input into the sixth fully connected module, output feature , the feature Input into the seventh fully connected module, output feature , the feature Input into the second Softmax activation module, output feature ; Among them, the characteristics ,feature , Represents the dimension of the feature.

9. The international standard transformation evaluation method based on cross-granularity feature cross-fusion according to claim 2 is characterized in that: The specific implementation process of step 6 is: Step 61: Split the international standard text to be evaluated into several groups of standard text sentences, input all standard text sentences into the trained cross-granularity feature cross-fusion index classification model, and obtain a set of features for each group of standard text sentences. and Features ; Step 62: Based on the characteristics of each set of standard text sentences Match the corresponding primary index, set the mask matrix according to the secondary index that has a mapping relationship with the matched primary index, and map the mask matrix to the feature Multiply to obtain the discarded vector A; Step 63: Convert the indicator scoring rule into a value matrix G; Step 64: Calculate the product of the discard vector A and the score matrix G of each standard text sentence; Step 65: Calculate the evaluation score based on the average value of the product results of all standard text sentences, and obtain the evaluation result in combination with the set evaluation threshold.

10. The international standard transformation evaluation method based on cross-granularity feature cross-fusion according to claim 9 is characterized in that: Step 62 Features is a one-dimensional vector consisting of 4 values, representing the predicted classification result of label one, feature is a one-dimensional vector consisting of 14 values, representing the predicted classification result of label 2; when feature When the maximum value is the first value, set the mask matrix to a vector of all zeros. When the maximum value is not the first value, all the label 2s corresponding to the secondary indicators mapped by the primary indicator matching the maximum value are set to 1, and all the label 2s corresponding to other secondary indicators are set to 0 to construct a mask matrix.

Images