English reading difficulty classification method based on convergent difference neural network

Abstract

The application discloses an English reading difficulty classification method based on a convergent differential neural network. The method comprises the following steps: collecting an English text data set and performing pretreatment; performing feature aggregation on the English text of the pretreated data set; constructing a convergent differential neural network combined with a voting strategy, and training the convergent differential neural network combined with the voting strategy by using the English text data set after the feature aggregation; inputting new English text data into the trained convergent differential neural network combined with the voting strategy after pretreatment, and obtaining corresponding English reading difficulty classification. Compared with the prior art, the method has higher classification efficiency and classification accuracy for the reading difficulty of English texts.

Description

A method for classifying English reading difficulty based on convergent difference neural networks Technical Field

[0001] This invention belongs to the technical field of neural networks, specifically relating to a method for classifying the difficulty of English reading based on a convergent differential neural network. Background Technology

[0002] Reading comprehension is a crucial language learning skill and a key criterion for assessing language proficiency. In language learning, learners' reading levels and comprehension abilities vary significantly. If reading materials are too difficult, they can create excessive psychological burden, causing learners to lose interest and confidence in learning the language; conversely, overly simple materials not only fail to improve reading skills but also fail to truly assess learners' reading abilities. Therefore, the purpose of categorizing text reading difficulty is to provide learners with reading materials tailored to their individual learning stages and levels.

[0003] In current technologies, the measurement of reading difficulty in English texts is mainly achieved by statistically analyzing factors such as word length and sentence length. This merely treats the text as a string of characters of varying lengths, without delving into the internal workings of the language. Therefore, its accuracy and applicability are greatly limited. For example, to measure text readability, researchers typically divide the text into strings of different lengths, then use computers to perform statistical analysis, and finally calculate readability based on the results. This method has limitations; it only provides static data and cannot delve into the internal structure of the language, resulting in poor accuracy.

[0004] Existing technologies also include some research methods based on statistical language models, which rely on large corpora and calculate information entropy at the lexical, sentence, and discourse levels to construct information theory-based language models and design corresponding automatic English text difficulty measurement systems. However, this method depends on manual extraction of text features and has the disadvantage of requiring a large amount of annotation.

[0005] Existing technologies also include text reading difficulty detection methods based on machine learning and complex features, such as monolingual models and Naive Bayes classification. While machine learning can improve classification efficiency and its performance in terms of classification accuracy meets expectations, it still relies on manual extraction of text features and has not achieved automation of feature extraction. Furthermore, there are currently a large number of different types of text features (semantic, syntactic, etc.) available for manual extraction, making the extraction of new features more difficult. Manual annotation introduces more redundant information, which is detrimental to parallel computing and the rapid convergence of classification models. Summary of the Invention

[0006] To overcome one or more defects and shortcomings of existing technologies, this invention provides an English reading difficulty classification method based on a convergent differential neural network. It employs neural dynamics (NDA) to train the English reading difficulty classification model, which reduces training parameters, improves computational efficiency, and increases accuracy. Furthermore, it combines NDA with a voting strategy to obtain a convergent differential neural network (V-CDNN). K-clustering and principal component analysis techniques are used to weight and extract features from the dataset, transforming the task of measuring text reading difficulty into a classification task. Based on the discrete dynamic equation defined by the derivative of the error, the neural network model is iterated, dynamically adjusting the network's weight parameters. This is then combined with a voting statistics strategy, reducing training time and further improving computational efficiency and classification accuracy.

[0007] To achieve the above objectives, the present invention adopts the following technical solution.

[0008] A method for classifying the difficulty of English reading based on a convergent differential neural network includes the following steps:

[0009] Collect and preprocess English text datasets;

[0010] Feature aggregation is performed on the English text in the preprocessed dataset;

[0011] Converging differential neural networks incorporating voting strategies are constructed and trained using an English text dataset with aggregated features.

[0012] New English text data is acquired, preprocessed, and then input into a convergent differential neural network that has been trained and incorporates a voting strategy to obtain the corresponding English reading difficulty classification.

[0013] Preferably, the specific process of preprocessing is as follows:

[0014] Perform text segmentation, non-English text removal, lowercase letter conversion, and stop word removal on English text data;

[0015] Text segmentation was performed using the word_tokenize tool from the Nltk tokenizer;

[0016] To remove non-English text, the regularization function `re` is introduced to replace non-English text with whitespace.

[0017] Stop words are removed by creating a stop word list and filtering stop words using a word segmentation dictionary based on character matching.

[0018] Furthermore, feature aggregation is performed on the preprocessed English text dataset. The specific process is as follows:

[0019] We define features to represent the reading difficulty of English text, and then use the min-max normalization method to process the English text data;

[0020] The standardized dataset is weighted according to the defined feature terms.

[0021] Furthermore, the features include the character-to-speech ratio, sentence length, and document length of the English text.

[0022] Furthermore, the standardized dataset is weighted according to the defined feature terms, specifically as follows:

[0023] Based on the feature weighting method of K-means clustering, the English text data of each feature category is aggregated.

[0024] Furthermore, a convergent differential neural network incorporating a voting strategy is constructed and trained, as follows:

[0025] Construct three parallel convergent differential neural networks; each convergent differential neural network consists of a fully connected input layer, a hidden layer, and an output layer.

[0026] The error of the output layer of the convergent differential neural network is corrected by backpropagation weights layer by layer according to the neural dynamics formula NDA;

[0027] The three convergent differential neural networks each employ different activation functions;

[0028] Based on three convergent differential neural networks using different activation functions, a convergent differential neural network combining a voting strategy is constructed.

[0029] The dataset after feature aggregation is divided into a training set and a test set. The dataset is then fed into a convergent differential neural network that incorporates a voting strategy for training. The predicted classification of each sample in the test set is obtained, and the accuracy is verified.

[0030] Furthermore, the convergent differential neural network uses the NDA method to iterate the error using the time derivative, continuously updating the weight matrix W. When the error stably converges to the minimum value, the final weights of the convergent differential neural network are obtained.

[0031] Furthermore, the activation functions for the three convergent differential neural networks are the Sinh function, the Tanh function, and the Power function, respectively.

[0032] Furthermore, the convergent differential neural network combined with the voting strategy employs a relative majority voting method to collect the classification results of the three convergent differential neural networks and predict the final category of the sample.

[0033] Furthermore, the relative majority voting method specifically includes:

[0034] Use an initialized zero vector a t =[a t1 … a tq This is used for voting statistics, and the test set is initialized as a 1×m vector x. t =[x t1 … x tq When the convergent differential neural network determines x t When it belongs to the r-th reading difficulty category, a tr Add 1 to accumulate, and finally, based on the majority vote principle in the three convergent difference neural networks, find a. t The position index of the maximum value among all elements is used as the final predicted classification.

[0035] Compared with the prior art, the technical solution of the present invention has the following beneficial effects:

[0036] This invention employs a neurodynamics approach for training the neural network, which reduces training parameters and improves computational efficiency and accuracy. It utilizes a convergent differential neural network constructed using a voting strategy, combining neurodynamics with K-means clustering to weight and extract features from the dataset, transforming the task of measuring text reading difficulty into a classification task, thus avoiding the need for extensive manual annotation required in existing technologies. Furthermore, by combining the iteration of the convergent differential neural network with a voting statistics strategy, this invention further reduces the classifier's training time while achieving higher classification accuracy, demonstrating superior performance in classifying the reading difficulty of English text compared to existing technologies. Attached Figure Description

[0037] Figure 1 is a flowchart illustrating an English reading difficulty classification method based on a convergent differential neural network according to the present invention.

[0038] Figure 2 is a schematic diagram of the structure and principle of a convergent differential neural network;

[0039] Figure 3(a) shows the working characteristic curve of the Sinh function mapping;

[0040] Figure 3(b) shows the working characteristic curve of the Tanh function mapping;

[0041] Figure 3(c) shows the working characteristic curve of the Power function mapping. Detailed Implementation

[0042] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0043] Example

[0044] As shown in Figure 1, the English reading difficulty classification method based on a convergent differential neural network in this embodiment includes the following steps:

[0045] S1. Collect and preprocess the English text dataset; the specific process is as follows:

[0046] S11. Use a Python web crawler to crawl publicly available English text from the Internet to form a dataset;

[0047] S12. In order to filter out noisy words in the text, reduce the computational burden, facilitate subsequent data analysis, and improve the accuracy of the results, the crawled text data is preprocessed. The preprocessing includes text segmentation, removal of non-English text, unification of letters to lowercase, and removal of stop words.

[0048] Text segmentation is as follows: words are the most basic semantic units of sentences. Before analyzing English text content, word segmentation must be performed. In Python web crawler programs, Nltk word segmenter is used for English word segmentation. The word_tokenize tool provided by Nltk word segmenter can also take into account punctuation marks in English.

[0049] Removing non-English text: In most cases, the text contains a lot of content that is not meaningful for classification. For example, the crawled text contains HTML, CSS, and other code tags, or some special symbols that affect the computer's recognition performance. By introducing the regularization function re to replace non-English text with whitespace, non-English text can be removed.

[0050] Convert letters to lowercase: To avoid affecting classification accuracy due to the program's inability to distinguish between uppercase and lowercase letters, all English letters need to be converted to lowercase.

[0051] Stop words removal: Words that appear frequently in the text but have no actual meaning for understanding the sentence, such as "a", "the", and "of", as well as words that also frequently appear in other texts in the dataset, can be called stop words. Stop words do not have discriminative power. Stop words are filtered by building a stop word list and scanning the word segmentation dictionary using character matching.

[0052] S2. Perform feature aggregation on the preprocessed English text dataset; the specific process is as follows:

[0053] S21. Define features representing the reading difficulty of English text, and then process the English text data using the min-max normalization method;

[0054] The difficulty of reading English texts is usually quantified by statistical data on passage length, sentence length, word length, and the number of unfamiliar words in the text. The selection of feature items needs to consider aspects such as vocabulary difficulty, sentence length, sentence complexity, and grammatical structure.

[0055] At the level of vocabulary difficulty, we use statistical analysis of word length and word structure (such as the number of syllables) as the basis for judging vocabulary difficulty, and then calculate the corresponding class-to-form ratio. The class-to-form ratio reflects the richness of vocabulary in English texts. The class-to-form ratio refers to the ratio of different words (classes) to all word forms (forms) in a specific corpus. It is an important indicator for measuring the breadth of the text's vocabulary. The larger the ratio, the lower the vocabulary recurrence rate and the greater the difficulty of the English text. Therefore, the class-to-form ratio is used as a feature item.

[0056] In terms of sentence length and complexity, longer sentences in English texts generally indicate more complex sentence structures, greater information content, and more logical relationships. Considering the impact of sentence complexity on text difficulty, long and complex sentences significantly increase the reader's comprehension difficulty. Since there is a positive correlation between average sentence length (calculated in terms of word count) and sentence complexity, the average sentence length of the text is used as an indicator of text sentence complexity. Therefore, sentence length is considered as another feature.

[0057] At the discourse level, English texts are usually presented in discourse units. It is generally believed that the longer the text, the richer the information it contains, and the higher the cognitive requirements for the reader, thus increasing the difficulty of reading the text. Therefore, the discourse length of English texts is also considered as a feature.

[0058] S22. Perform feature weighting on the standardized dataset based on the defined feature terms; specifically as follows:

[0059] Based on the feature weighting method of K-means clustering, the English text data of each feature category are aggregated, thereby reducing the variance within the feature category and increasing the distinction between feature categories;

[0060] The steps of K-means clustering weighting are as follows: using the k-means clustering algorithm, calculate the cluster center of each attribute (i.e. feature item) in the dataset, then calculate the absolute difference between the data in each attribute and the cluster center, then calculate the average of the absolute differences in each attribute, and divide the average by the cluster center to obtain the corresponding weight coefficient. Finally, multiply the obtained weight coefficient by the attribute value of the corresponding standardized English text data to perform feature weighting.

[0061] S3. Referring to Figure 2, construct a convergent differential neural network combining a voting strategy, and train the convergent differential neural network combining a voting strategy using the English text dataset after feature aggregation; the specific process is as follows:

[0062] S31. Construct three parallel Convergent Difference Neural Networks (CDNNs). CDNNs are feedforward neural networks, possessing both forward propagation of information and backward propagation of error. A CDNN consists of a fully connected input layer, a hidden layer, and an output layer. When the actual output of a CDNN differs from the expected output, the error in the output layer is corrected by backpropagating the weights layer by layer according to the Neural Dynamics Algorithm (NDA).

[0063] Referring to Figure 2, in this embodiment, the preferred input layer of the convergent differential neural network is a 1×m sample vector x = (x1, x2, ... x...). m The output of the input layer is stored in a 1×q vector. The hidden layer f has n neurons. The weights between the input layer and the hidden layer are an m×n matrix v(k). The output of the hidden layer is shown in the following equation:

[0064] u(k)=f(v(k)x) (1)

[0065] Here, k represents the training round; the number of neurons in the output layer is q; the weights between the hidden layers and the output layer are an n×q matrix w(k); and the output of the output layer is shown in the following formula:

[0066] Y=g(f(v(k)x)w(k)) (2)

[0067] In this context, matrix v(k) is randomly initialized and remains unchanged, and f(·) and g(·) are both softsign functions. The softsign function has a flat curve and a slower descent derivative, exhibiting efficient learning capabilities. The softsign function is shown in the following equation:

[0068]

[0069] Where z is the input to the softsign function; unlike the traditional gradient descent learning algorithm, the CDNN in this embodiment uses the NDA method to iterate the error as a time derivative, as shown in the following equation:

[0070]

[0071] in, Let represent the time derivative after weight iteration, and e(t) represent the time derivative of the weights; the discrete form of the NDA method is shown in the following equation:

[0072]

[0073] Wherein, the learning rate λ>0 is a scalar that controls the convergence rate. Adjusting the value of λ can make the convergence smoother during training. The function Φ(·) is monotonically increasing.

[0074] In this embodiment, the activation functions of the three convergent differential neural networks are preferably the Sinh function, the Tanh function, and the Power function, respectively. The specific calculations of the three activation functions are shown in the following formulas:

[0075]

[0076]

[0077] Power: z i ,i≥3 (8)

[0078] The loss function of a convergent differential neural network is shown in the following equation:

[0079]

[0080] Here, k represents the number of times the weights are iterated and updated, and Y(k) represents the prediction result obtained through the convergent differential neural network. Indicates the actual category labels of the test set;

[0081] Substituting the loss function into the discrete NDA formula (5), the weights are iteratively corrected. The weight correction process is as follows: The classification bias in the (k+1)th training round is shown in the following formula:

[0082] E(k+1)=E(k)-λΦ(E(k)) (10)

[0083] Combining equations (9) and (10), the discrete neural dynamics equation is obtained as follows:

[0084] Y(k+1)=Y(k)-λΦ(E(k)) (11)

[0085] Define the input matrix X, the output matrix U of the hidden layer, and the weight matrix W. In the (k+1)th round of training, given U and W(k), equation (11) can be transformed into the following equation:

[0086] g(UW(k+1))=g(UW(k))-λΦ(E(k)) (12)

[0087] Solving equation (12), the update formula for the weight matrix W is shown below:

[0088]

[0089] Among them, U + For the generalized inverse of U, Let g(·) be the inverse function. Since the inverse function of softsign does not exist, an approximate inverse function is used instead, as shown in the following equation:

[0090]

[0091] By continuously updating the weight matrix W, the final weights of the convergent differential neural network can be obtained when the error stabilizes and converges to the minimum value.

[0092] S32. Based on three convergent differential neural networks using different activation functions, construct a convergent differential neural network (V-CDNN classifier) ​​that combines a voting strategy.

[0093] The V-CDNN classifier uses a relative majority voting method to collect the classification results of three convergent differential neural networks and predicts the final class of a sample. The V-CDNN classifier uses an initialized zero vector a. t =[a t1 … a tq This is used for voting statistics, and the test sample is initialized as a 1×m vector x. t =[x t1 … x tq When a convergent differential neural network determines x t When it belongs to the r-th reading difficulty category, a tr Add 1 to accumulate, and finally, based on the majority vote principle in the three convergent difference neural networks, find a. t The position index of the maximum value among all elements is used as the final predicted classification, thus completing the classification of the reading difficulty of English text;

[0094] S33. Finally, the V-CDNN classifier is used to divide the dataset after feature aggregation into a training set and a test set. The training set is used to train the convergent differential neural network with combined voting strategy. Then, the test set is input into the convergent differential neural network with combined voting strategy in the training to obtain the predicted category of each sample in the test set and verify the accuracy, thus completing the training of the convergent differential neural network with combined voting strategy.

[0095] S4. Obtain new English text data, perform the preprocessing in step S1, and input it into the convergent differential neural network with voting strategy trained in step S3 to obtain the corresponding English reading difficulty classification.

[0096] To further verify the effectiveness of this embodiment, a simulation experiment was preferably conducted according to the above steps. The simulation experiment is as follows:

[0097] In this embodiment, the corresponding methods and procedures are preferably written using MATLAB language, and data processing and analysis are performed on the MATLAB R2021B software platform. The English text dataset is obtained from the reading comprehension materials of the past years' national college entrance examination English papers, the national college English test papers, and the English major level 8 papers that are currently publicly available on the Internet. A total of 421 articles were collected in the experiment, including 110 college entrance examination papers, 92 college English test papers, 116 college English test papers, and 103 English major level 8 papers. 337 samples were randomly selected from the total dataset, and CDNNs of the three mapping functions mentioned above were trained respectively. Then, each network independently predicted the English reading difficulty classification of the remaining 84 test samples.

[0098] After the dataset is divided, 5x cross-validation is used. That is, the training set is randomly selected from all datasets, divided into 5 parts on average, and four parts are selected for training in turn, while the remaining parts are used for validation. 5x cross-validation is used to adjust the parameters of CDNN. The obtained 5 sets of parameters are used to train CDNNs with different types of mapping functions in turn, and the highest accuracy is calculated. According to Figures 3(a), 3(b), and 3(c), when the initial value of λ in Equation (5) is 0.025 and the number of hidden layer nodes of the corresponding network model is 12, the network is trained for different mapping functions. Table 1 lists the calculation parameters of different types of CDNN:

[0099] Table 1. Accuracy and training parameters of CDNNs with different mapping functions

[0100]

[0101] These parameters correspond to the highest test accuracy of this CDNN, while the average accuracy is obtained by using different test samples randomly collected from the entire dataset. The experimental results show that the highest accuracy on the test set can reach 100%, and the average accuracy exceeds 98%. This indicates that CDNN has high accuracy in classifying the reading difficulty level of English text, and the longest CPU computation time (i.e., model training time) is only 0.5750s, which shows that the computational efficiency of this embodiment is high.

[0102] During the experiment, the receiver operating characteristic (ROC) curve is a comprehensive indicator reflecting the sensitivity and specificity of continuous variables. The closer the curve is to the upper left corner, the higher its classification accuracy. In the ROC curve, the X-axis represents the false positive rate, and the Y-axis represents the true positive rate. The true positive rate (TPR) represents the proportion of samples that are predicted to be positive and actually are positive out of all positive samples, which is the sensitivity. The false positive rate (FPR) represents the proportion of samples that are predicted to be positive but actually are negative out of all negative samples. The calculation of TPR and FPR are shown in the following formulas:

[0103]

[0104]

[0105] The range of TPR and FRP is [0,1]. The ROC curve reflects the relationship between TRP and FRP. The closer the curve is to the point (0,1), the better the classification performance of this embodiment.

[0106] During training and testing, cross-entropy was used to calculate the trend of training and testing errors of CDNN with three types of mapping functions. The test results showed that the errors decreased monotonically and smoothly as K increased, which indicates that the CDNN and V-CDNN used did not exhibit underfitting or overfitting. Furthermore, when K is large enough, the errors gradually converged.

[0107] This embodiment further verifies the classification performance of V-CDNN by comparing and analyzing classification accuracy, computation time, average precision, and highest precision using a traditional backpropagation neural network (BPNN). The results are shown in Table 2.

[0108] Table 2 Comparison of experimental results between V-CDNN and BPNN classifiers

[0109]

[0110] The evaluation metrics in Table 2 are all weighted averages, with the weight being the proportion of the number of samples in the corresponding category in the total number of samples. Compared with BPNN, V-CDNN has higher test accuracy, shorter computation time, and better performance evaluation metrics. In addition, the trend of training error and test error calculated by cross-entropy shows that when achieving the same training objective and classification performance, the number of training iterations for CDNN with different types of mapping functions is much smaller than that for BPNN. Therefore, it can be seen that the training efficiency of V-CDNN is much higher than that of the BPNN method.

[0111] For classification tasks, it is necessary to objectively determine the quality of a classifier by calculating the numerical values ​​of classification evaluation metrics. Higher scores indicate better classifier performance. Therefore, during training, the values ​​of each evaluation metric should be made as close to 1 as possible. This embodiment uses four common evaluation metrics: Accuracy (ratio of correctly classified samples to the total number of samples), Precision (proportion of correctly predicted positive samples out of all samples predicted as positive, typically used to measure the accuracy of the classifier's positive predictions), Recall (proportion of samples predicted as positive out of all samples actually predicted as positive, typically used to measure the reliability of the classifier's positive predictions), and F1 Score (harmonic mean of precision and recall). The formulas for calculating these four metrics are shown below.

[0112]

[0113]

[0114]

[0115]

[0116] Wherein, TP is the number of samples where both the actual and predicted results are positive, TF is the number of samples where both the actual and predicted results are negative, FP is the number of samples where the actual result is negative and the predicted result is positive, and FN is the number of samples where the actual result is positive and the predicted result is negative.

[0117] This embodiment further verifies the advantages of V-CDNN by comparing its classification performance with several commonly used algorithms. The results are shown in Table 3.

[0118] Table 3 Comparison of experimental results for different classifiers

[0119]

[0120] Experimental results show that V-CDNN performs significantly better than other models, indicating that the V-CDNN designed in this embodiment has significant optimization and improvement.

[0121] Compared with the prior art, the beneficial effects of this embodiment are as follows:

[0122] This embodiment employs a neurodynamics approach to train the neural network, which reduces training parameters and improves computational efficiency and accuracy. It utilizes a convergent differential neural network constructed using a voting strategy, employing K-means clustering to weight and extract features from the dataset, transforming the task of measuring text reading difficulty into a classification task, thus avoiding the need for extensive manual annotation required in existing technologies. Furthermore, by combining the iteration of the convergent differential neural network with a voting statistics strategy, this embodiment further reduces the classifier's training time while achieving higher classification accuracy, demonstrating superior performance in classifying the reading difficulty of English text compared to existing technologies.

[0123] The above embodiments are preferred embodiments of the present invention, but the embodiments of the present invention are not limited to the above embodiments. Any changes, modifications, substitutions, combinations, or simplifications made without departing from the spirit and principle of the present invention shall be considered equivalent substitutions and shall be included within the protection scope of the present invention.

Claims

1. A method for classifying the difficulty of English reading based on a convergent differential neural network, characterized in that, The steps include: collecting and preprocessing an English text dataset; performing feature aggregation on the preprocessed English text dataset; constructing a convergent differential neural network combining a voting strategy, and training the convergent differential neural network using the feature-aggregated English text dataset; acquiring new English text data, preprocessing it, and inputting it into the trained convergent differential neural network combining a voting strategy to obtain the corresponding English reading difficulty classification; constructing and training a convergent differential neural network combining a voting strategy, specifically as follows: building three parallel convergent differential neural networks; each convergent differential neural network consists of a fully connected input layer, a hidden layer, and an output layer; the input layer of the convergent differential neural network is... The sample vector x = ( , ,... The output of the input layer is stored in In the vector, the hidden layer f has n neurons, and the weights between the input layer and the hidden layer are... matrix The number of neurons in the output layer is The weights between the hidden layer and the output layer are matrix ； matrix Random initialization is performed and kept unchanged; the error of the output layer of the convergent differential neural network is corrected by backpropagation weights layer by layer according to the neural dynamics formula NDA; the activation functions of the three convergent differential neural networks are Sinh function, Tanh function, and Power function, respectively; based on the three convergent differential neural networks with different activation functions, a convergent differential neural network combined with a voting strategy is constructed. The dataset after feature aggregation is divided into training and test sets, and then fed into a convergent differential neural network (NDN) incorporating a voting strategy for training. The predicted classification of each sample in the test set is obtained, and the accuracy is verified. The NDN employs the NDA method, iteratively updating the weight matrix by applying the error as a time derivative. Once the error has converged to its minimum, the final weights of the convergent differential neural network are obtained. The convergent differential neural network combined with the voting strategy uses a relative majority voting method to collect the classification results of the three convergent differential neural networks and predict the final category of the sample. The relative majority voting method, specifically, uses an initialized zero vector. Used for voting statistics, and the test set is initialized to one. vector When the convergent differential neural network is determined Belongs to the When classifying reading difficulty, The process involves adding 1 to accumulate results, and finally, a vote is cast based on the majority decision principle of the three convergent difference neural networks to find the winner. The position index of the maximum value among all elements is used as the final predicted classification.

2. The English reading difficulty classification method based on a convergent differential neural network according to claim 1, characterized in that, The specific preprocessing process is as follows: the English text data is segmented, non-English text is removed, letters are uniformly lowercase, and stop words are removed; the text segmentation uses the word_tokenize tool of the Nltk segmenter; the removal of non-English text involves introducing the regularization function re to replace non-English text with whitespace; the removal of stop words involves establishing a stop word list and performing character matching to scan the segmentation dictionary to filter stop words.

3. The English reading difficulty classification method based on a convergent differential neural network according to claim 2, characterized in that, The specific process of feature aggregation on the preprocessed dataset of English text is as follows: First, feature terms representing the reading difficulty of the English text are defined, and then the English text data is processed using the min-max normalization method. Next, feature weighting is applied to the normalized dataset based on the defined feature terms.

4. The English reading difficulty classification method based on a convergent differential neural network according to claim 3, characterized in that, The features include the character-to-spelling ratio, sentence length, and passage length of the English text.

5. The English reading difficulty classification method based on a convergent differential neural network according to claim 4, characterized in that, The step of weighting the standardized dataset based on the set feature terms specifically involves a feature weighting method based on K-means clustering to aggregate the English text data for each feature term category.

Images