Intelligent classification method and system for natural language text data based on deep learning

Abstract

The application provides a natural language text data intelligent classification method and system based on deep learning, and relates to the technical field of natural language processing.The method comprises the following steps: in step 1, the real semantics of a target word is analyzed by using a context perception mechanism in view of an adversarial variant existing in the text, a pre-training process is combined with a word meaning library and a dynamic learning rate adjustment strategy, and a set of candidate replacement words with consistent semantics is generated; in step 2, multi-dimensional semantic similarity calculation and sentiment tendency discrimination are carried out based on the set of candidate replacement words, a suitable word conforming to the original cultural background is determined through a context adaptation strategy, and a standardized text sequence is generated. Through multi-dimensional semantic analysis, cultural context fusion, cross-granularity feature construction and dynamic parameter correction, the accuracy and adaptability of natural language text classification are realized.

Description

TECHNICAL FIELD

[0001] The application relates to the technical field of natural language processing, in particular to a natural language text data intelligent classification method and system based on deep learning. BACKGROUND

[0002] In the classification of daily mass text on a news information platform, traditional technologies mostly rely on artificial design features to realize text classification. When the text content involves complex semantics or domain-specific expressions, the classification accuracy may be reduced due to incomplete feature design. For example, when classifying the subfields under the sports news category, if only general sports keywords such as goals and matches are designed as features by artificial means, the traditional technology may not be able to accurately distinguish between football news containing corner kick tactics adjustment and basketball news containing pick-and-roll coordination scoring, resulting in the misclassification of some news into non-corresponding subfields.

[0003] In addition, once the parameters and classification logic of traditional text classification technology are determined, the adaptability to changes in text data distribution and new types of text is usually weak, and it is difficult to quickly adjust to meet new classification requirements. For example, when a news information platform adds an e-sports news category, if the traditional technology does not manually adjust the feature library and classification rules, it may not be able to accurately identify and classify the text containing e-sports teams, live events, hero equipment, etc. into the new category, but still misclassify it into sports news or entertainment news, resulting in a mismatch between the classification results and the actual content. SUMMARY

[0004] The technical problem to be solved by the application is to provide a natural language text data intelligent classification method and system based on deep learning, which automatically extracts deep semantic features of text and improves the classification accuracy and efficiency of complex semantic text and new types of text.

[0005] To solve the above technical problems, the technical solutions of the application are as follows:

[0006] In a first aspect, the application provides a natural language text data intelligent classification method based on deep learning, which comprises the following steps:

[0007] Step 1: For the existence of adversarial variants in the text, a context-aware mechanism is used to analyze the true semantics of the target vocabulary, and a pre-training process combining a vocabulary meaning library and a dynamic learning rate adjustment strategy is used to generate a set of candidate replacement vocabulary with consistent semantics.

[0008] Step 2: Based on the set of candidate replacement vocabulary, multi-dimensional semantic similarity calculation and sentiment orientation discrimination are performed, and a context adaptation strategy is used to determine the applicable vocabulary that conforms to the original cultural background to generate a standardized text sequence.

[0009] Step 3: Map word vectors to the standardized text sequence, and fuse local semantic information through multi-scale convolution operations to construct cross-granularity adversarial feature representations;

[0010] Step 4: Based on cross-granularity adversarial feature representation, a context-sensitive self-attention mechanism is used to obtain contextual dependencies and generate a classification feature vector carrying a measure of cultural context importance.

[0011] Step 5: Input the classification feature vector into the multilingual adaptive classification process to complete the violation type identification and generate a classification result with a credibility assessment;

[0012] Step 6: Based on the classification results, analyze and reconstruct the features of the discrimination error samples, construct an error distribution representation, generate parameter correction quantities based on the temporal correlation characteristics of the error distribution, and dynamically integrate the parameter correction quantities into the learning process using a multi-level feedback loop structure to form the final classification result.

[0013] Furthermore, for adversarial variants in the text, a context-aware mechanism is used to analyze the true semantics of the target words. Combined with a word sense database and a pre-training process using a dynamic learning rate adjustment strategy, a set of semantically consistent candidate replacement words is generated, including:

[0014] The original input text is encoded by a pre-trained language processing mechanism to obtain the contextual semantic vector representation of the target words;

[0015] Using contextual semantic vector representations as retrieval conditions, vector similarity matching is performed in the constructed word meaning resources to obtain an initial candidate word set associated with the target word;

[0016] Based on the semantic distribution characteristics of the initial candidate vocabulary set, a dynamic learning rate adjustment strategy is adopted to perform supervised fine-tuning of the pre-trained language processing mechanism in order to obtain semantic representation capabilities.

[0017] Based on semantic representation capabilities, the cosine similarity between each word in the initial candidate word set and the target word is calculated in a high-dimensional semantic space to generate a set of semantically consistent candidate replacement words.

[0018] Furthermore, based on the candidate replacement word set, multi-dimensional semantic similarity calculation and sentiment discrimination are performed. Applicable words that conform to the original cultural background are determined through a context-adaptation strategy, generating a standardized text sequence, including:

[0019] For each word in the candidate replacement word set, the similarity scores with the target word in the grammatical, semantic and pragmatic dimensions are calculated, and the similarity scores are weighted and fused according to the preset weights to generate a comprehensive semantic similarity score.

[0020] Based on the comprehensive semantic similarity score, the sentiment classifier is used to determine the sentiment tendency of candidate words, calculate the sentiment polarity probability distribution, and compare it with the sentiment polarity of the original text to select words with the same sentiment tendency and form a candidate subset with the same sentiment.

[0021] Based on the candidate subsets of emotional consistency, and combined with the cultural background and contextual features extracted from the original text, the subsets are evaluated for contextual suitability in order to obtain the contextual suitability evaluation results.

[0022] Based on the context fit assessment results, the selected applicable words replace the target words in the original text, generating a standardized text sequence that is grammatically correct, semantically coherent, and context-fitted.

[0023] Furthermore, word vector mapping is performed on the standardized text sequence, and local semantic information is fused through multi-scale convolution operations to construct cross-granularity adversarial feature representations, including:

[0024] Map each word in the standardized text sequence to a high-dimensional distributed vector and generate the corresponding word vector sequence;

[0025] The word vector sequence is input into a multi-scale convolutional neural network, and local semantic features of different spans in the text are extracted using convolutional kernels of different widths.

[0026] Local semantic features are concatenated and dimensionality reduced to obtain a unified feature representation after fusion;

[0027] Based on the unified feature representation after fusion, a cross-granularity adversarial feature representation is constructed to characterize the deep semantics and adversarial attributes of text.

[0028] Furthermore, based on cross-granularity adversarial feature representation, a context-sensitive self-attention mechanism is used to obtain contextual dependencies, generating classification feature vectors carrying a measure of cultural contextual importance, including:

[0029] The cross-granularity adversarial feature vectors are transformed into query vectors, key vectors, and value vectors respectively through a linear projection layer;

[0030] The cultural background features are linearly transformed to generate a cultural context prior vector, and then the cultural context prior vector is injected into the key vector to form a cultural context-enhanced key vector.

[0031] Based on query vectors and cultural context-enhanced key vectors, the importance weight distribution of each word in a specific cultural context is calculated through a scaling dot product attention mechanism;

[0032] The importance weight distribution is applied to the value vector, and a weighted summation is performed to generate a classification feature vector that includes semantic information and cultural context importance measurement.

[0033] Furthermore, the classification feature vector is input into the multilingual adaptive classification process to complete the violation type identification and generate classification results with credibility assessment, including:

[0034] The classification feature vectors are processed to obtain initial discrimination results for natural language text data in various violation types;

[0035] Based on the initial discrimination results, the probability distribution used to characterize the likelihood of belonging to each violation type is calculated;

[0036] The probability distribution is calibrated to generate a calibrated probability distribution and a corresponding confidence assessment.

[0037] Based on the calibrated probability distribution and the corresponding credibility assessment, classification results including violation type labels and credibility scores are determined and generated.

[0038] Furthermore, based on the classification results, the error samples are analyzed and their features reconstructed to construct an error distribution representation. Parameter correction values ​​are generated based on the temporal correlation characteristics of the error distribution, and a multi-level feedback loop structure is used to dynamically integrate these parameter correction values ​​into the learning process, forming the final classification result, including:

[0039] The discrimination error samples in the initial classification results are analyzed to identify the inherent patterns and extract the error features of the discrimination error samples at the semantic and contextual levels.

[0040] Based on error characteristics, a high-dimensional representation vector describing the error distribution pattern is constructed through a feature reconstruction process.

[0041] The correlation characteristics of the evolution of high-dimensional representation vectors over time are analyzed, and parameter correction quantities are generated using time series analysis.

[0042] Through a multi-level feedback loop structure, parameter corrections are passed layer by layer and dynamically integrated into the core learning process of classification to update internal parameters and form the final classification result.

[0043] Secondly, a deep learning-based intelligent classification system for natural language text data includes:

[0044] The semantic restoration module is used to analyze the true semantics of target words using a context-aware mechanism for adversarial variants in the text. It combines a word sense library with a pre-training process using a dynamic learning rate adjustment strategy to generate a set of semantically consistent candidate replacement words.

[0045] The text standard module is used to perform multi-dimensional semantic similarity calculation and sentiment judgment based on the candidate replacement word set, determine the applicable words that conform to the original cultural background through the context adaptation strategy, and generate a standardized text sequence.

[0046] The feature extraction module is used to map word vectors to standardized text sequences and fuse local semantic information through multi-scale convolution operations to construct cross-granularity adversarial feature representations.

[0047] The context modeling module is used to obtain contextual dependencies based on cross-granularity adversarial feature representations and employ a context-sensitive self-attention mechanism to generate classification feature vectors carrying a measure of cultural context importance.

[0048] The classification decision module is used to input the classification feature vector into the multilingual adaptive classification process, complete the violation type identification, and generate classification results with credibility assessment.

[0049] The dynamic feedback module is used to analyze and reconstruct the features of the discrimination error samples based on the classification results, construct an error distribution representation, generate parameter correction quantities based on the temporal correlation characteristics of the error distribution, and dynamically integrate the parameter correction quantities into the learning process using a multi-level feedback loop structure to form the final classification result.

[0050] Thirdly, a computing device includes:

[0051] One or more processors;

[0052] A storage device for storing one or more programs that, when executed by one or more processors, cause the one or more processors to implement the method.

[0053] Fourthly, a computer-readable storage medium storing a program that, when executed by a processor, implements the method.

[0054] The above-described solution of the present invention has at least the following beneficial effects:

[0055] By combining a context-aware mechanism with dynamic learning rate pre-training, adversarial variants in texts can be accurately identified, such as homophones and similar-looking words that violate regulations. Semantically consistent candidate replacement words are generated, effectively avoiding classification errors caused by variant interference and improving the classification accuracy of texts with adversarial features. By using multi-dimensional semantic similarity calculation and sentiment discrimination, and combining the original text's cultural background for context adaptation, the generated standardized text sequence is ensured to conform to the logic of specific cultural contexts, such as ambiguous sentences in multiple languages ​​and regionally distinctive expressions, thus solving the problem of contextual disconnect in cross-cultural and cross-language text classification.

[0056] By fusing local semantic information through multi-scale convolutional operations, a cross-granularity adversarial feature representation is constructed. Coupled with a context-sensitive self-attention mechanism, it can efficiently capture deep semantic relationships and cultural contextual importance of text, improving the classification reliability of complex texts. In the process of multilingual adaptive classification, a reliability assessment is generated by probability distribution reliability calibration, so that the classification results, such as violation type labels, have clear reliability indicators, making it easier for users to judge the usability of the classification results and reducing the risk of misjudgment and omission. Based on error sample analysis, an error distribution representation is constructed, and parameter correction is generated by combining temporal correlation characteristics. This is integrated into the learning process through a multi-level feedback loop, which can update parameters in real time. With the accumulation of data, the adaptability to new types of text and new adversarial variants is continuously improved. Attached Figure Description

[0057] Figure 1 This is a flowchart illustrating an intelligent classification method for natural language text data based on deep learning, provided in an embodiment of the present invention.

[0058] Figure 2 This is a schematic diagram of a deep learning-based intelligent classification system for natural language text data provided in an embodiment of the present invention. Detailed Implementation

[0059] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the disclosure to those skilled in the art.

[0060] like Figure 1 As shown, embodiments of the present invention propose an intelligent classification method for natural language text data based on deep learning, the method comprising the following steps:

[0061] Step 1: For adversarial variants in the text, a context-aware mechanism is used to analyze the true semantics of the target words. Combined with the pre-training process of the word sense database and dynamic learning rate adjustment strategy, a set of semantically consistent candidate replacement words is generated.

[0062] Step 2: Based on the candidate replacement word set, perform multi-dimensional semantic similarity calculation and sentiment judgment, determine applicable words that conform to the original cultural background through context adaptation strategy, and generate standardized text sequence;

[0063] Step 3: Map word vectors to the standardized text sequence, and fuse local semantic information through multi-scale convolution operations to construct cross-granularity adversarial feature representations;

[0064] Step 4: Based on cross-granularity adversarial feature representation, a context-sensitive self-attention mechanism is used to obtain contextual dependencies and generate a classification feature vector carrying a measure of cultural context importance.

[0065] Step 5: Input the classification feature vector into the multilingual adaptive classification process to complete the violation type identification and generate a classification result with a credibility assessment;

[0066] Step 6: Based on the classification results, analyze and reconstruct the features of the discrimination error samples, construct an error distribution representation, generate parameter correction quantities based on the temporal correlation characteristics of the error distribution, and dynamically integrate the parameter correction quantities into the learning process using a multi-level feedback loop structure to form the final classification result.

[0067] In this embodiment of the invention, by combining a context-aware mechanism with a dynamic pre-training strategy, the true semantics of target words can be accurately captured, generating semantically consistent candidate words. This effectively mitigates the interference of adversarial variants in the text on semantic understanding. By employing multi-dimensional semantic similarity calculation, sentiment discrimination, and context adaptation strategies, applicable words that conform to the original cultural background are selected, generating standardized text sequences. This ensures both the standardization and uniformity of text expression and avoids semantic deviations caused by cultural differences, thereby improving the cultural compatibility of text processing. Through word vector mapping and multi-scale convolution operations, local semantic information is fused to construct cross-granularity adversarial feature representations, which can comprehensively capture semantic features and adversarial information at different levels of the text, making the features more representative.

[0068] By leveraging a context-sensitive self-attention mechanism to acquire contextual dependencies, a classification feature vector carrying a measure of cultural context importance is generated. This allows the features to accurately reflect the influence of cultural context on text categories, enhancing the matching degree between classification features and task requirements. Through a multilingual adaptive classification process, violation type identification can be accurately completed, while generating credibility assessment results. This ensures that the identification results are both accurate and reliable through credibility indicators, making it easier for users to judge the reference value of the results. By analyzing and identifying error samples, constructing error distribution representations, and generating parameter correction quantities, combined with multi-level feedback loops to integrate the correction quantities into the learning process, biases can be continuously corrected, classification errors reduced, and the accuracy and stability of the final classification results continuously improved.

[0069] In a preferred embodiment of the present invention, step 1 above, which involves analyzing the true semantics of target words using a context-aware mechanism for adversarial variants in the text, and generating a semantically consistent set of candidate replacement words by combining a word sense database with a dynamic learning rate adjustment strategy during pre-training, may include:

[0070] In this embodiment of the invention, step 110, encoding the original input text using a pre-trained language processing mechanism to obtain the contextual semantic vector representation of the target word, specifically includes: First, determining the target word in the original input text, i.e., the word with adversarial variants, and extracting fixed-length text segments before and after the target word as context content. For example, if the target word is located at the 5th position in the text, words 1-4 and 6-9 can be extracted as context. Next, using the encoding of the pre-trained language processing mechanism, performing word-by-word semantic transformation on the original input text, first converting each word in the text, including the target word and context words, into a corresponding basic word vector. The dimension of the basic word vector is fixed according to the setting of the pre-training mechanism, such as 512 dimensions or 768 dimensions. Next, the attention calculation unit in the encoding calculates the association weight between the target word and each word in the context. Specifically, the base word vector of the target word is multiplied by the base word vector of each context word to obtain the initial association value. Then, all the initial association values ​​are Softmax normalized so that the sum of the normalized association weights is 1. The larger the value, the stronger the influence of the corresponding context word on the semantics of the target word. Finally, the base word vector of each context word is multiplied by the corresponding association weight to obtain the weighted context word vector. Then, all the weighted context word vectors are added to the base word vector of the target word dimension by dimension to obtain the context semantic vector representation of the target word. The dimension of this vector is consistent with the dimension of the base word vector.

[0071] Step 111: Using the context semantic vector representation as a retrieval condition, vector similarity matching is performed in the constructed word meaning resources to obtain an initial candidate word set associated with the target word. Specifically, this includes: first, determining the structure of the constructed word meaning resources, which contain a massive number of standard words and their corresponding standard semantic vectors, and all standard semantic vectors have the same dimension as the context semantic vectors obtained in step 110; using the target word context semantic vector generated in step 110 as the retrieval vector; traversing all standard semantic vectors in the word meaning resources and performing vector similarity matching calculations; for each standard semantic vector, calculating the square of the difference between its corresponding dimension value and the retrieval vector; summing the squared differences of all dimensions to obtain the sum of squares; then taking the square root of the sum of squares to obtain the Euclidean distance between the two; judging similarity based on the Euclidean distance, the smaller the Euclidean distance, the higher the similarity; setting a fixed distance threshold, such as 0.5; filtering out the standard words corresponding to the standard semantic vectors in the word meaning resources whose Euclidean distance is less than the threshold, and summing them to form an initial candidate word set associated with the target word.

[0072] Step 112: Based on the semantic distribution characteristics of the initial candidate vocabulary set, a dynamic learning rate adjustment strategy is used to supervise the fine-tuning of the pre-trained language processing mechanism to obtain semantic representation capabilities. Specifically, this includes: First, semantically labeling the initial candidate vocabulary set obtained in Step 111. The labeling content includes the semantic association degree between each candidate vocabulary and the target vocabulary, such as dividing it into three levels: strong association, medium association, and weak association. The labeling results are used as a supervision signal. Next, the adjustment rules for the dynamic learning rate are determined. The proportion of words with different semantic association levels in the initial candidate vocabulary set is statistically analyzed. If the proportion of strongly associated words is less than 30%, it indicates that the current pre-training mechanism's ability to capture strongly associated semantics is insufficient, and the learning rate is increased to 1.2 times the original learning rate. If the proportion of weakly associated words is greater than 40%, it indicates semantic confusion, and the learning rate is reduced to 0.8 times the original learning rate. If both are within a reasonable range, the learning rate remains unchanged. Then, the initial candidate vocabulary set and corresponding language processing mechanism are used for supervised fine-tuning. Using semantic annotations as training data, the pre-trained language processing mechanism is fine-tuned. Each time, a candidate word and its corresponding context are input, simulating the context of the target word in the original text. The mechanism outputs the semantic vector of the candidate word. The output semantic vector is compared with the ideal semantic vector annotated by humans (set according to the semantic association level; the ideal semantic vector of strongly associated words has a smaller difference from the semantic vector of the target word's context). The mean squared error between the two is calculated, which is the sum of the squares of the differences in the corresponding dimensions divided by the vector dimension. Based on the magnitude of the mean squared error and the dynamically adjusted learning rate, the parameters encoded in the pre-trained mechanism, such as the weight matrix of the attention calculation unit, are updated. If the mean squared error is greater than 0.3, the parameter update magnitude is the learning rate multiplied by the absolute value of the error gradient; if the mean squared error is less than 0.1, the parameter update magnitude is the learning rate multiplied by 0.5 times the error gradient. Through multiple rounds, such as 100 rounds of fine-tuning, the semantic representation capability of the pre-trained mechanism is gradually improved.

[0073] Step 113: Based on semantic representation capabilities, calculate the cosine similarity between each word in the initial candidate word set and the target word in the high-dimensional semantic space to generate a semantically consistent set of candidate replacement words. Specifically, this includes: first, utilizing the improved semantic representation capabilities from step 112, re-encoding each word in the initial candidate word set; constructing a simulated context for each candidate word, consistent with the original context structure of the target word, only replacing the target word with the candidate word; repeating the calculation process of step 110 through the fine-tuned pre-training mechanism, including basic word vector transformation, attention weight calculation, and weighted summation, to obtain the optimized semantic vector for each candidate word; simultaneously, the context semantic vector of the target word is... Step 110 results are used as the baseline vector. Next, the cosine similarity between the optimized semantic vector of each candidate word and the baseline vector is calculated in the high-dimensional semantic space. First, the product of the corresponding dimension values ​​of the two vectors is calculated, and the products of all dimensions are added together to obtain the vector dot product. Then, the magnitude of each vector is calculated separately, which is the square root of the sum of the squares of the values ​​of each dimension of each vector. Finally, the vector dot product is divided by the product of the magnitudes of the two vectors to obtain the cosine similarity value, which ranges from -1 to 1. The closer the value is to 1, the more consistent the semantics. A cosine similarity threshold is set, such as 0.8. Words with cosine similarity values ​​greater than this threshold in the initial candidate word set are filtered out and then aggregated to generate a set of semantically consistent candidate replacement words.

[0074] By encoding and calculating the context of the target vocabulary, the semantic information of the text can be transformed into a quantifiable vector form, accurately capturing the true semantic tendency of the target vocabulary in a specific context. This avoids the semantic isolation problem caused by adversarial variants and reduces semantic misjudgments caused by detachment from context. Vector similarity-based matching calculations can quickly filter out initial candidate words semantically related to the target vocabulary from massive word semantic resources, which is more efficient than manual screening. Simultaneously, through Euclidean distance quantification, it ensures that the semantic relevance between the initial candidate words and the target vocabulary is within a reasonable range, reducing the interference of irrelevant words on subsequent processes. The dynamic learning rate adjustment strategy can be adjusted based on the initial candidate words... The aggregated semantic distribution flexibly optimizes training parameters, avoiding the low training efficiency or overfitting problems caused by a fixed learning rate. Through supervised fine-tuning, the pre-training mechanism can be specifically improved to capture the semantics related to the target words, making the semantic representation of the mechanism more in line with the current text processing needs. Based on the cosine similarity calculation of semantic vectors, the semantic consistency between candidate words and target words can be more accurately measured in high-dimensional semantic space, reducing semantic bias. By setting reasonable thresholds to select words, the final generated candidate replacement word set is highly consistent with the semantics of the target words, reducing the interference of adversarial variants and providing a reliable basis for word selection to generate standardized text sequences.

[0075] In a preferred embodiment of the present invention, step 2, based on the candidate replacement vocabulary set, performs multi-dimensional semantic similarity calculation and sentiment tendency discrimination, and determines applicable vocabulary that conforms to the original cultural background through a context adaptation strategy to generate a standardized text sequence, which may include:

[0076] In an embodiment of the present invention, in step 220, for each vocabulary in the candidate replacement vocabulary set, calculate the similarity values in the grammar, semantics, and pragmatics dimensions with the target vocabulary respectively, and weightedly fuse the similarity values according to the preset weights to generate a comprehensive semantic similarity score, which specifically includes: First, for each vocabulary in the candidate replacement vocabulary set, calculate the similarity values in the grammar, semantics, and pragmatics dimensions respectively. In the grammar dimension, first clarify the词性 of the target vocabulary, such as noun, verb, adjective, etc., tense. If it is a verb, distinguish the simple present tense, simple past tense, etc., singular and plural forms. If it is a noun, judge singular or plural and collocation rules, such as which prepositions and adverbs the vocabulary often collocates with. Then compare the matching degree of the candidate vocabulary with the target vocabulary in these grammatical attributes. For example, if the target vocabulary is the verb "run" (simple present tense, original form), the candidate vocabulary "ran" (simple past tense) does not match in the tense attribute, and the grammar dimension similarity value will decrease; while the candidate vocabulary "operate" (also a verb and with similar collocation rules) has a high matching degree in the grammatical attributes, and the grammar dimension similarity value will increase. The calculation of the grammar dimension similarity value will refer to a large number of vocabulary corpora with annotated grammatical attributes, and determine by counting the proportion of the matching times of the candidate vocabulary and the target vocabulary in each grammatical attribute to the total number of grammatical attributes. For example, if there are 5 total grammatical attributes and the candidate vocabulary and the target vocabulary match 4 items, then the initial grammar dimension similarity value is 4 / 5 = 0.8, and then fine-tune according to the importance of different grammatical attributes. For example, the importance of the词性 attribute is higher than the tense attribute. If the词性 matches but the tense does not match, the fine-tuning range will be smaller, and finally determine the grammar dimension similarity value.

[0077] It should be noted that there is an unclear "词性" in the original text which might be a typo. You may want to double-check and correct it if necessary.In the semantic dimension, a semantic vector space mechanism for target and candidate words is first constructed. Words are mapped into the semantic space to form vectors. The construction of semantic vectors is based on a large-scale text corpus. The values ​​of each dimension of the vector are determined by calculating the co-occurrence frequency of words with other words in the text, as well as the semantic relevance of words, such as synonyms, near-synonyms, and antonyms. For example, the target word "happy" and the candidate word "joyful" have similar co-occurrence frequencies with words like "happy" and "happy" in a large amount of text, and their semantic relevance is high. The numerical difference between their semantic vectors in each dimension... The differences between the target word and the candidate word "happiness" are relatively small; however, the co-occurrence frequency and semantic relevance of the candidate words "sadness" and "happiness" are significantly different, as are the numerical differences in the semantic vector dimensions. Next, the cosine similarity between the semantic vectors of the target word and the candidate word is calculated. The cosine similarity calculation process is to first calculate the dot product of the two vectors, that is, multiply the corresponding dimension values ​​and then sum them, then calculate the magnitude of the two vectors, the square root of the sum of the squares of the values ​​of each dimension, and finally divide the dot product result by the product of the magnitudes of the two vectors. The result is the semantic similarity value. The closer the value is to 1, the higher the semantic similarity between the two words.

[0078] In the pragmatic dimension, the applicable scope, usage frequency, and pragmatic functions of the target word and candidate words in actual usage scenarios will be analyzed, such as whether they are used in formal or informal occasions, in written or spoken language, and whether they have specific pragmatic intentions such as requests, suggestions, statements, etc. For example, the target word "商榷" is often used in formal written language and business communication occasions, and its pragmatic function is to put forward opinions for discussion; the candidate word "商量" is often used in informal daily communication occasions, and its pragmatic function is also discussion, but the applicable scenarios and formality levels are different. If the usage scenario of the original text is formal business communication, then the similarity value of "商量" and "商榷" in the pragmatic dimension will be relatively low; while the candidate word "商议" is also applicable to formal occasions and has a similar pragmatic function, so the similarity value in the pragmatic dimension will be relatively high. The calculation of the similarity value in the pragmatic dimension will be based on collecting a large amount of text corpora in different usage scenarios, and statistically analyzing the proportion of the occurrence frequencies of the candidate word and the target word in the same usage scenario, as well as the matching degree in pragmatic functions. For example, in 100 scenarios where the target word is applicable, the candidate word appears 70 times, and the pragmatic functions match 60 times. First, calculate the proportion of the scene matching frequency as 70 / 100 = 0.7, and the proportion of the pragmatic function matching as 60 / 100 = 0.6. Then, sum them up with a weight ratio of 4:6 to get the similarity value in the pragmatic dimension as 0.7×0.4 + 0.6×0.6 = 0.64. After calculating the similarity values in the three dimensions, according to the preset weights, such as the grammar dimension weight of 0.3, the semantic dimension weight of 0.5, and the pragmatic dimension weight of 0.2, the similarity values in the three dimensions are weighted and fused. For example, for a certain candidate word, the similarity value in the grammar dimension is 0.8, the semantic dimension is 0.9, and the pragmatic dimension is 0.64. Then the comprehensive semantic similarity score = 0.8×0.3 + 0.9×0.5 + 0.64×0.2 = 0.24 + 0.45 + 0.128 = 0.818.

[0079] Step 221: Based on the comprehensive semantic similarity score, use a sentiment classifier to determine the sentiment tendency of the candidate words, calculate the sentiment polarity probability distribution, and compare it with the sentiment polarity of the original text to screen out the words with consistent sentiment tendencies, forming a candidate subset with consistent sentiment. Specifically, it includes: Based on the comprehensive semantic similarity score obtained in Step 220, first conduct a preliminary screening of the candidate words. Usually, a comprehensive semantic similarity score threshold is set, such as 0.6. Exclude the candidate words with scores lower than the threshold, and only retain the candidate words with scores higher than or equal to the threshold to enter the sentiment tendency determination link; then, use the sentiment classifier to determine the sentiment tendency of the screened candidate words. The sentiment classifier will first perform text preprocessing on the candidate words, including removing irrelevant characters in the words, such as special symbols, punctuation marks, etc., and converting the words into a unified case form. Then, extract the sentiment features of the candidate words, such as sentiment affixes contained in the words, and the semantic association with known sentiment words. For example, words with a high semantic association with the positive sentiment word "wonderful" probably have a positive sentiment tendency. The sentiment classifier will calculate the probability distribution of the candidate words belonging to positive sentiment, negative sentiment, and neutral sentiment, that is, the sentiment polarity probability distribution, based on these sentiment features and the trained sentiment classification framework. For example, after calculation, the probability of a certain candidate word belonging to positive sentiment is 0.8, the probability of negative sentiment is 0.1, and the probability of neutral sentiment is 0.1. Then its sentiment polarity probability distribution is (positive: 0.8, negative: 0.1, neutral: 0.1).

[0080] Calculate the sentiment polarity of the original text. The calculation of the sentiment polarity of the original text will first perform sentence splitting and word segmentation on the original text, remove the stop words in the text, such as words without actual sentiment meaning like "de", "le", "zai", etc., then extract the sentiment words in the text, count the number of positive sentiment words, negative sentiment words, and neutral sentiment words, and then combine the sentiment intensities of different sentiment words. For example, the positive sentiment intensity of "ecstatic" is higher than "happy", and calculate the overall sentiment score of the original text. If the positive sentiment score is higher than the negative sentiment score and the difference exceeds the preset threshold, such as 0.3, then the sentiment polarity of the original text is positive; if the negative sentiment score is higher than the positive sentiment score and the difference exceeds the preset threshold, then the sentiment polarity of the original text is negative; if the difference between the two scores is less than or equal to the preset threshold, or the proportion of neutral sentiment words in the total number of words exceeds 70%, then the sentiment polarity of the original text is neutral.

[0081] The sentiment polarity probability distribution of candidate words is compared with the sentiment polarity of the original text. If the sentiment polarity of the original text is positive, candidate words with the highest positive sentiment probability and above the preset positive probability threshold are selected. If the sentiment polarity of the original text is negative, candidate words with the highest negative sentiment probability and above the preset negative probability threshold are selected. If the sentiment polarity of the original text is neutral, candidate words with the highest neutral sentiment probability and above the preset neutral probability threshold are selected. These selected words with sentiment tendencies consistent with the original text constitute a candidate subset with consistent sentiment.

[0082] Step 222: Based on the candidate subsets of sentiment consistency, and incorporating the cultural background and contextual features extracted from the original text, a context fit assessment is performed on the subsets to obtain the context fit assessment results. Specifically, this includes: based on the candidate subsets of sentiment consistency, firstly, extracting cultural background and contextual features from the original text. The extraction of cultural background involves analyzing the cultural elements involved in the original text, such as historical events, cultural customs, religious beliefs, and regional characteristics. For example, if the original text mentions pasting Spring Festival couplets during the Spring Festival, the extracted cultural background is traditional Chinese Spring Festival culture; if it mentions hanging Christmas trees during Christmas, the cultural background is Western Christmas. The extraction of cultural and contextual features includes the text's contextual information, such as the words before and after the target words, sentence structure, the text's theme (e.g., science and technology, literature, news), the text's target audience (e.g., professionals, the general public, children), and the text's purpose (e.g., introducing knowledge, expressing opinions, telling stories). For example, if the original text is a scientific paper about artificial intelligence, with the target word being "algorithm" and the words before and after it being "deep learning" and "data training," then the contextual information would be: deep learning data training, the text's theme is science and technology, the target audience is professional researchers, and the purpose is to introduce the technical principles.

[0083] Combining extracted cultural background and contextual features, the contextual suitability of candidate words within the subset is evaluated. In terms of cultural background suitability, it is determined whether the candidate words conform to the cultural background of the original text. If the candidate words are related to the cultural background and there is no cultural conflict, the cultural background suitability score is high; if the candidate words are unrelated to the cultural background or there is a cultural conflict, such as the candidate word "Christmas gift" in a text involving traditional Chinese New Year culture, the cultural background suitability score is low. The calculation of the cultural background suitability score will refer to the vocabulary corpus related to the cultural background, count the frequency of the candidate words in the corresponding cultural background text, and the evaluation scores of cultural experts on the suitability of the candidate words to the cultural background. The two are weighted and summed in a 6:4 ratio to obtain the cultural background suitability score, with a maximum score of 1 point.

[0084] In assessing contextual feature suitability, the first step is to determine whether candidate words fit the contextual information of the text. This means assessing whether the sentence is fluent and semantically coherent after the candidate word is inserted into the sentence containing the target word. For example, if the sentence containing the target word is "deep learning data training," inserting the candidate word "optimize" results in the sentence "deep learning optimizes data training," which is semantically coherent and fluent, resulting in a high contextual suitability score. Conversely, inserting the candidate word "singing" results in a semantically incoherent sentence and a low contextual suitability score. Next, the assessment considers whether candidate words fit the text's theme. For instance, in a science and technology text, the candidate word "programming" fits the theme and has a high suitability score, while the candidate word "poetry" does not fit the theme and has a low suitability score. Finally, the assessment also considers whether candidate words fit the communicative context. For example, in children's story texts, the purpose of expression is entertainment and enlightenment. Candidate words like "cute" and "fun" meet the requirements and have high fit scores, while candidate words like "complex" and "abstract" do not meet the requirements and have low fit scores. The context feature fit score is obtained by weighting and summing context fit, theme fit, communicator fit, and expression purpose fit according to a weight of 4:3:1.5:1.5. Each fit score is obtained through corresponding evaluation criteria and corpus statistics, with a maximum score of 1 point. The cultural background fit score and the context feature fit score are weighted and summed according to a weight of 5:5 to obtain the context fit evaluation result of the candidate words, with a maximum score of 1 point. The higher the score, the better the context fit of the candidate words.

[0085] Step 223: Based on the context fit assessment results, replace the target words in the original text with the selected applicable words to generate a standardized text sequence that is grammatically correct, semantically coherent, and context-fitted. Specifically, this includes: based on the context fit assessment results obtained in step 222, setting a context fit assessment result threshold, such as 0.7, and selecting candidate words with assessment results higher than or equal to the threshold as applicable words. If there are multiple applicable words, their comprehensive semantic similarity scores will be compared, and the applicable word with the highest comprehensive semantic similarity score will be selected. If the comprehensive semantic similarity scores are the same, their sentiment polarity probabilities will be compared with the degree of matching of the sentiment polarity of the original text, and the applicable word with the highest degree of matching will be selected.

[0086] After identifying the appropriate vocabulary, the target vocabulary in the original text is replaced with it. After the replacement, the generated new text is grammatically checked to ensure that the subject-verb-object structure of the sentences is complete, the tenses are consistent, and the parts of speech are appropriate. If grammatical errors are found, such as subject-verb disagreement or tense confusion, the appropriate vocabulary is reselected and replaced until the grammar is correct. Next, a semantic coherence check is performed. The new text is read to determine whether the semantics between sentences and paragraphs are coherent and whether there are semantic gaps or contradictions. If semantic problems are found, the cause is analyzed to see if it is due to vocabulary replacement. If so, the semantic similarity and contextual suitability of the candidate vocabulary are re-evaluated, and a more suitable vocabulary is selected for replacement. Finally, the new text is reconfirmed to ensure that it conforms to the original cultural background and contextual characteristics, and that there are no cultural conflicts or contextual inconsistencies. Finally, a standardized text sequence that is grammatically correct, semantically coherent, and context-appropriate is generated.

[0087] By calculating similarity scores across grammatical, semantic, and pragmatic dimensions, and then weighting and fusing these scores with preset weights to generate a comprehensive semantic similarity score, this approach comprehensively and multidimensionally assesses the matching degree between candidate replacement words and target words. This avoids the limitations of single-dimensional evaluation. Preliminary screening based on the comprehensive semantic similarity score reduces the computational load of the sentiment tendency determination stage, saves computational resources, and improves processing efficiency. A sentiment classifier is used to calculate the sentiment polarity probability distribution of candidate words and compare it with the sentiment polarity of the original text. Words with consistent sentiment tendencies are selected to form a candidate subset, ensuring that the replaced words maintain a high degree of consistency with the original text in terms of sentiment expression. This avoids deviations or misunderstandings in the overall emotional communication of the text due to inconsistent sentiment tendencies. Cultural background and contextual features are extracted from the original text, and these features are used to evaluate the contextual adaptability of candidate words within the subset. This fully considers the cultural attributes and specific usage scenarios of the text, ensuring that the selected candidate words not only match the original text semantically and emotionally but also are highly adaptable to the original text in terms of cultural background and context.

[0088] In a preferred embodiment of the present invention, step 3 above, which involves mapping word vectors to the standardized text sequence and fusing local semantic information through multi-scale convolution operations to construct a cross-granularity adversarial feature representation, may include:

[0089] In this embodiment of the invention, step 330 involves mapping each word in the standardized text sequence to a high-order word.

[0090] This process involves generating a high-dimensional distributed vector and corresponding word vector sequences. Specifically, it includes: First, determining the basic information of each word in the standardized text sequence, including its spelling, position in the text, and frequency. Next, matching each word with a pre-defined word index table, which contains a large number of common words and their corresponding unique numerical identifiers (e.g., "technology" corresponds to identifier 156, "intelligence" to identifier 289). Each word is converted into a unique numerical identifier through index matching. Then, based on pre-trained word embedding rules, each numerical identifier is mapped to a high-dimensional distributed vector. The vector dimension is typically set to 100-300 dimensions, and the value of each dimension is calculated through the contextual association of the word in a large-scale corpus. The process begins by statistically analyzing the relationship between the word and other words in the text. The co-occurrence frequency of other words in the corpus is calculated, such as artificial intelligence often co-occurring with algorithms and data. Then, the weight of each dimension is calculated based on the co-occurrence frequency. The higher the co-occurrence frequency of related words, the closer the corresponding dimension values ​​are. At the same time, the semantic attributes of the words are combined, such as the relationship between synonyms and antonyms, to fine-tune the values. For example, computer and computer are semantically similar, and the difference in the values ​​of each dimension of their vectors is controlled within 0.1. Finally, according to the order of the words in the standardized text sequence, the high-dimensional distributed vectors corresponding to each word are arranged in sequence to generate a word vector sequence with the same length as the original text. For example, the vector sequence corresponding to the text "Intelligent devices improve efficiency" is [vector 1 (intelligent), vector 2 (device), vector 3 (improvement), vector 4 (efficiency)].

[0091] Step 331: Input the word vector sequence into a multi-scale convolutional neural network. Use convolutional kernels of different widths to extract local semantic features across different ranges in the text. Specifically, after inputting the word vector sequence generated in step 330 into the multi-scale convolutional neural network, first determine 3-5 different widths of convolutional kernels. Common widths are 2, 3, and 4 (corresponding to covering 2, 3, and 4 consecutive word vectors, respectively). Taking a convolutional kernel with a width of 2 as an example, during calculation, first fix the initial weights of the convolutional kernel (these weights are preset based on the previous training data, and the weights for each dimension are different). Starting from the first word vector in the word vector sequence, multiply the convolutional kernel by the corresponding dimension values ​​of the first two consecutive word vectors. For example, if the convolutional kernel dimension weights are [0.2, 0.5, 0.3, ...] (a total of 300 dimensions), the first... If the first word vector has dimensions [0.8, 0.1, 0.6, ...], and the second word vector has dimensions [0.7, 0.3, 0.4, ...], then the corresponding dimension products are 0.2 × 0.8 = 0.16, 0.5 × 0.1 = 0.05, 0.3 × 0.6 = 0.18, and so on, multiplying all 300 dimensions. Then, sum the products of all dimensions and add a preset bias value, usually 0.1-0.3, to obtain a single feature value representing the local semantic features of the two consecutive word vectors. Next, the convolution kernel moves one word vector position to the right, and the above calculation is repeated for the 2nd and 3rd word vectors until the entire word vector sequence is traversed, generating a set of local semantic feature vectors containing multiple feature values ​​(length = original word vector sequence length - convolution kernel width + ...). 1) Similarly, a convolutional kernel with a width of 3 covers 3 consecutive word vectors and generates another set of local semantic feature vectors according to the same calculation logic, capturing the semantics of a phrase composed of 3 words; a convolutional kernel with a width of 4 captures the semantics of a phrase composed of 4 words, and finally obtains multiple sets of local semantic feature vectors with different spans.

[0092] Step 332 involves concatenating and reducing the dimensionality of the local semantic features to obtain a unified feature representation after fusion. Specifically, this includes concatenating the multiple sets of local semantic feature vectors obtained in step 331. Assuming a convolutional kernel with a width of 2 generates a vector with 80 feature values, a kernel with a width of 3 generates a vector with 70 feature values, and a kernel with a width of 4 generates a vector with 60 feature values, the feature values ​​of the three sets of vectors are arranged sequentially according to the convolutional kernel width from smallest to largest, forming a fused feature vector containing 80 + 70 + 60 = 210 feature values. Next, dimensionality reduction is performed using principal component analysis. The first step is to calculate the mean of all feature values ​​in the fused feature vector, and then subtract the mean from each feature value. The first step involves obtaining the centered eigenvalues. The second step is to calculate the covariance matrix of the centered eigenvalues. By statistically analyzing the difference between the mean of the products of any two eigenvalues ​​and the product of the means, the correlation between the features is determined. For example, if the covariance between feature 1 and feature 2 is positive, it indicates that their trends are consistent. The third step is to solve for the eigenvalues ​​and eigenvectors of the covariance matrix. The eigenvalues ​​are sorted from largest to smallest, and the eigenvectors corresponding to the N largest eigenvalues ​​are selected as principal components. The fourth step is to multiply the fused feature vector with these N principal component vectors. After multiplying the corresponding elements of each principal component vector with the fused feature vector, the results are summed to obtain N new eigenvalues, forming a unified feature representation of dimension N. This achieves dimensionality reduction while preserving core semantic information.

[0093] Step 333: Based on the fused unified feature representation, construct a cross-granularity adversarial feature representation to characterize the deep semantics and adversarial attributes of the text. Specifically, this includes: Based on the unified feature representation obtained in step 332, firstly, extract deep semantic features, inputting the unified feature representation into a multi-layer numerical transformation process. In the first layer transformation, first determine the dimensions of the unified feature, such as 100 dimensions, and the dimensions of the preset weight set for that layer, such as a 100-row, 120-column numerical matrix. Then, multiply the specific value of each dimension of the unified feature by the 120 weight values ​​of the corresponding row of the weight set one by one; for example, if the value of the first dimension of the unified feature is 0.8, the weight... If the value in the first row and first column of the reassembled set is 0.2 and the value in the first row and second column is 0.3, then the corresponding calculations are 0.8 × 0.2 = 0.16 and 0.8 × 0.3 = 0.24. This completes the multiplication operation of all dimensions and their corresponding weights. Next, the results of all multiplication operations are summed, and a preset fixed bias value, such as 0.2, is added to the sum to form the preliminary transformation result. Then, the preliminary transformation result is processed according to the preset value filtering rules. If a value is less than 0, it is adjusted to 0; if the value is greater than or equal to 0, the original value is kept unchanged. Through this filtering process, a 120-dimensional intermediate feature is finally obtained.

[0094] In the second layer of transformation, the above calculation logic is repeated. Based on the 120-dimensional intermediate features, the corresponding weight sets of each dimension are matched, such as a 120-row, 80-column numerical matrix. Multiplication, summation, and bias addition are performed on the numerical values ​​and weights of each dimension. Then, the same numerical filtering rules are applied to transform the 120-dimensional intermediate features into 80-dimensional features. Following this process, after 3-4 layers of similar numerical transformations and filtering, the final features reflect the overall meaning of the text, including its core theme and key viewpoints—that is, deep semantic features. While extracting deep semantic features, adversarial attribute features are also developed. The construction process begins by adding a small adjustment to each dimension of the unified feature representation. The size of this adjustment is determined as follows: First, all dimension values ​​of the unified feature are input into a preset calculation process to obtain a set of prediction results. These prediction results are then compared with the actual results corresponding to the text, and the difference between the two is calculated. Next, for each feature dimension, the rate of change of the difference value with the change of that dimension value is calculated. The larger the absolute value of the rate of change, the more significant the impact of that dimension value on the final prediction result, and the larger the corresponding adjustment amount. Typically, the adjustment amount is 0.01-0.05 times the rate of change value.

[0095] The adjusted feature values ​​are input into the above calculation process to obtain new prediction results. The prediction results corresponding to the original features and the adjusted features are compared, and the difference between the two is statistically analyzed. From all feature dimensions, the top 20-30 dimensions with the largest differences are selected. The values ​​of these dimensions can reflect the sensitivity of the features to external interference and constitute adversarial attribute features. The deep semantic features and adversarial attribute features are concatenated in dimensional order. The 80 dimensional values ​​of the deep semantic features are arranged sequentially, and then the 30 dimensional values ​​of the adversarial attribute features are arranged immediately afterward to form a feature representation with a total dimension of 110. This feature representation contains both coarse-grained information reflecting the overall semantics of the text and fine-grained information reflecting local adversarial attributes, i.e., cross-granularity adversarial feature representation.

[0096] By accurately characterizing the semantics of words through high-dimensional distributed vectors, words with similar meanings are placed closer together in the vector space, avoiding semantic misjudgments caused by superficial differences in words. The word vector sequence strictly follows the temporal order of the original text, ensuring that the contextual relationship of words is preserved during feature extraction, laying the foundation for accurately capturing the semantic logic of the text. Convolutional kernels of different widths cover word combinations of different spans, which can capture the local semantics of two-word combinations such as smart devices, as well as the semantic information of multi-word phrases such as artificial intelligence technology. This avoids missing key semantic segments by single-scale convolution, comprehensively mining the local semantic details of the text, improving the richness of features, and concatenating multi-scale local features to integrate semantic information of different spans, making the features more comprehensive.

[0097] In a preferred embodiment of the present invention, step 4 above, based on cross-granularity adversarial feature representation, uses a context-sensitive self-attention mechanism to obtain contextual dependencies and generate a classification feature vector carrying a measure of cultural context importance, may include:

[0098] In this embodiment of the invention, step 440, which converts the cross-granularity adversarial feature vector into a query vector, a key vector, and a value vector respectively through a linear projection layer, specifically includes: First, determining the dimension of the cross-granularity adversarial feature vector; preparing three independent numerical transformation tables for linear projection; each transformation table having the same number of rows as the feature vector dimension and the same number of columns as the target vector dimension; the value at each position in the table being pre-set; and multiplying each dimension value of the cross-granularity adversarial feature vector by the 64 values ​​in the corresponding row of the first transformation table; for example, if the value of the 5th dimension of the feature vector is 0.6, the values ​​in the 5th row of the transformation table are multiplied sequentially... If the values ​​are 0.1, 0.3, ..., 0.2, then calculate 0.6 × 0.1 = 0.06, 0.6 × 0.3 = 0.18, ..., 0.6 × 0.2 = 0.12. Add these 64 product results together, and add a fixed base value to obtain the value of the first dimension of the query vector. Process all 110 dimensions of the feature vector in the same way to finally generate a 64-dimensional query vector. Perform the same calculation with the feature vector and the second transformation table to obtain a 64-dimensional key vector; calculate with the third transformation table to obtain a 64-dimensional value vector. The three vectors have the same dimensions but different values, and each plays a different computational role.

[0099] Step 441 involves linearly transforming the cultural background features to generate a cultural context prior vector, and then injecting this prior vector into the key vector to form a cultural context-enhanced key vector. Specifically, the cultural background features are cultural attribute values ​​extracted from the original text. A 50-row, 64-column numerical transformation table is prepared for this purpose. Each dimension value of the cultural background features is multiplied by the corresponding 64 values ​​in the transformation table. All products are summed and a bias value is added to convert the result into a 64-dimensional cultural context prior vector. This prior vector is then fused with the key vector generated in step 440 by directly adding the values ​​of corresponding dimensions of the two vectors. For example, if the 20th dimension of the prior vector is 0.4 and the 20th dimension of the key vector is 0.5, then the fused value for that dimension is 0.4 + 0.5 = 0.9. This method yields a cultural context-enhanced key vector, allowing the key vector to carry cultural background information.

[0100] Step 442: Based on the query vector and the cultural context enhancement key vector, calculate the importance weight distribution of each word in a specific cultural context using a scaled dot product attention mechanism. Specifically, for each word, calculate the dot product of its query vector and the cultural context enhancement key vector. Multiply each of the 64 dimensions of the query vector by the 64 dimensions of the enhancement key vector, and then sum all the product results. For example, if the query vector is [0.2, 0.3, ..., 0.1] and the enhancement key vector is [0.5, 0.4, ..., 0.2], then the dot product is 0.2 × 0.5 + 0.3 × 0.4 + ... Add 0.1 × 0.2 to obtain a total value. Scale this sum by dividing it by the square root of the vector dimension to get the scaled value. Perform the same calculation on all words in the text to obtain a set of scaled values. Normalize this set of scaled values ​​by first calculating the exponent of each value, and then dividing each exponent by the sum of all exponents. For example, if the scaled values ​​of three words are 2, 1, and 3, their exponents are approximately 7.389, 2.718, and 20.085, with a sum of 30.192. The weight of the first word is 7.389 ÷ 30.192 ≈ 0.245. This gives the importance weight of each word. All weights are the weight distribution in a specific cultural context; the larger the weight, the more important the word.

[0101] Step 443 involves applying the importance weight distribution to the value vector, performing weighted summation and fusion to generate a classification feature vector that includes semantic information and cultural context importance measurement. Specifically, this includes multiplying the importance weight of each word obtained in step 442 by the values ​​of each dimension of that word's value vector. For example, if a word has a weight of 0.3 and its value in the 10th dimension is 0.6, then 0.3 × 0.6 = 0.18 is calculated. For the same dimension of all words, the weighted values ​​are summed. For example, the weighted results for all words in the 10th dimension are 0.18, 0.25, and 0.12, respectively, with a sum of 0.55. This sum is the value of the 10th dimension of the classification feature vector. By processing all 64 dimensions in this way, the final generated vector contains both the semantic information of the text and, through weights, reflects the importance of different words in the cultural context, i.e., a classification feature vector carrying a cultural context importance measurement.

[0102] Three specialized vectors are generated through linear projection, each with its own function, accurately capturing the relationships between features and laying the foundation for effectively extracting contextual dependencies. This ensures the targeted and accurate nature of the computation. Cultural background information is integrated into the key vectors, giving them cultural context awareness capabilities. This ensures that attention calculations prioritize culturally relevant content, preventing cultural information from being ignored and enhancing the sensitivity and adaptability of features to cultural context. Scaling the dot product calculations reasonably measures the strength of associations between words, and normalization processing makes the weight distribution more reasonable, accurately quantifying the importance of each word in a specific cultural context. This helps highlight key information and improves the ability to identify the core content of cultural context. The weighted fusion process organically combines semantic information with the importance weights of cultural context. The generated classification feature vectors not only fully preserve the semantics of the text but also clearly reflect the influence of cultural factors, improving the accuracy and reliability of task processing.

[0103] In a preferred embodiment of the present invention, step 5 above, which involves inputting the classification feature vector into a multilingual adaptive classification process to complete violation type identification and generate a classification result with credibility assessment, may include:

[0104] In this embodiment of the invention, step 550 involves classifying the classification feature vector to obtain initial discrimination results for natural language text data across multiple violation types. Specifically, this includes: first, determining the violation type categories to be identified, such as false advertising, information leakage, copyright infringement, and inappropriate speech, with a total of 6 categories. Each violation type is assigned a unique numerical identifier, such as 1 for false advertising, 2 for information leakage, and so on, with 6 for inappropriate speech. Next, the classification feature vector is input into a preset classification calculation process. This process includes calculation branches corresponding to the number of violation types. In each branch, the value of each dimension of the classification feature vector is multiplied by the preset weight value of that branch. Each dimension corresponds to one weight, and a total of 64 weights are multiplied one by one. For example, the value of the 8th dimension of the feature vector is... If the weight of the 8th dimension of a certain branch is 0.7, then calculate 0.7 × 0.3 = 0.21. Next, sum the products of all dimensions to obtain the branch sum, and add a fixed bias value to generate the discrimination score for that branch. After calculating for each of the six branches, six discrimination scores are obtained, each corresponding to a violation type. Finally, compare the discrimination score of each branch with a preset branch threshold. If the score is higher than the threshold, the text is preliminarily determined to likely belong to that violation type and marked as a candidate violation type; if the score is lower than the threshold, the violation type is excluded and marked as a non-candidate violation type. All the marking results together constitute the initial discrimination result, for example, "Candidate violation types: false advertising, copyright infringement; Non-candidate violation types: information leakage, inappropriate speech...".

[0105] Step 551: Based on the initial discrimination results, calculate the probability distribution used to characterize the likelihood of belonging to each violation type. Specifically, this includes: based on the six violation type discrimination scores obtained in step 550, first preprocess all scores by subtracting the minimum value among all scores for each score to eliminate the influence of negative scores and ensure that all scores are non-negative. Then, calculate the sum of all preprocessed scores. For example, if the preprocessed scores are 0.8, 0.3, 1.2, 0.1, 0.9, and 0.5, the sum is 3.8. Then, divide each preprocessed score by this sum to obtain the probability distribution for each violation type. The probability values ​​for each violation type are generated, such as the first score 0.8 ÷ 3.8 ≈ 0.21, the second score 0.3 ÷ 3.8 ≈ 0.08, and so on, generating 6 probability values. The sum of these probability values ​​is 1, and the magnitude of each probability value directly reflects the likelihood that the text belongs to the corresponding violation type. The larger the probability value, the higher the likelihood. The 6 probability values ​​are arranged in order of violation type to form a probability distribution used to represent the likelihood of each violation type, for example, "False advertising: 0.08, Information leakage: 0.05, Copyright infringement: 0.21, Inappropriate remarks: 0.32...".

[0106] Step 552 involves performing confidence calibration on the probability distribution to generate a calibrated probability distribution and its corresponding confidence assessment. This includes: first, collecting a large number of text samples labeled with actual violation types; each sample is processed through steps 550-551 to obtain the corresponding probability distribution and actual violation type label; based on these sample data, establishing a calibration mapping relationship; and statistically analyzing the correct prediction probability corresponding to different probability values ​​for each violation type. For example, in the copyright infringement category, there are 100 samples with probability values ​​between 0.2 and 0.3, of which 85 actually constitute copyright infringement. Therefore, the correct prediction probability for this probability range is 85 ÷ 100 = 0.85.

[0107] For the probability distribution of the current text, the probability value of each violation type is processed one by one. If the probability value of inappropriate speech is 0.32, and the correct prediction probability corresponding to the probability interval of 0.3-0.4 under the inappropriate speech category in the calibration mapping relationship is 0.88, then the original probability value of 0.32 is calibrated to 0.88. The probability values ​​of other violation types are processed in the same way to obtain the calibrated probability distribution. At the same time, the credibility assessment index is calculated. First, the maximum value in the calibrated probability distribution is found, and then the difference between the maximum value and the second largest value is calculated. The larger the difference, the higher the discrimination of the classification result. Combined with the overall variance of the calibrated probability distribution, the smaller the variance, the more concentrated the probability distribution, and the clearer the classification, the credibility assessment level is comprehensively determined, such as high credibility, medium credibility, low credibility, or a specific credibility score. Among them, the larger the maximum value, the larger the difference, and the smaller the variance, the higher the credibility assessment.

[0108] Step 553: Based on the calibrated probability distribution and corresponding credibility assessment, determine and generate a classification result including violation type labels and credibility scores. Specifically, this includes: extracting the violation type with the highest probability value from the calibrated probability distribution and using its corresponding type name as the violation type label. For example, the probability value of inappropriate speech after calibration is 0.88, which is the highest, so the label is "Violation Type: Inappropriate Speech". Next, convert the credibility assessment generated in step 552 into specific credibility scores: if the assessment is "high credibility", the corresponding score range is 80-100 points, further refined by combining the maximum value and the difference, such as a maximum value of 0.88 and a difference of 0.23, corresponding to a score of 88 points; if it is "medium credibility", the score range is 50-79 points; and "low credibility" is 0-49 points. Finally, integrate the violation type labels and credibility scores to generate a complete classification result, in the format "Classification Result: Violation Type - Inappropriate Speech; Credibility Score - 88 points". If the calibrated probability values ​​of all violation types are lower than the preset minimum threshold, they are marked as no violation type, and the corresponding credibility score is given.

[0109] By employing multi-branch classification calculations and threshold comparisons, potential violation types are quickly screened out, excluding obviously irrelevant categories. This reduces the workload of probability calculations and calibration, while providing a clear candidate range for initial classification, avoiding omissions of possible violation types and improving the comprehensiveness of violation identification. The discrimination scores are transformed into probability distributions, using intuitive numerical values ​​to quantify the likelihood of text belonging to each violation type, making the classification results more interpretable and ensuring a clear relative relationship between the probabilities of each violation type. This provides a basis for distinguishing between primary and secondary violation types. Confidence calibration corrects the deviation between the original probability values ​​and the actual correct probabilities, making the probability distribution more closely reflect reality and improving the accuracy of the classification results. Credibility assessment quantifies the reliability of the classification results through multi-dimensional indicators, avoiding misleading decisions due to a single probability value and enhancing the practicality of the results. Clear violation type labels directly provide core classification conclusions, meeting the need for rapid identification of violation categories.

[0110] In a preferred embodiment of the present invention, step 6 above, which involves analyzing and reconstructing the features of the discrimination error samples based on the classification results, constructing an error distribution representation, generating parameter correction quantities based on the temporal correlation characteristics of the error distribution, and dynamically integrating the parameter correction quantities into the learning process using a multi-level feedback loop structure to form the final classification result, may include:

[0111] In this embodiment of the invention, step 660 involves analyzing the discrimination error samples in the initial classification results, identifying their inherent patterns, and extracting the error features of the discrimination error samples at the semantic and contextual levels. Specifically, this includes: first, screening discrimination error samples from the initial classification results; comparing the classification results with the manually labeled true results; if the violation type label is inconsistent, or the confidence score deviates from the manually reviewed result by more than a preset threshold, then marking them as discrimination error samples; identifying the inherent patterns of the error samples; statistically analyzing the common features of the error samples; and determining the patterns through counting statistics. For example, in 100 error samples, 60 belong to... For product review texts containing 45 vague expressions, vague expressions are the main inherent pattern of product review topics. Error features at the semantic and contextual levels are extracted. Semantically, the differences between the semantic annotations of core words in the error samples and their true meanings are compared. Contextually, the logical flow of key sentences in the error samples is analyzed. For example, if the first half of a text explains legal use, and the second half mentions borrowed content, ignoring the preceding logic may lead to misjudgment. Features of missing contextual logic are extracted. The frequency of related words in key sentences is statistically analyzed to quantify the features. Finally, all quantified error features are integrated to form an error feature set.

[0112] Step 661: Based on the error features, a high-dimensional representation vector describing the error distribution pattern is constructed through a feature reconstruction process. Specifically, this includes: based on the error feature set extracted in step 660, first determining the feature dimensions. If four types of error features are extracted—semantic annotation bias, contextual logic omission, topic matching error, and cultural context adaptation bias—each type contains five quantized sub-features, then the total feature dimension is 4 × 5 = 20 dimensions. Next, feature reconstruction is performed, standardizing the values ​​of each sub-feature. For example, the original value range of a sub-feature of semantic annotation bias is 0-10. The standardized features are then converted to values ​​in the 0-1 range. Next, the standardized features are combined and enhanced by multiplying the highly correlated sub-features to generate new features. For example, the sub-feature of topic classification bias in topic matching error (0.3) is multiplied with the sub-feature of missing logical association in context logic (0.4) to obtain a new feature value of 0.12, which is then added to the original feature set to expand the feature dimension from 20 to 30. Finally, all standardized and enhanced features are arranged in a preset order to form a 30-dimensional high-dimensional representation vector, with each dimension value corresponding to a quantitative description of an error distribution pattern.

[0113] Step 662 involves analyzing the correlation characteristics of high-dimensional representation vectors over time and generating parameter correction values ​​using time series analysis. Specifically, this includes: collecting high-dimensional representation vectors from different time periods to form time series data; analyzing the temporal correlation characteristics; calculating the similarity between vectors from adjacent time periods (e.g., subtracting the corresponding dimension values ​​of the vector from day 1 and day 2, summing the absolute values, and then dividing by the dimension number (30) to obtain the similarity value, such as 0.15. A smaller value indicates a higher correlation. It also involves statistically analyzing the changing trends of the dimensions in the vectors, for example, the dimension value corresponding to the cultural context adaptation deviation from 0.3 on day 1. The value rose to 0.7 on day 7, indicating a continuous upward trend, and was marked as trend feature value 1. Time series analysis was used to generate parameter correction values. First, the core parameters requiring correction were determined, such as weight values ​​and threshold values ​​in classification calculations. For example, to address cultural context adaptation bias errors, the weight parameters of the cultural context prior vector needed correction. Correction rules were set based on temporal correlation characteristics. If a certain dimension feature showed a continuous upward trend and the similarity between adjacent vectors was less than 0.2, it was determined that the error accumulated over time, and the correction magnitude of the corresponding parameter needed to be increased. Specifically, the correction value was calculated as follows: assuming the original parameter value was 0.5, and the current value of this dimension feature was 0.7 (higher than the average level of 0.4), then the correction value = (current feature value - average feature value) × correction coefficient (preset to 0.1), i.e., (0.7 - 0.4) × 0.1 = 0.03. The corrected parameter value was 0.5 + 0.03 = 0.53. The above calculation was repeated for all parameters requiring correction to generate a set of parameter correction values, each corresponding to the adjustment value of a core parameter.

[0114] Step 663 involves using a multi-level feedback loop structure to progressively pass parameter corrections and dynamically integrate them into the core learning process of classification, updating internal parameters to form the final classification result. Specifically, the multi-level feedback loop structure comprises three levels: a feature layer, a computation layer, and a result layer. First, at the feature layer, the error feature weight corrections generated in step 662 are passed to the error feature extraction stage. For example, the weight of the semantic annotation bias feature is adjusted from 0.2 to 0.2 + 0.02 = 0.22 (correction amount 0.02), making subsequent error feature extraction focus more on this type of feature. Next, at the computation layer, the classification branch weight corrections are passed to the classification calculation process. For example, the weight of the copyright infringement branch is adjusted from 0.3 to 0.3 + 0.04 = 0.34 (correction amount 0.04), optimizing the calculation accuracy of the branch discrimination score. Finally, at the result layer, the confidence threshold is passed... The correction amount is then transferred to the result determination stage. For example, the high confidence threshold is adjusted from 80 points to 80-2=78 points (correction amount-2) to adapt to changes in error distribution. During each level of transmission, the correction amount is dynamically integrated. First, the fusion ratio between the correction amount and the current parameter is calculated, with the preset ratio of correction amount to 30% and original parameter to 70%. For example, if the original parameter is 0.5 and the correction amount is 0.03, then the fused parameter = 0.5×0.7+0.03×0.3=0.35+0.009=0.359. After integration, the internal parameters are updated, and the classification process in step 5 (from processing the classification feature vector to generating the classification result) is re-executed to obtain the updated classification result. This feedback loop is repeated 2-3 times until the parameter adjustment amplitude is less than the preset minimum threshold, such as 0.001. At this point, the internal parameters have adapted to the error distribution pattern, and the final generated classification result is the final classification result.

[0115] By screening error samples and uncovering their inherent patterns, ineffective analysis of error-free samples is avoided, thus improving error processing efficiency. Quantitative error features are extracted from semantic and contextual levels, reducing processing biases caused by feature ambiguity and enhancing the depth and accuracy of error analysis. Feature reconstruction and high-dimensional representation vector construction integrate scattered error features into structured vectors, clearly describing the error distribution patterns. Standardization and enhancement processes enrich feature dimensions, avoiding the limitations of single features and making the error distribution representation more comprehensive. Temporal correlation analysis captures the changing trends of error distribution over time, avoiding cumulative and correlated errors ignored by static analysis. Trend-based parameter correction quantities are targeted, accurately adjusting core error-related parameters to ensure the effectiveness of parameter adjustments and providing a key basis for improving classification accuracy. Multi-level feedback loops achieve layer-by-layer penetration of parameter correction quantities, ensuring that parameters throughout the feature extraction, calculation, and result determination process are adapted to the error distribution, avoiding overall imbalance caused by correction in a single stage.

[0116] like Figure 2As shown, embodiments of the present invention also provide an intelligent classification system for natural language text data based on deep learning, including:

[0117] The semantic restoration module is used to analyze the true semantics of target words using a context-aware mechanism for adversarial variants in the text. It combines a word sense library with a pre-training process using a dynamic learning rate adjustment strategy to generate a set of semantically consistent candidate replacement words.

[0118] The text standard module is used to perform multi-dimensional semantic similarity calculation and sentiment judgment based on the candidate replacement word set, determine the applicable words that conform to the original cultural background through the context adaptation strategy, and generate a standardized text sequence.

[0119] The feature extraction module is used to map word vectors to standardized text sequences and fuse local semantic information through multi-scale convolution operations to construct cross-granularity adversarial feature representations.

[0120] The context modeling module is used to obtain contextual dependencies based on cross-granularity adversarial feature representations and employ a context-sensitive self-attention mechanism to generate classification feature vectors carrying a measure of cultural context importance.

[0121] The classification decision module is used to input the classification feature vector into the multilingual adaptive classification process, complete the violation type identification, and generate classification results with credibility assessment.

[0122] The dynamic feedback module is used to analyze and reconstruct the features of the discrimination error samples based on the classification results, construct an error distribution representation, generate parameter correction quantities based on the temporal correlation characteristics of the error distribution, and dynamically integrate the parameter correction quantities into the learning process using a multi-level feedback loop structure to form the final classification result.

[0123] It should be noted that this system is a system corresponding to the above method. All implementation methods in the above method embodiments are applicable to this embodiment and can achieve the same technical effect.

[0124] Embodiments of the present invention also provide a computing device, including: a processor and a memory storing a computer program, wherein the computer program, when executed by the processor, performs the method described above. All implementations in the above method embodiments are applicable to this embodiment and can achieve the same technical effects.

[0125] Embodiments of the present invention also provide a computer-readable storage medium storing instructions that, when executed on a computer, cause the computer to perform the method described above. All implementations in the above method embodiments are applicable to this embodiment and can achieve the same technical effects.

[0126] The above description represents the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A deep learning-based intelligent classification method for natural language text data, characterized in that, The method includes: Step 1: For adversarial variants in the text, a context-aware mechanism is used to analyze the true semantics of the target words. Combined with the pre-training process of the word sense database and dynamic learning rate adjustment strategy, a set of semantically consistent candidate replacement words is generated. Step 2: Based on the candidate replacement word set, perform multi-dimensional semantic similarity calculation and sentiment judgment. Use a context-adaptation strategy to determine suitable words that conform to the original cultural background and generate a standardized text sequence. Specifically, this includes: calculating the similarity scores between each word in the candidate replacement word set and the target word in grammatical, semantic, and pragmatic dimensions; weighting and fusing the similarity scores according to preset weights to generate a comprehensive semantic similarity score; based on the comprehensive semantic similarity score, using a sentiment classifier to judge the sentiment tendency of the candidate words, calculating the sentiment polarity probability distribution, and comparing it with the sentiment polarity of the original text to select words with consistent sentiment tendencies, forming a sentiment-consistent candidate subset; based on the sentiment-consistent candidate subset, combining the cultural background and contextual features extracted from the original text, evaluating the context adaptability of the subset to obtain context adaptability evaluation results; and replacing the target words in the original text with the selected suitable words according to the context adaptability evaluation results to generate a grammatically correct, semantically coherent, and context-adapted standardized text sequence. Step 3: Map word vectors to the standardized text sequence, and fuse local semantic information through multi-scale convolution operations to construct cross-granularity adversarial feature representations; Step 4: Based on cross-granularity adversarial feature representation, a context-sensitive self-attention mechanism is used to obtain contextual dependencies and generate a classification feature vector carrying a measure of cultural context importance. Step 5: Input the classification feature vector into the multilingual adaptive classification process to complete the violation type identification and generate a classification result with a credibility assessment; Step 6: Based on the classification results, analyze and reconstruct the features of the discrimination error samples, construct an error distribution representation, generate parameter correction quantities based on the temporal correlation characteristics of the error distribution, and dynamically integrate the parameter correction quantities into the learning process using a multi-level feedback loop structure to form the final classification result.

2. The intelligent classification method for natural language text data based on deep learning according to claim 1, characterized in that, To address adversarial variants in the text, a context-aware mechanism is employed to analyze the true semantics of the target words. Combined with a pre-training process involving a semantic database and a dynamic learning rate adjustment strategy, a set of semantically consistent candidate replacement words is generated, including: The original input text is encoded by a pre-trained language processing mechanism to obtain the contextual semantic vector representation of the target words; Using contextual semantic vector representations as retrieval conditions, vector similarity matching is performed in the constructed word meaning resources to obtain an initial candidate word set associated with the target word; Based on the semantic distribution characteristics of the initial candidate vocabulary set, a dynamic learning rate adjustment strategy is adopted to perform supervised fine-tuning of the pre-trained language processing mechanism in order to obtain semantic representation capabilities. Based on semantic representation capabilities, the cosine similarity between each word in the initial candidate word set and the target word is calculated in a high-dimensional semantic space to generate a set of semantically consistent candidate replacement words.

3. The intelligent classification method for natural language text data based on deep learning according to claim 2, characterized in that, Word vector mapping is performed on standardized text sequences, and local semantic information is fused through multi-scale convolution operations to construct cross-granularity adversarial feature representations, including: Each word in the standardized text sequence is mapped to a high-dimensional distributed vector, and the corresponding word is generated. Vector sequence; The word vector sequence is input into a multi-scale convolutional neural network, and local semantic features of different spans in the text are extracted using convolutional kernels of different widths. Local semantic features are concatenated and dimensionality reduced to obtain a unified feature representation after fusion; Based on the unified feature representation after fusion, a cross-granularity adversarial feature representation is constructed to characterize the deep semantics and adversarial attributes of text.

4. The intelligent classification method for natural language text data based on deep learning according to claim 3, characterized in that, Based on cross-granularity adversarial feature representation, a context-sensitive self-attention mechanism is used to obtain contextual dependencies, generating classification feature vectors carrying a measure of cultural contextual importance, including: The cross-granularity adversarial feature vectors are transformed into query vectors, key vectors, and value vectors respectively through a linear projection layer; The cultural background features are linearly transformed to generate a cultural context prior vector, and then the cultural context prior vector is injected into the key vector to form a cultural context-enhanced key vector. Based on query vectors and cultural context-enhanced key vectors, the importance weight distribution of each word in a specific cultural context is calculated through a scaling dot product attention mechanism; The importance weight distribution is applied to the value vector, and a weighted summation is performed to generate a classification feature vector that includes semantic information and cultural context importance measurement.

5. The intelligent classification method for natural language text data based on deep learning according to claim 4, characterized in that, The classification feature vector is input into a multilingual adaptive classification process to complete violation type identification and generate classification results with confidence assessment, including: The classification feature vectors are processed to obtain initial discrimination results for natural language text data in various violation types; Based on the initial discrimination results, the probability distribution used to characterize the likelihood of belonging to each violation type is calculated; The probability distribution is calibrated to generate a calibrated probability distribution and a corresponding confidence assessment. Based on the calibrated probability distribution and the corresponding credibility assessment, classification results including violation type labels and credibility scores are determined and generated.

6. The intelligent classification method for natural language text data based on deep learning according to claim 5, characterized in that, Based on the classification results, the error samples are analyzed and their features reconstructed to construct an error distribution representation. Parameter correction values ​​are generated based on the temporal correlation characteristics of the error distribution, and a multi-level feedback loop structure is used to dynamically integrate these parameter correction values ​​into the learning process, forming the final classification result, including: The discrimination error samples in the initial classification results are analyzed to identify the inherent patterns and extract the error features of the discrimination error samples at the semantic and contextual levels. Based on error characteristics, a high-dimensional representation vector describing the error distribution pattern is constructed through a feature reconstruction process. The correlation characteristics of the evolution of high-dimensional representation vectors over time are analyzed, and parameter correction quantities are generated using time series analysis. Through a multi-level feedback loop structure, parameter corrections are passed layer by layer and dynamically integrated into the core learning process of classification to update internal parameters and form the final classification result.

7. A deep learning-based intelligent classification system for natural language text data, wherein the system implements the method as described in any one of claims 1 to 6, characterized in that, include: The semantic restoration module is used to analyze the true semantics of target words using a context-aware mechanism for adversarial variants in the text. It combines a word sense library with a pre-training process using a dynamic learning rate adjustment strategy to generate a set of semantically consistent candidate replacement words. The text standard module is used to perform multi-dimensional semantic similarity calculation and sentiment judgment based on the candidate replacement word set, determine the applicable words that conform to the original cultural background through the context adaptation strategy, and generate a standardized text sequence. The feature extraction module is used to map word vectors to standardized text sequences and fuse local semantic information through multi-scale convolution operations to construct cross-granularity adversarial feature representations. The context modeling module is used to obtain contextual dependencies based on cross-granularity adversarial feature representations and employ a context-sensitive self-attention mechanism to generate classification feature vectors carrying a measure of cultural context importance. The classification decision module is used to input the classification feature vector into the multilingual adaptive classification process, complete the violation type identification, and generate classification results with credibility assessment. The dynamic feedback module is used to analyze and reconstruct the features of the discrimination error samples based on the classification results, construct an error distribution representation, generate parameter correction quantities based on the temporal correlation characteristics of the error distribution, and dynamically integrate the parameter correction quantities into the learning process using a multi-level feedback loop structure to form the final classification result.

8. A computing device, characterized in that, include: One or more processors; A storage device for storing one or more programs that, when executed by one or more processors, cause the one or more processors to implement the method as described in any one of claims 1 to 6.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a program that, when executed by a processor, implements the method as described in any one of claims 1 to 6.

Images